1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *  mm/userfaultfd.c
   4 *
   5 *  Copyright (C) 2015  Red Hat, Inc.
   6 */
   7
   8#include <linux/mm.h>
   9#include <linux/sched/signal.h>
  10#include <linux/pagemap.h>
  11#include <linux/rmap.h>
  12#include <linux/swap.h>
  13#include <linux/swapops.h>
  14#include <linux/userfaultfd_k.h>
  15#include <linux/mmu_notifier.h>
  16#include <linux/hugetlb.h>
  17#include <linux/shmem_fs.h>
  18#include <asm/tlbflush.h>
  19#include <asm/tlb.h>
  20#include "internal.h"
  21
  22static __always_inline
  23bool validate_dst_vma(struct vm_area_struct *dst_vma, unsigned long dst_end)
 
 
  24{
  25	/* Make sure that the dst range is fully within dst_vma. */
  26	if (dst_end > dst_vma->vm_end)
  27		return false;
 
 
 
 
 
 
 
 
 
 
  28
  29	/*
  30	 * Check the vma is registered in uffd, this is required to
  31	 * enforce the VM_MAYWRITE check done at uffd registration
  32	 * time.
  33	 */
  34	if (!dst_vma->vm_userfaultfd_ctx.ctx)
  35		return false;
  36
  37	return true;
  38}
  39
  40static __always_inline
  41struct vm_area_struct *find_vma_and_prepare_anon(struct mm_struct *mm,
  42						 unsigned long addr)
  43{
  44	struct vm_area_struct *vma;
  45
  46	mmap_assert_locked(mm);
  47	vma = vma_lookup(mm, addr);
  48	if (!vma)
  49		vma = ERR_PTR(-ENOENT);
  50	else if (!(vma->vm_flags & VM_SHARED) &&
  51		 unlikely(anon_vma_prepare(vma)))
  52		vma = ERR_PTR(-ENOMEM);
  53
  54	return vma;
  55}
  56
  57#ifdef CONFIG_PER_VMA_LOCK
  58/*
  59 * lock_vma() - Lookup and lock vma corresponding to @address.
  60 * @mm: mm to search vma in.
  61 * @address: address that the vma should contain.
  62 *
  63 * Should be called without holding mmap_lock. vma should be unlocked after use
  64 * with unlock_vma().
  65 *
  66 * Return: A locked vma containing @address, -ENOENT if no vma is found, or
  67 * -ENOMEM if anon_vma couldn't be allocated.
  68 */
  69static struct vm_area_struct *lock_vma(struct mm_struct *mm,
  70				       unsigned long address)
  71{
  72	struct vm_area_struct *vma;
  73
  74	vma = lock_vma_under_rcu(mm, address);
  75	if (vma) {
  76		/*
  77		 * lock_vma_under_rcu() only checks anon_vma for private
  78		 * anonymous mappings. But we need to ensure it is assigned in
  79		 * private file-backed vmas as well.
  80		 */
  81		if (!(vma->vm_flags & VM_SHARED) && unlikely(!vma->anon_vma))
  82			vma_end_read(vma);
  83		else
  84			return vma;
  85	}
  86
  87	mmap_read_lock(mm);
  88	vma = find_vma_and_prepare_anon(mm, address);
  89	if (!IS_ERR(vma)) {
  90		/*
  91		 * We cannot use vma_start_read() as it may fail due to
  92		 * false locked (see comment in vma_start_read()). We
  93		 * can avoid that by directly locking vm_lock under
  94		 * mmap_lock, which guarantees that nobody can lock the
  95		 * vma for write (vma_start_write()) under us.
  96		 */
  97		down_read(&vma->vm_lock->lock);
  98	}
  99
 100	mmap_read_unlock(mm);
 101	return vma;
 102}
 103
 104static struct vm_area_struct *uffd_mfill_lock(struct mm_struct *dst_mm,
 105					      unsigned long dst_start,
 106					      unsigned long len)
 107{
 108	struct vm_area_struct *dst_vma;
 109
 110	dst_vma = lock_vma(dst_mm, dst_start);
 111	if (IS_ERR(dst_vma) || validate_dst_vma(dst_vma, dst_start + len))
 112		return dst_vma;
 113
 114	vma_end_read(dst_vma);
 115	return ERR_PTR(-ENOENT);
 116}
 117
 118static void uffd_mfill_unlock(struct vm_area_struct *vma)
 119{
 120	vma_end_read(vma);
 121}
 122
 123#else
 124
 125static struct vm_area_struct *uffd_mfill_lock(struct mm_struct *dst_mm,
 126					      unsigned long dst_start,
 127					      unsigned long len)
 128{
 129	struct vm_area_struct *dst_vma;
 130
 131	mmap_read_lock(dst_mm);
 132	dst_vma = find_vma_and_prepare_anon(dst_mm, dst_start);
 133	if (IS_ERR(dst_vma))
 134		goto out_unlock;
 135
 136	if (validate_dst_vma(dst_vma, dst_start + len))
 137		return dst_vma;
 138
 139	dst_vma = ERR_PTR(-ENOENT);
 140out_unlock:
 141	mmap_read_unlock(dst_mm);
 142	return dst_vma;
 143}
 144
 145static void uffd_mfill_unlock(struct vm_area_struct *vma)
 146{
 147	mmap_read_unlock(vma->vm_mm);
 148}
 149#endif
 150
 151/* Check if dst_addr is outside of file's size. Must be called with ptl held. */
 152static bool mfill_file_over_size(struct vm_area_struct *dst_vma,
 153				 unsigned long dst_addr)
 154{
 155	struct inode *inode;
 156	pgoff_t offset, max_off;
 157
 158	if (!dst_vma->vm_file)
 159		return false;
 160
 161	inode = dst_vma->vm_file->f_inode;
 162	offset = linear_page_index(dst_vma, dst_addr);
 163	max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 164	return offset >= max_off;
 165}
 166
 167/*
 168 * Install PTEs, to map dst_addr (within dst_vma) to page.
 169 *
 170 * This function handles both MCOPY_ATOMIC_NORMAL and _CONTINUE for both shmem
 171 * and anon, and for both shared and private VMAs.
 172 */
 173int mfill_atomic_install_pte(pmd_t *dst_pmd,
 174			     struct vm_area_struct *dst_vma,
 175			     unsigned long dst_addr, struct page *page,
 176			     bool newly_allocated, uffd_flags_t flags)
 177{
 178	int ret;
 179	struct mm_struct *dst_mm = dst_vma->vm_mm;
 180	pte_t _dst_pte, *dst_pte;
 181	bool writable = dst_vma->vm_flags & VM_WRITE;
 182	bool vm_shared = dst_vma->vm_flags & VM_SHARED;
 183	bool page_in_cache = page_mapping(page);
 184	spinlock_t *ptl;
 185	struct folio *folio;
 186
 187	_dst_pte = mk_pte(page, dst_vma->vm_page_prot);
 188	_dst_pte = pte_mkdirty(_dst_pte);
 189	if (page_in_cache && !vm_shared)
 190		writable = false;
 191	if (writable)
 192		_dst_pte = pte_mkwrite(_dst_pte, dst_vma);
 193	if (flags & MFILL_ATOMIC_WP)
 194		_dst_pte = pte_mkuffd_wp(_dst_pte);
 195
 196	ret = -EAGAIN;
 197	dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 198	if (!dst_pte)
 199		goto out;
 200
 201	if (mfill_file_over_size(dst_vma, dst_addr)) {
 202		ret = -EFAULT;
 203		goto out_unlock;
 204	}
 205
 206	ret = -EEXIST;
 207	/*
 208	 * We allow to overwrite a pte marker: consider when both MISSING|WP
 209	 * registered, we firstly wr-protect a none pte which has no page cache
 210	 * page backing it, then access the page.
 211	 */
 212	if (!pte_none_mostly(ptep_get(dst_pte)))
 213		goto out_unlock;
 214
 215	folio = page_folio(page);
 216	if (page_in_cache) {
 217		/* Usually, cache pages are already added to LRU */
 218		if (newly_allocated)
 219			folio_add_lru(folio);
 220		folio_add_file_rmap_pte(folio, page, dst_vma);
 221	} else {
 222		folio_add_new_anon_rmap(folio, dst_vma, dst_addr);
 223		folio_add_lru_vma(folio, dst_vma);
 224	}
 225
 226	/*
 227	 * Must happen after rmap, as mm_counter() checks mapping (via
 228	 * PageAnon()), which is set by __page_set_anon_rmap().
 229	 */
 230	inc_mm_counter(dst_mm, mm_counter(folio));
 231
 232	set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 233
 234	/* No need to invalidate - it was non-present before */
 235	update_mmu_cache(dst_vma, dst_addr, dst_pte);
 236	ret = 0;
 237out_unlock:
 238	pte_unmap_unlock(dst_pte, ptl);
 239out:
 240	return ret;
 241}
 242
 243static int mfill_atomic_pte_copy(pmd_t *dst_pmd,
 244				 struct vm_area_struct *dst_vma,
 245				 unsigned long dst_addr,
 246				 unsigned long src_addr,
 247				 uffd_flags_t flags,
 248				 struct folio **foliop)
 249{
 250	void *kaddr;
 251	int ret;
 252	struct folio *folio;
 
 
 253
 254	if (!*foliop) {
 255		ret = -ENOMEM;
 256		folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, dst_vma,
 257					dst_addr, false);
 258		if (!folio)
 259			goto out;
 260
 261		kaddr = kmap_local_folio(folio, 0);
 262		/*
 263		 * The read mmap_lock is held here.  Despite the
 264		 * mmap_lock being read recursive a deadlock is still
 265		 * possible if a writer has taken a lock.  For example:
 266		 *
 267		 * process A thread 1 takes read lock on own mmap_lock
 268		 * process A thread 2 calls mmap, blocks taking write lock
 269		 * process B thread 1 takes page fault, read lock on own mmap lock
 270		 * process B thread 2 calls mmap, blocks taking write lock
 271		 * process A thread 1 blocks taking read lock on process B
 272		 * process B thread 1 blocks taking read lock on process A
 273		 *
 274		 * Disable page faults to prevent potential deadlock
 275		 * and retry the copy outside the mmap_lock.
 276		 */
 277		pagefault_disable();
 278		ret = copy_from_user(kaddr, (const void __user *) src_addr,
 279				     PAGE_SIZE);
 280		pagefault_enable();
 281		kunmap_local(kaddr);
 282
 283		/* fallback to copy_from_user outside mmap_lock */
 284		if (unlikely(ret)) {
 285			ret = -ENOENT;
 286			*foliop = folio;
 287			/* don't free the page */
 288			goto out;
 289		}
 290
 291		flush_dcache_folio(folio);
 292	} else {
 293		folio = *foliop;
 294		*foliop = NULL;
 295	}
 296
 297	/*
 298	 * The memory barrier inside __folio_mark_uptodate makes sure that
 299	 * preceding stores to the page contents become visible before
 300	 * the set_pte_at() write.
 301	 */
 302	__folio_mark_uptodate(folio);
 303
 304	ret = -ENOMEM;
 305	if (mem_cgroup_charge(folio, dst_vma->vm_mm, GFP_KERNEL))
 306		goto out_release;
 307
 308	ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
 309				       &folio->page, true, flags);
 310	if (ret)
 311		goto out_release;
 312out:
 313	return ret;
 314out_release:
 315	folio_put(folio);
 316	goto out;
 317}
 318
 319static int mfill_atomic_pte_zeroed_folio(pmd_t *dst_pmd,
 320					 struct vm_area_struct *dst_vma,
 321					 unsigned long dst_addr)
 322{
 323	struct folio *folio;
 324	int ret = -ENOMEM;
 325
 326	folio = vma_alloc_zeroed_movable_folio(dst_vma, dst_addr);
 327	if (!folio)
 328		return ret;
 329
 330	if (mem_cgroup_charge(folio, dst_vma->vm_mm, GFP_KERNEL))
 331		goto out_put;
 332
 333	/*
 334	 * The memory barrier inside __folio_mark_uptodate makes sure that
 335	 * zeroing out the folio become visible before mapping the page
 336	 * using set_pte_at(). See do_anonymous_page().
 337	 */
 338	__folio_mark_uptodate(folio);
 339
 340	ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
 341				       &folio->page, true, 0);
 342	if (ret)
 343		goto out_put;
 344
 345	return 0;
 346out_put:
 347	folio_put(folio);
 348	return ret;
 349}
 350
 351static int mfill_atomic_pte_zeropage(pmd_t *dst_pmd,
 352				     struct vm_area_struct *dst_vma,
 353				     unsigned long dst_addr)
 354{
 355	pte_t _dst_pte, *dst_pte;
 356	spinlock_t *ptl;
 357	int ret;
 358
 359	if (mm_forbids_zeropage(dst_vma->vm_mm))
 360		return mfill_atomic_pte_zeroed_folio(dst_pmd, dst_vma, dst_addr);
 361
 362	_dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 363					 dst_vma->vm_page_prot));
 364	ret = -EAGAIN;
 365	dst_pte = pte_offset_map_lock(dst_vma->vm_mm, dst_pmd, dst_addr, &ptl);
 366	if (!dst_pte)
 367		goto out;
 368	if (mfill_file_over_size(dst_vma, dst_addr)) {
 369		ret = -EFAULT;
 370		goto out_unlock;
 
 371	}
 372	ret = -EEXIST;
 373	if (!pte_none(ptep_get(dst_pte)))
 374		goto out_unlock;
 375	set_pte_at(dst_vma->vm_mm, dst_addr, dst_pte, _dst_pte);
 376	/* No need to invalidate - it was non-present before */
 377	update_mmu_cache(dst_vma, dst_addr, dst_pte);
 378	ret = 0;
 379out_unlock:
 380	pte_unmap_unlock(dst_pte, ptl);
 381out:
 382	return ret;
 383}
 384
 385/* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
 386static int mfill_atomic_pte_continue(pmd_t *dst_pmd,
 387				     struct vm_area_struct *dst_vma,
 388				     unsigned long dst_addr,
 389				     uffd_flags_t flags)
 390{
 391	struct inode *inode = file_inode(dst_vma->vm_file);
 392	pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
 393	struct folio *folio;
 394	struct page *page;
 395	int ret;
 396
 397	ret = shmem_get_folio(inode, pgoff, &folio, SGP_NOALLOC);
 398	/* Our caller expects us to return -EFAULT if we failed to find folio */
 399	if (ret == -ENOENT)
 400		ret = -EFAULT;
 401	if (ret)
 402		goto out;
 403	if (!folio) {
 404		ret = -EFAULT;
 405		goto out;
 406	}
 407
 408	page = folio_file_page(folio, pgoff);
 409	if (PageHWPoison(page)) {
 410		ret = -EIO;
 411		goto out_release;
 412	}
 413
 414	ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
 415				       page, false, flags);
 416	if (ret)
 417		goto out_release;
 418
 419	folio_unlock(folio);
 420	ret = 0;
 421out:
 422	return ret;
 
 
 423out_release:
 424	folio_unlock(folio);
 425	folio_put(folio);
 426	goto out;
 427}
 428
 429/* Handles UFFDIO_POISON for all non-hugetlb VMAs. */
 430static int mfill_atomic_pte_poison(pmd_t *dst_pmd,
 431				   struct vm_area_struct *dst_vma,
 432				   unsigned long dst_addr,
 433				   uffd_flags_t flags)
 434{
 435	int ret;
 436	struct mm_struct *dst_mm = dst_vma->vm_mm;
 437	pte_t _dst_pte, *dst_pte;
 438	spinlock_t *ptl;
 
 
 
 439
 440	_dst_pte = make_pte_marker(PTE_MARKER_POISONED);
 441	ret = -EAGAIN;
 442	dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 443	if (!dst_pte)
 444		goto out;
 445
 446	if (mfill_file_over_size(dst_vma, dst_addr)) {
 
 447		ret = -EFAULT;
 448		goto out_unlock;
 
 449	}
 450
 451	ret = -EEXIST;
 452	/* Refuse to overwrite any PTE, even a PTE marker (e.g. UFFD WP). */
 453	if (!pte_none(ptep_get(dst_pte)))
 454		goto out_unlock;
 455
 456	set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 457
 458	/* No need to invalidate - it was non-present before */
 459	update_mmu_cache(dst_vma, dst_addr, dst_pte);
 460	ret = 0;
 461out_unlock:
 462	pte_unmap_unlock(dst_pte, ptl);
 463out:
 464	return ret;
 465}
 466
 467static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 468{
 469	pgd_t *pgd;
 470	p4d_t *p4d;
 471	pud_t *pud;
 472
 473	pgd = pgd_offset(mm, address);
 474	p4d = p4d_alloc(mm, pgd, address);
 475	if (!p4d)
 476		return NULL;
 477	pud = pud_alloc(mm, p4d, address);
 478	if (!pud)
 479		return NULL;
 480	/*
 481	 * Note that we didn't run this because the pmd was
 482	 * missing, the *pmd may be already established and in
 483	 * turn it may also be a trans_huge_pmd.
 484	 */
 485	return pmd_alloc(mm, pud, address);
 486}
 487
 488#ifdef CONFIG_HUGETLB_PAGE
 489/*
 490 * mfill_atomic processing for HUGETLB vmas.  Note that this routine is
 491 * called with either vma-lock or mmap_lock held, it will release the lock
 492 * before returning.
 493 */
 494static __always_inline ssize_t mfill_atomic_hugetlb(
 495					      struct userfaultfd_ctx *ctx,
 496					      struct vm_area_struct *dst_vma,
 497					      unsigned long dst_start,
 498					      unsigned long src_start,
 499					      unsigned long len,
 500					      uffd_flags_t flags)
 501{
 502	struct mm_struct *dst_mm = dst_vma->vm_mm;
 
 503	ssize_t err;
 504	pte_t *dst_pte;
 505	unsigned long src_addr, dst_addr;
 506	long copied;
 507	struct folio *folio;
 508	unsigned long vma_hpagesize;
 509	pgoff_t idx;
 510	u32 hash;
 511	struct address_space *mapping;
 512
 513	/*
 514	 * There is no default zero huge page for all huge page sizes as
 515	 * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 516	 * by THP.  Since we can not reliably insert a zero page, this
 517	 * feature is not supported.
 518	 */
 519	if (uffd_flags_mode_is(flags, MFILL_ATOMIC_ZEROPAGE)) {
 520		up_read(&ctx->map_changing_lock);
 521		uffd_mfill_unlock(dst_vma);
 522		return -EINVAL;
 523	}
 524
 525	src_addr = src_start;
 526	dst_addr = dst_start;
 527	copied = 0;
 528	folio = NULL;
 529	vma_hpagesize = vma_kernel_pagesize(dst_vma);
 530
 531	/*
 532	 * Validate alignment based on huge page size
 533	 */
 534	err = -EINVAL;
 535	if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 536		goto out_unlock;
 537
 538retry:
 539	/*
 540	 * On routine entry dst_vma is set.  If we had to drop mmap_lock and
 541	 * retry, dst_vma will be set to NULL and we must lookup again.
 542	 */
 543	if (!dst_vma) {
 544		dst_vma = uffd_mfill_lock(dst_mm, dst_start, len);
 545		if (IS_ERR(dst_vma)) {
 546			err = PTR_ERR(dst_vma);
 547			goto out;
 548		}
 549
 550		err = -ENOENT;
 551		if (!is_vm_hugetlb_page(dst_vma))
 552			goto out_unlock_vma;
 
 553
 554		err = -EINVAL;
 555		if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 556			goto out_unlock_vma;
 557
 558		/*
 559		 * If memory mappings are changing because of non-cooperative
 560		 * operation (e.g. mremap) running in parallel, bail out and
 561		 * request the user to retry later
 562		 */
 563		down_read(&ctx->map_changing_lock);
 564		err = -EAGAIN;
 565		if (atomic_read(&ctx->mmap_changing))
 
 566			goto out_unlock;
 567	}
 568
 569	while (src_addr < src_start + len) {
 
 
 570		BUG_ON(dst_addr >= dst_start + len);
 571
 572		/*
 573		 * Serialize via vma_lock and hugetlb_fault_mutex.
 574		 * vma_lock ensures the dst_pte remains valid even
 575		 * in the case of shared pmds.  fault mutex prevents
 576		 * races with other faulting threads.
 577		 */
 578		idx = linear_page_index(dst_vma, dst_addr);
 579		mapping = dst_vma->vm_file->f_mapping;
 
 
 580		hash = hugetlb_fault_mutex_hash(mapping, idx);
 581		mutex_lock(&hugetlb_fault_mutex_table[hash]);
 582		hugetlb_vma_lock_read(dst_vma);
 583
 584		err = -ENOMEM;
 585		dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
 586		if (!dst_pte) {
 587			hugetlb_vma_unlock_read(dst_vma);
 588			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 
 589			goto out_unlock;
 590		}
 591
 592		if (!uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE) &&
 593		    !huge_pte_none_mostly(huge_ptep_get(dst_pte))) {
 594			err = -EEXIST;
 595			hugetlb_vma_unlock_read(dst_vma);
 596			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 
 597			goto out_unlock;
 598		}
 599
 600		err = hugetlb_mfill_atomic_pte(dst_pte, dst_vma, dst_addr,
 601					       src_addr, flags, &folio);
 602
 603		hugetlb_vma_unlock_read(dst_vma);
 604		mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 
 
 605
 606		cond_resched();
 607
 608		if (unlikely(err == -ENOENT)) {
 609			up_read(&ctx->map_changing_lock);
 610			uffd_mfill_unlock(dst_vma);
 611			BUG_ON(!folio);
 612
 613			err = copy_folio_from_user(folio,
 614						   (const void __user *)src_addr, true);
 
 
 615			if (unlikely(err)) {
 616				err = -EFAULT;
 617				goto out;
 618			}
 
 619
 620			dst_vma = NULL;
 621			goto retry;
 622		} else
 623			BUG_ON(folio);
 624
 625		if (!err) {
 626			dst_addr += vma_hpagesize;
 627			src_addr += vma_hpagesize;
 628			copied += vma_hpagesize;
 629
 630			if (fatal_signal_pending(current))
 631				err = -EINTR;
 632		}
 633		if (err)
 634			break;
 635	}
 636
 637out_unlock:
 638	up_read(&ctx->map_changing_lock);
 639out_unlock_vma:
 640	uffd_mfill_unlock(dst_vma);
 641out:
 642	if (folio)
 643		folio_put(folio);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 644	BUG_ON(copied < 0);
 645	BUG_ON(err > 0);
 646	BUG_ON(!copied && !err);
 647	return copied ? copied : err;
 648}
 649#else /* !CONFIG_HUGETLB_PAGE */
 650/* fail at build time if gcc attempts to use this */
 651extern ssize_t mfill_atomic_hugetlb(struct userfaultfd_ctx *ctx,
 652				    struct vm_area_struct *dst_vma,
 653				    unsigned long dst_start,
 654				    unsigned long src_start,
 655				    unsigned long len,
 656				    uffd_flags_t flags);
 657#endif /* CONFIG_HUGETLB_PAGE */
 658
 659static __always_inline ssize_t mfill_atomic_pte(pmd_t *dst_pmd,
 
 660						struct vm_area_struct *dst_vma,
 661						unsigned long dst_addr,
 662						unsigned long src_addr,
 663						uffd_flags_t flags,
 664						struct folio **foliop)
 
 665{
 666	ssize_t err;
 667
 668	if (uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE)) {
 669		return mfill_atomic_pte_continue(dst_pmd, dst_vma,
 670						 dst_addr, flags);
 671	} else if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
 672		return mfill_atomic_pte_poison(dst_pmd, dst_vma,
 673					       dst_addr, flags);
 674	}
 675
 676	/*
 677	 * The normal page fault path for a shmem will invoke the
 678	 * fault, fill the hole in the file and COW it right away. The
 679	 * result generates plain anonymous memory. So when we are
 680	 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 681	 * generate anonymous memory directly without actually filling
 682	 * the hole. For the MAP_PRIVATE case the robustness check
 683	 * only happens in the pagetable (to verify it's still none)
 684	 * and not in the radix tree.
 685	 */
 686	if (!(dst_vma->vm_flags & VM_SHARED)) {
 687		if (uffd_flags_mode_is(flags, MFILL_ATOMIC_COPY))
 688			err = mfill_atomic_pte_copy(dst_pmd, dst_vma,
 689						    dst_addr, src_addr,
 690						    flags, foliop);
 691		else
 692			err = mfill_atomic_pte_zeropage(dst_pmd,
 693						 dst_vma, dst_addr);
 694	} else {
 695		err = shmem_mfill_atomic_pte(dst_pmd, dst_vma,
 696					     dst_addr, src_addr,
 697					     flags, foliop);
 
 
 
 
 
 698	}
 699
 700	return err;
 701}
 702
 703static __always_inline ssize_t mfill_atomic(struct userfaultfd_ctx *ctx,
 704					    unsigned long dst_start,
 705					    unsigned long src_start,
 706					    unsigned long len,
 707					    uffd_flags_t flags)
 
 
 708{
 709	struct mm_struct *dst_mm = ctx->mm;
 710	struct vm_area_struct *dst_vma;
 711	ssize_t err;
 712	pmd_t *dst_pmd;
 713	unsigned long src_addr, dst_addr;
 714	long copied;
 715	struct folio *folio;
 
 716
 717	/*
 718	 * Sanitize the command parameters:
 719	 */
 720	BUG_ON(dst_start & ~PAGE_MASK);
 721	BUG_ON(len & ~PAGE_MASK);
 722
 723	/* Does the address range wrap, or is the span zero-sized? */
 724	BUG_ON(src_start + len <= src_start);
 725	BUG_ON(dst_start + len <= dst_start);
 726
 727	src_addr = src_start;
 728	dst_addr = dst_start;
 729	copied = 0;
 730	folio = NULL;
 731retry:
 732	/*
 733	 * Make sure the vma is not shared, that the dst range is
 734	 * both valid and fully within a single existing vma.
 735	 */
 736	dst_vma = uffd_mfill_lock(dst_mm, dst_start, len);
 737	if (IS_ERR(dst_vma)) {
 738		err = PTR_ERR(dst_vma);
 739		goto out;
 740	}
 741
 742	/*
 743	 * If memory mappings are changing because of non-cooperative
 744	 * operation (e.g. mremap) running in parallel, bail out and
 745	 * request the user to retry later
 746	 */
 747	down_read(&ctx->map_changing_lock);
 748	err = -EAGAIN;
 749	if (atomic_read(&ctx->mmap_changing))
 
 
 
 
 
 
 
 
 
 750		goto out_unlock;
 751
 752	err = -EINVAL;
 753	/*
 754	 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 755	 * it will overwrite vm_ops, so vma_is_anonymous must return false.
 756	 */
 757	if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 758	    dst_vma->vm_flags & VM_SHARED))
 759		goto out_unlock;
 760
 761	/*
 762	 * validate 'mode' now that we know the dst_vma: don't allow
 763	 * a wrprotect copy if the userfaultfd didn't register as WP.
 764	 */
 765	if ((flags & MFILL_ATOMIC_WP) && !(dst_vma->vm_flags & VM_UFFD_WP))
 
 766		goto out_unlock;
 767
 768	/*
 769	 * If this is a HUGETLB vma, pass off to appropriate routine
 770	 */
 771	if (is_vm_hugetlb_page(dst_vma))
 772		return  mfill_atomic_hugetlb(ctx, dst_vma, dst_start,
 773					     src_start, len, flags);
 774
 775	if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 776		goto out_unlock;
 777	if (!vma_is_shmem(dst_vma) &&
 778	    uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE))
 
 
 
 
 
 
 
 779		goto out_unlock;
 780
 781	while (src_addr < src_start + len) {
 782		pmd_t dst_pmdval;
 783
 784		BUG_ON(dst_addr >= dst_start + len);
 785
 786		dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 787		if (unlikely(!dst_pmd)) {
 788			err = -ENOMEM;
 789			break;
 790		}
 791
 792		dst_pmdval = pmdp_get_lockless(dst_pmd);
 793		/*
 794		 * If the dst_pmd is mapped as THP don't
 795		 * override it and just be strict.
 796		 */
 797		if (unlikely(pmd_trans_huge(dst_pmdval))) {
 798			err = -EEXIST;
 799			break;
 800		}
 801		if (unlikely(pmd_none(dst_pmdval)) &&
 802		    unlikely(__pte_alloc(dst_mm, dst_pmd))) {
 803			err = -ENOMEM;
 804			break;
 805		}
 806		/* If an huge pmd materialized from under us fail */
 807		if (unlikely(pmd_trans_huge(*dst_pmd))) {
 808			err = -EFAULT;
 809			break;
 810		}
 811
 812		BUG_ON(pmd_none(*dst_pmd));
 813		BUG_ON(pmd_trans_huge(*dst_pmd));
 814
 815		err = mfill_atomic_pte(dst_pmd, dst_vma, dst_addr,
 816				       src_addr, flags, &folio);
 817		cond_resched();
 818
 819		if (unlikely(err == -ENOENT)) {
 820			void *kaddr;
 821
 822			up_read(&ctx->map_changing_lock);
 823			uffd_mfill_unlock(dst_vma);
 824			BUG_ON(!folio);
 825
 826			kaddr = kmap_local_folio(folio, 0);
 827			err = copy_from_user(kaddr,
 828					     (const void __user *) src_addr,
 829					     PAGE_SIZE);
 830			kunmap_local(kaddr);
 831			if (unlikely(err)) {
 832				err = -EFAULT;
 833				goto out;
 834			}
 835			flush_dcache_folio(folio);
 836			goto retry;
 837		} else
 838			BUG_ON(folio);
 839
 840		if (!err) {
 841			dst_addr += PAGE_SIZE;
 842			src_addr += PAGE_SIZE;
 843			copied += PAGE_SIZE;
 844
 845			if (fatal_signal_pending(current))
 846				err = -EINTR;
 847		}
 848		if (err)
 849			break;
 850	}
 851
 852out_unlock:
 853	up_read(&ctx->map_changing_lock);
 854	uffd_mfill_unlock(dst_vma);
 855out:
 856	if (folio)
 857		folio_put(folio);
 858	BUG_ON(copied < 0);
 859	BUG_ON(err > 0);
 860	BUG_ON(!copied && !err);
 861	return copied ? copied : err;
 862}
 863
 864ssize_t mfill_atomic_copy(struct userfaultfd_ctx *ctx, unsigned long dst_start,
 865			  unsigned long src_start, unsigned long len,
 866			  uffd_flags_t flags)
 867{
 868	return mfill_atomic(ctx, dst_start, src_start, len,
 869			    uffd_flags_set_mode(flags, MFILL_ATOMIC_COPY));
 870}
 871
 872ssize_t mfill_atomic_zeropage(struct userfaultfd_ctx *ctx,
 873			      unsigned long start,
 874			      unsigned long len)
 875{
 876	return mfill_atomic(ctx, start, 0, len,
 877			    uffd_flags_set_mode(0, MFILL_ATOMIC_ZEROPAGE));
 878}
 879
 880ssize_t mfill_atomic_continue(struct userfaultfd_ctx *ctx, unsigned long start,
 881			      unsigned long len, uffd_flags_t flags)
 882{
 883
 884	/*
 885	 * A caller might reasonably assume that UFFDIO_CONTINUE contains an
 886	 * smp_wmb() to ensure that any writes to the about-to-be-mapped page by
 887	 * the thread doing the UFFDIO_CONTINUE are guaranteed to be visible to
 888	 * subsequent loads from the page through the newly mapped address range.
 889	 */
 890	smp_wmb();
 891
 892	return mfill_atomic(ctx, start, 0, len,
 893			    uffd_flags_set_mode(flags, MFILL_ATOMIC_CONTINUE));
 894}
 895
 896ssize_t mfill_atomic_poison(struct userfaultfd_ctx *ctx, unsigned long start,
 897			    unsigned long len, uffd_flags_t flags)
 898{
 899	return mfill_atomic(ctx, start, 0, len,
 900			    uffd_flags_set_mode(flags, MFILL_ATOMIC_POISON));
 901}
 902
 903long uffd_wp_range(struct vm_area_struct *dst_vma,
 904		   unsigned long start, unsigned long len, bool enable_wp)
 905{
 906	unsigned int mm_cp_flags;
 907	struct mmu_gather tlb;
 908	long ret;
 909
 910	VM_WARN_ONCE(start < dst_vma->vm_start || start + len > dst_vma->vm_end,
 911			"The address range exceeds VMA boundary.\n");
 912	if (enable_wp)
 913		mm_cp_flags = MM_CP_UFFD_WP;
 914	else
 915		mm_cp_flags = MM_CP_UFFD_WP_RESOLVE;
 916
 917	/*
 918	 * vma->vm_page_prot already reflects that uffd-wp is enabled for this
 919	 * VMA (see userfaultfd_set_vm_flags()) and that all PTEs are supposed
 920	 * to be write-protected as default whenever protection changes.
 921	 * Try upgrading write permissions manually.
 922	 */
 923	if (!enable_wp && vma_wants_manual_pte_write_upgrade(dst_vma))
 924		mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
 925	tlb_gather_mmu(&tlb, dst_vma->vm_mm);
 926	ret = change_protection(&tlb, dst_vma, start, start + len, mm_cp_flags);
 927	tlb_finish_mmu(&tlb);
 928
 929	return ret;
 930}
 931
 932int mwriteprotect_range(struct userfaultfd_ctx *ctx, unsigned long start,
 933			unsigned long len, bool enable_wp)
 934{
 935	struct mm_struct *dst_mm = ctx->mm;
 936	unsigned long end = start + len;
 937	unsigned long _start, _end;
 938	struct vm_area_struct *dst_vma;
 939	unsigned long page_mask;
 940	long err;
 941	VMA_ITERATOR(vmi, dst_mm, start);
 942
 943	/*
 944	 * Sanitize the command parameters:
 945	 */
 946	BUG_ON(start & ~PAGE_MASK);
 947	BUG_ON(len & ~PAGE_MASK);
 948
 949	/* Does the address range wrap, or is the span zero-sized? */
 950	BUG_ON(start + len <= start);
 951
 952	mmap_read_lock(dst_mm);
 953
 954	/*
 955	 * If memory mappings are changing because of non-cooperative
 956	 * operation (e.g. mremap) running in parallel, bail out and
 957	 * request the user to retry later
 958	 */
 959	down_read(&ctx->map_changing_lock);
 960	err = -EAGAIN;
 961	if (atomic_read(&ctx->mmap_changing))
 962		goto out_unlock;
 963
 964	err = -ENOENT;
 965	for_each_vma_range(vmi, dst_vma, end) {
 966
 967		if (!userfaultfd_wp(dst_vma)) {
 968			err = -ENOENT;
 969			break;
 970		}
 971
 972		if (is_vm_hugetlb_page(dst_vma)) {
 973			err = -EINVAL;
 974			page_mask = vma_kernel_pagesize(dst_vma) - 1;
 975			if ((start & page_mask) || (len & page_mask))
 976				break;
 977		}
 978
 979		_start = max(dst_vma->vm_start, start);
 980		_end = min(dst_vma->vm_end, end);
 981
 982		err = uffd_wp_range(dst_vma, _start, _end - _start, enable_wp);
 983
 984		/* Return 0 on success, <0 on failures */
 985		if (err < 0)
 986			break;
 987		err = 0;
 988	}
 989out_unlock:
 990	up_read(&ctx->map_changing_lock);
 991	mmap_read_unlock(dst_mm);
 992	return err;
 993}
 994
 995
 996void double_pt_lock(spinlock_t *ptl1,
 997		    spinlock_t *ptl2)
 998	__acquires(ptl1)
 999	__acquires(ptl2)
1000{
1001	spinlock_t *ptl_tmp;
1002
1003	if (ptl1 > ptl2) {
1004		/* exchange ptl1 and ptl2 */
1005		ptl_tmp = ptl1;
1006		ptl1 = ptl2;
1007		ptl2 = ptl_tmp;
1008	}
1009	/* lock in virtual address order to avoid lock inversion */
1010	spin_lock(ptl1);
1011	if (ptl1 != ptl2)
1012		spin_lock_nested(ptl2, SINGLE_DEPTH_NESTING);
1013	else
1014		__acquire(ptl2);
1015}
1016
1017void double_pt_unlock(spinlock_t *ptl1,
1018		      spinlock_t *ptl2)
1019	__releases(ptl1)
1020	__releases(ptl2)
1021{
1022	spin_unlock(ptl1);
1023	if (ptl1 != ptl2)
1024		spin_unlock(ptl2);
1025	else
1026		__release(ptl2);
1027}
1028
1029
1030static int move_present_pte(struct mm_struct *mm,
1031			    struct vm_area_struct *dst_vma,
1032			    struct vm_area_struct *src_vma,
1033			    unsigned long dst_addr, unsigned long src_addr,
1034			    pte_t *dst_pte, pte_t *src_pte,
1035			    pte_t orig_dst_pte, pte_t orig_src_pte,
1036			    spinlock_t *dst_ptl, spinlock_t *src_ptl,
1037			    struct folio *src_folio)
1038{
1039	int err = 0;
1040
1041	double_pt_lock(dst_ptl, src_ptl);
1042
1043	if (!pte_same(ptep_get(src_pte), orig_src_pte) ||
1044	    !pte_same(ptep_get(dst_pte), orig_dst_pte)) {
1045		err = -EAGAIN;
1046		goto out;
1047	}
1048	if (folio_test_large(src_folio) ||
1049	    folio_maybe_dma_pinned(src_folio) ||
1050	    !PageAnonExclusive(&src_folio->page)) {
1051		err = -EBUSY;
1052		goto out;
1053	}
1054
1055	orig_src_pte = ptep_clear_flush(src_vma, src_addr, src_pte);
1056	/* Folio got pinned from under us. Put it back and fail the move. */
1057	if (folio_maybe_dma_pinned(src_folio)) {
1058		set_pte_at(mm, src_addr, src_pte, orig_src_pte);
1059		err = -EBUSY;
1060		goto out;
1061	}
1062
1063	folio_move_anon_rmap(src_folio, dst_vma);
1064	WRITE_ONCE(src_folio->index, linear_page_index(dst_vma, dst_addr));
1065
1066	orig_dst_pte = mk_pte(&src_folio->page, dst_vma->vm_page_prot);
1067	/* Follow mremap() behavior and treat the entry dirty after the move */
1068	orig_dst_pte = pte_mkwrite(pte_mkdirty(orig_dst_pte), dst_vma);
1069
1070	set_pte_at(mm, dst_addr, dst_pte, orig_dst_pte);
1071out:
1072	double_pt_unlock(dst_ptl, src_ptl);
1073	return err;
1074}
1075
1076static int move_swap_pte(struct mm_struct *mm,
1077			 unsigned long dst_addr, unsigned long src_addr,
1078			 pte_t *dst_pte, pte_t *src_pte,
1079			 pte_t orig_dst_pte, pte_t orig_src_pte,
1080			 spinlock_t *dst_ptl, spinlock_t *src_ptl)
1081{
1082	if (!pte_swp_exclusive(orig_src_pte))
1083		return -EBUSY;
1084
1085	double_pt_lock(dst_ptl, src_ptl);
1086
1087	if (!pte_same(ptep_get(src_pte), orig_src_pte) ||
1088	    !pte_same(ptep_get(dst_pte), orig_dst_pte)) {
1089		double_pt_unlock(dst_ptl, src_ptl);
1090		return -EAGAIN;
1091	}
1092
1093	orig_src_pte = ptep_get_and_clear(mm, src_addr, src_pte);
1094	set_pte_at(mm, dst_addr, dst_pte, orig_src_pte);
1095	double_pt_unlock(dst_ptl, src_ptl);
1096
1097	return 0;
1098}
1099
1100static int move_zeropage_pte(struct mm_struct *mm,
1101			     struct vm_area_struct *dst_vma,
1102			     struct vm_area_struct *src_vma,
1103			     unsigned long dst_addr, unsigned long src_addr,
1104			     pte_t *dst_pte, pte_t *src_pte,
1105			     pte_t orig_dst_pte, pte_t orig_src_pte,
1106			     spinlock_t *dst_ptl, spinlock_t *src_ptl)
1107{
1108	pte_t zero_pte;
1109
1110	double_pt_lock(dst_ptl, src_ptl);
1111	if (!pte_same(ptep_get(src_pte), orig_src_pte) ||
1112	    !pte_same(ptep_get(dst_pte), orig_dst_pte)) {
1113		double_pt_unlock(dst_ptl, src_ptl);
1114		return -EAGAIN;
1115	}
1116
1117	zero_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
1118					 dst_vma->vm_page_prot));
1119	ptep_clear_flush(src_vma, src_addr, src_pte);
1120	set_pte_at(mm, dst_addr, dst_pte, zero_pte);
1121	double_pt_unlock(dst_ptl, src_ptl);
1122
1123	return 0;
1124}
1125
1126
1127/*
1128 * The mmap_lock for reading is held by the caller. Just move the page
1129 * from src_pmd to dst_pmd if possible, and return true if succeeded
1130 * in moving the page.
1131 */
1132static int move_pages_pte(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd,
1133			  struct vm_area_struct *dst_vma,
1134			  struct vm_area_struct *src_vma,
1135			  unsigned long dst_addr, unsigned long src_addr,
1136			  __u64 mode)
1137{
1138	swp_entry_t entry;
1139	pte_t orig_src_pte, orig_dst_pte;
1140	pte_t src_folio_pte;
1141	spinlock_t *src_ptl, *dst_ptl;
1142	pte_t *src_pte = NULL;
1143	pte_t *dst_pte = NULL;
1144
1145	struct folio *src_folio = NULL;
1146	struct anon_vma *src_anon_vma = NULL;
1147	struct mmu_notifier_range range;
1148	int err = 0;
1149
1150	flush_cache_range(src_vma, src_addr, src_addr + PAGE_SIZE);
1151	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
1152				src_addr, src_addr + PAGE_SIZE);
1153	mmu_notifier_invalidate_range_start(&range);
1154retry:
1155	dst_pte = pte_offset_map_nolock(mm, dst_pmd, dst_addr, &dst_ptl);
1156
1157	/* Retry if a huge pmd materialized from under us */
1158	if (unlikely(!dst_pte)) {
1159		err = -EAGAIN;
1160		goto out;
1161	}
1162
1163	src_pte = pte_offset_map_nolock(mm, src_pmd, src_addr, &src_ptl);
1164
1165	/*
1166	 * We held the mmap_lock for reading so MADV_DONTNEED
1167	 * can zap transparent huge pages under us, or the
1168	 * transparent huge page fault can establish new
1169	 * transparent huge pages under us.
1170	 */
1171	if (unlikely(!src_pte)) {
1172		err = -EAGAIN;
1173		goto out;
1174	}
1175
1176	/* Sanity checks before the operation */
1177	if (WARN_ON_ONCE(pmd_none(*dst_pmd)) ||	WARN_ON_ONCE(pmd_none(*src_pmd)) ||
1178	    WARN_ON_ONCE(pmd_trans_huge(*dst_pmd)) || WARN_ON_ONCE(pmd_trans_huge(*src_pmd))) {
1179		err = -EINVAL;
1180		goto out;
1181	}
1182
1183	spin_lock(dst_ptl);
1184	orig_dst_pte = ptep_get(dst_pte);
1185	spin_unlock(dst_ptl);
1186	if (!pte_none(orig_dst_pte)) {
1187		err = -EEXIST;
1188		goto out;
1189	}
1190
1191	spin_lock(src_ptl);
1192	orig_src_pte = ptep_get(src_pte);
1193	spin_unlock(src_ptl);
1194	if (pte_none(orig_src_pte)) {
1195		if (!(mode & UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES))
1196			err = -ENOENT;
1197		else /* nothing to do to move a hole */
1198			err = 0;
1199		goto out;
1200	}
1201
1202	/* If PTE changed after we locked the folio them start over */
1203	if (src_folio && unlikely(!pte_same(src_folio_pte, orig_src_pte))) {
1204		err = -EAGAIN;
1205		goto out;
1206	}
1207
1208	if (pte_present(orig_src_pte)) {
1209		if (is_zero_pfn(pte_pfn(orig_src_pte))) {
1210			err = move_zeropage_pte(mm, dst_vma, src_vma,
1211					       dst_addr, src_addr, dst_pte, src_pte,
1212					       orig_dst_pte, orig_src_pte,
1213					       dst_ptl, src_ptl);
1214			goto out;
1215		}
1216
1217		/*
1218		 * Pin and lock both source folio and anon_vma. Since we are in
1219		 * RCU read section, we can't block, so on contention have to
1220		 * unmap the ptes, obtain the lock and retry.
1221		 */
1222		if (!src_folio) {
1223			struct folio *folio;
1224
1225			/*
1226			 * Pin the page while holding the lock to be sure the
1227			 * page isn't freed under us
1228			 */
1229			spin_lock(src_ptl);
1230			if (!pte_same(orig_src_pte, ptep_get(src_pte))) {
1231				spin_unlock(src_ptl);
1232				err = -EAGAIN;
1233				goto out;
1234			}
1235
1236			folio = vm_normal_folio(src_vma, src_addr, orig_src_pte);
1237			if (!folio || !PageAnonExclusive(&folio->page)) {
1238				spin_unlock(src_ptl);
1239				err = -EBUSY;
1240				goto out;
1241			}
1242
1243			folio_get(folio);
1244			src_folio = folio;
1245			src_folio_pte = orig_src_pte;
1246			spin_unlock(src_ptl);
1247
1248			if (!folio_trylock(src_folio)) {
1249				pte_unmap(&orig_src_pte);
1250				pte_unmap(&orig_dst_pte);
1251				src_pte = dst_pte = NULL;
1252				/* now we can block and wait */
1253				folio_lock(src_folio);
1254				goto retry;
1255			}
1256
1257			if (WARN_ON_ONCE(!folio_test_anon(src_folio))) {
1258				err = -EBUSY;
1259				goto out;
1260			}
1261		}
1262
1263		/* at this point we have src_folio locked */
1264		if (folio_test_large(src_folio)) {
1265			/* split_folio() can block */
1266			pte_unmap(&orig_src_pte);
1267			pte_unmap(&orig_dst_pte);
1268			src_pte = dst_pte = NULL;
1269			err = split_folio(src_folio);
1270			if (err)
1271				goto out;
1272			/* have to reacquire the folio after it got split */
1273			folio_unlock(src_folio);
1274			folio_put(src_folio);
1275			src_folio = NULL;
1276			goto retry;
1277		}
1278
1279		if (!src_anon_vma) {
1280			/*
1281			 * folio_referenced walks the anon_vma chain
1282			 * without the folio lock. Serialize against it with
1283			 * the anon_vma lock, the folio lock is not enough.
1284			 */
1285			src_anon_vma = folio_get_anon_vma(src_folio);
1286			if (!src_anon_vma) {
1287				/* page was unmapped from under us */
1288				err = -EAGAIN;
1289				goto out;
1290			}
1291			if (!anon_vma_trylock_write(src_anon_vma)) {
1292				pte_unmap(&orig_src_pte);
1293				pte_unmap(&orig_dst_pte);
1294				src_pte = dst_pte = NULL;
1295				/* now we can block and wait */
1296				anon_vma_lock_write(src_anon_vma);
1297				goto retry;
1298			}
1299		}
1300
1301		err = move_present_pte(mm,  dst_vma, src_vma,
1302				       dst_addr, src_addr, dst_pte, src_pte,
1303				       orig_dst_pte, orig_src_pte,
1304				       dst_ptl, src_ptl, src_folio);
1305	} else {
1306		entry = pte_to_swp_entry(orig_src_pte);
1307		if (non_swap_entry(entry)) {
1308			if (is_migration_entry(entry)) {
1309				pte_unmap(&orig_src_pte);
1310				pte_unmap(&orig_dst_pte);
1311				src_pte = dst_pte = NULL;
1312				migration_entry_wait(mm, src_pmd, src_addr);
1313				err = -EAGAIN;
1314			} else
1315				err = -EFAULT;
1316			goto out;
1317		}
1318
1319		err = move_swap_pte(mm, dst_addr, src_addr,
1320				    dst_pte, src_pte,
1321				    orig_dst_pte, orig_src_pte,
1322				    dst_ptl, src_ptl);
1323	}
1324
1325out:
1326	if (src_anon_vma) {
1327		anon_vma_unlock_write(src_anon_vma);
1328		put_anon_vma(src_anon_vma);
1329	}
1330	if (src_folio) {
1331		folio_unlock(src_folio);
1332		folio_put(src_folio);
1333	}
1334	if (dst_pte)
1335		pte_unmap(dst_pte);
1336	if (src_pte)
1337		pte_unmap(src_pte);
1338	mmu_notifier_invalidate_range_end(&range);
1339
1340	return err;
1341}
1342
1343#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1344static inline bool move_splits_huge_pmd(unsigned long dst_addr,
1345					unsigned long src_addr,
1346					unsigned long src_end)
1347{
1348	return (src_addr & ~HPAGE_PMD_MASK) || (dst_addr & ~HPAGE_PMD_MASK) ||
1349		src_end - src_addr < HPAGE_PMD_SIZE;
1350}
1351#else
1352static inline bool move_splits_huge_pmd(unsigned long dst_addr,
1353					unsigned long src_addr,
1354					unsigned long src_end)
1355{
1356	/* This is unreachable anyway, just to avoid warnings when HPAGE_PMD_SIZE==0 */
1357	return false;
1358}
1359#endif
1360
1361static inline bool vma_move_compatible(struct vm_area_struct *vma)
1362{
1363	return !(vma->vm_flags & (VM_PFNMAP | VM_IO |  VM_HUGETLB |
1364				  VM_MIXEDMAP | VM_SHADOW_STACK));
1365}
1366
1367static int validate_move_areas(struct userfaultfd_ctx *ctx,
1368			       struct vm_area_struct *src_vma,
1369			       struct vm_area_struct *dst_vma)
1370{
1371	/* Only allow moving if both have the same access and protection */
1372	if ((src_vma->vm_flags & VM_ACCESS_FLAGS) != (dst_vma->vm_flags & VM_ACCESS_FLAGS) ||
1373	    pgprot_val(src_vma->vm_page_prot) != pgprot_val(dst_vma->vm_page_prot))
1374		return -EINVAL;
1375
1376	/* Only allow moving if both are mlocked or both aren't */
1377	if ((src_vma->vm_flags & VM_LOCKED) != (dst_vma->vm_flags & VM_LOCKED))
1378		return -EINVAL;
1379
1380	/*
1381	 * For now, we keep it simple and only move between writable VMAs.
1382	 * Access flags are equal, therefore cheching only the source is enough.
1383	 */
1384	if (!(src_vma->vm_flags & VM_WRITE))
1385		return -EINVAL;
1386
1387	/* Check if vma flags indicate content which can be moved */
1388	if (!vma_move_compatible(src_vma) || !vma_move_compatible(dst_vma))
1389		return -EINVAL;
1390
1391	/* Ensure dst_vma is registered in uffd we are operating on */
1392	if (!dst_vma->vm_userfaultfd_ctx.ctx ||
1393	    dst_vma->vm_userfaultfd_ctx.ctx != ctx)
1394		return -EINVAL;
1395
1396	/* Only allow moving across anonymous vmas */
1397	if (!vma_is_anonymous(src_vma) || !vma_is_anonymous(dst_vma))
1398		return -EINVAL;
1399
1400	return 0;
1401}
1402
1403static __always_inline
1404int find_vmas_mm_locked(struct mm_struct *mm,
1405			unsigned long dst_start,
1406			unsigned long src_start,
1407			struct vm_area_struct **dst_vmap,
1408			struct vm_area_struct **src_vmap)
1409{
1410	struct vm_area_struct *vma;
1411
1412	mmap_assert_locked(mm);
1413	vma = find_vma_and_prepare_anon(mm, dst_start);
1414	if (IS_ERR(vma))
1415		return PTR_ERR(vma);
1416
1417	*dst_vmap = vma;
1418	/* Skip finding src_vma if src_start is in dst_vma */
1419	if (src_start >= vma->vm_start && src_start < vma->vm_end)
1420		goto out_success;
1421
1422	vma = vma_lookup(mm, src_start);
1423	if (!vma)
1424		return -ENOENT;
1425out_success:
1426	*src_vmap = vma;
1427	return 0;
1428}
1429
1430#ifdef CONFIG_PER_VMA_LOCK
1431static int uffd_move_lock(struct mm_struct *mm,
1432			  unsigned long dst_start,
1433			  unsigned long src_start,
1434			  struct vm_area_struct **dst_vmap,
1435			  struct vm_area_struct **src_vmap)
1436{
1437	struct vm_area_struct *vma;
1438	int err;
1439
1440	vma = lock_vma(mm, dst_start);
1441	if (IS_ERR(vma))
1442		return PTR_ERR(vma);
1443
1444	*dst_vmap = vma;
1445	/*
1446	 * Skip finding src_vma if src_start is in dst_vma. This also ensures
1447	 * that we don't lock the same vma twice.
1448	 */
1449	if (src_start >= vma->vm_start && src_start < vma->vm_end) {
1450		*src_vmap = vma;
1451		return 0;
1452	}
1453
1454	/*
1455	 * Using lock_vma() to get src_vma can lead to following deadlock:
1456	 *
1457	 * Thread1				Thread2
1458	 * -------				-------
1459	 * vma_start_read(dst_vma)
1460	 *					mmap_write_lock(mm)
1461	 *					vma_start_write(src_vma)
1462	 * vma_start_read(src_vma)
1463	 * mmap_read_lock(mm)
1464	 *					vma_start_write(dst_vma)
1465	 */
1466	*src_vmap = lock_vma_under_rcu(mm, src_start);
1467	if (likely(*src_vmap))
1468		return 0;
1469
1470	/* Undo any locking and retry in mmap_lock critical section */
1471	vma_end_read(*dst_vmap);
1472
1473	mmap_read_lock(mm);
1474	err = find_vmas_mm_locked(mm, dst_start, src_start, dst_vmap, src_vmap);
1475	if (!err) {
1476		/*
1477		 * See comment in lock_vma() as to why not using
1478		 * vma_start_read() here.
1479		 */
1480		down_read(&(*dst_vmap)->vm_lock->lock);
1481		if (*dst_vmap != *src_vmap)
1482			down_read_nested(&(*src_vmap)->vm_lock->lock,
1483					 SINGLE_DEPTH_NESTING);
1484	}
1485	mmap_read_unlock(mm);
1486	return err;
1487}
1488
1489static void uffd_move_unlock(struct vm_area_struct *dst_vma,
1490			     struct vm_area_struct *src_vma)
1491{
1492	vma_end_read(src_vma);
1493	if (src_vma != dst_vma)
1494		vma_end_read(dst_vma);
1495}
1496
1497#else
1498
1499static int uffd_move_lock(struct mm_struct *mm,
1500			  unsigned long dst_start,
1501			  unsigned long src_start,
1502			  struct vm_area_struct **dst_vmap,
1503			  struct vm_area_struct **src_vmap)
1504{
1505	int err;
1506
1507	mmap_read_lock(mm);
1508	err = find_vmas_mm_locked(mm, dst_start, src_start, dst_vmap, src_vmap);
1509	if (err)
1510		mmap_read_unlock(mm);
1511	return err;
1512}
1513
1514static void uffd_move_unlock(struct vm_area_struct *dst_vma,
1515			     struct vm_area_struct *src_vma)
1516{
1517	mmap_assert_locked(src_vma->vm_mm);
1518	mmap_read_unlock(dst_vma->vm_mm);
1519}
1520#endif
1521
1522/**
1523 * move_pages - move arbitrary anonymous pages of an existing vma
1524 * @ctx: pointer to the userfaultfd context
1525 * @dst_start: start of the destination virtual memory range
1526 * @src_start: start of the source virtual memory range
1527 * @len: length of the virtual memory range
1528 * @mode: flags from uffdio_move.mode
1529 *
1530 * It will either use the mmap_lock in read mode or per-vma locks
1531 *
1532 * move_pages() remaps arbitrary anonymous pages atomically in zero
1533 * copy. It only works on non shared anonymous pages because those can
1534 * be relocated without generating non linear anon_vmas in the rmap
1535 * code.
1536 *
1537 * It provides a zero copy mechanism to handle userspace page faults.
1538 * The source vma pages should have mapcount == 1, which can be
1539 * enforced by using madvise(MADV_DONTFORK) on src vma.
1540 *
1541 * The thread receiving the page during the userland page fault
1542 * will receive the faulting page in the source vma through the network,
1543 * storage or any other I/O device (MADV_DONTFORK in the source vma
1544 * avoids move_pages() to fail with -EBUSY if the process forks before
1545 * move_pages() is called), then it will call move_pages() to map the
1546 * page in the faulting address in the destination vma.
1547 *
1548 * This userfaultfd command works purely via pagetables, so it's the
1549 * most efficient way to move physical non shared anonymous pages
1550 * across different virtual addresses. Unlike mremap()/mmap()/munmap()
1551 * it does not create any new vmas. The mapping in the destination
1552 * address is atomic.
1553 *
1554 * It only works if the vma protection bits are identical from the
1555 * source and destination vma.
1556 *
1557 * It can remap non shared anonymous pages within the same vma too.
1558 *
1559 * If the source virtual memory range has any unmapped holes, or if
1560 * the destination virtual memory range is not a whole unmapped hole,
1561 * move_pages() will fail respectively with -ENOENT or -EEXIST. This
1562 * provides a very strict behavior to avoid any chance of memory
1563 * corruption going unnoticed if there are userland race conditions.
1564 * Only one thread should resolve the userland page fault at any given
1565 * time for any given faulting address. This means that if two threads
1566 * try to both call move_pages() on the same destination address at the
1567 * same time, the second thread will get an explicit error from this
1568 * command.
1569 *
1570 * The command retval will return "len" is successful. The command
1571 * however can be interrupted by fatal signals or errors. If
1572 * interrupted it will return the number of bytes successfully
1573 * remapped before the interruption if any, or the negative error if
1574 * none. It will never return zero. Either it will return an error or
1575 * an amount of bytes successfully moved. If the retval reports a
1576 * "short" remap, the move_pages() command should be repeated by
1577 * userland with src+retval, dst+reval, len-retval if it wants to know
1578 * about the error that interrupted it.
1579 *
1580 * The UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES flag can be specified to
1581 * prevent -ENOENT errors to materialize if there are holes in the
1582 * source virtual range that is being remapped. The holes will be
1583 * accounted as successfully remapped in the retval of the
1584 * command. This is mostly useful to remap hugepage naturally aligned
1585 * virtual regions without knowing if there are transparent hugepage
1586 * in the regions or not, but preventing the risk of having to split
1587 * the hugepmd during the remap.
1588 *
1589 * If there's any rmap walk that is taking the anon_vma locks without
1590 * first obtaining the folio lock (the only current instance is
1591 * folio_referenced), they will have to verify if the folio->mapping
1592 * has changed after taking the anon_vma lock. If it changed they
1593 * should release the lock and retry obtaining a new anon_vma, because
1594 * it means the anon_vma was changed by move_pages() before the lock
1595 * could be obtained. This is the only additional complexity added to
1596 * the rmap code to provide this anonymous page remapping functionality.
1597 */
1598ssize_t move_pages(struct userfaultfd_ctx *ctx, unsigned long dst_start,
1599		   unsigned long src_start, unsigned long len, __u64 mode)
1600{
1601	struct mm_struct *mm = ctx->mm;
1602	struct vm_area_struct *src_vma, *dst_vma;
1603	unsigned long src_addr, dst_addr;
1604	pmd_t *src_pmd, *dst_pmd;
1605	long err = -EINVAL;
1606	ssize_t moved = 0;
1607
1608	/* Sanitize the command parameters. */
1609	if (WARN_ON_ONCE(src_start & ~PAGE_MASK) ||
1610	    WARN_ON_ONCE(dst_start & ~PAGE_MASK) ||
1611	    WARN_ON_ONCE(len & ~PAGE_MASK))
1612		goto out;
1613
1614	/* Does the address range wrap, or is the span zero-sized? */
1615	if (WARN_ON_ONCE(src_start + len <= src_start) ||
1616	    WARN_ON_ONCE(dst_start + len <= dst_start))
1617		goto out;
1618
1619	err = uffd_move_lock(mm, dst_start, src_start, &dst_vma, &src_vma);
1620	if (err)
1621		goto out;
1622
1623	/* Re-check after taking map_changing_lock */
1624	err = -EAGAIN;
1625	down_read(&ctx->map_changing_lock);
1626	if (likely(atomic_read(&ctx->mmap_changing)))
1627		goto out_unlock;
1628	/*
1629	 * Make sure the vma is not shared, that the src and dst remap
1630	 * ranges are both valid and fully within a single existing
1631	 * vma.
1632	 */
1633	err = -EINVAL;
1634	if (src_vma->vm_flags & VM_SHARED)
1635		goto out_unlock;
1636	if (src_start + len > src_vma->vm_end)
1637		goto out_unlock;
1638
1639	if (dst_vma->vm_flags & VM_SHARED)
1640		goto out_unlock;
1641	if (dst_start + len > dst_vma->vm_end)
1642		goto out_unlock;
1643
1644	err = validate_move_areas(ctx, src_vma, dst_vma);
1645	if (err)
1646		goto out_unlock;
1647
1648	for (src_addr = src_start, dst_addr = dst_start;
1649	     src_addr < src_start + len;) {
1650		spinlock_t *ptl;
1651		pmd_t dst_pmdval;
1652		unsigned long step_size;
1653
1654		/*
1655		 * Below works because anonymous area would not have a
1656		 * transparent huge PUD. If file-backed support is added,
1657		 * that case would need to be handled here.
1658		 */
1659		src_pmd = mm_find_pmd(mm, src_addr);
1660		if (unlikely(!src_pmd)) {
1661			if (!(mode & UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES)) {
1662				err = -ENOENT;
1663				break;
1664			}
1665			src_pmd = mm_alloc_pmd(mm, src_addr);
1666			if (unlikely(!src_pmd)) {
1667				err = -ENOMEM;
1668				break;
1669			}
1670		}
1671		dst_pmd = mm_alloc_pmd(mm, dst_addr);
1672		if (unlikely(!dst_pmd)) {
1673			err = -ENOMEM;
1674			break;
1675		}
1676
1677		dst_pmdval = pmdp_get_lockless(dst_pmd);
1678		/*
1679		 * If the dst_pmd is mapped as THP don't override it and just
1680		 * be strict. If dst_pmd changes into TPH after this check, the
1681		 * move_pages_huge_pmd() will detect the change and retry
1682		 * while move_pages_pte() will detect the change and fail.
1683		 */
1684		if (unlikely(pmd_trans_huge(dst_pmdval))) {
1685			err = -EEXIST;
1686			break;
1687		}
1688
1689		ptl = pmd_trans_huge_lock(src_pmd, src_vma);
1690		if (ptl) {
1691			if (pmd_devmap(*src_pmd)) {
1692				spin_unlock(ptl);
1693				err = -ENOENT;
1694				break;
1695			}
1696
1697			/* Check if we can move the pmd without splitting it. */
1698			if (move_splits_huge_pmd(dst_addr, src_addr, src_start + len) ||
1699			    !pmd_none(dst_pmdval)) {
1700				struct folio *folio = pfn_folio(pmd_pfn(*src_pmd));
1701
1702				if (!folio || (!is_huge_zero_page(&folio->page) &&
1703					       !PageAnonExclusive(&folio->page))) {
1704					spin_unlock(ptl);
1705					err = -EBUSY;
1706					break;
1707				}
1708
1709				spin_unlock(ptl);
1710				split_huge_pmd(src_vma, src_pmd, src_addr);
1711				/* The folio will be split by move_pages_pte() */
1712				continue;
1713			}
1714
1715			err = move_pages_huge_pmd(mm, dst_pmd, src_pmd,
1716						  dst_pmdval, dst_vma, src_vma,
1717						  dst_addr, src_addr);
1718			step_size = HPAGE_PMD_SIZE;
1719		} else {
1720			if (pmd_none(*src_pmd)) {
1721				if (!(mode & UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES)) {
1722					err = -ENOENT;
1723					break;
1724				}
1725				if (unlikely(__pte_alloc(mm, src_pmd))) {
1726					err = -ENOMEM;
1727					break;
1728				}
1729			}
1730
1731			if (unlikely(pte_alloc(mm, dst_pmd))) {
1732				err = -ENOMEM;
1733				break;
1734			}
1735
1736			err = move_pages_pte(mm, dst_pmd, src_pmd,
1737					     dst_vma, src_vma,
1738					     dst_addr, src_addr, mode);
1739			step_size = PAGE_SIZE;
1740		}
1741
1742		cond_resched();
1743
1744		if (fatal_signal_pending(current)) {
1745			/* Do not override an error */
1746			if (!err || err == -EAGAIN)
1747				err = -EINTR;
1748			break;
1749		}
1750
1751		if (err) {
1752			if (err == -EAGAIN)
1753				continue;
1754			break;
1755		}
1756
1757		/* Proceed to the next page */
1758		dst_addr += step_size;
1759		src_addr += step_size;
1760		moved += step_size;
1761	}
1762
 
1763out_unlock:
1764	up_read(&ctx->map_changing_lock);
1765	uffd_move_unlock(dst_vma, src_vma);
1766out:
1767	VM_WARN_ON(moved < 0);
1768	VM_WARN_ON(err > 0);
1769	VM_WARN_ON(!moved && !err);
1770	return moved ? moved : err;
1771}