Linux Audio

Check our new training course

Open-source upstreaming

Need help get the support for your hardware in upstream Linux?
Loading...
   1/*
   2 * mm/rmap.c - physical to virtual reverse mappings
   3 *
   4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
   5 * Released under the General Public License (GPL).
   6 *
   7 * Simple, low overhead reverse mapping scheme.
   8 * Please try to keep this thing as modular as possible.
   9 *
  10 * Provides methods for unmapping each kind of mapped page:
  11 * the anon methods track anonymous pages, and
  12 * the file methods track pages belonging to an inode.
  13 *
  14 * Original design by Rik van Riel <riel@conectiva.com.br> 2001
  15 * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
  16 * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
  17 * Contributions by Hugh Dickins 2003, 2004
  18 */
  19
  20/*
  21 * Lock ordering in mm:
  22 *
  23 * inode->i_mutex	(while writing or truncating, not reading or faulting)
  24 *   mm->mmap_sem
  25 *     page->flags PG_locked (lock_page)
  26 *       mapping->i_mmap_mutex
  27 *         anon_vma->rwsem
  28 *           mm->page_table_lock or pte_lock
  29 *             zone->lru_lock (in mark_page_accessed, isolate_lru_page)
  30 *             swap_lock (in swap_duplicate, swap_info_get)
  31 *               mmlist_lock (in mmput, drain_mmlist and others)
  32 *               mapping->private_lock (in __set_page_dirty_buffers)
  33 *               inode->i_lock (in set_page_dirty's __mark_inode_dirty)
  34 *               bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
  35 *                 sb_lock (within inode_lock in fs/fs-writeback.c)
  36 *                 mapping->tree_lock (widely used, in set_page_dirty,
  37 *                           in arch-dependent flush_dcache_mmap_lock,
  38 *                           within bdi.wb->list_lock in __sync_single_inode)
 
 
 
  39 *
  40 * anon_vma->rwsem,mapping->i_mutex      (memory_failure, collect_procs_anon)
  41 *   ->tasklist_lock
  42 *     pte map lock
  43 */
  44
  45#include <linux/mm.h>
 
 
  46#include <linux/pagemap.h>
  47#include <linux/swap.h>
  48#include <linux/swapops.h>
  49#include <linux/slab.h>
  50#include <linux/init.h>
  51#include <linux/ksm.h>
  52#include <linux/rmap.h>
  53#include <linux/rcupdate.h>
  54#include <linux/export.h>
  55#include <linux/memcontrol.h>
  56#include <linux/mmu_notifier.h>
  57#include <linux/migrate.h>
  58#include <linux/hugetlb.h>
 
  59#include <linux/backing-dev.h>
 
 
 
  60
  61#include <asm/tlbflush.h>
  62
 
 
  63#include "internal.h"
  64
  65static struct kmem_cache *anon_vma_cachep;
  66static struct kmem_cache *anon_vma_chain_cachep;
  67
  68static inline struct anon_vma *anon_vma_alloc(void)
  69{
  70	struct anon_vma *anon_vma;
  71
  72	anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
  73	if (anon_vma) {
  74		atomic_set(&anon_vma->refcount, 1);
 
 
  75		/*
  76		 * Initialise the anon_vma root to point to itself. If called
  77		 * from fork, the root will be reset to the parents anon_vma.
  78		 */
  79		anon_vma->root = anon_vma;
  80	}
  81
  82	return anon_vma;
  83}
  84
  85static inline void anon_vma_free(struct anon_vma *anon_vma)
  86{
  87	VM_BUG_ON(atomic_read(&anon_vma->refcount));
  88
  89	/*
  90	 * Synchronize against page_lock_anon_vma_read() such that
  91	 * we can safely hold the lock without the anon_vma getting
  92	 * freed.
  93	 *
  94	 * Relies on the full mb implied by the atomic_dec_and_test() from
  95	 * put_anon_vma() against the acquire barrier implied by
  96	 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
  97	 *
  98	 * page_lock_anon_vma_read()	VS	put_anon_vma()
  99	 *   down_read_trylock()		  atomic_dec_and_test()
 100	 *   LOCK				  MB
 101	 *   atomic_read()			  rwsem_is_locked()
 102	 *
 103	 * LOCK should suffice since the actual taking of the lock must
 104	 * happen _before_ what follows.
 105	 */
 
 106	if (rwsem_is_locked(&anon_vma->root->rwsem)) {
 107		anon_vma_lock_write(anon_vma);
 108		anon_vma_unlock_write(anon_vma);
 109	}
 110
 111	kmem_cache_free(anon_vma_cachep, anon_vma);
 112}
 113
 114static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
 115{
 116	return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
 117}
 118
 119static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
 120{
 121	kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
 122}
 123
 124static void anon_vma_chain_link(struct vm_area_struct *vma,
 125				struct anon_vma_chain *avc,
 126				struct anon_vma *anon_vma)
 127{
 128	avc->vma = vma;
 129	avc->anon_vma = anon_vma;
 130	list_add(&avc->same_vma, &vma->anon_vma_chain);
 131	anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
 132}
 133
 134/**
 135 * anon_vma_prepare - attach an anon_vma to a memory region
 136 * @vma: the memory region in question
 137 *
 138 * This makes sure the memory mapping described by 'vma' has
 139 * an 'anon_vma' attached to it, so that we can associate the
 140 * anonymous pages mapped into it with that anon_vma.
 141 *
 142 * The common case will be that we already have one, but if
 
 143 * not we either need to find an adjacent mapping that we
 144 * can re-use the anon_vma from (very common when the only
 145 * reason for splitting a vma has been mprotect()), or we
 146 * allocate a new one.
 147 *
 148 * Anon-vma allocations are very subtle, because we may have
 149 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
 150 * and that may actually touch the spinlock even in the newly
 151 * allocated vma (it depends on RCU to make sure that the
 152 * anon_vma isn't actually destroyed).
 153 *
 154 * As a result, we need to do proper anon_vma locking even
 155 * for the new allocation. At the same time, we do not want
 156 * to do any locking for the common case of already having
 157 * an anon_vma.
 158 *
 159 * This must be called with the mmap_sem held for reading.
 160 */
 161int anon_vma_prepare(struct vm_area_struct *vma)
 162{
 163	struct anon_vma *anon_vma = vma->anon_vma;
 
 164	struct anon_vma_chain *avc;
 165
 166	might_sleep();
 167	if (unlikely(!anon_vma)) {
 168		struct mm_struct *mm = vma->vm_mm;
 169		struct anon_vma *allocated;
 170
 171		avc = anon_vma_chain_alloc(GFP_KERNEL);
 172		if (!avc)
 173			goto out_enomem;
 174
 175		anon_vma = find_mergeable_anon_vma(vma);
 176		allocated = NULL;
 177		if (!anon_vma) {
 178			anon_vma = anon_vma_alloc();
 179			if (unlikely(!anon_vma))
 180				goto out_enomem_free_avc;
 181			allocated = anon_vma;
 182		}
 183
 184		anon_vma_lock_write(anon_vma);
 185		/* page_table_lock to protect against threads */
 186		spin_lock(&mm->page_table_lock);
 187		if (likely(!vma->anon_vma)) {
 188			vma->anon_vma = anon_vma;
 189			anon_vma_chain_link(vma, avc, anon_vma);
 190			allocated = NULL;
 191			avc = NULL;
 192		}
 193		spin_unlock(&mm->page_table_lock);
 194		anon_vma_unlock_write(anon_vma);
 195
 196		if (unlikely(allocated))
 197			put_anon_vma(allocated);
 198		if (unlikely(avc))
 199			anon_vma_chain_free(avc);
 
 
 
 
 
 
 200	}
 
 
 
 
 
 
 
 
 201	return 0;
 202
 203 out_enomem_free_avc:
 204	anon_vma_chain_free(avc);
 205 out_enomem:
 206	return -ENOMEM;
 207}
 208
 209/*
 210 * This is a useful helper function for locking the anon_vma root as
 211 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
 212 * have the same vma.
 213 *
 214 * Such anon_vma's should have the same root, so you'd expect to see
 215 * just a single mutex_lock for the whole traversal.
 216 */
 217static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
 218{
 219	struct anon_vma *new_root = anon_vma->root;
 220	if (new_root != root) {
 221		if (WARN_ON_ONCE(root))
 222			up_write(&root->rwsem);
 223		root = new_root;
 224		down_write(&root->rwsem);
 225	}
 226	return root;
 227}
 228
 229static inline void unlock_anon_vma_root(struct anon_vma *root)
 230{
 231	if (root)
 232		up_write(&root->rwsem);
 233}
 234
 235/*
 236 * Attach the anon_vmas from src to dst.
 237 * Returns 0 on success, -ENOMEM on failure.
 
 
 
 
 
 
 
 
 238 */
 239int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
 240{
 241	struct anon_vma_chain *avc, *pavc;
 242	struct anon_vma *root = NULL;
 243
 244	list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
 245		struct anon_vma *anon_vma;
 246
 247		avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
 248		if (unlikely(!avc)) {
 249			unlock_anon_vma_root(root);
 250			root = NULL;
 251			avc = anon_vma_chain_alloc(GFP_KERNEL);
 252			if (!avc)
 253				goto enomem_failure;
 254		}
 255		anon_vma = pavc->anon_vma;
 256		root = lock_anon_vma_root(root, anon_vma);
 257		anon_vma_chain_link(dst, avc, anon_vma);
 
 
 
 
 
 
 
 
 
 
 
 
 258	}
 
 
 259	unlock_anon_vma_root(root);
 260	return 0;
 261
 262 enomem_failure:
 
 
 
 
 
 
 
 263	unlink_anon_vmas(dst);
 264	return -ENOMEM;
 265}
 266
 267/*
 268 * Attach vma to its own anon_vma, as well as to the anon_vmas that
 269 * the corresponding VMA in the parent process is attached to.
 270 * Returns 0 on success, non-zero on failure.
 271 */
 272int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
 273{
 274	struct anon_vma_chain *avc;
 275	struct anon_vma *anon_vma;
 
 276
 277	/* Don't bother if the parent process has no anon_vma here. */
 278	if (!pvma->anon_vma)
 279		return 0;
 280
 
 
 
 281	/*
 282	 * First, attach the new VMA to the parent VMA's anon_vmas,
 283	 * so rmap can find non-COWed pages in child processes.
 284	 */
 285	if (anon_vma_clone(vma, pvma))
 286		return -ENOMEM;
 
 
 
 
 
 287
 288	/* Then add our own anon_vma. */
 289	anon_vma = anon_vma_alloc();
 290	if (!anon_vma)
 291		goto out_error;
 292	avc = anon_vma_chain_alloc(GFP_KERNEL);
 293	if (!avc)
 294		goto out_error_free_anon_vma;
 295
 296	/*
 297	 * The root anon_vma's spinlock is the lock actually used when we
 298	 * lock any of the anon_vmas in this anon_vma tree.
 299	 */
 300	anon_vma->root = pvma->anon_vma->root;
 
 301	/*
 302	 * With refcounts, an anon_vma can stay around longer than the
 303	 * process it belongs to. The root anon_vma needs to be pinned until
 304	 * this anon_vma is freed, because the lock lives in the root.
 305	 */
 306	get_anon_vma(anon_vma->root);
 307	/* Mark this anon_vma as the one where our new (COWed) pages go. */
 308	vma->anon_vma = anon_vma;
 309	anon_vma_lock_write(anon_vma);
 310	anon_vma_chain_link(vma, avc, anon_vma);
 
 311	anon_vma_unlock_write(anon_vma);
 312
 313	return 0;
 314
 315 out_error_free_anon_vma:
 316	put_anon_vma(anon_vma);
 317 out_error:
 318	unlink_anon_vmas(vma);
 319	return -ENOMEM;
 320}
 321
 322void unlink_anon_vmas(struct vm_area_struct *vma)
 323{
 324	struct anon_vma_chain *avc, *next;
 325	struct anon_vma *root = NULL;
 326
 327	/*
 328	 * Unlink each anon_vma chained to the VMA.  This list is ordered
 329	 * from newest to oldest, ensuring the root anon_vma gets freed last.
 330	 */
 331	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 332		struct anon_vma *anon_vma = avc->anon_vma;
 333
 334		root = lock_anon_vma_root(root, anon_vma);
 335		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
 336
 337		/*
 338		 * Leave empty anon_vmas on the list - we'll need
 339		 * to free them outside the lock.
 340		 */
 341		if (RB_EMPTY_ROOT(&anon_vma->rb_root))
 
 342			continue;
 
 343
 344		list_del(&avc->same_vma);
 345		anon_vma_chain_free(avc);
 346	}
 
 
 347	unlock_anon_vma_root(root);
 348
 349	/*
 350	 * Iterate the list once more, it now only contains empty and unlinked
 351	 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
 352	 * needing to write-acquire the anon_vma->root->rwsem.
 353	 */
 354	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 355		struct anon_vma *anon_vma = avc->anon_vma;
 356
 
 357		put_anon_vma(anon_vma);
 358
 359		list_del(&avc->same_vma);
 360		anon_vma_chain_free(avc);
 361	}
 362}
 363
 364static void anon_vma_ctor(void *data)
 365{
 366	struct anon_vma *anon_vma = data;
 367
 368	init_rwsem(&anon_vma->rwsem);
 369	atomic_set(&anon_vma->refcount, 0);
 370	anon_vma->rb_root = RB_ROOT;
 371}
 372
 373void __init anon_vma_init(void)
 374{
 375	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
 376			0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
 377	anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
 
 
 378}
 379
 380/*
 381 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
 382 *
 383 * Since there is no serialization what so ever against page_remove_rmap()
 384 * the best this function can do is return a locked anon_vma that might
 385 * have been relevant to this page.
 386 *
 387 * The page might have been remapped to a different anon_vma or the anon_vma
 388 * returned may already be freed (and even reused).
 389 *
 390 * In case it was remapped to a different anon_vma, the new anon_vma will be a
 391 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
 392 * ensure that any anon_vma obtained from the page will still be valid for as
 393 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
 394 *
 395 * All users of this function must be very careful when walking the anon_vma
 396 * chain and verify that the page in question is indeed mapped in it
 397 * [ something equivalent to page_mapped_in_vma() ].
 398 *
 399 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
 400 * that the anon_vma pointer from page->mapping is valid if there is a
 401 * mapcount, we can dereference the anon_vma after observing those.
 402 */
 403struct anon_vma *page_get_anon_vma(struct page *page)
 404{
 405	struct anon_vma *anon_vma = NULL;
 406	unsigned long anon_mapping;
 407
 408	rcu_read_lock();
 409	anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
 410	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 411		goto out;
 412	if (!page_mapped(page))
 413		goto out;
 414
 415	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 416	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 417		anon_vma = NULL;
 418		goto out;
 419	}
 420
 421	/*
 422	 * If this page is still mapped, then its anon_vma cannot have been
 423	 * freed.  But if it has been unmapped, we have no security against the
 424	 * anon_vma structure being freed and reused (for another anon_vma:
 425	 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
 426	 * above cannot corrupt).
 427	 */
 428	if (!page_mapped(page)) {
 
 429		put_anon_vma(anon_vma);
 430		anon_vma = NULL;
 431	}
 432out:
 433	rcu_read_unlock();
 434
 435	return anon_vma;
 436}
 437
 438/*
 439 * Similar to page_get_anon_vma() except it locks the anon_vma.
 440 *
 441 * Its a little more complex as it tries to keep the fast path to a single
 442 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
 443 * reference like with page_get_anon_vma() and then block on the mutex.
 444 */
 445struct anon_vma *page_lock_anon_vma_read(struct page *page)
 446{
 447	struct anon_vma *anon_vma = NULL;
 448	struct anon_vma *root_anon_vma;
 449	unsigned long anon_mapping;
 450
 451	rcu_read_lock();
 452	anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
 453	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 454		goto out;
 455	if (!page_mapped(page))
 456		goto out;
 457
 458	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 459	root_anon_vma = ACCESS_ONCE(anon_vma->root);
 460	if (down_read_trylock(&root_anon_vma->rwsem)) {
 461		/*
 462		 * If the page is still mapped, then this anon_vma is still
 463		 * its anon_vma, and holding the mutex ensures that it will
 464		 * not go away, see anon_vma_free().
 465		 */
 466		if (!page_mapped(page)) {
 467			up_read(&root_anon_vma->rwsem);
 468			anon_vma = NULL;
 469		}
 470		goto out;
 471	}
 472
 473	/* trylock failed, we got to sleep */
 474	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 475		anon_vma = NULL;
 476		goto out;
 477	}
 478
 479	if (!page_mapped(page)) {
 
 480		put_anon_vma(anon_vma);
 481		anon_vma = NULL;
 482		goto out;
 483	}
 484
 485	/* we pinned the anon_vma, its safe to sleep */
 486	rcu_read_unlock();
 487	anon_vma_lock_read(anon_vma);
 488
 489	if (atomic_dec_and_test(&anon_vma->refcount)) {
 490		/*
 491		 * Oops, we held the last refcount, release the lock
 492		 * and bail -- can't simply use put_anon_vma() because
 493		 * we'll deadlock on the anon_vma_lock_write() recursion.
 494		 */
 495		anon_vma_unlock_read(anon_vma);
 496		__put_anon_vma(anon_vma);
 497		anon_vma = NULL;
 498	}
 499
 500	return anon_vma;
 501
 502out:
 503	rcu_read_unlock();
 504	return anon_vma;
 505}
 506
 507void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
 508{
 509	anon_vma_unlock_read(anon_vma);
 510}
 511
 
 512/*
 513 * At what user virtual address is page expected in @vma?
 
 
 
 514 */
 515static inline unsigned long
 516__vma_address(struct page *page, struct vm_area_struct *vma)
 517{
 518	pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 519
 520	if (unlikely(is_vm_hugetlb_page(vma)))
 521		pgoff = page->index << huge_page_order(page_hstate(page));
 522
 523	return vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
 
 
 524}
 525
 526inline unsigned long
 527vma_address(struct page *page, struct vm_area_struct *vma)
 528{
 529	unsigned long address = __vma_address(page, vma);
 530
 531	/* page should be within @vma mapping range */
 532	VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
 
 533
 534	return address;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 535}
 536
 537/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 538 * At what user virtual address is page expected in vma?
 539 * Caller should check the page is actually part of the vma.
 540 */
 541unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
 542{
 543	unsigned long address;
 544	if (PageAnon(page)) {
 545		struct anon_vma *page__anon_vma = page_anon_vma(page);
 546		/*
 547		 * Note: swapoff's unuse_vma() is more efficient with this
 548		 * check, and needs it to match anon_vma when KSM is active.
 549		 */
 550		if (!vma->anon_vma || !page__anon_vma ||
 551		    vma->anon_vma->root != page__anon_vma->root)
 552			return -EFAULT;
 553	} else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
 554		if (!vma->vm_file ||
 555		    vma->vm_file->f_mapping != page->mapping)
 556			return -EFAULT;
 557	} else
 558		return -EFAULT;
 559	address = __vma_address(page, vma);
 560	if (unlikely(address < vma->vm_start || address >= vma->vm_end))
 561		return -EFAULT;
 562	return address;
 563}
 564
 565pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
 566{
 567	pgd_t *pgd;
 
 568	pud_t *pud;
 569	pmd_t *pmd = NULL;
 
 570
 571	pgd = pgd_offset(mm, address);
 572	if (!pgd_present(*pgd))
 573		goto out;
 574
 575	pud = pud_offset(pgd, address);
 
 
 
 
 576	if (!pud_present(*pud))
 577		goto out;
 578
 579	pmd = pmd_offset(pud, address);
 580	if (!pmd_present(*pmd))
 
 
 
 
 
 
 
 581		pmd = NULL;
 582out:
 583	return pmd;
 584}
 585
 586/*
 587 * Check that @page is mapped at @address into @mm.
 588 *
 589 * If @sync is false, page_check_address may perform a racy check to avoid
 590 * the page table lock when the pte is not present (helpful when reclaiming
 591 * highly shared pages).
 592 *
 593 * On success returns with pte mapped and locked.
 594 */
 595pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
 596			  unsigned long address, spinlock_t **ptlp, int sync)
 597{
 598	pmd_t *pmd;
 599	pte_t *pte;
 600	spinlock_t *ptl;
 601
 602	if (unlikely(PageHuge(page))) {
 603		/* when pud is not present, pte will be NULL */
 604		pte = huge_pte_offset(mm, address);
 605		if (!pte)
 606			return NULL;
 607
 608		ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
 609		goto check;
 610	}
 611
 612	pmd = mm_find_pmd(mm, address);
 613	if (!pmd)
 614		return NULL;
 615
 616	if (pmd_trans_huge(*pmd))
 617		return NULL;
 618
 619	pte = pte_offset_map(pmd, address);
 620	/* Make a quick check before getting the lock */
 621	if (!sync && !pte_present(*pte)) {
 622		pte_unmap(pte);
 623		return NULL;
 624	}
 625
 626	ptl = pte_lockptr(mm, pmd);
 627check:
 628	spin_lock(ptl);
 629	if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
 630		*ptlp = ptl;
 631		return pte;
 632	}
 633	pte_unmap_unlock(pte, ptl);
 634	return NULL;
 635}
 636
 637/**
 638 * page_mapped_in_vma - check whether a page is really mapped in a VMA
 639 * @page: the page to test
 640 * @vma: the VMA to test
 641 *
 642 * Returns 1 if the page is mapped into the page tables of the VMA, 0
 643 * if the page is not mapped into the page tables of this VMA.  Only
 644 * valid for normal file or anonymous VMAs.
 645 */
 646int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
 647{
 648	unsigned long address;
 649	pte_t *pte;
 650	spinlock_t *ptl;
 651
 652	address = __vma_address(page, vma);
 653	if (unlikely(address < vma->vm_start || address >= vma->vm_end))
 654		return 0;
 655	pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
 656	if (!pte)			/* the page is not in this mm */
 657		return 0;
 658	pte_unmap_unlock(pte, ptl);
 659
 660	return 1;
 661}
 662
 663struct page_referenced_arg {
 664	int mapcount;
 665	int referenced;
 666	unsigned long vm_flags;
 667	struct mem_cgroup *memcg;
 668};
 669/*
 670 * arg: page_referenced_arg will be passed
 671 */
 672int page_referenced_one(struct page *page, struct vm_area_struct *vma,
 673			unsigned long address, void *arg)
 674{
 675	struct mm_struct *mm = vma->vm_mm;
 676	spinlock_t *ptl;
 677	int referenced = 0;
 678	struct page_referenced_arg *pra = arg;
 
 
 
 
 
 
 679
 680	if (unlikely(PageTransHuge(page))) {
 681		pmd_t *pmd;
 682
 683		/*
 684		 * rmap might return false positives; we must filter
 685		 * these out using page_check_address_pmd().
 686		 */
 687		pmd = page_check_address_pmd(page, mm, address,
 688					     PAGE_CHECK_ADDRESS_PMD_FLAG, &ptl);
 689		if (!pmd)
 690			return SWAP_AGAIN;
 691
 692		if (vma->vm_flags & VM_LOCKED) {
 693			spin_unlock(ptl);
 694			pra->vm_flags |= VM_LOCKED;
 695			return SWAP_FAIL; /* To break the loop */
 696		}
 697
 698		/* go ahead even if the pmd is pmd_trans_splitting() */
 699		if (pmdp_clear_flush_young_notify(vma, address, pmd))
 700			referenced++;
 701		spin_unlock(ptl);
 702	} else {
 703		pte_t *pte;
 704
 705		/*
 706		 * rmap might return false positives; we must filter
 707		 * these out using page_check_address().
 708		 */
 709		pte = page_check_address(page, mm, address, &ptl, 0);
 710		if (!pte)
 711			return SWAP_AGAIN;
 712
 713		if (vma->vm_flags & VM_LOCKED) {
 714			pte_unmap_unlock(pte, ptl);
 715			pra->vm_flags |= VM_LOCKED;
 716			return SWAP_FAIL; /* To break the loop */
 717		}
 718
 719		if (ptep_clear_flush_young_notify(vma, address, pte)) {
 720			/*
 721			 * Don't treat a reference through a sequentially read
 722			 * mapping as such.  If the page has been used in
 723			 * another mapping, we will catch it; if this other
 724			 * mapping is already gone, the unmap path will have
 725			 * set PG_referenced or activated the page.
 726			 */
 727			if (likely(!(vma->vm_flags & VM_SEQ_READ)))
 
 
 
 
 
 
 
 
 728				referenced++;
 
 
 
 729		}
 730		pte_unmap_unlock(pte, ptl);
 
 731	}
 732
 
 
 
 
 
 733	if (referenced) {
 734		pra->referenced++;
 735		pra->vm_flags |= vma->vm_flags;
 736	}
 737
 738	pra->mapcount--;
 739	if (!pra->mapcount)
 740		return SWAP_SUCCESS; /* To break the loop */
 741
 742	return SWAP_AGAIN;
 743}
 744
 745static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
 746{
 747	struct page_referenced_arg *pra = arg;
 748	struct mem_cgroup *memcg = pra->memcg;
 749
 750	if (!mm_match_cgroup(vma->vm_mm, memcg))
 751		return true;
 752
 753	return false;
 754}
 755
 756/**
 757 * page_referenced - test if the page was referenced
 758 * @page: the page to test
 759 * @is_locked: caller holds lock on the page
 760 * @memcg: target memory cgroup
 761 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
 762 *
 763 * Quick test_and_clear_referenced for all mappings to a page,
 764 * returns the number of ptes which referenced the page.
 765 */
 766int page_referenced(struct page *page,
 767		    int is_locked,
 768		    struct mem_cgroup *memcg,
 769		    unsigned long *vm_flags)
 770{
 771	int ret;
 772	int we_locked = 0;
 773	struct page_referenced_arg pra = {
 774		.mapcount = page_mapcount(page),
 775		.memcg = memcg,
 776	};
 777	struct rmap_walk_control rwc = {
 778		.rmap_one = page_referenced_one,
 779		.arg = (void *)&pra,
 780		.anon_lock = page_lock_anon_vma_read,
 781	};
 782
 783	*vm_flags = 0;
 784	if (!page_mapped(page))
 785		return 0;
 786
 787	if (!page_rmapping(page))
 788		return 0;
 789
 790	if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
 791		we_locked = trylock_page(page);
 792		if (!we_locked)
 793			return 1;
 794	}
 795
 796	/*
 797	 * If we are reclaiming on behalf of a cgroup, skip
 798	 * counting on behalf of references from different
 799	 * cgroups
 800	 */
 801	if (memcg) {
 802		rwc.invalid_vma = invalid_page_referenced_vma;
 803	}
 804
 805	ret = rmap_walk(page, &rwc);
 806	*vm_flags = pra.vm_flags;
 807
 808	if (we_locked)
 809		unlock_page(page);
 810
 811	return pra.referenced;
 812}
 813
 814static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
 815			    unsigned long address, void *arg)
 816{
 817	struct mm_struct *mm = vma->vm_mm;
 818	pte_t *pte;
 819	spinlock_t *ptl;
 820	int ret = 0;
 
 
 
 821	int *cleaned = arg;
 822
 823	pte = page_check_address(page, mm, address, &ptl, 1);
 824	if (!pte)
 825		goto out;
 826
 827	if (pte_dirty(*pte) || pte_write(*pte)) {
 828		pte_t entry;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 829
 830		flush_cache_page(vma, address, pte_pfn(*pte));
 831		entry = ptep_clear_flush(vma, address, pte);
 832		entry = pte_wrprotect(entry);
 833		entry = pte_mkclean(entry);
 834		set_pte_at(mm, address, pte, entry);
 835		ret = 1;
 836	}
 837
 838	pte_unmap_unlock(pte, ptl);
 839
 840	if (ret) {
 841		mmu_notifier_invalidate_page(mm, address);
 842		(*cleaned)++;
 843	}
 844out:
 845	return SWAP_AGAIN;
 
 
 846}
 847
 848static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
 849{
 850	if (vma->vm_flags & VM_SHARED)
 851		return false;
 852
 853	return true;
 854}
 855
 856int page_mkclean(struct page *page)
 857{
 858	int cleaned = 0;
 859	struct address_space *mapping;
 860	struct rmap_walk_control rwc = {
 861		.arg = (void *)&cleaned,
 862		.rmap_one = page_mkclean_one,
 863		.invalid_vma = invalid_mkclean_vma,
 864	};
 865
 866	BUG_ON(!PageLocked(page));
 867
 868	if (!page_mapped(page))
 869		return 0;
 870
 871	mapping = page_mapping(page);
 872	if (!mapping)
 873		return 0;
 874
 875	rmap_walk(page, &rwc);
 876
 877	return cleaned;
 878}
 879EXPORT_SYMBOL_GPL(page_mkclean);
 880
 881/**
 882 * page_move_anon_rmap - move a page to our anon_vma
 883 * @page:	the page to move to our anon_vma
 884 * @vma:	the vma the page belongs to
 885 * @address:	the user virtual address mapped
 886 *
 887 * When a page belongs exclusively to one process after a COW event,
 888 * that page can be moved into the anon_vma that belongs to just that
 889 * process, so the rmap code will not search the parent or sibling
 890 * processes.
 891 */
 892void page_move_anon_rmap(struct page *page,
 893	struct vm_area_struct *vma, unsigned long address)
 894{
 895	struct anon_vma *anon_vma = vma->anon_vma;
 896
 
 
 897	VM_BUG_ON_PAGE(!PageLocked(page), page);
 898	VM_BUG_ON(!anon_vma);
 899	VM_BUG_ON_PAGE(page->index != linear_page_index(vma, address), page);
 900
 901	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
 902	page->mapping = (struct address_space *) anon_vma;
 
 
 
 
 
 903}
 904
 905/**
 906 * __page_set_anon_rmap - set up new anonymous rmap
 907 * @page:	Page to add to rmap	
 908 * @vma:	VM area to add page to.
 909 * @address:	User virtual address of the mapping	
 910 * @exclusive:	the page is exclusively owned by the current process
 911 */
 912static void __page_set_anon_rmap(struct page *page,
 913	struct vm_area_struct *vma, unsigned long address, int exclusive)
 914{
 915	struct anon_vma *anon_vma = vma->anon_vma;
 916
 917	BUG_ON(!anon_vma);
 918
 919	if (PageAnon(page))
 920		return;
 921
 922	/*
 923	 * If the page isn't exclusively mapped into this vma,
 924	 * we must use the _oldest_ possible anon_vma for the
 925	 * page mapping!
 926	 */
 927	if (!exclusive)
 928		anon_vma = anon_vma->root;
 929
 930	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
 931	page->mapping = (struct address_space *) anon_vma;
 932	page->index = linear_page_index(vma, address);
 933}
 934
 935/**
 936 * __page_check_anon_rmap - sanity check anonymous rmap addition
 937 * @page:	the page to add the mapping to
 938 * @vma:	the vm area in which the mapping is added
 939 * @address:	the user virtual address mapped
 940 */
 941static void __page_check_anon_rmap(struct page *page,
 942	struct vm_area_struct *vma, unsigned long address)
 943{
 944#ifdef CONFIG_DEBUG_VM
 945	/*
 946	 * The page's anon-rmap details (mapping and index) are guaranteed to
 947	 * be set up correctly at this point.
 948	 *
 949	 * We have exclusion against page_add_anon_rmap because the caller
 950	 * always holds the page locked, except if called from page_dup_rmap,
 951	 * in which case the page is already known to be setup.
 952	 *
 953	 * We have exclusion against page_add_new_anon_rmap because those pages
 954	 * are initially only visible via the pagetables, and the pte is locked
 955	 * over the call to page_add_new_anon_rmap.
 956	 */
 957	BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
 958	BUG_ON(page->index != linear_page_index(vma, address));
 959#endif
 960}
 961
 962/**
 963 * page_add_anon_rmap - add pte mapping to an anonymous page
 964 * @page:	the page to add the mapping to
 965 * @vma:	the vm area in which the mapping is added
 966 * @address:	the user virtual address mapped
 
 967 *
 968 * The caller needs to hold the pte lock, and the page must be locked in
 969 * the anon_vma case: to serialize mapping,index checking after setting,
 970 * and to ensure that PageAnon is not being upgraded racily to PageKsm
 971 * (but PageKsm is never downgraded to PageAnon).
 972 */
 973void page_add_anon_rmap(struct page *page,
 974	struct vm_area_struct *vma, unsigned long address)
 975{
 976	do_page_add_anon_rmap(page, vma, address, 0);
 977}
 978
 979/*
 980 * Special version of the above for do_swap_page, which often runs
 981 * into pages that are exclusively owned by the current process.
 982 * Everybody else should continue to use page_add_anon_rmap above.
 983 */
 984void do_page_add_anon_rmap(struct page *page,
 985	struct vm_area_struct *vma, unsigned long address, int exclusive)
 986{
 987	int first = atomic_inc_and_test(&page->_mapcount);
 
 
 
 
 
 
 
 
 
 
 
 
 988	if (first) {
 989		if (PageTransHuge(page))
 990			__inc_zone_page_state(page,
 991					      NR_ANON_TRANSPARENT_HUGEPAGES);
 992		__mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
 993				hpage_nr_pages(page));
 
 
 
 
 
 994	}
 995	if (unlikely(PageKsm(page)))
 996		return;
 997
 998	VM_BUG_ON_PAGE(!PageLocked(page), page);
 
 999	/* address might be in next vma when migration races vma_adjust */
1000	if (first)
1001		__page_set_anon_rmap(page, vma, address, exclusive);
 
1002	else
1003		__page_check_anon_rmap(page, vma, address);
1004}
1005
1006/**
1007 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1008 * @page:	the page to add the mapping to
1009 * @vma:	the vm area in which the mapping is added
1010 * @address:	the user virtual address mapped
 
1011 *
1012 * Same as page_add_anon_rmap but must only be called on *new* pages.
1013 * This means the inc-and-test can be bypassed.
1014 * Page does not have to be locked.
1015 */
1016void page_add_new_anon_rmap(struct page *page,
1017	struct vm_area_struct *vma, unsigned long address)
1018{
1019	VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1020	SetPageSwapBacked(page);
1021	atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1022	if (PageTransHuge(page))
1023		__inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1024	__mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1025			hpage_nr_pages(page));
 
 
 
 
 
 
 
 
 
1026	__page_set_anon_rmap(page, vma, address, 1);
1027	if (!mlocked_vma_newpage(vma, page)) {
1028		SetPageActive(page);
1029		lru_cache_add(page);
1030	} else
1031		add_page_to_unevictable_list(page);
1032}
1033
1034/**
1035 * page_add_file_rmap - add pte mapping to a file page
1036 * @page: the page to add the mapping to
 
1037 *
1038 * The caller needs to hold the pte lock.
1039 */
1040void page_add_file_rmap(struct page *page)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1041{
1042	bool locked;
1043	unsigned long flags;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1044
1045	mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1046	if (atomic_inc_and_test(&page->_mapcount)) {
1047		__inc_zone_page_state(page, NR_FILE_MAPPED);
1048		mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED);
 
 
1049	}
1050	mem_cgroup_end_update_page_stat(page, &locked, &flags);
1051}
1052
1053/**
1054 * page_remove_rmap - take down pte mapping from a page
1055 * @page: page to remove mapping from
 
1056 *
1057 * The caller needs to hold the pte lock.
1058 */
1059void page_remove_rmap(struct page *page)
1060{
1061	bool anon = PageAnon(page);
1062	bool locked;
1063	unsigned long flags;
1064
1065	/*
1066	 * The anon case has no mem_cgroup page_stat to update; but may
1067	 * uncharge_page() below, where the lock ordering can deadlock if
1068	 * we hold the lock against page_stat move: so avoid it on anon.
1069	 */
1070	if (!anon)
1071		mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1072
1073	/* page still mapped by someone else? */
1074	if (!atomic_add_negative(-1, &page->_mapcount))
1075		goto out;
1076
1077	/*
1078	 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1079	 * and not charged by memcg for now.
 
1080	 */
1081	if (unlikely(PageHuge(page)))
1082		goto out;
1083	if (anon) {
1084		mem_cgroup_uncharge_page(page);
1085		if (PageTransHuge(page))
1086			__dec_zone_page_state(page,
1087					      NR_ANON_TRANSPARENT_HUGEPAGES);
1088		__mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1089				-hpage_nr_pages(page));
1090	} else {
1091		__dec_zone_page_state(page, NR_FILE_MAPPED);
1092		mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED);
1093		mem_cgroup_end_update_page_stat(page, &locked, &flags);
1094	}
1095	if (unlikely(PageMlocked(page)))
1096		clear_page_mlock(page);
 
 
 
 
1097	/*
1098	 * It would be tidy to reset the PageAnon mapping here,
1099	 * but that might overwrite a racing page_add_anon_rmap
1100	 * which increments mapcount after us but sets mapping
1101	 * before us: so leave the reset to free_hot_cold_page,
1102	 * and remember that it's only reliable while mapped.
1103	 * Leaving it set also helps swapoff to reinstate ptes
1104	 * faster for those pages still in swapcache.
1105	 */
1106	return;
1107out:
1108	if (!anon)
1109		mem_cgroup_end_update_page_stat(page, &locked, &flags);
1110}
1111
1112/*
1113 * @arg: enum ttu_flags will be passed to this argument
1114 */
1115int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1116		     unsigned long address, void *arg)
1117{
1118	struct mm_struct *mm = vma->vm_mm;
1119	pte_t *pte;
 
 
 
 
1120	pte_t pteval;
1121	spinlock_t *ptl;
1122	int ret = SWAP_AGAIN;
 
1123	enum ttu_flags flags = (enum ttu_flags)arg;
1124
1125	pte = page_check_address(page, mm, address, &ptl, 0);
1126	if (!pte)
1127		goto out;
 
 
 
 
 
 
 
 
 
1128
1129	/*
1130	 * If the page is mlock()d, we cannot swap it out.
1131	 * If it's recently referenced (perhaps page_referenced
1132	 * skipped over this mm) then we should reactivate it.
1133	 */
1134	if (!(flags & TTU_IGNORE_MLOCK)) {
1135		if (vma->vm_flags & VM_LOCKED)
1136			goto out_mlock;
1137
1138		if (TTU_ACTION(flags) == TTU_MUNLOCK)
1139			goto out_unmap;
1140	}
1141	if (!(flags & TTU_IGNORE_ACCESS)) {
1142		if (ptep_clear_flush_young_notify(vma, address, pte)) {
1143			ret = SWAP_FAIL;
1144			goto out_unmap;
1145		}
1146  	}
 
 
 
 
 
 
 
 
1147
1148	/* Nuke the page table entry. */
1149	flush_cache_page(vma, address, page_to_pfn(page));
1150	pteval = ptep_clear_flush(vma, address, pte);
1151
1152	/* Move the dirty bit to the physical page now the pte is gone. */
1153	if (pte_dirty(pteval))
1154		set_page_dirty(page);
1155
1156	/* Update high watermark before we lower rss */
1157	update_hiwater_rss(mm);
1158
1159	if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1160		if (!PageHuge(page)) {
1161			if (PageAnon(page))
1162				dec_mm_counter(mm, MM_ANONPAGES);
1163			else
1164				dec_mm_counter(mm, MM_FILEPAGES);
1165		}
1166		set_pte_at(mm, address, pte,
1167			   swp_entry_to_pte(make_hwpoison_entry(page)));
1168	} else if (pte_unused(pteval)) {
1169		/*
1170		 * The guest indicated that the page content is of no
1171		 * interest anymore. Simply discard the pte, vmscan
1172		 * will take care of the rest.
1173		 */
1174		if (PageAnon(page))
1175			dec_mm_counter(mm, MM_ANONPAGES);
1176		else
1177			dec_mm_counter(mm, MM_FILEPAGES);
1178	} else if (PageAnon(page)) {
1179		swp_entry_t entry = { .val = page_private(page) };
1180		pte_t swp_pte;
1181
1182		if (PageSwapCache(page)) {
1183			/*
1184			 * Store the swap location in the pte.
1185			 * See handle_pte_fault() ...
1186			 */
1187			if (swap_duplicate(entry) < 0) {
1188				set_pte_at(mm, address, pte, pteval);
1189				ret = SWAP_FAIL;
1190				goto out_unmap;
1191			}
1192			if (list_empty(&mm->mmlist)) {
1193				spin_lock(&mmlist_lock);
1194				if (list_empty(&mm->mmlist))
1195					list_add(&mm->mmlist, &init_mm.mmlist);
1196				spin_unlock(&mmlist_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1197			}
1198			dec_mm_counter(mm, MM_ANONPAGES);
1199			inc_mm_counter(mm, MM_SWAPENTS);
1200		} else if (IS_ENABLED(CONFIG_MIGRATION)) {
 
 
 
 
 
 
 
1201			/*
1202			 * Store the pfn of the page in a special migration
1203			 * pte. do_swap_page() will wait until the migration
1204			 * pte is removed and then restart fault handling.
1205			 */
1206			BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
1207			entry = make_migration_entry(page, pte_write(pteval));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1208		}
1209		swp_pte = swp_entry_to_pte(entry);
1210		if (pte_soft_dirty(pteval))
1211			swp_pte = pte_swp_mksoft_dirty(swp_pte);
1212		set_pte_at(mm, address, pte, swp_pte);
1213		BUG_ON(pte_file(*pte));
1214	} else if (IS_ENABLED(CONFIG_MIGRATION) &&
1215		   (TTU_ACTION(flags) == TTU_MIGRATION)) {
1216		/* Establish migration entry for a file page */
1217		swp_entry_t entry;
1218		entry = make_migration_entry(page, pte_write(pteval));
1219		set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1220	} else
1221		dec_mm_counter(mm, MM_FILEPAGES);
1222
1223	page_remove_rmap(page);
1224	page_cache_release(page);
1225
1226out_unmap:
1227	pte_unmap_unlock(pte, ptl);
1228	if (ret != SWAP_FAIL)
1229		mmu_notifier_invalidate_page(mm, address);
1230out:
1231	return ret;
1232
1233out_mlock:
1234	pte_unmap_unlock(pte, ptl);
1235
1236
1237	/*
1238	 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1239	 * unstable result and race. Plus, We can't wait here because
1240	 * we now hold anon_vma->rwsem or mapping->i_mmap_mutex.
1241	 * if trylock failed, the page remain in evictable lru and later
1242	 * vmscan could retry to move the page to unevictable lru if the
1243	 * page is actually mlocked.
1244	 */
1245	if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1246		if (vma->vm_flags & VM_LOCKED) {
1247			mlock_vma_page(page);
1248			ret = SWAP_MLOCK;
1249		}
1250		up_read(&vma->vm_mm->mmap_sem);
1251	}
1252	return ret;
1253}
1254
1255/*
1256 * objrmap doesn't work for nonlinear VMAs because the assumption that
1257 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1258 * Consequently, given a particular page and its ->index, we cannot locate the
1259 * ptes which are mapping that page without an exhaustive linear search.
1260 *
1261 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1262 * maps the file to which the target page belongs.  The ->vm_private_data field
1263 * holds the current cursor into that scan.  Successive searches will circulate
1264 * around the vma's virtual address space.
1265 *
1266 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1267 * more scanning pressure is placed against them as well.   Eventually pages
1268 * will become fully unmapped and are eligible for eviction.
1269 *
1270 * For very sparsely populated VMAs this is a little inefficient - chances are
1271 * there there won't be many ptes located within the scan cluster.  In this case
1272 * maybe we could scan further - to the end of the pte page, perhaps.
1273 *
1274 * Mlocked pages:  check VM_LOCKED under mmap_sem held for read, if we can
1275 * acquire it without blocking.  If vma locked, mlock the pages in the cluster,
1276 * rather than unmapping them.  If we encounter the "check_page" that vmscan is
1277 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1278 */
1279#define CLUSTER_SIZE	min(32*PAGE_SIZE, PMD_SIZE)
1280#define CLUSTER_MASK	(~(CLUSTER_SIZE - 1))
1281
1282static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1283		struct vm_area_struct *vma, struct page *check_page)
1284{
1285	struct mm_struct *mm = vma->vm_mm;
1286	pmd_t *pmd;
1287	pte_t *pte;
1288	pte_t pteval;
1289	spinlock_t *ptl;
1290	struct page *page;
1291	unsigned long address;
1292	unsigned long mmun_start;	/* For mmu_notifiers */
1293	unsigned long mmun_end;		/* For mmu_notifiers */
1294	unsigned long end;
1295	int ret = SWAP_AGAIN;
1296	int locked_vma = 0;
1297
1298	address = (vma->vm_start + cursor) & CLUSTER_MASK;
1299	end = address + CLUSTER_SIZE;
1300	if (address < vma->vm_start)
1301		address = vma->vm_start;
1302	if (end > vma->vm_end)
1303		end = vma->vm_end;
1304
1305	pmd = mm_find_pmd(mm, address);
1306	if (!pmd)
1307		return ret;
1308
1309	mmun_start = address;
1310	mmun_end   = end;
1311	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1312
1313	/*
1314	 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1315	 * keep the sem while scanning the cluster for mlocking pages.
1316	 */
1317	if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1318		locked_vma = (vma->vm_flags & VM_LOCKED);
1319		if (!locked_vma)
1320			up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1321	}
 
 
1322
1323	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1324
1325	/* Update high watermark before we lower rss */
1326	update_hiwater_rss(mm);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1327
1328	for (; address < end; pte++, address += PAGE_SIZE) {
1329		if (!pte_present(*pte))
1330			continue;
1331		page = vm_normal_page(vma, address, *pte);
1332		BUG_ON(!page || PageAnon(page));
 
 
 
 
1333
1334		if (locked_vma) {
1335			if (page == check_page) {
1336				/* we know we have check_page locked */
1337				mlock_vma_page(page);
1338				ret = SWAP_MLOCK;
1339			} else if (trylock_page(page)) {
1340				/*
1341				 * If we can lock the page, perform mlock.
1342				 * Otherwise leave the page alone, it will be
1343				 * eventually encountered again later.
1344				 */
1345				mlock_vma_page(page);
1346				unlock_page(page);
 
 
 
1347			}
1348			continue;	/* don't unmap */
1349		}
1350
1351		if (ptep_clear_flush_young_notify(vma, address, pte))
1352			continue;
1353
1354		/* Nuke the page table entry. */
1355		flush_cache_page(vma, address, pte_pfn(*pte));
1356		pteval = ptep_clear_flush(vma, address, pte);
1357
1358		/* If nonlinear, store the file page offset in the pte. */
1359		if (page->index != linear_page_index(vma, address)) {
1360			pte_t ptfile = pgoff_to_pte(page->index);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1361			if (pte_soft_dirty(pteval))
1362				pte_file_mksoft_dirty(ptfile);
1363			set_pte_at(mm, address, pte, ptfile);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1364		}
1365
1366		/* Move the dirty bit to the physical page now the pte is gone. */
1367		if (pte_dirty(pteval))
1368			set_page_dirty(page);
1369
1370		page_remove_rmap(page);
1371		page_cache_release(page);
1372		dec_mm_counter(mm, MM_FILEPAGES);
1373		(*mapcount)--;
1374	}
1375	pte_unmap_unlock(pte - 1, ptl);
1376	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1377	if (locked_vma)
1378		up_read(&vma->vm_mm->mmap_sem);
1379	return ret;
1380}
1381
1382static int try_to_unmap_nonlinear(struct page *page,
1383		struct address_space *mapping, void *arg)
1384{
1385	struct vm_area_struct *vma;
1386	int ret = SWAP_AGAIN;
1387	unsigned long cursor;
1388	unsigned long max_nl_cursor = 0;
1389	unsigned long max_nl_size = 0;
1390	unsigned int mapcount;
1391
1392	list_for_each_entry(vma,
1393		&mapping->i_mmap_nonlinear, shared.nonlinear) {
1394
1395		cursor = (unsigned long) vma->vm_private_data;
1396		if (cursor > max_nl_cursor)
1397			max_nl_cursor = cursor;
1398		cursor = vma->vm_end - vma->vm_start;
1399		if (cursor > max_nl_size)
1400			max_nl_size = cursor;
1401	}
1402
1403	if (max_nl_size == 0) {	/* all nonlinears locked or reserved ? */
1404		return SWAP_FAIL;
1405	}
1406
1407	/*
1408	 * We don't try to search for this page in the nonlinear vmas,
1409	 * and page_referenced wouldn't have found it anyway.  Instead
1410	 * just walk the nonlinear vmas trying to age and unmap some.
1411	 * The mapcount of the page we came in with is irrelevant,
1412	 * but even so use it as a guide to how hard we should try?
1413	 */
1414	mapcount = page_mapcount(page);
1415	if (!mapcount)
1416		return ret;
1417
1418	cond_resched();
1419
1420	max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1421	if (max_nl_cursor == 0)
1422		max_nl_cursor = CLUSTER_SIZE;
1423
1424	do {
1425		list_for_each_entry(vma,
1426			&mapping->i_mmap_nonlinear, shared.nonlinear) {
1427
1428			cursor = (unsigned long) vma->vm_private_data;
1429			while (cursor < max_nl_cursor &&
1430				cursor < vma->vm_end - vma->vm_start) {
1431				if (try_to_unmap_cluster(cursor, &mapcount,
1432						vma, page) == SWAP_MLOCK)
1433					ret = SWAP_MLOCK;
1434				cursor += CLUSTER_SIZE;
1435				vma->vm_private_data = (void *) cursor;
1436				if ((int)mapcount <= 0)
1437					return ret;
1438			}
1439			vma->vm_private_data = (void *) max_nl_cursor;
1440		}
1441		cond_resched();
1442		max_nl_cursor += CLUSTER_SIZE;
1443	} while (max_nl_cursor <= max_nl_size);
1444
1445	/*
1446	 * Don't loop forever (perhaps all the remaining pages are
1447	 * in locked vmas).  Reset cursor on all unreserved nonlinear
1448	 * vmas, now forgetting on which ones it had fallen behind.
1449	 */
1450	list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.nonlinear)
1451		vma->vm_private_data = NULL;
1452
1453	return ret;
1454}
1455
1456bool is_vma_temporary_stack(struct vm_area_struct *vma)
1457{
1458	int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1459
1460	if (!maybe_stack)
1461		return false;
1462
1463	if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1464						VM_STACK_INCOMPLETE_SETUP)
1465		return true;
1466
1467	return false;
1468}
1469
1470static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1471{
1472	return is_vma_temporary_stack(vma);
1473}
1474
1475static int page_not_mapped(struct page *page)
1476{
1477	return !page_mapped(page);
1478};
1479
1480/**
1481 * try_to_unmap - try to remove all page table mappings to a page
1482 * @page: the page to get unmapped
1483 * @flags: action and flags
1484 *
1485 * Tries to remove all the page table entries which are mapping this
1486 * page, used in the pageout path.  Caller must hold the page lock.
1487 * Return values are:
1488 *
1489 * SWAP_SUCCESS	- we succeeded in removing all mappings
1490 * SWAP_AGAIN	- we missed a mapping, try again later
1491 * SWAP_FAIL	- the page is unswappable
1492 * SWAP_MLOCK	- page is mlocked.
1493 */
1494int try_to_unmap(struct page *page, enum ttu_flags flags)
1495{
1496	int ret;
1497	struct rmap_walk_control rwc = {
1498		.rmap_one = try_to_unmap_one,
1499		.arg = (void *)flags,
1500		.done = page_not_mapped,
1501		.file_nonlinear = try_to_unmap_nonlinear,
1502		.anon_lock = page_lock_anon_vma_read,
1503	};
1504
1505	VM_BUG_ON_PAGE(!PageHuge(page) && PageTransHuge(page), page);
1506
1507	/*
1508	 * During exec, a temporary VMA is setup and later moved.
1509	 * The VMA is moved under the anon_vma lock but not the
1510	 * page tables leading to a race where migration cannot
1511	 * find the migration ptes. Rather than increasing the
1512	 * locking requirements of exec(), migration skips
1513	 * temporary VMAs until after exec() completes.
1514	 */
1515	if (flags & TTU_MIGRATION && !PageKsm(page) && PageAnon(page))
 
1516		rwc.invalid_vma = invalid_migration_vma;
1517
1518	ret = rmap_walk(page, &rwc);
 
 
 
1519
1520	if (ret != SWAP_MLOCK && !page_mapped(page))
1521		ret = SWAP_SUCCESS;
1522	return ret;
1523}
1524
 
 
 
 
 
1525/**
1526 * try_to_munlock - try to munlock a page
1527 * @page: the page to be munlocked
1528 *
1529 * Called from munlock code.  Checks all of the VMAs mapping the page
1530 * to make sure nobody else has this page mlocked. The page will be
1531 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1532 *
1533 * Return values are:
1534 *
1535 * SWAP_AGAIN	- no vma is holding page mlocked, or,
1536 * SWAP_AGAIN	- page mapped in mlocked vma -- couldn't acquire mmap sem
1537 * SWAP_FAIL	- page cannot be located at present
1538 * SWAP_MLOCK	- page is now mlocked.
1539 */
1540int try_to_munlock(struct page *page)
 
1541{
1542	int ret;
1543	struct rmap_walk_control rwc = {
1544		.rmap_one = try_to_unmap_one,
1545		.arg = (void *)TTU_MUNLOCK,
1546		.done = page_not_mapped,
1547		/*
1548		 * We don't bother to try to find the munlocked page in
1549		 * nonlinears. It's costly. Instead, later, page reclaim logic
1550		 * may call try_to_unmap() and recover PG_mlocked lazily.
1551		 */
1552		.file_nonlinear = NULL,
1553		.anon_lock = page_lock_anon_vma_read,
1554
1555	};
1556
1557	VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
 
1558
1559	ret = rmap_walk(page, &rwc);
1560	return ret;
1561}
1562
1563void __put_anon_vma(struct anon_vma *anon_vma)
1564{
1565	struct anon_vma *root = anon_vma->root;
1566
1567	anon_vma_free(anon_vma);
1568	if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1569		anon_vma_free(root);
1570}
1571
1572static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1573					struct rmap_walk_control *rwc)
1574{
1575	struct anon_vma *anon_vma;
1576
1577	if (rwc->anon_lock)
1578		return rwc->anon_lock(page);
1579
1580	/*
1581	 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1582	 * because that depends on page_mapped(); but not all its usages
1583	 * are holding mmap_sem. Users without mmap_sem are required to
1584	 * take a reference count to prevent the anon_vma disappearing
1585	 */
1586	anon_vma = page_anon_vma(page);
1587	if (!anon_vma)
1588		return NULL;
1589
1590	anon_vma_lock_read(anon_vma);
1591	return anon_vma;
1592}
1593
1594/*
1595 * rmap_walk_anon - do something to anonymous page using the object-based
1596 * rmap method
1597 * @page: the page to be handled
1598 * @rwc: control variable according to each walk type
1599 *
1600 * Find all the mappings of a page using the mapping pointer and the vma chains
1601 * contained in the anon_vma struct it points to.
1602 *
1603 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1604 * where the page was found will be held for write.  So, we won't recheck
1605 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1606 * LOCKED.
1607 */
1608static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc)
 
1609{
1610	struct anon_vma *anon_vma;
1611	pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1612	struct anon_vma_chain *avc;
1613	int ret = SWAP_AGAIN;
1614
1615	anon_vma = rmap_walk_anon_lock(page, rwc);
 
 
 
 
 
 
1616	if (!anon_vma)
1617		return ret;
1618
1619	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
 
 
 
1620		struct vm_area_struct *vma = avc->vma;
1621		unsigned long address = vma_address(page, vma);
1622
 
 
1623		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1624			continue;
1625
1626		ret = rwc->rmap_one(page, vma, address, rwc->arg);
1627		if (ret != SWAP_AGAIN)
1628			break;
1629		if (rwc->done && rwc->done(page))
1630			break;
1631	}
1632	anon_vma_unlock_read(anon_vma);
1633	return ret;
 
1634}
1635
1636/*
1637 * rmap_walk_file - do something to file page using the object-based rmap method
1638 * @page: the page to be handled
1639 * @rwc: control variable according to each walk type
1640 *
1641 * Find all the mappings of a page using the mapping pointer and the vma chains
1642 * contained in the address_space struct it points to.
1643 *
1644 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1645 * where the page was found will be held for write.  So, we won't recheck
1646 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1647 * LOCKED.
1648 */
1649static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc)
 
1650{
1651	struct address_space *mapping = page->mapping;
1652	pgoff_t pgoff = page->index << compound_order(page);
1653	struct vm_area_struct *vma;
1654	int ret = SWAP_AGAIN;
1655
1656	/*
1657	 * The page lock not only makes sure that page->mapping cannot
1658	 * suddenly be NULLified by truncation, it makes sure that the
1659	 * structure at mapping cannot be freed and reused yet,
1660	 * so we can safely take mapping->i_mmap_mutex.
1661	 */
1662	VM_BUG_ON(!PageLocked(page));
1663
1664	if (!mapping)
1665		return ret;
1666	mutex_lock(&mapping->i_mmap_mutex);
1667	vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
 
 
 
 
 
1668		unsigned long address = vma_address(page, vma);
1669
 
 
1670		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1671			continue;
1672
1673		ret = rwc->rmap_one(page, vma, address, rwc->arg);
1674		if (ret != SWAP_AGAIN)
1675			goto done;
1676		if (rwc->done && rwc->done(page))
1677			goto done;
1678	}
1679
1680	if (!rwc->file_nonlinear)
1681		goto done;
1682
1683	if (list_empty(&mapping->i_mmap_nonlinear))
1684		goto done;
1685
1686	ret = rwc->file_nonlinear(page, mapping, rwc->arg);
1687
1688done:
1689	mutex_unlock(&mapping->i_mmap_mutex);
1690	return ret;
1691}
1692
1693int rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1694{
1695	if (unlikely(PageKsm(page)))
1696		return rmap_walk_ksm(page, rwc);
1697	else if (PageAnon(page))
1698		return rmap_walk_anon(page, rwc);
1699	else
1700		return rmap_walk_file(page, rwc);
 
 
 
 
 
 
 
 
 
 
 
1701}
1702
1703#ifdef CONFIG_HUGETLB_PAGE
1704/*
1705 * The following three functions are for anonymous (private mapped) hugepages.
1706 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1707 * and no lru code, because we handle hugepages differently from common pages.
1708 */
1709static void __hugepage_set_anon_rmap(struct page *page,
1710	struct vm_area_struct *vma, unsigned long address, int exclusive)
1711{
1712	struct anon_vma *anon_vma = vma->anon_vma;
1713
1714	BUG_ON(!anon_vma);
1715
1716	if (PageAnon(page))
1717		return;
1718	if (!exclusive)
1719		anon_vma = anon_vma->root;
1720
1721	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1722	page->mapping = (struct address_space *) anon_vma;
1723	page->index = linear_page_index(vma, address);
1724}
1725
1726void hugepage_add_anon_rmap(struct page *page,
1727			    struct vm_area_struct *vma, unsigned long address)
1728{
1729	struct anon_vma *anon_vma = vma->anon_vma;
1730	int first;
1731
1732	BUG_ON(!PageLocked(page));
1733	BUG_ON(!anon_vma);
1734	/* address might be in next vma when migration races vma_adjust */
1735	first = atomic_inc_and_test(&page->_mapcount);
1736	if (first)
1737		__hugepage_set_anon_rmap(page, vma, address, 0);
1738}
1739
1740void hugepage_add_new_anon_rmap(struct page *page,
1741			struct vm_area_struct *vma, unsigned long address)
1742{
1743	BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1744	atomic_set(&page->_mapcount, 0);
1745	__hugepage_set_anon_rmap(page, vma, address, 1);
1746}
1747#endif /* CONFIG_HUGETLB_PAGE */
   1/*
   2 * mm/rmap.c - physical to virtual reverse mappings
   3 *
   4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
   5 * Released under the General Public License (GPL).
   6 *
   7 * Simple, low overhead reverse mapping scheme.
   8 * Please try to keep this thing as modular as possible.
   9 *
  10 * Provides methods for unmapping each kind of mapped page:
  11 * the anon methods track anonymous pages, and
  12 * the file methods track pages belonging to an inode.
  13 *
  14 * Original design by Rik van Riel <riel@conectiva.com.br> 2001
  15 * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
  16 * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
  17 * Contributions by Hugh Dickins 2003, 2004
  18 */
  19
  20/*
  21 * Lock ordering in mm:
  22 *
  23 * inode->i_mutex	(while writing or truncating, not reading or faulting)
  24 *   mm->mmap_sem
  25 *     page->flags PG_locked (lock_page)
  26 *       hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
  27 *         mapping->i_mmap_rwsem
  28 *           anon_vma->rwsem
  29 *             mm->page_table_lock or pte_lock
  30 *               pgdat->lru_lock (in mark_page_accessed, isolate_lru_page)
  31 *               swap_lock (in swap_duplicate, swap_info_get)
  32 *                 mmlist_lock (in mmput, drain_mmlist and others)
  33 *                 mapping->private_lock (in __set_page_dirty_buffers)
  34 *                   mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
  35 *                     i_pages lock (widely used)
  36 *                 inode->i_lock (in set_page_dirty's __mark_inode_dirty)
  37 *                 bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
  38 *                   sb_lock (within inode_lock in fs/fs-writeback.c)
  39 *                   i_pages lock (widely used, in set_page_dirty,
  40 *                             in arch-dependent flush_dcache_mmap_lock,
  41 *                             within bdi.wb->list_lock in __sync_single_inode)
  42 *
  43 * anon_vma->rwsem,mapping->i_mutex      (memory_failure, collect_procs_anon)
  44 *   ->tasklist_lock
  45 *     pte map lock
  46 */
  47
  48#include <linux/mm.h>
  49#include <linux/sched/mm.h>
  50#include <linux/sched/task.h>
  51#include <linux/pagemap.h>
  52#include <linux/swap.h>
  53#include <linux/swapops.h>
  54#include <linux/slab.h>
  55#include <linux/init.h>
  56#include <linux/ksm.h>
  57#include <linux/rmap.h>
  58#include <linux/rcupdate.h>
  59#include <linux/export.h>
  60#include <linux/memcontrol.h>
  61#include <linux/mmu_notifier.h>
  62#include <linux/migrate.h>
  63#include <linux/hugetlb.h>
  64#include <linux/huge_mm.h>
  65#include <linux/backing-dev.h>
  66#include <linux/page_idle.h>
  67#include <linux/memremap.h>
  68#include <linux/userfaultfd_k.h>
  69
  70#include <asm/tlbflush.h>
  71
  72#include <trace/events/tlb.h>
  73
  74#include "internal.h"
  75
  76static struct kmem_cache *anon_vma_cachep;
  77static struct kmem_cache *anon_vma_chain_cachep;
  78
  79static inline struct anon_vma *anon_vma_alloc(void)
  80{
  81	struct anon_vma *anon_vma;
  82
  83	anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
  84	if (anon_vma) {
  85		atomic_set(&anon_vma->refcount, 1);
  86		anon_vma->degree = 1;	/* Reference for first vma */
  87		anon_vma->parent = anon_vma;
  88		/*
  89		 * Initialise the anon_vma root to point to itself. If called
  90		 * from fork, the root will be reset to the parents anon_vma.
  91		 */
  92		anon_vma->root = anon_vma;
  93	}
  94
  95	return anon_vma;
  96}
  97
  98static inline void anon_vma_free(struct anon_vma *anon_vma)
  99{
 100	VM_BUG_ON(atomic_read(&anon_vma->refcount));
 101
 102	/*
 103	 * Synchronize against page_lock_anon_vma_read() such that
 104	 * we can safely hold the lock without the anon_vma getting
 105	 * freed.
 106	 *
 107	 * Relies on the full mb implied by the atomic_dec_and_test() from
 108	 * put_anon_vma() against the acquire barrier implied by
 109	 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
 110	 *
 111	 * page_lock_anon_vma_read()	VS	put_anon_vma()
 112	 *   down_read_trylock()		  atomic_dec_and_test()
 113	 *   LOCK				  MB
 114	 *   atomic_read()			  rwsem_is_locked()
 115	 *
 116	 * LOCK should suffice since the actual taking of the lock must
 117	 * happen _before_ what follows.
 118	 */
 119	might_sleep();
 120	if (rwsem_is_locked(&anon_vma->root->rwsem)) {
 121		anon_vma_lock_write(anon_vma);
 122		anon_vma_unlock_write(anon_vma);
 123	}
 124
 125	kmem_cache_free(anon_vma_cachep, anon_vma);
 126}
 127
 128static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
 129{
 130	return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
 131}
 132
 133static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
 134{
 135	kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
 136}
 137
 138static void anon_vma_chain_link(struct vm_area_struct *vma,
 139				struct anon_vma_chain *avc,
 140				struct anon_vma *anon_vma)
 141{
 142	avc->vma = vma;
 143	avc->anon_vma = anon_vma;
 144	list_add(&avc->same_vma, &vma->anon_vma_chain);
 145	anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
 146}
 147
 148/**
 149 * __anon_vma_prepare - attach an anon_vma to a memory region
 150 * @vma: the memory region in question
 151 *
 152 * This makes sure the memory mapping described by 'vma' has
 153 * an 'anon_vma' attached to it, so that we can associate the
 154 * anonymous pages mapped into it with that anon_vma.
 155 *
 156 * The common case will be that we already have one, which
 157 * is handled inline by anon_vma_prepare(). But if
 158 * not we either need to find an adjacent mapping that we
 159 * can re-use the anon_vma from (very common when the only
 160 * reason for splitting a vma has been mprotect()), or we
 161 * allocate a new one.
 162 *
 163 * Anon-vma allocations are very subtle, because we may have
 164 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
 165 * and that may actually touch the spinlock even in the newly
 166 * allocated vma (it depends on RCU to make sure that the
 167 * anon_vma isn't actually destroyed).
 168 *
 169 * As a result, we need to do proper anon_vma locking even
 170 * for the new allocation. At the same time, we do not want
 171 * to do any locking for the common case of already having
 172 * an anon_vma.
 173 *
 174 * This must be called with the mmap_sem held for reading.
 175 */
 176int __anon_vma_prepare(struct vm_area_struct *vma)
 177{
 178	struct mm_struct *mm = vma->vm_mm;
 179	struct anon_vma *anon_vma, *allocated;
 180	struct anon_vma_chain *avc;
 181
 182	might_sleep();
 
 
 
 
 
 
 
 183
 184	avc = anon_vma_chain_alloc(GFP_KERNEL);
 185	if (!avc)
 186		goto out_enomem;
 
 
 
 
 
 187
 188	anon_vma = find_mergeable_anon_vma(vma);
 189	allocated = NULL;
 190	if (!anon_vma) {
 191		anon_vma = anon_vma_alloc();
 192		if (unlikely(!anon_vma))
 193			goto out_enomem_free_avc;
 194		allocated = anon_vma;
 195	}
 
 
 
 196
 197	anon_vma_lock_write(anon_vma);
 198	/* page_table_lock to protect against threads */
 199	spin_lock(&mm->page_table_lock);
 200	if (likely(!vma->anon_vma)) {
 201		vma->anon_vma = anon_vma;
 202		anon_vma_chain_link(vma, avc, anon_vma);
 203		/* vma reference or self-parent link for new root */
 204		anon_vma->degree++;
 205		allocated = NULL;
 206		avc = NULL;
 207	}
 208	spin_unlock(&mm->page_table_lock);
 209	anon_vma_unlock_write(anon_vma);
 210
 211	if (unlikely(allocated))
 212		put_anon_vma(allocated);
 213	if (unlikely(avc))
 214		anon_vma_chain_free(avc);
 215
 216	return 0;
 217
 218 out_enomem_free_avc:
 219	anon_vma_chain_free(avc);
 220 out_enomem:
 221	return -ENOMEM;
 222}
 223
 224/*
 225 * This is a useful helper function for locking the anon_vma root as
 226 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
 227 * have the same vma.
 228 *
 229 * Such anon_vma's should have the same root, so you'd expect to see
 230 * just a single mutex_lock for the whole traversal.
 231 */
 232static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
 233{
 234	struct anon_vma *new_root = anon_vma->root;
 235	if (new_root != root) {
 236		if (WARN_ON_ONCE(root))
 237			up_write(&root->rwsem);
 238		root = new_root;
 239		down_write(&root->rwsem);
 240	}
 241	return root;
 242}
 243
 244static inline void unlock_anon_vma_root(struct anon_vma *root)
 245{
 246	if (root)
 247		up_write(&root->rwsem);
 248}
 249
 250/*
 251 * Attach the anon_vmas from src to dst.
 252 * Returns 0 on success, -ENOMEM on failure.
 253 *
 254 * If dst->anon_vma is NULL this function tries to find and reuse existing
 255 * anon_vma which has no vmas and only one child anon_vma. This prevents
 256 * degradation of anon_vma hierarchy to endless linear chain in case of
 257 * constantly forking task. On the other hand, an anon_vma with more than one
 258 * child isn't reused even if there was no alive vma, thus rmap walker has a
 259 * good chance of avoiding scanning the whole hierarchy when it searches where
 260 * page is mapped.
 261 */
 262int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
 263{
 264	struct anon_vma_chain *avc, *pavc;
 265	struct anon_vma *root = NULL;
 266
 267	list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
 268		struct anon_vma *anon_vma;
 269
 270		avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
 271		if (unlikely(!avc)) {
 272			unlock_anon_vma_root(root);
 273			root = NULL;
 274			avc = anon_vma_chain_alloc(GFP_KERNEL);
 275			if (!avc)
 276				goto enomem_failure;
 277		}
 278		anon_vma = pavc->anon_vma;
 279		root = lock_anon_vma_root(root, anon_vma);
 280		anon_vma_chain_link(dst, avc, anon_vma);
 281
 282		/*
 283		 * Reuse existing anon_vma if its degree lower than two,
 284		 * that means it has no vma and only one anon_vma child.
 285		 *
 286		 * Do not chose parent anon_vma, otherwise first child
 287		 * will always reuse it. Root anon_vma is never reused:
 288		 * it has self-parent reference and at least one child.
 289		 */
 290		if (!dst->anon_vma && anon_vma != src->anon_vma &&
 291				anon_vma->degree < 2)
 292			dst->anon_vma = anon_vma;
 293	}
 294	if (dst->anon_vma)
 295		dst->anon_vma->degree++;
 296	unlock_anon_vma_root(root);
 297	return 0;
 298
 299 enomem_failure:
 300	/*
 301	 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
 302	 * decremented in unlink_anon_vmas().
 303	 * We can safely do this because callers of anon_vma_clone() don't care
 304	 * about dst->anon_vma if anon_vma_clone() failed.
 305	 */
 306	dst->anon_vma = NULL;
 307	unlink_anon_vmas(dst);
 308	return -ENOMEM;
 309}
 310
 311/*
 312 * Attach vma to its own anon_vma, as well as to the anon_vmas that
 313 * the corresponding VMA in the parent process is attached to.
 314 * Returns 0 on success, non-zero on failure.
 315 */
 316int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
 317{
 318	struct anon_vma_chain *avc;
 319	struct anon_vma *anon_vma;
 320	int error;
 321
 322	/* Don't bother if the parent process has no anon_vma here. */
 323	if (!pvma->anon_vma)
 324		return 0;
 325
 326	/* Drop inherited anon_vma, we'll reuse existing or allocate new. */
 327	vma->anon_vma = NULL;
 328
 329	/*
 330	 * First, attach the new VMA to the parent VMA's anon_vmas,
 331	 * so rmap can find non-COWed pages in child processes.
 332	 */
 333	error = anon_vma_clone(vma, pvma);
 334	if (error)
 335		return error;
 336
 337	/* An existing anon_vma has been reused, all done then. */
 338	if (vma->anon_vma)
 339		return 0;
 340
 341	/* Then add our own anon_vma. */
 342	anon_vma = anon_vma_alloc();
 343	if (!anon_vma)
 344		goto out_error;
 345	avc = anon_vma_chain_alloc(GFP_KERNEL);
 346	if (!avc)
 347		goto out_error_free_anon_vma;
 348
 349	/*
 350	 * The root anon_vma's spinlock is the lock actually used when we
 351	 * lock any of the anon_vmas in this anon_vma tree.
 352	 */
 353	anon_vma->root = pvma->anon_vma->root;
 354	anon_vma->parent = pvma->anon_vma;
 355	/*
 356	 * With refcounts, an anon_vma can stay around longer than the
 357	 * process it belongs to. The root anon_vma needs to be pinned until
 358	 * this anon_vma is freed, because the lock lives in the root.
 359	 */
 360	get_anon_vma(anon_vma->root);
 361	/* Mark this anon_vma as the one where our new (COWed) pages go. */
 362	vma->anon_vma = anon_vma;
 363	anon_vma_lock_write(anon_vma);
 364	anon_vma_chain_link(vma, avc, anon_vma);
 365	anon_vma->parent->degree++;
 366	anon_vma_unlock_write(anon_vma);
 367
 368	return 0;
 369
 370 out_error_free_anon_vma:
 371	put_anon_vma(anon_vma);
 372 out_error:
 373	unlink_anon_vmas(vma);
 374	return -ENOMEM;
 375}
 376
 377void unlink_anon_vmas(struct vm_area_struct *vma)
 378{
 379	struct anon_vma_chain *avc, *next;
 380	struct anon_vma *root = NULL;
 381
 382	/*
 383	 * Unlink each anon_vma chained to the VMA.  This list is ordered
 384	 * from newest to oldest, ensuring the root anon_vma gets freed last.
 385	 */
 386	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 387		struct anon_vma *anon_vma = avc->anon_vma;
 388
 389		root = lock_anon_vma_root(root, anon_vma);
 390		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
 391
 392		/*
 393		 * Leave empty anon_vmas on the list - we'll need
 394		 * to free them outside the lock.
 395		 */
 396		if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
 397			anon_vma->parent->degree--;
 398			continue;
 399		}
 400
 401		list_del(&avc->same_vma);
 402		anon_vma_chain_free(avc);
 403	}
 404	if (vma->anon_vma)
 405		vma->anon_vma->degree--;
 406	unlock_anon_vma_root(root);
 407
 408	/*
 409	 * Iterate the list once more, it now only contains empty and unlinked
 410	 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
 411	 * needing to write-acquire the anon_vma->root->rwsem.
 412	 */
 413	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 414		struct anon_vma *anon_vma = avc->anon_vma;
 415
 416		VM_WARN_ON(anon_vma->degree);
 417		put_anon_vma(anon_vma);
 418
 419		list_del(&avc->same_vma);
 420		anon_vma_chain_free(avc);
 421	}
 422}
 423
 424static void anon_vma_ctor(void *data)
 425{
 426	struct anon_vma *anon_vma = data;
 427
 428	init_rwsem(&anon_vma->rwsem);
 429	atomic_set(&anon_vma->refcount, 0);
 430	anon_vma->rb_root = RB_ROOT_CACHED;
 431}
 432
 433void __init anon_vma_init(void)
 434{
 435	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
 436			0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
 437			anon_vma_ctor);
 438	anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
 439			SLAB_PANIC|SLAB_ACCOUNT);
 440}
 441
 442/*
 443 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
 444 *
 445 * Since there is no serialization what so ever against page_remove_rmap()
 446 * the best this function can do is return a locked anon_vma that might
 447 * have been relevant to this page.
 448 *
 449 * The page might have been remapped to a different anon_vma or the anon_vma
 450 * returned may already be freed (and even reused).
 451 *
 452 * In case it was remapped to a different anon_vma, the new anon_vma will be a
 453 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
 454 * ensure that any anon_vma obtained from the page will still be valid for as
 455 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
 456 *
 457 * All users of this function must be very careful when walking the anon_vma
 458 * chain and verify that the page in question is indeed mapped in it
 459 * [ something equivalent to page_mapped_in_vma() ].
 460 *
 461 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
 462 * that the anon_vma pointer from page->mapping is valid if there is a
 463 * mapcount, we can dereference the anon_vma after observing those.
 464 */
 465struct anon_vma *page_get_anon_vma(struct page *page)
 466{
 467	struct anon_vma *anon_vma = NULL;
 468	unsigned long anon_mapping;
 469
 470	rcu_read_lock();
 471	anon_mapping = (unsigned long)READ_ONCE(page->mapping);
 472	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 473		goto out;
 474	if (!page_mapped(page))
 475		goto out;
 476
 477	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 478	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 479		anon_vma = NULL;
 480		goto out;
 481	}
 482
 483	/*
 484	 * If this page is still mapped, then its anon_vma cannot have been
 485	 * freed.  But if it has been unmapped, we have no security against the
 486	 * anon_vma structure being freed and reused (for another anon_vma:
 487	 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
 488	 * above cannot corrupt).
 489	 */
 490	if (!page_mapped(page)) {
 491		rcu_read_unlock();
 492		put_anon_vma(anon_vma);
 493		return NULL;
 494	}
 495out:
 496	rcu_read_unlock();
 497
 498	return anon_vma;
 499}
 500
 501/*
 502 * Similar to page_get_anon_vma() except it locks the anon_vma.
 503 *
 504 * Its a little more complex as it tries to keep the fast path to a single
 505 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
 506 * reference like with page_get_anon_vma() and then block on the mutex.
 507 */
 508struct anon_vma *page_lock_anon_vma_read(struct page *page)
 509{
 510	struct anon_vma *anon_vma = NULL;
 511	struct anon_vma *root_anon_vma;
 512	unsigned long anon_mapping;
 513
 514	rcu_read_lock();
 515	anon_mapping = (unsigned long)READ_ONCE(page->mapping);
 516	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 517		goto out;
 518	if (!page_mapped(page))
 519		goto out;
 520
 521	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 522	root_anon_vma = READ_ONCE(anon_vma->root);
 523	if (down_read_trylock(&root_anon_vma->rwsem)) {
 524		/*
 525		 * If the page is still mapped, then this anon_vma is still
 526		 * its anon_vma, and holding the mutex ensures that it will
 527		 * not go away, see anon_vma_free().
 528		 */
 529		if (!page_mapped(page)) {
 530			up_read(&root_anon_vma->rwsem);
 531			anon_vma = NULL;
 532		}
 533		goto out;
 534	}
 535
 536	/* trylock failed, we got to sleep */
 537	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 538		anon_vma = NULL;
 539		goto out;
 540	}
 541
 542	if (!page_mapped(page)) {
 543		rcu_read_unlock();
 544		put_anon_vma(anon_vma);
 545		return NULL;
 
 546	}
 547
 548	/* we pinned the anon_vma, its safe to sleep */
 549	rcu_read_unlock();
 550	anon_vma_lock_read(anon_vma);
 551
 552	if (atomic_dec_and_test(&anon_vma->refcount)) {
 553		/*
 554		 * Oops, we held the last refcount, release the lock
 555		 * and bail -- can't simply use put_anon_vma() because
 556		 * we'll deadlock on the anon_vma_lock_write() recursion.
 557		 */
 558		anon_vma_unlock_read(anon_vma);
 559		__put_anon_vma(anon_vma);
 560		anon_vma = NULL;
 561	}
 562
 563	return anon_vma;
 564
 565out:
 566	rcu_read_unlock();
 567	return anon_vma;
 568}
 569
 570void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
 571{
 572	anon_vma_unlock_read(anon_vma);
 573}
 574
 575#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
 576/*
 577 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
 578 * important if a PTE was dirty when it was unmapped that it's flushed
 579 * before any IO is initiated on the page to prevent lost writes. Similarly,
 580 * it must be flushed before freeing to prevent data leakage.
 581 */
 582void try_to_unmap_flush(void)
 
 583{
 584	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 585
 586	if (!tlb_ubc->flush_required)
 587		return;
 588
 589	arch_tlbbatch_flush(&tlb_ubc->arch);
 590	tlb_ubc->flush_required = false;
 591	tlb_ubc->writable = false;
 592}
 593
 594/* Flush iff there are potentially writable TLB entries that can race with IO */
 595void try_to_unmap_flush_dirty(void)
 596{
 597	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 598
 599	if (tlb_ubc->writable)
 600		try_to_unmap_flush();
 601}
 602
 603static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
 604{
 605	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 606
 607	arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
 608	tlb_ubc->flush_required = true;
 609
 610	/*
 611	 * Ensure compiler does not re-order the setting of tlb_flush_batched
 612	 * before the PTE is cleared.
 613	 */
 614	barrier();
 615	mm->tlb_flush_batched = true;
 616
 617	/*
 618	 * If the PTE was dirty then it's best to assume it's writable. The
 619	 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
 620	 * before the page is queued for IO.
 621	 */
 622	if (writable)
 623		tlb_ubc->writable = true;
 624}
 625
 626/*
 627 * Returns true if the TLB flush should be deferred to the end of a batch of
 628 * unmap operations to reduce IPIs.
 629 */
 630static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
 631{
 632	bool should_defer = false;
 633
 634	if (!(flags & TTU_BATCH_FLUSH))
 635		return false;
 636
 637	/* If remote CPUs need to be flushed then defer batch the flush */
 638	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
 639		should_defer = true;
 640	put_cpu();
 641
 642	return should_defer;
 643}
 644
 645/*
 646 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
 647 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
 648 * operation such as mprotect or munmap to race between reclaim unmapping
 649 * the page and flushing the page. If this race occurs, it potentially allows
 650 * access to data via a stale TLB entry. Tracking all mm's that have TLB
 651 * batching in flight would be expensive during reclaim so instead track
 652 * whether TLB batching occurred in the past and if so then do a flush here
 653 * if required. This will cost one additional flush per reclaim cycle paid
 654 * by the first operation at risk such as mprotect and mumap.
 655 *
 656 * This must be called under the PTL so that an access to tlb_flush_batched
 657 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
 658 * via the PTL.
 659 */
 660void flush_tlb_batched_pending(struct mm_struct *mm)
 661{
 662	if (mm->tlb_flush_batched) {
 663		flush_tlb_mm(mm);
 664
 665		/*
 666		 * Do not allow the compiler to re-order the clearing of
 667		 * tlb_flush_batched before the tlb is flushed.
 668		 */
 669		barrier();
 670		mm->tlb_flush_batched = false;
 671	}
 672}
 673#else
 674static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
 675{
 676}
 677
 678static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
 679{
 680	return false;
 681}
 682#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
 683
 684/*
 685 * At what user virtual address is page expected in vma?
 686 * Caller should check the page is actually part of the vma.
 687 */
 688unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
 689{
 690	unsigned long address;
 691	if (PageAnon(page)) {
 692		struct anon_vma *page__anon_vma = page_anon_vma(page);
 693		/*
 694		 * Note: swapoff's unuse_vma() is more efficient with this
 695		 * check, and needs it to match anon_vma when KSM is active.
 696		 */
 697		if (!vma->anon_vma || !page__anon_vma ||
 698		    vma->anon_vma->root != page__anon_vma->root)
 699			return -EFAULT;
 700	} else if (page->mapping) {
 701		if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
 
 702			return -EFAULT;
 703	} else
 704		return -EFAULT;
 705	address = __vma_address(page, vma);
 706	if (unlikely(address < vma->vm_start || address >= vma->vm_end))
 707		return -EFAULT;
 708	return address;
 709}
 710
 711pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
 712{
 713	pgd_t *pgd;
 714	p4d_t *p4d;
 715	pud_t *pud;
 716	pmd_t *pmd = NULL;
 717	pmd_t pmde;
 718
 719	pgd = pgd_offset(mm, address);
 720	if (!pgd_present(*pgd))
 721		goto out;
 722
 723	p4d = p4d_offset(pgd, address);
 724	if (!p4d_present(*p4d))
 725		goto out;
 726
 727	pud = pud_offset(p4d, address);
 728	if (!pud_present(*pud))
 729		goto out;
 730
 731	pmd = pmd_offset(pud, address);
 732	/*
 733	 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
 734	 * without holding anon_vma lock for write.  So when looking for a
 735	 * genuine pmde (in which to find pte), test present and !THP together.
 736	 */
 737	pmde = *pmd;
 738	barrier();
 739	if (!pmd_present(pmde) || pmd_trans_huge(pmde))
 740		pmd = NULL;
 741out:
 742	return pmd;
 743}
 744
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 745struct page_referenced_arg {
 746	int mapcount;
 747	int referenced;
 748	unsigned long vm_flags;
 749	struct mem_cgroup *memcg;
 750};
 751/*
 752 * arg: page_referenced_arg will be passed
 753 */
 754static bool page_referenced_one(struct page *page, struct vm_area_struct *vma,
 755			unsigned long address, void *arg)
 756{
 
 
 
 757	struct page_referenced_arg *pra = arg;
 758	struct page_vma_mapped_walk pvmw = {
 759		.page = page,
 760		.vma = vma,
 761		.address = address,
 762	};
 763	int referenced = 0;
 764
 765	while (page_vma_mapped_walk(&pvmw)) {
 766		address = pvmw.address;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 767
 768		if (vma->vm_flags & VM_LOCKED) {
 769			page_vma_mapped_walk_done(&pvmw);
 770			pra->vm_flags |= VM_LOCKED;
 771			return false; /* To break the loop */
 772		}
 773
 774		if (pvmw.pte) {
 775			if (ptep_clear_flush_young_notify(vma, address,
 776						pvmw.pte)) {
 777				/*
 778				 * Don't treat a reference through
 779				 * a sequentially read mapping as such.
 780				 * If the page has been used in another mapping,
 781				 * we will catch it; if this other mapping is
 782				 * already gone, the unmap path will have set
 783				 * PG_referenced or activated the page.
 784				 */
 785				if (likely(!(vma->vm_flags & VM_SEQ_READ)))
 786					referenced++;
 787			}
 788		} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
 789			if (pmdp_clear_flush_young_notify(vma, address,
 790						pvmw.pmd))
 791				referenced++;
 792		} else {
 793			/* unexpected pmd-mapped page? */
 794			WARN_ON_ONCE(1);
 795		}
 796
 797		pra->mapcount--;
 798	}
 799
 800	if (referenced)
 801		clear_page_idle(page);
 802	if (test_and_clear_page_young(page))
 803		referenced++;
 804
 805	if (referenced) {
 806		pra->referenced++;
 807		pra->vm_flags |= vma->vm_flags;
 808	}
 809
 
 810	if (!pra->mapcount)
 811		return false; /* To break the loop */
 812
 813	return true;
 814}
 815
 816static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
 817{
 818	struct page_referenced_arg *pra = arg;
 819	struct mem_cgroup *memcg = pra->memcg;
 820
 821	if (!mm_match_cgroup(vma->vm_mm, memcg))
 822		return true;
 823
 824	return false;
 825}
 826
 827/**
 828 * page_referenced - test if the page was referenced
 829 * @page: the page to test
 830 * @is_locked: caller holds lock on the page
 831 * @memcg: target memory cgroup
 832 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
 833 *
 834 * Quick test_and_clear_referenced for all mappings to a page,
 835 * returns the number of ptes which referenced the page.
 836 */
 837int page_referenced(struct page *page,
 838		    int is_locked,
 839		    struct mem_cgroup *memcg,
 840		    unsigned long *vm_flags)
 841{
 
 842	int we_locked = 0;
 843	struct page_referenced_arg pra = {
 844		.mapcount = total_mapcount(page),
 845		.memcg = memcg,
 846	};
 847	struct rmap_walk_control rwc = {
 848		.rmap_one = page_referenced_one,
 849		.arg = (void *)&pra,
 850		.anon_lock = page_lock_anon_vma_read,
 851	};
 852
 853	*vm_flags = 0;
 854	if (!pra.mapcount)
 855		return 0;
 856
 857	if (!page_rmapping(page))
 858		return 0;
 859
 860	if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
 861		we_locked = trylock_page(page);
 862		if (!we_locked)
 863			return 1;
 864	}
 865
 866	/*
 867	 * If we are reclaiming on behalf of a cgroup, skip
 868	 * counting on behalf of references from different
 869	 * cgroups
 870	 */
 871	if (memcg) {
 872		rwc.invalid_vma = invalid_page_referenced_vma;
 873	}
 874
 875	rmap_walk(page, &rwc);
 876	*vm_flags = pra.vm_flags;
 877
 878	if (we_locked)
 879		unlock_page(page);
 880
 881	return pra.referenced;
 882}
 883
 884static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
 885			    unsigned long address, void *arg)
 886{
 887	struct page_vma_mapped_walk pvmw = {
 888		.page = page,
 889		.vma = vma,
 890		.address = address,
 891		.flags = PVMW_SYNC,
 892	};
 893	struct mmu_notifier_range range;
 894	int *cleaned = arg;
 895
 896	/*
 897	 * We have to assume the worse case ie pmd for invalidation. Note that
 898	 * the page can not be free from this function.
 899	 */
 900	mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
 901				0, vma, vma->vm_mm, address,
 902				min(vma->vm_end, address + page_size(page)));
 903	mmu_notifier_invalidate_range_start(&range);
 904
 905	while (page_vma_mapped_walk(&pvmw)) {
 906		int ret = 0;
 907
 908		address = pvmw.address;
 909		if (pvmw.pte) {
 910			pte_t entry;
 911			pte_t *pte = pvmw.pte;
 912
 913			if (!pte_dirty(*pte) && !pte_write(*pte))
 914				continue;
 915
 916			flush_cache_page(vma, address, pte_pfn(*pte));
 917			entry = ptep_clear_flush(vma, address, pte);
 918			entry = pte_wrprotect(entry);
 919			entry = pte_mkclean(entry);
 920			set_pte_at(vma->vm_mm, address, pte, entry);
 921			ret = 1;
 922		} else {
 923#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
 924			pmd_t *pmd = pvmw.pmd;
 925			pmd_t entry;
 926
 927			if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
 928				continue;
 929
 930			flush_cache_page(vma, address, page_to_pfn(page));
 931			entry = pmdp_invalidate(vma, address, pmd);
 932			entry = pmd_wrprotect(entry);
 933			entry = pmd_mkclean(entry);
 934			set_pmd_at(vma->vm_mm, address, pmd, entry);
 935			ret = 1;
 936#else
 937			/* unexpected pmd-mapped page? */
 938			WARN_ON_ONCE(1);
 939#endif
 940		}
 941
 942		/*
 943		 * No need to call mmu_notifier_invalidate_range() as we are
 944		 * downgrading page table protection not changing it to point
 945		 * to a new page.
 946		 *
 947		 * See Documentation/vm/mmu_notifier.rst
 948		 */
 949		if (ret)
 950			(*cleaned)++;
 
 
 
 
 951	}
 952
 953	mmu_notifier_invalidate_range_end(&range);
 954
 955	return true;
 956}
 957
 958static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
 959{
 960	if (vma->vm_flags & VM_SHARED)
 961		return false;
 962
 963	return true;
 964}
 965
 966int page_mkclean(struct page *page)
 967{
 968	int cleaned = 0;
 969	struct address_space *mapping;
 970	struct rmap_walk_control rwc = {
 971		.arg = (void *)&cleaned,
 972		.rmap_one = page_mkclean_one,
 973		.invalid_vma = invalid_mkclean_vma,
 974	};
 975
 976	BUG_ON(!PageLocked(page));
 977
 978	if (!page_mapped(page))
 979		return 0;
 980
 981	mapping = page_mapping(page);
 982	if (!mapping)
 983		return 0;
 984
 985	rmap_walk(page, &rwc);
 986
 987	return cleaned;
 988}
 989EXPORT_SYMBOL_GPL(page_mkclean);
 990
 991/**
 992 * page_move_anon_rmap - move a page to our anon_vma
 993 * @page:	the page to move to our anon_vma
 994 * @vma:	the vma the page belongs to
 
 995 *
 996 * When a page belongs exclusively to one process after a COW event,
 997 * that page can be moved into the anon_vma that belongs to just that
 998 * process, so the rmap code will not search the parent or sibling
 999 * processes.
1000 */
1001void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
 
1002{
1003	struct anon_vma *anon_vma = vma->anon_vma;
1004
1005	page = compound_head(page);
1006
1007	VM_BUG_ON_PAGE(!PageLocked(page), page);
1008	VM_BUG_ON_VMA(!anon_vma, vma);
 
1009
1010	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1011	/*
1012	 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1013	 * simultaneously, so a concurrent reader (eg page_referenced()'s
1014	 * PageAnon()) will not see one without the other.
1015	 */
1016	WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1017}
1018
1019/**
1020 * __page_set_anon_rmap - set up new anonymous rmap
1021 * @page:	Page or Hugepage to add to rmap
1022 * @vma:	VM area to add page to.
1023 * @address:	User virtual address of the mapping	
1024 * @exclusive:	the page is exclusively owned by the current process
1025 */
1026static void __page_set_anon_rmap(struct page *page,
1027	struct vm_area_struct *vma, unsigned long address, int exclusive)
1028{
1029	struct anon_vma *anon_vma = vma->anon_vma;
1030
1031	BUG_ON(!anon_vma);
1032
1033	if (PageAnon(page))
1034		return;
1035
1036	/*
1037	 * If the page isn't exclusively mapped into this vma,
1038	 * we must use the _oldest_ possible anon_vma for the
1039	 * page mapping!
1040	 */
1041	if (!exclusive)
1042		anon_vma = anon_vma->root;
1043
1044	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1045	page->mapping = (struct address_space *) anon_vma;
1046	page->index = linear_page_index(vma, address);
1047}
1048
1049/**
1050 * __page_check_anon_rmap - sanity check anonymous rmap addition
1051 * @page:	the page to add the mapping to
1052 * @vma:	the vm area in which the mapping is added
1053 * @address:	the user virtual address mapped
1054 */
1055static void __page_check_anon_rmap(struct page *page,
1056	struct vm_area_struct *vma, unsigned long address)
1057{
1058#ifdef CONFIG_DEBUG_VM
1059	/*
1060	 * The page's anon-rmap details (mapping and index) are guaranteed to
1061	 * be set up correctly at this point.
1062	 *
1063	 * We have exclusion against page_add_anon_rmap because the caller
1064	 * always holds the page locked, except if called from page_dup_rmap,
1065	 * in which case the page is already known to be setup.
1066	 *
1067	 * We have exclusion against page_add_new_anon_rmap because those pages
1068	 * are initially only visible via the pagetables, and the pte is locked
1069	 * over the call to page_add_new_anon_rmap.
1070	 */
1071	BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1072	BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
1073#endif
1074}
1075
1076/**
1077 * page_add_anon_rmap - add pte mapping to an anonymous page
1078 * @page:	the page to add the mapping to
1079 * @vma:	the vm area in which the mapping is added
1080 * @address:	the user virtual address mapped
1081 * @compound:	charge the page as compound or small page
1082 *
1083 * The caller needs to hold the pte lock, and the page must be locked in
1084 * the anon_vma case: to serialize mapping,index checking after setting,
1085 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1086 * (but PageKsm is never downgraded to PageAnon).
1087 */
1088void page_add_anon_rmap(struct page *page,
1089	struct vm_area_struct *vma, unsigned long address, bool compound)
1090{
1091	do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1092}
1093
1094/*
1095 * Special version of the above for do_swap_page, which often runs
1096 * into pages that are exclusively owned by the current process.
1097 * Everybody else should continue to use page_add_anon_rmap above.
1098 */
1099void do_page_add_anon_rmap(struct page *page,
1100	struct vm_area_struct *vma, unsigned long address, int flags)
1101{
1102	bool compound = flags & RMAP_COMPOUND;
1103	bool first;
1104
1105	if (compound) {
1106		atomic_t *mapcount;
1107		VM_BUG_ON_PAGE(!PageLocked(page), page);
1108		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1109		mapcount = compound_mapcount_ptr(page);
1110		first = atomic_inc_and_test(mapcount);
1111	} else {
1112		first = atomic_inc_and_test(&page->_mapcount);
1113	}
1114
1115	if (first) {
1116		int nr = compound ? hpage_nr_pages(page) : 1;
1117		/*
1118		 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1119		 * these counters are not modified in interrupt context, and
1120		 * pte lock(a spinlock) is held, which implies preemption
1121		 * disabled.
1122		 */
1123		if (compound)
1124			__inc_node_page_state(page, NR_ANON_THPS);
1125		__mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1126	}
1127	if (unlikely(PageKsm(page)))
1128		return;
1129
1130	VM_BUG_ON_PAGE(!PageLocked(page), page);
1131
1132	/* address might be in next vma when migration races vma_adjust */
1133	if (first)
1134		__page_set_anon_rmap(page, vma, address,
1135				flags & RMAP_EXCLUSIVE);
1136	else
1137		__page_check_anon_rmap(page, vma, address);
1138}
1139
1140/**
1141 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1142 * @page:	the page to add the mapping to
1143 * @vma:	the vm area in which the mapping is added
1144 * @address:	the user virtual address mapped
1145 * @compound:	charge the page as compound or small page
1146 *
1147 * Same as page_add_anon_rmap but must only be called on *new* pages.
1148 * This means the inc-and-test can be bypassed.
1149 * Page does not have to be locked.
1150 */
1151void page_add_new_anon_rmap(struct page *page,
1152	struct vm_area_struct *vma, unsigned long address, bool compound)
1153{
1154	int nr = compound ? hpage_nr_pages(page) : 1;
1155
1156	VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1157	__SetPageSwapBacked(page);
1158	if (compound) {
1159		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1160		/* increment count (starts at -1) */
1161		atomic_set(compound_mapcount_ptr(page), 0);
1162		__inc_node_page_state(page, NR_ANON_THPS);
1163	} else {
1164		/* Anon THP always mapped first with PMD */
1165		VM_BUG_ON_PAGE(PageTransCompound(page), page);
1166		/* increment count (starts at -1) */
1167		atomic_set(&page->_mapcount, 0);
1168	}
1169	__mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1170	__page_set_anon_rmap(page, vma, address, 1);
 
 
 
 
 
1171}
1172
1173/**
1174 * page_add_file_rmap - add pte mapping to a file page
1175 * @page: the page to add the mapping to
1176 * @compound: charge the page as compound or small page
1177 *
1178 * The caller needs to hold the pte lock.
1179 */
1180void page_add_file_rmap(struct page *page, bool compound)
1181{
1182	int i, nr = 1;
1183
1184	VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1185	lock_page_memcg(page);
1186	if (compound && PageTransHuge(page)) {
1187		for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1188			if (atomic_inc_and_test(&page[i]._mapcount))
1189				nr++;
1190		}
1191		if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1192			goto out;
1193		if (PageSwapBacked(page))
1194			__inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1195		else
1196			__inc_node_page_state(page, NR_FILE_PMDMAPPED);
1197	} else {
1198		if (PageTransCompound(page) && page_mapping(page)) {
1199			VM_WARN_ON_ONCE(!PageLocked(page));
1200
1201			SetPageDoubleMap(compound_head(page));
1202			if (PageMlocked(page))
1203				clear_page_mlock(compound_head(page));
1204		}
1205		if (!atomic_inc_and_test(&page->_mapcount))
1206			goto out;
1207	}
1208	__mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1209out:
1210	unlock_page_memcg(page);
1211}
1212
1213static void page_remove_file_rmap(struct page *page, bool compound)
1214{
1215	int i, nr = 1;
1216
1217	VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1218	lock_page_memcg(page);
1219
1220	/* Hugepages are not counted in NR_FILE_MAPPED for now. */
1221	if (unlikely(PageHuge(page))) {
1222		/* hugetlb pages are always mapped with pmds */
1223		atomic_dec(compound_mapcount_ptr(page));
1224		goto out;
1225	}
1226
1227	/* page still mapped by someone else? */
1228	if (compound && PageTransHuge(page)) {
1229		for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1230			if (atomic_add_negative(-1, &page[i]._mapcount))
1231				nr++;
1232		}
1233		if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1234			goto out;
1235		if (PageSwapBacked(page))
1236			__dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1237		else
1238			__dec_node_page_state(page, NR_FILE_PMDMAPPED);
1239	} else {
1240		if (!atomic_add_negative(-1, &page->_mapcount))
1241			goto out;
1242	}
1243
1244	/*
1245	 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1246	 * these counters are not modified in interrupt context, and
1247	 * pte lock(a spinlock) is held, which implies preemption disabled.
1248	 */
1249	__mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1250
1251	if (unlikely(PageMlocked(page)))
1252		clear_page_mlock(page);
1253out:
1254	unlock_page_memcg(page);
1255}
1256
1257static void page_remove_anon_compound_rmap(struct page *page)
1258{
1259	int i, nr;
1260
1261	if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1262		return;
1263
1264	/* Hugepages are not counted in NR_ANON_PAGES for now. */
1265	if (unlikely(PageHuge(page)))
1266		return;
1267
1268	if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1269		return;
1270
1271	__dec_node_page_state(page, NR_ANON_THPS);
1272
1273	if (TestClearPageDoubleMap(page)) {
1274		/*
1275		 * Subpages can be mapped with PTEs too. Check how many of
1276		 * themi are still mapped.
1277		 */
1278		for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1279			if (atomic_add_negative(-1, &page[i]._mapcount))
1280				nr++;
1281		}
1282	} else {
1283		nr = HPAGE_PMD_NR;
1284	}
1285
1286	if (unlikely(PageMlocked(page)))
1287		clear_page_mlock(page);
1288
1289	if (nr) {
1290		__mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
1291		deferred_split_huge_page(page);
1292	}
 
1293}
1294
1295/**
1296 * page_remove_rmap - take down pte mapping from a page
1297 * @page:	page to remove mapping from
1298 * @compound:	uncharge the page as compound or small page
1299 *
1300 * The caller needs to hold the pte lock.
1301 */
1302void page_remove_rmap(struct page *page, bool compound)
1303{
1304	if (!PageAnon(page))
1305		return page_remove_file_rmap(page, compound);
 
1306
1307	if (compound)
1308		return page_remove_anon_compound_rmap(page);
 
 
 
 
 
1309
1310	/* page still mapped by someone else? */
1311	if (!atomic_add_negative(-1, &page->_mapcount))
1312		return;
1313
1314	/*
1315	 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1316	 * these counters are not modified in interrupt context, and
1317	 * pte lock(a spinlock) is held, which implies preemption disabled.
1318	 */
1319	__dec_node_page_state(page, NR_ANON_MAPPED);
1320
 
 
 
 
 
 
 
 
 
 
 
 
1321	if (unlikely(PageMlocked(page)))
1322		clear_page_mlock(page);
1323
1324	if (PageTransCompound(page))
1325		deferred_split_huge_page(compound_head(page));
1326
1327	/*
1328	 * It would be tidy to reset the PageAnon mapping here,
1329	 * but that might overwrite a racing page_add_anon_rmap
1330	 * which increments mapcount after us but sets mapping
1331	 * before us: so leave the reset to free_unref_page,
1332	 * and remember that it's only reliable while mapped.
1333	 * Leaving it set also helps swapoff to reinstate ptes
1334	 * faster for those pages still in swapcache.
1335	 */
 
 
 
 
1336}
1337
1338/*
1339 * @arg: enum ttu_flags will be passed to this argument
1340 */
1341static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1342		     unsigned long address, void *arg)
1343{
1344	struct mm_struct *mm = vma->vm_mm;
1345	struct page_vma_mapped_walk pvmw = {
1346		.page = page,
1347		.vma = vma,
1348		.address = address,
1349	};
1350	pte_t pteval;
1351	struct page *subpage;
1352	bool ret = true;
1353	struct mmu_notifier_range range;
1354	enum ttu_flags flags = (enum ttu_flags)arg;
1355
1356	/* munlock has nothing to gain from examining un-locked vmas */
1357	if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1358		return true;
1359
1360	if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1361	    is_zone_device_page(page) && !is_device_private_page(page))
1362		return true;
1363
1364	if (flags & TTU_SPLIT_HUGE_PMD) {
1365		split_huge_pmd_address(vma, address,
1366				flags & TTU_SPLIT_FREEZE, page);
1367	}
1368
1369	/*
1370	 * For THP, we have to assume the worse case ie pmd for invalidation.
1371	 * For hugetlb, it could be much worse if we need to do pud
1372	 * invalidation in the case of pmd sharing.
1373	 *
1374	 * Note that the page can not be free in this function as call of
1375	 * try_to_unmap() must hold a reference on the page.
1376	 */
1377	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1378				address,
1379				min(vma->vm_end, address + page_size(page)));
1380	if (PageHuge(page)) {
1381		/*
1382		 * If sharing is possible, start and end will be adjusted
1383		 * accordingly.
1384		 */
1385		adjust_range_if_pmd_sharing_possible(vma, &range.start,
1386						     &range.end);
1387	}
1388	mmu_notifier_invalidate_range_start(&range);
1389
1390	while (page_vma_mapped_walk(&pvmw)) {
1391#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1392		/* PMD-mapped THP migration entry */
1393		if (!pvmw.pte && (flags & TTU_MIGRATION)) {
1394			VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
1395
1396			set_pmd_migration_entry(&pvmw, page);
1397			continue;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1398		}
1399#endif
1400
 
1401		/*
1402		 * If the page is mlock()d, we cannot swap it out.
1403		 * If it's recently referenced (perhaps page_referenced
1404		 * skipped over this mm) then we should reactivate it.
1405		 */
1406		if (!(flags & TTU_IGNORE_MLOCK)) {
1407			if (vma->vm_flags & VM_LOCKED) {
1408				/* PTE-mapped THP are never mlocked */
1409				if (!PageTransCompound(page)) {
1410					/*
1411					 * Holding pte lock, we do *not* need
1412					 * mmap_sem here
1413					 */
1414					mlock_vma_page(page);
1415				}
1416				ret = false;
1417				page_vma_mapped_walk_done(&pvmw);
1418				break;
 
 
 
 
1419			}
1420			if (flags & TTU_MUNLOCK)
1421				continue;
1422		}
1423
1424		/* Unexpected PMD-mapped THP? */
1425		VM_BUG_ON_PAGE(!pvmw.pte, page);
1426
1427		subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1428		address = pvmw.address;
1429
1430		if (PageHuge(page)) {
1431			if (huge_pmd_unshare(mm, &address, pvmw.pte)) {
1432				/*
1433				 * huge_pmd_unshare unmapped an entire PMD
1434				 * page.  There is no way of knowing exactly
1435				 * which PMDs may be cached for this mm, so
1436				 * we must flush them all.  start/end were
1437				 * already adjusted above to cover this range.
1438				 */
1439				flush_cache_range(vma, range.start, range.end);
1440				flush_tlb_range(vma, range.start, range.end);
1441				mmu_notifier_invalidate_range(mm, range.start,
1442							      range.end);
1443
1444				/*
1445				 * The ref count of the PMD page was dropped
1446				 * which is part of the way map counting
1447				 * is done for shared PMDs.  Return 'true'
1448				 * here.  When there is no other sharing,
1449				 * huge_pmd_unshare returns false and we will
1450				 * unmap the actual page and drop map count
1451				 * to zero.
1452				 */
1453				page_vma_mapped_walk_done(&pvmw);
1454				break;
1455			}
1456		}
1457
1458		if (IS_ENABLED(CONFIG_MIGRATION) &&
1459		    (flags & TTU_MIGRATION) &&
1460		    is_zone_device_page(page)) {
1461			swp_entry_t entry;
1462			pte_t swp_pte;
1463
1464			pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte);
1465
1466			/*
1467			 * Store the pfn of the page in a special migration
1468			 * pte. do_swap_page() will wait until the migration
1469			 * pte is removed and then restart fault handling.
1470			 */
1471			entry = make_migration_entry(page, 0);
1472			swp_pte = swp_entry_to_pte(entry);
1473			if (pte_soft_dirty(pteval))
1474				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1475			set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
1476			/*
1477			 * No need to invalidate here it will synchronize on
1478			 * against the special swap migration pte.
1479			 *
1480			 * The assignment to subpage above was computed from a
1481			 * swap PTE which results in an invalid pointer.
1482			 * Since only PAGE_SIZE pages can currently be
1483			 * migrated, just set it to page. This will need to be
1484			 * changed when hugepage migrations to device private
1485			 * memory are supported.
1486			 */
1487			subpage = page;
1488			goto discard;
1489		}
 
 
 
 
 
 
 
 
 
 
 
 
 
1490
1491		if (!(flags & TTU_IGNORE_ACCESS)) {
1492			if (ptep_clear_flush_young_notify(vma, address,
1493						pvmw.pte)) {
1494				ret = false;
1495				page_vma_mapped_walk_done(&pvmw);
1496				break;
1497			}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1498		}
 
 
 
 
1499
1500		/* Nuke the page table entry. */
1501		flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1502		if (should_defer_flush(mm, flags)) {
1503			/*
1504			 * We clear the PTE but do not flush so potentially
1505			 * a remote CPU could still be writing to the page.
1506			 * If the entry was previously clean then the
1507			 * architecture must guarantee that a clear->dirty
1508			 * transition on a cached TLB entry is written through
1509			 * and traps if the PTE is unmapped.
1510			 */
1511			pteval = ptep_get_and_clear(mm, address, pvmw.pte);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1512
1513			set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1514		} else {
1515			pteval = ptep_clear_flush(vma, address, pvmw.pte);
1516		}
 
 
1517
1518		/* Move the dirty bit to the page. Now the pte is gone. */
1519		if (pte_dirty(pteval))
1520			set_page_dirty(page);
1521
1522		/* Update high watermark before we lower rss */
1523		update_hiwater_rss(mm);
 
1524
1525		if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1526			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1527			if (PageHuge(page)) {
1528				hugetlb_count_sub(compound_nr(page), mm);
1529				set_huge_swap_pte_at(mm, address,
1530						     pvmw.pte, pteval,
1531						     vma_mmu_pagesize(vma));
1532			} else {
1533				dec_mm_counter(mm, mm_counter(page));
1534				set_pte_at(mm, address, pvmw.pte, pteval);
1535			}
1536
1537		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1538			/*
1539			 * The guest indicated that the page content is of no
1540			 * interest anymore. Simply discard the pte, vmscan
1541			 * will take care of the rest.
1542			 * A future reference will then fault in a new zero
1543			 * page. When userfaultfd is active, we must not drop
1544			 * this page though, as its main user (postcopy
1545			 * migration) will not expect userfaults on already
1546			 * copied pages.
1547			 */
1548			dec_mm_counter(mm, mm_counter(page));
1549			/* We have to invalidate as we cleared the pte */
1550			mmu_notifier_invalidate_range(mm, address,
1551						      address + PAGE_SIZE);
1552		} else if (IS_ENABLED(CONFIG_MIGRATION) &&
1553				(flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) {
1554			swp_entry_t entry;
1555			pte_t swp_pte;
1556
1557			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1558				set_pte_at(mm, address, pvmw.pte, pteval);
1559				ret = false;
1560				page_vma_mapped_walk_done(&pvmw);
1561				break;
1562			}
1563
1564			/*
1565			 * Store the pfn of the page in a special migration
1566			 * pte. do_swap_page() will wait until the migration
1567			 * pte is removed and then restart fault handling.
1568			 */
1569			entry = make_migration_entry(subpage,
1570					pte_write(pteval));
1571			swp_pte = swp_entry_to_pte(entry);
1572			if (pte_soft_dirty(pteval))
1573				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1574			set_pte_at(mm, address, pvmw.pte, swp_pte);
1575			/*
1576			 * No need to invalidate here it will synchronize on
1577			 * against the special swap migration pte.
1578			 */
1579		} else if (PageAnon(page)) {
1580			swp_entry_t entry = { .val = page_private(subpage) };
1581			pte_t swp_pte;
1582			/*
1583			 * Store the swap location in the pte.
1584			 * See handle_pte_fault() ...
1585			 */
1586			if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
1587				WARN_ON_ONCE(1);
1588				ret = false;
1589				/* We have to invalidate as we cleared the pte */
1590				mmu_notifier_invalidate_range(mm, address,
1591							address + PAGE_SIZE);
1592				page_vma_mapped_walk_done(&pvmw);
1593				break;
1594			}
1595
1596			/* MADV_FREE page check */
1597			if (!PageSwapBacked(page)) {
1598				if (!PageDirty(page)) {
1599					/* Invalidate as we cleared the pte */
1600					mmu_notifier_invalidate_range(mm,
1601						address, address + PAGE_SIZE);
1602					dec_mm_counter(mm, MM_ANONPAGES);
1603					goto discard;
1604				}
1605
 
 
 
 
 
 
1606				/*
1607				 * If the page was redirtied, it cannot be
1608				 * discarded. Remap the page to page table.
 
1609				 */
1610				set_pte_at(mm, address, pvmw.pte, pteval);
1611				SetPageSwapBacked(page);
1612				ret = false;
1613				page_vma_mapped_walk_done(&pvmw);
1614				break;
1615			}
 
 
 
 
 
 
 
 
 
1616
1617			if (swap_duplicate(entry) < 0) {
1618				set_pte_at(mm, address, pvmw.pte, pteval);
1619				ret = false;
1620				page_vma_mapped_walk_done(&pvmw);
1621				break;
1622			}
1623			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1624				set_pte_at(mm, address, pvmw.pte, pteval);
1625				ret = false;
1626				page_vma_mapped_walk_done(&pvmw);
1627				break;
1628			}
1629			if (list_empty(&mm->mmlist)) {
1630				spin_lock(&mmlist_lock);
1631				if (list_empty(&mm->mmlist))
1632					list_add(&mm->mmlist, &init_mm.mmlist);
1633				spin_unlock(&mmlist_lock);
1634			}
1635			dec_mm_counter(mm, MM_ANONPAGES);
1636			inc_mm_counter(mm, MM_SWAPENTS);
1637			swp_pte = swp_entry_to_pte(entry);
1638			if (pte_soft_dirty(pteval))
1639				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1640			set_pte_at(mm, address, pvmw.pte, swp_pte);
1641			/* Invalidate as we cleared the pte */
1642			mmu_notifier_invalidate_range(mm, address,
1643						      address + PAGE_SIZE);
1644		} else {
1645			/*
1646			 * This is a locked file-backed page, thus it cannot
1647			 * be removed from the page cache and replaced by a new
1648			 * page before mmu_notifier_invalidate_range_end, so no
1649			 * concurrent thread might update its page table to
1650			 * point at new page while a device still is using this
1651			 * page.
1652			 *
1653			 * See Documentation/vm/mmu_notifier.rst
1654			 */
1655			dec_mm_counter(mm, mm_counter_file(page));
1656		}
1657discard:
1658		/*
1659		 * No need to call mmu_notifier_invalidate_range() it has be
1660		 * done above for all cases requiring it to happen under page
1661		 * table lock before mmu_notifier_invalidate_range_end()
1662		 *
1663		 * See Documentation/vm/mmu_notifier.rst
1664		 */
1665		page_remove_rmap(subpage, PageHuge(page));
1666		put_page(page);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1667	}
1668
1669	mmu_notifier_invalidate_range_end(&range);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1670
1671	return ret;
1672}
1673
1674bool is_vma_temporary_stack(struct vm_area_struct *vma)
1675{
1676	int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1677
1678	if (!maybe_stack)
1679		return false;
1680
1681	if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1682						VM_STACK_INCOMPLETE_SETUP)
1683		return true;
1684
1685	return false;
1686}
1687
1688static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1689{
1690	return is_vma_temporary_stack(vma);
1691}
1692
1693static int page_mapcount_is_zero(struct page *page)
1694{
1695	return !total_mapcount(page);
1696}
1697
1698/**
1699 * try_to_unmap - try to remove all page table mappings to a page
1700 * @page: the page to get unmapped
1701 * @flags: action and flags
1702 *
1703 * Tries to remove all the page table entries which are mapping this
1704 * page, used in the pageout path.  Caller must hold the page lock.
 
1705 *
1706 * If unmap is successful, return true. Otherwise, false.
 
 
 
1707 */
1708bool try_to_unmap(struct page *page, enum ttu_flags flags)
1709{
 
1710	struct rmap_walk_control rwc = {
1711		.rmap_one = try_to_unmap_one,
1712		.arg = (void *)flags,
1713		.done = page_mapcount_is_zero,
 
1714		.anon_lock = page_lock_anon_vma_read,
1715	};
1716
 
 
1717	/*
1718	 * During exec, a temporary VMA is setup and later moved.
1719	 * The VMA is moved under the anon_vma lock but not the
1720	 * page tables leading to a race where migration cannot
1721	 * find the migration ptes. Rather than increasing the
1722	 * locking requirements of exec(), migration skips
1723	 * temporary VMAs until after exec() completes.
1724	 */
1725	if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))
1726	    && !PageKsm(page) && PageAnon(page))
1727		rwc.invalid_vma = invalid_migration_vma;
1728
1729	if (flags & TTU_RMAP_LOCKED)
1730		rmap_walk_locked(page, &rwc);
1731	else
1732		rmap_walk(page, &rwc);
1733
1734	return !page_mapcount(page) ? true : false;
 
 
1735}
1736
1737static int page_not_mapped(struct page *page)
1738{
1739	return !page_mapped(page);
1740};
1741
1742/**
1743 * try_to_munlock - try to munlock a page
1744 * @page: the page to be munlocked
1745 *
1746 * Called from munlock code.  Checks all of the VMAs mapping the page
1747 * to make sure nobody else has this page mlocked. The page will be
1748 * returned with PG_mlocked cleared if no other vmas have it mlocked.
 
 
 
 
 
 
 
1749 */
1750
1751void try_to_munlock(struct page *page)
1752{
 
1753	struct rmap_walk_control rwc = {
1754		.rmap_one = try_to_unmap_one,
1755		.arg = (void *)TTU_MUNLOCK,
1756		.done = page_not_mapped,
 
 
 
 
 
 
1757		.anon_lock = page_lock_anon_vma_read,
1758
1759	};
1760
1761	VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1762	VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
1763
1764	rmap_walk(page, &rwc);
 
1765}
1766
1767void __put_anon_vma(struct anon_vma *anon_vma)
1768{
1769	struct anon_vma *root = anon_vma->root;
1770
1771	anon_vma_free(anon_vma);
1772	if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1773		anon_vma_free(root);
1774}
1775
1776static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1777					struct rmap_walk_control *rwc)
1778{
1779	struct anon_vma *anon_vma;
1780
1781	if (rwc->anon_lock)
1782		return rwc->anon_lock(page);
1783
1784	/*
1785	 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1786	 * because that depends on page_mapped(); but not all its usages
1787	 * are holding mmap_sem. Users without mmap_sem are required to
1788	 * take a reference count to prevent the anon_vma disappearing
1789	 */
1790	anon_vma = page_anon_vma(page);
1791	if (!anon_vma)
1792		return NULL;
1793
1794	anon_vma_lock_read(anon_vma);
1795	return anon_vma;
1796}
1797
1798/*
1799 * rmap_walk_anon - do something to anonymous page using the object-based
1800 * rmap method
1801 * @page: the page to be handled
1802 * @rwc: control variable according to each walk type
1803 *
1804 * Find all the mappings of a page using the mapping pointer and the vma chains
1805 * contained in the anon_vma struct it points to.
1806 *
1807 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1808 * where the page was found will be held for write.  So, we won't recheck
1809 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1810 * LOCKED.
1811 */
1812static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1813		bool locked)
1814{
1815	struct anon_vma *anon_vma;
1816	pgoff_t pgoff_start, pgoff_end;
1817	struct anon_vma_chain *avc;
 
1818
1819	if (locked) {
1820		anon_vma = page_anon_vma(page);
1821		/* anon_vma disappear under us? */
1822		VM_BUG_ON_PAGE(!anon_vma, page);
1823	} else {
1824		anon_vma = rmap_walk_anon_lock(page, rwc);
1825	}
1826	if (!anon_vma)
1827		return;
1828
1829	pgoff_start = page_to_pgoff(page);
1830	pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1831	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1832			pgoff_start, pgoff_end) {
1833		struct vm_area_struct *vma = avc->vma;
1834		unsigned long address = vma_address(page, vma);
1835
1836		cond_resched();
1837
1838		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1839			continue;
1840
1841		if (!rwc->rmap_one(page, vma, address, rwc->arg))
 
1842			break;
1843		if (rwc->done && rwc->done(page))
1844			break;
1845	}
1846
1847	if (!locked)
1848		anon_vma_unlock_read(anon_vma);
1849}
1850
1851/*
1852 * rmap_walk_file - do something to file page using the object-based rmap method
1853 * @page: the page to be handled
1854 * @rwc: control variable according to each walk type
1855 *
1856 * Find all the mappings of a page using the mapping pointer and the vma chains
1857 * contained in the address_space struct it points to.
1858 *
1859 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1860 * where the page was found will be held for write.  So, we won't recheck
1861 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1862 * LOCKED.
1863 */
1864static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1865		bool locked)
1866{
1867	struct address_space *mapping = page_mapping(page);
1868	pgoff_t pgoff_start, pgoff_end;
1869	struct vm_area_struct *vma;
 
1870
1871	/*
1872	 * The page lock not only makes sure that page->mapping cannot
1873	 * suddenly be NULLified by truncation, it makes sure that the
1874	 * structure at mapping cannot be freed and reused yet,
1875	 * so we can safely take mapping->i_mmap_rwsem.
1876	 */
1877	VM_BUG_ON_PAGE(!PageLocked(page), page);
1878
1879	if (!mapping)
1880		return;
1881
1882	pgoff_start = page_to_pgoff(page);
1883	pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1884	if (!locked)
1885		i_mmap_lock_read(mapping);
1886	vma_interval_tree_foreach(vma, &mapping->i_mmap,
1887			pgoff_start, pgoff_end) {
1888		unsigned long address = vma_address(page, vma);
1889
1890		cond_resched();
1891
1892		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1893			continue;
1894
1895		if (!rwc->rmap_one(page, vma, address, rwc->arg))
 
1896			goto done;
1897		if (rwc->done && rwc->done(page))
1898			goto done;
1899	}
1900
 
 
 
 
 
 
 
 
1901done:
1902	if (!locked)
1903		i_mmap_unlock_read(mapping);
1904}
1905
1906void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1907{
1908	if (unlikely(PageKsm(page)))
1909		rmap_walk_ksm(page, rwc);
1910	else if (PageAnon(page))
1911		rmap_walk_anon(page, rwc, false);
1912	else
1913		rmap_walk_file(page, rwc, false);
1914}
1915
1916/* Like rmap_walk, but caller holds relevant rmap lock */
1917void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1918{
1919	/* no ksm support for now */
1920	VM_BUG_ON_PAGE(PageKsm(page), page);
1921	if (PageAnon(page))
1922		rmap_walk_anon(page, rwc, true);
1923	else
1924		rmap_walk_file(page, rwc, true);
1925}
1926
1927#ifdef CONFIG_HUGETLB_PAGE
1928/*
1929 * The following two functions are for anonymous (private mapped) hugepages.
1930 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1931 * and no lru code, because we handle hugepages differently from common pages.
1932 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1933void hugepage_add_anon_rmap(struct page *page,
1934			    struct vm_area_struct *vma, unsigned long address)
1935{
1936	struct anon_vma *anon_vma = vma->anon_vma;
1937	int first;
1938
1939	BUG_ON(!PageLocked(page));
1940	BUG_ON(!anon_vma);
1941	/* address might be in next vma when migration races vma_adjust */
1942	first = atomic_inc_and_test(compound_mapcount_ptr(page));
1943	if (first)
1944		__page_set_anon_rmap(page, vma, address, 0);
1945}
1946
1947void hugepage_add_new_anon_rmap(struct page *page,
1948			struct vm_area_struct *vma, unsigned long address)
1949{
1950	BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1951	atomic_set(compound_mapcount_ptr(page), 0);
1952	__page_set_anon_rmap(page, vma, address, 1);
1953}
1954#endif /* CONFIG_HUGETLB_PAGE */