Linux Audio

Check our new training course

Loading...
v5.9
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *	linux/mm/madvise.c
   4 *
   5 * Copyright (C) 1999  Linus Torvalds
   6 * Copyright (C) 2002  Christoph Hellwig
   7 */
   8
   9#include <linux/mman.h>
  10#include <linux/pagemap.h>
  11#include <linux/syscalls.h>
  12#include <linux/mempolicy.h>
  13#include <linux/page-isolation.h>
  14#include <linux/page_idle.h>
  15#include <linux/userfaultfd_k.h>
  16#include <linux/hugetlb.h>
  17#include <linux/falloc.h>
  18#include <linux/fadvise.h>
  19#include <linux/sched.h>
 
 
 
 
  20#include <linux/ksm.h>
  21#include <linux/fs.h>
  22#include <linux/file.h>
  23#include <linux/blkdev.h>
  24#include <linux/backing-dev.h>
  25#include <linux/pagewalk.h>
  26#include <linux/swap.h>
  27#include <linux/swapops.h>
  28#include <linux/shmem_fs.h>
  29#include <linux/mmu_notifier.h>
  30#include <linux/sched/mm.h>
  31
  32#include <asm/tlb.h>
  33
  34#include "internal.h"
 
 
 
 
 
 
 
  35
  36struct madvise_walk_private {
  37	struct mmu_gather *tlb;
  38	bool pageout;
  39};
  40
  41/*
  42 * Any behaviour which results in changes to the vma->vm_flags needs to
  43 * take mmap_lock for writing. Others, which simply traverse vmas, need
  44 * to only take it for reading.
  45 */
  46static int madvise_need_mmap_write(int behavior)
  47{
  48	switch (behavior) {
  49	case MADV_REMOVE:
  50	case MADV_WILLNEED:
  51	case MADV_DONTNEED:
 
  52	case MADV_COLD:
  53	case MADV_PAGEOUT:
  54	case MADV_FREE:
 
 
 
 
 
  55		return 0;
  56	default:
  57		/* be safe, default to 1. list exceptions explicitly */
  58		return 1;
  59	}
  60}
  61
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  62/*
  63 * We can potentially split a vm area into separate
  64 * areas, each area with its own behavior.
 
 
  65 */
  66static long madvise_behavior(struct vm_area_struct *vma,
  67		     struct vm_area_struct **prev,
  68		     unsigned long start, unsigned long end, int behavior)
 
  69{
  70	struct mm_struct *mm = vma->vm_mm;
  71	int error = 0;
  72	pgoff_t pgoff;
  73	unsigned long new_flags = vma->vm_flags;
  74
  75	switch (behavior) {
  76	case MADV_NORMAL:
  77		new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  78		break;
  79	case MADV_SEQUENTIAL:
  80		new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  81		break;
  82	case MADV_RANDOM:
  83		new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  84		break;
  85	case MADV_DONTFORK:
  86		new_flags |= VM_DONTCOPY;
  87		break;
  88	case MADV_DOFORK:
  89		if (vma->vm_flags & VM_IO) {
  90			error = -EINVAL;
  91			goto out;
  92		}
  93		new_flags &= ~VM_DONTCOPY;
  94		break;
  95	case MADV_WIPEONFORK:
  96		/* MADV_WIPEONFORK is only supported on anonymous memory. */
  97		if (vma->vm_file || vma->vm_flags & VM_SHARED) {
  98			error = -EINVAL;
  99			goto out;
 100		}
 101		new_flags |= VM_WIPEONFORK;
 102		break;
 103	case MADV_KEEPONFORK:
 104		new_flags &= ~VM_WIPEONFORK;
 105		break;
 106	case MADV_DONTDUMP:
 107		new_flags |= VM_DONTDUMP;
 108		break;
 109	case MADV_DODUMP:
 110		if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) {
 111			error = -EINVAL;
 112			goto out;
 113		}
 114		new_flags &= ~VM_DONTDUMP;
 115		break;
 116	case MADV_MERGEABLE:
 117	case MADV_UNMERGEABLE:
 118		error = ksm_madvise(vma, start, end, behavior, &new_flags);
 119		if (error)
 120			goto out_convert_errno;
 121		break;
 122	case MADV_HUGEPAGE:
 123	case MADV_NOHUGEPAGE:
 124		error = hugepage_madvise(vma, &new_flags, behavior);
 125		if (error)
 126			goto out_convert_errno;
 127		break;
 128	}
 129
 130	if (new_flags == vma->vm_flags) {
 131		*prev = vma;
 132		goto out;
 133	}
 134
 135	pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 136	*prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 137			  vma->vm_file, pgoff, vma_policy(vma),
 138			  vma->vm_userfaultfd_ctx);
 139	if (*prev) {
 140		vma = *prev;
 141		goto success;
 142	}
 143
 144	*prev = vma;
 145
 146	if (start != vma->vm_start) {
 147		if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 148			error = -ENOMEM;
 149			goto out;
 150		}
 151		error = __split_vma(mm, vma, start, 1);
 152		if (error)
 153			goto out_convert_errno;
 154	}
 155
 156	if (end != vma->vm_end) {
 157		if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 158			error = -ENOMEM;
 159			goto out;
 160		}
 161		error = __split_vma(mm, vma, end, 0);
 162		if (error)
 163			goto out_convert_errno;
 164	}
 165
 166success:
 167	/*
 168	 * vm_flags is protected by the mmap_lock held in write mode.
 169	 */
 170	vma->vm_flags = new_flags;
 171
 172out_convert_errno:
 173	/*
 174	 * madvise() returns EAGAIN if kernel resources, such as
 175	 * slab, are temporarily unavailable.
 176	 */
 177	if (error == -ENOMEM)
 178		error = -EAGAIN;
 179out:
 180	return error;
 181}
 182
 183#ifdef CONFIG_SWAP
 184static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 185	unsigned long end, struct mm_walk *walk)
 186{
 187	pte_t *orig_pte;
 188	struct vm_area_struct *vma = walk->private;
 189	unsigned long index;
 190
 191	if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 192		return 0;
 193
 194	for (index = start; index != end; index += PAGE_SIZE) {
 195		pte_t pte;
 196		swp_entry_t entry;
 197		struct page *page;
 198		spinlock_t *ptl;
 199
 200		orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 201		pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 202		pte_unmap_unlock(orig_pte, ptl);
 
 
 203
 204		if (pte_present(pte) || pte_none(pte))
 
 205			continue;
 206		entry = pte_to_swp_entry(pte);
 207		if (unlikely(non_swap_entry(entry)))
 208			continue;
 209
 210		page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 211							vma, index, false);
 212		if (page)
 213			put_page(page);
 
 
 
 214	}
 215
 
 
 
 
 
 216	return 0;
 217}
 218
 219static const struct mm_walk_ops swapin_walk_ops = {
 220	.pmd_entry		= swapin_walk_pmd_entry,
 
 221};
 222
 223static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 224		unsigned long start, unsigned long end,
 225		struct address_space *mapping)
 226{
 227	pgoff_t index;
 228	struct page *page;
 229	swp_entry_t swap;
 230
 231	for (; start < end; start += PAGE_SIZE) {
 232		index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 233
 234		page = find_get_entry(mapping, index);
 235		if (!xa_is_value(page)) {
 236			if (page)
 237				put_page(page);
 
 
 
 
 238			continue;
 239		}
 240		swap = radix_to_swp_entry(page);
 241		page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 242							NULL, 0, false);
 243		if (page)
 244			put_page(page);
 245	}
 246
 247	lru_add_drain();	/* Push any new pages onto the LRU now */
 
 
 
 
 
 
 
 
 
 
 
 
 
 248}
 249#endif		/* CONFIG_SWAP */
 250
 251/*
 252 * Schedule all required I/O operations.  Do not wait for completion.
 253 */
 254static long madvise_willneed(struct vm_area_struct *vma,
 255			     struct vm_area_struct **prev,
 256			     unsigned long start, unsigned long end)
 257{
 
 258	struct file *file = vma->vm_file;
 259	loff_t offset;
 260
 261	*prev = vma;
 262#ifdef CONFIG_SWAP
 263	if (!file) {
 264		walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
 265		lru_add_drain(); /* Push any new pages onto the LRU now */
 266		return 0;
 267	}
 268
 269	if (shmem_mapping(file->f_mapping)) {
 270		force_shm_swapin_readahead(vma, start, end,
 271					file->f_mapping);
 272		return 0;
 273	}
 274#else
 275	if (!file)
 276		return -EBADF;
 277#endif
 278
 279	if (IS_DAX(file_inode(file))) {
 280		/* no bad return value, but ignore advice */
 281		return 0;
 282	}
 283
 284	/*
 285	 * Filesystem's fadvise may need to take various locks.  We need to
 286	 * explicitly grab a reference because the vma (and hence the
 287	 * vma's reference to the file) can go away as soon as we drop
 288	 * mmap_lock.
 289	 */
 290	*prev = NULL;	/* tell sys_madvise we drop mmap_lock */
 291	get_file(file);
 292	offset = (loff_t)(start - vma->vm_start)
 293			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 294	mmap_read_unlock(current->mm);
 295	vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
 296	fput(file);
 297	mmap_read_lock(current->mm);
 298	return 0;
 299}
 300
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 301static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
 302				unsigned long addr, unsigned long end,
 303				struct mm_walk *walk)
 304{
 305	struct madvise_walk_private *private = walk->private;
 306	struct mmu_gather *tlb = private->tlb;
 307	bool pageout = private->pageout;
 308	struct mm_struct *mm = tlb->mm;
 309	struct vm_area_struct *vma = walk->vma;
 310	pte_t *orig_pte, *pte, ptent;
 311	spinlock_t *ptl;
 312	struct page *page = NULL;
 313	LIST_HEAD(page_list);
 
 
 
 314
 315	if (fatal_signal_pending(current))
 316		return -EINTR;
 317
 
 
 
 318#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 319	if (pmd_trans_huge(*pmd)) {
 320		pmd_t orig_pmd;
 321		unsigned long next = pmd_addr_end(addr, end);
 322
 323		tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
 324		ptl = pmd_trans_huge_lock(pmd, vma);
 325		if (!ptl)
 326			return 0;
 327
 328		orig_pmd = *pmd;
 329		if (is_huge_zero_pmd(orig_pmd))
 330			goto huge_unlock;
 331
 332		if (unlikely(!pmd_present(orig_pmd))) {
 333			VM_BUG_ON(thp_migration_supported() &&
 334					!is_pmd_migration_entry(orig_pmd));
 335			goto huge_unlock;
 336		}
 337
 338		page = pmd_page(orig_pmd);
 339
 340		/* Do not interfere with other mappings of this page */
 341		if (page_mapcount(page) != 1)
 
 
 
 342			goto huge_unlock;
 343
 344		if (next - addr != HPAGE_PMD_SIZE) {
 345			int err;
 346
 347			get_page(page);
 348			spin_unlock(ptl);
 349			lock_page(page);
 350			err = split_huge_page(page);
 351			unlock_page(page);
 352			put_page(page);
 353			if (!err)
 354				goto regular_page;
 355			return 0;
 356		}
 357
 358		if (pmd_young(orig_pmd)) {
 359			pmdp_invalidate(vma, addr, pmd);
 360			orig_pmd = pmd_mkold(orig_pmd);
 361
 362			set_pmd_at(mm, addr, pmd, orig_pmd);
 363			tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
 364		}
 365
 366		ClearPageReferenced(page);
 367		test_and_clear_page_young(page);
 
 
 368		if (pageout) {
 369			if (!isolate_lru_page(page)) {
 370				if (PageUnevictable(page))
 371					putback_lru_page(page);
 372				else
 373					list_add(&page->lru, &page_list);
 374			}
 375		} else
 376			deactivate_page(page);
 377huge_unlock:
 378		spin_unlock(ptl);
 379		if (pageout)
 380			reclaim_pages(&page_list);
 381		return 0;
 382	}
 383
 384regular_page:
 385	if (pmd_trans_unstable(pmd))
 386		return 0;
 387#endif
 388	tlb_change_page_size(tlb, PAGE_SIZE);
 389	orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 
 
 
 390	flush_tlb_batched_pending(mm);
 391	arch_enter_lazy_mmu_mode();
 392	for (; addr < end; pte++, addr += PAGE_SIZE) {
 393		ptent = *pte;
 
 
 
 
 
 
 
 
 
 
 
 394
 395		if (pte_none(ptent))
 396			continue;
 397
 398		if (!pte_present(ptent))
 399			continue;
 400
 401		page = vm_normal_page(vma, addr, ptent);
 402		if (!page)
 403			continue;
 404
 405		/*
 406		 * Creating a THP page is expensive so split it only if we
 407		 * are sure it's worth. Split it if we are only owner.
 
 
 
 408		 */
 409		if (PageTransCompound(page)) {
 410			if (page_mapcount(page) != 1)
 411				break;
 412			get_page(page);
 413			if (!trylock_page(page)) {
 414				put_page(page);
 415				break;
 416			}
 417			pte_unmap_unlock(orig_pte, ptl);
 418			if (split_huge_page(page)) {
 419				unlock_page(page);
 420				put_page(page);
 421				pte_offset_map_lock(mm, pmd, addr, &ptl);
 422				break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 423			}
 424			unlock_page(page);
 425			put_page(page);
 426			pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 427			pte--;
 428			addr -= PAGE_SIZE;
 429			continue;
 430		}
 431
 432		/* Do not interfere with other mappings of this page */
 433		if (page_mapcount(page) != 1)
 
 
 
 
 
 
 434			continue;
 435
 436		VM_BUG_ON_PAGE(PageTransCompound(page), page);
 
 437
 438		if (pte_young(ptent)) {
 439			ptent = ptep_get_and_clear_full(mm, addr, pte,
 440							tlb->fullmm);
 441			ptent = pte_mkold(ptent);
 442			set_pte_at(mm, addr, pte, ptent);
 443			tlb_remove_tlb_entry(tlb, pte, addr);
 444		}
 445
 446		/*
 447		 * We are deactivating a page for accelerating reclaiming.
 448		 * VM couldn't reclaim the page unless we clear PG_young.
 449		 * As a side effect, it makes confuse idle-page tracking
 450		 * because they will miss recent referenced history.
 451		 */
 452		ClearPageReferenced(page);
 453		test_and_clear_page_young(page);
 
 
 454		if (pageout) {
 455			if (!isolate_lru_page(page)) {
 456				if (PageUnevictable(page))
 457					putback_lru_page(page);
 458				else
 459					list_add(&page->lru, &page_list);
 460			}
 461		} else
 462			deactivate_page(page);
 463	}
 464
 465	arch_leave_lazy_mmu_mode();
 466	pte_unmap_unlock(orig_pte, ptl);
 
 
 467	if (pageout)
 468		reclaim_pages(&page_list);
 469	cond_resched();
 470
 471	return 0;
 472}
 473
 474static const struct mm_walk_ops cold_walk_ops = {
 475	.pmd_entry = madvise_cold_or_pageout_pte_range,
 
 476};
 477
 478static void madvise_cold_page_range(struct mmu_gather *tlb,
 479			     struct vm_area_struct *vma,
 480			     unsigned long addr, unsigned long end)
 481{
 482	struct madvise_walk_private walk_private = {
 483		.pageout = false,
 484		.tlb = tlb,
 485	};
 486
 487	tlb_start_vma(tlb, vma);
 488	walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 489	tlb_end_vma(tlb, vma);
 490}
 491
 
 
 
 
 
 492static long madvise_cold(struct vm_area_struct *vma,
 493			struct vm_area_struct **prev,
 494			unsigned long start_addr, unsigned long end_addr)
 495{
 496	struct mm_struct *mm = vma->vm_mm;
 497	struct mmu_gather tlb;
 498
 499	*prev = vma;
 500	if (!can_madv_lru_vma(vma))
 501		return -EINVAL;
 502
 503	lru_add_drain();
 504	tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 505	madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 506	tlb_finish_mmu(&tlb, start_addr, end_addr);
 507
 508	return 0;
 509}
 510
 511static void madvise_pageout_page_range(struct mmu_gather *tlb,
 512			     struct vm_area_struct *vma,
 513			     unsigned long addr, unsigned long end)
 514{
 515	struct madvise_walk_private walk_private = {
 516		.pageout = true,
 517		.tlb = tlb,
 518	};
 519
 520	tlb_start_vma(tlb, vma);
 521	walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 522	tlb_end_vma(tlb, vma);
 523}
 524
 525static inline bool can_do_pageout(struct vm_area_struct *vma)
 526{
 527	if (vma_is_anonymous(vma))
 528		return true;
 529	if (!vma->vm_file)
 530		return false;
 531	/*
 532	 * paging out pagecache only for non-anonymous mappings that correspond
 533	 * to the files the calling process could (if tried) open for writing;
 534	 * otherwise we'd be including shared non-exclusive mappings, which
 535	 * opens a side channel.
 536	 */
 537	return inode_owner_or_capable(file_inode(vma->vm_file)) ||
 538		inode_permission(file_inode(vma->vm_file), MAY_WRITE) == 0;
 539}
 540
 541static long madvise_pageout(struct vm_area_struct *vma,
 542			struct vm_area_struct **prev,
 543			unsigned long start_addr, unsigned long end_addr)
 544{
 545	struct mm_struct *mm = vma->vm_mm;
 546	struct mmu_gather tlb;
 547
 548	*prev = vma;
 549	if (!can_madv_lru_vma(vma))
 550		return -EINVAL;
 551
 552	if (!can_do_pageout(vma))
 
 
 
 
 
 
 
 553		return 0;
 554
 555	lru_add_drain();
 556	tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 557	madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
 558	tlb_finish_mmu(&tlb, start_addr, end_addr);
 559
 560	return 0;
 561}
 562
 563static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 564				unsigned long end, struct mm_walk *walk)
 565
 566{
 
 567	struct mmu_gather *tlb = walk->private;
 568	struct mm_struct *mm = tlb->mm;
 569	struct vm_area_struct *vma = walk->vma;
 570	spinlock_t *ptl;
 571	pte_t *orig_pte, *pte, ptent;
 572	struct page *page;
 573	int nr_swap = 0;
 574	unsigned long next;
 
 575
 576	next = pmd_addr_end(addr, end);
 577	if (pmd_trans_huge(*pmd))
 578		if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 579			goto next;
 580
 581	if (pmd_trans_unstable(pmd))
 582		return 0;
 583
 584	tlb_change_page_size(tlb, PAGE_SIZE);
 585	orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 
 
 586	flush_tlb_batched_pending(mm);
 587	arch_enter_lazy_mmu_mode();
 588	for (; addr != end; pte++, addr += PAGE_SIZE) {
 589		ptent = *pte;
 
 590
 591		if (pte_none(ptent))
 592			continue;
 593		/*
 594		 * If the pte has swp_entry, just clear page table to
 595		 * prevent swap-in which is more expensive rather than
 596		 * (page allocation + zeroing).
 597		 */
 598		if (!pte_present(ptent)) {
 599			swp_entry_t entry;
 600
 601			entry = pte_to_swp_entry(ptent);
 602			if (non_swap_entry(entry))
 603				continue;
 604			nr_swap--;
 605			free_swap_and_cache(entry);
 606			pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 
 
 
 
 
 607			continue;
 608		}
 609
 610		page = vm_normal_page(vma, addr, ptent);
 611		if (!page)
 612			continue;
 613
 614		/*
 615		 * If pmd isn't transhuge but the page is THP and
 616		 * is owned by only this process, split it and
 617		 * deactivate all pages.
 
 
 618		 */
 619		if (PageTransCompound(page)) {
 620			if (page_mapcount(page) != 1)
 621				goto out;
 622			get_page(page);
 623			if (!trylock_page(page)) {
 624				put_page(page);
 625				goto out;
 626			}
 627			pte_unmap_unlock(orig_pte, ptl);
 628			if (split_huge_page(page)) {
 629				unlock_page(page);
 630				put_page(page);
 631				pte_offset_map_lock(mm, pmd, addr, &ptl);
 632				goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 633			}
 634			unlock_page(page);
 635			put_page(page);
 636			pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 637			pte--;
 638			addr -= PAGE_SIZE;
 639			continue;
 640		}
 641
 642		VM_BUG_ON_PAGE(PageTransCompound(page), page);
 
 
 
 
 643
 644		if (PageSwapCache(page) || PageDirty(page)) {
 645			if (!trylock_page(page))
 646				continue;
 647			/*
 648			 * If page is shared with others, we couldn't clear
 649			 * PG_dirty of the page.
 
 650			 */
 651			if (page_mapcount(page) != 1) {
 652				unlock_page(page);
 653				continue;
 654			}
 655
 656			if (PageSwapCache(page) && !try_to_free_swap(page)) {
 657				unlock_page(page);
 
 658				continue;
 659			}
 660
 661			ClearPageDirty(page);
 662			unlock_page(page);
 663		}
 664
 665		if (pte_young(ptent) || pte_dirty(ptent)) {
 666			/*
 667			 * Some of architecture(ex, PPC) don't update TLB
 668			 * with set_pte_at and tlb_remove_tlb_entry so for
 669			 * the portability, remap the pte with old|clean
 670			 * after pte clearing.
 671			 */
 672			ptent = ptep_get_and_clear_full(mm, addr, pte,
 673							tlb->fullmm);
 674
 675			ptent = pte_mkold(ptent);
 676			ptent = pte_mkclean(ptent);
 677			set_pte_at(mm, addr, pte, ptent);
 678			tlb_remove_tlb_entry(tlb, pte, addr);
 679		}
 680		mark_page_lazyfree(page);
 681	}
 682out:
 683	if (nr_swap) {
 684		if (current->mm == mm)
 685			sync_mm_rss(mm);
 686
 
 687		add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 
 
 
 688	}
 689	arch_leave_lazy_mmu_mode();
 690	pte_unmap_unlock(orig_pte, ptl);
 691	cond_resched();
 692next:
 693	return 0;
 694}
 695
 696static const struct mm_walk_ops madvise_free_walk_ops = {
 697	.pmd_entry		= madvise_free_pte_range,
 
 698};
 699
 700static int madvise_free_single_vma(struct vm_area_struct *vma,
 701			unsigned long start_addr, unsigned long end_addr)
 702{
 703	struct mm_struct *mm = vma->vm_mm;
 704	struct mmu_notifier_range range;
 705	struct mmu_gather tlb;
 706
 707	/* MADV_FREE works for only anon vma at the moment */
 708	if (!vma_is_anonymous(vma))
 709		return -EINVAL;
 710
 711	range.start = max(vma->vm_start, start_addr);
 712	if (range.start >= vma->vm_end)
 713		return -EINVAL;
 714	range.end = min(vma->vm_end, end_addr);
 715	if (range.end <= vma->vm_start)
 716		return -EINVAL;
 717	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
 718				range.start, range.end);
 719
 720	lru_add_drain();
 721	tlb_gather_mmu(&tlb, mm, range.start, range.end);
 722	update_hiwater_rss(mm);
 723
 724	mmu_notifier_invalidate_range_start(&range);
 725	tlb_start_vma(&tlb, vma);
 726	walk_page_range(vma->vm_mm, range.start, range.end,
 727			&madvise_free_walk_ops, &tlb);
 728	tlb_end_vma(&tlb, vma);
 729	mmu_notifier_invalidate_range_end(&range);
 730	tlb_finish_mmu(&tlb, range.start, range.end);
 731
 732	return 0;
 733}
 734
 735/*
 736 * Application no longer needs these pages.  If the pages are dirty,
 737 * it's OK to just throw them away.  The app will be more careful about
 738 * data it wants to keep.  Be sure to free swap resources too.  The
 739 * zap_page_range call sets things up for shrink_active_list to actually free
 740 * these pages later if no one else has touched them in the meantime,
 741 * although we could add these pages to a global reuse list for
 742 * shrink_active_list to pick up before reclaiming other pages.
 743 *
 744 * NB: This interface discards data rather than pushes it out to swap,
 745 * as some implementations do.  This has performance implications for
 746 * applications like large transactional databases which want to discard
 747 * pages in anonymous maps after committing to backing store the data
 748 * that was kept in them.  There is no reason to write this data out to
 749 * the swap area if the application is discarding it.
 750 *
 751 * An interface that causes the system to free clean pages and flush
 752 * dirty pages is already available as msync(MS_INVALIDATE).
 753 */
 754static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 755					unsigned long start, unsigned long end)
 756{
 757	zap_page_range(vma, start, end - start);
 758	return 0;
 759}
 760
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 761static long madvise_dontneed_free(struct vm_area_struct *vma,
 762				  struct vm_area_struct **prev,
 763				  unsigned long start, unsigned long end,
 764				  int behavior)
 765{
 
 
 766	*prev = vma;
 767	if (!can_madv_lru_vma(vma))
 768		return -EINVAL;
 769
 
 
 
 770	if (!userfaultfd_remove(vma, start, end)) {
 771		*prev = NULL; /* mmap_lock has been dropped, prev is stale */
 772
 773		mmap_read_lock(current->mm);
 774		vma = find_vma(current->mm, start);
 775		if (!vma)
 776			return -ENOMEM;
 777		if (start < vma->vm_start) {
 778			/*
 779			 * This "vma" under revalidation is the one
 780			 * with the lowest vma->vm_start where start
 781			 * is also < vma->vm_end. If start <
 782			 * vma->vm_start it means an hole materialized
 783			 * in the user address space within the
 784			 * virtual range passed to MADV_DONTNEED
 785			 * or MADV_FREE.
 786			 */
 787			return -ENOMEM;
 788		}
 789		if (!can_madv_lru_vma(vma))
 790			return -EINVAL;
 791		if (end > vma->vm_end) {
 792			/*
 793			 * Don't fail if end > vma->vm_end. If the old
 794			 * vma was splitted while the mmap_lock was
 795			 * released the effect of the concurrent
 796			 * operation may not cause madvise() to
 797			 * have an undefined result. There may be an
 798			 * adjacent next vma that we'll walk
 799			 * next. userfaultfd_remove() will generate an
 800			 * UFFD_EVENT_REMOVE repetition on the
 801			 * end-vma->vm_end range, but the manager can
 802			 * handle a repetition fine.
 803			 */
 804			end = vma->vm_end;
 805		}
 806		VM_WARN_ON(start >= end);
 
 
 
 
 
 
 
 
 
 807	}
 808
 809	if (behavior == MADV_DONTNEED)
 810		return madvise_dontneed_single_vma(vma, start, end);
 811	else if (behavior == MADV_FREE)
 812		return madvise_free_single_vma(vma, start, end);
 813	else
 814		return -EINVAL;
 815}
 816
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 817/*
 818 * Application wants to free up the pages and associated backing store.
 819 * This is effectively punching a hole into the middle of a file.
 820 */
 821static long madvise_remove(struct vm_area_struct *vma,
 822				struct vm_area_struct **prev,
 823				unsigned long start, unsigned long end)
 824{
 825	loff_t offset;
 826	int error;
 827	struct file *f;
 
 828
 829	*prev = NULL;	/* tell sys_madvise we drop mmap_lock */
 830
 831	if (vma->vm_flags & VM_LOCKED)
 832		return -EINVAL;
 833
 834	f = vma->vm_file;
 835
 836	if (!f || !f->f_mapping || !f->f_mapping->host) {
 837			return -EINVAL;
 838	}
 839
 840	if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 841		return -EACCES;
 842
 843	offset = (loff_t)(start - vma->vm_start)
 844			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 845
 846	/*
 847	 * Filesystem's fallocate may need to take i_mutex.  We need to
 848	 * explicitly grab a reference because the vma (and hence the
 849	 * vma's reference to the file) can go away as soon as we drop
 850	 * mmap_lock.
 851	 */
 852	get_file(f);
 853	if (userfaultfd_remove(vma, start, end)) {
 854		/* mmap_lock was not released by userfaultfd_remove() */
 855		mmap_read_unlock(current->mm);
 856	}
 857	error = vfs_fallocate(f,
 858				FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 859				offset, end - start);
 860	fput(f);
 861	mmap_read_lock(current->mm);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 862	return error;
 863}
 864
 865#ifdef CONFIG_MEMORY_FAILURE
 866/*
 867 * Error injection support for memory error handling.
 868 */
 869static int madvise_inject_error(int behavior,
 870		unsigned long start, unsigned long end)
 871{
 872	struct page *page;
 873	struct zone *zone;
 874	unsigned long size;
 875
 876	if (!capable(CAP_SYS_ADMIN))
 877		return -EPERM;
 878
 879
 880	for (; start < end; start += size) {
 881		unsigned long pfn;
 
 882		int ret;
 883
 884		ret = get_user_pages_fast(start, 1, 0, &page);
 885		if (ret != 1)
 886			return ret;
 887		pfn = page_to_pfn(page);
 888
 889		/*
 890		 * When soft offlining hugepages, after migrating the page
 891		 * we dissolve it, therefore in the second loop "page" will
 892		 * no longer be a compound page.
 893		 */
 894		size = page_size(compound_head(page));
 895
 896		if (PageHWPoison(page)) {
 897			put_page(page);
 898			continue;
 899		}
 900
 901		if (behavior == MADV_SOFT_OFFLINE) {
 902			pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
 903					pfn, start);
 904
 905			ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
 906			if (ret)
 907				return ret;
 908			continue;
 
 
 
 909		}
 910
 911		pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
 912				pfn, start);
 913
 914		/*
 915		 * Drop the page reference taken by get_user_pages_fast(). In
 916		 * the absence of MF_COUNT_INCREASED the memory_failure()
 917		 * routine is responsible for pinning the page to prevent it
 918		 * from being released back to the page allocator.
 919		 */
 920		put_page(page);
 921		ret = memory_failure(pfn, 0);
 922		if (ret)
 923			return ret;
 924	}
 925
 926	/* Ensure that all poisoned pages are removed from per-cpu lists */
 927	for_each_populated_zone(zone)
 928		drain_all_pages(zone);
 929
 930	return 0;
 931}
 932#endif
 933
 934static long
 935madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 936		unsigned long start, unsigned long end, int behavior)
 937{
 938	switch (behavior) {
 939	case MADV_REMOVE:
 940		return madvise_remove(vma, prev, start, end);
 941	case MADV_WILLNEED:
 942		return madvise_willneed(vma, prev, start, end);
 943	case MADV_COLD:
 944		return madvise_cold(vma, prev, start, end);
 945	case MADV_PAGEOUT:
 946		return madvise_pageout(vma, prev, start, end);
 947	case MADV_FREE:
 948	case MADV_DONTNEED:
 949		return madvise_dontneed_free(vma, prev, start, end, behavior);
 950	default:
 951		return madvise_behavior(vma, prev, start, end, behavior);
 952	}
 953}
 954
 955static bool
 956madvise_behavior_valid(int behavior)
 957{
 958	switch (behavior) {
 959	case MADV_DOFORK:
 960	case MADV_DONTFORK:
 961	case MADV_NORMAL:
 962	case MADV_SEQUENTIAL:
 963	case MADV_RANDOM:
 964	case MADV_REMOVE:
 965	case MADV_WILLNEED:
 966	case MADV_DONTNEED:
 
 967	case MADV_FREE:
 968	case MADV_COLD:
 969	case MADV_PAGEOUT:
 
 
 970#ifdef CONFIG_KSM
 971	case MADV_MERGEABLE:
 972	case MADV_UNMERGEABLE:
 973#endif
 974#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 975	case MADV_HUGEPAGE:
 976	case MADV_NOHUGEPAGE:
 
 977#endif
 978	case MADV_DONTDUMP:
 979	case MADV_DODUMP:
 980	case MADV_WIPEONFORK:
 981	case MADV_KEEPONFORK:
 
 
 982#ifdef CONFIG_MEMORY_FAILURE
 983	case MADV_SOFT_OFFLINE:
 984	case MADV_HWPOISON:
 985#endif
 986		return true;
 987
 988	default:
 989		return false;
 990	}
 991}
 992
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 993/*
 994 * The madvise(2) system call.
 995 *
 996 * Applications can use madvise() to advise the kernel how it should
 997 * handle paging I/O in this VM area.  The idea is to help the kernel
 998 * use appropriate read-ahead and caching techniques.  The information
 999 * provided is advisory only, and can be safely disregarded by the
1000 * kernel without affecting the correct operation of the application.
1001 *
1002 * behavior values:
1003 *  MADV_NORMAL - the default behavior is to read clusters.  This
1004 *		results in some read-ahead and read-behind.
1005 *  MADV_RANDOM - the system should read the minimum amount of data
1006 *		on any access, since it is unlikely that the appli-
1007 *		cation will need more than what it asks for.
1008 *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1009 *		once, so they can be aggressively read ahead, and
1010 *		can be freed soon after they are accessed.
1011 *  MADV_WILLNEED - the application is notifying the system to read
1012 *		some pages ahead.
1013 *  MADV_DONTNEED - the application is finished with the given range,
1014 *		so the kernel can free resources associated with it.
1015 *  MADV_FREE - the application marks pages in the given range as lazy free,
1016 *		where actual purges are postponed until memory pressure happens.
1017 *  MADV_REMOVE - the application wants to free up the given range of
1018 *		pages and associated backing store.
1019 *  MADV_DONTFORK - omit this area from child's address space when forking:
1020 *		typically, to avoid COWing pages pinned by get_user_pages().
1021 *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1022 *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1023 *              range after a fork.
1024 *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1025 *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1026 *		were corrupted by unrecoverable hardware memory failure.
1027 *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1028 *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1029 *		this area with pages of identical content from other such areas.
1030 *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1031 *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1032 *		huge pages in the future. Existing pages might be coalesced and
1033 *		new pages might be allocated as THP.
1034 *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1035 *		transparent huge pages so the existing pages will not be
1036 *		coalesced into THP and new pages will not be allocated as THP.
 
1037 *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1038 *		from being included in its core dump.
1039 *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
 
 
 
 
 
 
 
 
 
1040 *
1041 * return values:
1042 *  zero    - success
1043 *  -EINVAL - start + len < 0, start is not page-aligned,
1044 *		"behavior" is not a valid value, or application
1045 *		is attempting to release locked or shared pages,
1046 *		or the specified address range includes file, Huge TLB,
1047 *		MAP_SHARED or VMPFNMAP range.
1048 *  -ENOMEM - addresses in the specified range are not currently
1049 *		mapped, or are outside the AS of the process.
1050 *  -EIO    - an I/O error occurred while paging in data.
1051 *  -EBADF  - map exists, but area maps something that isn't a file.
1052 *  -EAGAIN - a kernel resource was temporarily unavailable.
 
1053 */
1054int do_madvise(unsigned long start, size_t len_in, int behavior)
1055{
1056	unsigned long end, tmp;
1057	struct vm_area_struct *vma, *prev;
1058	int unmapped_error = 0;
1059	int error = -EINVAL;
1060	int write;
1061	size_t len;
1062	struct blk_plug plug;
1063
1064	start = untagged_addr(start);
1065
1066	if (!madvise_behavior_valid(behavior))
1067		return error;
1068
1069	if (!PAGE_ALIGNED(start))
1070		return error;
1071	len = PAGE_ALIGN(len_in);
1072
1073	/* Check to see whether len was rounded up from small -ve to zero */
1074	if (len_in && !len)
1075		return error;
1076
1077	end = start + len;
1078	if (end < start)
1079		return error;
1080
1081	error = 0;
1082	if (end == start)
1083		return error;
1084
1085#ifdef CONFIG_MEMORY_FAILURE
1086	if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1087		return madvise_inject_error(behavior, start, start + len_in);
1088#endif
1089
1090	write = madvise_need_mmap_write(behavior);
1091	if (write) {
1092		if (mmap_write_lock_killable(current->mm))
1093			return -EINTR;
1094
1095		/*
1096		 * We may have stolen the mm from another process
1097		 * that is undergoing core dumping.
1098		 *
1099		 * Right now that's io_ring, in the future it may
1100		 * be remote process management and not "current"
1101		 * at all.
1102		 *
1103		 * We need to fix core dumping to not do this,
1104		 * but for now we have the mmget_still_valid()
1105		 * model.
1106		 */
1107		if (!mmget_still_valid(current->mm)) {
1108			mmap_write_unlock(current->mm);
1109			return -EINTR;
1110		}
1111	} else {
1112		mmap_read_lock(current->mm);
1113	}
1114
1115	/*
1116	 * If the interval [start,end) covers some unmapped address
1117	 * ranges, just ignore them, but return -ENOMEM at the end.
1118	 * - different from the way of handling in mlock etc.
1119	 */
1120	vma = find_vma_prev(current->mm, start, &prev);
1121	if (vma && start > vma->vm_start)
1122		prev = vma;
1123
1124	blk_start_plug(&plug);
1125	for (;;) {
1126		/* Still start < end. */
1127		error = -ENOMEM;
1128		if (!vma)
1129			goto out;
1130
1131		/* Here start < (end|vma->vm_end). */
1132		if (start < vma->vm_start) {
1133			unmapped_error = -ENOMEM;
1134			start = vma->vm_start;
1135			if (start >= end)
1136				goto out;
1137		}
1138
1139		/* Here vma->vm_start <= start < (end|vma->vm_end) */
1140		tmp = vma->vm_end;
1141		if (end < tmp)
1142			tmp = end;
1143
1144		/* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1145		error = madvise_vma(vma, &prev, start, tmp, behavior);
1146		if (error)
1147			goto out;
1148		start = tmp;
1149		if (prev && start < prev->vm_end)
1150			start = prev->vm_end;
1151		error = unmapped_error;
1152		if (start >= end)
1153			goto out;
1154		if (prev)
1155			vma = prev->vm_next;
1156		else	/* madvise_remove dropped mmap_lock */
1157			vma = find_vma(current->mm, start);
1158	}
1159out:
1160	blk_finish_plug(&plug);
 
1161	if (write)
1162		mmap_write_unlock(current->mm);
1163	else
1164		mmap_read_unlock(current->mm);
1165
1166	return error;
1167}
1168
1169SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1170{
1171	return do_madvise(start, len_in, behavior);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1172}
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *	linux/mm/madvise.c
   4 *
   5 * Copyright (C) 1999  Linus Torvalds
   6 * Copyright (C) 2002  Christoph Hellwig
   7 */
   8
   9#include <linux/mman.h>
  10#include <linux/pagemap.h>
  11#include <linux/syscalls.h>
  12#include <linux/mempolicy.h>
  13#include <linux/page-isolation.h>
  14#include <linux/page_idle.h>
  15#include <linux/userfaultfd_k.h>
  16#include <linux/hugetlb.h>
  17#include <linux/falloc.h>
  18#include <linux/fadvise.h>
  19#include <linux/sched.h>
  20#include <linux/sched/mm.h>
  21#include <linux/mm_inline.h>
  22#include <linux/string.h>
  23#include <linux/uio.h>
  24#include <linux/ksm.h>
  25#include <linux/fs.h>
  26#include <linux/file.h>
  27#include <linux/blkdev.h>
  28#include <linux/backing-dev.h>
  29#include <linux/pagewalk.h>
  30#include <linux/swap.h>
  31#include <linux/swapops.h>
  32#include <linux/shmem_fs.h>
  33#include <linux/mmu_notifier.h>
 
  34
  35#include <asm/tlb.h>
  36
  37#include "internal.h"
  38#include "swap.h"
  39
  40/*
  41 * Maximum number of attempts we make to install guard pages before we give up
  42 * and return -ERESTARTNOINTR to have userspace try again.
  43 */
  44#define MAX_MADVISE_GUARD_RETRIES 3
  45
  46struct madvise_walk_private {
  47	struct mmu_gather *tlb;
  48	bool pageout;
  49};
  50
  51/*
  52 * Any behaviour which results in changes to the vma->vm_flags needs to
  53 * take mmap_lock for writing. Others, which simply traverse vmas, need
  54 * to only take it for reading.
  55 */
  56static int madvise_need_mmap_write(int behavior)
  57{
  58	switch (behavior) {
  59	case MADV_REMOVE:
  60	case MADV_WILLNEED:
  61	case MADV_DONTNEED:
  62	case MADV_DONTNEED_LOCKED:
  63	case MADV_COLD:
  64	case MADV_PAGEOUT:
  65	case MADV_FREE:
  66	case MADV_POPULATE_READ:
  67	case MADV_POPULATE_WRITE:
  68	case MADV_COLLAPSE:
  69	case MADV_GUARD_INSTALL:
  70	case MADV_GUARD_REMOVE:
  71		return 0;
  72	default:
  73		/* be safe, default to 1. list exceptions explicitly */
  74		return 1;
  75	}
  76}
  77
  78#ifdef CONFIG_ANON_VMA_NAME
  79struct anon_vma_name *anon_vma_name_alloc(const char *name)
  80{
  81	struct anon_vma_name *anon_name;
  82	size_t count;
  83
  84	/* Add 1 for NUL terminator at the end of the anon_name->name */
  85	count = strlen(name) + 1;
  86	anon_name = kmalloc(struct_size(anon_name, name, count), GFP_KERNEL);
  87	if (anon_name) {
  88		kref_init(&anon_name->kref);
  89		memcpy(anon_name->name, name, count);
  90	}
  91
  92	return anon_name;
  93}
  94
  95void anon_vma_name_free(struct kref *kref)
  96{
  97	struct anon_vma_name *anon_name =
  98			container_of(kref, struct anon_vma_name, kref);
  99	kfree(anon_name);
 100}
 101
 102struct anon_vma_name *anon_vma_name(struct vm_area_struct *vma)
 103{
 104	mmap_assert_locked(vma->vm_mm);
 105
 106	return vma->anon_name;
 107}
 108
 109/* mmap_lock should be write-locked */
 110static int replace_anon_vma_name(struct vm_area_struct *vma,
 111				 struct anon_vma_name *anon_name)
 112{
 113	struct anon_vma_name *orig_name = anon_vma_name(vma);
 114
 115	if (!anon_name) {
 116		vma->anon_name = NULL;
 117		anon_vma_name_put(orig_name);
 118		return 0;
 119	}
 120
 121	if (anon_vma_name_eq(orig_name, anon_name))
 122		return 0;
 123
 124	vma->anon_name = anon_vma_name_reuse(anon_name);
 125	anon_vma_name_put(orig_name);
 126
 127	return 0;
 128}
 129#else /* CONFIG_ANON_VMA_NAME */
 130static int replace_anon_vma_name(struct vm_area_struct *vma,
 131				 struct anon_vma_name *anon_name)
 132{
 133	if (anon_name)
 134		return -EINVAL;
 135
 136	return 0;
 137}
 138#endif /* CONFIG_ANON_VMA_NAME */
 139/*
 140 * Update the vm_flags on region of a vma, splitting it or merging it as
 141 * necessary.  Must be called with mmap_lock held for writing;
 142 * Caller should ensure anon_name stability by raising its refcount even when
 143 * anon_name belongs to a valid vma because this function might free that vma.
 144 */
 145static int madvise_update_vma(struct vm_area_struct *vma,
 146			      struct vm_area_struct **prev, unsigned long start,
 147			      unsigned long end, unsigned long new_flags,
 148			      struct anon_vma_name *anon_name)
 149{
 150	struct mm_struct *mm = vma->vm_mm;
 151	int error;
 152	VMA_ITERATOR(vmi, mm, start);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 153
 154	if (new_flags == vma->vm_flags && anon_vma_name_eq(anon_vma_name(vma), anon_name)) {
 155		*prev = vma;
 156		return 0;
 157	}
 158
 159	vma = vma_modify_flags_name(&vmi, *prev, vma, start, end, new_flags,
 160				    anon_name);
 161	if (IS_ERR(vma))
 162		return PTR_ERR(vma);
 
 
 
 
 163
 164	*prev = vma;
 165
 166	/* vm_flags is protected by the mmap_lock held in write mode. */
 167	vma_start_write(vma);
 168	vm_flags_reset(vma, new_flags);
 169	if (!vma->vm_file || vma_is_anon_shmem(vma)) {
 170		error = replace_anon_vma_name(vma, anon_name);
 
 171		if (error)
 172			return error;
 173	}
 174
 175	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 176}
 177
 178#ifdef CONFIG_SWAP
 179static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 180		unsigned long end, struct mm_walk *walk)
 181{
 
 182	struct vm_area_struct *vma = walk->private;
 183	struct swap_iocb *splug = NULL;
 184	pte_t *ptep = NULL;
 185	spinlock_t *ptl;
 186	unsigned long addr;
 187
 188	for (addr = start; addr < end; addr += PAGE_SIZE) {
 189		pte_t pte;
 190		swp_entry_t entry;
 191		struct folio *folio;
 
 192
 193		if (!ptep++) {
 194			ptep = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 195			if (!ptep)
 196				break;
 197		}
 198
 199		pte = ptep_get(ptep);
 200		if (!is_swap_pte(pte))
 201			continue;
 202		entry = pte_to_swp_entry(pte);
 203		if (unlikely(non_swap_entry(entry)))
 204			continue;
 205
 206		pte_unmap_unlock(ptep, ptl);
 207		ptep = NULL;
 208
 209		folio = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 210					     vma, addr, &splug);
 211		if (folio)
 212			folio_put(folio);
 213	}
 214
 215	if (ptep)
 216		pte_unmap_unlock(ptep, ptl);
 217	swap_read_unplug(splug);
 218	cond_resched();
 219
 220	return 0;
 221}
 222
 223static const struct mm_walk_ops swapin_walk_ops = {
 224	.pmd_entry		= swapin_walk_pmd_entry,
 225	.walk_lock		= PGWALK_RDLOCK,
 226};
 227
 228static void shmem_swapin_range(struct vm_area_struct *vma,
 229		unsigned long start, unsigned long end,
 230		struct address_space *mapping)
 231{
 232	XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
 233	pgoff_t end_index = linear_page_index(vma, end) - 1;
 234	struct folio *folio;
 235	struct swap_iocb *splug = NULL;
 236
 237	rcu_read_lock();
 238	xas_for_each(&xas, folio, end_index) {
 239		unsigned long addr;
 240		swp_entry_t entry;
 241
 242		if (!xa_is_value(folio))
 243			continue;
 244		entry = radix_to_swp_entry(folio);
 245		/* There might be swapin error entries in shmem mapping. */
 246		if (non_swap_entry(entry))
 247			continue;
 
 
 
 
 
 
 
 248
 249		addr = vma->vm_start +
 250			((xas.xa_index - vma->vm_pgoff) << PAGE_SHIFT);
 251		xas_pause(&xas);
 252		rcu_read_unlock();
 253
 254		folio = read_swap_cache_async(entry, mapping_gfp_mask(mapping),
 255					     vma, addr, &splug);
 256		if (folio)
 257			folio_put(folio);
 258
 259		rcu_read_lock();
 260	}
 261	rcu_read_unlock();
 262	swap_read_unplug(splug);
 263}
 264#endif		/* CONFIG_SWAP */
 265
 266/*
 267 * Schedule all required I/O operations.  Do not wait for completion.
 268 */
 269static long madvise_willneed(struct vm_area_struct *vma,
 270			     struct vm_area_struct **prev,
 271			     unsigned long start, unsigned long end)
 272{
 273	struct mm_struct *mm = vma->vm_mm;
 274	struct file *file = vma->vm_file;
 275	loff_t offset;
 276
 277	*prev = vma;
 278#ifdef CONFIG_SWAP
 279	if (!file) {
 280		walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
 281		lru_add_drain(); /* Push any new pages onto the LRU now */
 282		return 0;
 283	}
 284
 285	if (shmem_mapping(file->f_mapping)) {
 286		shmem_swapin_range(vma, start, end, file->f_mapping);
 287		lru_add_drain(); /* Push any new pages onto the LRU now */
 288		return 0;
 289	}
 290#else
 291	if (!file)
 292		return -EBADF;
 293#endif
 294
 295	if (IS_DAX(file_inode(file))) {
 296		/* no bad return value, but ignore advice */
 297		return 0;
 298	}
 299
 300	/*
 301	 * Filesystem's fadvise may need to take various locks.  We need to
 302	 * explicitly grab a reference because the vma (and hence the
 303	 * vma's reference to the file) can go away as soon as we drop
 304	 * mmap_lock.
 305	 */
 306	*prev = NULL;	/* tell sys_madvise we drop mmap_lock */
 307	get_file(file);
 308	offset = (loff_t)(start - vma->vm_start)
 309			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 310	mmap_read_unlock(mm);
 311	vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
 312	fput(file);
 313	mmap_read_lock(mm);
 314	return 0;
 315}
 316
 317static inline bool can_do_file_pageout(struct vm_area_struct *vma)
 318{
 319	if (!vma->vm_file)
 320		return false;
 321	/*
 322	 * paging out pagecache only for non-anonymous mappings that correspond
 323	 * to the files the calling process could (if tried) open for writing;
 324	 * otherwise we'd be including shared non-exclusive mappings, which
 325	 * opens a side channel.
 326	 */
 327	return inode_owner_or_capable(&nop_mnt_idmap,
 328				      file_inode(vma->vm_file)) ||
 329	       file_permission(vma->vm_file, MAY_WRITE) == 0;
 330}
 331
 332static inline int madvise_folio_pte_batch(unsigned long addr, unsigned long end,
 333					  struct folio *folio, pte_t *ptep,
 334					  pte_t pte, bool *any_young,
 335					  bool *any_dirty)
 336{
 337	const fpb_t fpb_flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
 338	int max_nr = (end - addr) / PAGE_SIZE;
 339
 340	return folio_pte_batch(folio, addr, ptep, pte, max_nr, fpb_flags, NULL,
 341			       any_young, any_dirty);
 342}
 343
 344static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
 345				unsigned long addr, unsigned long end,
 346				struct mm_walk *walk)
 347{
 348	struct madvise_walk_private *private = walk->private;
 349	struct mmu_gather *tlb = private->tlb;
 350	bool pageout = private->pageout;
 351	struct mm_struct *mm = tlb->mm;
 352	struct vm_area_struct *vma = walk->vma;
 353	pte_t *start_pte, *pte, ptent;
 354	spinlock_t *ptl;
 355	struct folio *folio = NULL;
 356	LIST_HEAD(folio_list);
 357	bool pageout_anon_only_filter;
 358	unsigned int batch_count = 0;
 359	int nr;
 360
 361	if (fatal_signal_pending(current))
 362		return -EINTR;
 363
 364	pageout_anon_only_filter = pageout && !vma_is_anonymous(vma) &&
 365					!can_do_file_pageout(vma);
 366
 367#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 368	if (pmd_trans_huge(*pmd)) {
 369		pmd_t orig_pmd;
 370		unsigned long next = pmd_addr_end(addr, end);
 371
 372		tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
 373		ptl = pmd_trans_huge_lock(pmd, vma);
 374		if (!ptl)
 375			return 0;
 376
 377		orig_pmd = *pmd;
 378		if (is_huge_zero_pmd(orig_pmd))
 379			goto huge_unlock;
 380
 381		if (unlikely(!pmd_present(orig_pmd))) {
 382			VM_BUG_ON(thp_migration_supported() &&
 383					!is_pmd_migration_entry(orig_pmd));
 384			goto huge_unlock;
 385		}
 386
 387		folio = pmd_folio(orig_pmd);
 388
 389		/* Do not interfere with other mappings of this folio */
 390		if (folio_likely_mapped_shared(folio))
 391			goto huge_unlock;
 392
 393		if (pageout_anon_only_filter && !folio_test_anon(folio))
 394			goto huge_unlock;
 395
 396		if (next - addr != HPAGE_PMD_SIZE) {
 397			int err;
 398
 399			folio_get(folio);
 400			spin_unlock(ptl);
 401			folio_lock(folio);
 402			err = split_folio(folio);
 403			folio_unlock(folio);
 404			folio_put(folio);
 405			if (!err)
 406				goto regular_folio;
 407			return 0;
 408		}
 409
 410		if (!pageout && pmd_young(orig_pmd)) {
 411			pmdp_invalidate(vma, addr, pmd);
 412			orig_pmd = pmd_mkold(orig_pmd);
 413
 414			set_pmd_at(mm, addr, pmd, orig_pmd);
 415			tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
 416		}
 417
 418		folio_clear_referenced(folio);
 419		folio_test_clear_young(folio);
 420		if (folio_test_active(folio))
 421			folio_set_workingset(folio);
 422		if (pageout) {
 423			if (folio_isolate_lru(folio)) {
 424				if (folio_test_unevictable(folio))
 425					folio_putback_lru(folio);
 426				else
 427					list_add(&folio->lru, &folio_list);
 428			}
 429		} else
 430			folio_deactivate(folio);
 431huge_unlock:
 432		spin_unlock(ptl);
 433		if (pageout)
 434			reclaim_pages(&folio_list);
 435		return 0;
 436	}
 437
 438regular_folio:
 
 
 439#endif
 440	tlb_change_page_size(tlb, PAGE_SIZE);
 441restart:
 442	start_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 443	if (!start_pte)
 444		return 0;
 445	flush_tlb_batched_pending(mm);
 446	arch_enter_lazy_mmu_mode();
 447	for (; addr < end; pte += nr, addr += nr * PAGE_SIZE) {
 448		nr = 1;
 449		ptent = ptep_get(pte);
 450
 451		if (++batch_count == SWAP_CLUSTER_MAX) {
 452			batch_count = 0;
 453			if (need_resched()) {
 454				arch_leave_lazy_mmu_mode();
 455				pte_unmap_unlock(start_pte, ptl);
 456				cond_resched();
 457				goto restart;
 458			}
 459		}
 460
 461		if (pte_none(ptent))
 462			continue;
 463
 464		if (!pte_present(ptent))
 465			continue;
 466
 467		folio = vm_normal_folio(vma, addr, ptent);
 468		if (!folio || folio_is_zone_device(folio))
 469			continue;
 470
 471		/*
 472		 * If we encounter a large folio, only split it if it is not
 473		 * fully mapped within the range we are operating on. Otherwise
 474		 * leave it as is so that it can be swapped out whole. If we
 475		 * fail to split a folio, leave it in place and advance to the
 476		 * next pte in the range.
 477		 */
 478		if (folio_test_large(folio)) {
 479			bool any_young;
 480
 481			nr = madvise_folio_pte_batch(addr, end, folio, pte,
 482						     ptent, &any_young, NULL);
 483			if (any_young)
 484				ptent = pte_mkyoung(ptent);
 485
 486			if (nr < folio_nr_pages(folio)) {
 487				int err;
 488
 489				if (folio_likely_mapped_shared(folio))
 490					continue;
 491				if (pageout_anon_only_filter && !folio_test_anon(folio))
 492					continue;
 493				if (!folio_trylock(folio))
 494					continue;
 495				folio_get(folio);
 496				arch_leave_lazy_mmu_mode();
 497				pte_unmap_unlock(start_pte, ptl);
 498				start_pte = NULL;
 499				err = split_folio(folio);
 500				folio_unlock(folio);
 501				folio_put(folio);
 502				start_pte = pte =
 503					pte_offset_map_lock(mm, pmd, addr, &ptl);
 504				if (!start_pte)
 505					break;
 506				arch_enter_lazy_mmu_mode();
 507				if (!err)
 508					nr = 0;
 509				continue;
 510			}
 
 
 
 
 
 
 511		}
 512
 513		/*
 514		 * Do not interfere with other mappings of this folio and
 515		 * non-LRU folio. If we have a large folio at this point, we
 516		 * know it is fully mapped so if its mapcount is the same as its
 517		 * number of pages, it must be exclusive.
 518		 */
 519		if (!folio_test_lru(folio) ||
 520		    folio_mapcount(folio) != folio_nr_pages(folio))
 521			continue;
 522
 523		if (pageout_anon_only_filter && !folio_test_anon(folio))
 524			continue;
 525
 526		if (!pageout && pte_young(ptent)) {
 527			clear_young_dirty_ptes(vma, addr, pte, nr,
 528					       CYDP_CLEAR_YOUNG);
 529			tlb_remove_tlb_entries(tlb, pte, nr, addr);
 
 
 530		}
 531
 532		/*
 533		 * We are deactivating a folio for accelerating reclaiming.
 534		 * VM couldn't reclaim the folio unless we clear PG_young.
 535		 * As a side effect, it makes confuse idle-page tracking
 536		 * because they will miss recent referenced history.
 537		 */
 538		folio_clear_referenced(folio);
 539		folio_test_clear_young(folio);
 540		if (folio_test_active(folio))
 541			folio_set_workingset(folio);
 542		if (pageout) {
 543			if (folio_isolate_lru(folio)) {
 544				if (folio_test_unevictable(folio))
 545					folio_putback_lru(folio);
 546				else
 547					list_add(&folio->lru, &folio_list);
 548			}
 549		} else
 550			folio_deactivate(folio);
 551	}
 552
 553	if (start_pte) {
 554		arch_leave_lazy_mmu_mode();
 555		pte_unmap_unlock(start_pte, ptl);
 556	}
 557	if (pageout)
 558		reclaim_pages(&folio_list);
 559	cond_resched();
 560
 561	return 0;
 562}
 563
 564static const struct mm_walk_ops cold_walk_ops = {
 565	.pmd_entry = madvise_cold_or_pageout_pte_range,
 566	.walk_lock = PGWALK_RDLOCK,
 567};
 568
 569static void madvise_cold_page_range(struct mmu_gather *tlb,
 570			     struct vm_area_struct *vma,
 571			     unsigned long addr, unsigned long end)
 572{
 573	struct madvise_walk_private walk_private = {
 574		.pageout = false,
 575		.tlb = tlb,
 576	};
 577
 578	tlb_start_vma(tlb, vma);
 579	walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 580	tlb_end_vma(tlb, vma);
 581}
 582
 583static inline bool can_madv_lru_vma(struct vm_area_struct *vma)
 584{
 585	return !(vma->vm_flags & (VM_LOCKED|VM_PFNMAP|VM_HUGETLB));
 586}
 587
 588static long madvise_cold(struct vm_area_struct *vma,
 589			struct vm_area_struct **prev,
 590			unsigned long start_addr, unsigned long end_addr)
 591{
 592	struct mm_struct *mm = vma->vm_mm;
 593	struct mmu_gather tlb;
 594
 595	*prev = vma;
 596	if (!can_madv_lru_vma(vma))
 597		return -EINVAL;
 598
 599	lru_add_drain();
 600	tlb_gather_mmu(&tlb, mm);
 601	madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 602	tlb_finish_mmu(&tlb);
 603
 604	return 0;
 605}
 606
 607static void madvise_pageout_page_range(struct mmu_gather *tlb,
 608			     struct vm_area_struct *vma,
 609			     unsigned long addr, unsigned long end)
 610{
 611	struct madvise_walk_private walk_private = {
 612		.pageout = true,
 613		.tlb = tlb,
 614	};
 615
 616	tlb_start_vma(tlb, vma);
 617	walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 618	tlb_end_vma(tlb, vma);
 619}
 620
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 621static long madvise_pageout(struct vm_area_struct *vma,
 622			struct vm_area_struct **prev,
 623			unsigned long start_addr, unsigned long end_addr)
 624{
 625	struct mm_struct *mm = vma->vm_mm;
 626	struct mmu_gather tlb;
 627
 628	*prev = vma;
 629	if (!can_madv_lru_vma(vma))
 630		return -EINVAL;
 631
 632	/*
 633	 * If the VMA belongs to a private file mapping, there can be private
 634	 * dirty pages which can be paged out if even this process is neither
 635	 * owner nor write capable of the file. We allow private file mappings
 636	 * further to pageout dirty anon pages.
 637	 */
 638	if (!vma_is_anonymous(vma) && (!can_do_file_pageout(vma) &&
 639				(vma->vm_flags & VM_MAYSHARE)))
 640		return 0;
 641
 642	lru_add_drain();
 643	tlb_gather_mmu(&tlb, mm);
 644	madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
 645	tlb_finish_mmu(&tlb);
 646
 647	return 0;
 648}
 649
 650static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 651				unsigned long end, struct mm_walk *walk)
 652
 653{
 654	const cydp_t cydp_flags = CYDP_CLEAR_YOUNG | CYDP_CLEAR_DIRTY;
 655	struct mmu_gather *tlb = walk->private;
 656	struct mm_struct *mm = tlb->mm;
 657	struct vm_area_struct *vma = walk->vma;
 658	spinlock_t *ptl;
 659	pte_t *start_pte, *pte, ptent;
 660	struct folio *folio;
 661	int nr_swap = 0;
 662	unsigned long next;
 663	int nr, max_nr;
 664
 665	next = pmd_addr_end(addr, end);
 666	if (pmd_trans_huge(*pmd))
 667		if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 668			return 0;
 
 
 
 669
 670	tlb_change_page_size(tlb, PAGE_SIZE);
 671	start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 672	if (!start_pte)
 673		return 0;
 674	flush_tlb_batched_pending(mm);
 675	arch_enter_lazy_mmu_mode();
 676	for (; addr != end; pte += nr, addr += PAGE_SIZE * nr) {
 677		nr = 1;
 678		ptent = ptep_get(pte);
 679
 680		if (pte_none(ptent))
 681			continue;
 682		/*
 683		 * If the pte has swp_entry, just clear page table to
 684		 * prevent swap-in which is more expensive rather than
 685		 * (page allocation + zeroing).
 686		 */
 687		if (!pte_present(ptent)) {
 688			swp_entry_t entry;
 689
 690			entry = pte_to_swp_entry(ptent);
 691			if (!non_swap_entry(entry)) {
 692				max_nr = (end - addr) / PAGE_SIZE;
 693				nr = swap_pte_batch(pte, max_nr, ptent);
 694				nr_swap -= nr;
 695				free_swap_and_cache_nr(entry, nr);
 696				clear_not_present_full_ptes(mm, addr, pte, nr, tlb->fullmm);
 697			} else if (is_hwpoison_entry(entry) ||
 698				   is_poisoned_swp_entry(entry)) {
 699				pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 700			}
 701			continue;
 702		}
 703
 704		folio = vm_normal_folio(vma, addr, ptent);
 705		if (!folio || folio_is_zone_device(folio))
 706			continue;
 707
 708		/*
 709		 * If we encounter a large folio, only split it if it is not
 710		 * fully mapped within the range we are operating on. Otherwise
 711		 * leave it as is so that it can be marked as lazyfree. If we
 712		 * fail to split a folio, leave it in place and advance to the
 713		 * next pte in the range.
 714		 */
 715		if (folio_test_large(folio)) {
 716			bool any_young, any_dirty;
 717
 718			nr = madvise_folio_pte_batch(addr, end, folio, pte,
 719						     ptent, &any_young, &any_dirty);
 720
 721			if (nr < folio_nr_pages(folio)) {
 722				int err;
 723
 724				if (folio_likely_mapped_shared(folio))
 725					continue;
 726				if (!folio_trylock(folio))
 727					continue;
 728				folio_get(folio);
 729				arch_leave_lazy_mmu_mode();
 730				pte_unmap_unlock(start_pte, ptl);
 731				start_pte = NULL;
 732				err = split_folio(folio);
 733				folio_unlock(folio);
 734				folio_put(folio);
 735				pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 736				start_pte = pte;
 737				if (!start_pte)
 738					break;
 739				arch_enter_lazy_mmu_mode();
 740				if (!err)
 741					nr = 0;
 742				continue;
 743			}
 
 
 
 
 
 
 
 744
 745			if (any_young)
 746				ptent = pte_mkyoung(ptent);
 747			if (any_dirty)
 748				ptent = pte_mkdirty(ptent);
 749		}
 750
 751		if (folio_test_swapcache(folio) || folio_test_dirty(folio)) {
 752			if (!folio_trylock(folio))
 753				continue;
 754			/*
 755			 * If we have a large folio at this point, we know it is
 756			 * fully mapped so if its mapcount is the same as its
 757			 * number of pages, it must be exclusive.
 758			 */
 759			if (folio_mapcount(folio) != folio_nr_pages(folio)) {
 760				folio_unlock(folio);
 761				continue;
 762			}
 763
 764			if (folio_test_swapcache(folio) &&
 765			    !folio_free_swap(folio)) {
 766				folio_unlock(folio);
 767				continue;
 768			}
 769
 770			folio_clear_dirty(folio);
 771			folio_unlock(folio);
 772		}
 773
 774		if (pte_young(ptent) || pte_dirty(ptent)) {
 775			clear_young_dirty_ptes(vma, addr, pte, nr, cydp_flags);
 776			tlb_remove_tlb_entries(tlb, pte, nr, addr);
 
 
 
 
 
 
 
 
 
 
 
 777		}
 778		folio_mark_lazyfree(folio);
 779	}
 
 
 
 
 780
 781	if (nr_swap)
 782		add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 783	if (start_pte) {
 784		arch_leave_lazy_mmu_mode();
 785		pte_unmap_unlock(start_pte, ptl);
 786	}
 
 
 787	cond_resched();
 788
 789	return 0;
 790}
 791
 792static const struct mm_walk_ops madvise_free_walk_ops = {
 793	.pmd_entry		= madvise_free_pte_range,
 794	.walk_lock		= PGWALK_RDLOCK,
 795};
 796
 797static int madvise_free_single_vma(struct vm_area_struct *vma,
 798			unsigned long start_addr, unsigned long end_addr)
 799{
 800	struct mm_struct *mm = vma->vm_mm;
 801	struct mmu_notifier_range range;
 802	struct mmu_gather tlb;
 803
 804	/* MADV_FREE works for only anon vma at the moment */
 805	if (!vma_is_anonymous(vma))
 806		return -EINVAL;
 807
 808	range.start = max(vma->vm_start, start_addr);
 809	if (range.start >= vma->vm_end)
 810		return -EINVAL;
 811	range.end = min(vma->vm_end, end_addr);
 812	if (range.end <= vma->vm_start)
 813		return -EINVAL;
 814	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
 815				range.start, range.end);
 816
 817	lru_add_drain();
 818	tlb_gather_mmu(&tlb, mm);
 819	update_hiwater_rss(mm);
 820
 821	mmu_notifier_invalidate_range_start(&range);
 822	tlb_start_vma(&tlb, vma);
 823	walk_page_range(vma->vm_mm, range.start, range.end,
 824			&madvise_free_walk_ops, &tlb);
 825	tlb_end_vma(&tlb, vma);
 826	mmu_notifier_invalidate_range_end(&range);
 827	tlb_finish_mmu(&tlb);
 828
 829	return 0;
 830}
 831
 832/*
 833 * Application no longer needs these pages.  If the pages are dirty,
 834 * it's OK to just throw them away.  The app will be more careful about
 835 * data it wants to keep.  Be sure to free swap resources too.  The
 836 * zap_page_range_single call sets things up for shrink_active_list to actually
 837 * free these pages later if no one else has touched them in the meantime,
 838 * although we could add these pages to a global reuse list for
 839 * shrink_active_list to pick up before reclaiming other pages.
 840 *
 841 * NB: This interface discards data rather than pushes it out to swap,
 842 * as some implementations do.  This has performance implications for
 843 * applications like large transactional databases which want to discard
 844 * pages in anonymous maps after committing to backing store the data
 845 * that was kept in them.  There is no reason to write this data out to
 846 * the swap area if the application is discarding it.
 847 *
 848 * An interface that causes the system to free clean pages and flush
 849 * dirty pages is already available as msync(MS_INVALIDATE).
 850 */
 851static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 852					unsigned long start, unsigned long end)
 853{
 854	zap_page_range_single(vma, start, end - start, NULL);
 855	return 0;
 856}
 857
 858static bool madvise_dontneed_free_valid_vma(struct vm_area_struct *vma,
 859					    unsigned long start,
 860					    unsigned long *end,
 861					    int behavior)
 862{
 863	if (!is_vm_hugetlb_page(vma)) {
 864		unsigned int forbidden = VM_PFNMAP;
 865
 866		if (behavior != MADV_DONTNEED_LOCKED)
 867			forbidden |= VM_LOCKED;
 868
 869		return !(vma->vm_flags & forbidden);
 870	}
 871
 872	if (behavior != MADV_DONTNEED && behavior != MADV_DONTNEED_LOCKED)
 873		return false;
 874	if (start & ~huge_page_mask(hstate_vma(vma)))
 875		return false;
 876
 877	/*
 878	 * Madvise callers expect the length to be rounded up to PAGE_SIZE
 879	 * boundaries, and may be unaware that this VMA uses huge pages.
 880	 * Avoid unexpected data loss by rounding down the number of
 881	 * huge pages freed.
 882	 */
 883	*end = ALIGN_DOWN(*end, huge_page_size(hstate_vma(vma)));
 884
 885	return true;
 886}
 887
 888static long madvise_dontneed_free(struct vm_area_struct *vma,
 889				  struct vm_area_struct **prev,
 890				  unsigned long start, unsigned long end,
 891				  int behavior)
 892{
 893	struct mm_struct *mm = vma->vm_mm;
 894
 895	*prev = vma;
 896	if (!madvise_dontneed_free_valid_vma(vma, start, &end, behavior))
 897		return -EINVAL;
 898
 899	if (start == end)
 900		return 0;
 901
 902	if (!userfaultfd_remove(vma, start, end)) {
 903		*prev = NULL; /* mmap_lock has been dropped, prev is stale */
 904
 905		mmap_read_lock(mm);
 906		vma = vma_lookup(mm, start);
 907		if (!vma)
 908			return -ENOMEM;
 909		/*
 910		 * Potential end adjustment for hugetlb vma is OK as
 911		 * the check below keeps end within vma.
 912		 */
 913		if (!madvise_dontneed_free_valid_vma(vma, start, &end,
 914						     behavior))
 
 
 
 
 
 
 
 915			return -EINVAL;
 916		if (end > vma->vm_end) {
 917			/*
 918			 * Don't fail if end > vma->vm_end. If the old
 919			 * vma was split while the mmap_lock was
 920			 * released the effect of the concurrent
 921			 * operation may not cause madvise() to
 922			 * have an undefined result. There may be an
 923			 * adjacent next vma that we'll walk
 924			 * next. userfaultfd_remove() will generate an
 925			 * UFFD_EVENT_REMOVE repetition on the
 926			 * end-vma->vm_end range, but the manager can
 927			 * handle a repetition fine.
 928			 */
 929			end = vma->vm_end;
 930		}
 931		/*
 932		 * If the memory region between start and end was
 933		 * originally backed by 4kB pages and then remapped to
 934		 * be backed by hugepages while mmap_lock was dropped,
 935		 * the adjustment for hugetlb vma above may have rounded
 936		 * end down to the start address.
 937		 */
 938		if (start == end)
 939			return 0;
 940		VM_WARN_ON(start > end);
 941	}
 942
 943	if (behavior == MADV_DONTNEED || behavior == MADV_DONTNEED_LOCKED)
 944		return madvise_dontneed_single_vma(vma, start, end);
 945	else if (behavior == MADV_FREE)
 946		return madvise_free_single_vma(vma, start, end);
 947	else
 948		return -EINVAL;
 949}
 950
 951static long madvise_populate(struct mm_struct *mm, unsigned long start,
 952		unsigned long end, int behavior)
 953{
 954	const bool write = behavior == MADV_POPULATE_WRITE;
 955	int locked = 1;
 956	long pages;
 957
 958	while (start < end) {
 959		/* Populate (prefault) page tables readable/writable. */
 960		pages = faultin_page_range(mm, start, end, write, &locked);
 961		if (!locked) {
 962			mmap_read_lock(mm);
 963			locked = 1;
 964		}
 965		if (pages < 0) {
 966			switch (pages) {
 967			case -EINTR:
 968				return -EINTR;
 969			case -EINVAL: /* Incompatible mappings / permissions. */
 970				return -EINVAL;
 971			case -EHWPOISON:
 972				return -EHWPOISON;
 973			case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
 974				return -EFAULT;
 975			default:
 976				pr_warn_once("%s: unhandled return value: %ld\n",
 977					     __func__, pages);
 978				fallthrough;
 979			case -ENOMEM: /* No VMA or out of memory. */
 980				return -ENOMEM;
 981			}
 982		}
 983		start += pages * PAGE_SIZE;
 984	}
 985	return 0;
 986}
 987
 988/*
 989 * Application wants to free up the pages and associated backing store.
 990 * This is effectively punching a hole into the middle of a file.
 991 */
 992static long madvise_remove(struct vm_area_struct *vma,
 993				struct vm_area_struct **prev,
 994				unsigned long start, unsigned long end)
 995{
 996	loff_t offset;
 997	int error;
 998	struct file *f;
 999	struct mm_struct *mm = vma->vm_mm;
1000
1001	*prev = NULL;	/* tell sys_madvise we drop mmap_lock */
1002
1003	if (vma->vm_flags & VM_LOCKED)
1004		return -EINVAL;
1005
1006	f = vma->vm_file;
1007
1008	if (!f || !f->f_mapping || !f->f_mapping->host) {
1009			return -EINVAL;
1010	}
1011
1012	if (!vma_is_shared_maywrite(vma))
1013		return -EACCES;
1014
1015	offset = (loff_t)(start - vma->vm_start)
1016			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
1017
1018	/*
1019	 * Filesystem's fallocate may need to take i_rwsem.  We need to
1020	 * explicitly grab a reference because the vma (and hence the
1021	 * vma's reference to the file) can go away as soon as we drop
1022	 * mmap_lock.
1023	 */
1024	get_file(f);
1025	if (userfaultfd_remove(vma, start, end)) {
1026		/* mmap_lock was not released by userfaultfd_remove() */
1027		mmap_read_unlock(mm);
1028	}
1029	error = vfs_fallocate(f,
1030				FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1031				offset, end - start);
1032	fput(f);
1033	mmap_read_lock(mm);
1034	return error;
1035}
1036
1037static bool is_valid_guard_vma(struct vm_area_struct *vma, bool allow_locked)
1038{
1039	vm_flags_t disallowed = VM_SPECIAL | VM_HUGETLB;
1040
1041	/*
1042	 * A user could lock after setting a guard range but that's fine, as
1043	 * they'd not be able to fault in. The issue arises when we try to zap
1044	 * existing locked VMAs. We don't want to do that.
1045	 */
1046	if (!allow_locked)
1047		disallowed |= VM_LOCKED;
1048
1049	if (!vma_is_anonymous(vma))
1050		return false;
1051
1052	if ((vma->vm_flags & (VM_MAYWRITE | disallowed)) != VM_MAYWRITE)
1053		return false;
1054
1055	return true;
1056}
1057
1058static bool is_guard_pte_marker(pte_t ptent)
1059{
1060	return is_pte_marker(ptent) &&
1061		is_guard_swp_entry(pte_to_swp_entry(ptent));
1062}
1063
1064static int guard_install_pud_entry(pud_t *pud, unsigned long addr,
1065				   unsigned long next, struct mm_walk *walk)
1066{
1067	pud_t pudval = pudp_get(pud);
1068
1069	/* If huge return >0 so we abort the operation + zap. */
1070	return pud_trans_huge(pudval) || pud_devmap(pudval);
1071}
1072
1073static int guard_install_pmd_entry(pmd_t *pmd, unsigned long addr,
1074				   unsigned long next, struct mm_walk *walk)
1075{
1076	pmd_t pmdval = pmdp_get(pmd);
1077
1078	/* If huge return >0 so we abort the operation + zap. */
1079	return pmd_trans_huge(pmdval) || pmd_devmap(pmdval);
1080}
1081
1082static int guard_install_pte_entry(pte_t *pte, unsigned long addr,
1083				   unsigned long next, struct mm_walk *walk)
1084{
1085	pte_t pteval = ptep_get(pte);
1086	unsigned long *nr_pages = (unsigned long *)walk->private;
1087
1088	/* If there is already a guard page marker, we have nothing to do. */
1089	if (is_guard_pte_marker(pteval)) {
1090		(*nr_pages)++;
1091
1092		return 0;
1093	}
1094
1095	/* If populated return >0 so we abort the operation + zap. */
1096	return 1;
1097}
1098
1099static int guard_install_set_pte(unsigned long addr, unsigned long next,
1100				 pte_t *ptep, struct mm_walk *walk)
1101{
1102	unsigned long *nr_pages = (unsigned long *)walk->private;
1103
1104	/* Simply install a PTE marker, this causes segfault on access. */
1105	*ptep = make_pte_marker(PTE_MARKER_GUARD);
1106	(*nr_pages)++;
1107
1108	return 0;
1109}
1110
1111static const struct mm_walk_ops guard_install_walk_ops = {
1112	.pud_entry		= guard_install_pud_entry,
1113	.pmd_entry		= guard_install_pmd_entry,
1114	.pte_entry		= guard_install_pte_entry,
1115	.install_pte		= guard_install_set_pte,
1116	.walk_lock		= PGWALK_RDLOCK,
1117};
1118
1119static long madvise_guard_install(struct vm_area_struct *vma,
1120				 struct vm_area_struct **prev,
1121				 unsigned long start, unsigned long end)
1122{
1123	long err;
1124	int i;
1125
1126	*prev = vma;
1127	if (!is_valid_guard_vma(vma, /* allow_locked = */false))
1128		return -EINVAL;
1129
1130	/*
1131	 * If we install guard markers, then the range is no longer
1132	 * empty from a page table perspective and therefore it's
1133	 * appropriate to have an anon_vma.
1134	 *
1135	 * This ensures that on fork, we copy page tables correctly.
1136	 */
1137	err = anon_vma_prepare(vma);
1138	if (err)
1139		return err;
1140
1141	/*
1142	 * Optimistically try to install the guard marker pages first. If any
1143	 * non-guard pages are encountered, give up and zap the range before
1144	 * trying again.
1145	 *
1146	 * We try a few times before giving up and releasing back to userland to
1147	 * loop around, releasing locks in the process to avoid contention. This
1148	 * would only happen if there was a great many racing page faults.
1149	 *
1150	 * In most cases we should simply install the guard markers immediately
1151	 * with no zap or looping.
1152	 */
1153	for (i = 0; i < MAX_MADVISE_GUARD_RETRIES; i++) {
1154		unsigned long nr_pages = 0;
1155
1156		/* Returns < 0 on error, == 0 if success, > 0 if zap needed. */
1157		err = walk_page_range_mm(vma->vm_mm, start, end,
1158					 &guard_install_walk_ops, &nr_pages);
1159		if (err < 0)
1160			return err;
1161
1162		if (err == 0) {
1163			unsigned long nr_expected_pages = PHYS_PFN(end - start);
1164
1165			VM_WARN_ON(nr_pages != nr_expected_pages);
1166			return 0;
1167		}
1168
1169		/*
1170		 * OK some of the range have non-guard pages mapped, zap
1171		 * them. This leaves existing guard pages in place.
1172		 */
1173		zap_page_range_single(vma, start, end - start, NULL);
1174	}
1175
1176	/*
1177	 * We were unable to install the guard pages due to being raced by page
1178	 * faults. This should not happen ordinarily. We return to userspace and
1179	 * immediately retry, relieving lock contention.
1180	 */
1181	return restart_syscall();
1182}
1183
1184static int guard_remove_pud_entry(pud_t *pud, unsigned long addr,
1185				  unsigned long next, struct mm_walk *walk)
1186{
1187	pud_t pudval = pudp_get(pud);
1188
1189	/* If huge, cannot have guard pages present, so no-op - skip. */
1190	if (pud_trans_huge(pudval) || pud_devmap(pudval))
1191		walk->action = ACTION_CONTINUE;
1192
1193	return 0;
1194}
1195
1196static int guard_remove_pmd_entry(pmd_t *pmd, unsigned long addr,
1197				  unsigned long next, struct mm_walk *walk)
1198{
1199	pmd_t pmdval = pmdp_get(pmd);
1200
1201	/* If huge, cannot have guard pages present, so no-op - skip. */
1202	if (pmd_trans_huge(pmdval) || pmd_devmap(pmdval))
1203		walk->action = ACTION_CONTINUE;
1204
1205	return 0;
1206}
1207
1208static int guard_remove_pte_entry(pte_t *pte, unsigned long addr,
1209				  unsigned long next, struct mm_walk *walk)
1210{
1211	pte_t ptent = ptep_get(pte);
1212
1213	if (is_guard_pte_marker(ptent)) {
1214		/* Simply clear the PTE marker. */
1215		pte_clear_not_present_full(walk->mm, addr, pte, false);
1216		update_mmu_cache(walk->vma, addr, pte);
1217	}
1218
1219	return 0;
1220}
1221
1222static const struct mm_walk_ops guard_remove_walk_ops = {
1223	.pud_entry		= guard_remove_pud_entry,
1224	.pmd_entry		= guard_remove_pmd_entry,
1225	.pte_entry		= guard_remove_pte_entry,
1226	.walk_lock		= PGWALK_RDLOCK,
1227};
1228
1229static long madvise_guard_remove(struct vm_area_struct *vma,
1230				 struct vm_area_struct **prev,
1231				 unsigned long start, unsigned long end)
1232{
1233	*prev = vma;
1234	/*
1235	 * We're ok with removing guards in mlock()'d ranges, as this is a
1236	 * non-destructive action.
1237	 */
1238	if (!is_valid_guard_vma(vma, /* allow_locked = */true))
1239		return -EINVAL;
1240
1241	return walk_page_range(vma->vm_mm, start, end,
1242			       &guard_remove_walk_ops, NULL);
1243}
1244
1245/*
1246 * Apply an madvise behavior to a region of a vma.  madvise_update_vma
1247 * will handle splitting a vm area into separate areas, each area with its own
1248 * behavior.
1249 */
1250static int madvise_vma_behavior(struct vm_area_struct *vma,
1251				struct vm_area_struct **prev,
1252				unsigned long start, unsigned long end,
1253				unsigned long behavior)
1254{
1255	int error;
1256	struct anon_vma_name *anon_name;
1257	unsigned long new_flags = vma->vm_flags;
1258
1259	if (unlikely(!can_modify_vma_madv(vma, behavior)))
1260		return -EPERM;
1261
1262	switch (behavior) {
1263	case MADV_REMOVE:
1264		return madvise_remove(vma, prev, start, end);
1265	case MADV_WILLNEED:
1266		return madvise_willneed(vma, prev, start, end);
1267	case MADV_COLD:
1268		return madvise_cold(vma, prev, start, end);
1269	case MADV_PAGEOUT:
1270		return madvise_pageout(vma, prev, start, end);
1271	case MADV_FREE:
1272	case MADV_DONTNEED:
1273	case MADV_DONTNEED_LOCKED:
1274		return madvise_dontneed_free(vma, prev, start, end, behavior);
1275	case MADV_NORMAL:
1276		new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
1277		break;
1278	case MADV_SEQUENTIAL:
1279		new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
1280		break;
1281	case MADV_RANDOM:
1282		new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
1283		break;
1284	case MADV_DONTFORK:
1285		new_flags |= VM_DONTCOPY;
1286		break;
1287	case MADV_DOFORK:
1288		if (vma->vm_flags & VM_IO)
1289			return -EINVAL;
1290		new_flags &= ~VM_DONTCOPY;
1291		break;
1292	case MADV_WIPEONFORK:
1293		/* MADV_WIPEONFORK is only supported on anonymous memory. */
1294		if (vma->vm_file || vma->vm_flags & VM_SHARED)
1295			return -EINVAL;
1296		new_flags |= VM_WIPEONFORK;
1297		break;
1298	case MADV_KEEPONFORK:
1299		if (vma->vm_flags & VM_DROPPABLE)
1300			return -EINVAL;
1301		new_flags &= ~VM_WIPEONFORK;
1302		break;
1303	case MADV_DONTDUMP:
1304		new_flags |= VM_DONTDUMP;
1305		break;
1306	case MADV_DODUMP:
1307		if ((!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) ||
1308		    (vma->vm_flags & VM_DROPPABLE))
1309			return -EINVAL;
1310		new_flags &= ~VM_DONTDUMP;
1311		break;
1312	case MADV_MERGEABLE:
1313	case MADV_UNMERGEABLE:
1314		error = ksm_madvise(vma, start, end, behavior, &new_flags);
1315		if (error)
1316			goto out;
1317		break;
1318	case MADV_HUGEPAGE:
1319	case MADV_NOHUGEPAGE:
1320		error = hugepage_madvise(vma, &new_flags, behavior);
1321		if (error)
1322			goto out;
1323		break;
1324	case MADV_COLLAPSE:
1325		return madvise_collapse(vma, prev, start, end);
1326	case MADV_GUARD_INSTALL:
1327		return madvise_guard_install(vma, prev, start, end);
1328	case MADV_GUARD_REMOVE:
1329		return madvise_guard_remove(vma, prev, start, end);
1330	}
1331
1332	anon_name = anon_vma_name(vma);
1333	anon_vma_name_get(anon_name);
1334	error = madvise_update_vma(vma, prev, start, end, new_flags,
1335				   anon_name);
1336	anon_vma_name_put(anon_name);
1337
1338out:
1339	/*
1340	 * madvise() returns EAGAIN if kernel resources, such as
1341	 * slab, are temporarily unavailable.
1342	 */
1343	if (error == -ENOMEM)
1344		error = -EAGAIN;
1345	return error;
1346}
1347
1348#ifdef CONFIG_MEMORY_FAILURE
1349/*
1350 * Error injection support for memory error handling.
1351 */
1352static int madvise_inject_error(int behavior,
1353		unsigned long start, unsigned long end)
1354{
 
 
1355	unsigned long size;
1356
1357	if (!capable(CAP_SYS_ADMIN))
1358		return -EPERM;
1359
1360
1361	for (; start < end; start += size) {
1362		unsigned long pfn;
1363		struct page *page;
1364		int ret;
1365
1366		ret = get_user_pages_fast(start, 1, 0, &page);
1367		if (ret != 1)
1368			return ret;
1369		pfn = page_to_pfn(page);
1370
1371		/*
1372		 * When soft offlining hugepages, after migrating the page
1373		 * we dissolve it, therefore in the second loop "page" will
1374		 * no longer be a compound page.
1375		 */
1376		size = page_size(compound_head(page));
1377
 
 
 
 
 
1378		if (behavior == MADV_SOFT_OFFLINE) {
1379			pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
1380				 pfn, start);
 
1381			ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
1382		} else {
1383			pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
1384				 pfn, start);
1385			ret = memory_failure(pfn, MF_ACTION_REQUIRED | MF_COUNT_INCREASED | MF_SW_SIMULATED);
1386			if (ret == -EOPNOTSUPP)
1387				ret = 0;
1388		}
1389
 
 
 
 
 
 
 
 
 
 
 
1390		if (ret)
1391			return ret;
1392	}
1393
 
 
 
 
1394	return 0;
1395}
1396#endif
1397
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1398static bool
1399madvise_behavior_valid(int behavior)
1400{
1401	switch (behavior) {
1402	case MADV_DOFORK:
1403	case MADV_DONTFORK:
1404	case MADV_NORMAL:
1405	case MADV_SEQUENTIAL:
1406	case MADV_RANDOM:
1407	case MADV_REMOVE:
1408	case MADV_WILLNEED:
1409	case MADV_DONTNEED:
1410	case MADV_DONTNEED_LOCKED:
1411	case MADV_FREE:
1412	case MADV_COLD:
1413	case MADV_PAGEOUT:
1414	case MADV_POPULATE_READ:
1415	case MADV_POPULATE_WRITE:
1416#ifdef CONFIG_KSM
1417	case MADV_MERGEABLE:
1418	case MADV_UNMERGEABLE:
1419#endif
1420#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1421	case MADV_HUGEPAGE:
1422	case MADV_NOHUGEPAGE:
1423	case MADV_COLLAPSE:
1424#endif
1425	case MADV_DONTDUMP:
1426	case MADV_DODUMP:
1427	case MADV_WIPEONFORK:
1428	case MADV_KEEPONFORK:
1429	case MADV_GUARD_INSTALL:
1430	case MADV_GUARD_REMOVE:
1431#ifdef CONFIG_MEMORY_FAILURE
1432	case MADV_SOFT_OFFLINE:
1433	case MADV_HWPOISON:
1434#endif
1435		return true;
1436
1437	default:
1438		return false;
1439	}
1440}
1441
1442/* Can we invoke process_madvise() on a remote mm for the specified behavior? */
1443static bool process_madvise_remote_valid(int behavior)
1444{
1445	switch (behavior) {
1446	case MADV_COLD:
1447	case MADV_PAGEOUT:
1448	case MADV_WILLNEED:
1449	case MADV_COLLAPSE:
1450		return true;
1451	default:
1452		return false;
1453	}
1454}
1455
1456/*
1457 * Walk the vmas in range [start,end), and call the visit function on each one.
1458 * The visit function will get start and end parameters that cover the overlap
1459 * between the current vma and the original range.  Any unmapped regions in the
1460 * original range will result in this function returning -ENOMEM while still
1461 * calling the visit function on all of the existing vmas in the range.
1462 * Must be called with the mmap_lock held for reading or writing.
1463 */
1464static
1465int madvise_walk_vmas(struct mm_struct *mm, unsigned long start,
1466		      unsigned long end, unsigned long arg,
1467		      int (*visit)(struct vm_area_struct *vma,
1468				   struct vm_area_struct **prev, unsigned long start,
1469				   unsigned long end, unsigned long arg))
1470{
1471	struct vm_area_struct *vma;
1472	struct vm_area_struct *prev;
1473	unsigned long tmp;
1474	int unmapped_error = 0;
1475
1476	/*
1477	 * If the interval [start,end) covers some unmapped address
1478	 * ranges, just ignore them, but return -ENOMEM at the end.
1479	 * - different from the way of handling in mlock etc.
1480	 */
1481	vma = find_vma_prev(mm, start, &prev);
1482	if (vma && start > vma->vm_start)
1483		prev = vma;
1484
1485	for (;;) {
1486		int error;
1487
1488		/* Still start < end. */
1489		if (!vma)
1490			return -ENOMEM;
1491
1492		/* Here start < (end|vma->vm_end). */
1493		if (start < vma->vm_start) {
1494			unmapped_error = -ENOMEM;
1495			start = vma->vm_start;
1496			if (start >= end)
1497				break;
1498		}
1499
1500		/* Here vma->vm_start <= start < (end|vma->vm_end) */
1501		tmp = vma->vm_end;
1502		if (end < tmp)
1503			tmp = end;
1504
1505		/* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1506		error = visit(vma, &prev, start, tmp, arg);
1507		if (error)
1508			return error;
1509		start = tmp;
1510		if (prev && start < prev->vm_end)
1511			start = prev->vm_end;
1512		if (start >= end)
1513			break;
1514		if (prev)
1515			vma = find_vma(mm, prev->vm_end);
1516		else	/* madvise_remove dropped mmap_lock */
1517			vma = find_vma(mm, start);
1518	}
1519
1520	return unmapped_error;
1521}
1522
1523#ifdef CONFIG_ANON_VMA_NAME
1524static int madvise_vma_anon_name(struct vm_area_struct *vma,
1525				 struct vm_area_struct **prev,
1526				 unsigned long start, unsigned long end,
1527				 unsigned long anon_name)
1528{
1529	int error;
1530
1531	/* Only anonymous mappings can be named */
1532	if (vma->vm_file && !vma_is_anon_shmem(vma))
1533		return -EBADF;
1534
1535	error = madvise_update_vma(vma, prev, start, end, vma->vm_flags,
1536				   (struct anon_vma_name *)anon_name);
1537
1538	/*
1539	 * madvise() returns EAGAIN if kernel resources, such as
1540	 * slab, are temporarily unavailable.
1541	 */
1542	if (error == -ENOMEM)
1543		error = -EAGAIN;
1544	return error;
1545}
1546
1547int madvise_set_anon_name(struct mm_struct *mm, unsigned long start,
1548			  unsigned long len_in, struct anon_vma_name *anon_name)
1549{
1550	unsigned long end;
1551	unsigned long len;
1552
1553	if (start & ~PAGE_MASK)
1554		return -EINVAL;
1555	len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1556
1557	/* Check to see whether len was rounded up from small -ve to zero */
1558	if (len_in && !len)
1559		return -EINVAL;
1560
1561	end = start + len;
1562	if (end < start)
1563		return -EINVAL;
1564
1565	if (end == start)
1566		return 0;
1567
1568	return madvise_walk_vmas(mm, start, end, (unsigned long)anon_name,
1569				 madvise_vma_anon_name);
1570}
1571#endif /* CONFIG_ANON_VMA_NAME */
1572/*
1573 * The madvise(2) system call.
1574 *
1575 * Applications can use madvise() to advise the kernel how it should
1576 * handle paging I/O in this VM area.  The idea is to help the kernel
1577 * use appropriate read-ahead and caching techniques.  The information
1578 * provided is advisory only, and can be safely disregarded by the
1579 * kernel without affecting the correct operation of the application.
1580 *
1581 * behavior values:
1582 *  MADV_NORMAL - the default behavior is to read clusters.  This
1583 *		results in some read-ahead and read-behind.
1584 *  MADV_RANDOM - the system should read the minimum amount of data
1585 *		on any access, since it is unlikely that the appli-
1586 *		cation will need more than what it asks for.
1587 *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1588 *		once, so they can be aggressively read ahead, and
1589 *		can be freed soon after they are accessed.
1590 *  MADV_WILLNEED - the application is notifying the system to read
1591 *		some pages ahead.
1592 *  MADV_DONTNEED - the application is finished with the given range,
1593 *		so the kernel can free resources associated with it.
1594 *  MADV_FREE - the application marks pages in the given range as lazy free,
1595 *		where actual purges are postponed until memory pressure happens.
1596 *  MADV_REMOVE - the application wants to free up the given range of
1597 *		pages and associated backing store.
1598 *  MADV_DONTFORK - omit this area from child's address space when forking:
1599 *		typically, to avoid COWing pages pinned by get_user_pages().
1600 *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1601 *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1602 *              range after a fork.
1603 *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1604 *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1605 *		were corrupted by unrecoverable hardware memory failure.
1606 *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1607 *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1608 *		this area with pages of identical content from other such areas.
1609 *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1610 *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1611 *		huge pages in the future. Existing pages might be coalesced and
1612 *		new pages might be allocated as THP.
1613 *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1614 *		transparent huge pages so the existing pages will not be
1615 *		coalesced into THP and new pages will not be allocated as THP.
1616 *  MADV_COLLAPSE - synchronously coalesce pages into new THP.
1617 *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1618 *		from being included in its core dump.
1619 *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1620 *  MADV_COLD - the application is not expected to use this memory soon,
1621 *		deactivate pages in this range so that they can be reclaimed
1622 *		easily if memory pressure happens.
1623 *  MADV_PAGEOUT - the application is not expected to use this memory soon,
1624 *		page out the pages in this range immediately.
1625 *  MADV_POPULATE_READ - populate (prefault) page tables readable by
1626 *		triggering read faults if required
1627 *  MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1628 *		triggering write faults if required
1629 *
1630 * return values:
1631 *  zero    - success
1632 *  -EINVAL - start + len < 0, start is not page-aligned,
1633 *		"behavior" is not a valid value, or application
1634 *		is attempting to release locked or shared pages,
1635 *		or the specified address range includes file, Huge TLB,
1636 *		MAP_SHARED or VMPFNMAP range.
1637 *  -ENOMEM - addresses in the specified range are not currently
1638 *		mapped, or are outside the AS of the process.
1639 *  -EIO    - an I/O error occurred while paging in data.
1640 *  -EBADF  - map exists, but area maps something that isn't a file.
1641 *  -EAGAIN - a kernel resource was temporarily unavailable.
1642 *  -EPERM  - memory is sealed.
1643 */
1644int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1645{
1646	unsigned long end;
1647	int error;
 
 
1648	int write;
1649	size_t len;
1650	struct blk_plug plug;
1651
 
 
1652	if (!madvise_behavior_valid(behavior))
1653		return -EINVAL;
1654
1655	if (!PAGE_ALIGNED(start))
1656		return -EINVAL;
1657	len = PAGE_ALIGN(len_in);
1658
1659	/* Check to see whether len was rounded up from small -ve to zero */
1660	if (len_in && !len)
1661		return -EINVAL;
1662
1663	end = start + len;
1664	if (end < start)
1665		return -EINVAL;
1666
 
1667	if (end == start)
1668		return 0;
1669
1670#ifdef CONFIG_MEMORY_FAILURE
1671	if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1672		return madvise_inject_error(behavior, start, start + len_in);
1673#endif
1674
1675	write = madvise_need_mmap_write(behavior);
1676	if (write) {
1677		if (mmap_write_lock_killable(mm))
1678			return -EINTR;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1679	} else {
1680		mmap_read_lock(mm);
1681	}
1682
1683	start = untagged_addr_remote(mm, start);
1684	end = start + len;
 
 
 
 
 
 
1685
1686	blk_start_plug(&plug);
1687	switch (behavior) {
1688	case MADV_POPULATE_READ:
1689	case MADV_POPULATE_WRITE:
1690		error = madvise_populate(mm, start, end, behavior);
1691		break;
1692	default:
1693		error = madvise_walk_vmas(mm, start, end, behavior,
1694					  madvise_vma_behavior);
1695		break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1696	}
 
1697	blk_finish_plug(&plug);
1698
1699	if (write)
1700		mmap_write_unlock(mm);
1701	else
1702		mmap_read_unlock(mm);
1703
1704	return error;
1705}
1706
1707SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1708{
1709	return do_madvise(current->mm, start, len_in, behavior);
1710}
1711
1712/* Perform an madvise operation over a vector of addresses and lengths. */
1713static ssize_t vector_madvise(struct mm_struct *mm, struct iov_iter *iter,
1714			      int behavior)
1715{
1716	ssize_t ret = 0;
1717	size_t total_len;
1718
1719	total_len = iov_iter_count(iter);
1720
1721	while (iov_iter_count(iter)) {
1722		ret = do_madvise(mm, (unsigned long)iter_iov_addr(iter),
1723				 iter_iov_len(iter), behavior);
1724		/*
1725		 * An madvise operation is attempting to restart the syscall,
1726		 * but we cannot proceed as it would not be correct to repeat
1727		 * the operation in aggregate, and would be surprising to the
1728		 * user.
1729		 *
1730		 * As we have already dropped locks, it is safe to just loop and
1731		 * try again. We check for fatal signals in case we need exit
1732		 * early anyway.
1733		 */
1734		if (ret == -ERESTARTNOINTR) {
1735			if (fatal_signal_pending(current)) {
1736				ret = -EINTR;
1737				break;
1738			}
1739			continue;
1740		}
1741		if (ret < 0)
1742			break;
1743		iov_iter_advance(iter, iter_iov_len(iter));
1744	}
1745
1746	ret = (total_len - iov_iter_count(iter)) ? : ret;
1747
1748	return ret;
1749}
1750
1751SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1752		size_t, vlen, int, behavior, unsigned int, flags)
1753{
1754	ssize_t ret;
1755	struct iovec iovstack[UIO_FASTIOV];
1756	struct iovec *iov = iovstack;
1757	struct iov_iter iter;
1758	struct task_struct *task;
1759	struct mm_struct *mm;
1760	unsigned int f_flags;
1761
1762	if (flags != 0) {
1763		ret = -EINVAL;
1764		goto out;
1765	}
1766
1767	ret = import_iovec(ITER_DEST, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1768	if (ret < 0)
1769		goto out;
1770
1771	task = pidfd_get_task(pidfd, &f_flags);
1772	if (IS_ERR(task)) {
1773		ret = PTR_ERR(task);
1774		goto free_iov;
1775	}
1776
1777	/* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1778	mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1779	if (IS_ERR(mm)) {
1780		ret = PTR_ERR(mm);
1781		goto release_task;
1782	}
1783
1784	/*
1785	 * We need only perform this check if we are attempting to manipulate a
1786	 * remote process's address space.
1787	 */
1788	if (mm != current->mm && !process_madvise_remote_valid(behavior)) {
1789		ret = -EINVAL;
1790		goto release_mm;
1791	}
1792
1793	/*
1794	 * Require CAP_SYS_NICE for influencing process performance. Note that
1795	 * only non-destructive hints are currently supported for remote
1796	 * processes.
1797	 */
1798	if (mm != current->mm && !capable(CAP_SYS_NICE)) {
1799		ret = -EPERM;
1800		goto release_mm;
1801	}
1802
1803	ret = vector_madvise(mm, &iter, behavior);
1804
1805release_mm:
1806	mmput(mm);
1807release_task:
1808	put_task_struct(task);
1809free_iov:
1810	kfree(iov);
1811out:
1812	return ret;
1813}