Linux Audio

Check our new training course

Loading...
v6.2
   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 *       page->flags PG_locked (lock_page)
  27 *         hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share, see hugetlbfs below)
  28 *           mapping->i_mmap_rwsem
  29 *             anon_vma->rwsem
  30 *               mm->page_table_lock or pte_lock
  31 *                 swap_lock (in swap_duplicate, swap_info_get)
  32 *                   mmlist_lock (in mmput, drain_mmlist and others)
  33 *                   mapping->private_lock (in block_dirty_folio)
  34 *                     folio_lock_memcg move_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 *         page->flags PG_locked (lock_page)
  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
  78#include <asm/tlbflush.h>
  79
  80#define CREATE_TRACE_POINTS
  81#include <trace/events/tlb.h>
  82#include <trace/events/migrate.h>
  83
  84#include "internal.h"
  85
  86static struct kmem_cache *anon_vma_cachep;
  87static struct kmem_cache *anon_vma_chain_cachep;
  88
  89static inline struct anon_vma *anon_vma_alloc(void)
  90{
  91	struct anon_vma *anon_vma;
  92
  93	anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
  94	if (anon_vma) {
  95		atomic_set(&anon_vma->refcount, 1);
  96		anon_vma->num_children = 0;
  97		anon_vma->num_active_vmas = 0;
  98		anon_vma->parent = anon_vma;
  99		/*
 100		 * Initialise the anon_vma root to point to itself. If called
 101		 * from fork, the root will be reset to the parents anon_vma.
 102		 */
 103		anon_vma->root = anon_vma;
 104	}
 105
 106	return anon_vma;
 107}
 108
 109static inline void anon_vma_free(struct anon_vma *anon_vma)
 110{
 111	VM_BUG_ON(atomic_read(&anon_vma->refcount));
 112
 113	/*
 114	 * Synchronize against folio_lock_anon_vma_read() such that
 115	 * we can safely hold the lock without the anon_vma getting
 116	 * freed.
 117	 *
 118	 * Relies on the full mb implied by the atomic_dec_and_test() from
 119	 * put_anon_vma() against the acquire barrier implied by
 120	 * down_read_trylock() from folio_lock_anon_vma_read(). This orders:
 121	 *
 122	 * folio_lock_anon_vma_read()	VS	put_anon_vma()
 123	 *   down_read_trylock()		  atomic_dec_and_test()
 124	 *   LOCK				  MB
 125	 *   atomic_read()			  rwsem_is_locked()
 126	 *
 127	 * LOCK should suffice since the actual taking of the lock must
 128	 * happen _before_ what follows.
 129	 */
 130	might_sleep();
 131	if (rwsem_is_locked(&anon_vma->root->rwsem)) {
 132		anon_vma_lock_write(anon_vma);
 133		anon_vma_unlock_write(anon_vma);
 134	}
 135
 136	kmem_cache_free(anon_vma_cachep, anon_vma);
 137}
 138
 139static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
 140{
 141	return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
 142}
 143
 144static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
 145{
 146	kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
 147}
 148
 149static void anon_vma_chain_link(struct vm_area_struct *vma,
 150				struct anon_vma_chain *avc,
 151				struct anon_vma *anon_vma)
 152{
 153	avc->vma = vma;
 154	avc->anon_vma = anon_vma;
 155	list_add(&avc->same_vma, &vma->anon_vma_chain);
 156	anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
 157}
 158
 159/**
 160 * __anon_vma_prepare - attach an anon_vma to a memory region
 161 * @vma: the memory region in question
 162 *
 163 * This makes sure the memory mapping described by 'vma' has
 164 * an 'anon_vma' attached to it, so that we can associate the
 165 * anonymous pages mapped into it with that anon_vma.
 166 *
 167 * The common case will be that we already have one, which
 168 * is handled inline by anon_vma_prepare(). But if
 169 * not we either need to find an adjacent mapping that we
 170 * can re-use the anon_vma from (very common when the only
 171 * reason for splitting a vma has been mprotect()), or we
 172 * allocate a new one.
 173 *
 174 * Anon-vma allocations are very subtle, because we may have
 175 * optimistically looked up an anon_vma in folio_lock_anon_vma_read()
 176 * and that may actually touch the rwsem even in the newly
 177 * allocated vma (it depends on RCU to make sure that the
 178 * anon_vma isn't actually destroyed).
 179 *
 180 * As a result, we need to do proper anon_vma locking even
 181 * for the new allocation. At the same time, we do not want
 182 * to do any locking for the common case of already having
 183 * an anon_vma.
 184 *
 185 * This must be called with the mmap_lock held for reading.
 186 */
 187int __anon_vma_prepare(struct vm_area_struct *vma)
 188{
 189	struct mm_struct *mm = vma->vm_mm;
 190	struct anon_vma *anon_vma, *allocated;
 191	struct anon_vma_chain *avc;
 192
 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_adjust(), __split_vma(), copy_vma() and
 266 * anon_vma_fork(). The first three want an exact copy of src, while the last
 267 * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
 268 * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
 269 * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
 
 270 *
 271 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
 272 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
 273 * This prevents degradation of anon_vma hierarchy to endless linear chain in
 274 * case of constantly forking task. On the other hand, an anon_vma with more
 275 * than one child isn't reused even if there was no alive vma, thus rmap
 276 * walker has a good chance of avoiding scanning the whole hierarchy when it
 277 * searches where page is mapped.
 278 */
 279int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
 280{
 281	struct anon_vma_chain *avc, *pavc;
 282	struct anon_vma *root = NULL;
 283
 284	list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
 285		struct anon_vma *anon_vma;
 286
 287		avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
 288		if (unlikely(!avc)) {
 289			unlock_anon_vma_root(root);
 290			root = NULL;
 291			avc = anon_vma_chain_alloc(GFP_KERNEL);
 292			if (!avc)
 293				goto enomem_failure;
 294		}
 295		anon_vma = pavc->anon_vma;
 296		root = lock_anon_vma_root(root, anon_vma);
 297		anon_vma_chain_link(dst, avc, anon_vma);
 298
 299		/*
 300		 * Reuse existing anon_vma if it has no vma and only one
 301		 * anon_vma child.
 302		 *
 303		 * Root anon_vma is never reused:
 304		 * it has self-parent reference and at least one child.
 305		 */
 306		if (!dst->anon_vma && src->anon_vma &&
 307		    anon_vma->num_children < 2 &&
 308		    anon_vma->num_active_vmas == 0)
 309			dst->anon_vma = anon_vma;
 310	}
 311	if (dst->anon_vma)
 312		dst->anon_vma->num_active_vmas++;
 313	unlock_anon_vma_root(root);
 314	return 0;
 315
 316 enomem_failure:
 317	/*
 318	 * dst->anon_vma is dropped here otherwise its num_active_vmas can
 319	 * be incorrectly decremented in unlink_anon_vmas().
 320	 * We can safely do this because callers of anon_vma_clone() don't care
 321	 * about dst->anon_vma if anon_vma_clone() failed.
 322	 */
 323	dst->anon_vma = NULL;
 324	unlink_anon_vmas(dst);
 325	return -ENOMEM;
 326}
 327
 328/*
 329 * Attach vma to its own anon_vma, as well as to the anon_vmas that
 330 * the corresponding VMA in the parent process is attached to.
 331 * Returns 0 on success, non-zero on failure.
 332 */
 333int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
 334{
 335	struct anon_vma_chain *avc;
 336	struct anon_vma *anon_vma;
 337	int error;
 338
 339	/* Don't bother if the parent process has no anon_vma here. */
 340	if (!pvma->anon_vma)
 341		return 0;
 342
 343	/* Drop inherited anon_vma, we'll reuse existing or allocate new. */
 344	vma->anon_vma = NULL;
 345
 346	/*
 347	 * First, attach the new VMA to the parent VMA's anon_vmas,
 348	 * so rmap can find non-COWed pages in child processes.
 349	 */
 350	error = anon_vma_clone(vma, pvma);
 351	if (error)
 352		return error;
 353
 354	/* An existing anon_vma has been reused, all done then. */
 355	if (vma->anon_vma)
 356		return 0;
 357
 358	/* Then add our own anon_vma. */
 359	anon_vma = anon_vma_alloc();
 360	if (!anon_vma)
 361		goto out_error;
 362	anon_vma->num_active_vmas++;
 363	avc = anon_vma_chain_alloc(GFP_KERNEL);
 364	if (!avc)
 365		goto out_error_free_anon_vma;
 366
 367	/*
 368	 * The root anon_vma's rwsem is the lock actually used when we
 369	 * lock any of the anon_vmas in this anon_vma tree.
 370	 */
 371	anon_vma->root = pvma->anon_vma->root;
 372	anon_vma->parent = pvma->anon_vma;
 373	/*
 374	 * With refcounts, an anon_vma can stay around longer than the
 375	 * process it belongs to. The root anon_vma needs to be pinned until
 376	 * this anon_vma is freed, because the lock lives in the root.
 377	 */
 378	get_anon_vma(anon_vma->root);
 379	/* Mark this anon_vma as the one where our new (COWed) pages go. */
 380	vma->anon_vma = anon_vma;
 381	anon_vma_lock_write(anon_vma);
 382	anon_vma_chain_link(vma, avc, anon_vma);
 383	anon_vma->parent->num_children++;
 384	anon_vma_unlock_write(anon_vma);
 385
 386	return 0;
 387
 388 out_error_free_anon_vma:
 389	put_anon_vma(anon_vma);
 390 out_error:
 391	unlink_anon_vmas(vma);
 392	return -ENOMEM;
 393}
 394
 395void unlink_anon_vmas(struct vm_area_struct *vma)
 396{
 397	struct anon_vma_chain *avc, *next;
 398	struct anon_vma *root = NULL;
 399
 400	/*
 401	 * Unlink each anon_vma chained to the VMA.  This list is ordered
 402	 * from newest to oldest, ensuring the root anon_vma gets freed last.
 403	 */
 404	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 405		struct anon_vma *anon_vma = avc->anon_vma;
 406
 407		root = lock_anon_vma_root(root, anon_vma);
 408		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
 409
 410		/*
 411		 * Leave empty anon_vmas on the list - we'll need
 412		 * to free them outside the lock.
 413		 */
 414		if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
 415			anon_vma->parent->num_children--;
 416			continue;
 417		}
 418
 419		list_del(&avc->same_vma);
 420		anon_vma_chain_free(avc);
 421	}
 422	if (vma->anon_vma) {
 423		vma->anon_vma->num_active_vmas--;
 424
 425		/*
 426		 * vma would still be needed after unlink, and anon_vma will be prepared
 427		 * when handle fault.
 428		 */
 429		vma->anon_vma = NULL;
 430	}
 431	unlock_anon_vma_root(root);
 432
 433	/*
 434	 * Iterate the list once more, it now only contains empty and unlinked
 435	 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
 436	 * needing to write-acquire the anon_vma->root->rwsem.
 437	 */
 438	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 439		struct anon_vma *anon_vma = avc->anon_vma;
 440
 441		VM_WARN_ON(anon_vma->num_children);
 442		VM_WARN_ON(anon_vma->num_active_vmas);
 443		put_anon_vma(anon_vma);
 444
 445		list_del(&avc->same_vma);
 446		anon_vma_chain_free(avc);
 447	}
 448}
 449
 450static void anon_vma_ctor(void *data)
 451{
 452	struct anon_vma *anon_vma = data;
 453
 454	init_rwsem(&anon_vma->rwsem);
 455	atomic_set(&anon_vma->refcount, 0);
 456	anon_vma->rb_root = RB_ROOT_CACHED;
 457}
 458
 459void __init anon_vma_init(void)
 460{
 461	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
 462			0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
 463			anon_vma_ctor);
 464	anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
 465			SLAB_PANIC|SLAB_ACCOUNT);
 466}
 467
 468/*
 469 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
 470 *
 471 * Since there is no serialization what so ever against page_remove_rmap()
 472 * the best this function can do is return a refcount increased anon_vma
 473 * that might have been relevant to this page.
 474 *
 475 * The page might have been remapped to a different anon_vma or the anon_vma
 476 * returned may already be freed (and even reused).
 477 *
 478 * In case it was remapped to a different anon_vma, the new anon_vma will be a
 479 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
 480 * ensure that any anon_vma obtained from the page will still be valid for as
 481 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
 482 *
 483 * All users of this function must be very careful when walking the anon_vma
 484 * chain and verify that the page in question is indeed mapped in it
 485 * [ something equivalent to page_mapped_in_vma() ].
 486 *
 487 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
 488 * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
 489 * if there is a mapcount, we can dereference the anon_vma after observing
 490 * those.
 
 
 
 
 
 
 491 */
 492struct anon_vma *folio_get_anon_vma(struct folio *folio)
 493{
 494	struct anon_vma *anon_vma = NULL;
 495	unsigned long anon_mapping;
 496
 497	rcu_read_lock();
 498	anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
 499	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 500		goto out;
 501	if (!folio_mapped(folio))
 502		goto out;
 503
 504	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 505	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 506		anon_vma = NULL;
 507		goto out;
 508	}
 509
 510	/*
 511	 * If this folio is still mapped, then its anon_vma cannot have been
 512	 * freed.  But if it has been unmapped, we have no security against the
 513	 * anon_vma structure being freed and reused (for another anon_vma:
 514	 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
 515	 * above cannot corrupt).
 516	 */
 517	if (!folio_mapped(folio)) {
 518		rcu_read_unlock();
 519		put_anon_vma(anon_vma);
 520		return NULL;
 521	}
 522out:
 523	rcu_read_unlock();
 524
 525	return anon_vma;
 526}
 527
 528/*
 529 * Similar to folio_get_anon_vma() except it locks the anon_vma.
 530 *
 531 * Its a little more complex as it tries to keep the fast path to a single
 532 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
 533 * reference like with folio_get_anon_vma() and then block on the mutex
 534 * on !rwc->try_lock case.
 535 */
 536struct anon_vma *folio_lock_anon_vma_read(struct folio *folio,
 537					  struct rmap_walk_control *rwc)
 538{
 539	struct anon_vma *anon_vma = NULL;
 540	struct anon_vma *root_anon_vma;
 541	unsigned long anon_mapping;
 542
 
 543	rcu_read_lock();
 544	anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
 545	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 546		goto out;
 547	if (!folio_mapped(folio))
 548		goto out;
 549
 550	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 551	root_anon_vma = READ_ONCE(anon_vma->root);
 552	if (down_read_trylock(&root_anon_vma->rwsem)) {
 553		/*
 
 
 
 
 
 
 
 
 
 
 
 554		 * If the folio is still mapped, then this anon_vma is still
 555		 * its anon_vma, and holding the mutex ensures that it will
 556		 * not go away, see anon_vma_free().
 557		 */
 558		if (!folio_mapped(folio)) {
 559			up_read(&root_anon_vma->rwsem);
 560			anon_vma = NULL;
 561		}
 562		goto out;
 563	}
 564
 565	if (rwc && rwc->try_lock) {
 566		anon_vma = NULL;
 567		rwc->contended = true;
 568		goto out;
 569	}
 570
 571	/* trylock failed, we got to sleep */
 572	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 573		anon_vma = NULL;
 574		goto out;
 575	}
 576
 577	if (!folio_mapped(folio)) {
 578		rcu_read_unlock();
 579		put_anon_vma(anon_vma);
 580		return NULL;
 581	}
 582
 583	/* we pinned the anon_vma, its safe to sleep */
 584	rcu_read_unlock();
 585	anon_vma_lock_read(anon_vma);
 586
 
 
 
 
 
 
 
 
 
 
 
 
 587	if (atomic_dec_and_test(&anon_vma->refcount)) {
 588		/*
 589		 * Oops, we held the last refcount, release the lock
 590		 * and bail -- can't simply use put_anon_vma() because
 591		 * we'll deadlock on the anon_vma_lock_write() recursion.
 592		 */
 593		anon_vma_unlock_read(anon_vma);
 594		__put_anon_vma(anon_vma);
 595		anon_vma = NULL;
 596	}
 597
 598	return anon_vma;
 599
 600out:
 601	rcu_read_unlock();
 602	return anon_vma;
 603}
 604
 605#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
 606/*
 607 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
 608 * important if a PTE was dirty when it was unmapped that it's flushed
 609 * before any IO is initiated on the page to prevent lost writes. Similarly,
 610 * it must be flushed before freeing to prevent data leakage.
 611 */
 612void try_to_unmap_flush(void)
 613{
 614	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 615
 616	if (!tlb_ubc->flush_required)
 617		return;
 618
 619	arch_tlbbatch_flush(&tlb_ubc->arch);
 620	tlb_ubc->flush_required = false;
 621	tlb_ubc->writable = false;
 622}
 623
 624/* Flush iff there are potentially writable TLB entries that can race with IO */
 625void try_to_unmap_flush_dirty(void)
 626{
 627	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 628
 629	if (tlb_ubc->writable)
 630		try_to_unmap_flush();
 631}
 632
 633/*
 634 * Bits 0-14 of mm->tlb_flush_batched record pending generations.
 635 * Bits 16-30 of mm->tlb_flush_batched bit record flushed generations.
 636 */
 637#define TLB_FLUSH_BATCH_FLUSHED_SHIFT	16
 638#define TLB_FLUSH_BATCH_PENDING_MASK			\
 639	((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT - 1)) - 1)
 640#define TLB_FLUSH_BATCH_PENDING_LARGE			\
 641	(TLB_FLUSH_BATCH_PENDING_MASK / 2)
 642
 643static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
 
 644{
 645	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 646	int batch, nbatch;
 
 
 
 
 647
 648	arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
 649	tlb_ubc->flush_required = true;
 650
 651	/*
 652	 * Ensure compiler does not re-order the setting of tlb_flush_batched
 653	 * before the PTE is cleared.
 654	 */
 655	barrier();
 656	batch = atomic_read(&mm->tlb_flush_batched);
 657retry:
 658	if ((batch & TLB_FLUSH_BATCH_PENDING_MASK) > TLB_FLUSH_BATCH_PENDING_LARGE) {
 659		/*
 660		 * Prevent `pending' from catching up with `flushed' because of
 661		 * overflow.  Reset `pending' and `flushed' to be 1 and 0 if
 662		 * `pending' becomes large.
 663		 */
 664		nbatch = atomic_cmpxchg(&mm->tlb_flush_batched, batch, 1);
 665		if (nbatch != batch) {
 666			batch = nbatch;
 667			goto retry;
 668		}
 669	} else {
 670		atomic_inc(&mm->tlb_flush_batched);
 671	}
 672
 673	/*
 674	 * If the PTE was dirty then it's best to assume it's writable. The
 675	 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
 676	 * before the page is queued for IO.
 677	 */
 678	if (writable)
 679		tlb_ubc->writable = true;
 680}
 681
 682/*
 683 * Returns true if the TLB flush should be deferred to the end of a batch of
 684 * unmap operations to reduce IPIs.
 685 */
 686static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
 687{
 688	bool should_defer = false;
 689
 690	if (!(flags & TTU_BATCH_FLUSH))
 691		return false;
 692
 693	/* If remote CPUs need to be flushed then defer batch the flush */
 694	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
 695		should_defer = true;
 696	put_cpu();
 697
 698	return should_defer;
 699}
 700
 701/*
 702 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
 703 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
 704 * operation such as mprotect or munmap to race between reclaim unmapping
 705 * the page and flushing the page. If this race occurs, it potentially allows
 706 * access to data via a stale TLB entry. Tracking all mm's that have TLB
 707 * batching in flight would be expensive during reclaim so instead track
 708 * whether TLB batching occurred in the past and if so then do a flush here
 709 * if required. This will cost one additional flush per reclaim cycle paid
 710 * by the first operation at risk such as mprotect and mumap.
 711 *
 712 * This must be called under the PTL so that an access to tlb_flush_batched
 713 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
 714 * via the PTL.
 715 */
 716void flush_tlb_batched_pending(struct mm_struct *mm)
 717{
 718	int batch = atomic_read(&mm->tlb_flush_batched);
 719	int pending = batch & TLB_FLUSH_BATCH_PENDING_MASK;
 720	int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT;
 721
 722	if (pending != flushed) {
 723		flush_tlb_mm(mm);
 724		/*
 725		 * If the new TLB flushing is pending during flushing, leave
 726		 * mm->tlb_flush_batched as is, to avoid losing flushing.
 727		 */
 728		atomic_cmpxchg(&mm->tlb_flush_batched, batch,
 729			       pending | (pending << TLB_FLUSH_BATCH_FLUSHED_SHIFT));
 730	}
 731}
 732#else
 733static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
 
 734{
 735}
 736
 737static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
 738{
 739	return false;
 740}
 741#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
 742
 743/*
 744 * At what user virtual address is page expected in vma?
 745 * Caller should check the page is actually part of the vma.
 746 */
 747unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
 748{
 749	struct folio *folio = page_folio(page);
 750	if (folio_test_anon(folio)) {
 751		struct anon_vma *page__anon_vma = folio_anon_vma(folio);
 752		/*
 753		 * Note: swapoff's unuse_vma() is more efficient with this
 754		 * check, and needs it to match anon_vma when KSM is active.
 755		 */
 756		if (!vma->anon_vma || !page__anon_vma ||
 757		    vma->anon_vma->root != page__anon_vma->root)
 758			return -EFAULT;
 759	} else if (!vma->vm_file) {
 760		return -EFAULT;
 761	} else if (vma->vm_file->f_mapping != folio->mapping) {
 762		return -EFAULT;
 763	}
 764
 765	return vma_address(page, vma);
 766}
 767
 768/*
 769 * Returns the actual pmd_t* where we expect 'address' to be mapped from, or
 770 * NULL if it doesn't exist.  No guarantees / checks on what the pmd_t*
 771 * represents.
 772 */
 773pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
 774{
 775	pgd_t *pgd;
 776	p4d_t *p4d;
 777	pud_t *pud;
 778	pmd_t *pmd = NULL;
 779
 780	pgd = pgd_offset(mm, address);
 781	if (!pgd_present(*pgd))
 782		goto out;
 783
 784	p4d = p4d_offset(pgd, address);
 785	if (!p4d_present(*p4d))
 786		goto out;
 787
 788	pud = pud_offset(p4d, address);
 789	if (!pud_present(*pud))
 790		goto out;
 791
 792	pmd = pmd_offset(pud, address);
 793out:
 794	return pmd;
 795}
 796
 797struct folio_referenced_arg {
 798	int mapcount;
 799	int referenced;
 800	unsigned long vm_flags;
 801	struct mem_cgroup *memcg;
 802};
 
 803/*
 804 * arg: folio_referenced_arg will be passed
 805 */
 806static bool folio_referenced_one(struct folio *folio,
 807		struct vm_area_struct *vma, unsigned long address, void *arg)
 808{
 809	struct folio_referenced_arg *pra = arg;
 810	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
 811	int referenced = 0;
 
 812
 813	while (page_vma_mapped_walk(&pvmw)) {
 814		address = pvmw.address;
 815
 816		if ((vma->vm_flags & VM_LOCKED) &&
 817		    (!folio_test_large(folio) || !pvmw.pte)) {
 818			/* Restore the mlock which got missed */
 819			mlock_vma_folio(folio, vma, !pvmw.pte);
 820			page_vma_mapped_walk_done(&pvmw);
 821			pra->vm_flags |= VM_LOCKED;
 822			return false; /* To break the loop */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 823		}
 824
 825		if (pvmw.pte) {
 826			if (lru_gen_enabled() && pte_young(*pvmw.pte) &&
 827			    !(vma->vm_flags & (VM_SEQ_READ | VM_RAND_READ))) {
 828				lru_gen_look_around(&pvmw);
 829				referenced++;
 830			}
 831
 832			if (ptep_clear_flush_young_notify(vma, address,
 833						pvmw.pte)) {
 834				/*
 835				 * Don't treat a reference through
 836				 * a sequentially read mapping as such.
 837				 * If the folio has been used in another mapping,
 838				 * we will catch it; if this other mapping is
 839				 * already gone, the unmap path will have set
 840				 * the referenced flag or activated the folio.
 841				 */
 842				if (likely(!(vma->vm_flags & VM_SEQ_READ)))
 843					referenced++;
 844			}
 845		} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
 846			if (pmdp_clear_flush_young_notify(vma, address,
 847						pvmw.pmd))
 848				referenced++;
 849		} else {
 850			/* unexpected pmd-mapped folio? */
 851			WARN_ON_ONCE(1);
 852		}
 853
 854		pra->mapcount--;
 855	}
 856
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 857	if (referenced)
 858		folio_clear_idle(folio);
 859	if (folio_test_clear_young(folio))
 860		referenced++;
 861
 862	if (referenced) {
 863		pra->referenced++;
 864		pra->vm_flags |= vma->vm_flags & ~VM_LOCKED;
 865	}
 866
 867	if (!pra->mapcount)
 868		return false; /* To break the loop */
 869
 870	return true;
 871}
 872
 873static bool invalid_folio_referenced_vma(struct vm_area_struct *vma, void *arg)
 874{
 875	struct folio_referenced_arg *pra = arg;
 876	struct mem_cgroup *memcg = pra->memcg;
 877
 878	if (!mm_match_cgroup(vma->vm_mm, memcg))
 
 
 
 
 
 
 
 
 
 
 
 
 
 879		return true;
 880
 881	return false;
 882}
 883
 884/**
 885 * folio_referenced() - Test if the folio was referenced.
 886 * @folio: The folio to test.
 887 * @is_locked: Caller holds lock on the folio.
 888 * @memcg: target memory cgroup
 889 * @vm_flags: A combination of all the vma->vm_flags which referenced the folio.
 890 *
 891 * Quick test_and_clear_referenced for all mappings of a folio,
 892 *
 893 * Return: The number of mappings which referenced the folio. Return -1 if
 894 * the function bailed out due to rmap lock contention.
 895 */
 896int folio_referenced(struct folio *folio, int is_locked,
 897		     struct mem_cgroup *memcg, unsigned long *vm_flags)
 898{
 899	int we_locked = 0;
 900	struct folio_referenced_arg pra = {
 901		.mapcount = folio_mapcount(folio),
 902		.memcg = memcg,
 903	};
 904	struct rmap_walk_control rwc = {
 905		.rmap_one = folio_referenced_one,
 906		.arg = (void *)&pra,
 907		.anon_lock = folio_lock_anon_vma_read,
 908		.try_lock = true,
 
 909	};
 910
 911	*vm_flags = 0;
 912	if (!pra.mapcount)
 913		return 0;
 914
 915	if (!folio_raw_mapping(folio))
 916		return 0;
 917
 918	if (!is_locked && (!folio_test_anon(folio) || folio_test_ksm(folio))) {
 919		we_locked = folio_trylock(folio);
 920		if (!we_locked)
 921			return 1;
 922	}
 923
 924	/*
 925	 * If we are reclaiming on behalf of a cgroup, skip
 926	 * counting on behalf of references from different
 927	 * cgroups
 928	 */
 929	if (memcg) {
 930		rwc.invalid_vma = invalid_folio_referenced_vma;
 931	}
 932
 933	rmap_walk(folio, &rwc);
 934	*vm_flags = pra.vm_flags;
 935
 936	if (we_locked)
 937		folio_unlock(folio);
 938
 939	return rwc.contended ? -1 : pra.referenced;
 940}
 941
 942static int page_vma_mkclean_one(struct page_vma_mapped_walk *pvmw)
 943{
 944	int cleaned = 0;
 945	struct vm_area_struct *vma = pvmw->vma;
 946	struct mmu_notifier_range range;
 947	unsigned long address = pvmw->address;
 948
 949	/*
 950	 * We have to assume the worse case ie pmd for invalidation. Note that
 951	 * the folio can not be freed from this function.
 952	 */
 953	mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
 954				0, vma, vma->vm_mm, address,
 955				vma_address_end(pvmw));
 956	mmu_notifier_invalidate_range_start(&range);
 957
 958	while (page_vma_mapped_walk(pvmw)) {
 959		int ret = 0;
 960
 961		address = pvmw->address;
 962		if (pvmw->pte) {
 963			pte_t entry;
 964			pte_t *pte = pvmw->pte;
 
 965
 966			if (!pte_dirty(*pte) && !pte_write(*pte))
 967				continue;
 968
 969			flush_cache_page(vma, address, pte_pfn(*pte));
 970			entry = ptep_clear_flush(vma, address, pte);
 971			entry = pte_wrprotect(entry);
 972			entry = pte_mkclean(entry);
 973			set_pte_at(vma->vm_mm, address, pte, entry);
 974			ret = 1;
 975		} else {
 976#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 977			pmd_t *pmd = pvmw->pmd;
 978			pmd_t entry;
 979
 980			if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
 981				continue;
 982
 983			flush_cache_range(vma, address,
 984					  address + HPAGE_PMD_SIZE);
 985			entry = pmdp_invalidate(vma, address, pmd);
 986			entry = pmd_wrprotect(entry);
 987			entry = pmd_mkclean(entry);
 988			set_pmd_at(vma->vm_mm, address, pmd, entry);
 989			ret = 1;
 990#else
 991			/* unexpected pmd-mapped folio? */
 992			WARN_ON_ONCE(1);
 993#endif
 994		}
 995
 996		/*
 997		 * No need to call mmu_notifier_invalidate_range() as we are
 998		 * downgrading page table protection not changing it to point
 999		 * to a new page.
1000		 *
1001		 * See Documentation/mm/mmu_notifier.rst
1002		 */
1003		if (ret)
1004			cleaned++;
1005	}
1006
1007	mmu_notifier_invalidate_range_end(&range);
1008
1009	return cleaned;
1010}
1011
1012static bool page_mkclean_one(struct folio *folio, struct vm_area_struct *vma,
1013			     unsigned long address, void *arg)
1014{
1015	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC);
1016	int *cleaned = arg;
1017
1018	*cleaned += page_vma_mkclean_one(&pvmw);
1019
1020	return true;
1021}
1022
1023static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1024{
1025	if (vma->vm_flags & VM_SHARED)
1026		return false;
1027
1028	return true;
1029}
1030
1031int folio_mkclean(struct folio *folio)
1032{
1033	int cleaned = 0;
1034	struct address_space *mapping;
1035	struct rmap_walk_control rwc = {
1036		.arg = (void *)&cleaned,
1037		.rmap_one = page_mkclean_one,
1038		.invalid_vma = invalid_mkclean_vma,
1039	};
1040
1041	BUG_ON(!folio_test_locked(folio));
1042
1043	if (!folio_mapped(folio))
1044		return 0;
1045
1046	mapping = folio_mapping(folio);
1047	if (!mapping)
1048		return 0;
1049
1050	rmap_walk(folio, &rwc);
1051
1052	return cleaned;
1053}
1054EXPORT_SYMBOL_GPL(folio_mkclean);
1055
1056/**
1057 * pfn_mkclean_range - Cleans the PTEs (including PMDs) mapped with range of
1058 *                     [@pfn, @pfn + @nr_pages) at the specific offset (@pgoff)
1059 *                     within the @vma of shared mappings. And since clean PTEs
1060 *                     should also be readonly, write protects them too.
1061 * @pfn: start pfn.
1062 * @nr_pages: number of physically contiguous pages srarting with @pfn.
1063 * @pgoff: page offset that the @pfn mapped with.
1064 * @vma: vma that @pfn mapped within.
1065 *
1066 * Returns the number of cleaned PTEs (including PMDs).
1067 */
1068int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff,
1069		      struct vm_area_struct *vma)
1070{
1071	struct page_vma_mapped_walk pvmw = {
1072		.pfn		= pfn,
1073		.nr_pages	= nr_pages,
1074		.pgoff		= pgoff,
1075		.vma		= vma,
1076		.flags		= PVMW_SYNC,
1077	};
1078
1079	if (invalid_mkclean_vma(vma, NULL))
1080		return 0;
1081
1082	pvmw.address = vma_pgoff_address(pgoff, nr_pages, vma);
1083	VM_BUG_ON_VMA(pvmw.address == -EFAULT, vma);
1084
1085	return page_vma_mkclean_one(&pvmw);
1086}
1087
1088int total_compound_mapcount(struct page *head)
1089{
1090	int mapcount = head_compound_mapcount(head);
1091	int nr_subpages;
1092	int i;
1093
1094	/* In the common case, avoid the loop when no subpages mapped by PTE */
1095	if (head_subpages_mapcount(head) == 0)
1096		return mapcount;
1097	/*
1098	 * Add all the PTE mappings of those subpages mapped by PTE.
1099	 * Limit the loop, knowing that only subpages_mapcount are mapped?
1100	 * Perhaps: given all the raciness, that may be a good or a bad idea.
1101	 */
1102	nr_subpages = thp_nr_pages(head);
1103	for (i = 0; i < nr_subpages; i++)
1104		mapcount += atomic_read(&head[i]._mapcount);
1105
1106	/* But each of those _mapcounts was based on -1 */
1107	mapcount += nr_subpages;
1108	return mapcount;
1109}
1110
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1111/**
1112 * page_move_anon_rmap - move a page to our anon_vma
1113 * @page:	the page to move to our anon_vma
1114 * @vma:	the vma the page belongs to
1115 *
1116 * When a page belongs exclusively to one process after a COW event,
1117 * that page can be moved into the anon_vma that belongs to just that
1118 * process, so the rmap code will not search the parent or sibling
1119 * processes.
1120 */
1121void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1122{
1123	void *anon_vma = vma->anon_vma;
1124	struct folio *folio = page_folio(page);
1125
1126	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1127	VM_BUG_ON_VMA(!anon_vma, vma);
1128
1129	anon_vma += PAGE_MAPPING_ANON;
1130	/*
1131	 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1132	 * simultaneously, so a concurrent reader (eg folio_referenced()'s
1133	 * folio_test_anon()) will not see one without the other.
1134	 */
1135	WRITE_ONCE(folio->mapping, anon_vma);
1136	SetPageAnonExclusive(page);
1137}
1138
1139/**
1140 * __page_set_anon_rmap - set up new anonymous rmap
1141 * @page:	Page or Hugepage to add to rmap
1142 * @vma:	VM area to add page to.
1143 * @address:	User virtual address of the mapping	
1144 * @exclusive:	the page is exclusively owned by the current process
1145 */
1146static void __page_set_anon_rmap(struct page *page,
1147	struct vm_area_struct *vma, unsigned long address, int exclusive)
1148{
1149	struct anon_vma *anon_vma = vma->anon_vma;
1150
1151	BUG_ON(!anon_vma);
1152
1153	if (PageAnon(page))
1154		goto out;
1155
1156	/*
1157	 * If the page isn't exclusively mapped into this vma,
1158	 * we must use the _oldest_ possible anon_vma for the
1159	 * page mapping!
1160	 */
1161	if (!exclusive)
1162		anon_vma = anon_vma->root;
1163
1164	/*
1165	 * page_idle does a lockless/optimistic rmap scan on page->mapping.
1166	 * Make sure the compiler doesn't split the stores of anon_vma and
1167	 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1168	 * could mistake the mapping for a struct address_space and crash.
1169	 */
1170	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1171	WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1172	page->index = linear_page_index(vma, address);
1173out:
1174	if (exclusive)
1175		SetPageAnonExclusive(page);
1176}
1177
1178/**
1179 * __page_check_anon_rmap - sanity check anonymous rmap addition
1180 * @page:	the page to add the mapping to
 
1181 * @vma:	the vm area in which the mapping is added
1182 * @address:	the user virtual address mapped
1183 */
1184static void __page_check_anon_rmap(struct page *page,
1185	struct vm_area_struct *vma, unsigned long address)
1186{
1187	struct folio *folio = page_folio(page);
1188	/*
1189	 * The page's anon-rmap details (mapping and index) are guaranteed to
1190	 * be set up correctly at this point.
1191	 *
1192	 * We have exclusion against page_add_anon_rmap because the caller
1193	 * always holds the page locked.
1194	 *
1195	 * We have exclusion against page_add_new_anon_rmap because those pages
1196	 * are initially only visible via the pagetables, and the pte is locked
1197	 * over the call to page_add_new_anon_rmap.
1198	 */
1199	VM_BUG_ON_FOLIO(folio_anon_vma(folio)->root != vma->anon_vma->root,
1200			folio);
1201	VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1202		       page);
1203}
1204
1205/**
1206 * page_add_anon_rmap - add pte mapping to an anonymous page
1207 * @page:	the page to add the mapping to
1208 * @vma:	the vm area in which the mapping is added
1209 * @address:	the user virtual address mapped
1210 * @flags:	the rmap flags
1211 *
1212 * The caller needs to hold the pte lock, and the page must be locked in
1213 * the anon_vma case: to serialize mapping,index checking after setting,
1214 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1215 * (but PageKsm is never downgraded to PageAnon).
1216 */
1217void page_add_anon_rmap(struct page *page,
1218	struct vm_area_struct *vma, unsigned long address, rmap_t flags)
1219{
1220	atomic_t *mapped;
1221	int nr = 0, nr_pmdmapped = 0;
1222	bool compound = flags & RMAP_COMPOUND;
1223	bool first = true;
1224
1225	if (unlikely(PageKsm(page)))
1226		lock_page_memcg(page);
1227
1228	/* Is page being mapped by PTE? Is this its first map to be added? */
1229	if (likely(!compound)) {
1230		first = atomic_inc_and_test(&page->_mapcount);
1231		nr = first;
1232		if (first && PageCompound(page)) {
1233			mapped = subpages_mapcount_ptr(compound_head(page));
1234			nr = atomic_inc_return_relaxed(mapped);
1235			nr = (nr < COMPOUND_MAPPED);
1236		}
1237	} else if (PageTransHuge(page)) {
1238		/* That test is redundant: it's for safety or to optimize out */
1239
1240		first = atomic_inc_and_test(compound_mapcount_ptr(page));
1241		if (first) {
1242			mapped = subpages_mapcount_ptr(page);
1243			nr = atomic_add_return_relaxed(COMPOUND_MAPPED, mapped);
1244			if (likely(nr < COMPOUND_MAPPED + COMPOUND_MAPPED)) {
1245				nr_pmdmapped = thp_nr_pages(page);
1246				nr = nr_pmdmapped - (nr & SUBPAGES_MAPPED);
1247				/* Raced ahead of a remove and another add? */
1248				if (unlikely(nr < 0))
1249					nr = 0;
1250			} else {
1251				/* Raced ahead of a remove of COMPOUND_MAPPED */
1252				nr = 0;
1253			}
1254		}
 
 
 
 
 
1255	}
1256
1257	VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page);
1258	VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page);
 
 
 
 
 
 
 
 
 
 
 
1259
1260	if (nr_pmdmapped)
1261		__mod_lruvec_page_state(page, NR_ANON_THPS, nr_pmdmapped);
1262	if (nr)
1263		__mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
 
 
1264
1265	if (unlikely(PageKsm(page)))
1266		unlock_page_memcg(page);
 
 
 
 
 
 
 
1267
1268	/* address might be in next vma when migration races vma_adjust */
1269	else if (first)
1270		__page_set_anon_rmap(page, vma, address,
1271				     !!(flags & RMAP_EXCLUSIVE));
1272	else
1273		__page_check_anon_rmap(page, vma, address);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1274
1275	mlock_vma_page(page, vma, compound);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1276}
1277
1278/**
1279 * page_add_new_anon_rmap - add mapping to a new anonymous page
1280 * @page:	the page to add the mapping to
1281 * @vma:	the vm area in which the mapping is added
1282 * @address:	the user virtual address mapped
1283 *
1284 * If it's a compound page, it is accounted as a compound page. As the page
1285 * is new, it's assume to get mapped exclusively by a single process.
1286 *
1287 * Same as page_add_anon_rmap but must only be called on *new* pages.
1288 * This means the inc-and-test can be bypassed.
1289 * Page does not have to be locked.
 
 
 
1290 */
1291void page_add_new_anon_rmap(struct page *page,
1292	struct vm_area_struct *vma, unsigned long address)
1293{
1294	int nr;
1295
1296	VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1297	__SetPageSwapBacked(page);
 
 
 
1298
1299	if (likely(!PageCompound(page))) {
1300		/* increment count (starts at -1) */
1301		atomic_set(&page->_mapcount, 0);
1302		nr = 1;
 
 
 
 
 
 
 
 
 
 
 
 
1303	} else {
1304		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1305		/* increment count (starts at -1) */
1306		atomic_set(compound_mapcount_ptr(page), 0);
1307		atomic_set(subpages_mapcount_ptr(page), COMPOUND_MAPPED);
1308		nr = thp_nr_pages(page);
1309		__mod_lruvec_page_state(page, NR_ANON_THPS, nr);
1310	}
1311
1312	__mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1313	__page_set_anon_rmap(page, vma, address, 1);
1314}
1315
1316/**
1317 * page_add_file_rmap - add pte mapping to a file page
1318 * @page:	the page to add the mapping to
1319 * @vma:	the vm area in which the mapping is added
1320 * @compound:	charge the page as compound or small page
1321 *
1322 * The caller needs to hold the pte lock.
1323 */
1324void page_add_file_rmap(struct page *page,
1325	struct vm_area_struct *vma, bool compound)
1326{
1327	atomic_t *mapped;
1328	int nr = 0, nr_pmdmapped = 0;
1329	bool first;
1330
1331	VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1332	lock_page_memcg(page);
1333
1334	/* Is page being mapped by PTE? Is this its first map to be added? */
1335	if (likely(!compound)) {
1336		first = atomic_inc_and_test(&page->_mapcount);
1337		nr = first;
1338		if (first && PageCompound(page)) {
1339			mapped = subpages_mapcount_ptr(compound_head(page));
1340			nr = atomic_inc_return_relaxed(mapped);
1341			nr = (nr < COMPOUND_MAPPED);
1342		}
1343	} else if (PageTransHuge(page)) {
1344		/* That test is redundant: it's for safety or to optimize out */
1345
1346		first = atomic_inc_and_test(compound_mapcount_ptr(page));
1347		if (first) {
1348			mapped = subpages_mapcount_ptr(page);
1349			nr = atomic_add_return_relaxed(COMPOUND_MAPPED, mapped);
1350			if (likely(nr < COMPOUND_MAPPED + COMPOUND_MAPPED)) {
1351				nr_pmdmapped = thp_nr_pages(page);
1352				nr = nr_pmdmapped - (nr & SUBPAGES_MAPPED);
1353				/* Raced ahead of a remove and another add? */
1354				if (unlikely(nr < 0))
1355					nr = 0;
1356			} else {
1357				/* Raced ahead of a remove of COMPOUND_MAPPED */
1358				nr = 0;
1359			}
1360		}
1361	}
1362
 
1363	if (nr_pmdmapped)
1364		__mod_lruvec_page_state(page, PageSwapBacked(page) ?
1365			NR_SHMEM_PMDMAPPED : NR_FILE_PMDMAPPED, nr_pmdmapped);
1366	if (nr)
1367		__mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1368	unlock_page_memcg(page);
1369
1370	mlock_vma_page(page, vma, compound);
 
 
1371}
1372
1373/**
1374 * page_remove_rmap - take down pte mapping from a page
1375 * @page:	page to remove mapping from
1376 * @vma:	the vm area from which the mapping is removed
1377 * @compound:	uncharge the page as compound or small page
1378 *
1379 * The caller needs to hold the pte lock.
1380 */
1381void page_remove_rmap(struct page *page,
1382	struct vm_area_struct *vma, bool compound)
1383{
1384	atomic_t *mapped;
1385	int nr = 0, nr_pmdmapped = 0;
1386	bool last;
1387
1388	VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1389
1390	/* Hugetlb pages are not counted in NR_*MAPPED */
1391	if (unlikely(PageHuge(page))) {
1392		/* hugetlb pages are always mapped with pmds */
1393		atomic_dec(compound_mapcount_ptr(page));
1394		return;
1395	}
1396
1397	lock_page_memcg(page);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1398
1399	/* Is page being unmapped by PTE? Is this its last map to be removed? */
1400	if (likely(!compound)) {
1401		last = atomic_add_negative(-1, &page->_mapcount);
1402		nr = last;
1403		if (last && PageCompound(page)) {
1404			mapped = subpages_mapcount_ptr(compound_head(page));
1405			nr = atomic_dec_return_relaxed(mapped);
1406			nr = (nr < COMPOUND_MAPPED);
1407		}
1408	} else if (PageTransHuge(page)) {
1409		/* That test is redundant: it's for safety or to optimize out */
 
 
 
 
 
 
 
1410
1411		last = atomic_add_negative(-1, compound_mapcount_ptr(page));
 
 
 
 
 
1412		if (last) {
1413			mapped = subpages_mapcount_ptr(page);
1414			nr = atomic_sub_return_relaxed(COMPOUND_MAPPED, mapped);
1415			if (likely(nr < COMPOUND_MAPPED)) {
1416				nr_pmdmapped = thp_nr_pages(page);
1417				nr = nr_pmdmapped - (nr & SUBPAGES_MAPPED);
1418				/* Raced ahead of another remove and an add? */
1419				if (unlikely(nr < 0))
1420					nr = 0;
1421			} else {
1422				/* An add of COMPOUND_MAPPED raced ahead */
1423				nr = 0;
1424			}
1425		}
 
1426	}
1427
1428	if (nr_pmdmapped) {
1429		__mod_lruvec_page_state(page, PageAnon(page) ? NR_ANON_THPS :
1430				(PageSwapBacked(page) ? NR_SHMEM_PMDMAPPED :
1431				NR_FILE_PMDMAPPED), -nr_pmdmapped);
 
 
 
 
1432	}
1433	if (nr) {
1434		__mod_lruvec_page_state(page, PageAnon(page) ? NR_ANON_MAPPED :
1435				NR_FILE_MAPPED, -nr);
 
1436		/*
1437		 * Queue anon THP for deferred split if at least one small
1438		 * page of the compound page is unmapped, but at least one
1439		 * small page is still mapped.
1440		 */
1441		if (PageTransCompound(page) && PageAnon(page))
1442			if (!compound || nr < nr_pmdmapped)
1443				deferred_split_huge_page(compound_head(page));
1444	}
1445
1446	/*
1447	 * It would be tidy to reset PageAnon mapping when fully unmapped,
1448	 * but that might overwrite a racing page_add_anon_rmap
1449	 * which increments mapcount after us but sets mapping
1450	 * before us: so leave the reset to free_pages_prepare,
1451	 * and remember that it's only reliable while mapped.
1452	 */
1453
1454	unlock_page_memcg(page);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1455
1456	munlock_vma_page(page, vma, compound);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1457}
1458
1459/*
1460 * @arg: enum ttu_flags will be passed to this argument
1461 */
1462static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
1463		     unsigned long address, void *arg)
1464{
1465	struct mm_struct *mm = vma->vm_mm;
1466	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1467	pte_t pteval;
1468	struct page *subpage;
1469	bool anon_exclusive, ret = true;
1470	struct mmu_notifier_range range;
1471	enum ttu_flags flags = (enum ttu_flags)(long)arg;
 
 
1472
1473	/*
1474	 * When racing against e.g. zap_pte_range() on another cpu,
1475	 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1476	 * try_to_unmap() may return before page_mapped() has become false,
1477	 * if page table locking is skipped: use TTU_SYNC to wait for that.
1478	 */
1479	if (flags & TTU_SYNC)
1480		pvmw.flags = PVMW_SYNC;
1481
1482	if (flags & TTU_SPLIT_HUGE_PMD)
1483		split_huge_pmd_address(vma, address, false, folio);
1484
1485	/*
1486	 * For THP, we have to assume the worse case ie pmd for invalidation.
1487	 * For hugetlb, it could be much worse if we need to do pud
1488	 * invalidation in the case of pmd sharing.
1489	 *
1490	 * Note that the folio can not be freed in this function as call of
1491	 * try_to_unmap() must hold a reference on the folio.
1492	 */
1493	range.end = vma_address_end(&pvmw);
1494	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1495				address, range.end);
1496	if (folio_test_hugetlb(folio)) {
1497		/*
1498		 * If sharing is possible, start and end will be adjusted
1499		 * accordingly.
1500		 */
1501		adjust_range_if_pmd_sharing_possible(vma, &range.start,
1502						     &range.end);
 
 
 
1503	}
1504	mmu_notifier_invalidate_range_start(&range);
1505
1506	while (page_vma_mapped_walk(&pvmw)) {
1507		/* Unexpected PMD-mapped THP? */
1508		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
1509
1510		/*
1511		 * If the folio is in an mlock()d vma, we must not swap it out.
1512		 */
1513		if (!(flags & TTU_IGNORE_MLOCK) &&
1514		    (vma->vm_flags & VM_LOCKED)) {
1515			/* Restore the mlock which got missed */
1516			mlock_vma_folio(folio, vma, false);
 
1517			page_vma_mapped_walk_done(&pvmw);
1518			ret = false;
1519			break;
1520		}
1521
1522		subpage = folio_page(folio,
1523					pte_pfn(*pvmw.pte) - folio_pfn(folio));
1524		address = pvmw.address;
1525		anon_exclusive = folio_test_anon(folio) &&
1526				 PageAnonExclusive(subpage);
1527
1528		if (folio_test_hugetlb(folio)) {
1529			bool anon = folio_test_anon(folio);
1530
1531			/*
1532			 * The try_to_unmap() is only passed a hugetlb page
1533			 * in the case where the hugetlb page is poisoned.
1534			 */
1535			VM_BUG_ON_PAGE(!PageHWPoison(subpage), subpage);
1536			/*
1537			 * huge_pmd_unshare may unmap an entire PMD page.
1538			 * There is no way of knowing exactly which PMDs may
1539			 * be cached for this mm, so we must flush them all.
1540			 * start/end were already adjusted above to cover this
1541			 * range.
1542			 */
1543			flush_cache_range(vma, range.start, range.end);
1544
1545			/*
1546			 * To call huge_pmd_unshare, i_mmap_rwsem must be
1547			 * held in write mode.  Caller needs to explicitly
1548			 * do this outside rmap routines.
1549			 *
1550			 * We also must hold hugetlb vma_lock in write mode.
1551			 * Lock order dictates acquiring vma_lock BEFORE
1552			 * i_mmap_rwsem.  We can only try lock here and fail
1553			 * if unsuccessful.
1554			 */
1555			if (!anon) {
1556				VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1557				if (!hugetlb_vma_trylock_write(vma)) {
1558					page_vma_mapped_walk_done(&pvmw);
1559					ret = false;
1560					break;
1561				}
1562				if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1563					hugetlb_vma_unlock_write(vma);
1564					flush_tlb_range(vma,
1565						range.start, range.end);
1566					mmu_notifier_invalidate_range(mm,
1567						range.start, range.end);
1568					/*
1569					 * The ref count of the PMD page was
1570					 * dropped which is part of the way map
1571					 * counting is done for shared PMDs.
1572					 * Return 'true' here.  When there is
1573					 * no other sharing, huge_pmd_unshare
1574					 * returns false and we will unmap the
1575					 * actual page and drop map count
1576					 * to zero.
1577					 */
1578					page_vma_mapped_walk_done(&pvmw);
1579					break;
1580				}
1581				hugetlb_vma_unlock_write(vma);
1582			}
1583			pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
1584		} else {
1585			flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1586			/* Nuke the page table entry. */
1587			if (should_defer_flush(mm, flags)) {
1588				/*
1589				 * We clear the PTE but do not flush so potentially
1590				 * a remote CPU could still be writing to the folio.
1591				 * If the entry was previously clean then the
1592				 * architecture must guarantee that a clear->dirty
1593				 * transition on a cached TLB entry is written through
1594				 * and traps if the PTE is unmapped.
1595				 */
1596				pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1597
1598				set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1599			} else {
1600				pteval = ptep_clear_flush(vma, address, pvmw.pte);
1601			}
1602		}
1603
1604		/*
1605		 * Now the pte is cleared. If this pte was uffd-wp armed,
1606		 * we may want to replace a none pte with a marker pte if
1607		 * it's file-backed, so we don't lose the tracking info.
1608		 */
1609		pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval);
1610
1611		/* Set the dirty flag on the folio now the pte is gone. */
1612		if (pte_dirty(pteval))
1613			folio_mark_dirty(folio);
1614
1615		/* Update high watermark before we lower rss */
1616		update_hiwater_rss(mm);
1617
1618		if (PageHWPoison(subpage) && !(flags & TTU_IGNORE_HWPOISON)) {
1619			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1620			if (folio_test_hugetlb(folio)) {
1621				hugetlb_count_sub(folio_nr_pages(folio), mm);
1622				set_huge_pte_at(mm, address, pvmw.pte, pteval);
 
1623			} else {
1624				dec_mm_counter(mm, mm_counter(&folio->page));
1625				set_pte_at(mm, address, pvmw.pte, pteval);
1626			}
1627
1628		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1629			/*
1630			 * The guest indicated that the page content is of no
1631			 * interest anymore. Simply discard the pte, vmscan
1632			 * will take care of the rest.
1633			 * A future reference will then fault in a new zero
1634			 * page. When userfaultfd is active, we must not drop
1635			 * this page though, as its main user (postcopy
1636			 * migration) will not expect userfaults on already
1637			 * copied pages.
1638			 */
1639			dec_mm_counter(mm, mm_counter(&folio->page));
1640			/* We have to invalidate as we cleared the pte */
1641			mmu_notifier_invalidate_range(mm, address,
1642						      address + PAGE_SIZE);
1643		} else if (folio_test_anon(folio)) {
1644			swp_entry_t entry = { .val = page_private(subpage) };
1645			pte_t swp_pte;
1646			/*
1647			 * Store the swap location in the pte.
1648			 * See handle_pte_fault() ...
1649			 */
1650			if (unlikely(folio_test_swapbacked(folio) !=
1651					folio_test_swapcache(folio))) {
1652				WARN_ON_ONCE(1);
1653				ret = false;
1654				/* We have to invalidate as we cleared the pte */
1655				mmu_notifier_invalidate_range(mm, address,
1656							address + PAGE_SIZE);
1657				page_vma_mapped_walk_done(&pvmw);
1658				break;
1659			}
1660
1661			/* MADV_FREE page check */
1662			if (!folio_test_swapbacked(folio)) {
1663				int ref_count, map_count;
1664
1665				/*
1666				 * Synchronize with gup_pte_range():
1667				 * - clear PTE; barrier; read refcount
1668				 * - inc refcount; barrier; read PTE
1669				 */
1670				smp_mb();
1671
1672				ref_count = folio_ref_count(folio);
1673				map_count = folio_mapcount(folio);
1674
1675				/*
1676				 * Order reads for page refcount and dirty flag
1677				 * (see comments in __remove_mapping()).
1678				 */
1679				smp_rmb();
1680
1681				/*
1682				 * The only page refs must be one from isolation
1683				 * plus the rmap(s) (dropped by discard:).
1684				 */
1685				if (ref_count == 1 + map_count &&
1686				    !folio_test_dirty(folio)) {
1687					/* Invalidate as we cleared the pte */
1688					mmu_notifier_invalidate_range(mm,
1689						address, address + PAGE_SIZE);
1690					dec_mm_counter(mm, MM_ANONPAGES);
1691					goto discard;
1692				}
1693
1694				/*
1695				 * If the folio was redirtied, it cannot be
1696				 * discarded. Remap the page to page table.
1697				 */
1698				set_pte_at(mm, address, pvmw.pte, pteval);
1699				folio_set_swapbacked(folio);
1700				ret = false;
1701				page_vma_mapped_walk_done(&pvmw);
1702				break;
1703			}
1704
1705			if (swap_duplicate(entry) < 0) {
1706				set_pte_at(mm, address, pvmw.pte, pteval);
1707				ret = false;
1708				page_vma_mapped_walk_done(&pvmw);
1709				break;
1710			}
1711			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1712				swap_free(entry);
1713				set_pte_at(mm, address, pvmw.pte, pteval);
1714				ret = false;
1715				page_vma_mapped_walk_done(&pvmw);
1716				break;
1717			}
1718
1719			/* See page_try_share_anon_rmap(): clear PTE first. */
1720			if (anon_exclusive &&
1721			    page_try_share_anon_rmap(subpage)) {
1722				swap_free(entry);
1723				set_pte_at(mm, address, pvmw.pte, pteval);
1724				ret = false;
1725				page_vma_mapped_walk_done(&pvmw);
1726				break;
1727			}
1728			/*
1729			 * Note: We *don't* remember if the page was mapped
1730			 * exclusively in the swap pte if the architecture
1731			 * doesn't support __HAVE_ARCH_PTE_SWP_EXCLUSIVE. In
1732			 * that case, swapin code has to re-determine that
1733			 * manually and might detect the page as possibly
1734			 * shared, for example, if there are other references on
1735			 * the page or if the page is under writeback. We made
1736			 * sure that there are no GUP pins on the page that
1737			 * would rely on it, so for GUP pins this is fine.
1738			 */
1739			if (list_empty(&mm->mmlist)) {
1740				spin_lock(&mmlist_lock);
1741				if (list_empty(&mm->mmlist))
1742					list_add(&mm->mmlist, &init_mm.mmlist);
1743				spin_unlock(&mmlist_lock);
1744			}
1745			dec_mm_counter(mm, MM_ANONPAGES);
1746			inc_mm_counter(mm, MM_SWAPENTS);
1747			swp_pte = swp_entry_to_pte(entry);
1748			if (anon_exclusive)
1749				swp_pte = pte_swp_mkexclusive(swp_pte);
1750			if (pte_soft_dirty(pteval))
1751				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1752			if (pte_uffd_wp(pteval))
1753				swp_pte = pte_swp_mkuffd_wp(swp_pte);
1754			set_pte_at(mm, address, pvmw.pte, swp_pte);
1755			/* Invalidate as we cleared the pte */
1756			mmu_notifier_invalidate_range(mm, address,
1757						      address + PAGE_SIZE);
1758		} else {
1759			/*
1760			 * This is a locked file-backed folio,
1761			 * so it cannot be removed from the page
1762			 * cache and replaced by a new folio before
1763			 * mmu_notifier_invalidate_range_end, so no
1764			 * concurrent thread might update its page table
1765			 * to point at a new folio while a device is
1766			 * still using this folio.
1767			 *
1768			 * See Documentation/mm/mmu_notifier.rst
1769			 */
1770			dec_mm_counter(mm, mm_counter_file(&folio->page));
1771		}
1772discard:
1773		/*
1774		 * No need to call mmu_notifier_invalidate_range() it has be
1775		 * done above for all cases requiring it to happen under page
1776		 * table lock before mmu_notifier_invalidate_range_end()
1777		 *
1778		 * See Documentation/mm/mmu_notifier.rst
1779		 */
1780		page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
1781		if (vma->vm_flags & VM_LOCKED)
1782			mlock_page_drain_local();
1783		folio_put(folio);
1784	}
1785
1786	mmu_notifier_invalidate_range_end(&range);
1787
1788	return ret;
1789}
1790
1791static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1792{
1793	return vma_is_temporary_stack(vma);
1794}
1795
1796static int folio_not_mapped(struct folio *folio)
1797{
1798	return !folio_mapped(folio);
1799}
1800
1801/**
1802 * try_to_unmap - Try to remove all page table mappings to a folio.
1803 * @folio: The folio to unmap.
1804 * @flags: action and flags
1805 *
1806 * Tries to remove all the page table entries which are mapping this
1807 * folio.  It is the caller's responsibility to check if the folio is
1808 * still mapped if needed (use TTU_SYNC to prevent accounting races).
1809 *
1810 * Context: Caller must hold the folio lock.
1811 */
1812void try_to_unmap(struct folio *folio, enum ttu_flags flags)
1813{
1814	struct rmap_walk_control rwc = {
1815		.rmap_one = try_to_unmap_one,
1816		.arg = (void *)flags,
1817		.done = folio_not_mapped,
1818		.anon_lock = folio_lock_anon_vma_read,
1819	};
1820
1821	if (flags & TTU_RMAP_LOCKED)
1822		rmap_walk_locked(folio, &rwc);
1823	else
1824		rmap_walk(folio, &rwc);
1825}
1826
1827/*
1828 * @arg: enum ttu_flags will be passed to this argument.
1829 *
1830 * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
1831 * containing migration entries.
1832 */
1833static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
1834		     unsigned long address, void *arg)
1835{
1836	struct mm_struct *mm = vma->vm_mm;
1837	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1838	pte_t pteval;
1839	struct page *subpage;
1840	bool anon_exclusive, ret = true;
1841	struct mmu_notifier_range range;
1842	enum ttu_flags flags = (enum ttu_flags)(long)arg;
 
 
1843
1844	/*
1845	 * When racing against e.g. zap_pte_range() on another cpu,
1846	 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1847	 * try_to_migrate() may return before page_mapped() has become false,
1848	 * if page table locking is skipped: use TTU_SYNC to wait for that.
1849	 */
1850	if (flags & TTU_SYNC)
1851		pvmw.flags = PVMW_SYNC;
1852
1853	/*
1854	 * unmap_page() in mm/huge_memory.c is the only user of migration with
1855	 * TTU_SPLIT_HUGE_PMD and it wants to freeze.
1856	 */
1857	if (flags & TTU_SPLIT_HUGE_PMD)
1858		split_huge_pmd_address(vma, address, true, folio);
1859
1860	/*
1861	 * For THP, we have to assume the worse case ie pmd for invalidation.
1862	 * For hugetlb, it could be much worse if we need to do pud
1863	 * invalidation in the case of pmd sharing.
1864	 *
1865	 * Note that the page can not be free in this function as call of
1866	 * try_to_unmap() must hold a reference on the page.
1867	 */
1868	range.end = vma_address_end(&pvmw);
1869	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1870				address, range.end);
1871	if (folio_test_hugetlb(folio)) {
1872		/*
1873		 * If sharing is possible, start and end will be adjusted
1874		 * accordingly.
1875		 */
1876		adjust_range_if_pmd_sharing_possible(vma, &range.start,
1877						     &range.end);
 
 
 
1878	}
1879	mmu_notifier_invalidate_range_start(&range);
1880
1881	while (page_vma_mapped_walk(&pvmw)) {
1882#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1883		/* PMD-mapped THP migration entry */
1884		if (!pvmw.pte) {
1885			subpage = folio_page(folio,
1886				pmd_pfn(*pvmw.pmd) - folio_pfn(folio));
1887			VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
1888					!folio_test_pmd_mappable(folio), folio);
1889
1890			if (set_pmd_migration_entry(&pvmw, subpage)) {
1891				ret = false;
1892				page_vma_mapped_walk_done(&pvmw);
1893				break;
1894			}
1895			continue;
1896		}
1897#endif
1898
1899		/* Unexpected PMD-mapped THP? */
1900		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
1901
 
 
1902		if (folio_is_zone_device(folio)) {
1903			/*
1904			 * Our PTE is a non-present device exclusive entry and
1905			 * calculating the subpage as for the common case would
1906			 * result in an invalid pointer.
1907			 *
1908			 * Since only PAGE_SIZE pages can currently be
1909			 * migrated, just set it to page. This will need to be
1910			 * changed when hugepage migrations to device private
1911			 * memory are supported.
1912			 */
1913			VM_BUG_ON_FOLIO(folio_nr_pages(folio) > 1, folio);
1914			subpage = &folio->page;
1915		} else {
1916			subpage = folio_page(folio,
1917					pte_pfn(*pvmw.pte) - folio_pfn(folio));
1918		}
1919		address = pvmw.address;
1920		anon_exclusive = folio_test_anon(folio) &&
1921				 PageAnonExclusive(subpage);
1922
1923		if (folio_test_hugetlb(folio)) {
1924			bool anon = folio_test_anon(folio);
1925
1926			/*
1927			 * huge_pmd_unshare may unmap an entire PMD page.
1928			 * There is no way of knowing exactly which PMDs may
1929			 * be cached for this mm, so we must flush them all.
1930			 * start/end were already adjusted above to cover this
1931			 * range.
1932			 */
1933			flush_cache_range(vma, range.start, range.end);
1934
1935			/*
1936			 * To call huge_pmd_unshare, i_mmap_rwsem must be
1937			 * held in write mode.  Caller needs to explicitly
1938			 * do this outside rmap routines.
1939			 *
1940			 * We also must hold hugetlb vma_lock in write mode.
1941			 * Lock order dictates acquiring vma_lock BEFORE
1942			 * i_mmap_rwsem.  We can only try lock here and
1943			 * fail if unsuccessful.
1944			 */
1945			if (!anon) {
1946				VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1947				if (!hugetlb_vma_trylock_write(vma)) {
1948					page_vma_mapped_walk_done(&pvmw);
1949					ret = false;
1950					break;
1951				}
1952				if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1953					hugetlb_vma_unlock_write(vma);
1954					flush_tlb_range(vma,
1955						range.start, range.end);
1956					mmu_notifier_invalidate_range(mm,
1957						range.start, range.end);
1958
1959					/*
1960					 * The ref count of the PMD page was
1961					 * dropped which is part of the way map
1962					 * counting is done for shared PMDs.
1963					 * Return 'true' here.  When there is
1964					 * no other sharing, huge_pmd_unshare
1965					 * returns false and we will unmap the
1966					 * actual page and drop map count
1967					 * to zero.
1968					 */
1969					page_vma_mapped_walk_done(&pvmw);
1970					break;
1971				}
1972				hugetlb_vma_unlock_write(vma);
1973			}
1974			/* Nuke the hugetlb page table entry */
1975			pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
1976		} else {
1977			flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1978			/* Nuke the page table entry. */
1979			pteval = ptep_clear_flush(vma, address, pvmw.pte);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1980		}
1981
1982		/* Set the dirty flag on the folio now the pte is gone. */
1983		if (pte_dirty(pteval))
1984			folio_mark_dirty(folio);
1985
1986		/* Update high watermark before we lower rss */
1987		update_hiwater_rss(mm);
1988
1989		if (folio_is_device_private(folio)) {
1990			unsigned long pfn = folio_pfn(folio);
1991			swp_entry_t entry;
1992			pte_t swp_pte;
1993
1994			if (anon_exclusive)
1995				BUG_ON(page_try_share_anon_rmap(subpage));
 
1996
1997			/*
1998			 * Store the pfn of the page in a special migration
1999			 * pte. do_swap_page() will wait until the migration
2000			 * pte is removed and then restart fault handling.
2001			 */
2002			entry = pte_to_swp_entry(pteval);
2003			if (is_writable_device_private_entry(entry))
2004				entry = make_writable_migration_entry(pfn);
2005			else if (anon_exclusive)
2006				entry = make_readable_exclusive_migration_entry(pfn);
2007			else
2008				entry = make_readable_migration_entry(pfn);
2009			swp_pte = swp_entry_to_pte(entry);
2010
2011			/*
2012			 * pteval maps a zone device page and is therefore
2013			 * a swap pte.
2014			 */
2015			if (pte_swp_soft_dirty(pteval))
2016				swp_pte = pte_swp_mksoft_dirty(swp_pte);
2017			if (pte_swp_uffd_wp(pteval))
2018				swp_pte = pte_swp_mkuffd_wp(swp_pte);
2019			set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
2020			trace_set_migration_pte(pvmw.address, pte_val(swp_pte),
2021						compound_order(&folio->page));
2022			/*
2023			 * No need to invalidate here it will synchronize on
2024			 * against the special swap migration pte.
2025			 */
2026		} else if (PageHWPoison(subpage)) {
2027			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
2028			if (folio_test_hugetlb(folio)) {
2029				hugetlb_count_sub(folio_nr_pages(folio), mm);
2030				set_huge_pte_at(mm, address, pvmw.pte, pteval);
 
2031			} else {
2032				dec_mm_counter(mm, mm_counter(&folio->page));
2033				set_pte_at(mm, address, pvmw.pte, pteval);
2034			}
2035
2036		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
2037			/*
2038			 * The guest indicated that the page content is of no
2039			 * interest anymore. Simply discard the pte, vmscan
2040			 * will take care of the rest.
2041			 * A future reference will then fault in a new zero
2042			 * page. When userfaultfd is active, we must not drop
2043			 * this page though, as its main user (postcopy
2044			 * migration) will not expect userfaults on already
2045			 * copied pages.
2046			 */
2047			dec_mm_counter(mm, mm_counter(&folio->page));
2048			/* We have to invalidate as we cleared the pte */
2049			mmu_notifier_invalidate_range(mm, address,
2050						      address + PAGE_SIZE);
2051		} else {
2052			swp_entry_t entry;
2053			pte_t swp_pte;
2054
2055			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
2056				if (folio_test_hugetlb(folio))
2057					set_huge_pte_at(mm, address, pvmw.pte, pteval);
 
2058				else
2059					set_pte_at(mm, address, pvmw.pte, pteval);
2060				ret = false;
2061				page_vma_mapped_walk_done(&pvmw);
2062				break;
2063			}
2064			VM_BUG_ON_PAGE(pte_write(pteval) && folio_test_anon(folio) &&
2065				       !anon_exclusive, subpage);
2066
2067			/* See page_try_share_anon_rmap(): clear PTE first. */
2068			if (anon_exclusive &&
2069			    page_try_share_anon_rmap(subpage)) {
2070				if (folio_test_hugetlb(folio))
2071					set_huge_pte_at(mm, address, pvmw.pte, pteval);
2072				else
2073					set_pte_at(mm, address, pvmw.pte, pteval);
 
 
 
 
 
 
2074				ret = false;
2075				page_vma_mapped_walk_done(&pvmw);
2076				break;
2077			}
2078
2079			/*
2080			 * Store the pfn of the page in a special migration
2081			 * pte. do_swap_page() will wait until the migration
2082			 * pte is removed and then restart fault handling.
2083			 */
2084			if (pte_write(pteval))
2085				entry = make_writable_migration_entry(
2086							page_to_pfn(subpage));
2087			else if (anon_exclusive)
2088				entry = make_readable_exclusive_migration_entry(
2089							page_to_pfn(subpage));
2090			else
2091				entry = make_readable_migration_entry(
2092							page_to_pfn(subpage));
2093			if (pte_young(pteval))
2094				entry = make_migration_entry_young(entry);
2095			if (pte_dirty(pteval))
2096				entry = make_migration_entry_dirty(entry);
2097			swp_pte = swp_entry_to_pte(entry);
2098			if (pte_soft_dirty(pteval))
2099				swp_pte = pte_swp_mksoft_dirty(swp_pte);
2100			if (pte_uffd_wp(pteval))
2101				swp_pte = pte_swp_mkuffd_wp(swp_pte);
2102			if (folio_test_hugetlb(folio))
2103				set_huge_pte_at(mm, address, pvmw.pte, swp_pte);
 
2104			else
2105				set_pte_at(mm, address, pvmw.pte, swp_pte);
2106			trace_set_migration_pte(address, pte_val(swp_pte),
2107						compound_order(&folio->page));
2108			/*
2109			 * No need to invalidate here it will synchronize on
2110			 * against the special swap migration pte.
2111			 */
2112		}
2113
2114		/*
2115		 * No need to call mmu_notifier_invalidate_range() it has be
2116		 * done above for all cases requiring it to happen under page
2117		 * table lock before mmu_notifier_invalidate_range_end()
2118		 *
2119		 * See Documentation/mm/mmu_notifier.rst
2120		 */
2121		page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
2122		if (vma->vm_flags & VM_LOCKED)
2123			mlock_page_drain_local();
2124		folio_put(folio);
2125	}
2126
2127	mmu_notifier_invalidate_range_end(&range);
2128
2129	return ret;
2130}
2131
2132/**
2133 * try_to_migrate - try to replace all page table mappings with swap entries
2134 * @folio: the folio to replace page table entries for
2135 * @flags: action and flags
2136 *
2137 * Tries to remove all the page table entries which are mapping this folio and
2138 * replace them with special swap entries. Caller must hold the folio lock.
2139 */
2140void try_to_migrate(struct folio *folio, enum ttu_flags flags)
2141{
2142	struct rmap_walk_control rwc = {
2143		.rmap_one = try_to_migrate_one,
2144		.arg = (void *)flags,
2145		.done = folio_not_mapped,
2146		.anon_lock = folio_lock_anon_vma_read,
2147	};
2148
2149	/*
2150	 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
2151	 * TTU_SPLIT_HUGE_PMD and TTU_SYNC flags.
2152	 */
2153	if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2154					TTU_SYNC)))
2155		return;
2156
2157	if (folio_is_zone_device(folio) &&
2158	    (!folio_is_device_private(folio) && !folio_is_device_coherent(folio)))
2159		return;
2160
2161	/*
2162	 * During exec, a temporary VMA is setup and later moved.
2163	 * The VMA is moved under the anon_vma lock but not the
2164	 * page tables leading to a race where migration cannot
2165	 * find the migration ptes. Rather than increasing the
2166	 * locking requirements of exec(), migration skips
2167	 * temporary VMAs until after exec() completes.
2168	 */
2169	if (!folio_test_ksm(folio) && folio_test_anon(folio))
2170		rwc.invalid_vma = invalid_migration_vma;
2171
2172	if (flags & TTU_RMAP_LOCKED)
2173		rmap_walk_locked(folio, &rwc);
2174	else
2175		rmap_walk(folio, &rwc);
2176}
2177
2178#ifdef CONFIG_DEVICE_PRIVATE
2179struct make_exclusive_args {
2180	struct mm_struct *mm;
2181	unsigned long address;
2182	void *owner;
2183	bool valid;
2184};
2185
2186static bool page_make_device_exclusive_one(struct folio *folio,
2187		struct vm_area_struct *vma, unsigned long address, void *priv)
2188{
2189	struct mm_struct *mm = vma->vm_mm;
2190	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
2191	struct make_exclusive_args *args = priv;
2192	pte_t pteval;
2193	struct page *subpage;
2194	bool ret = true;
2195	struct mmu_notifier_range range;
2196	swp_entry_t entry;
2197	pte_t swp_pte;
 
2198
2199	mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
2200				      vma->vm_mm, address, min(vma->vm_end,
2201				      address + folio_size(folio)),
2202				      args->owner);
2203	mmu_notifier_invalidate_range_start(&range);
2204
2205	while (page_vma_mapped_walk(&pvmw)) {
2206		/* Unexpected PMD-mapped THP? */
2207		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2208
2209		if (!pte_present(*pvmw.pte)) {
 
2210			ret = false;
2211			page_vma_mapped_walk_done(&pvmw);
2212			break;
2213		}
2214
2215		subpage = folio_page(folio,
2216				pte_pfn(*pvmw.pte) - folio_pfn(folio));
2217		address = pvmw.address;
2218
2219		/* Nuke the page table entry. */
2220		flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
2221		pteval = ptep_clear_flush(vma, address, pvmw.pte);
2222
2223		/* Set the dirty flag on the folio now the pte is gone. */
2224		if (pte_dirty(pteval))
2225			folio_mark_dirty(folio);
2226
2227		/*
2228		 * Check that our target page is still mapped at the expected
2229		 * address.
2230		 */
2231		if (args->mm == mm && args->address == address &&
2232		    pte_write(pteval))
2233			args->valid = true;
2234
2235		/*
2236		 * Store the pfn of the page in a special migration
2237		 * pte. do_swap_page() will wait until the migration
2238		 * pte is removed and then restart fault handling.
2239		 */
2240		if (pte_write(pteval))
2241			entry = make_writable_device_exclusive_entry(
2242							page_to_pfn(subpage));
2243		else
2244			entry = make_readable_device_exclusive_entry(
2245							page_to_pfn(subpage));
2246		swp_pte = swp_entry_to_pte(entry);
2247		if (pte_soft_dirty(pteval))
2248			swp_pte = pte_swp_mksoft_dirty(swp_pte);
2249		if (pte_uffd_wp(pteval))
2250			swp_pte = pte_swp_mkuffd_wp(swp_pte);
2251
2252		set_pte_at(mm, address, pvmw.pte, swp_pte);
2253
2254		/*
2255		 * There is a reference on the page for the swap entry which has
2256		 * been removed, so shouldn't take another.
2257		 */
2258		page_remove_rmap(subpage, vma, false);
2259	}
2260
2261	mmu_notifier_invalidate_range_end(&range);
2262
2263	return ret;
2264}
2265
2266/**
2267 * folio_make_device_exclusive - Mark the folio exclusively owned by a device.
2268 * @folio: The folio to replace page table entries for.
2269 * @mm: The mm_struct where the folio is expected to be mapped.
2270 * @address: Address where the folio is expected to be mapped.
2271 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
2272 *
2273 * Tries to remove all the page table entries which are mapping this
2274 * folio and replace them with special device exclusive swap entries to
2275 * grant a device exclusive access to the folio.
2276 *
2277 * Context: Caller must hold the folio lock.
2278 * Return: false if the page is still mapped, or if it could not be unmapped
2279 * from the expected address. Otherwise returns true (success).
2280 */
2281static bool folio_make_device_exclusive(struct folio *folio,
2282		struct mm_struct *mm, unsigned long address, void *owner)
2283{
2284	struct make_exclusive_args args = {
2285		.mm = mm,
2286		.address = address,
2287		.owner = owner,
2288		.valid = false,
2289	};
2290	struct rmap_walk_control rwc = {
2291		.rmap_one = page_make_device_exclusive_one,
2292		.done = folio_not_mapped,
2293		.anon_lock = folio_lock_anon_vma_read,
2294		.arg = &args,
2295	};
2296
2297	/*
2298	 * Restrict to anonymous folios for now to avoid potential writeback
2299	 * issues.
2300	 */
2301	if (!folio_test_anon(folio))
2302		return false;
2303
2304	rmap_walk(folio, &rwc);
2305
2306	return args.valid && !folio_mapcount(folio);
2307}
2308
2309/**
2310 * make_device_exclusive_range() - Mark a range for exclusive use by a device
2311 * @mm: mm_struct of associated target process
2312 * @start: start of the region to mark for exclusive device access
2313 * @end: end address of region
2314 * @pages: returns the pages which were successfully marked for exclusive access
2315 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
2316 *
2317 * Returns: number of pages found in the range by GUP. A page is marked for
2318 * exclusive access only if the page pointer is non-NULL.
2319 *
2320 * This function finds ptes mapping page(s) to the given address range, locks
2321 * them and replaces mappings with special swap entries preventing userspace CPU
2322 * access. On fault these entries are replaced with the original mapping after
2323 * calling MMU notifiers.
2324 *
2325 * A driver using this to program access from a device must use a mmu notifier
2326 * critical section to hold a device specific lock during programming. Once
2327 * programming is complete it should drop the page lock and reference after
2328 * which point CPU access to the page will revoke the exclusive access.
2329 */
2330int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
2331				unsigned long end, struct page **pages,
2332				void *owner)
2333{
2334	long npages = (end - start) >> PAGE_SHIFT;
2335	long i;
2336
2337	npages = get_user_pages_remote(mm, start, npages,
2338				       FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
2339				       pages, NULL, NULL);
2340	if (npages < 0)
2341		return npages;
2342
2343	for (i = 0; i < npages; i++, start += PAGE_SIZE) {
2344		struct folio *folio = page_folio(pages[i]);
2345		if (PageTail(pages[i]) || !folio_trylock(folio)) {
2346			folio_put(folio);
2347			pages[i] = NULL;
2348			continue;
2349		}
2350
2351		if (!folio_make_device_exclusive(folio, mm, start, owner)) {
2352			folio_unlock(folio);
2353			folio_put(folio);
2354			pages[i] = NULL;
2355		}
2356	}
2357
2358	return npages;
2359}
2360EXPORT_SYMBOL_GPL(make_device_exclusive_range);
2361#endif
2362
2363void __put_anon_vma(struct anon_vma *anon_vma)
2364{
2365	struct anon_vma *root = anon_vma->root;
2366
2367	anon_vma_free(anon_vma);
2368	if (root != anon_vma && atomic_dec_and_test(&root->refcount))
2369		anon_vma_free(root);
2370}
2371
2372static struct anon_vma *rmap_walk_anon_lock(struct folio *folio,
2373					    struct rmap_walk_control *rwc)
2374{
2375	struct anon_vma *anon_vma;
2376
2377	if (rwc->anon_lock)
2378		return rwc->anon_lock(folio, rwc);
2379
2380	/*
2381	 * Note: remove_migration_ptes() cannot use folio_lock_anon_vma_read()
2382	 * because that depends on page_mapped(); but not all its usages
2383	 * are holding mmap_lock. Users without mmap_lock are required to
2384	 * take a reference count to prevent the anon_vma disappearing
2385	 */
2386	anon_vma = folio_anon_vma(folio);
2387	if (!anon_vma)
2388		return NULL;
2389
2390	if (anon_vma_trylock_read(anon_vma))
2391		goto out;
2392
2393	if (rwc->try_lock) {
2394		anon_vma = NULL;
2395		rwc->contended = true;
2396		goto out;
2397	}
2398
2399	anon_vma_lock_read(anon_vma);
2400out:
2401	return anon_vma;
2402}
2403
2404/*
2405 * rmap_walk_anon - do something to anonymous page using the object-based
2406 * rmap method
2407 * @page: the page to be handled
2408 * @rwc: control variable according to each walk type
 
2409 *
2410 * Find all the mappings of a page using the mapping pointer and the vma chains
2411 * contained in the anon_vma struct it points to.
2412 */
2413static void rmap_walk_anon(struct folio *folio,
2414		struct rmap_walk_control *rwc, bool locked)
2415{
2416	struct anon_vma *anon_vma;
2417	pgoff_t pgoff_start, pgoff_end;
2418	struct anon_vma_chain *avc;
2419
2420	if (locked) {
2421		anon_vma = folio_anon_vma(folio);
2422		/* anon_vma disappear under us? */
2423		VM_BUG_ON_FOLIO(!anon_vma, folio);
2424	} else {
2425		anon_vma = rmap_walk_anon_lock(folio, rwc);
2426	}
2427	if (!anon_vma)
2428		return;
2429
2430	pgoff_start = folio_pgoff(folio);
2431	pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2432	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
2433			pgoff_start, pgoff_end) {
2434		struct vm_area_struct *vma = avc->vma;
2435		unsigned long address = vma_address(&folio->page, vma);
2436
2437		VM_BUG_ON_VMA(address == -EFAULT, vma);
2438		cond_resched();
2439
2440		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2441			continue;
2442
2443		if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2444			break;
2445		if (rwc->done && rwc->done(folio))
2446			break;
2447	}
2448
2449	if (!locked)
2450		anon_vma_unlock_read(anon_vma);
2451}
2452
2453/*
2454 * rmap_walk_file - do something to file page using the object-based rmap method
2455 * @page: the page to be handled
2456 * @rwc: control variable according to each walk type
 
2457 *
2458 * Find all the mappings of a page using the mapping pointer and the vma chains
2459 * contained in the address_space struct it points to.
2460 */
2461static void rmap_walk_file(struct folio *folio,
2462		struct rmap_walk_control *rwc, bool locked)
2463{
2464	struct address_space *mapping = folio_mapping(folio);
2465	pgoff_t pgoff_start, pgoff_end;
2466	struct vm_area_struct *vma;
2467
2468	/*
2469	 * The page lock not only makes sure that page->mapping cannot
2470	 * suddenly be NULLified by truncation, it makes sure that the
2471	 * structure at mapping cannot be freed and reused yet,
2472	 * so we can safely take mapping->i_mmap_rwsem.
2473	 */
2474	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2475
2476	if (!mapping)
2477		return;
2478
2479	pgoff_start = folio_pgoff(folio);
2480	pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2481	if (!locked) {
2482		if (i_mmap_trylock_read(mapping))
2483			goto lookup;
2484
2485		if (rwc->try_lock) {
2486			rwc->contended = true;
2487			return;
2488		}
2489
2490		i_mmap_lock_read(mapping);
2491	}
2492lookup:
2493	vma_interval_tree_foreach(vma, &mapping->i_mmap,
2494			pgoff_start, pgoff_end) {
2495		unsigned long address = vma_address(&folio->page, vma);
2496
2497		VM_BUG_ON_VMA(address == -EFAULT, vma);
2498		cond_resched();
2499
2500		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2501			continue;
2502
2503		if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2504			goto done;
2505		if (rwc->done && rwc->done(folio))
2506			goto done;
2507	}
2508
2509done:
2510	if (!locked)
2511		i_mmap_unlock_read(mapping);
2512}
2513
2514void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc)
2515{
2516	if (unlikely(folio_test_ksm(folio)))
2517		rmap_walk_ksm(folio, rwc);
2518	else if (folio_test_anon(folio))
2519		rmap_walk_anon(folio, rwc, false);
2520	else
2521		rmap_walk_file(folio, rwc, false);
2522}
2523
2524/* Like rmap_walk, but caller holds relevant rmap lock */
2525void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc)
2526{
2527	/* no ksm support for now */
2528	VM_BUG_ON_FOLIO(folio_test_ksm(folio), folio);
2529	if (folio_test_anon(folio))
2530		rmap_walk_anon(folio, rwc, true);
2531	else
2532		rmap_walk_file(folio, rwc, true);
2533}
2534
2535#ifdef CONFIG_HUGETLB_PAGE
2536/*
2537 * The following two functions are for anonymous (private mapped) hugepages.
2538 * Unlike common anonymous pages, anonymous hugepages have no accounting code
2539 * and no lru code, because we handle hugepages differently from common pages.
2540 *
2541 * RMAP_COMPOUND is ignored.
2542 */
2543void hugepage_add_anon_rmap(struct page *page, struct vm_area_struct *vma,
2544			    unsigned long address, rmap_t flags)
2545{
2546	struct anon_vma *anon_vma = vma->anon_vma;
2547	int first;
2548
2549	BUG_ON(!PageLocked(page));
2550	BUG_ON(!anon_vma);
2551	/* address might be in next vma when migration races vma_adjust */
2552	first = atomic_inc_and_test(compound_mapcount_ptr(page));
2553	VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page);
2554	VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page);
2555	if (first)
2556		__page_set_anon_rmap(page, vma, address,
2557				     !!(flags & RMAP_EXCLUSIVE));
2558}
2559
2560void hugepage_add_new_anon_rmap(struct page *page,
2561			struct vm_area_struct *vma, unsigned long address)
2562{
 
 
2563	BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2564	/* increment count (starts at -1) */
2565	atomic_set(compound_mapcount_ptr(page), 0);
2566	ClearHPageRestoreReserve(page);
2567	__page_set_anon_rmap(page, vma, address, 1);
 
2568}
2569#endif /* CONFIG_HUGETLB_PAGE */
v6.8
   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 *       page->flags PG_locked (lock_page)
  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 *                       folio_lock_memcg move_lock (in block_dirty_folio)
  36 *                         i_pages lock (widely used)
  37 *                           lruvec->lru_lock (in folio_lruvec_lock_irq)
  38 *                     inode->i_lock (in set_page_dirty's __mark_inode_dirty)
  39 *                     bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
  40 *                       sb_lock (within inode_lock in fs/fs-writeback.c)
  41 *                       i_pages lock (widely used, in set_page_dirty,
  42 *                                 in arch-dependent flush_dcache_mmap_lock,
  43 *                                 within bdi.wb->list_lock in __sync_single_inode)
  44 *
  45 * anon_vma->rwsem,mapping->i_mmap_rwsem   (memory_failure, collect_procs_anon)
  46 *   ->tasklist_lock
  47 *     pte map lock
  48 *
  49 * hugetlbfs PageHuge() take locks in this order:
  50 *   hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
  51 *     vma_lock (hugetlb specific lock for pmd_sharing)
  52 *       mapping->i_mmap_rwsem (also used for hugetlb pmd sharing)
  53 *         page->flags PG_locked (lock_page)
  54 */
  55
  56#include <linux/mm.h>
  57#include <linux/sched/mm.h>
  58#include <linux/sched/task.h>
  59#include <linux/pagemap.h>
  60#include <linux/swap.h>
  61#include <linux/swapops.h>
  62#include <linux/slab.h>
  63#include <linux/init.h>
  64#include <linux/ksm.h>
  65#include <linux/rmap.h>
  66#include <linux/rcupdate.h>
  67#include <linux/export.h>
  68#include <linux/memcontrol.h>
  69#include <linux/mmu_notifier.h>
  70#include <linux/migrate.h>
  71#include <linux/hugetlb.h>
  72#include <linux/huge_mm.h>
  73#include <linux/backing-dev.h>
  74#include <linux/page_idle.h>
  75#include <linux/memremap.h>
  76#include <linux/userfaultfd_k.h>
  77#include <linux/mm_inline.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 *
 186 * This must be called with the mmap_lock held for reading.
 187 */
 188int __anon_vma_prepare(struct vm_area_struct *vma)
 189{
 190	struct mm_struct *mm = vma->vm_mm;
 191	struct anon_vma *anon_vma, *allocated;
 192	struct anon_vma_chain *avc;
 193
 194	might_sleep();
 195
 196	avc = anon_vma_chain_alloc(GFP_KERNEL);
 197	if (!avc)
 198		goto out_enomem;
 199
 200	anon_vma = find_mergeable_anon_vma(vma);
 201	allocated = NULL;
 202	if (!anon_vma) {
 203		anon_vma = anon_vma_alloc();
 204		if (unlikely(!anon_vma))
 205			goto out_enomem_free_avc;
 206		anon_vma->num_children++; /* self-parent link for new root */
 207		allocated = anon_vma;
 208	}
 209
 210	anon_vma_lock_write(anon_vma);
 211	/* page_table_lock to protect against threads */
 212	spin_lock(&mm->page_table_lock);
 213	if (likely(!vma->anon_vma)) {
 214		vma->anon_vma = anon_vma;
 215		anon_vma_chain_link(vma, avc, anon_vma);
 216		anon_vma->num_active_vmas++;
 217		allocated = NULL;
 218		avc = NULL;
 219	}
 220	spin_unlock(&mm->page_table_lock);
 221	anon_vma_unlock_write(anon_vma);
 222
 223	if (unlikely(allocated))
 224		put_anon_vma(allocated);
 225	if (unlikely(avc))
 226		anon_vma_chain_free(avc);
 227
 228	return 0;
 229
 230 out_enomem_free_avc:
 231	anon_vma_chain_free(avc);
 232 out_enomem:
 233	return -ENOMEM;
 234}
 235
 236/*
 237 * This is a useful helper function for locking the anon_vma root as
 238 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
 239 * have the same vma.
 240 *
 241 * Such anon_vma's should have the same root, so you'd expect to see
 242 * just a single mutex_lock for the whole traversal.
 243 */
 244static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
 245{
 246	struct anon_vma *new_root = anon_vma->root;
 247	if (new_root != root) {
 248		if (WARN_ON_ONCE(root))
 249			up_write(&root->rwsem);
 250		root = new_root;
 251		down_write(&root->rwsem);
 252	}
 253	return root;
 254}
 255
 256static inline void unlock_anon_vma_root(struct anon_vma *root)
 257{
 258	if (root)
 259		up_write(&root->rwsem);
 260}
 261
 262/*
 263 * Attach the anon_vmas from src to dst.
 264 * Returns 0 on success, -ENOMEM on failure.
 265 *
 266 * anon_vma_clone() is called by vma_expand(), vma_merge(), __split_vma(),
 267 * copy_vma() and anon_vma_fork(). The first four want an exact copy of src,
 268 * while the last one, anon_vma_fork(), may try to reuse an existing anon_vma to
 269 * prevent endless growth of anon_vma. Since dst->anon_vma is set to NULL before
 270 * call, we can identify this case by checking (!dst->anon_vma &&
 271 * src->anon_vma).
 272 *
 273 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
 274 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
 275 * This prevents degradation of anon_vma hierarchy to endless linear chain in
 276 * case of constantly forking task. On the other hand, an anon_vma with more
 277 * than one child isn't reused even if there was no alive vma, thus rmap
 278 * walker has a good chance of avoiding scanning the whole hierarchy when it
 279 * searches where page is mapped.
 280 */
 281int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
 282{
 283	struct anon_vma_chain *avc, *pavc;
 284	struct anon_vma *root = NULL;
 285
 286	list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
 287		struct anon_vma *anon_vma;
 288
 289		avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
 290		if (unlikely(!avc)) {
 291			unlock_anon_vma_root(root);
 292			root = NULL;
 293			avc = anon_vma_chain_alloc(GFP_KERNEL);
 294			if (!avc)
 295				goto enomem_failure;
 296		}
 297		anon_vma = pavc->anon_vma;
 298		root = lock_anon_vma_root(root, anon_vma);
 299		anon_vma_chain_link(dst, avc, anon_vma);
 300
 301		/*
 302		 * Reuse existing anon_vma if it has no vma and only one
 303		 * anon_vma child.
 304		 *
 305		 * Root anon_vma is never reused:
 306		 * it has self-parent reference and at least one child.
 307		 */
 308		if (!dst->anon_vma && src->anon_vma &&
 309		    anon_vma->num_children < 2 &&
 310		    anon_vma->num_active_vmas == 0)
 311			dst->anon_vma = anon_vma;
 312	}
 313	if (dst->anon_vma)
 314		dst->anon_vma->num_active_vmas++;
 315	unlock_anon_vma_root(root);
 316	return 0;
 317
 318 enomem_failure:
 319	/*
 320	 * dst->anon_vma is dropped here otherwise its num_active_vmas can
 321	 * be incorrectly decremented in unlink_anon_vmas().
 322	 * We can safely do this because callers of anon_vma_clone() don't care
 323	 * about dst->anon_vma if anon_vma_clone() failed.
 324	 */
 325	dst->anon_vma = NULL;
 326	unlink_anon_vmas(dst);
 327	return -ENOMEM;
 328}
 329
 330/*
 331 * Attach vma to its own anon_vma, as well as to the anon_vmas that
 332 * the corresponding VMA in the parent process is attached to.
 333 * Returns 0 on success, non-zero on failure.
 334 */
 335int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
 336{
 337	struct anon_vma_chain *avc;
 338	struct anon_vma *anon_vma;
 339	int error;
 340
 341	/* Don't bother if the parent process has no anon_vma here. */
 342	if (!pvma->anon_vma)
 343		return 0;
 344
 345	/* Drop inherited anon_vma, we'll reuse existing or allocate new. */
 346	vma->anon_vma = NULL;
 347
 348	/*
 349	 * First, attach the new VMA to the parent VMA's anon_vmas,
 350	 * so rmap can find non-COWed pages in child processes.
 351	 */
 352	error = anon_vma_clone(vma, pvma);
 353	if (error)
 354		return error;
 355
 356	/* An existing anon_vma has been reused, all done then. */
 357	if (vma->anon_vma)
 358		return 0;
 359
 360	/* Then add our own anon_vma. */
 361	anon_vma = anon_vma_alloc();
 362	if (!anon_vma)
 363		goto out_error;
 364	anon_vma->num_active_vmas++;
 365	avc = anon_vma_chain_alloc(GFP_KERNEL);
 366	if (!avc)
 367		goto out_error_free_anon_vma;
 368
 369	/*
 370	 * The root anon_vma's rwsem is the lock actually used when we
 371	 * lock any of the anon_vmas in this anon_vma tree.
 372	 */
 373	anon_vma->root = pvma->anon_vma->root;
 374	anon_vma->parent = pvma->anon_vma;
 375	/*
 376	 * With refcounts, an anon_vma can stay around longer than the
 377	 * process it belongs to. The root anon_vma needs to be pinned until
 378	 * this anon_vma is freed, because the lock lives in the root.
 379	 */
 380	get_anon_vma(anon_vma->root);
 381	/* Mark this anon_vma as the one where our new (COWed) pages go. */
 382	vma->anon_vma = anon_vma;
 383	anon_vma_lock_write(anon_vma);
 384	anon_vma_chain_link(vma, avc, anon_vma);
 385	anon_vma->parent->num_children++;
 386	anon_vma_unlock_write(anon_vma);
 387
 388	return 0;
 389
 390 out_error_free_anon_vma:
 391	put_anon_vma(anon_vma);
 392 out_error:
 393	unlink_anon_vmas(vma);
 394	return -ENOMEM;
 395}
 396
 397void unlink_anon_vmas(struct vm_area_struct *vma)
 398{
 399	struct anon_vma_chain *avc, *next;
 400	struct anon_vma *root = NULL;
 401
 402	/*
 403	 * Unlink each anon_vma chained to the VMA.  This list is ordered
 404	 * from newest to oldest, ensuring the root anon_vma gets freed last.
 405	 */
 406	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 407		struct anon_vma *anon_vma = avc->anon_vma;
 408
 409		root = lock_anon_vma_root(root, anon_vma);
 410		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
 411
 412		/*
 413		 * Leave empty anon_vmas on the list - we'll need
 414		 * to free them outside the lock.
 415		 */
 416		if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
 417			anon_vma->parent->num_children--;
 418			continue;
 419		}
 420
 421		list_del(&avc->same_vma);
 422		anon_vma_chain_free(avc);
 423	}
 424	if (vma->anon_vma) {
 425		vma->anon_vma->num_active_vmas--;
 426
 427		/*
 428		 * vma would still be needed after unlink, and anon_vma will be prepared
 429		 * when handle fault.
 430		 */
 431		vma->anon_vma = NULL;
 432	}
 433	unlock_anon_vma_root(root);
 434
 435	/*
 436	 * Iterate the list once more, it now only contains empty and unlinked
 437	 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
 438	 * needing to write-acquire the anon_vma->root->rwsem.
 439	 */
 440	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 441		struct anon_vma *anon_vma = avc->anon_vma;
 442
 443		VM_WARN_ON(anon_vma->num_children);
 444		VM_WARN_ON(anon_vma->num_active_vmas);
 445		put_anon_vma(anon_vma);
 446
 447		list_del(&avc->same_vma);
 448		anon_vma_chain_free(avc);
 449	}
 450}
 451
 452static void anon_vma_ctor(void *data)
 453{
 454	struct anon_vma *anon_vma = data;
 455
 456	init_rwsem(&anon_vma->rwsem);
 457	atomic_set(&anon_vma->refcount, 0);
 458	anon_vma->rb_root = RB_ROOT_CACHED;
 459}
 460
 461void __init anon_vma_init(void)
 462{
 463	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
 464			0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
 465			anon_vma_ctor);
 466	anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
 467			SLAB_PANIC|SLAB_ACCOUNT);
 468}
 469
 470/*
 471 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
 472 *
 473 * Since there is no serialization what so ever against folio_remove_rmap_*()
 474 * the best this function can do is return a refcount increased anon_vma
 475 * that might have been relevant to this page.
 476 *
 477 * The page might have been remapped to a different anon_vma or the anon_vma
 478 * returned may already be freed (and even reused).
 479 *
 480 * In case it was remapped to a different anon_vma, the new anon_vma will be a
 481 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
 482 * ensure that any anon_vma obtained from the page will still be valid for as
 483 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
 484 *
 485 * All users of this function must be very careful when walking the anon_vma
 486 * chain and verify that the page in question is indeed mapped in it
 487 * [ something equivalent to page_mapped_in_vma() ].
 488 *
 489 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
 490 * folio_remove_rmap_*() that the anon_vma pointer from page->mapping is valid
 491 * if there is a mapcount, we can dereference the anon_vma after observing
 492 * those.
 493 *
 494 * NOTE: the caller should normally hold folio lock when calling this.  If
 495 * not, the caller needs to double check the anon_vma didn't change after
 496 * taking the anon_vma lock for either read or write (UFFDIO_MOVE can modify it
 497 * concurrently without folio lock protection). See folio_lock_anon_vma_read()
 498 * which has already covered that, and comment above remap_pages().
 499 */
 500struct anon_vma *folio_get_anon_vma(struct folio *folio)
 501{
 502	struct anon_vma *anon_vma = NULL;
 503	unsigned long anon_mapping;
 504
 505	rcu_read_lock();
 506	anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
 507	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 508		goto out;
 509	if (!folio_mapped(folio))
 510		goto out;
 511
 512	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 513	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 514		anon_vma = NULL;
 515		goto out;
 516	}
 517
 518	/*
 519	 * If this folio is still mapped, then its anon_vma cannot have been
 520	 * freed.  But if it has been unmapped, we have no security against the
 521	 * anon_vma structure being freed and reused (for another anon_vma:
 522	 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
 523	 * above cannot corrupt).
 524	 */
 525	if (!folio_mapped(folio)) {
 526		rcu_read_unlock();
 527		put_anon_vma(anon_vma);
 528		return NULL;
 529	}
 530out:
 531	rcu_read_unlock();
 532
 533	return anon_vma;
 534}
 535
 536/*
 537 * Similar to folio_get_anon_vma() except it locks the anon_vma.
 538 *
 539 * Its a little more complex as it tries to keep the fast path to a single
 540 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
 541 * reference like with folio_get_anon_vma() and then block on the mutex
 542 * on !rwc->try_lock case.
 543 */
 544struct anon_vma *folio_lock_anon_vma_read(struct folio *folio,
 545					  struct rmap_walk_control *rwc)
 546{
 547	struct anon_vma *anon_vma = NULL;
 548	struct anon_vma *root_anon_vma;
 549	unsigned long anon_mapping;
 550
 551retry:
 552	rcu_read_lock();
 553	anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
 554	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 555		goto out;
 556	if (!folio_mapped(folio))
 557		goto out;
 558
 559	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 560	root_anon_vma = READ_ONCE(anon_vma->root);
 561	if (down_read_trylock(&root_anon_vma->rwsem)) {
 562		/*
 563		 * folio_move_anon_rmap() might have changed the anon_vma as we
 564		 * might not hold the folio lock here.
 565		 */
 566		if (unlikely((unsigned long)READ_ONCE(folio->mapping) !=
 567			     anon_mapping)) {
 568			up_read(&root_anon_vma->rwsem);
 569			rcu_read_unlock();
 570			goto retry;
 571		}
 572
 573		/*
 574		 * If the folio is still mapped, then this anon_vma is still
 575		 * its anon_vma, and holding the mutex ensures that it will
 576		 * not go away, see anon_vma_free().
 577		 */
 578		if (!folio_mapped(folio)) {
 579			up_read(&root_anon_vma->rwsem);
 580			anon_vma = NULL;
 581		}
 582		goto out;
 583	}
 584
 585	if (rwc && rwc->try_lock) {
 586		anon_vma = NULL;
 587		rwc->contended = true;
 588		goto out;
 589	}
 590
 591	/* trylock failed, we got to sleep */
 592	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 593		anon_vma = NULL;
 594		goto out;
 595	}
 596
 597	if (!folio_mapped(folio)) {
 598		rcu_read_unlock();
 599		put_anon_vma(anon_vma);
 600		return NULL;
 601	}
 602
 603	/* we pinned the anon_vma, its safe to sleep */
 604	rcu_read_unlock();
 605	anon_vma_lock_read(anon_vma);
 606
 607	/*
 608	 * folio_move_anon_rmap() might have changed the anon_vma as we might
 609	 * not hold the folio lock here.
 610	 */
 611	if (unlikely((unsigned long)READ_ONCE(folio->mapping) !=
 612		     anon_mapping)) {
 613		anon_vma_unlock_read(anon_vma);
 614		put_anon_vma(anon_vma);
 615		anon_vma = NULL;
 616		goto retry;
 617	}
 618
 619	if (atomic_dec_and_test(&anon_vma->refcount)) {
 620		/*
 621		 * Oops, we held the last refcount, release the lock
 622		 * and bail -- can't simply use put_anon_vma() because
 623		 * we'll deadlock on the anon_vma_lock_write() recursion.
 624		 */
 625		anon_vma_unlock_read(anon_vma);
 626		__put_anon_vma(anon_vma);
 627		anon_vma = NULL;
 628	}
 629
 630	return anon_vma;
 631
 632out:
 633	rcu_read_unlock();
 634	return anon_vma;
 635}
 636
 637#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
 638/*
 639 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
 640 * important if a PTE was dirty when it was unmapped that it's flushed
 641 * before any IO is initiated on the page to prevent lost writes. Similarly,
 642 * it must be flushed before freeing to prevent data leakage.
 643 */
 644void try_to_unmap_flush(void)
 645{
 646	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 647
 648	if (!tlb_ubc->flush_required)
 649		return;
 650
 651	arch_tlbbatch_flush(&tlb_ubc->arch);
 652	tlb_ubc->flush_required = false;
 653	tlb_ubc->writable = false;
 654}
 655
 656/* Flush iff there are potentially writable TLB entries that can race with IO */
 657void try_to_unmap_flush_dirty(void)
 658{
 659	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 660
 661	if (tlb_ubc->writable)
 662		try_to_unmap_flush();
 663}
 664
 665/*
 666 * Bits 0-14 of mm->tlb_flush_batched record pending generations.
 667 * Bits 16-30 of mm->tlb_flush_batched bit record flushed generations.
 668 */
 669#define TLB_FLUSH_BATCH_FLUSHED_SHIFT	16
 670#define TLB_FLUSH_BATCH_PENDING_MASK			\
 671	((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT - 1)) - 1)
 672#define TLB_FLUSH_BATCH_PENDING_LARGE			\
 673	(TLB_FLUSH_BATCH_PENDING_MASK / 2)
 674
 675static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval,
 676				      unsigned long uaddr)
 677{
 678	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 679	int batch;
 680	bool writable = pte_dirty(pteval);
 681
 682	if (!pte_accessible(mm, pteval))
 683		return;
 684
 685	arch_tlbbatch_add_pending(&tlb_ubc->arch, mm, uaddr);
 686	tlb_ubc->flush_required = true;
 687
 688	/*
 689	 * Ensure compiler does not re-order the setting of tlb_flush_batched
 690	 * before the PTE is cleared.
 691	 */
 692	barrier();
 693	batch = atomic_read(&mm->tlb_flush_batched);
 694retry:
 695	if ((batch & TLB_FLUSH_BATCH_PENDING_MASK) > TLB_FLUSH_BATCH_PENDING_LARGE) {
 696		/*
 697		 * Prevent `pending' from catching up with `flushed' because of
 698		 * overflow.  Reset `pending' and `flushed' to be 1 and 0 if
 699		 * `pending' becomes large.
 700		 */
 701		if (!atomic_try_cmpxchg(&mm->tlb_flush_batched, &batch, 1))
 
 
 702			goto retry;
 
 703	} else {
 704		atomic_inc(&mm->tlb_flush_batched);
 705	}
 706
 707	/*
 708	 * If the PTE was dirty then it's best to assume it's writable. The
 709	 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
 710	 * before the page is queued for IO.
 711	 */
 712	if (writable)
 713		tlb_ubc->writable = true;
 714}
 715
 716/*
 717 * Returns true if the TLB flush should be deferred to the end of a batch of
 718 * unmap operations to reduce IPIs.
 719 */
 720static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
 721{
 
 
 722	if (!(flags & TTU_BATCH_FLUSH))
 723		return false;
 724
 725	return arch_tlbbatch_should_defer(mm);
 
 
 
 
 
 726}
 727
 728/*
 729 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
 730 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
 731 * operation such as mprotect or munmap to race between reclaim unmapping
 732 * the page and flushing the page. If this race occurs, it potentially allows
 733 * access to data via a stale TLB entry. Tracking all mm's that have TLB
 734 * batching in flight would be expensive during reclaim so instead track
 735 * whether TLB batching occurred in the past and if so then do a flush here
 736 * if required. This will cost one additional flush per reclaim cycle paid
 737 * by the first operation at risk such as mprotect and mumap.
 738 *
 739 * This must be called under the PTL so that an access to tlb_flush_batched
 740 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
 741 * via the PTL.
 742 */
 743void flush_tlb_batched_pending(struct mm_struct *mm)
 744{
 745	int batch = atomic_read(&mm->tlb_flush_batched);
 746	int pending = batch & TLB_FLUSH_BATCH_PENDING_MASK;
 747	int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT;
 748
 749	if (pending != flushed) {
 750		arch_flush_tlb_batched_pending(mm);
 751		/*
 752		 * If the new TLB flushing is pending during flushing, leave
 753		 * mm->tlb_flush_batched as is, to avoid losing flushing.
 754		 */
 755		atomic_cmpxchg(&mm->tlb_flush_batched, batch,
 756			       pending | (pending << TLB_FLUSH_BATCH_FLUSHED_SHIFT));
 757	}
 758}
 759#else
 760static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval,
 761				      unsigned long uaddr)
 762{
 763}
 764
 765static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
 766{
 767	return false;
 768}
 769#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
 770
 771/*
 772 * At what user virtual address is page expected in vma?
 773 * Caller should check the page is actually part of the vma.
 774 */
 775unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
 776{
 777	struct folio *folio = page_folio(page);
 778	if (folio_test_anon(folio)) {
 779		struct anon_vma *page__anon_vma = folio_anon_vma(folio);
 780		/*
 781		 * Note: swapoff's unuse_vma() is more efficient with this
 782		 * check, and needs it to match anon_vma when KSM is active.
 783		 */
 784		if (!vma->anon_vma || !page__anon_vma ||
 785		    vma->anon_vma->root != page__anon_vma->root)
 786			return -EFAULT;
 787	} else if (!vma->vm_file) {
 788		return -EFAULT;
 789	} else if (vma->vm_file->f_mapping != folio->mapping) {
 790		return -EFAULT;
 791	}
 792
 793	return vma_address(page, vma);
 794}
 795
 796/*
 797 * Returns the actual pmd_t* where we expect 'address' to be mapped from, or
 798 * NULL if it doesn't exist.  No guarantees / checks on what the pmd_t*
 799 * represents.
 800 */
 801pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
 802{
 803	pgd_t *pgd;
 804	p4d_t *p4d;
 805	pud_t *pud;
 806	pmd_t *pmd = NULL;
 807
 808	pgd = pgd_offset(mm, address);
 809	if (!pgd_present(*pgd))
 810		goto out;
 811
 812	p4d = p4d_offset(pgd, address);
 813	if (!p4d_present(*p4d))
 814		goto out;
 815
 816	pud = pud_offset(p4d, address);
 817	if (!pud_present(*pud))
 818		goto out;
 819
 820	pmd = pmd_offset(pud, address);
 821out:
 822	return pmd;
 823}
 824
 825struct folio_referenced_arg {
 826	int mapcount;
 827	int referenced;
 828	unsigned long vm_flags;
 829	struct mem_cgroup *memcg;
 830};
 831
 832/*
 833 * arg: folio_referenced_arg will be passed
 834 */
 835static bool folio_referenced_one(struct folio *folio,
 836		struct vm_area_struct *vma, unsigned long address, void *arg)
 837{
 838	struct folio_referenced_arg *pra = arg;
 839	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
 840	int referenced = 0;
 841	unsigned long start = address, ptes = 0;
 842
 843	while (page_vma_mapped_walk(&pvmw)) {
 844		address = pvmw.address;
 845
 846		if (vma->vm_flags & VM_LOCKED) {
 847			if (!folio_test_large(folio) || !pvmw.pte) {
 848				/* Restore the mlock which got missed */
 849				mlock_vma_folio(folio, vma);
 850				page_vma_mapped_walk_done(&pvmw);
 851				pra->vm_flags |= VM_LOCKED;
 852				return false; /* To break the loop */
 853			}
 854			/*
 855			 * For large folio fully mapped to VMA, will
 856			 * be handled after the pvmw loop.
 857			 *
 858			 * For large folio cross VMA boundaries, it's
 859			 * expected to be picked  by page reclaim. But
 860			 * should skip reference of pages which are in
 861			 * the range of VM_LOCKED vma. As page reclaim
 862			 * should just count the reference of pages out
 863			 * the range of VM_LOCKED vma.
 864			 */
 865			ptes++;
 866			pra->mapcount--;
 867			continue;
 868		}
 869
 870		if (pvmw.pte) {
 871			if (lru_gen_enabled() &&
 872			    pte_young(ptep_get(pvmw.pte))) {
 873				lru_gen_look_around(&pvmw);
 874				referenced++;
 875			}
 876
 877			if (ptep_clear_flush_young_notify(vma, address,
 878						pvmw.pte))
 879				referenced++;
 
 
 
 
 
 
 
 
 
 
 880		} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
 881			if (pmdp_clear_flush_young_notify(vma, address,
 882						pvmw.pmd))
 883				referenced++;
 884		} else {
 885			/* unexpected pmd-mapped folio? */
 886			WARN_ON_ONCE(1);
 887		}
 888
 889		pra->mapcount--;
 890	}
 891
 892	if ((vma->vm_flags & VM_LOCKED) &&
 893			folio_test_large(folio) &&
 894			folio_within_vma(folio, vma)) {
 895		unsigned long s_align, e_align;
 896
 897		s_align = ALIGN_DOWN(start, PMD_SIZE);
 898		e_align = ALIGN_DOWN(start + folio_size(folio) - 1, PMD_SIZE);
 899
 900		/* folio doesn't cross page table boundary and fully mapped */
 901		if ((s_align == e_align) && (ptes == folio_nr_pages(folio))) {
 902			/* Restore the mlock which got missed */
 903			mlock_vma_folio(folio, vma);
 904			pra->vm_flags |= VM_LOCKED;
 905			return false; /* To break the loop */
 906		}
 907	}
 908
 909	if (referenced)
 910		folio_clear_idle(folio);
 911	if (folio_test_clear_young(folio))
 912		referenced++;
 913
 914	if (referenced) {
 915		pra->referenced++;
 916		pra->vm_flags |= vma->vm_flags & ~VM_LOCKED;
 917	}
 918
 919	if (!pra->mapcount)
 920		return false; /* To break the loop */
 921
 922	return true;
 923}
 924
 925static bool invalid_folio_referenced_vma(struct vm_area_struct *vma, void *arg)
 926{
 927	struct folio_referenced_arg *pra = arg;
 928	struct mem_cgroup *memcg = pra->memcg;
 929
 930	/*
 931	 * Ignore references from this mapping if it has no recency. If the
 932	 * folio has been used in another mapping, we will catch it; if this
 933	 * other mapping is already gone, the unmap path will have set the
 934	 * referenced flag or activated the folio in zap_pte_range().
 935	 */
 936	if (!vma_has_recency(vma))
 937		return true;
 938
 939	/*
 940	 * If we are reclaiming on behalf of a cgroup, skip counting on behalf
 941	 * of references from different cgroups.
 942	 */
 943	if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
 944		return true;
 945
 946	return false;
 947}
 948
 949/**
 950 * folio_referenced() - Test if the folio was referenced.
 951 * @folio: The folio to test.
 952 * @is_locked: Caller holds lock on the folio.
 953 * @memcg: target memory cgroup
 954 * @vm_flags: A combination of all the vma->vm_flags which referenced the folio.
 955 *
 956 * Quick test_and_clear_referenced for all mappings of a folio,
 957 *
 958 * Return: The number of mappings which referenced the folio. Return -1 if
 959 * the function bailed out due to rmap lock contention.
 960 */
 961int folio_referenced(struct folio *folio, int is_locked,
 962		     struct mem_cgroup *memcg, unsigned long *vm_flags)
 963{
 964	int we_locked = 0;
 965	struct folio_referenced_arg pra = {
 966		.mapcount = folio_mapcount(folio),
 967		.memcg = memcg,
 968	};
 969	struct rmap_walk_control rwc = {
 970		.rmap_one = folio_referenced_one,
 971		.arg = (void *)&pra,
 972		.anon_lock = folio_lock_anon_vma_read,
 973		.try_lock = true,
 974		.invalid_vma = invalid_folio_referenced_vma,
 975	};
 976
 977	*vm_flags = 0;
 978	if (!pra.mapcount)
 979		return 0;
 980
 981	if (!folio_raw_mapping(folio))
 982		return 0;
 983
 984	if (!is_locked && (!folio_test_anon(folio) || folio_test_ksm(folio))) {
 985		we_locked = folio_trylock(folio);
 986		if (!we_locked)
 987			return 1;
 988	}
 989
 
 
 
 
 
 
 
 
 
 990	rmap_walk(folio, &rwc);
 991	*vm_flags = pra.vm_flags;
 992
 993	if (we_locked)
 994		folio_unlock(folio);
 995
 996	return rwc.contended ? -1 : pra.referenced;
 997}
 998
 999static int page_vma_mkclean_one(struct page_vma_mapped_walk *pvmw)
1000{
1001	int cleaned = 0;
1002	struct vm_area_struct *vma = pvmw->vma;
1003	struct mmu_notifier_range range;
1004	unsigned long address = pvmw->address;
1005
1006	/*
1007	 * We have to assume the worse case ie pmd for invalidation. Note that
1008	 * the folio can not be freed from this function.
1009	 */
1010	mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 0,
1011				vma->vm_mm, address, vma_address_end(pvmw));
 
1012	mmu_notifier_invalidate_range_start(&range);
1013
1014	while (page_vma_mapped_walk(pvmw)) {
1015		int ret = 0;
1016
1017		address = pvmw->address;
1018		if (pvmw->pte) {
 
1019			pte_t *pte = pvmw->pte;
1020			pte_t entry = ptep_get(pte);
1021
1022			if (!pte_dirty(entry) && !pte_write(entry))
1023				continue;
1024
1025			flush_cache_page(vma, address, pte_pfn(entry));
1026			entry = ptep_clear_flush(vma, address, pte);
1027			entry = pte_wrprotect(entry);
1028			entry = pte_mkclean(entry);
1029			set_pte_at(vma->vm_mm, address, pte, entry);
1030			ret = 1;
1031		} else {
1032#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1033			pmd_t *pmd = pvmw->pmd;
1034			pmd_t entry;
1035
1036			if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
1037				continue;
1038
1039			flush_cache_range(vma, address,
1040					  address + HPAGE_PMD_SIZE);
1041			entry = pmdp_invalidate(vma, address, pmd);
1042			entry = pmd_wrprotect(entry);
1043			entry = pmd_mkclean(entry);
1044			set_pmd_at(vma->vm_mm, address, pmd, entry);
1045			ret = 1;
1046#else
1047			/* unexpected pmd-mapped folio? */
1048			WARN_ON_ONCE(1);
1049#endif
1050		}
1051
 
 
 
 
 
 
 
1052		if (ret)
1053			cleaned++;
1054	}
1055
1056	mmu_notifier_invalidate_range_end(&range);
1057
1058	return cleaned;
1059}
1060
1061static bool page_mkclean_one(struct folio *folio, struct vm_area_struct *vma,
1062			     unsigned long address, void *arg)
1063{
1064	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC);
1065	int *cleaned = arg;
1066
1067	*cleaned += page_vma_mkclean_one(&pvmw);
1068
1069	return true;
1070}
1071
1072static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1073{
1074	if (vma->vm_flags & VM_SHARED)
1075		return false;
1076
1077	return true;
1078}
1079
1080int folio_mkclean(struct folio *folio)
1081{
1082	int cleaned = 0;
1083	struct address_space *mapping;
1084	struct rmap_walk_control rwc = {
1085		.arg = (void *)&cleaned,
1086		.rmap_one = page_mkclean_one,
1087		.invalid_vma = invalid_mkclean_vma,
1088	};
1089
1090	BUG_ON(!folio_test_locked(folio));
1091
1092	if (!folio_mapped(folio))
1093		return 0;
1094
1095	mapping = folio_mapping(folio);
1096	if (!mapping)
1097		return 0;
1098
1099	rmap_walk(folio, &rwc);
1100
1101	return cleaned;
1102}
1103EXPORT_SYMBOL_GPL(folio_mkclean);
1104
1105/**
1106 * pfn_mkclean_range - Cleans the PTEs (including PMDs) mapped with range of
1107 *                     [@pfn, @pfn + @nr_pages) at the specific offset (@pgoff)
1108 *                     within the @vma of shared mappings. And since clean PTEs
1109 *                     should also be readonly, write protects them too.
1110 * @pfn: start pfn.
1111 * @nr_pages: number of physically contiguous pages srarting with @pfn.
1112 * @pgoff: page offset that the @pfn mapped with.
1113 * @vma: vma that @pfn mapped within.
1114 *
1115 * Returns the number of cleaned PTEs (including PMDs).
1116 */
1117int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff,
1118		      struct vm_area_struct *vma)
1119{
1120	struct page_vma_mapped_walk pvmw = {
1121		.pfn		= pfn,
1122		.nr_pages	= nr_pages,
1123		.pgoff		= pgoff,
1124		.vma		= vma,
1125		.flags		= PVMW_SYNC,
1126	};
1127
1128	if (invalid_mkclean_vma(vma, NULL))
1129		return 0;
1130
1131	pvmw.address = vma_pgoff_address(pgoff, nr_pages, vma);
1132	VM_BUG_ON_VMA(pvmw.address == -EFAULT, vma);
1133
1134	return page_vma_mkclean_one(&pvmw);
1135}
1136
1137int folio_total_mapcount(struct folio *folio)
1138{
1139	int mapcount = folio_entire_mapcount(folio);
1140	int nr_pages;
1141	int i;
1142
1143	/* In the common case, avoid the loop when no pages mapped by PTE */
1144	if (folio_nr_pages_mapped(folio) == 0)
1145		return mapcount;
1146	/*
1147	 * Add all the PTE mappings of those pages mapped by PTE.
1148	 * Limit the loop to folio_nr_pages_mapped()?
1149	 * Perhaps: given all the raciness, that may be a good or a bad idea.
1150	 */
1151	nr_pages = folio_nr_pages(folio);
1152	for (i = 0; i < nr_pages; i++)
1153		mapcount += atomic_read(&folio_page(folio, i)->_mapcount);
1154
1155	/* But each of those _mapcounts was based on -1 */
1156	mapcount += nr_pages;
1157	return mapcount;
1158}
1159
1160static __always_inline unsigned int __folio_add_rmap(struct folio *folio,
1161		struct page *page, int nr_pages, enum rmap_level level,
1162		int *nr_pmdmapped)
1163{
1164	atomic_t *mapped = &folio->_nr_pages_mapped;
1165	int first, nr = 0;
1166
1167	__folio_rmap_sanity_checks(folio, page, nr_pages, level);
1168
1169	switch (level) {
1170	case RMAP_LEVEL_PTE:
1171		do {
1172			first = atomic_inc_and_test(&page->_mapcount);
1173			if (first && folio_test_large(folio)) {
1174				first = atomic_inc_return_relaxed(mapped);
1175				first = (first < ENTIRELY_MAPPED);
1176			}
1177
1178			if (first)
1179				nr++;
1180		} while (page++, --nr_pages > 0);
1181		break;
1182	case RMAP_LEVEL_PMD:
1183		first = atomic_inc_and_test(&folio->_entire_mapcount);
1184		if (first) {
1185			nr = atomic_add_return_relaxed(ENTIRELY_MAPPED, mapped);
1186			if (likely(nr < ENTIRELY_MAPPED + ENTIRELY_MAPPED)) {
1187				*nr_pmdmapped = folio_nr_pages(folio);
1188				nr = *nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED);
1189				/* Raced ahead of a remove and another add? */
1190				if (unlikely(nr < 0))
1191					nr = 0;
1192			} else {
1193				/* Raced ahead of a remove of ENTIRELY_MAPPED */
1194				nr = 0;
1195			}
1196		}
1197		break;
1198	}
1199	return nr;
1200}
1201
1202/**
1203 * folio_move_anon_rmap - move a folio to our anon_vma
1204 * @folio:	The folio to move to our anon_vma
1205 * @vma:	The vma the folio belongs to
1206 *
1207 * When a folio belongs exclusively to one process after a COW event,
1208 * that folio can be moved into the anon_vma that belongs to just that
1209 * process, so the rmap code will not search the parent or sibling processes.
 
1210 */
1211void folio_move_anon_rmap(struct folio *folio, struct vm_area_struct *vma)
1212{
1213	void *anon_vma = vma->anon_vma;
 
1214
1215	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1216	VM_BUG_ON_VMA(!anon_vma, vma);
1217
1218	anon_vma += PAGE_MAPPING_ANON;
1219	/*
1220	 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1221	 * simultaneously, so a concurrent reader (eg folio_referenced()'s
1222	 * folio_test_anon()) will not see one without the other.
1223	 */
1224	WRITE_ONCE(folio->mapping, anon_vma);
 
1225}
1226
1227/**
1228 * __folio_set_anon - set up a new anonymous rmap for a folio
1229 * @folio:	The folio to set up the new anonymous rmap for.
1230 * @vma:	VM area to add the folio to.
1231 * @address:	User virtual address of the mapping
1232 * @exclusive:	Whether the folio is exclusive to the process.
1233 */
1234static void __folio_set_anon(struct folio *folio, struct vm_area_struct *vma,
1235			     unsigned long address, bool exclusive)
1236{
1237	struct anon_vma *anon_vma = vma->anon_vma;
1238
1239	BUG_ON(!anon_vma);
1240
 
 
 
1241	/*
1242	 * If the folio isn't exclusive to this vma, we must use the _oldest_
1243	 * possible anon_vma for the folio mapping!
 
1244	 */
1245	if (!exclusive)
1246		anon_vma = anon_vma->root;
1247
1248	/*
1249	 * page_idle does a lockless/optimistic rmap scan on folio->mapping.
1250	 * Make sure the compiler doesn't split the stores of anon_vma and
1251	 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1252	 * could mistake the mapping for a struct address_space and crash.
1253	 */
1254	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1255	WRITE_ONCE(folio->mapping, (struct address_space *) anon_vma);
1256	folio->index = linear_page_index(vma, address);
 
 
 
1257}
1258
1259/**
1260 * __page_check_anon_rmap - sanity check anonymous rmap addition
1261 * @folio:	The folio containing @page.
1262 * @page:	the page to check the mapping of
1263 * @vma:	the vm area in which the mapping is added
1264 * @address:	the user virtual address mapped
1265 */
1266static void __page_check_anon_rmap(struct folio *folio, struct page *page,
1267	struct vm_area_struct *vma, unsigned long address)
1268{
 
1269	/*
1270	 * The page's anon-rmap details (mapping and index) are guaranteed to
1271	 * be set up correctly at this point.
1272	 *
1273	 * We have exclusion against folio_add_anon_rmap_*() because the caller
1274	 * always holds the page locked.
1275	 *
1276	 * We have exclusion against folio_add_new_anon_rmap because those pages
1277	 * are initially only visible via the pagetables, and the pte is locked
1278	 * over the call to folio_add_new_anon_rmap.
1279	 */
1280	VM_BUG_ON_FOLIO(folio_anon_vma(folio)->root != vma->anon_vma->root,
1281			folio);
1282	VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1283		       page);
1284}
1285
1286static __always_inline void __folio_add_anon_rmap(struct folio *folio,
1287		struct page *page, int nr_pages, struct vm_area_struct *vma,
1288		unsigned long address, rmap_t flags, enum rmap_level level)
 
 
 
 
 
 
 
 
 
 
 
1289{
1290	int i, nr, nr_pmdmapped = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1291
1292	nr = __folio_add_rmap(folio, page, nr_pages, level, &nr_pmdmapped);
1293	if (nr_pmdmapped)
1294		__lruvec_stat_mod_folio(folio, NR_ANON_THPS, nr_pmdmapped);
1295	if (nr)
1296		__lruvec_stat_mod_folio(folio, NR_ANON_MAPPED, nr);
1297
1298	if (unlikely(!folio_test_anon(folio))) {
1299		VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
1300		/*
1301		 * For a PTE-mapped large folio, we only know that the single
1302		 * PTE is exclusive. Further, __folio_set_anon() might not get
1303		 * folio->index right when not given the address of the head
1304		 * page.
1305		 */
1306		VM_WARN_ON_FOLIO(folio_test_large(folio) &&
1307				 level != RMAP_LEVEL_PMD, folio);
1308		__folio_set_anon(folio, vma, address,
1309				 !!(flags & RMAP_EXCLUSIVE));
1310	} else if (likely(!folio_test_ksm(folio))) {
1311		__page_check_anon_rmap(folio, page, vma, address);
1312	}
1313
1314	if (flags & RMAP_EXCLUSIVE) {
1315		switch (level) {
1316		case RMAP_LEVEL_PTE:
1317			for (i = 0; i < nr_pages; i++)
1318				SetPageAnonExclusive(page + i);
1319			break;
1320		case RMAP_LEVEL_PMD:
1321			SetPageAnonExclusive(page);
1322			break;
1323		}
1324	}
1325	for (i = 0; i < nr_pages; i++) {
1326		struct page *cur_page = page + i;
1327
1328		/* While PTE-mapping a THP we have a PMD and a PTE mapping. */
1329		VM_WARN_ON_FOLIO((atomic_read(&cur_page->_mapcount) > 0 ||
1330				  (folio_test_large(folio) &&
1331				   folio_entire_mapcount(folio) > 1)) &&
1332				 PageAnonExclusive(cur_page), folio);
1333	}
1334
1335	/*
1336	 * For large folio, only mlock it if it's fully mapped to VMA. It's
1337	 * not easy to check whether the large folio is fully mapped to VMA
1338	 * here. Only mlock normal 4K folio and leave page reclaim to handle
1339	 * large folio.
1340	 */
1341	if (!folio_test_large(folio))
1342		mlock_vma_folio(folio, vma);
1343}
1344
1345/**
1346 * folio_add_anon_rmap_ptes - add PTE mappings to a page range of an anon folio
1347 * @folio:	The folio to add the mappings to
1348 * @page:	The first page to add
1349 * @nr_pages:	The number of pages which will be mapped
1350 * @vma:	The vm area in which the mappings are added
1351 * @address:	The user virtual address of the first page to map
1352 * @flags:	The rmap flags
1353 *
1354 * The page range of folio is defined by [first_page, first_page + nr_pages)
1355 *
1356 * The caller needs to hold the page table lock, and the page must be locked in
1357 * the anon_vma case: to serialize mapping,index checking after setting,
1358 * and to ensure that an anon folio is not being upgraded racily to a KSM folio
1359 * (but KSM folios are never downgraded).
1360 */
1361void folio_add_anon_rmap_ptes(struct folio *folio, struct page *page,
1362		int nr_pages, struct vm_area_struct *vma, unsigned long address,
1363		rmap_t flags)
1364{
1365	__folio_add_anon_rmap(folio, page, nr_pages, vma, address, flags,
1366			      RMAP_LEVEL_PTE);
1367}
1368
1369/**
1370 * folio_add_anon_rmap_pmd - add a PMD mapping to a page range of an anon folio
1371 * @folio:	The folio to add the mapping to
1372 * @page:	The first page to add
1373 * @vma:	The vm area in which the mapping is added
1374 * @address:	The user virtual address of the first page to map
1375 * @flags:	The rmap flags
1376 *
1377 * The page range of folio is defined by [first_page, first_page + HPAGE_PMD_NR)
1378 *
1379 * The caller needs to hold the page table lock, and the page must be locked in
1380 * the anon_vma case: to serialize mapping,index checking after setting.
1381 */
1382void folio_add_anon_rmap_pmd(struct folio *folio, struct page *page,
1383		struct vm_area_struct *vma, unsigned long address, rmap_t flags)
1384{
1385#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1386	__folio_add_anon_rmap(folio, page, HPAGE_PMD_NR, vma, address, flags,
1387			      RMAP_LEVEL_PMD);
1388#else
1389	WARN_ON_ONCE(true);
1390#endif
1391}
1392
1393/**
1394 * folio_add_new_anon_rmap - Add mapping to a new anonymous folio.
1395 * @folio:	The folio to add the mapping to.
1396 * @vma:	the vm area in which the mapping is added
1397 * @address:	the user virtual address mapped
1398 *
1399 * Like folio_add_anon_rmap_*() but must only be called on *new* folios.
 
 
 
1400 * This means the inc-and-test can be bypassed.
1401 * The folio does not have to be locked.
1402 *
1403 * If the folio is pmd-mappable, it is accounted as a THP.  As the folio
1404 * is new, it's assumed to be mapped exclusively by a single process.
1405 */
1406void folio_add_new_anon_rmap(struct folio *folio, struct vm_area_struct *vma,
1407		unsigned long address)
1408{
1409	int nr = folio_nr_pages(folio);
1410
1411	VM_WARN_ON_FOLIO(folio_test_hugetlb(folio), folio);
1412	VM_BUG_ON_VMA(address < vma->vm_start ||
1413			address + (nr << PAGE_SHIFT) > vma->vm_end, vma);
1414	__folio_set_swapbacked(folio);
1415	__folio_set_anon(folio, vma, address, true);
1416
1417	if (likely(!folio_test_large(folio))) {
1418		/* increment count (starts at -1) */
1419		atomic_set(&folio->_mapcount, 0);
1420		SetPageAnonExclusive(&folio->page);
1421	} else if (!folio_test_pmd_mappable(folio)) {
1422		int i;
1423
1424		for (i = 0; i < nr; i++) {
1425			struct page *page = folio_page(folio, i);
1426
1427			/* increment count (starts at -1) */
1428			atomic_set(&page->_mapcount, 0);
1429			SetPageAnonExclusive(page);
1430		}
1431
1432		atomic_set(&folio->_nr_pages_mapped, nr);
1433	} else {
 
1434		/* increment count (starts at -1) */
1435		atomic_set(&folio->_entire_mapcount, 0);
1436		atomic_set(&folio->_nr_pages_mapped, ENTIRELY_MAPPED);
1437		SetPageAnonExclusive(&folio->page);
1438		__lruvec_stat_mod_folio(folio, NR_ANON_THPS, nr);
1439	}
1440
1441	__lruvec_stat_mod_folio(folio, NR_ANON_MAPPED, nr);
 
1442}
1443
1444static __always_inline void __folio_add_file_rmap(struct folio *folio,
1445		struct page *page, int nr_pages, struct vm_area_struct *vma,
1446		enum rmap_level level)
 
 
 
 
 
 
 
1447{
1448	int nr, nr_pmdmapped = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1449
1450	VM_WARN_ON_FOLIO(folio_test_anon(folio), folio);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1451
1452	nr = __folio_add_rmap(folio, page, nr_pages, level, &nr_pmdmapped);
1453	if (nr_pmdmapped)
1454		__lruvec_stat_mod_folio(folio, folio_test_swapbacked(folio) ?
1455			NR_SHMEM_PMDMAPPED : NR_FILE_PMDMAPPED, nr_pmdmapped);
1456	if (nr)
1457		__lruvec_stat_mod_folio(folio, NR_FILE_MAPPED, nr);
 
1458
1459	/* See comments in folio_add_anon_rmap_*() */
1460	if (!folio_test_large(folio))
1461		mlock_vma_folio(folio, vma);
1462}
1463
1464/**
1465 * folio_add_file_rmap_ptes - add PTE mappings to a page range of a folio
1466 * @folio:	The folio to add the mappings to
1467 * @page:	The first page to add
1468 * @nr_pages:	The number of pages that will be mapped using PTEs
1469 * @vma:	The vm area in which the mappings are added
1470 *
1471 * The page range of the folio is defined by [page, page + nr_pages)
1472 *
1473 * The caller needs to hold the page table lock.
1474 */
1475void folio_add_file_rmap_ptes(struct folio *folio, struct page *page,
1476		int nr_pages, struct vm_area_struct *vma)
1477{
1478	__folio_add_file_rmap(folio, page, nr_pages, vma, RMAP_LEVEL_PTE);
1479}
 
 
 
 
 
 
 
1480
1481/**
1482 * folio_add_file_rmap_pmd - add a PMD mapping to a page range of a folio
1483 * @folio:	The folio to add the mapping to
1484 * @page:	The first page to add
1485 * @vma:	The vm area in which the mapping is added
1486 *
1487 * The page range of the folio is defined by [page, page + HPAGE_PMD_NR)
1488 *
1489 * The caller needs to hold the page table lock.
1490 */
1491void folio_add_file_rmap_pmd(struct folio *folio, struct page *page,
1492		struct vm_area_struct *vma)
1493{
1494#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1495	__folio_add_file_rmap(folio, page, HPAGE_PMD_NR, vma, RMAP_LEVEL_PMD);
1496#else
1497	WARN_ON_ONCE(true);
1498#endif
1499}
1500
1501static __always_inline void __folio_remove_rmap(struct folio *folio,
1502		struct page *page, int nr_pages, struct vm_area_struct *vma,
1503		enum rmap_level level)
1504{
1505	atomic_t *mapped = &folio->_nr_pages_mapped;
1506	int last, nr = 0, nr_pmdmapped = 0;
1507	enum node_stat_item idx;
1508
1509	__folio_rmap_sanity_checks(folio, page, nr_pages, level);
1510
1511	switch (level) {
1512	case RMAP_LEVEL_PTE:
1513		do {
1514			last = atomic_add_negative(-1, &page->_mapcount);
1515			if (last && folio_test_large(folio)) {
1516				last = atomic_dec_return_relaxed(mapped);
1517				last = (last < ENTIRELY_MAPPED);
1518			}
1519
1520			if (last)
1521				nr++;
1522		} while (page++, --nr_pages > 0);
1523		break;
1524	case RMAP_LEVEL_PMD:
1525		last = atomic_add_negative(-1, &folio->_entire_mapcount);
1526		if (last) {
1527			nr = atomic_sub_return_relaxed(ENTIRELY_MAPPED, mapped);
1528			if (likely(nr < ENTIRELY_MAPPED)) {
1529				nr_pmdmapped = folio_nr_pages(folio);
1530				nr = nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED);
 
1531				/* Raced ahead of another remove and an add? */
1532				if (unlikely(nr < 0))
1533					nr = 0;
1534			} else {
1535				/* An add of ENTIRELY_MAPPED raced ahead */
1536				nr = 0;
1537			}
1538		}
1539		break;
1540	}
1541
1542	if (nr_pmdmapped) {
1543		if (folio_test_anon(folio))
1544			idx = NR_ANON_THPS;
1545		else if (folio_test_swapbacked(folio))
1546			idx = NR_SHMEM_PMDMAPPED;
1547		else
1548			idx = NR_FILE_PMDMAPPED;
1549		__lruvec_stat_mod_folio(folio, idx, -nr_pmdmapped);
1550	}
1551	if (nr) {
1552		idx = folio_test_anon(folio) ? NR_ANON_MAPPED : NR_FILE_MAPPED;
1553		__lruvec_stat_mod_folio(folio, idx, -nr);
1554
1555		/*
1556		 * Queue anon large folio for deferred split if at least one
1557		 * page of the folio is unmapped and at least one page
1558		 * is still mapped.
1559		 */
1560		if (folio_test_large(folio) && folio_test_anon(folio))
1561			if (level == RMAP_LEVEL_PTE || nr < nr_pmdmapped)
1562				deferred_split_folio(folio);
1563	}
1564
1565	/*
1566	 * It would be tidy to reset folio_test_anon mapping when fully
1567	 * unmapped, but that might overwrite a racing folio_add_anon_rmap_*()
1568	 * which increments mapcount after us but sets mapping before us:
1569	 * so leave the reset to free_pages_prepare, and remember that
1570	 * it's only reliable while mapped.
1571	 */
1572
1573	munlock_vma_folio(folio, vma);
1574}
1575
1576/**
1577 * folio_remove_rmap_ptes - remove PTE mappings from a page range of a folio
1578 * @folio:	The folio to remove the mappings from
1579 * @page:	The first page to remove
1580 * @nr_pages:	The number of pages that will be removed from the mapping
1581 * @vma:	The vm area from which the mappings are removed
1582 *
1583 * The page range of the folio is defined by [page, page + nr_pages)
1584 *
1585 * The caller needs to hold the page table lock.
1586 */
1587void folio_remove_rmap_ptes(struct folio *folio, struct page *page,
1588		int nr_pages, struct vm_area_struct *vma)
1589{
1590	__folio_remove_rmap(folio, page, nr_pages, vma, RMAP_LEVEL_PTE);
1591}
1592
1593/**
1594 * folio_remove_rmap_pmd - remove a PMD mapping from a page range of a folio
1595 * @folio:	The folio to remove the mapping from
1596 * @page:	The first page to remove
1597 * @vma:	The vm area from which the mapping is removed
1598 *
1599 * The page range of the folio is defined by [page, page + HPAGE_PMD_NR)
1600 *
1601 * The caller needs to hold the page table lock.
1602 */
1603void folio_remove_rmap_pmd(struct folio *folio, struct page *page,
1604		struct vm_area_struct *vma)
1605{
1606#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1607	__folio_remove_rmap(folio, page, HPAGE_PMD_NR, vma, RMAP_LEVEL_PMD);
1608#else
1609	WARN_ON_ONCE(true);
1610#endif
1611}
1612
1613/*
1614 * @arg: enum ttu_flags will be passed to this argument
1615 */
1616static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
1617		     unsigned long address, void *arg)
1618{
1619	struct mm_struct *mm = vma->vm_mm;
1620	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1621	pte_t pteval;
1622	struct page *subpage;
1623	bool anon_exclusive, ret = true;
1624	struct mmu_notifier_range range;
1625	enum ttu_flags flags = (enum ttu_flags)(long)arg;
1626	unsigned long pfn;
1627	unsigned long hsz = 0;
1628
1629	/*
1630	 * When racing against e.g. zap_pte_range() on another cpu,
1631	 * in between its ptep_get_and_clear_full() and folio_remove_rmap_*(),
1632	 * try_to_unmap() may return before page_mapped() has become false,
1633	 * if page table locking is skipped: use TTU_SYNC to wait for that.
1634	 */
1635	if (flags & TTU_SYNC)
1636		pvmw.flags = PVMW_SYNC;
1637
1638	if (flags & TTU_SPLIT_HUGE_PMD)
1639		split_huge_pmd_address(vma, address, false, folio);
1640
1641	/*
1642	 * For THP, we have to assume the worse case ie pmd for invalidation.
1643	 * For hugetlb, it could be much worse if we need to do pud
1644	 * invalidation in the case of pmd sharing.
1645	 *
1646	 * Note that the folio can not be freed in this function as call of
1647	 * try_to_unmap() must hold a reference on the folio.
1648	 */
1649	range.end = vma_address_end(&pvmw);
1650	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1651				address, range.end);
1652	if (folio_test_hugetlb(folio)) {
1653		/*
1654		 * If sharing is possible, start and end will be adjusted
1655		 * accordingly.
1656		 */
1657		adjust_range_if_pmd_sharing_possible(vma, &range.start,
1658						     &range.end);
1659
1660		/* We need the huge page size for set_huge_pte_at() */
1661		hsz = huge_page_size(hstate_vma(vma));
1662	}
1663	mmu_notifier_invalidate_range_start(&range);
1664
1665	while (page_vma_mapped_walk(&pvmw)) {
1666		/* Unexpected PMD-mapped THP? */
1667		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
1668
1669		/*
1670		 * If the folio is in an mlock()d vma, we must not swap it out.
1671		 */
1672		if (!(flags & TTU_IGNORE_MLOCK) &&
1673		    (vma->vm_flags & VM_LOCKED)) {
1674			/* Restore the mlock which got missed */
1675			if (!folio_test_large(folio))
1676				mlock_vma_folio(folio, vma);
1677			page_vma_mapped_walk_done(&pvmw);
1678			ret = false;
1679			break;
1680		}
1681
1682		pfn = pte_pfn(ptep_get(pvmw.pte));
1683		subpage = folio_page(folio, pfn - folio_pfn(folio));
1684		address = pvmw.address;
1685		anon_exclusive = folio_test_anon(folio) &&
1686				 PageAnonExclusive(subpage);
1687
1688		if (folio_test_hugetlb(folio)) {
1689			bool anon = folio_test_anon(folio);
1690
1691			/*
1692			 * The try_to_unmap() is only passed a hugetlb page
1693			 * in the case where the hugetlb page is poisoned.
1694			 */
1695			VM_BUG_ON_PAGE(!PageHWPoison(subpage), subpage);
1696			/*
1697			 * huge_pmd_unshare may unmap an entire PMD page.
1698			 * There is no way of knowing exactly which PMDs may
1699			 * be cached for this mm, so we must flush them all.
1700			 * start/end were already adjusted above to cover this
1701			 * range.
1702			 */
1703			flush_cache_range(vma, range.start, range.end);
1704
1705			/*
1706			 * To call huge_pmd_unshare, i_mmap_rwsem must be
1707			 * held in write mode.  Caller needs to explicitly
1708			 * do this outside rmap routines.
1709			 *
1710			 * We also must hold hugetlb vma_lock in write mode.
1711			 * Lock order dictates acquiring vma_lock BEFORE
1712			 * i_mmap_rwsem.  We can only try lock here and fail
1713			 * if unsuccessful.
1714			 */
1715			if (!anon) {
1716				VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1717				if (!hugetlb_vma_trylock_write(vma)) {
1718					page_vma_mapped_walk_done(&pvmw);
1719					ret = false;
1720					break;
1721				}
1722				if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1723					hugetlb_vma_unlock_write(vma);
1724					flush_tlb_range(vma,
1725						range.start, range.end);
 
 
1726					/*
1727					 * The ref count of the PMD page was
1728					 * dropped which is part of the way map
1729					 * counting is done for shared PMDs.
1730					 * Return 'true' here.  When there is
1731					 * no other sharing, huge_pmd_unshare
1732					 * returns false and we will unmap the
1733					 * actual page and drop map count
1734					 * to zero.
1735					 */
1736					page_vma_mapped_walk_done(&pvmw);
1737					break;
1738				}
1739				hugetlb_vma_unlock_write(vma);
1740			}
1741			pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
1742		} else {
1743			flush_cache_page(vma, address, pfn);
1744			/* Nuke the page table entry. */
1745			if (should_defer_flush(mm, flags)) {
1746				/*
1747				 * We clear the PTE but do not flush so potentially
1748				 * a remote CPU could still be writing to the folio.
1749				 * If the entry was previously clean then the
1750				 * architecture must guarantee that a clear->dirty
1751				 * transition on a cached TLB entry is written through
1752				 * and traps if the PTE is unmapped.
1753				 */
1754				pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1755
1756				set_tlb_ubc_flush_pending(mm, pteval, address);
1757			} else {
1758				pteval = ptep_clear_flush(vma, address, pvmw.pte);
1759			}
1760		}
1761
1762		/*
1763		 * Now the pte is cleared. If this pte was uffd-wp armed,
1764		 * we may want to replace a none pte with a marker pte if
1765		 * it's file-backed, so we don't lose the tracking info.
1766		 */
1767		pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval);
1768
1769		/* Set the dirty flag on the folio now the pte is gone. */
1770		if (pte_dirty(pteval))
1771			folio_mark_dirty(folio);
1772
1773		/* Update high watermark before we lower rss */
1774		update_hiwater_rss(mm);
1775
1776		if (PageHWPoison(subpage) && (flags & TTU_HWPOISON)) {
1777			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1778			if (folio_test_hugetlb(folio)) {
1779				hugetlb_count_sub(folio_nr_pages(folio), mm);
1780				set_huge_pte_at(mm, address, pvmw.pte, pteval,
1781						hsz);
1782			} else {
1783				dec_mm_counter(mm, mm_counter(&folio->page));
1784				set_pte_at(mm, address, pvmw.pte, pteval);
1785			}
1786
1787		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1788			/*
1789			 * The guest indicated that the page content is of no
1790			 * interest anymore. Simply discard the pte, vmscan
1791			 * will take care of the rest.
1792			 * A future reference will then fault in a new zero
1793			 * page. When userfaultfd is active, we must not drop
1794			 * this page though, as its main user (postcopy
1795			 * migration) will not expect userfaults on already
1796			 * copied pages.
1797			 */
1798			dec_mm_counter(mm, mm_counter(&folio->page));
 
 
 
1799		} else if (folio_test_anon(folio)) {
1800			swp_entry_t entry = page_swap_entry(subpage);
1801			pte_t swp_pte;
1802			/*
1803			 * Store the swap location in the pte.
1804			 * See handle_pte_fault() ...
1805			 */
1806			if (unlikely(folio_test_swapbacked(folio) !=
1807					folio_test_swapcache(folio))) {
1808				WARN_ON_ONCE(1);
1809				ret = false;
 
 
 
1810				page_vma_mapped_walk_done(&pvmw);
1811				break;
1812			}
1813
1814			/* MADV_FREE page check */
1815			if (!folio_test_swapbacked(folio)) {
1816				int ref_count, map_count;
1817
1818				/*
1819				 * Synchronize with gup_pte_range():
1820				 * - clear PTE; barrier; read refcount
1821				 * - inc refcount; barrier; read PTE
1822				 */
1823				smp_mb();
1824
1825				ref_count = folio_ref_count(folio);
1826				map_count = folio_mapcount(folio);
1827
1828				/*
1829				 * Order reads for page refcount and dirty flag
1830				 * (see comments in __remove_mapping()).
1831				 */
1832				smp_rmb();
1833
1834				/*
1835				 * The only page refs must be one from isolation
1836				 * plus the rmap(s) (dropped by discard:).
1837				 */
1838				if (ref_count == 1 + map_count &&
1839				    !folio_test_dirty(folio)) {
 
 
 
1840					dec_mm_counter(mm, MM_ANONPAGES);
1841					goto discard;
1842				}
1843
1844				/*
1845				 * If the folio was redirtied, it cannot be
1846				 * discarded. Remap the page to page table.
1847				 */
1848				set_pte_at(mm, address, pvmw.pte, pteval);
1849				folio_set_swapbacked(folio);
1850				ret = false;
1851				page_vma_mapped_walk_done(&pvmw);
1852				break;
1853			}
1854
1855			if (swap_duplicate(entry) < 0) {
1856				set_pte_at(mm, address, pvmw.pte, pteval);
1857				ret = false;
1858				page_vma_mapped_walk_done(&pvmw);
1859				break;
1860			}
1861			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1862				swap_free(entry);
1863				set_pte_at(mm, address, pvmw.pte, pteval);
1864				ret = false;
1865				page_vma_mapped_walk_done(&pvmw);
1866				break;
1867			}
1868
1869			/* See folio_try_share_anon_rmap(): clear PTE first. */
1870			if (anon_exclusive &&
1871			    folio_try_share_anon_rmap_pte(folio, subpage)) {
1872				swap_free(entry);
1873				set_pte_at(mm, address, pvmw.pte, pteval);
1874				ret = false;
1875				page_vma_mapped_walk_done(&pvmw);
1876				break;
1877			}
 
 
 
 
 
 
 
 
 
 
 
1878			if (list_empty(&mm->mmlist)) {
1879				spin_lock(&mmlist_lock);
1880				if (list_empty(&mm->mmlist))
1881					list_add(&mm->mmlist, &init_mm.mmlist);
1882				spin_unlock(&mmlist_lock);
1883			}
1884			dec_mm_counter(mm, MM_ANONPAGES);
1885			inc_mm_counter(mm, MM_SWAPENTS);
1886			swp_pte = swp_entry_to_pte(entry);
1887			if (anon_exclusive)
1888				swp_pte = pte_swp_mkexclusive(swp_pte);
1889			if (pte_soft_dirty(pteval))
1890				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1891			if (pte_uffd_wp(pteval))
1892				swp_pte = pte_swp_mkuffd_wp(swp_pte);
1893			set_pte_at(mm, address, pvmw.pte, swp_pte);
 
 
 
1894		} else {
1895			/*
1896			 * This is a locked file-backed folio,
1897			 * so it cannot be removed from the page
1898			 * cache and replaced by a new folio before
1899			 * mmu_notifier_invalidate_range_end, so no
1900			 * concurrent thread might update its page table
1901			 * to point at a new folio while a device is
1902			 * still using this folio.
1903			 *
1904			 * See Documentation/mm/mmu_notifier.rst
1905			 */
1906			dec_mm_counter(mm, mm_counter_file(&folio->page));
1907		}
1908discard:
1909		if (unlikely(folio_test_hugetlb(folio)))
1910			hugetlb_remove_rmap(folio);
1911		else
1912			folio_remove_rmap_pte(folio, subpage, vma);
 
 
 
 
1913		if (vma->vm_flags & VM_LOCKED)
1914			mlock_drain_local();
1915		folio_put(folio);
1916	}
1917
1918	mmu_notifier_invalidate_range_end(&range);
1919
1920	return ret;
1921}
1922
1923static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1924{
1925	return vma_is_temporary_stack(vma);
1926}
1927
1928static int folio_not_mapped(struct folio *folio)
1929{
1930	return !folio_mapped(folio);
1931}
1932
1933/**
1934 * try_to_unmap - Try to remove all page table mappings to a folio.
1935 * @folio: The folio to unmap.
1936 * @flags: action and flags
1937 *
1938 * Tries to remove all the page table entries which are mapping this
1939 * folio.  It is the caller's responsibility to check if the folio is
1940 * still mapped if needed (use TTU_SYNC to prevent accounting races).
1941 *
1942 * Context: Caller must hold the folio lock.
1943 */
1944void try_to_unmap(struct folio *folio, enum ttu_flags flags)
1945{
1946	struct rmap_walk_control rwc = {
1947		.rmap_one = try_to_unmap_one,
1948		.arg = (void *)flags,
1949		.done = folio_not_mapped,
1950		.anon_lock = folio_lock_anon_vma_read,
1951	};
1952
1953	if (flags & TTU_RMAP_LOCKED)
1954		rmap_walk_locked(folio, &rwc);
1955	else
1956		rmap_walk(folio, &rwc);
1957}
1958
1959/*
1960 * @arg: enum ttu_flags will be passed to this argument.
1961 *
1962 * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
1963 * containing migration entries.
1964 */
1965static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
1966		     unsigned long address, void *arg)
1967{
1968	struct mm_struct *mm = vma->vm_mm;
1969	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1970	pte_t pteval;
1971	struct page *subpage;
1972	bool anon_exclusive, ret = true;
1973	struct mmu_notifier_range range;
1974	enum ttu_flags flags = (enum ttu_flags)(long)arg;
1975	unsigned long pfn;
1976	unsigned long hsz = 0;
1977
1978	/*
1979	 * When racing against e.g. zap_pte_range() on another cpu,
1980	 * in between its ptep_get_and_clear_full() and folio_remove_rmap_*(),
1981	 * try_to_migrate() may return before page_mapped() has become false,
1982	 * if page table locking is skipped: use TTU_SYNC to wait for that.
1983	 */
1984	if (flags & TTU_SYNC)
1985		pvmw.flags = PVMW_SYNC;
1986
1987	/*
1988	 * unmap_page() in mm/huge_memory.c is the only user of migration with
1989	 * TTU_SPLIT_HUGE_PMD and it wants to freeze.
1990	 */
1991	if (flags & TTU_SPLIT_HUGE_PMD)
1992		split_huge_pmd_address(vma, address, true, folio);
1993
1994	/*
1995	 * For THP, we have to assume the worse case ie pmd for invalidation.
1996	 * For hugetlb, it could be much worse if we need to do pud
1997	 * invalidation in the case of pmd sharing.
1998	 *
1999	 * Note that the page can not be free in this function as call of
2000	 * try_to_unmap() must hold a reference on the page.
2001	 */
2002	range.end = vma_address_end(&pvmw);
2003	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2004				address, range.end);
2005	if (folio_test_hugetlb(folio)) {
2006		/*
2007		 * If sharing is possible, start and end will be adjusted
2008		 * accordingly.
2009		 */
2010		adjust_range_if_pmd_sharing_possible(vma, &range.start,
2011						     &range.end);
2012
2013		/* We need the huge page size for set_huge_pte_at() */
2014		hsz = huge_page_size(hstate_vma(vma));
2015	}
2016	mmu_notifier_invalidate_range_start(&range);
2017
2018	while (page_vma_mapped_walk(&pvmw)) {
2019#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2020		/* PMD-mapped THP migration entry */
2021		if (!pvmw.pte) {
2022			subpage = folio_page(folio,
2023				pmd_pfn(*pvmw.pmd) - folio_pfn(folio));
2024			VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
2025					!folio_test_pmd_mappable(folio), folio);
2026
2027			if (set_pmd_migration_entry(&pvmw, subpage)) {
2028				ret = false;
2029				page_vma_mapped_walk_done(&pvmw);
2030				break;
2031			}
2032			continue;
2033		}
2034#endif
2035
2036		/* Unexpected PMD-mapped THP? */
2037		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2038
2039		pfn = pte_pfn(ptep_get(pvmw.pte));
2040
2041		if (folio_is_zone_device(folio)) {
2042			/*
2043			 * Our PTE is a non-present device exclusive entry and
2044			 * calculating the subpage as for the common case would
2045			 * result in an invalid pointer.
2046			 *
2047			 * Since only PAGE_SIZE pages can currently be
2048			 * migrated, just set it to page. This will need to be
2049			 * changed when hugepage migrations to device private
2050			 * memory are supported.
2051			 */
2052			VM_BUG_ON_FOLIO(folio_nr_pages(folio) > 1, folio);
2053			subpage = &folio->page;
2054		} else {
2055			subpage = folio_page(folio, pfn - folio_pfn(folio));
 
2056		}
2057		address = pvmw.address;
2058		anon_exclusive = folio_test_anon(folio) &&
2059				 PageAnonExclusive(subpage);
2060
2061		if (folio_test_hugetlb(folio)) {
2062			bool anon = folio_test_anon(folio);
2063
2064			/*
2065			 * huge_pmd_unshare may unmap an entire PMD page.
2066			 * There is no way of knowing exactly which PMDs may
2067			 * be cached for this mm, so we must flush them all.
2068			 * start/end were already adjusted above to cover this
2069			 * range.
2070			 */
2071			flush_cache_range(vma, range.start, range.end);
2072
2073			/*
2074			 * To call huge_pmd_unshare, i_mmap_rwsem must be
2075			 * held in write mode.  Caller needs to explicitly
2076			 * do this outside rmap routines.
2077			 *
2078			 * We also must hold hugetlb vma_lock in write mode.
2079			 * Lock order dictates acquiring vma_lock BEFORE
2080			 * i_mmap_rwsem.  We can only try lock here and
2081			 * fail if unsuccessful.
2082			 */
2083			if (!anon) {
2084				VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
2085				if (!hugetlb_vma_trylock_write(vma)) {
2086					page_vma_mapped_walk_done(&pvmw);
2087					ret = false;
2088					break;
2089				}
2090				if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
2091					hugetlb_vma_unlock_write(vma);
2092					flush_tlb_range(vma,
2093						range.start, range.end);
 
 
2094
2095					/*
2096					 * The ref count of the PMD page was
2097					 * dropped which is part of the way map
2098					 * counting is done for shared PMDs.
2099					 * Return 'true' here.  When there is
2100					 * no other sharing, huge_pmd_unshare
2101					 * returns false and we will unmap the
2102					 * actual page and drop map count
2103					 * to zero.
2104					 */
2105					page_vma_mapped_walk_done(&pvmw);
2106					break;
2107				}
2108				hugetlb_vma_unlock_write(vma);
2109			}
2110			/* Nuke the hugetlb page table entry */
2111			pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
2112		} else {
2113			flush_cache_page(vma, address, pfn);
2114			/* Nuke the page table entry. */
2115			if (should_defer_flush(mm, flags)) {
2116				/*
2117				 * We clear the PTE but do not flush so potentially
2118				 * a remote CPU could still be writing to the folio.
2119				 * If the entry was previously clean then the
2120				 * architecture must guarantee that a clear->dirty
2121				 * transition on a cached TLB entry is written through
2122				 * and traps if the PTE is unmapped.
2123				 */
2124				pteval = ptep_get_and_clear(mm, address, pvmw.pte);
2125
2126				set_tlb_ubc_flush_pending(mm, pteval, address);
2127			} else {
2128				pteval = ptep_clear_flush(vma, address, pvmw.pte);
2129			}
2130		}
2131
2132		/* Set the dirty flag on the folio now the pte is gone. */
2133		if (pte_dirty(pteval))
2134			folio_mark_dirty(folio);
2135
2136		/* Update high watermark before we lower rss */
2137		update_hiwater_rss(mm);
2138
2139		if (folio_is_device_private(folio)) {
2140			unsigned long pfn = folio_pfn(folio);
2141			swp_entry_t entry;
2142			pte_t swp_pte;
2143
2144			if (anon_exclusive)
2145				WARN_ON_ONCE(folio_try_share_anon_rmap_pte(folio,
2146									   subpage));
2147
2148			/*
2149			 * Store the pfn of the page in a special migration
2150			 * pte. do_swap_page() will wait until the migration
2151			 * pte is removed and then restart fault handling.
2152			 */
2153			entry = pte_to_swp_entry(pteval);
2154			if (is_writable_device_private_entry(entry))
2155				entry = make_writable_migration_entry(pfn);
2156			else if (anon_exclusive)
2157				entry = make_readable_exclusive_migration_entry(pfn);
2158			else
2159				entry = make_readable_migration_entry(pfn);
2160			swp_pte = swp_entry_to_pte(entry);
2161
2162			/*
2163			 * pteval maps a zone device page and is therefore
2164			 * a swap pte.
2165			 */
2166			if (pte_swp_soft_dirty(pteval))
2167				swp_pte = pte_swp_mksoft_dirty(swp_pte);
2168			if (pte_swp_uffd_wp(pteval))
2169				swp_pte = pte_swp_mkuffd_wp(swp_pte);
2170			set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
2171			trace_set_migration_pte(pvmw.address, pte_val(swp_pte),
2172						compound_order(&folio->page));
2173			/*
2174			 * No need to invalidate here it will synchronize on
2175			 * against the special swap migration pte.
2176			 */
2177		} else if (PageHWPoison(subpage)) {
2178			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
2179			if (folio_test_hugetlb(folio)) {
2180				hugetlb_count_sub(folio_nr_pages(folio), mm);
2181				set_huge_pte_at(mm, address, pvmw.pte, pteval,
2182						hsz);
2183			} else {
2184				dec_mm_counter(mm, mm_counter(&folio->page));
2185				set_pte_at(mm, address, pvmw.pte, pteval);
2186			}
2187
2188		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
2189			/*
2190			 * The guest indicated that the page content is of no
2191			 * interest anymore. Simply discard the pte, vmscan
2192			 * will take care of the rest.
2193			 * A future reference will then fault in a new zero
2194			 * page. When userfaultfd is active, we must not drop
2195			 * this page though, as its main user (postcopy
2196			 * migration) will not expect userfaults on already
2197			 * copied pages.
2198			 */
2199			dec_mm_counter(mm, mm_counter(&folio->page));
 
 
 
2200		} else {
2201			swp_entry_t entry;
2202			pte_t swp_pte;
2203
2204			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
2205				if (folio_test_hugetlb(folio))
2206					set_huge_pte_at(mm, address, pvmw.pte,
2207							pteval, hsz);
2208				else
2209					set_pte_at(mm, address, pvmw.pte, pteval);
2210				ret = false;
2211				page_vma_mapped_walk_done(&pvmw);
2212				break;
2213			}
2214			VM_BUG_ON_PAGE(pte_write(pteval) && folio_test_anon(folio) &&
2215				       !anon_exclusive, subpage);
2216
2217			/* See folio_try_share_anon_rmap_pte(): clear PTE first. */
2218			if (folio_test_hugetlb(folio)) {
2219				if (anon_exclusive &&
2220				    hugetlb_try_share_anon_rmap(folio)) {
2221					set_huge_pte_at(mm, address, pvmw.pte,
2222							pteval, hsz);
2223					ret = false;
2224					page_vma_mapped_walk_done(&pvmw);
2225					break;
2226				}
2227			} else if (anon_exclusive &&
2228				   folio_try_share_anon_rmap_pte(folio, subpage)) {
2229				set_pte_at(mm, address, pvmw.pte, pteval);
2230				ret = false;
2231				page_vma_mapped_walk_done(&pvmw);
2232				break;
2233			}
2234
2235			/*
2236			 * Store the pfn of the page in a special migration
2237			 * pte. do_swap_page() will wait until the migration
2238			 * pte is removed and then restart fault handling.
2239			 */
2240			if (pte_write(pteval))
2241				entry = make_writable_migration_entry(
2242							page_to_pfn(subpage));
2243			else if (anon_exclusive)
2244				entry = make_readable_exclusive_migration_entry(
2245							page_to_pfn(subpage));
2246			else
2247				entry = make_readable_migration_entry(
2248							page_to_pfn(subpage));
2249			if (pte_young(pteval))
2250				entry = make_migration_entry_young(entry);
2251			if (pte_dirty(pteval))
2252				entry = make_migration_entry_dirty(entry);
2253			swp_pte = swp_entry_to_pte(entry);
2254			if (pte_soft_dirty(pteval))
2255				swp_pte = pte_swp_mksoft_dirty(swp_pte);
2256			if (pte_uffd_wp(pteval))
2257				swp_pte = pte_swp_mkuffd_wp(swp_pte);
2258			if (folio_test_hugetlb(folio))
2259				set_huge_pte_at(mm, address, pvmw.pte, swp_pte,
2260						hsz);
2261			else
2262				set_pte_at(mm, address, pvmw.pte, swp_pte);
2263			trace_set_migration_pte(address, pte_val(swp_pte),
2264						compound_order(&folio->page));
2265			/*
2266			 * No need to invalidate here it will synchronize on
2267			 * against the special swap migration pte.
2268			 */
2269		}
2270
2271		if (unlikely(folio_test_hugetlb(folio)))
2272			hugetlb_remove_rmap(folio);
2273		else
2274			folio_remove_rmap_pte(folio, subpage, vma);
 
 
 
 
2275		if (vma->vm_flags & VM_LOCKED)
2276			mlock_drain_local();
2277		folio_put(folio);
2278	}
2279
2280	mmu_notifier_invalidate_range_end(&range);
2281
2282	return ret;
2283}
2284
2285/**
2286 * try_to_migrate - try to replace all page table mappings with swap entries
2287 * @folio: the folio to replace page table entries for
2288 * @flags: action and flags
2289 *
2290 * Tries to remove all the page table entries which are mapping this folio and
2291 * replace them with special swap entries. Caller must hold the folio lock.
2292 */
2293void try_to_migrate(struct folio *folio, enum ttu_flags flags)
2294{
2295	struct rmap_walk_control rwc = {
2296		.rmap_one = try_to_migrate_one,
2297		.arg = (void *)flags,
2298		.done = folio_not_mapped,
2299		.anon_lock = folio_lock_anon_vma_read,
2300	};
2301
2302	/*
2303	 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
2304	 * TTU_SPLIT_HUGE_PMD, TTU_SYNC, and TTU_BATCH_FLUSH flags.
2305	 */
2306	if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2307					TTU_SYNC | TTU_BATCH_FLUSH)))
2308		return;
2309
2310	if (folio_is_zone_device(folio) &&
2311	    (!folio_is_device_private(folio) && !folio_is_device_coherent(folio)))
2312		return;
2313
2314	/*
2315	 * During exec, a temporary VMA is setup and later moved.
2316	 * The VMA is moved under the anon_vma lock but not the
2317	 * page tables leading to a race where migration cannot
2318	 * find the migration ptes. Rather than increasing the
2319	 * locking requirements of exec(), migration skips
2320	 * temporary VMAs until after exec() completes.
2321	 */
2322	if (!folio_test_ksm(folio) && folio_test_anon(folio))
2323		rwc.invalid_vma = invalid_migration_vma;
2324
2325	if (flags & TTU_RMAP_LOCKED)
2326		rmap_walk_locked(folio, &rwc);
2327	else
2328		rmap_walk(folio, &rwc);
2329}
2330
2331#ifdef CONFIG_DEVICE_PRIVATE
2332struct make_exclusive_args {
2333	struct mm_struct *mm;
2334	unsigned long address;
2335	void *owner;
2336	bool valid;
2337};
2338
2339static bool page_make_device_exclusive_one(struct folio *folio,
2340		struct vm_area_struct *vma, unsigned long address, void *priv)
2341{
2342	struct mm_struct *mm = vma->vm_mm;
2343	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
2344	struct make_exclusive_args *args = priv;
2345	pte_t pteval;
2346	struct page *subpage;
2347	bool ret = true;
2348	struct mmu_notifier_range range;
2349	swp_entry_t entry;
2350	pte_t swp_pte;
2351	pte_t ptent;
2352
2353	mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0,
2354				      vma->vm_mm, address, min(vma->vm_end,
2355				      address + folio_size(folio)),
2356				      args->owner);
2357	mmu_notifier_invalidate_range_start(&range);
2358
2359	while (page_vma_mapped_walk(&pvmw)) {
2360		/* Unexpected PMD-mapped THP? */
2361		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2362
2363		ptent = ptep_get(pvmw.pte);
2364		if (!pte_present(ptent)) {
2365			ret = false;
2366			page_vma_mapped_walk_done(&pvmw);
2367			break;
2368		}
2369
2370		subpage = folio_page(folio,
2371				pte_pfn(ptent) - folio_pfn(folio));
2372		address = pvmw.address;
2373
2374		/* Nuke the page table entry. */
2375		flush_cache_page(vma, address, pte_pfn(ptent));
2376		pteval = ptep_clear_flush(vma, address, pvmw.pte);
2377
2378		/* Set the dirty flag on the folio now the pte is gone. */
2379		if (pte_dirty(pteval))
2380			folio_mark_dirty(folio);
2381
2382		/*
2383		 * Check that our target page is still mapped at the expected
2384		 * address.
2385		 */
2386		if (args->mm == mm && args->address == address &&
2387		    pte_write(pteval))
2388			args->valid = true;
2389
2390		/*
2391		 * Store the pfn of the page in a special migration
2392		 * pte. do_swap_page() will wait until the migration
2393		 * pte is removed and then restart fault handling.
2394		 */
2395		if (pte_write(pteval))
2396			entry = make_writable_device_exclusive_entry(
2397							page_to_pfn(subpage));
2398		else
2399			entry = make_readable_device_exclusive_entry(
2400							page_to_pfn(subpage));
2401		swp_pte = swp_entry_to_pte(entry);
2402		if (pte_soft_dirty(pteval))
2403			swp_pte = pte_swp_mksoft_dirty(swp_pte);
2404		if (pte_uffd_wp(pteval))
2405			swp_pte = pte_swp_mkuffd_wp(swp_pte);
2406
2407		set_pte_at(mm, address, pvmw.pte, swp_pte);
2408
2409		/*
2410		 * There is a reference on the page for the swap entry which has
2411		 * been removed, so shouldn't take another.
2412		 */
2413		folio_remove_rmap_pte(folio, subpage, vma);
2414	}
2415
2416	mmu_notifier_invalidate_range_end(&range);
2417
2418	return ret;
2419}
2420
2421/**
2422 * folio_make_device_exclusive - Mark the folio exclusively owned by a device.
2423 * @folio: The folio to replace page table entries for.
2424 * @mm: The mm_struct where the folio is expected to be mapped.
2425 * @address: Address where the folio is expected to be mapped.
2426 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
2427 *
2428 * Tries to remove all the page table entries which are mapping this
2429 * folio and replace them with special device exclusive swap entries to
2430 * grant a device exclusive access to the folio.
2431 *
2432 * Context: Caller must hold the folio lock.
2433 * Return: false if the page is still mapped, or if it could not be unmapped
2434 * from the expected address. Otherwise returns true (success).
2435 */
2436static bool folio_make_device_exclusive(struct folio *folio,
2437		struct mm_struct *mm, unsigned long address, void *owner)
2438{
2439	struct make_exclusive_args args = {
2440		.mm = mm,
2441		.address = address,
2442		.owner = owner,
2443		.valid = false,
2444	};
2445	struct rmap_walk_control rwc = {
2446		.rmap_one = page_make_device_exclusive_one,
2447		.done = folio_not_mapped,
2448		.anon_lock = folio_lock_anon_vma_read,
2449		.arg = &args,
2450	};
2451
2452	/*
2453	 * Restrict to anonymous folios for now to avoid potential writeback
2454	 * issues.
2455	 */
2456	if (!folio_test_anon(folio))
2457		return false;
2458
2459	rmap_walk(folio, &rwc);
2460
2461	return args.valid && !folio_mapcount(folio);
2462}
2463
2464/**
2465 * make_device_exclusive_range() - Mark a range for exclusive use by a device
2466 * @mm: mm_struct of associated target process
2467 * @start: start of the region to mark for exclusive device access
2468 * @end: end address of region
2469 * @pages: returns the pages which were successfully marked for exclusive access
2470 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
2471 *
2472 * Returns: number of pages found in the range by GUP. A page is marked for
2473 * exclusive access only if the page pointer is non-NULL.
2474 *
2475 * This function finds ptes mapping page(s) to the given address range, locks
2476 * them and replaces mappings with special swap entries preventing userspace CPU
2477 * access. On fault these entries are replaced with the original mapping after
2478 * calling MMU notifiers.
2479 *
2480 * A driver using this to program access from a device must use a mmu notifier
2481 * critical section to hold a device specific lock during programming. Once
2482 * programming is complete it should drop the page lock and reference after
2483 * which point CPU access to the page will revoke the exclusive access.
2484 */
2485int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
2486				unsigned long end, struct page **pages,
2487				void *owner)
2488{
2489	long npages = (end - start) >> PAGE_SHIFT;
2490	long i;
2491
2492	npages = get_user_pages_remote(mm, start, npages,
2493				       FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
2494				       pages, NULL);
2495	if (npages < 0)
2496		return npages;
2497
2498	for (i = 0; i < npages; i++, start += PAGE_SIZE) {
2499		struct folio *folio = page_folio(pages[i]);
2500		if (PageTail(pages[i]) || !folio_trylock(folio)) {
2501			folio_put(folio);
2502			pages[i] = NULL;
2503			continue;
2504		}
2505
2506		if (!folio_make_device_exclusive(folio, mm, start, owner)) {
2507			folio_unlock(folio);
2508			folio_put(folio);
2509			pages[i] = NULL;
2510		}
2511	}
2512
2513	return npages;
2514}
2515EXPORT_SYMBOL_GPL(make_device_exclusive_range);
2516#endif
2517
2518void __put_anon_vma(struct anon_vma *anon_vma)
2519{
2520	struct anon_vma *root = anon_vma->root;
2521
2522	anon_vma_free(anon_vma);
2523	if (root != anon_vma && atomic_dec_and_test(&root->refcount))
2524		anon_vma_free(root);
2525}
2526
2527static struct anon_vma *rmap_walk_anon_lock(struct folio *folio,
2528					    struct rmap_walk_control *rwc)
2529{
2530	struct anon_vma *anon_vma;
2531
2532	if (rwc->anon_lock)
2533		return rwc->anon_lock(folio, rwc);
2534
2535	/*
2536	 * Note: remove_migration_ptes() cannot use folio_lock_anon_vma_read()
2537	 * because that depends on page_mapped(); but not all its usages
2538	 * are holding mmap_lock. Users without mmap_lock are required to
2539	 * take a reference count to prevent the anon_vma disappearing
2540	 */
2541	anon_vma = folio_anon_vma(folio);
2542	if (!anon_vma)
2543		return NULL;
2544
2545	if (anon_vma_trylock_read(anon_vma))
2546		goto out;
2547
2548	if (rwc->try_lock) {
2549		anon_vma = NULL;
2550		rwc->contended = true;
2551		goto out;
2552	}
2553
2554	anon_vma_lock_read(anon_vma);
2555out:
2556	return anon_vma;
2557}
2558
2559/*
2560 * rmap_walk_anon - do something to anonymous page using the object-based
2561 * rmap method
2562 * @folio: the folio to be handled
2563 * @rwc: control variable according to each walk type
2564 * @locked: caller holds relevant rmap lock
2565 *
2566 * Find all the mappings of a folio using the mapping pointer and the vma
2567 * chains contained in the anon_vma struct it points to.
2568 */
2569static void rmap_walk_anon(struct folio *folio,
2570		struct rmap_walk_control *rwc, bool locked)
2571{
2572	struct anon_vma *anon_vma;
2573	pgoff_t pgoff_start, pgoff_end;
2574	struct anon_vma_chain *avc;
2575
2576	if (locked) {
2577		anon_vma = folio_anon_vma(folio);
2578		/* anon_vma disappear under us? */
2579		VM_BUG_ON_FOLIO(!anon_vma, folio);
2580	} else {
2581		anon_vma = rmap_walk_anon_lock(folio, rwc);
2582	}
2583	if (!anon_vma)
2584		return;
2585
2586	pgoff_start = folio_pgoff(folio);
2587	pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2588	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
2589			pgoff_start, pgoff_end) {
2590		struct vm_area_struct *vma = avc->vma;
2591		unsigned long address = vma_address(&folio->page, vma);
2592
2593		VM_BUG_ON_VMA(address == -EFAULT, vma);
2594		cond_resched();
2595
2596		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2597			continue;
2598
2599		if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2600			break;
2601		if (rwc->done && rwc->done(folio))
2602			break;
2603	}
2604
2605	if (!locked)
2606		anon_vma_unlock_read(anon_vma);
2607}
2608
2609/*
2610 * rmap_walk_file - do something to file page using the object-based rmap method
2611 * @folio: the folio to be handled
2612 * @rwc: control variable according to each walk type
2613 * @locked: caller holds relevant rmap lock
2614 *
2615 * Find all the mappings of a folio using the mapping pointer and the vma chains
2616 * contained in the address_space struct it points to.
2617 */
2618static void rmap_walk_file(struct folio *folio,
2619		struct rmap_walk_control *rwc, bool locked)
2620{
2621	struct address_space *mapping = folio_mapping(folio);
2622	pgoff_t pgoff_start, pgoff_end;
2623	struct vm_area_struct *vma;
2624
2625	/*
2626	 * The page lock not only makes sure that page->mapping cannot
2627	 * suddenly be NULLified by truncation, it makes sure that the
2628	 * structure at mapping cannot be freed and reused yet,
2629	 * so we can safely take mapping->i_mmap_rwsem.
2630	 */
2631	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2632
2633	if (!mapping)
2634		return;
2635
2636	pgoff_start = folio_pgoff(folio);
2637	pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2638	if (!locked) {
2639		if (i_mmap_trylock_read(mapping))
2640			goto lookup;
2641
2642		if (rwc->try_lock) {
2643			rwc->contended = true;
2644			return;
2645		}
2646
2647		i_mmap_lock_read(mapping);
2648	}
2649lookup:
2650	vma_interval_tree_foreach(vma, &mapping->i_mmap,
2651			pgoff_start, pgoff_end) {
2652		unsigned long address = vma_address(&folio->page, vma);
2653
2654		VM_BUG_ON_VMA(address == -EFAULT, vma);
2655		cond_resched();
2656
2657		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2658			continue;
2659
2660		if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2661			goto done;
2662		if (rwc->done && rwc->done(folio))
2663			goto done;
2664	}
2665
2666done:
2667	if (!locked)
2668		i_mmap_unlock_read(mapping);
2669}
2670
2671void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc)
2672{
2673	if (unlikely(folio_test_ksm(folio)))
2674		rmap_walk_ksm(folio, rwc);
2675	else if (folio_test_anon(folio))
2676		rmap_walk_anon(folio, rwc, false);
2677	else
2678		rmap_walk_file(folio, rwc, false);
2679}
2680
2681/* Like rmap_walk, but caller holds relevant rmap lock */
2682void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc)
2683{
2684	/* no ksm support for now */
2685	VM_BUG_ON_FOLIO(folio_test_ksm(folio), folio);
2686	if (folio_test_anon(folio))
2687		rmap_walk_anon(folio, rwc, true);
2688	else
2689		rmap_walk_file(folio, rwc, true);
2690}
2691
2692#ifdef CONFIG_HUGETLB_PAGE
2693/*
2694 * The following two functions are for anonymous (private mapped) hugepages.
2695 * Unlike common anonymous pages, anonymous hugepages have no accounting code
2696 * and no lru code, because we handle hugepages differently from common pages.
 
 
2697 */
2698void hugetlb_add_anon_rmap(struct folio *folio, struct vm_area_struct *vma,
2699		unsigned long address, rmap_t flags)
2700{
2701	VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio);
2702	VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
2703
2704	atomic_inc(&folio->_entire_mapcount);
2705	if (flags & RMAP_EXCLUSIVE)
2706		SetPageAnonExclusive(&folio->page);
2707	VM_WARN_ON_FOLIO(folio_entire_mapcount(folio) > 1 &&
2708			 PageAnonExclusive(&folio->page), folio);
 
 
 
 
2709}
2710
2711void hugetlb_add_new_anon_rmap(struct folio *folio,
2712		struct vm_area_struct *vma, unsigned long address)
2713{
2714	VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio);
2715
2716	BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2717	/* increment count (starts at -1) */
2718	atomic_set(&folio->_entire_mapcount, 0);
2719	folio_clear_hugetlb_restore_reserve(folio);
2720	__folio_set_anon(folio, vma, address, true);
2721	SetPageAnonExclusive(&folio->page);
2722}
2723#endif /* CONFIG_HUGETLB_PAGE */