Linux Audio

Check our new training course

Loading...
v6.8
   1
   2// SPDX-License-Identifier: GPL-2.0-only
   3/*
   4 *  linux/mm/memory.c
   5 *
   6 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   7 */
   8
   9/*
  10 * demand-loading started 01.12.91 - seems it is high on the list of
  11 * things wanted, and it should be easy to implement. - Linus
  12 */
  13
  14/*
  15 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
  16 * pages started 02.12.91, seems to work. - Linus.
  17 *
  18 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
  19 * would have taken more than the 6M I have free, but it worked well as
  20 * far as I could see.
  21 *
  22 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
  23 */
  24
  25/*
  26 * Real VM (paging to/from disk) started 18.12.91. Much more work and
  27 * thought has to go into this. Oh, well..
  28 * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
  29 *		Found it. Everything seems to work now.
  30 * 20.12.91  -  Ok, making the swap-device changeable like the root.
  31 */
  32
  33/*
  34 * 05.04.94  -  Multi-page memory management added for v1.1.
  35 *              Idea by Alex Bligh (alex@cconcepts.co.uk)
  36 *
  37 * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
  38 *		(Gerhard.Wichert@pdb.siemens.de)
  39 *
  40 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
  41 */
  42
  43#include <linux/kernel_stat.h>
  44#include <linux/mm.h>
  45#include <linux/mm_inline.h>
  46#include <linux/sched/mm.h>
  47#include <linux/sched/coredump.h>
  48#include <linux/sched/numa_balancing.h>
  49#include <linux/sched/task.h>
  50#include <linux/hugetlb.h>
  51#include <linux/mman.h>
  52#include <linux/swap.h>
  53#include <linux/highmem.h>
  54#include <linux/pagemap.h>
  55#include <linux/memremap.h>
  56#include <linux/kmsan.h>
  57#include <linux/ksm.h>
  58#include <linux/rmap.h>
  59#include <linux/export.h>
  60#include <linux/delayacct.h>
  61#include <linux/init.h>
  62#include <linux/pfn_t.h>
  63#include <linux/writeback.h>
  64#include <linux/memcontrol.h>
  65#include <linux/mmu_notifier.h>
  66#include <linux/swapops.h>
  67#include <linux/elf.h>
  68#include <linux/gfp.h>
  69#include <linux/migrate.h>
  70#include <linux/string.h>
  71#include <linux/memory-tiers.h>
  72#include <linux/debugfs.h>
  73#include <linux/userfaultfd_k.h>
  74#include <linux/dax.h>
  75#include <linux/oom.h>
  76#include <linux/numa.h>
  77#include <linux/perf_event.h>
  78#include <linux/ptrace.h>
  79#include <linux/vmalloc.h>
  80#include <linux/sched/sysctl.h>
  81
  82#include <trace/events/kmem.h>
  83
  84#include <asm/io.h>
  85#include <asm/mmu_context.h>
  86#include <asm/pgalloc.h>
  87#include <linux/uaccess.h>
  88#include <asm/tlb.h>
  89#include <asm/tlbflush.h>
 
  90
  91#include "pgalloc-track.h"
  92#include "internal.h"
  93#include "swap.h"
  94
  95#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
  96#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
  97#endif
  98
  99#ifndef CONFIG_NUMA
 
 100unsigned long max_mapnr;
 101EXPORT_SYMBOL(max_mapnr);
 102
 103struct page *mem_map;
 104EXPORT_SYMBOL(mem_map);
 105#endif
 106
 107static vm_fault_t do_fault(struct vm_fault *vmf);
 108static vm_fault_t do_anonymous_page(struct vm_fault *vmf);
 109static bool vmf_pte_changed(struct vm_fault *vmf);
 110
 111/*
 112 * Return true if the original pte was a uffd-wp pte marker (so the pte was
 113 * wr-protected).
 114 */
 115static bool vmf_orig_pte_uffd_wp(struct vm_fault *vmf)
 116{
 117	if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID))
 118		return false;
 119
 120	return pte_marker_uffd_wp(vmf->orig_pte);
 121}
 122
 123/*
 124 * A number of key systems in x86 including ioremap() rely on the assumption
 125 * that high_memory defines the upper bound on direct map memory, then end
 126 * of ZONE_NORMAL.
 
 
 127 */
 128void *high_memory;
 129EXPORT_SYMBOL(high_memory);
 130
 131/*
 132 * Randomize the address space (stacks, mmaps, brk, etc.).
 133 *
 134 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
 135 *   as ancient (libc5 based) binaries can segfault. )
 136 */
 137int randomize_va_space __read_mostly =
 138#ifdef CONFIG_COMPAT_BRK
 139					1;
 140#else
 141					2;
 142#endif
 143
 144#ifndef arch_wants_old_prefaulted_pte
 145static inline bool arch_wants_old_prefaulted_pte(void)
 146{
 147	/*
 148	 * Transitioning a PTE from 'old' to 'young' can be expensive on
 149	 * some architectures, even if it's performed in hardware. By
 150	 * default, "false" means prefaulted entries will be 'young'.
 151	 */
 152	return false;
 153}
 154#endif
 155
 156static int __init disable_randmaps(char *s)
 157{
 158	randomize_va_space = 0;
 159	return 1;
 160}
 161__setup("norandmaps", disable_randmaps);
 162
 163unsigned long zero_pfn __read_mostly;
 164EXPORT_SYMBOL(zero_pfn);
 165
 166unsigned long highest_memmap_pfn __read_mostly;
 167
 168/*
 169 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
 170 */
 171static int __init init_zero_pfn(void)
 172{
 173	zero_pfn = page_to_pfn(ZERO_PAGE(0));
 174	return 0;
 175}
 176early_initcall(init_zero_pfn);
 
 
 
 177
 178void mm_trace_rss_stat(struct mm_struct *mm, int member)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 179{
 180	trace_rss_stat(mm, member);
 181}
 182
 
 
 183/*
 184 * Note: this doesn't free the actual pages themselves. That
 185 * has been handled earlier when unmapping all the memory regions.
 186 */
 187static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 188			   unsigned long addr)
 189{
 190	pgtable_t token = pmd_pgtable(*pmd);
 191	pmd_clear(pmd);
 192	pte_free_tlb(tlb, token, addr);
 193	mm_dec_nr_ptes(tlb->mm);
 194}
 195
 196static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 197				unsigned long addr, unsigned long end,
 198				unsigned long floor, unsigned long ceiling)
 199{
 200	pmd_t *pmd;
 201	unsigned long next;
 202	unsigned long start;
 203
 204	start = addr;
 205	pmd = pmd_offset(pud, addr);
 206	do {
 207		next = pmd_addr_end(addr, end);
 208		if (pmd_none_or_clear_bad(pmd))
 209			continue;
 210		free_pte_range(tlb, pmd, addr);
 211	} while (pmd++, addr = next, addr != end);
 212
 213	start &= PUD_MASK;
 214	if (start < floor)
 215		return;
 216	if (ceiling) {
 217		ceiling &= PUD_MASK;
 218		if (!ceiling)
 219			return;
 220	}
 221	if (end - 1 > ceiling - 1)
 222		return;
 223
 224	pmd = pmd_offset(pud, start);
 225	pud_clear(pud);
 226	pmd_free_tlb(tlb, pmd, start);
 227	mm_dec_nr_pmds(tlb->mm);
 228}
 229
 230static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
 231				unsigned long addr, unsigned long end,
 232				unsigned long floor, unsigned long ceiling)
 233{
 234	pud_t *pud;
 235	unsigned long next;
 236	unsigned long start;
 237
 238	start = addr;
 239	pud = pud_offset(p4d, addr);
 240	do {
 241		next = pud_addr_end(addr, end);
 242		if (pud_none_or_clear_bad(pud))
 243			continue;
 244		free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 245	} while (pud++, addr = next, addr != end);
 246
 247	start &= P4D_MASK;
 248	if (start < floor)
 249		return;
 250	if (ceiling) {
 251		ceiling &= P4D_MASK;
 252		if (!ceiling)
 253			return;
 254	}
 255	if (end - 1 > ceiling - 1)
 256		return;
 257
 258	pud = pud_offset(p4d, start);
 259	p4d_clear(p4d);
 260	pud_free_tlb(tlb, pud, start);
 261	mm_dec_nr_puds(tlb->mm);
 262}
 263
 264static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
 265				unsigned long addr, unsigned long end,
 266				unsigned long floor, unsigned long ceiling)
 267{
 268	p4d_t *p4d;
 269	unsigned long next;
 270	unsigned long start;
 271
 272	start = addr;
 273	p4d = p4d_offset(pgd, addr);
 274	do {
 275		next = p4d_addr_end(addr, end);
 276		if (p4d_none_or_clear_bad(p4d))
 277			continue;
 278		free_pud_range(tlb, p4d, addr, next, floor, ceiling);
 279	} while (p4d++, addr = next, addr != end);
 280
 281	start &= PGDIR_MASK;
 282	if (start < floor)
 283		return;
 284	if (ceiling) {
 285		ceiling &= PGDIR_MASK;
 286		if (!ceiling)
 287			return;
 288	}
 289	if (end - 1 > ceiling - 1)
 290		return;
 291
 292	p4d = p4d_offset(pgd, start);
 293	pgd_clear(pgd);
 294	p4d_free_tlb(tlb, p4d, start);
 295}
 296
 297/*
 298 * This function frees user-level page tables of a process.
 299 */
 300void free_pgd_range(struct mmu_gather *tlb,
 301			unsigned long addr, unsigned long end,
 302			unsigned long floor, unsigned long ceiling)
 303{
 304	pgd_t *pgd;
 305	unsigned long next;
 306
 307	/*
 308	 * The next few lines have given us lots of grief...
 309	 *
 310	 * Why are we testing PMD* at this top level?  Because often
 311	 * there will be no work to do at all, and we'd prefer not to
 312	 * go all the way down to the bottom just to discover that.
 313	 *
 314	 * Why all these "- 1"s?  Because 0 represents both the bottom
 315	 * of the address space and the top of it (using -1 for the
 316	 * top wouldn't help much: the masks would do the wrong thing).
 317	 * The rule is that addr 0 and floor 0 refer to the bottom of
 318	 * the address space, but end 0 and ceiling 0 refer to the top
 319	 * Comparisons need to use "end - 1" and "ceiling - 1" (though
 320	 * that end 0 case should be mythical).
 321	 *
 322	 * Wherever addr is brought up or ceiling brought down, we must
 323	 * be careful to reject "the opposite 0" before it confuses the
 324	 * subsequent tests.  But what about where end is brought down
 325	 * by PMD_SIZE below? no, end can't go down to 0 there.
 326	 *
 327	 * Whereas we round start (addr) and ceiling down, by different
 328	 * masks at different levels, in order to test whether a table
 329	 * now has no other vmas using it, so can be freed, we don't
 330	 * bother to round floor or end up - the tests don't need that.
 331	 */
 332
 333	addr &= PMD_MASK;
 334	if (addr < floor) {
 335		addr += PMD_SIZE;
 336		if (!addr)
 337			return;
 338	}
 339	if (ceiling) {
 340		ceiling &= PMD_MASK;
 341		if (!ceiling)
 342			return;
 343	}
 344	if (end - 1 > ceiling - 1)
 345		end -= PMD_SIZE;
 346	if (addr > end - 1)
 347		return;
 348	/*
 349	 * We add page table cache pages with PAGE_SIZE,
 350	 * (see pte_free_tlb()), flush the tlb if we need
 351	 */
 352	tlb_change_page_size(tlb, PAGE_SIZE);
 353	pgd = pgd_offset(tlb->mm, addr);
 354	do {
 355		next = pgd_addr_end(addr, end);
 356		if (pgd_none_or_clear_bad(pgd))
 357			continue;
 358		free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
 359	} while (pgd++, addr = next, addr != end);
 360}
 361
 362void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,
 363		   struct vm_area_struct *vma, unsigned long floor,
 364		   unsigned long ceiling, bool mm_wr_locked)
 365{
 366	do {
 
 367		unsigned long addr = vma->vm_start;
 368		struct vm_area_struct *next;
 369
 370		/*
 371		 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may
 372		 * be 0.  This will underflow and is okay.
 373		 */
 374		next = mas_find(mas, ceiling - 1);
 375		if (unlikely(xa_is_zero(next)))
 376			next = NULL;
 377
 378		/*
 379		 * Hide vma from rmap and truncate_pagecache before freeing
 380		 * pgtables
 381		 */
 382		if (mm_wr_locked)
 383			vma_start_write(vma);
 384		unlink_anon_vmas(vma);
 385		unlink_file_vma(vma);
 386
 387		if (is_vm_hugetlb_page(vma)) {
 388			hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 389				floor, next ? next->vm_start : ceiling);
 390		} else {
 391			/*
 392			 * Optimization: gather nearby vmas into one call down
 393			 */
 394			while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 395			       && !is_vm_hugetlb_page(next)) {
 396				vma = next;
 397				next = mas_find(mas, ceiling - 1);
 398				if (unlikely(xa_is_zero(next)))
 399					next = NULL;
 400				if (mm_wr_locked)
 401					vma_start_write(vma);
 402				unlink_anon_vmas(vma);
 403				unlink_file_vma(vma);
 404			}
 405			free_pgd_range(tlb, addr, vma->vm_end,
 406				floor, next ? next->vm_start : ceiling);
 407		}
 408		vma = next;
 409	} while (vma);
 410}
 411
 412void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte)
 413{
 414	spinlock_t *ptl = pmd_lock(mm, pmd);
 415
 416	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
 417		mm_inc_nr_ptes(mm);
 418		/*
 419		 * Ensure all pte setup (eg. pte page lock and page clearing) are
 420		 * visible before the pte is made visible to other CPUs by being
 421		 * put into page tables.
 422		 *
 423		 * The other side of the story is the pointer chasing in the page
 424		 * table walking code (when walking the page table without locking;
 425		 * ie. most of the time). Fortunately, these data accesses consist
 426		 * of a chain of data-dependent loads, meaning most CPUs (alpha
 427		 * being the notable exception) will already guarantee loads are
 428		 * seen in-order. See the alpha page table accessors for the
 429		 * smp_rmb() barriers in page table walking code.
 430		 */
 431		smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
 432		pmd_populate(mm, pmd, *pte);
 433		*pte = NULL;
 434	}
 435	spin_unlock(ptl);
 436}
 437
 438int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
 439{
 
 440	pgtable_t new = pte_alloc_one(mm);
 441	if (!new)
 442		return -ENOMEM;
 443
 444	pmd_install(mm, pmd, &new);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 445	if (new)
 446		pte_free(mm, new);
 447	return 0;
 448}
 449
 450int __pte_alloc_kernel(pmd_t *pmd)
 451{
 452	pte_t *new = pte_alloc_one_kernel(&init_mm);
 453	if (!new)
 454		return -ENOMEM;
 455
 
 
 456	spin_lock(&init_mm.page_table_lock);
 457	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
 458		smp_wmb(); /* See comment in pmd_install() */
 459		pmd_populate_kernel(&init_mm, pmd, new);
 460		new = NULL;
 461	}
 462	spin_unlock(&init_mm.page_table_lock);
 463	if (new)
 464		pte_free_kernel(&init_mm, new);
 465	return 0;
 466}
 467
 468static inline void init_rss_vec(int *rss)
 469{
 470	memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
 471}
 472
 473static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
 474{
 475	int i;
 476
 
 
 477	for (i = 0; i < NR_MM_COUNTERS; i++)
 478		if (rss[i])
 479			add_mm_counter(mm, i, rss[i]);
 480}
 481
 482/*
 483 * This function is called to print an error when a bad pte
 484 * is found. For example, we might have a PFN-mapped pte in
 485 * a region that doesn't allow it.
 486 *
 487 * The calling function must still handle the error.
 488 */
 489static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
 490			  pte_t pte, struct page *page)
 491{
 492	pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
 493	p4d_t *p4d = p4d_offset(pgd, addr);
 494	pud_t *pud = pud_offset(p4d, addr);
 495	pmd_t *pmd = pmd_offset(pud, addr);
 496	struct address_space *mapping;
 497	pgoff_t index;
 498	static unsigned long resume;
 499	static unsigned long nr_shown;
 500	static unsigned long nr_unshown;
 501
 502	/*
 503	 * Allow a burst of 60 reports, then keep quiet for that minute;
 504	 * or allow a steady drip of one report per second.
 505	 */
 506	if (nr_shown == 60) {
 507		if (time_before(jiffies, resume)) {
 508			nr_unshown++;
 509			return;
 510		}
 511		if (nr_unshown) {
 512			pr_alert("BUG: Bad page map: %lu messages suppressed\n",
 513				 nr_unshown);
 514			nr_unshown = 0;
 515		}
 516		nr_shown = 0;
 517	}
 518	if (nr_shown++ == 0)
 519		resume = jiffies + 60 * HZ;
 520
 521	mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
 522	index = linear_page_index(vma, addr);
 523
 524	pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
 525		 current->comm,
 526		 (long long)pte_val(pte), (long long)pmd_val(*pmd));
 527	if (page)
 528		dump_page(page, "bad pte");
 529	pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
 530		 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
 531	pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n",
 532		 vma->vm_file,
 533		 vma->vm_ops ? vma->vm_ops->fault : NULL,
 534		 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
 535		 mapping ? mapping->a_ops->read_folio : NULL);
 536	dump_stack();
 537	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 538}
 539
 540/*
 541 * vm_normal_page -- This function gets the "struct page" associated with a pte.
 542 *
 543 * "Special" mappings do not wish to be associated with a "struct page" (either
 544 * it doesn't exist, or it exists but they don't want to touch it). In this
 545 * case, NULL is returned here. "Normal" mappings do have a struct page.
 546 *
 547 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
 548 * pte bit, in which case this function is trivial. Secondly, an architecture
 549 * may not have a spare pte bit, which requires a more complicated scheme,
 550 * described below.
 551 *
 552 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
 553 * special mapping (even if there are underlying and valid "struct pages").
 554 * COWed pages of a VM_PFNMAP are always normal.
 555 *
 556 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
 557 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
 558 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
 559 * mapping will always honor the rule
 560 *
 561 *	pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 562 *
 563 * And for normal mappings this is false.
 564 *
 565 * This restricts such mappings to be a linear translation from virtual address
 566 * to pfn. To get around this restriction, we allow arbitrary mappings so long
 567 * as the vma is not a COW mapping; in that case, we know that all ptes are
 568 * special (because none can have been COWed).
 569 *
 570 *
 571 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
 572 *
 573 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
 574 * page" backing, however the difference is that _all_ pages with a struct
 575 * page (that is, those where pfn_valid is true) are refcounted and considered
 576 * normal pages by the VM. The disadvantage is that pages are refcounted
 577 * (which can be slower and simply not an option for some PFNMAP users). The
 578 * advantage is that we don't have to follow the strict linearity rule of
 579 * PFNMAP mappings in order to support COWable mappings.
 580 *
 581 */
 582struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 583			    pte_t pte)
 584{
 585	unsigned long pfn = pte_pfn(pte);
 586
 587	if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
 588		if (likely(!pte_special(pte)))
 589			goto check_pfn;
 590		if (vma->vm_ops && vma->vm_ops->find_special_page)
 591			return vma->vm_ops->find_special_page(vma, addr);
 592		if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
 593			return NULL;
 594		if (is_zero_pfn(pfn))
 595			return NULL;
 596		if (pte_devmap(pte))
 597		/*
 598		 * NOTE: New users of ZONE_DEVICE will not set pte_devmap()
 599		 * and will have refcounts incremented on their struct pages
 600		 * when they are inserted into PTEs, thus they are safe to
 601		 * return here. Legacy ZONE_DEVICE pages that set pte_devmap()
 602		 * do not have refcounts. Example of legacy ZONE_DEVICE is
 603		 * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers.
 604		 */
 605			return NULL;
 606
 607		print_bad_pte(vma, addr, pte, NULL);
 608		return NULL;
 609	}
 610
 611	/* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
 612
 613	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 614		if (vma->vm_flags & VM_MIXEDMAP) {
 615			if (!pfn_valid(pfn))
 616				return NULL;
 617			goto out;
 618		} else {
 619			unsigned long off;
 620			off = (addr - vma->vm_start) >> PAGE_SHIFT;
 621			if (pfn == vma->vm_pgoff + off)
 622				return NULL;
 623			if (!is_cow_mapping(vma->vm_flags))
 624				return NULL;
 625		}
 626	}
 627
 628	if (is_zero_pfn(pfn))
 629		return NULL;
 630
 631check_pfn:
 632	if (unlikely(pfn > highest_memmap_pfn)) {
 633		print_bad_pte(vma, addr, pte, NULL);
 634		return NULL;
 635	}
 636
 637	/*
 638	 * NOTE! We still have PageReserved() pages in the page tables.
 639	 * eg. VDSO mappings can cause them to exist.
 640	 */
 641out:
 642	return pfn_to_page(pfn);
 643}
 644
 645struct folio *vm_normal_folio(struct vm_area_struct *vma, unsigned long addr,
 646			    pte_t pte)
 647{
 648	struct page *page = vm_normal_page(vma, addr, pte);
 649
 650	if (page)
 651		return page_folio(page);
 652	return NULL;
 653}
 654
 655#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 656struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
 657				pmd_t pmd)
 658{
 659	unsigned long pfn = pmd_pfn(pmd);
 660
 661	/*
 662	 * There is no pmd_special() but there may be special pmds, e.g.
 663	 * in a direct-access (dax) mapping, so let's just replicate the
 664	 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
 665	 */
 666	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 667		if (vma->vm_flags & VM_MIXEDMAP) {
 668			if (!pfn_valid(pfn))
 669				return NULL;
 670			goto out;
 671		} else {
 672			unsigned long off;
 673			off = (addr - vma->vm_start) >> PAGE_SHIFT;
 674			if (pfn == vma->vm_pgoff + off)
 675				return NULL;
 676			if (!is_cow_mapping(vma->vm_flags))
 677				return NULL;
 678		}
 679	}
 680
 681	if (pmd_devmap(pmd))
 682		return NULL;
 683	if (is_huge_zero_pmd(pmd))
 684		return NULL;
 685	if (unlikely(pfn > highest_memmap_pfn))
 686		return NULL;
 687
 688	/*
 689	 * NOTE! We still have PageReserved() pages in the page tables.
 690	 * eg. VDSO mappings can cause them to exist.
 691	 */
 692out:
 693	return pfn_to_page(pfn);
 694}
 695
 696struct folio *vm_normal_folio_pmd(struct vm_area_struct *vma,
 697				  unsigned long addr, pmd_t pmd)
 698{
 699	struct page *page = vm_normal_page_pmd(vma, addr, pmd);
 700
 701	if (page)
 702		return page_folio(page);
 703	return NULL;
 704}
 705#endif
 706
 707static void restore_exclusive_pte(struct vm_area_struct *vma,
 708				  struct page *page, unsigned long address,
 709				  pte_t *ptep)
 710{
 711	struct folio *folio = page_folio(page);
 712	pte_t orig_pte;
 713	pte_t pte;
 714	swp_entry_t entry;
 715
 716	orig_pte = ptep_get(ptep);
 717	pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
 718	if (pte_swp_soft_dirty(orig_pte))
 719		pte = pte_mksoft_dirty(pte);
 720
 721	entry = pte_to_swp_entry(orig_pte);
 722	if (pte_swp_uffd_wp(orig_pte))
 723		pte = pte_mkuffd_wp(pte);
 724	else if (is_writable_device_exclusive_entry(entry))
 725		pte = maybe_mkwrite(pte_mkdirty(pte), vma);
 726
 727	VM_BUG_ON_FOLIO(pte_write(pte) && (!folio_test_anon(folio) &&
 728					   PageAnonExclusive(page)), folio);
 729
 730	/*
 731	 * No need to take a page reference as one was already
 732	 * created when the swap entry was made.
 733	 */
 734	if (folio_test_anon(folio))
 735		folio_add_anon_rmap_pte(folio, page, vma, address, RMAP_NONE);
 736	else
 737		/*
 738		 * Currently device exclusive access only supports anonymous
 739		 * memory so the entry shouldn't point to a filebacked page.
 740		 */
 741		WARN_ON_ONCE(1);
 742
 743	set_pte_at(vma->vm_mm, address, ptep, pte);
 744
 745	/*
 746	 * No need to invalidate - it was non-present before. However
 747	 * secondary CPUs may have mappings that need invalidating.
 748	 */
 749	update_mmu_cache(vma, address, ptep);
 750}
 751
 752/*
 753 * Tries to restore an exclusive pte if the page lock can be acquired without
 754 * sleeping.
 755 */
 756static int
 757try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
 758			unsigned long addr)
 759{
 760	swp_entry_t entry = pte_to_swp_entry(ptep_get(src_pte));
 761	struct page *page = pfn_swap_entry_to_page(entry);
 762
 763	if (trylock_page(page)) {
 764		restore_exclusive_pte(vma, page, addr, src_pte);
 765		unlock_page(page);
 766		return 0;
 767	}
 768
 769	return -EBUSY;
 770}
 771
 772/*
 773 * copy one vm_area from one task to the other. Assumes the page tables
 774 * already present in the new task to be cleared in the whole range
 775 * covered by this vma.
 776 */
 777
 778static unsigned long
 779copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 780		pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
 781		struct vm_area_struct *src_vma, unsigned long addr, int *rss)
 782{
 783	unsigned long vm_flags = dst_vma->vm_flags;
 784	pte_t orig_pte = ptep_get(src_pte);
 785	pte_t pte = orig_pte;
 786	struct folio *folio;
 787	struct page *page;
 788	swp_entry_t entry = pte_to_swp_entry(orig_pte);
 789
 790	if (likely(!non_swap_entry(entry))) {
 791		if (swap_duplicate(entry) < 0)
 792			return -EIO;
 793
 794		/* make sure dst_mm is on swapoff's mmlist. */
 795		if (unlikely(list_empty(&dst_mm->mmlist))) {
 796			spin_lock(&mmlist_lock);
 797			if (list_empty(&dst_mm->mmlist))
 798				list_add(&dst_mm->mmlist,
 799						&src_mm->mmlist);
 800			spin_unlock(&mmlist_lock);
 801		}
 802		/* Mark the swap entry as shared. */
 803		if (pte_swp_exclusive(orig_pte)) {
 804			pte = pte_swp_clear_exclusive(orig_pte);
 805			set_pte_at(src_mm, addr, src_pte, pte);
 806		}
 807		rss[MM_SWAPENTS]++;
 808	} else if (is_migration_entry(entry)) {
 809		page = pfn_swap_entry_to_page(entry);
 
 810
 811		rss[mm_counter(page)]++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 812
 813		if (!is_readable_migration_entry(entry) &&
 814				is_cow_mapping(vm_flags)) {
 815			/*
 816			 * COW mappings require pages in both parent and child
 817			 * to be set to read. A previously exclusive entry is
 818			 * now shared.
 
 
 
 
 819			 */
 820			entry = make_readable_migration_entry(
 821							swp_offset(entry));
 822			pte = swp_entry_to_pte(entry);
 823			if (pte_swp_soft_dirty(orig_pte))
 824				pte = pte_swp_mksoft_dirty(pte);
 825			if (pte_swp_uffd_wp(orig_pte))
 826				pte = pte_swp_mkuffd_wp(pte);
 827			set_pte_at(src_mm, addr, src_pte, pte);
 828		}
 829	} else if (is_device_private_entry(entry)) {
 830		page = pfn_swap_entry_to_page(entry);
 831		folio = page_folio(page);
 832
 833		/*
 834		 * Update rss count even for unaddressable pages, as
 835		 * they should treated just like normal pages in this
 836		 * respect.
 837		 *
 838		 * We will likely want to have some new rss counters
 839		 * for unaddressable pages, at some point. But for now
 840		 * keep things as they are.
 841		 */
 842		folio_get(folio);
 843		rss[mm_counter(page)]++;
 844		/* Cannot fail as these pages cannot get pinned. */
 845		folio_try_dup_anon_rmap_pte(folio, page, src_vma);
 846
 847		/*
 848		 * We do not preserve soft-dirty information, because so
 849		 * far, checkpoint/restore is the only feature that
 850		 * requires that. And checkpoint/restore does not work
 851		 * when a device driver is involved (you cannot easily
 852		 * save and restore device driver state).
 853		 */
 854		if (is_writable_device_private_entry(entry) &&
 855		    is_cow_mapping(vm_flags)) {
 856			entry = make_readable_device_private_entry(
 857							swp_offset(entry));
 858			pte = swp_entry_to_pte(entry);
 859			if (pte_swp_uffd_wp(orig_pte))
 860				pte = pte_swp_mkuffd_wp(pte);
 861			set_pte_at(src_mm, addr, src_pte, pte);
 862		}
 863	} else if (is_device_exclusive_entry(entry)) {
 864		/*
 865		 * Make device exclusive entries present by restoring the
 866		 * original entry then copying as for a present pte. Device
 867		 * exclusive entries currently only support private writable
 868		 * (ie. COW) mappings.
 869		 */
 870		VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
 871		if (try_restore_exclusive_pte(src_pte, src_vma, addr))
 872			return -EBUSY;
 873		return -ENOENT;
 874	} else if (is_pte_marker_entry(entry)) {
 875		pte_marker marker = copy_pte_marker(entry, dst_vma);
 876
 877		if (marker)
 878			set_pte_at(dst_mm, addr, dst_pte,
 879				   make_pte_marker(marker));
 880		return 0;
 881	}
 882	if (!userfaultfd_wp(dst_vma))
 883		pte = pte_swp_clear_uffd_wp(pte);
 884	set_pte_at(dst_mm, addr, dst_pte, pte);
 885	return 0;
 886}
 887
 888/*
 889 * Copy a present and normal page.
 890 *
 891 * NOTE! The usual case is that this isn't required;
 892 * instead, the caller can just increase the page refcount
 893 * and re-use the pte the traditional way.
 894 *
 895 * And if we need a pre-allocated page but don't yet have
 896 * one, return a negative error to let the preallocation
 897 * code know so that it can do so outside the page table
 898 * lock.
 899 */
 900static inline int
 901copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
 902		  pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
 903		  struct folio **prealloc, struct page *page)
 904{
 905	struct folio *new_folio;
 906	pte_t pte;
 907
 908	new_folio = *prealloc;
 909	if (!new_folio)
 910		return -EAGAIN;
 911
 912	/*
 913	 * We have a prealloc page, all good!  Take it
 914	 * over and copy the page & arm it.
 915	 */
 916	*prealloc = NULL;
 917	copy_user_highpage(&new_folio->page, page, addr, src_vma);
 918	__folio_mark_uptodate(new_folio);
 919	folio_add_new_anon_rmap(new_folio, dst_vma, addr);
 920	folio_add_lru_vma(new_folio, dst_vma);
 921	rss[MM_ANONPAGES]++;
 922
 923	/* All done, just insert the new page copy in the child */
 924	pte = mk_pte(&new_folio->page, dst_vma->vm_page_prot);
 925	pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
 926	if (userfaultfd_pte_wp(dst_vma, ptep_get(src_pte)))
 927		/* Uffd-wp needs to be delivered to dest pte as well */
 928		pte = pte_mkuffd_wp(pte);
 929	set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
 930	return 0;
 931}
 932
 933/*
 934 * Copy one pte.  Returns 0 if succeeded, or -EAGAIN if one preallocated page
 935 * is required to copy this pte.
 936 */
 937static inline int
 938copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
 939		 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
 940		 struct folio **prealloc)
 941{
 942	struct mm_struct *src_mm = src_vma->vm_mm;
 943	unsigned long vm_flags = src_vma->vm_flags;
 944	pte_t pte = ptep_get(src_pte);
 945	struct page *page;
 946	struct folio *folio;
 947
 948	page = vm_normal_page(src_vma, addr, pte);
 949	if (page)
 950		folio = page_folio(page);
 951	if (page && folio_test_anon(folio)) {
 952		/*
 953		 * If this page may have been pinned by the parent process,
 954		 * copy the page immediately for the child so that we'll always
 955		 * guarantee the pinned page won't be randomly replaced in the
 956		 * future.
 957		 */
 958		folio_get(folio);
 959		if (unlikely(folio_try_dup_anon_rmap_pte(folio, page, src_vma))) {
 960			/* Page may be pinned, we have to copy. */
 961			folio_put(folio);
 962			return copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
 963						 addr, rss, prealloc, page);
 964		}
 965		rss[MM_ANONPAGES]++;
 966	} else if (page) {
 967		folio_get(folio);
 968		folio_dup_file_rmap_pte(folio, page);
 969		rss[mm_counter_file(page)]++;
 970	}
 971
 972	/*
 973	 * If it's a COW mapping, write protect it both
 974	 * in the parent and the child
 975	 */
 976	if (is_cow_mapping(vm_flags) && pte_write(pte)) {
 977		ptep_set_wrprotect(src_mm, addr, src_pte);
 978		pte = pte_wrprotect(pte);
 979	}
 980	VM_BUG_ON(page && folio_test_anon(folio) && PageAnonExclusive(page));
 981
 982	/*
 983	 * If it's a shared mapping, mark it clean in
 984	 * the child
 985	 */
 986	if (vm_flags & VM_SHARED)
 987		pte = pte_mkclean(pte);
 988	pte = pte_mkold(pte);
 989
 990	if (!userfaultfd_wp(dst_vma))
 991		pte = pte_clear_uffd_wp(pte);
 992
 993	set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
 994	return 0;
 995}
 996
 997static inline struct folio *folio_prealloc(struct mm_struct *src_mm,
 998		struct vm_area_struct *vma, unsigned long addr, bool need_zero)
 999{
1000	struct folio *new_folio;
1001
1002	if (need_zero)
1003		new_folio = vma_alloc_zeroed_movable_folio(vma, addr);
1004	else
1005		new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma,
1006					    addr, false);
1007
1008	if (!new_folio)
1009		return NULL;
1010
1011	if (mem_cgroup_charge(new_folio, src_mm, GFP_KERNEL)) {
1012		folio_put(new_folio);
1013		return NULL;
1014	}
1015	folio_throttle_swaprate(new_folio, GFP_KERNEL);
1016
1017	return new_folio;
 
 
1018}
1019
1020static int
1021copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1022	       pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1023	       unsigned long end)
1024{
1025	struct mm_struct *dst_mm = dst_vma->vm_mm;
1026	struct mm_struct *src_mm = src_vma->vm_mm;
1027	pte_t *orig_src_pte, *orig_dst_pte;
1028	pte_t *src_pte, *dst_pte;
1029	pte_t ptent;
1030	spinlock_t *src_ptl, *dst_ptl;
1031	int progress, ret = 0;
1032	int rss[NR_MM_COUNTERS];
1033	swp_entry_t entry = (swp_entry_t){0};
1034	struct folio *prealloc = NULL;
1035
1036again:
1037	progress = 0;
1038	init_rss_vec(rss);
1039
1040	/*
1041	 * copy_pmd_range()'s prior pmd_none_or_clear_bad(src_pmd), and the
1042	 * error handling here, assume that exclusive mmap_lock on dst and src
1043	 * protects anon from unexpected THP transitions; with shmem and file
1044	 * protected by mmap_lock-less collapse skipping areas with anon_vma
1045	 * (whereas vma_needs_copy() skips areas without anon_vma).  A rework
1046	 * can remove such assumptions later, but this is good enough for now.
1047	 */
1048	dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1049	if (!dst_pte) {
1050		ret = -ENOMEM;
1051		goto out;
1052	}
1053	src_pte = pte_offset_map_nolock(src_mm, src_pmd, addr, &src_ptl);
1054	if (!src_pte) {
1055		pte_unmap_unlock(dst_pte, dst_ptl);
1056		/* ret == 0 */
1057		goto out;
1058	}
1059	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1060	orig_src_pte = src_pte;
1061	orig_dst_pte = dst_pte;
1062	arch_enter_lazy_mmu_mode();
1063
1064	do {
1065		/*
1066		 * We are holding two locks at this point - either of them
1067		 * could generate latencies in another task on another CPU.
1068		 */
1069		if (progress >= 32) {
1070			progress = 0;
1071			if (need_resched() ||
1072			    spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1073				break;
1074		}
1075		ptent = ptep_get(src_pte);
1076		if (pte_none(ptent)) {
1077			progress++;
1078			continue;
1079		}
1080		if (unlikely(!pte_present(ptent))) {
1081			ret = copy_nonpresent_pte(dst_mm, src_mm,
1082						  dst_pte, src_pte,
1083						  dst_vma, src_vma,
1084						  addr, rss);
1085			if (ret == -EIO) {
1086				entry = pte_to_swp_entry(ptep_get(src_pte));
1087				break;
1088			} else if (ret == -EBUSY) {
1089				break;
1090			} else if (!ret) {
1091				progress += 8;
1092				continue;
1093			}
1094
1095			/*
1096			 * Device exclusive entry restored, continue by copying
1097			 * the now present pte.
1098			 */
1099			WARN_ON_ONCE(ret != -ENOENT);
1100		}
1101		/* copy_present_pte() will clear `*prealloc' if consumed */
1102		ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
1103				       addr, rss, &prealloc);
1104		/*
1105		 * If we need a pre-allocated page for this pte, drop the
1106		 * locks, allocate, and try again.
1107		 */
1108		if (unlikely(ret == -EAGAIN))
1109			break;
1110		if (unlikely(prealloc)) {
1111			/*
1112			 * pre-alloc page cannot be reused by next time so as
1113			 * to strictly follow mempolicy (e.g., alloc_page_vma()
1114			 * will allocate page according to address).  This
1115			 * could only happen if one pinned pte changed.
1116			 */
1117			folio_put(prealloc);
1118			prealloc = NULL;
1119		}
1120		progress += 8;
1121	} while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1122
1123	arch_leave_lazy_mmu_mode();
1124	pte_unmap_unlock(orig_src_pte, src_ptl);
 
1125	add_mm_rss_vec(dst_mm, rss);
1126	pte_unmap_unlock(orig_dst_pte, dst_ptl);
1127	cond_resched();
1128
1129	if (ret == -EIO) {
1130		VM_WARN_ON_ONCE(!entry.val);
1131		if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1132			ret = -ENOMEM;
1133			goto out;
1134		}
1135		entry.val = 0;
1136	} else if (ret == -EBUSY) {
1137		goto out;
1138	} else if (ret ==  -EAGAIN) {
1139		prealloc = folio_prealloc(src_mm, src_vma, addr, false);
1140		if (!prealloc)
1141			return -ENOMEM;
1142	} else if (ret) {
1143		VM_WARN_ON_ONCE(1);
1144	}
1145
1146	/* We've captured and resolved the error. Reset, try again. */
1147	ret = 0;
1148
1149	if (addr != end)
1150		goto again;
1151out:
1152	if (unlikely(prealloc))
1153		folio_put(prealloc);
1154	return ret;
1155}
1156
1157static inline int
1158copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1159	       pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1160	       unsigned long end)
1161{
1162	struct mm_struct *dst_mm = dst_vma->vm_mm;
1163	struct mm_struct *src_mm = src_vma->vm_mm;
1164	pmd_t *src_pmd, *dst_pmd;
1165	unsigned long next;
1166
1167	dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1168	if (!dst_pmd)
1169		return -ENOMEM;
1170	src_pmd = pmd_offset(src_pud, addr);
1171	do {
1172		next = pmd_addr_end(addr, end);
1173		if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1174			|| pmd_devmap(*src_pmd)) {
1175			int err;
1176			VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1177			err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1178					    addr, dst_vma, src_vma);
1179			if (err == -ENOMEM)
1180				return -ENOMEM;
1181			if (!err)
1182				continue;
1183			/* fall through */
1184		}
1185		if (pmd_none_or_clear_bad(src_pmd))
1186			continue;
1187		if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1188				   addr, next))
1189			return -ENOMEM;
1190	} while (dst_pmd++, src_pmd++, addr = next, addr != end);
1191	return 0;
1192}
1193
1194static inline int
1195copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1196	       p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1197	       unsigned long end)
1198{
1199	struct mm_struct *dst_mm = dst_vma->vm_mm;
1200	struct mm_struct *src_mm = src_vma->vm_mm;
1201	pud_t *src_pud, *dst_pud;
1202	unsigned long next;
1203
1204	dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1205	if (!dst_pud)
1206		return -ENOMEM;
1207	src_pud = pud_offset(src_p4d, addr);
1208	do {
1209		next = pud_addr_end(addr, end);
1210		if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1211			int err;
1212
1213			VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1214			err = copy_huge_pud(dst_mm, src_mm,
1215					    dst_pud, src_pud, addr, src_vma);
1216			if (err == -ENOMEM)
1217				return -ENOMEM;
1218			if (!err)
1219				continue;
1220			/* fall through */
1221		}
1222		if (pud_none_or_clear_bad(src_pud))
1223			continue;
1224		if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1225				   addr, next))
1226			return -ENOMEM;
1227	} while (dst_pud++, src_pud++, addr = next, addr != end);
1228	return 0;
1229}
1230
1231static inline int
1232copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1233	       pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1234	       unsigned long end)
1235{
1236	struct mm_struct *dst_mm = dst_vma->vm_mm;
1237	p4d_t *src_p4d, *dst_p4d;
1238	unsigned long next;
1239
1240	dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1241	if (!dst_p4d)
1242		return -ENOMEM;
1243	src_p4d = p4d_offset(src_pgd, addr);
1244	do {
1245		next = p4d_addr_end(addr, end);
1246		if (p4d_none_or_clear_bad(src_p4d))
1247			continue;
1248		if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1249				   addr, next))
1250			return -ENOMEM;
1251	} while (dst_p4d++, src_p4d++, addr = next, addr != end);
1252	return 0;
1253}
1254
1255/*
1256 * Return true if the vma needs to copy the pgtable during this fork().  Return
1257 * false when we can speed up fork() by allowing lazy page faults later until
1258 * when the child accesses the memory range.
1259 */
1260static bool
1261vma_needs_copy(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1262{
1263	/*
1264	 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1265	 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1266	 * contains uffd-wp protection information, that's something we can't
1267	 * retrieve from page cache, and skip copying will lose those info.
1268	 */
1269	if (userfaultfd_wp(dst_vma))
1270		return true;
1271
1272	if (src_vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
1273		return true;
1274
1275	if (src_vma->anon_vma)
1276		return true;
1277
1278	/*
1279	 * Don't copy ptes where a page fault will fill them correctly.  Fork
1280	 * becomes much lighter when there are big shared or private readonly
1281	 * mappings. The tradeoff is that copy_page_range is more efficient
1282	 * than faulting.
1283	 */
1284	return false;
1285}
1286
1287int
1288copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1289{
1290	pgd_t *src_pgd, *dst_pgd;
1291	unsigned long next;
1292	unsigned long addr = src_vma->vm_start;
1293	unsigned long end = src_vma->vm_end;
1294	struct mm_struct *dst_mm = dst_vma->vm_mm;
1295	struct mm_struct *src_mm = src_vma->vm_mm;
1296	struct mmu_notifier_range range;
1297	bool is_cow;
1298	int ret;
1299
1300	if (!vma_needs_copy(dst_vma, src_vma))
 
 
 
 
 
 
 
1301		return 0;
1302
1303	if (is_vm_hugetlb_page(src_vma))
1304		return copy_hugetlb_page_range(dst_mm, src_mm, dst_vma, src_vma);
1305
1306	if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1307		/*
1308		 * We do not free on error cases below as remove_vma
1309		 * gets called on error from higher level routine
1310		 */
1311		ret = track_pfn_copy(src_vma);
1312		if (ret)
1313			return ret;
1314	}
1315
1316	/*
1317	 * We need to invalidate the secondary MMU mappings only when
1318	 * there could be a permission downgrade on the ptes of the
1319	 * parent mm. And a permission downgrade will only happen if
1320	 * is_cow_mapping() returns true.
1321	 */
1322	is_cow = is_cow_mapping(src_vma->vm_flags);
1323
1324	if (is_cow) {
1325		mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1326					0, src_mm, addr, end);
1327		mmu_notifier_invalidate_range_start(&range);
1328		/*
1329		 * Disabling preemption is not needed for the write side, as
1330		 * the read side doesn't spin, but goes to the mmap_lock.
1331		 *
1332		 * Use the raw variant of the seqcount_t write API to avoid
1333		 * lockdep complaining about preemptibility.
1334		 */
1335		vma_assert_write_locked(src_vma);
1336		raw_write_seqcount_begin(&src_mm->write_protect_seq);
1337	}
1338
1339	ret = 0;
1340	dst_pgd = pgd_offset(dst_mm, addr);
1341	src_pgd = pgd_offset(src_mm, addr);
1342	do {
1343		next = pgd_addr_end(addr, end);
1344		if (pgd_none_or_clear_bad(src_pgd))
1345			continue;
1346		if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1347					    addr, next))) {
1348			untrack_pfn_clear(dst_vma);
1349			ret = -ENOMEM;
1350			break;
1351		}
1352	} while (dst_pgd++, src_pgd++, addr = next, addr != end);
1353
1354	if (is_cow) {
1355		raw_write_seqcount_end(&src_mm->write_protect_seq);
1356		mmu_notifier_invalidate_range_end(&range);
1357	}
1358	return ret;
1359}
1360
1361/* Whether we should zap all COWed (private) pages too */
1362static inline bool should_zap_cows(struct zap_details *details)
1363{
1364	/* By default, zap all pages */
1365	if (!details)
1366		return true;
1367
1368	/* Or, we zap COWed pages only if the caller wants to */
1369	return details->even_cows;
1370}
1371
1372/* Decides whether we should zap this page with the page pointer specified */
1373static inline bool should_zap_page(struct zap_details *details, struct page *page)
1374{
1375	/* If we can make a decision without *page.. */
1376	if (should_zap_cows(details))
1377		return true;
1378
1379	/* E.g. the caller passes NULL for the case of a zero page */
1380	if (!page)
1381		return true;
1382
1383	/* Otherwise we should only zap non-anon pages */
1384	return !PageAnon(page);
1385}
1386
1387static inline bool zap_drop_file_uffd_wp(struct zap_details *details)
1388{
1389	if (!details)
1390		return false;
1391
1392	return details->zap_flags & ZAP_FLAG_DROP_MARKER;
1393}
1394
1395/*
1396 * This function makes sure that we'll replace the none pte with an uffd-wp
1397 * swap special pte marker when necessary. Must be with the pgtable lock held.
1398 */
1399static inline void
1400zap_install_uffd_wp_if_needed(struct vm_area_struct *vma,
1401			      unsigned long addr, pte_t *pte,
1402			      struct zap_details *details, pte_t pteval)
1403{
1404	/* Zap on anonymous always means dropping everything */
1405	if (vma_is_anonymous(vma))
1406		return;
1407
1408	if (zap_drop_file_uffd_wp(details))
1409		return;
1410
1411	pte_install_uffd_wp_if_needed(vma, addr, pte, pteval);
1412}
1413
1414static unsigned long zap_pte_range(struct mmu_gather *tlb,
1415				struct vm_area_struct *vma, pmd_t *pmd,
1416				unsigned long addr, unsigned long end,
1417				struct zap_details *details)
1418{
1419	struct mm_struct *mm = tlb->mm;
1420	int force_flush = 0;
1421	int rss[NR_MM_COUNTERS];
1422	spinlock_t *ptl;
1423	pte_t *start_pte;
1424	pte_t *pte;
1425	swp_entry_t entry;
1426
1427	tlb_change_page_size(tlb, PAGE_SIZE);
 
1428	init_rss_vec(rss);
1429	start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1430	if (!pte)
1431		return addr;
1432
1433	flush_tlb_batched_pending(mm);
1434	arch_enter_lazy_mmu_mode();
1435	do {
1436		pte_t ptent = ptep_get(pte);
1437		struct folio *folio;
1438		struct page *page;
1439
1440		if (pte_none(ptent))
1441			continue;
1442
1443		if (need_resched())
1444			break;
1445
1446		if (pte_present(ptent)) {
1447			unsigned int delay_rmap;
1448
1449			page = vm_normal_page(vma, addr, ptent);
1450			if (unlikely(!should_zap_page(details, page)))
1451				continue;
 
 
 
 
 
 
 
 
1452			ptent = ptep_get_and_clear_full(mm, addr, pte,
1453							tlb->fullmm);
1454			arch_check_zapped_pte(vma, ptent);
1455			tlb_remove_tlb_entry(tlb, pte, addr);
1456			zap_install_uffd_wp_if_needed(vma, addr, pte, details,
1457						      ptent);
1458			if (unlikely(!page)) {
1459				ksm_might_unmap_zero_page(mm, ptent);
1460				continue;
1461			}
1462
1463			folio = page_folio(page);
1464			delay_rmap = 0;
1465			if (!folio_test_anon(folio)) {
1466				if (pte_dirty(ptent)) {
1467					folio_mark_dirty(folio);
1468					if (tlb_delay_rmap(tlb)) {
1469						delay_rmap = 1;
1470						force_flush = 1;
1471					}
1472				}
1473				if (pte_young(ptent) && likely(vma_has_recency(vma)))
1474					folio_mark_accessed(folio);
 
1475			}
1476			rss[mm_counter(page)]--;
1477			if (!delay_rmap) {
1478				folio_remove_rmap_pte(folio, page, vma);
1479				if (unlikely(page_mapcount(page) < 0))
1480					print_bad_pte(vma, addr, ptent, page);
1481			}
1482			if (unlikely(__tlb_remove_page(tlb, page, delay_rmap))) {
1483				force_flush = 1;
1484				addr += PAGE_SIZE;
1485				break;
1486			}
1487			continue;
1488		}
1489
1490		entry = pte_to_swp_entry(ptent);
1491		if (is_device_private_entry(entry) ||
1492		    is_device_exclusive_entry(entry)) {
1493			page = pfn_swap_entry_to_page(entry);
1494			folio = page_folio(page);
1495			if (unlikely(!should_zap_page(details, page)))
1496				continue;
1497			/*
1498			 * Both device private/exclusive mappings should only
1499			 * work with anonymous page so far, so we don't need to
1500			 * consider uffd-wp bit when zap. For more information,
1501			 * see zap_install_uffd_wp_if_needed().
1502			 */
1503			WARN_ON_ONCE(!vma_is_anonymous(vma));
 
 
1504			rss[mm_counter(page)]--;
1505			if (is_device_private_entry(entry))
1506				folio_remove_rmap_pte(folio, page, vma);
1507			folio_put(folio);
1508		} else if (!non_swap_entry(entry)) {
1509			/* Genuine swap entry, hence a private anon page */
1510			if (!should_zap_cows(details))
1511				continue;
 
 
 
1512			rss[MM_SWAPENTS]--;
1513			if (unlikely(!free_swap_and_cache(entry)))
1514				print_bad_pte(vma, addr, ptent, NULL);
1515		} else if (is_migration_entry(entry)) {
1516			page = pfn_swap_entry_to_page(entry);
1517			if (!should_zap_page(details, page))
1518				continue;
1519			rss[mm_counter(page)]--;
1520		} else if (pte_marker_entry_uffd_wp(entry)) {
1521			/*
1522			 * For anon: always drop the marker; for file: only
1523			 * drop the marker if explicitly requested.
1524			 */
1525			if (!vma_is_anonymous(vma) &&
1526			    !zap_drop_file_uffd_wp(details))
1527				continue;
1528		} else if (is_hwpoison_entry(entry) ||
1529			   is_poisoned_swp_entry(entry)) {
1530			if (!should_zap_cows(details))
1531				continue;
1532		} else {
1533			/* We should have covered all the swap entry types */
1534			pr_alert("unrecognized swap entry 0x%lx\n", entry.val);
1535			WARN_ON_ONCE(1);
1536		}
 
 
1537		pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1538		zap_install_uffd_wp_if_needed(vma, addr, pte, details, ptent);
1539	} while (pte++, addr += PAGE_SIZE, addr != end);
1540
1541	add_mm_rss_vec(mm, rss);
1542	arch_leave_lazy_mmu_mode();
1543
1544	/* Do the actual TLB flush before dropping ptl */
1545	if (force_flush) {
1546		tlb_flush_mmu_tlbonly(tlb);
1547		tlb_flush_rmaps(tlb, vma);
1548	}
1549	pte_unmap_unlock(start_pte, ptl);
1550
1551	/*
1552	 * If we forced a TLB flush (either due to running out of
1553	 * batch buffers or because we needed to flush dirty TLB
1554	 * entries before releasing the ptl), free the batched
1555	 * memory too. Come back again if we didn't do everything.
1556	 */
1557	if (force_flush)
 
1558		tlb_flush_mmu(tlb);
 
 
 
 
 
 
1559
1560	return addr;
1561}
1562
1563static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1564				struct vm_area_struct *vma, pud_t *pud,
1565				unsigned long addr, unsigned long end,
1566				struct zap_details *details)
1567{
1568	pmd_t *pmd;
1569	unsigned long next;
1570
1571	pmd = pmd_offset(pud, addr);
1572	do {
1573		next = pmd_addr_end(addr, end);
1574		if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1575			if (next - addr != HPAGE_PMD_SIZE)
1576				__split_huge_pmd(vma, pmd, addr, false, NULL);
1577			else if (zap_huge_pmd(tlb, vma, pmd, addr)) {
1578				addr = next;
1579				continue;
1580			}
1581			/* fall through */
1582		} else if (details && details->single_folio &&
1583			   folio_test_pmd_mappable(details->single_folio) &&
1584			   next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1585			spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1586			/*
1587			 * Take and drop THP pmd lock so that we cannot return
1588			 * prematurely, while zap_huge_pmd() has cleared *pmd,
1589			 * but not yet decremented compound_mapcount().
1590			 */
1591			spin_unlock(ptl);
1592		}
1593		if (pmd_none(*pmd)) {
1594			addr = next;
1595			continue;
1596		}
1597		addr = zap_pte_range(tlb, vma, pmd, addr, next, details);
1598		if (addr != next)
1599			pmd--;
1600	} while (pmd++, cond_resched(), addr != end);
 
 
 
 
 
 
 
 
 
1601
1602	return addr;
1603}
1604
1605static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1606				struct vm_area_struct *vma, p4d_t *p4d,
1607				unsigned long addr, unsigned long end,
1608				struct zap_details *details)
1609{
1610	pud_t *pud;
1611	unsigned long next;
1612
1613	pud = pud_offset(p4d, addr);
1614	do {
1615		next = pud_addr_end(addr, end);
1616		if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1617			if (next - addr != HPAGE_PUD_SIZE) {
1618				mmap_assert_locked(tlb->mm);
1619				split_huge_pud(vma, pud, addr);
1620			} else if (zap_huge_pud(tlb, vma, pud, addr))
1621				goto next;
1622			/* fall through */
1623		}
1624		if (pud_none_or_clear_bad(pud))
1625			continue;
1626		next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1627next:
1628		cond_resched();
1629	} while (pud++, addr = next, addr != end);
1630
1631	return addr;
1632}
1633
1634static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1635				struct vm_area_struct *vma, pgd_t *pgd,
1636				unsigned long addr, unsigned long end,
1637				struct zap_details *details)
1638{
1639	p4d_t *p4d;
1640	unsigned long next;
1641
1642	p4d = p4d_offset(pgd, addr);
1643	do {
1644		next = p4d_addr_end(addr, end);
1645		if (p4d_none_or_clear_bad(p4d))
1646			continue;
1647		next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1648	} while (p4d++, addr = next, addr != end);
1649
1650	return addr;
1651}
1652
1653void unmap_page_range(struct mmu_gather *tlb,
1654			     struct vm_area_struct *vma,
1655			     unsigned long addr, unsigned long end,
1656			     struct zap_details *details)
1657{
1658	pgd_t *pgd;
1659	unsigned long next;
1660
1661	BUG_ON(addr >= end);
1662	tlb_start_vma(tlb, vma);
1663	pgd = pgd_offset(vma->vm_mm, addr);
1664	do {
1665		next = pgd_addr_end(addr, end);
1666		if (pgd_none_or_clear_bad(pgd))
1667			continue;
1668		next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1669	} while (pgd++, addr = next, addr != end);
1670	tlb_end_vma(tlb, vma);
1671}
1672
1673
1674static void unmap_single_vma(struct mmu_gather *tlb,
1675		struct vm_area_struct *vma, unsigned long start_addr,
1676		unsigned long end_addr,
1677		struct zap_details *details, bool mm_wr_locked)
1678{
1679	unsigned long start = max(vma->vm_start, start_addr);
1680	unsigned long end;
1681
1682	if (start >= vma->vm_end)
1683		return;
1684	end = min(vma->vm_end, end_addr);
1685	if (end <= vma->vm_start)
1686		return;
1687
1688	if (vma->vm_file)
1689		uprobe_munmap(vma, start, end);
1690
1691	if (unlikely(vma->vm_flags & VM_PFNMAP))
1692		untrack_pfn(vma, 0, 0, mm_wr_locked);
1693
1694	if (start != end) {
1695		if (unlikely(is_vm_hugetlb_page(vma))) {
1696			/*
1697			 * It is undesirable to test vma->vm_file as it
1698			 * should be non-null for valid hugetlb area.
1699			 * However, vm_file will be NULL in the error
1700			 * cleanup path of mmap_region. When
1701			 * hugetlbfs ->mmap method fails,
1702			 * mmap_region() nullifies vma->vm_file
1703			 * before calling this function to clean up.
1704			 * Since no pte has actually been setup, it is
1705			 * safe to do nothing in this case.
1706			 */
1707			if (vma->vm_file) {
1708				zap_flags_t zap_flags = details ?
1709				    details->zap_flags : 0;
1710				__unmap_hugepage_range(tlb, vma, start, end,
1711							     NULL, zap_flags);
1712			}
1713		} else
1714			unmap_page_range(tlb, vma, start, end, details);
1715	}
1716}
1717
1718/**
1719 * unmap_vmas - unmap a range of memory covered by a list of vma's
1720 * @tlb: address of the caller's struct mmu_gather
1721 * @mas: the maple state
1722 * @vma: the starting vma
1723 * @start_addr: virtual address at which to start unmapping
1724 * @end_addr: virtual address at which to end unmapping
1725 * @tree_end: The maximum index to check
1726 * @mm_wr_locked: lock flag
1727 *
1728 * Unmap all pages in the vma list.
1729 *
1730 * Only addresses between `start' and `end' will be unmapped.
1731 *
1732 * The VMA list must be sorted in ascending virtual address order.
1733 *
1734 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1735 * range after unmap_vmas() returns.  So the only responsibility here is to
1736 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1737 * drops the lock and schedules.
1738 */
1739void unmap_vmas(struct mmu_gather *tlb, struct ma_state *mas,
1740		struct vm_area_struct *vma, unsigned long start_addr,
1741		unsigned long end_addr, unsigned long tree_end,
1742		bool mm_wr_locked)
1743{
1744	struct mmu_notifier_range range;
1745	struct zap_details details = {
1746		.zap_flags = ZAP_FLAG_DROP_MARKER | ZAP_FLAG_UNMAP,
1747		/* Careful - we need to zap private pages too! */
1748		.even_cows = true,
1749	};
1750
1751	mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma->vm_mm,
1752				start_addr, end_addr);
1753	mmu_notifier_invalidate_range_start(&range);
1754	do {
1755		unsigned long start = start_addr;
1756		unsigned long end = end_addr;
1757		hugetlb_zap_begin(vma, &start, &end);
1758		unmap_single_vma(tlb, vma, start, end, &details,
1759				 mm_wr_locked);
1760		hugetlb_zap_end(vma, &details);
1761		vma = mas_find(mas, tree_end - 1);
1762	} while (vma && likely(!xa_is_zero(vma)));
1763	mmu_notifier_invalidate_range_end(&range);
1764}
1765
1766/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1767 * zap_page_range_single - remove user pages in a given range
1768 * @vma: vm_area_struct holding the applicable pages
1769 * @address: starting address of pages to zap
1770 * @size: number of bytes to zap
1771 * @details: details of shared cache invalidation
1772 *
1773 * The range must fit into one VMA.
1774 */
1775void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1776		unsigned long size, struct zap_details *details)
1777{
1778	const unsigned long end = address + size;
1779	struct mmu_notifier_range range;
1780	struct mmu_gather tlb;
1781
1782	lru_add_drain();
1783	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1784				address, end);
1785	hugetlb_zap_begin(vma, &range.start, &range.end);
1786	tlb_gather_mmu(&tlb, vma->vm_mm);
1787	update_hiwater_rss(vma->vm_mm);
1788	mmu_notifier_invalidate_range_start(&range);
1789	/*
1790	 * unmap 'address-end' not 'range.start-range.end' as range
1791	 * could have been expanded for hugetlb pmd sharing.
1792	 */
1793	unmap_single_vma(&tlb, vma, address, end, details, false);
1794	mmu_notifier_invalidate_range_end(&range);
1795	tlb_finish_mmu(&tlb);
1796	hugetlb_zap_end(vma, details);
1797}
1798
1799/**
1800 * zap_vma_ptes - remove ptes mapping the vma
1801 * @vma: vm_area_struct holding ptes to be zapped
1802 * @address: starting address of pages to zap
1803 * @size: number of bytes to zap
1804 *
1805 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1806 *
1807 * The entire address range must be fully contained within the vma.
1808 *
1809 */
1810void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1811		unsigned long size)
1812{
1813	if (!range_in_vma(vma, address, address + size) ||
1814	    		!(vma->vm_flags & VM_PFNMAP))
1815		return;
1816
1817	zap_page_range_single(vma, address, size, NULL);
1818}
1819EXPORT_SYMBOL_GPL(zap_vma_ptes);
1820
1821static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
 
1822{
1823	pgd_t *pgd;
1824	p4d_t *p4d;
1825	pud_t *pud;
1826	pmd_t *pmd;
1827
1828	pgd = pgd_offset(mm, addr);
1829	p4d = p4d_alloc(mm, pgd, addr);
1830	if (!p4d)
1831		return NULL;
1832	pud = pud_alloc(mm, p4d, addr);
1833	if (!pud)
1834		return NULL;
1835	pmd = pmd_alloc(mm, pud, addr);
1836	if (!pmd)
1837		return NULL;
1838
1839	VM_BUG_ON(pmd_trans_huge(*pmd));
1840	return pmd;
1841}
1842
1843pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1844			spinlock_t **ptl)
1845{
1846	pmd_t *pmd = walk_to_pmd(mm, addr);
1847
1848	if (!pmd)
1849		return NULL;
1850	return pte_alloc_map_lock(mm, pmd, addr, ptl);
1851}
1852
1853static int validate_page_before_insert(struct page *page)
1854{
1855	struct folio *folio = page_folio(page);
1856
1857	if (folio_test_anon(folio) || folio_test_slab(folio) ||
1858	    page_has_type(page))
1859		return -EINVAL;
1860	flush_dcache_folio(folio);
1861	return 0;
1862}
1863
1864static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte,
1865			unsigned long addr, struct page *page, pgprot_t prot)
1866{
1867	struct folio *folio = page_folio(page);
1868
1869	if (!pte_none(ptep_get(pte)))
1870		return -EBUSY;
1871	/* Ok, finally just insert the thing.. */
1872	folio_get(folio);
1873	inc_mm_counter(vma->vm_mm, mm_counter_file(page));
1874	folio_add_file_rmap_pte(folio, page, vma);
1875	set_pte_at(vma->vm_mm, addr, pte, mk_pte(page, prot));
1876	return 0;
1877}
1878
1879/*
1880 * This is the old fallback for page remapping.
1881 *
1882 * For historical reasons, it only allows reserved pages. Only
1883 * old drivers should use this, and they needed to mark their
1884 * pages reserved for the old functions anyway.
1885 */
1886static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1887			struct page *page, pgprot_t prot)
1888{
 
1889	int retval;
1890	pte_t *pte;
1891	spinlock_t *ptl;
1892
1893	retval = validate_page_before_insert(page);
1894	if (retval)
1895		goto out;
1896	retval = -ENOMEM;
1897	pte = get_locked_pte(vma->vm_mm, addr, &ptl);
 
1898	if (!pte)
1899		goto out;
1900	retval = insert_page_into_pte_locked(vma, pte, addr, page, prot);
1901	pte_unmap_unlock(pte, ptl);
1902out:
1903	return retval;
1904}
1905
1906static int insert_page_in_batch_locked(struct vm_area_struct *vma, pte_t *pte,
1907			unsigned long addr, struct page *page, pgprot_t prot)
1908{
1909	int err;
1910
1911	if (!page_count(page))
1912		return -EINVAL;
1913	err = validate_page_before_insert(page);
1914	if (err)
1915		return err;
1916	return insert_page_into_pte_locked(vma, pte, addr, page, prot);
1917}
1918
1919/* insert_pages() amortizes the cost of spinlock operations
1920 * when inserting pages in a loop.
1921 */
1922static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1923			struct page **pages, unsigned long *num, pgprot_t prot)
1924{
1925	pmd_t *pmd = NULL;
1926	pte_t *start_pte, *pte;
1927	spinlock_t *pte_lock;
1928	struct mm_struct *const mm = vma->vm_mm;
1929	unsigned long curr_page_idx = 0;
1930	unsigned long remaining_pages_total = *num;
1931	unsigned long pages_to_write_in_pmd;
1932	int ret;
1933more:
1934	ret = -EFAULT;
1935	pmd = walk_to_pmd(mm, addr);
1936	if (!pmd)
1937		goto out;
1938
1939	pages_to_write_in_pmd = min_t(unsigned long,
1940		remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1941
1942	/* Allocate the PTE if necessary; takes PMD lock once only. */
1943	ret = -ENOMEM;
1944	if (pte_alloc(mm, pmd))
1945		goto out;
 
1946
1947	while (pages_to_write_in_pmd) {
1948		int pte_idx = 0;
1949		const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1950
1951		start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1952		if (!start_pte) {
1953			ret = -EFAULT;
1954			goto out;
1955		}
1956		for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1957			int err = insert_page_in_batch_locked(vma, pte,
1958				addr, pages[curr_page_idx], prot);
1959			if (unlikely(err)) {
1960				pte_unmap_unlock(start_pte, pte_lock);
1961				ret = err;
1962				remaining_pages_total -= pte_idx;
1963				goto out;
1964			}
1965			addr += PAGE_SIZE;
1966			++curr_page_idx;
1967		}
1968		pte_unmap_unlock(start_pte, pte_lock);
1969		pages_to_write_in_pmd -= batch_size;
1970		remaining_pages_total -= batch_size;
1971	}
1972	if (remaining_pages_total)
1973		goto more;
1974	ret = 0;
1975out:
1976	*num = remaining_pages_total;
1977	return ret;
1978}
1979
1980/**
1981 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1982 * @vma: user vma to map to
1983 * @addr: target start user address of these pages
1984 * @pages: source kernel pages
1985 * @num: in: number of pages to map. out: number of pages that were *not*
1986 * mapped. (0 means all pages were successfully mapped).
1987 *
1988 * Preferred over vm_insert_page() when inserting multiple pages.
1989 *
1990 * In case of error, we may have mapped a subset of the provided
1991 * pages. It is the caller's responsibility to account for this case.
1992 *
1993 * The same restrictions apply as in vm_insert_page().
1994 */
1995int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1996			struct page **pages, unsigned long *num)
1997{
1998	const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1999
2000	if (addr < vma->vm_start || end_addr >= vma->vm_end)
2001		return -EFAULT;
2002	if (!(vma->vm_flags & VM_MIXEDMAP)) {
2003		BUG_ON(mmap_read_trylock(vma->vm_mm));
2004		BUG_ON(vma->vm_flags & VM_PFNMAP);
2005		vm_flags_set(vma, VM_MIXEDMAP);
2006	}
2007	/* Defer page refcount checking till we're about to map that page. */
2008	return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
2009}
2010EXPORT_SYMBOL(vm_insert_pages);
2011
2012/**
2013 * vm_insert_page - insert single page into user vma
2014 * @vma: user vma to map to
2015 * @addr: target user address of this page
2016 * @page: source kernel page
2017 *
2018 * This allows drivers to insert individual pages they've allocated
2019 * into a user vma.
2020 *
2021 * The page has to be a nice clean _individual_ kernel allocation.
2022 * If you allocate a compound page, you need to have marked it as
2023 * such (__GFP_COMP), or manually just split the page up yourself
2024 * (see split_page()).
2025 *
2026 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2027 * took an arbitrary page protection parameter. This doesn't allow
2028 * that. Your vma protection will have to be set up correctly, which
2029 * means that if you want a shared writable mapping, you'd better
2030 * ask for a shared writable mapping!
2031 *
2032 * The page does not need to be reserved.
2033 *
2034 * Usually this function is called from f_op->mmap() handler
2035 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2036 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2037 * function from other places, for example from page-fault handler.
2038 *
2039 * Return: %0 on success, negative error code otherwise.
2040 */
2041int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2042			struct page *page)
2043{
2044	if (addr < vma->vm_start || addr >= vma->vm_end)
2045		return -EFAULT;
2046	if (!page_count(page))
2047		return -EINVAL;
2048	if (!(vma->vm_flags & VM_MIXEDMAP)) {
2049		BUG_ON(mmap_read_trylock(vma->vm_mm));
2050		BUG_ON(vma->vm_flags & VM_PFNMAP);
2051		vm_flags_set(vma, VM_MIXEDMAP);
2052	}
2053	return insert_page(vma, addr, page, vma->vm_page_prot);
2054}
2055EXPORT_SYMBOL(vm_insert_page);
2056
2057/*
2058 * __vm_map_pages - maps range of kernel pages into user vma
2059 * @vma: user vma to map to
2060 * @pages: pointer to array of source kernel pages
2061 * @num: number of pages in page array
2062 * @offset: user's requested vm_pgoff
2063 *
2064 * This allows drivers to map range of kernel pages into a user vma.
2065 *
2066 * Return: 0 on success and error code otherwise.
2067 */
2068static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2069				unsigned long num, unsigned long offset)
2070{
2071	unsigned long count = vma_pages(vma);
2072	unsigned long uaddr = vma->vm_start;
2073	int ret, i;
2074
2075	/* Fail if the user requested offset is beyond the end of the object */
2076	if (offset >= num)
2077		return -ENXIO;
2078
2079	/* Fail if the user requested size exceeds available object size */
2080	if (count > num - offset)
2081		return -ENXIO;
2082
2083	for (i = 0; i < count; i++) {
2084		ret = vm_insert_page(vma, uaddr, pages[offset + i]);
2085		if (ret < 0)
2086			return ret;
2087		uaddr += PAGE_SIZE;
2088	}
2089
2090	return 0;
2091}
2092
2093/**
2094 * vm_map_pages - maps range of kernel pages starts with non zero offset
2095 * @vma: user vma to map to
2096 * @pages: pointer to array of source kernel pages
2097 * @num: number of pages in page array
2098 *
2099 * Maps an object consisting of @num pages, catering for the user's
2100 * requested vm_pgoff
2101 *
2102 * If we fail to insert any page into the vma, the function will return
2103 * immediately leaving any previously inserted pages present.  Callers
2104 * from the mmap handler may immediately return the error as their caller
2105 * will destroy the vma, removing any successfully inserted pages. Other
2106 * callers should make their own arrangements for calling unmap_region().
2107 *
2108 * Context: Process context. Called by mmap handlers.
2109 * Return: 0 on success and error code otherwise.
2110 */
2111int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2112				unsigned long num)
2113{
2114	return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
2115}
2116EXPORT_SYMBOL(vm_map_pages);
2117
2118/**
2119 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2120 * @vma: user vma to map to
2121 * @pages: pointer to array of source kernel pages
2122 * @num: number of pages in page array
2123 *
2124 * Similar to vm_map_pages(), except that it explicitly sets the offset
2125 * to 0. This function is intended for the drivers that did not consider
2126 * vm_pgoff.
2127 *
2128 * Context: Process context. Called by mmap handlers.
2129 * Return: 0 on success and error code otherwise.
2130 */
2131int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2132				unsigned long num)
2133{
2134	return __vm_map_pages(vma, pages, num, 0);
2135}
2136EXPORT_SYMBOL(vm_map_pages_zero);
2137
2138static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2139			pfn_t pfn, pgprot_t prot, bool mkwrite)
2140{
2141	struct mm_struct *mm = vma->vm_mm;
2142	pte_t *pte, entry;
2143	spinlock_t *ptl;
2144
2145	pte = get_locked_pte(mm, addr, &ptl);
2146	if (!pte)
2147		return VM_FAULT_OOM;
2148	entry = ptep_get(pte);
2149	if (!pte_none(entry)) {
2150		if (mkwrite) {
2151			/*
2152			 * For read faults on private mappings the PFN passed
2153			 * in may not match the PFN we have mapped if the
2154			 * mapped PFN is a writeable COW page.  In the mkwrite
2155			 * case we are creating a writable PTE for a shared
2156			 * mapping and we expect the PFNs to match. If they
2157			 * don't match, we are likely racing with block
2158			 * allocation and mapping invalidation so just skip the
2159			 * update.
2160			 */
2161			if (pte_pfn(entry) != pfn_t_to_pfn(pfn)) {
2162				WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry)));
2163				goto out_unlock;
2164			}
2165			entry = pte_mkyoung(entry);
2166			entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2167			if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2168				update_mmu_cache(vma, addr, pte);
2169		}
2170		goto out_unlock;
2171	}
2172
2173	/* Ok, finally just insert the thing.. */
2174	if (pfn_t_devmap(pfn))
2175		entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2176	else
2177		entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2178
2179	if (mkwrite) {
2180		entry = pte_mkyoung(entry);
2181		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2182	}
2183
2184	set_pte_at(mm, addr, pte, entry);
2185	update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2186
2187out_unlock:
2188	pte_unmap_unlock(pte, ptl);
2189	return VM_FAULT_NOPAGE;
2190}
2191
2192/**
2193 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2194 * @vma: user vma to map to
2195 * @addr: target user address of this page
2196 * @pfn: source kernel pfn
2197 * @pgprot: pgprot flags for the inserted page
2198 *
2199 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2200 * to override pgprot on a per-page basis.
2201 *
2202 * This only makes sense for IO mappings, and it makes no sense for
2203 * COW mappings.  In general, using multiple vmas is preferable;
2204 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2205 * impractical.
2206 *
2207 * pgprot typically only differs from @vma->vm_page_prot when drivers set
2208 * caching- and encryption bits different than those of @vma->vm_page_prot,
2209 * because the caching- or encryption mode may not be known at mmap() time.
2210 *
2211 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2212 * to set caching and encryption bits for those vmas (except for COW pages).
2213 * This is ensured by core vm only modifying these page table entries using
2214 * functions that don't touch caching- or encryption bits, using pte_modify()
2215 * if needed. (See for example mprotect()).
2216 *
2217 * Also when new page-table entries are created, this is only done using the
2218 * fault() callback, and never using the value of vma->vm_page_prot,
2219 * except for page-table entries that point to anonymous pages as the result
2220 * of COW.
2221 *
2222 * Context: Process context.  May allocate using %GFP_KERNEL.
2223 * Return: vm_fault_t value.
2224 */
2225vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2226			unsigned long pfn, pgprot_t pgprot)
2227{
2228	/*
2229	 * Technically, architectures with pte_special can avoid all these
2230	 * restrictions (same for remap_pfn_range).  However we would like
2231	 * consistency in testing and feature parity among all, so we should
2232	 * try to keep these invariants in place for everybody.
2233	 */
2234	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2235	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2236						(VM_PFNMAP|VM_MIXEDMAP));
2237	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2238	BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2239
2240	if (addr < vma->vm_start || addr >= vma->vm_end)
2241		return VM_FAULT_SIGBUS;
2242
2243	if (!pfn_modify_allowed(pfn, pgprot))
2244		return VM_FAULT_SIGBUS;
2245
2246	track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2247
2248	return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2249			false);
2250}
2251EXPORT_SYMBOL(vmf_insert_pfn_prot);
2252
2253/**
2254 * vmf_insert_pfn - insert single pfn into user vma
2255 * @vma: user vma to map to
2256 * @addr: target user address of this page
2257 * @pfn: source kernel pfn
2258 *
2259 * Similar to vm_insert_page, this allows drivers to insert individual pages
2260 * they've allocated into a user vma. Same comments apply.
2261 *
2262 * This function should only be called from a vm_ops->fault handler, and
2263 * in that case the handler should return the result of this function.
2264 *
2265 * vma cannot be a COW mapping.
2266 *
2267 * As this is called only for pages that do not currently exist, we
2268 * do not need to flush old virtual caches or the TLB.
2269 *
2270 * Context: Process context.  May allocate using %GFP_KERNEL.
2271 * Return: vm_fault_t value.
2272 */
2273vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2274			unsigned long pfn)
2275{
2276	return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2277}
2278EXPORT_SYMBOL(vmf_insert_pfn);
2279
2280static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2281{
2282	/* these checks mirror the abort conditions in vm_normal_page */
2283	if (vma->vm_flags & VM_MIXEDMAP)
2284		return true;
2285	if (pfn_t_devmap(pfn))
2286		return true;
2287	if (pfn_t_special(pfn))
2288		return true;
2289	if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2290		return true;
2291	return false;
2292}
2293
2294static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2295		unsigned long addr, pfn_t pfn, bool mkwrite)
2296{
2297	pgprot_t pgprot = vma->vm_page_prot;
2298	int err;
2299
2300	BUG_ON(!vm_mixed_ok(vma, pfn));
2301
2302	if (addr < vma->vm_start || addr >= vma->vm_end)
2303		return VM_FAULT_SIGBUS;
2304
2305	track_pfn_insert(vma, &pgprot, pfn);
2306
2307	if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2308		return VM_FAULT_SIGBUS;
2309
2310	/*
2311	 * If we don't have pte special, then we have to use the pfn_valid()
2312	 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2313	 * refcount the page if pfn_valid is true (hence insert_page rather
2314	 * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2315	 * without pte special, it would there be refcounted as a normal page.
2316	 */
2317	if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2318	    !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2319		struct page *page;
2320
2321		/*
2322		 * At this point we are committed to insert_page()
2323		 * regardless of whether the caller specified flags that
2324		 * result in pfn_t_has_page() == false.
2325		 */
2326		page = pfn_to_page(pfn_t_to_pfn(pfn));
2327		err = insert_page(vma, addr, page, pgprot);
2328	} else {
2329		return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2330	}
2331
2332	if (err == -ENOMEM)
2333		return VM_FAULT_OOM;
2334	if (err < 0 && err != -EBUSY)
2335		return VM_FAULT_SIGBUS;
2336
2337	return VM_FAULT_NOPAGE;
2338}
2339
2340vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2341		pfn_t pfn)
2342{
2343	return __vm_insert_mixed(vma, addr, pfn, false);
2344}
2345EXPORT_SYMBOL(vmf_insert_mixed);
2346
2347/*
2348 *  If the insertion of PTE failed because someone else already added a
2349 *  different entry in the mean time, we treat that as success as we assume
2350 *  the same entry was actually inserted.
2351 */
2352vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2353		unsigned long addr, pfn_t pfn)
2354{
2355	return __vm_insert_mixed(vma, addr, pfn, true);
2356}
2357EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2358
2359/*
2360 * maps a range of physical memory into the requested pages. the old
2361 * mappings are removed. any references to nonexistent pages results
2362 * in null mappings (currently treated as "copy-on-access")
2363 */
2364static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2365			unsigned long addr, unsigned long end,
2366			unsigned long pfn, pgprot_t prot)
2367{
2368	pte_t *pte, *mapped_pte;
2369	spinlock_t *ptl;
2370	int err = 0;
2371
2372	mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2373	if (!pte)
2374		return -ENOMEM;
2375	arch_enter_lazy_mmu_mode();
2376	do {
2377		BUG_ON(!pte_none(ptep_get(pte)));
2378		if (!pfn_modify_allowed(pfn, prot)) {
2379			err = -EACCES;
2380			break;
2381		}
2382		set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2383		pfn++;
2384	} while (pte++, addr += PAGE_SIZE, addr != end);
2385	arch_leave_lazy_mmu_mode();
2386	pte_unmap_unlock(mapped_pte, ptl);
2387	return err;
2388}
2389
2390static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2391			unsigned long addr, unsigned long end,
2392			unsigned long pfn, pgprot_t prot)
2393{
2394	pmd_t *pmd;
2395	unsigned long next;
2396	int err;
2397
2398	pfn -= addr >> PAGE_SHIFT;
2399	pmd = pmd_alloc(mm, pud, addr);
2400	if (!pmd)
2401		return -ENOMEM;
2402	VM_BUG_ON(pmd_trans_huge(*pmd));
2403	do {
2404		next = pmd_addr_end(addr, end);
2405		err = remap_pte_range(mm, pmd, addr, next,
2406				pfn + (addr >> PAGE_SHIFT), prot);
2407		if (err)
2408			return err;
2409	} while (pmd++, addr = next, addr != end);
2410	return 0;
2411}
2412
2413static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2414			unsigned long addr, unsigned long end,
2415			unsigned long pfn, pgprot_t prot)
2416{
2417	pud_t *pud;
2418	unsigned long next;
2419	int err;
2420
2421	pfn -= addr >> PAGE_SHIFT;
2422	pud = pud_alloc(mm, p4d, addr);
2423	if (!pud)
2424		return -ENOMEM;
2425	do {
2426		next = pud_addr_end(addr, end);
2427		err = remap_pmd_range(mm, pud, addr, next,
2428				pfn + (addr >> PAGE_SHIFT), prot);
2429		if (err)
2430			return err;
2431	} while (pud++, addr = next, addr != end);
2432	return 0;
2433}
2434
2435static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2436			unsigned long addr, unsigned long end,
2437			unsigned long pfn, pgprot_t prot)
2438{
2439	p4d_t *p4d;
2440	unsigned long next;
2441	int err;
2442
2443	pfn -= addr >> PAGE_SHIFT;
2444	p4d = p4d_alloc(mm, pgd, addr);
2445	if (!p4d)
2446		return -ENOMEM;
2447	do {
2448		next = p4d_addr_end(addr, end);
2449		err = remap_pud_range(mm, p4d, addr, next,
2450				pfn + (addr >> PAGE_SHIFT), prot);
2451		if (err)
2452			return err;
2453	} while (p4d++, addr = next, addr != end);
2454	return 0;
2455}
2456
2457/*
2458 * Variant of remap_pfn_range that does not call track_pfn_remap.  The caller
2459 * must have pre-validated the caching bits of the pgprot_t.
 
 
 
 
 
 
 
 
2460 */
2461int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2462		unsigned long pfn, unsigned long size, pgprot_t prot)
2463{
2464	pgd_t *pgd;
2465	unsigned long next;
2466	unsigned long end = addr + PAGE_ALIGN(size);
2467	struct mm_struct *mm = vma->vm_mm;
 
2468	int err;
2469
2470	if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2471		return -EINVAL;
2472
2473	/*
2474	 * Physically remapped pages are special. Tell the
2475	 * rest of the world about it:
2476	 *   VM_IO tells people not to look at these pages
2477	 *	(accesses can have side effects).
2478	 *   VM_PFNMAP tells the core MM that the base pages are just
2479	 *	raw PFN mappings, and do not have a "struct page" associated
2480	 *	with them.
2481	 *   VM_DONTEXPAND
2482	 *      Disable vma merging and expanding with mremap().
2483	 *   VM_DONTDUMP
2484	 *      Omit vma from core dump, even when VM_IO turned off.
2485	 *
2486	 * There's a horrible special case to handle copy-on-write
2487	 * behaviour that some programs depend on. We mark the "original"
2488	 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2489	 * See vm_normal_page() for details.
2490	 */
2491	if (is_cow_mapping(vma->vm_flags)) {
2492		if (addr != vma->vm_start || end != vma->vm_end)
2493			return -EINVAL;
2494		vma->vm_pgoff = pfn;
2495	}
2496
2497	vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
 
 
 
 
2498
2499	BUG_ON(addr >= end);
2500	pfn -= addr >> PAGE_SHIFT;
2501	pgd = pgd_offset(mm, addr);
2502	flush_cache_range(vma, addr, end);
2503	do {
2504		next = pgd_addr_end(addr, end);
2505		err = remap_p4d_range(mm, pgd, addr, next,
2506				pfn + (addr >> PAGE_SHIFT), prot);
2507		if (err)
2508			return err;
2509	} while (pgd++, addr = next, addr != end);
2510
2511	return 0;
2512}
2513
2514/**
2515 * remap_pfn_range - remap kernel memory to userspace
2516 * @vma: user vma to map to
2517 * @addr: target page aligned user address to start at
2518 * @pfn: page frame number of kernel physical memory address
2519 * @size: size of mapping area
2520 * @prot: page protection flags for this mapping
2521 *
2522 * Note: this is only safe if the mm semaphore is held when called.
2523 *
2524 * Return: %0 on success, negative error code otherwise.
2525 */
2526int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2527		    unsigned long pfn, unsigned long size, pgprot_t prot)
2528{
2529	int err;
2530
2531	err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2532	if (err)
2533		return -EINVAL;
2534
2535	err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2536	if (err)
2537		untrack_pfn(vma, pfn, PAGE_ALIGN(size), true);
2538	return err;
2539}
2540EXPORT_SYMBOL(remap_pfn_range);
2541
2542/**
2543 * vm_iomap_memory - remap memory to userspace
2544 * @vma: user vma to map to
2545 * @start: start of the physical memory to be mapped
2546 * @len: size of area
2547 *
2548 * This is a simplified io_remap_pfn_range() for common driver use. The
2549 * driver just needs to give us the physical memory range to be mapped,
2550 * we'll figure out the rest from the vma information.
2551 *
2552 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2553 * whatever write-combining details or similar.
2554 *
2555 * Return: %0 on success, negative error code otherwise.
2556 */
2557int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2558{
2559	unsigned long vm_len, pfn, pages;
2560
2561	/* Check that the physical memory area passed in looks valid */
2562	if (start + len < start)
2563		return -EINVAL;
2564	/*
2565	 * You *really* shouldn't map things that aren't page-aligned,
2566	 * but we've historically allowed it because IO memory might
2567	 * just have smaller alignment.
2568	 */
2569	len += start & ~PAGE_MASK;
2570	pfn = start >> PAGE_SHIFT;
2571	pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2572	if (pfn + pages < pfn)
2573		return -EINVAL;
2574
2575	/* We start the mapping 'vm_pgoff' pages into the area */
2576	if (vma->vm_pgoff > pages)
2577		return -EINVAL;
2578	pfn += vma->vm_pgoff;
2579	pages -= vma->vm_pgoff;
2580
2581	/* Can we fit all of the mapping? */
2582	vm_len = vma->vm_end - vma->vm_start;
2583	if (vm_len >> PAGE_SHIFT > pages)
2584		return -EINVAL;
2585
2586	/* Ok, let it rip */
2587	return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2588}
2589EXPORT_SYMBOL(vm_iomap_memory);
2590
2591static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2592				     unsigned long addr, unsigned long end,
2593				     pte_fn_t fn, void *data, bool create,
2594				     pgtbl_mod_mask *mask)
2595{
2596	pte_t *pte, *mapped_pte;
2597	int err = 0;
2598	spinlock_t *ptl;
2599
2600	if (create) {
2601		mapped_pte = pte = (mm == &init_mm) ?
2602			pte_alloc_kernel_track(pmd, addr, mask) :
2603			pte_alloc_map_lock(mm, pmd, addr, &ptl);
2604		if (!pte)
2605			return -ENOMEM;
2606	} else {
2607		mapped_pte = pte = (mm == &init_mm) ?
2608			pte_offset_kernel(pmd, addr) :
2609			pte_offset_map_lock(mm, pmd, addr, &ptl);
2610		if (!pte)
2611			return -EINVAL;
2612	}
2613
2614	arch_enter_lazy_mmu_mode();
2615
2616	if (fn) {
2617		do {
2618			if (create || !pte_none(ptep_get(pte))) {
2619				err = fn(pte++, addr, data);
2620				if (err)
2621					break;
2622			}
2623		} while (addr += PAGE_SIZE, addr != end);
2624	}
2625	*mask |= PGTBL_PTE_MODIFIED;
2626
2627	arch_leave_lazy_mmu_mode();
2628
2629	if (mm != &init_mm)
2630		pte_unmap_unlock(mapped_pte, ptl);
2631	return err;
2632}
2633
2634static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2635				     unsigned long addr, unsigned long end,
2636				     pte_fn_t fn, void *data, bool create,
2637				     pgtbl_mod_mask *mask)
2638{
2639	pmd_t *pmd;
2640	unsigned long next;
2641	int err = 0;
2642
2643	BUG_ON(pud_huge(*pud));
2644
2645	if (create) {
2646		pmd = pmd_alloc_track(mm, pud, addr, mask);
2647		if (!pmd)
2648			return -ENOMEM;
2649	} else {
2650		pmd = pmd_offset(pud, addr);
2651	}
2652	do {
2653		next = pmd_addr_end(addr, end);
2654		if (pmd_none(*pmd) && !create)
2655			continue;
2656		if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2657			return -EINVAL;
2658		if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2659			if (!create)
2660				continue;
2661			pmd_clear_bad(pmd);
2662		}
2663		err = apply_to_pte_range(mm, pmd, addr, next,
2664					 fn, data, create, mask);
2665		if (err)
2666			break;
2667	} while (pmd++, addr = next, addr != end);
2668
2669	return err;
2670}
2671
2672static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2673				     unsigned long addr, unsigned long end,
2674				     pte_fn_t fn, void *data, bool create,
2675				     pgtbl_mod_mask *mask)
2676{
2677	pud_t *pud;
2678	unsigned long next;
2679	int err = 0;
2680
2681	if (create) {
2682		pud = pud_alloc_track(mm, p4d, addr, mask);
2683		if (!pud)
2684			return -ENOMEM;
2685	} else {
2686		pud = pud_offset(p4d, addr);
2687	}
2688	do {
2689		next = pud_addr_end(addr, end);
2690		if (pud_none(*pud) && !create)
2691			continue;
2692		if (WARN_ON_ONCE(pud_leaf(*pud)))
2693			return -EINVAL;
2694		if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2695			if (!create)
2696				continue;
2697			pud_clear_bad(pud);
2698		}
2699		err = apply_to_pmd_range(mm, pud, addr, next,
2700					 fn, data, create, mask);
2701		if (err)
2702			break;
2703	} while (pud++, addr = next, addr != end);
2704
2705	return err;
2706}
2707
2708static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2709				     unsigned long addr, unsigned long end,
2710				     pte_fn_t fn, void *data, bool create,
2711				     pgtbl_mod_mask *mask)
2712{
2713	p4d_t *p4d;
2714	unsigned long next;
2715	int err = 0;
2716
2717	if (create) {
2718		p4d = p4d_alloc_track(mm, pgd, addr, mask);
2719		if (!p4d)
2720			return -ENOMEM;
2721	} else {
2722		p4d = p4d_offset(pgd, addr);
2723	}
2724	do {
2725		next = p4d_addr_end(addr, end);
2726		if (p4d_none(*p4d) && !create)
2727			continue;
2728		if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2729			return -EINVAL;
2730		if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2731			if (!create)
2732				continue;
2733			p4d_clear_bad(p4d);
2734		}
2735		err = apply_to_pud_range(mm, p4d, addr, next,
2736					 fn, data, create, mask);
2737		if (err)
2738			break;
2739	} while (p4d++, addr = next, addr != end);
2740
2741	return err;
2742}
2743
2744static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2745				 unsigned long size, pte_fn_t fn,
2746				 void *data, bool create)
 
 
 
2747{
2748	pgd_t *pgd;
2749	unsigned long start = addr, next;
2750	unsigned long end = addr + size;
2751	pgtbl_mod_mask mask = 0;
2752	int err = 0;
2753
2754	if (WARN_ON(addr >= end))
2755		return -EINVAL;
2756
2757	pgd = pgd_offset(mm, addr);
2758	do {
2759		next = pgd_addr_end(addr, end);
2760		if (pgd_none(*pgd) && !create)
2761			continue;
2762		if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2763			return -EINVAL;
2764		if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2765			if (!create)
2766				continue;
2767			pgd_clear_bad(pgd);
2768		}
2769		err = apply_to_p4d_range(mm, pgd, addr, next,
2770					 fn, data, create, &mask);
2771		if (err)
2772			break;
2773	} while (pgd++, addr = next, addr != end);
2774
2775	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2776		arch_sync_kernel_mappings(start, start + size);
2777
2778	return err;
2779}
2780
2781/*
2782 * Scan a region of virtual memory, filling in page tables as necessary
2783 * and calling a provided function on each leaf page table.
2784 */
2785int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2786			unsigned long size, pte_fn_t fn, void *data)
2787{
2788	return __apply_to_page_range(mm, addr, size, fn, data, true);
2789}
2790EXPORT_SYMBOL_GPL(apply_to_page_range);
2791
2792/*
2793 * Scan a region of virtual memory, calling a provided function on
2794 * each leaf page table where it exists.
2795 *
2796 * Unlike apply_to_page_range, this does _not_ fill in page tables
2797 * where they are absent.
2798 */
2799int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2800				 unsigned long size, pte_fn_t fn, void *data)
2801{
2802	return __apply_to_page_range(mm, addr, size, fn, data, false);
2803}
2804EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2805
2806/*
2807 * handle_pte_fault chooses page fault handler according to an entry which was
2808 * read non-atomically.  Before making any commitment, on those architectures
2809 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2810 * parts, do_swap_page must check under lock before unmapping the pte and
2811 * proceeding (but do_wp_page is only called after already making such a check;
2812 * and do_anonymous_page can safely check later on).
2813 */
2814static inline int pte_unmap_same(struct vm_fault *vmf)
 
2815{
2816	int same = 1;
2817#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2818	if (sizeof(pte_t) > sizeof(unsigned long)) {
2819		spin_lock(vmf->ptl);
2820		same = pte_same(ptep_get(vmf->pte), vmf->orig_pte);
2821		spin_unlock(vmf->ptl);
 
2822	}
2823#endif
2824	pte_unmap(vmf->pte);
2825	vmf->pte = NULL;
2826	return same;
2827}
2828
2829/*
2830 * Return:
2831 *	0:		copied succeeded
2832 *	-EHWPOISON:	copy failed due to hwpoison in source page
2833 *	-EAGAIN:	copied failed (some other reason)
2834 */
2835static inline int __wp_page_copy_user(struct page *dst, struct page *src,
2836				      struct vm_fault *vmf)
2837{
2838	int ret;
2839	void *kaddr;
2840	void __user *uaddr;
2841	struct vm_area_struct *vma = vmf->vma;
2842	struct mm_struct *mm = vma->vm_mm;
2843	unsigned long addr = vmf->address;
2844
2845	if (likely(src)) {
2846		if (copy_mc_user_highpage(dst, src, addr, vma)) {
2847			memory_failure_queue(page_to_pfn(src), 0);
2848			return -EHWPOISON;
2849		}
2850		return 0;
2851	}
2852
2853	/*
2854	 * If the source page was a PFN mapping, we don't have
2855	 * a "struct page" for it. We do a best-effort copy by
2856	 * just copying from the original user address. If that
2857	 * fails, we just zero-fill it. Live with it.
2858	 */
2859	kaddr = kmap_local_page(dst);
2860	pagefault_disable();
2861	uaddr = (void __user *)(addr & PAGE_MASK);
2862
2863	/*
2864	 * On architectures with software "accessed" bits, we would
2865	 * take a double page fault, so mark it accessed here.
2866	 */
2867	vmf->pte = NULL;
2868	if (!arch_has_hw_pte_young() && !pte_young(vmf->orig_pte)) {
2869		pte_t entry;
2870
2871		vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2872		if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
2873			/*
2874			 * Other thread has already handled the fault
2875			 * and update local tlb only
2876			 */
2877			if (vmf->pte)
2878				update_mmu_tlb(vma, addr, vmf->pte);
2879			ret = -EAGAIN;
2880			goto pte_unlock;
2881		}
2882
2883		entry = pte_mkyoung(vmf->orig_pte);
2884		if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2885			update_mmu_cache_range(vmf, vma, addr, vmf->pte, 1);
2886	}
2887
2888	/*
2889	 * This really shouldn't fail, because the page is there
2890	 * in the page tables. But it might just be unreadable,
2891	 * in which case we just give up and fill the result with
2892	 * zeroes.
2893	 */
2894	if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2895		if (vmf->pte)
2896			goto warn;
2897
2898		/* Re-validate under PTL if the page is still mapped */
2899		vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2900		if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
2901			/* The PTE changed under us, update local tlb */
2902			if (vmf->pte)
2903				update_mmu_tlb(vma, addr, vmf->pte);
2904			ret = -EAGAIN;
2905			goto pte_unlock;
2906		}
2907
2908		/*
2909		 * The same page can be mapped back since last copy attempt.
2910		 * Try to copy again under PTL.
 
 
2911		 */
2912		if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2913			/*
2914			 * Give a warn in case there can be some obscure
2915			 * use-case
2916			 */
2917warn:
2918			WARN_ON_ONCE(1);
2919			clear_page(kaddr);
2920		}
2921	}
2922
2923	ret = 0;
2924
2925pte_unlock:
2926	if (vmf->pte)
2927		pte_unmap_unlock(vmf->pte, vmf->ptl);
2928	pagefault_enable();
2929	kunmap_local(kaddr);
2930	flush_dcache_page(dst);
2931
2932	return ret;
2933}
2934
2935static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2936{
2937	struct file *vm_file = vma->vm_file;
2938
2939	if (vm_file)
2940		return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2941
2942	/*
2943	 * Special mappings (e.g. VDSO) do not have any file so fake
2944	 * a default GFP_KERNEL for them.
2945	 */
2946	return GFP_KERNEL;
2947}
2948
2949/*
2950 * Notify the address space that the page is about to become writable so that
2951 * it can prohibit this or wait for the page to get into an appropriate state.
2952 *
2953 * We do this without the lock held, so that it can sleep if it needs to.
2954 */
2955static vm_fault_t do_page_mkwrite(struct vm_fault *vmf, struct folio *folio)
2956{
2957	vm_fault_t ret;
 
2958	unsigned int old_flags = vmf->flags;
2959
2960	vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2961
2962	if (vmf->vma->vm_file &&
2963	    IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2964		return VM_FAULT_SIGBUS;
2965
2966	ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2967	/* Restore original flags so that caller is not surprised */
2968	vmf->flags = old_flags;
2969	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2970		return ret;
2971	if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2972		folio_lock(folio);
2973		if (!folio->mapping) {
2974			folio_unlock(folio);
2975			return 0; /* retry */
2976		}
2977		ret |= VM_FAULT_LOCKED;
2978	} else
2979		VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2980	return ret;
2981}
2982
2983/*
2984 * Handle dirtying of a page in shared file mapping on a write fault.
2985 *
2986 * The function expects the page to be locked and unlocks it.
2987 */
2988static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
 
2989{
2990	struct vm_area_struct *vma = vmf->vma;
2991	struct address_space *mapping;
2992	struct folio *folio = page_folio(vmf->page);
2993	bool dirtied;
2994	bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2995
2996	dirtied = folio_mark_dirty(folio);
2997	VM_BUG_ON_FOLIO(folio_test_anon(folio), folio);
2998	/*
2999	 * Take a local copy of the address_space - folio.mapping may be zeroed
3000	 * by truncate after folio_unlock().   The address_space itself remains
3001	 * pinned by vma->vm_file's reference.  We rely on folio_unlock()'s
3002	 * release semantics to prevent the compiler from undoing this copying.
3003	 */
3004	mapping = folio_raw_mapping(folio);
3005	folio_unlock(folio);
3006
3007	if (!page_mkwrite)
3008		file_update_time(vma->vm_file);
3009
3010	/*
3011	 * Throttle page dirtying rate down to writeback speed.
3012	 *
3013	 * mapping may be NULL here because some device drivers do not
3014	 * set page.mapping but still dirty their pages
3015	 *
3016	 * Drop the mmap_lock before waiting on IO, if we can. The file
3017	 * is pinning the mapping, as per above.
3018	 */
3019	if ((dirtied || page_mkwrite) && mapping) {
3020		struct file *fpin;
3021
3022		fpin = maybe_unlock_mmap_for_io(vmf, NULL);
 
3023		balance_dirty_pages_ratelimited(mapping);
3024		if (fpin) {
3025			fput(fpin);
3026			return VM_FAULT_COMPLETED;
3027		}
3028	}
3029
3030	return 0;
 
3031}
3032
3033/*
3034 * Handle write page faults for pages that can be reused in the current vma
3035 *
3036 * This can happen either due to the mapping being with the VM_SHARED flag,
3037 * or due to us being the last reference standing to the page. In either
3038 * case, all we need to do here is to mark the page as writable and update
3039 * any related book-keeping.
3040 */
3041static inline void wp_page_reuse(struct vm_fault *vmf, struct folio *folio)
3042	__releases(vmf->ptl)
3043{
3044	struct vm_area_struct *vma = vmf->vma;
 
3045	pte_t entry;
3046
3047	VM_BUG_ON(!(vmf->flags & FAULT_FLAG_WRITE));
3048
3049	if (folio) {
3050		VM_BUG_ON(folio_test_anon(folio) &&
3051			  !PageAnonExclusive(vmf->page));
3052		/*
3053		 * Clear the folio's cpupid information as the existing
3054		 * information potentially belongs to a now completely
3055		 * unrelated process.
3056		 */
3057		folio_xchg_last_cpupid(folio, (1 << LAST_CPUPID_SHIFT) - 1);
3058	}
3059
3060	flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3061	entry = pte_mkyoung(vmf->orig_pte);
3062	entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3063	if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
3064		update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
3065	pte_unmap_unlock(vmf->pte, vmf->ptl);
3066	count_vm_event(PGREUSE);
3067}
3068
3069/*
3070 * We could add a bitflag somewhere, but for now, we know that all
3071 * vm_ops that have a ->map_pages have been audited and don't need
3072 * the mmap_lock to be held.
3073 */
3074static inline vm_fault_t vmf_can_call_fault(const struct vm_fault *vmf)
3075{
3076	struct vm_area_struct *vma = vmf->vma;
3077
3078	if (vma->vm_ops->map_pages || !(vmf->flags & FAULT_FLAG_VMA_LOCK))
3079		return 0;
3080	vma_end_read(vma);
3081	return VM_FAULT_RETRY;
3082}
3083
3084static vm_fault_t vmf_anon_prepare(struct vm_fault *vmf)
3085{
3086	struct vm_area_struct *vma = vmf->vma;
3087
3088	if (likely(vma->anon_vma))
3089		return 0;
3090	if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3091		vma_end_read(vma);
3092		return VM_FAULT_RETRY;
3093	}
3094	if (__anon_vma_prepare(vma))
3095		return VM_FAULT_OOM;
3096	return 0;
3097}
3098
3099/*
3100 * Handle the case of a page which we actually need to copy to a new page,
3101 * either due to COW or unsharing.
3102 *
3103 * Called with mmap_lock locked and the old page referenced, but
3104 * without the ptl held.
3105 *
3106 * High level logic flow:
3107 *
3108 * - Allocate a page, copy the content of the old page to the new one.
3109 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3110 * - Take the PTL. If the pte changed, bail out and release the allocated page
3111 * - If the pte is still the way we remember it, update the page table and all
3112 *   relevant references. This includes dropping the reference the page-table
3113 *   held to the old page, as well as updating the rmap.
3114 * - In any case, unlock the PTL and drop the reference we took to the old page.
3115 */
3116static vm_fault_t wp_page_copy(struct vm_fault *vmf)
3117{
3118	const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3119	struct vm_area_struct *vma = vmf->vma;
3120	struct mm_struct *mm = vma->vm_mm;
3121	struct folio *old_folio = NULL;
3122	struct folio *new_folio = NULL;
3123	pte_t entry;
3124	int page_copied = 0;
 
3125	struct mmu_notifier_range range;
3126	vm_fault_t ret;
3127	bool pfn_is_zero;
3128
3129	delayacct_wpcopy_start();
3130
3131	if (vmf->page)
3132		old_folio = page_folio(vmf->page);
3133	ret = vmf_anon_prepare(vmf);
3134	if (unlikely(ret))
3135		goto out;
3136
3137	pfn_is_zero = is_zero_pfn(pte_pfn(vmf->orig_pte));
3138	new_folio = folio_prealloc(mm, vma, vmf->address, pfn_is_zero);
3139	if (!new_folio)
3140		goto oom;
3141
3142	if (!pfn_is_zero) {
3143		int err;
3144
3145		err = __wp_page_copy_user(&new_folio->page, vmf->page, vmf);
3146		if (err) {
3147			/*
3148			 * COW failed, if the fault was solved by other,
3149			 * it's fine. If not, userspace would re-fault on
3150			 * the same address and we will handle the fault
3151			 * from the second attempt.
3152			 * The -EHWPOISON case will not be retried.
3153			 */
3154			folio_put(new_folio);
3155			if (old_folio)
3156				folio_put(old_folio);
3157
3158			delayacct_wpcopy_end();
3159			return err == -EHWPOISON ? VM_FAULT_HWPOISON : 0;
3160		}
3161		kmsan_copy_page_meta(&new_folio->page, vmf->page);
3162	}
3163
3164	__folio_mark_uptodate(new_folio);
 
 
 
3165
3166	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
3167				vmf->address & PAGE_MASK,
3168				(vmf->address & PAGE_MASK) + PAGE_SIZE);
3169	mmu_notifier_invalidate_range_start(&range);
3170
3171	/*
3172	 * Re-check the pte - we dropped the lock
3173	 */
3174	vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3175	if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3176		if (old_folio) {
3177			if (!folio_test_anon(old_folio)) {
3178				dec_mm_counter(mm, mm_counter_file(&old_folio->page));
3179				inc_mm_counter(mm, MM_ANONPAGES);
 
3180			}
3181		} else {
3182			ksm_might_unmap_zero_page(mm, vmf->orig_pte);
3183			inc_mm_counter(mm, MM_ANONPAGES);
3184		}
3185		flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3186		entry = mk_pte(&new_folio->page, vma->vm_page_prot);
3187		entry = pte_sw_mkyoung(entry);
3188		if (unlikely(unshare)) {
3189			if (pte_soft_dirty(vmf->orig_pte))
3190				entry = pte_mksoft_dirty(entry);
3191			if (pte_uffd_wp(vmf->orig_pte))
3192				entry = pte_mkuffd_wp(entry);
3193		} else {
3194			entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3195		}
3196
3197		/*
3198		 * Clear the pte entry and flush it first, before updating the
3199		 * pte with the new entry, to keep TLBs on different CPUs in
3200		 * sync. This code used to set the new PTE then flush TLBs, but
3201		 * that left a window where the new PTE could be loaded into
3202		 * some TLBs while the old PTE remains in others.
3203		 */
3204		ptep_clear_flush(vma, vmf->address, vmf->pte);
3205		folio_add_new_anon_rmap(new_folio, vma, vmf->address);
3206		folio_add_lru_vma(new_folio, vma);
 
3207		/*
3208		 * We call the notify macro here because, when using secondary
3209		 * mmu page tables (such as kvm shadow page tables), we want the
3210		 * new page to be mapped directly into the secondary page table.
3211		 */
3212		BUG_ON(unshare && pte_write(entry));
3213		set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
3214		update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
3215		if (old_folio) {
3216			/*
3217			 * Only after switching the pte to the new page may
3218			 * we remove the mapcount here. Otherwise another
3219			 * process may come and find the rmap count decremented
3220			 * before the pte is switched to the new page, and
3221			 * "reuse" the old page writing into it while our pte
3222			 * here still points into it and can be read by other
3223			 * threads.
3224			 *
3225			 * The critical issue is to order this
3226			 * folio_remove_rmap_pte() with the ptp_clear_flush
3227			 * above. Those stores are ordered by (if nothing else,)
3228			 * the barrier present in the atomic_add_negative
3229			 * in folio_remove_rmap_pte();
3230			 *
3231			 * Then the TLB flush in ptep_clear_flush ensures that
3232			 * no process can access the old page before the
3233			 * decremented mapcount is visible. And the old page
3234			 * cannot be reused until after the decremented
3235			 * mapcount is visible. So transitively, TLBs to
3236			 * old page will be flushed before it can be reused.
3237			 */
3238			folio_remove_rmap_pte(old_folio, vmf->page, vma);
3239		}
3240
3241		/* Free the old page.. */
3242		new_folio = old_folio;
3243		page_copied = 1;
3244		pte_unmap_unlock(vmf->pte, vmf->ptl);
3245	} else if (vmf->pte) {
3246		update_mmu_tlb(vma, vmf->address, vmf->pte);
3247		pte_unmap_unlock(vmf->pte, vmf->ptl);
3248	}
3249
3250	mmu_notifier_invalidate_range_end(&range);
3251
3252	if (new_folio)
3253		folio_put(new_folio);
3254	if (old_folio) {
3255		if (page_copied)
3256			free_swap_cache(&old_folio->page);
3257		folio_put(old_folio);
3258	}
3259
3260	delayacct_wpcopy_end();
3261	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3262oom:
3263	ret = VM_FAULT_OOM;
3264out:
3265	if (old_folio)
3266		folio_put(old_folio);
3267
3268	delayacct_wpcopy_end();
3269	return ret;
3270}
3271
3272/**
3273 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3274 *			  writeable once the page is prepared
3275 *
3276 * @vmf: structure describing the fault
3277 * @folio: the folio of vmf->page
3278 *
3279 * This function handles all that is needed to finish a write page fault in a
3280 * shared mapping due to PTE being read-only once the mapped page is prepared.
3281 * It handles locking of PTE and modifying it.
3282 *
3283 * The function expects the page to be locked or other protection against
3284 * concurrent faults / writeback (such as DAX radix tree locks).
3285 *
3286 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3287 * we acquired PTE lock.
3288 */
3289static vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf, struct folio *folio)
3290{
3291	WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3292	vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3293				       &vmf->ptl);
3294	if (!vmf->pte)
3295		return VM_FAULT_NOPAGE;
3296	/*
3297	 * We might have raced with another page fault while we released the
3298	 * pte_offset_map_lock.
3299	 */
3300	if (!pte_same(ptep_get(vmf->pte), vmf->orig_pte)) {
3301		update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3302		pte_unmap_unlock(vmf->pte, vmf->ptl);
3303		return VM_FAULT_NOPAGE;
3304	}
3305	wp_page_reuse(vmf, folio);
3306	return 0;
3307}
3308
3309/*
3310 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3311 * mapping
3312 */
3313static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3314{
3315	struct vm_area_struct *vma = vmf->vma;
3316
3317	if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3318		vm_fault_t ret;
3319
3320		pte_unmap_unlock(vmf->pte, vmf->ptl);
3321		ret = vmf_can_call_fault(vmf);
3322		if (ret)
3323			return ret;
3324
3325		vmf->flags |= FAULT_FLAG_MKWRITE;
3326		ret = vma->vm_ops->pfn_mkwrite(vmf);
3327		if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3328			return ret;
3329		return finish_mkwrite_fault(vmf, NULL);
3330	}
3331	wp_page_reuse(vmf, NULL);
3332	return 0;
3333}
3334
3335static vm_fault_t wp_page_shared(struct vm_fault *vmf, struct folio *folio)
3336	__releases(vmf->ptl)
3337{
3338	struct vm_area_struct *vma = vmf->vma;
3339	vm_fault_t ret = 0;
3340
3341	folio_get(folio);
3342
3343	if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3344		vm_fault_t tmp;
3345
3346		pte_unmap_unlock(vmf->pte, vmf->ptl);
3347		tmp = vmf_can_call_fault(vmf);
3348		if (tmp) {
3349			folio_put(folio);
3350			return tmp;
3351		}
3352
3353		tmp = do_page_mkwrite(vmf, folio);
3354		if (unlikely(!tmp || (tmp &
3355				      (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3356			folio_put(folio);
3357			return tmp;
3358		}
3359		tmp = finish_mkwrite_fault(vmf, folio);
3360		if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3361			folio_unlock(folio);
3362			folio_put(folio);
3363			return tmp;
3364		}
3365	} else {
3366		wp_page_reuse(vmf, folio);
3367		folio_lock(folio);
3368	}
3369	ret |= fault_dirty_shared_page(vmf);
3370	folio_put(folio);
3371
3372	return ret;
3373}
3374
3375static bool wp_can_reuse_anon_folio(struct folio *folio,
3376				    struct vm_area_struct *vma)
3377{
3378	/*
3379	 * We have to verify under folio lock: these early checks are
3380	 * just an optimization to avoid locking the folio and freeing
3381	 * the swapcache if there is little hope that we can reuse.
3382	 *
3383	 * KSM doesn't necessarily raise the folio refcount.
3384	 */
3385	if (folio_test_ksm(folio) || folio_ref_count(folio) > 3)
3386		return false;
3387	if (!folio_test_lru(folio))
3388		/*
3389		 * We cannot easily detect+handle references from
3390		 * remote LRU caches or references to LRU folios.
3391		 */
3392		lru_add_drain();
3393	if (folio_ref_count(folio) > 1 + folio_test_swapcache(folio))
3394		return false;
3395	if (!folio_trylock(folio))
3396		return false;
3397	if (folio_test_swapcache(folio))
3398		folio_free_swap(folio);
3399	if (folio_test_ksm(folio) || folio_ref_count(folio) != 1) {
3400		folio_unlock(folio);
3401		return false;
3402	}
3403	/*
3404	 * Ok, we've got the only folio reference from our mapping
3405	 * and the folio is locked, it's dark out, and we're wearing
3406	 * sunglasses. Hit it.
3407	 */
3408	folio_move_anon_rmap(folio, vma);
3409	folio_unlock(folio);
3410	return true;
3411}
3412
3413/*
3414 * This routine handles present pages, when
3415 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3416 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3417 *   (FAULT_FLAG_UNSHARE)
3418 *
3419 * It is done by copying the page to a new address and decrementing the
3420 * shared-page counter for the old page.
3421 *
3422 * Note that this routine assumes that the protection checks have been
3423 * done by the caller (the low-level page fault routine in most cases).
3424 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3425 * done any necessary COW.
3426 *
3427 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3428 * though the page will change only once the write actually happens. This
3429 * avoids a few races, and potentially makes it more efficient.
3430 *
3431 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3432 * but allow concurrent faults), with pte both mapped and locked.
3433 * We return with mmap_lock still held, but pte unmapped and unlocked.
3434 */
3435static vm_fault_t do_wp_page(struct vm_fault *vmf)
3436	__releases(vmf->ptl)
3437{
3438	const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3439	struct vm_area_struct *vma = vmf->vma;
3440	struct folio *folio = NULL;
3441	pte_t pte;
3442
3443	if (likely(!unshare)) {
3444		if (userfaultfd_pte_wp(vma, ptep_get(vmf->pte))) {
3445			if (!userfaultfd_wp_async(vma)) {
3446				pte_unmap_unlock(vmf->pte, vmf->ptl);
3447				return handle_userfault(vmf, VM_UFFD_WP);
3448			}
3449
3450			/*
3451			 * Nothing needed (cache flush, TLB invalidations,
3452			 * etc.) because we're only removing the uffd-wp bit,
3453			 * which is completely invisible to the user.
3454			 */
3455			pte = pte_clear_uffd_wp(ptep_get(vmf->pte));
3456
3457			set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3458			/*
3459			 * Update this to be prepared for following up CoW
3460			 * handling
3461			 */
3462			vmf->orig_pte = pte;
3463		}
3464
3465		/*
3466		 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3467		 * is flushed in this case before copying.
3468		 */
3469		if (unlikely(userfaultfd_wp(vmf->vma) &&
3470			     mm_tlb_flush_pending(vmf->vma->vm_mm)))
3471			flush_tlb_page(vmf->vma, vmf->address);
3472	}
3473
3474	vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3475
3476	if (vmf->page)
3477		folio = page_folio(vmf->page);
3478
3479	/*
3480	 * Shared mapping: we are guaranteed to have VM_WRITE and
3481	 * FAULT_FLAG_WRITE set at this point.
3482	 */
3483	if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
3484		/*
3485		 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3486		 * VM_PFNMAP VMA.
3487		 *
3488		 * We should not cow pages in a shared writeable mapping.
3489		 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3490		 */
3491		if (!vmf->page)
 
3492			return wp_pfn_shared(vmf);
3493		return wp_page_shared(vmf, folio);
 
 
3494	}
3495
3496	/*
3497	 * Private mapping: create an exclusive anonymous page copy if reuse
3498	 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
3499	 *
3500	 * If we encounter a page that is marked exclusive, we must reuse
3501	 * the page without further checks.
3502	 */
3503	if (folio && folio_test_anon(folio) &&
3504	    (PageAnonExclusive(vmf->page) || wp_can_reuse_anon_folio(folio, vma))) {
3505		if (!PageAnonExclusive(vmf->page))
3506			SetPageAnonExclusive(vmf->page);
3507		if (unlikely(unshare)) {
 
 
3508			pte_unmap_unlock(vmf->pte, vmf->ptl);
3509			return 0;
 
 
 
 
 
 
 
 
 
3510		}
3511		wp_page_reuse(vmf, folio);
3512		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3513	}
 
3514	/*
3515	 * Ok, we need to copy. Oh, well..
3516	 */
3517	if (folio)
3518		folio_get(folio);
3519
3520	pte_unmap_unlock(vmf->pte, vmf->ptl);
3521#ifdef CONFIG_KSM
3522	if (folio && folio_test_ksm(folio))
3523		count_vm_event(COW_KSM);
3524#endif
3525	return wp_page_copy(vmf);
3526}
3527
3528static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3529		unsigned long start_addr, unsigned long end_addr,
3530		struct zap_details *details)
3531{
3532	zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3533}
3534
3535static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3536					    pgoff_t first_index,
3537					    pgoff_t last_index,
3538					    struct zap_details *details)
3539{
3540	struct vm_area_struct *vma;
3541	pgoff_t vba, vea, zba, zea;
3542
3543	vma_interval_tree_foreach(vma, root, first_index, last_index) {
 
 
3544		vba = vma->vm_pgoff;
3545		vea = vba + vma_pages(vma) - 1;
3546		zba = max(first_index, vba);
3547		zea = min(last_index, vea);
 
 
 
 
3548
3549		unmap_mapping_range_vma(vma,
3550			((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3551			((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3552				details);
3553	}
3554}
3555
3556/**
3557 * unmap_mapping_folio() - Unmap single folio from processes.
3558 * @folio: The locked folio to be unmapped.
3559 *
3560 * Unmap this folio from any userspace process which still has it mmaped.
3561 * Typically, for efficiency, the range of nearby pages has already been
3562 * unmapped by unmap_mapping_pages() or unmap_mapping_range().  But once
3563 * truncation or invalidation holds the lock on a folio, it may find that
3564 * the page has been remapped again: and then uses unmap_mapping_folio()
3565 * to unmap it finally.
3566 */
3567void unmap_mapping_folio(struct folio *folio)
3568{
3569	struct address_space *mapping = folio->mapping;
3570	struct zap_details details = { };
3571	pgoff_t	first_index;
3572	pgoff_t	last_index;
3573
3574	VM_BUG_ON(!folio_test_locked(folio));
3575
3576	first_index = folio->index;
3577	last_index = folio_next_index(folio) - 1;
3578
3579	details.even_cows = false;
3580	details.single_folio = folio;
3581	details.zap_flags = ZAP_FLAG_DROP_MARKER;
3582
3583	i_mmap_lock_read(mapping);
3584	if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3585		unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3586					 last_index, &details);
3587	i_mmap_unlock_read(mapping);
3588}
3589
3590/**
3591 * unmap_mapping_pages() - Unmap pages from processes.
3592 * @mapping: The address space containing pages to be unmapped.
3593 * @start: Index of first page to be unmapped.
3594 * @nr: Number of pages to be unmapped.  0 to unmap to end of file.
3595 * @even_cows: Whether to unmap even private COWed pages.
3596 *
3597 * Unmap the pages in this address space from any userspace process which
3598 * has them mmaped.  Generally, you want to remove COWed pages as well when
3599 * a file is being truncated, but not when invalidating pages from the page
3600 * cache.
3601 */
3602void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3603		pgoff_t nr, bool even_cows)
3604{
3605	struct zap_details details = { };
3606	pgoff_t	first_index = start;
3607	pgoff_t	last_index = start + nr - 1;
3608
3609	details.even_cows = even_cows;
3610	if (last_index < first_index)
3611		last_index = ULONG_MAX;
 
 
3612
3613	i_mmap_lock_read(mapping);
3614	if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3615		unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3616					 last_index, &details);
3617	i_mmap_unlock_read(mapping);
3618}
3619EXPORT_SYMBOL_GPL(unmap_mapping_pages);
3620
3621/**
3622 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3623 * address_space corresponding to the specified byte range in the underlying
3624 * file.
3625 *
3626 * @mapping: the address space containing mmaps to be unmapped.
3627 * @holebegin: byte in first page to unmap, relative to the start of
3628 * the underlying file.  This will be rounded down to a PAGE_SIZE
3629 * boundary.  Note that this is different from truncate_pagecache(), which
3630 * must keep the partial page.  In contrast, we must get rid of
3631 * partial pages.
3632 * @holelen: size of prospective hole in bytes.  This will be rounded
3633 * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
3634 * end of the file.
3635 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3636 * but 0 when invalidating pagecache, don't throw away private data.
3637 */
3638void unmap_mapping_range(struct address_space *mapping,
3639		loff_t const holebegin, loff_t const holelen, int even_cows)
3640{
3641	pgoff_t hba = (pgoff_t)(holebegin) >> PAGE_SHIFT;
3642	pgoff_t hlen = ((pgoff_t)(holelen) + PAGE_SIZE - 1) >> PAGE_SHIFT;
3643
3644	/* Check for overflow. */
3645	if (sizeof(holelen) > sizeof(hlen)) {
3646		long long holeend =
3647			(holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3648		if (holeend & ~(long long)ULONG_MAX)
3649			hlen = ULONG_MAX - hba + 1;
3650	}
3651
3652	unmap_mapping_pages(mapping, hba, hlen, even_cows);
3653}
3654EXPORT_SYMBOL(unmap_mapping_range);
3655
3656/*
3657 * Restore a potential device exclusive pte to a working pte entry
3658 */
3659static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3660{
3661	struct folio *folio = page_folio(vmf->page);
3662	struct vm_area_struct *vma = vmf->vma;
3663	struct mmu_notifier_range range;
3664	vm_fault_t ret;
3665
3666	/*
3667	 * We need a reference to lock the folio because we don't hold
3668	 * the PTL so a racing thread can remove the device-exclusive
3669	 * entry and unmap it. If the folio is free the entry must
3670	 * have been removed already. If it happens to have already
3671	 * been re-allocated after being freed all we do is lock and
3672	 * unlock it.
3673	 */
3674	if (!folio_try_get(folio))
3675		return 0;
3676
3677	ret = folio_lock_or_retry(folio, vmf);
3678	if (ret) {
3679		folio_put(folio);
3680		return ret;
3681	}
3682	mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0,
3683				vma->vm_mm, vmf->address & PAGE_MASK,
3684				(vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3685	mmu_notifier_invalidate_range_start(&range);
3686
3687	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3688				&vmf->ptl);
3689	if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3690		restore_exclusive_pte(vma, vmf->page, vmf->address, vmf->pte);
3691
3692	if (vmf->pte)
3693		pte_unmap_unlock(vmf->pte, vmf->ptl);
3694	folio_unlock(folio);
3695	folio_put(folio);
3696
3697	mmu_notifier_invalidate_range_end(&range);
3698	return 0;
3699}
3700
3701static inline bool should_try_to_free_swap(struct folio *folio,
3702					   struct vm_area_struct *vma,
3703					   unsigned int fault_flags)
3704{
3705	if (!folio_test_swapcache(folio))
3706		return false;
3707	if (mem_cgroup_swap_full(folio) || (vma->vm_flags & VM_LOCKED) ||
3708	    folio_test_mlocked(folio))
3709		return true;
3710	/*
3711	 * If we want to map a page that's in the swapcache writable, we
3712	 * have to detect via the refcount if we're really the exclusive
3713	 * user. Try freeing the swapcache to get rid of the swapcache
3714	 * reference only in case it's likely that we'll be the exlusive user.
3715	 */
3716	return (fault_flags & FAULT_FLAG_WRITE) && !folio_test_ksm(folio) &&
3717		folio_ref_count(folio) == 2;
3718}
3719
3720static vm_fault_t pte_marker_clear(struct vm_fault *vmf)
3721{
3722	vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
3723				       vmf->address, &vmf->ptl);
3724	if (!vmf->pte)
3725		return 0;
3726	/*
3727	 * Be careful so that we will only recover a special uffd-wp pte into a
3728	 * none pte.  Otherwise it means the pte could have changed, so retry.
3729	 *
3730	 * This should also cover the case where e.g. the pte changed
3731	 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED.
3732	 * So is_pte_marker() check is not enough to safely drop the pte.
3733	 */
3734	if (pte_same(vmf->orig_pte, ptep_get(vmf->pte)))
3735		pte_clear(vmf->vma->vm_mm, vmf->address, vmf->pte);
3736	pte_unmap_unlock(vmf->pte, vmf->ptl);
3737	return 0;
3738}
3739
3740static vm_fault_t do_pte_missing(struct vm_fault *vmf)
3741{
3742	if (vma_is_anonymous(vmf->vma))
3743		return do_anonymous_page(vmf);
3744	else
3745		return do_fault(vmf);
3746}
3747
3748/*
3749 * This is actually a page-missing access, but with uffd-wp special pte
3750 * installed.  It means this pte was wr-protected before being unmapped.
3751 */
3752static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf)
3753{
3754	/*
3755	 * Just in case there're leftover special ptes even after the region
3756	 * got unregistered - we can simply clear them.
3757	 */
3758	if (unlikely(!userfaultfd_wp(vmf->vma)))
3759		return pte_marker_clear(vmf);
3760
3761	return do_pte_missing(vmf);
3762}
3763
3764static vm_fault_t handle_pte_marker(struct vm_fault *vmf)
3765{
3766	swp_entry_t entry = pte_to_swp_entry(vmf->orig_pte);
3767	unsigned long marker = pte_marker_get(entry);
3768
3769	/*
3770	 * PTE markers should never be empty.  If anything weird happened,
3771	 * the best thing to do is to kill the process along with its mm.
3772	 */
3773	if (WARN_ON_ONCE(!marker))
3774		return VM_FAULT_SIGBUS;
3775
3776	/* Higher priority than uffd-wp when data corrupted */
3777	if (marker & PTE_MARKER_POISONED)
3778		return VM_FAULT_HWPOISON;
3779
3780	if (pte_marker_entry_uffd_wp(entry))
3781		return pte_marker_handle_uffd_wp(vmf);
3782
3783	/* This is an unknown pte marker */
3784	return VM_FAULT_SIGBUS;
3785}
3786
3787/*
3788 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3789 * but allow concurrent faults), and pte mapped but not yet locked.
3790 * We return with pte unmapped and unlocked.
3791 *
3792 * We return with the mmap_lock locked or unlocked in the same cases
3793 * as does filemap_fault().
3794 */
3795vm_fault_t do_swap_page(struct vm_fault *vmf)
3796{
3797	struct vm_area_struct *vma = vmf->vma;
3798	struct folio *swapcache, *folio = NULL;
3799	struct page *page;
3800	struct swap_info_struct *si = NULL;
3801	rmap_t rmap_flags = RMAP_NONE;
3802	bool need_clear_cache = false;
3803	bool exclusive = false;
3804	swp_entry_t entry;
3805	pte_t pte;
 
 
3806	vm_fault_t ret = 0;
3807	void *shadow = NULL;
3808
3809	if (!pte_unmap_same(vmf))
3810		goto out;
3811
3812	entry = pte_to_swp_entry(vmf->orig_pte);
3813	if (unlikely(non_swap_entry(entry))) {
3814		if (is_migration_entry(entry)) {
3815			migration_entry_wait(vma->vm_mm, vmf->pmd,
3816					     vmf->address);
3817		} else if (is_device_exclusive_entry(entry)) {
3818			vmf->page = pfn_swap_entry_to_page(entry);
3819			ret = remove_device_exclusive_entry(vmf);
3820		} else if (is_device_private_entry(entry)) {
3821			if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3822				/*
3823				 * migrate_to_ram is not yet ready to operate
3824				 * under VMA lock.
3825				 */
3826				vma_end_read(vma);
3827				ret = VM_FAULT_RETRY;
3828				goto out;
3829			}
3830
3831			vmf->page = pfn_swap_entry_to_page(entry);
3832			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3833					vmf->address, &vmf->ptl);
3834			if (unlikely(!vmf->pte ||
3835				     !pte_same(ptep_get(vmf->pte),
3836							vmf->orig_pte)))
3837				goto unlock;
3838
3839			/*
3840			 * Get a page reference while we know the page can't be
3841			 * freed.
3842			 */
3843			get_page(vmf->page);
3844			pte_unmap_unlock(vmf->pte, vmf->ptl);
3845			ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3846			put_page(vmf->page);
3847		} else if (is_hwpoison_entry(entry)) {
3848			ret = VM_FAULT_HWPOISON;
3849		} else if (is_pte_marker_entry(entry)) {
3850			ret = handle_pte_marker(vmf);
3851		} else {
3852			print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3853			ret = VM_FAULT_SIGBUS;
3854		}
3855		goto out;
3856	}
3857
3858	/* Prevent swapoff from happening to us. */
3859	si = get_swap_device(entry);
3860	if (unlikely(!si))
3861		goto out;
3862
3863	folio = swap_cache_get_folio(entry, vma, vmf->address);
3864	if (folio)
3865		page = folio_file_page(folio, swp_offset(entry));
3866	swapcache = folio;
3867
3868	if (!folio) {
3869		if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3870		    __swap_count(entry) == 1) {
3871			/*
3872			 * Prevent parallel swapin from proceeding with
3873			 * the cache flag. Otherwise, another thread may
3874			 * finish swapin first, free the entry, and swapout
3875			 * reusing the same entry. It's undetectable as
3876			 * pte_same() returns true due to entry reuse.
3877			 */
3878			if (swapcache_prepare(entry)) {
3879				/* Relax a bit to prevent rapid repeated page faults */
3880				schedule_timeout_uninterruptible(1);
3881				goto out;
3882			}
3883			need_clear_cache = true;
3884
3885			/* skip swapcache */
3886			folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0,
3887						vma, vmf->address, false);
3888			page = &folio->page;
3889			if (folio) {
3890				__folio_set_locked(folio);
3891				__folio_set_swapbacked(folio);
3892
3893				if (mem_cgroup_swapin_charge_folio(folio,
3894							vma->vm_mm, GFP_KERNEL,
3895							entry)) {
3896					ret = VM_FAULT_OOM;
3897					goto out_page;
3898				}
3899				mem_cgroup_swapin_uncharge_swap(entry);
3900
3901				shadow = get_shadow_from_swap_cache(entry);
3902				if (shadow)
3903					workingset_refault(folio, shadow);
3904
3905				folio_add_lru(folio);
3906
3907				/* To provide entry to swap_read_folio() */
3908				folio->swap = entry;
3909				swap_read_folio(folio, true, NULL);
3910				folio->private = NULL;
 
3911			}
3912		} else {
3913			page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3914						vmf);
3915			if (page)
3916				folio = page_folio(page);
3917			swapcache = folio;
3918		}
3919
3920		if (!folio) {
3921			/*
3922			 * Back out if somebody else faulted in this pte
3923			 * while we released the pte lock.
3924			 */
3925			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3926					vmf->address, &vmf->ptl);
3927			if (likely(vmf->pte &&
3928				   pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3929				ret = VM_FAULT_OOM;
 
3930			goto unlock;
3931		}
3932
3933		/* Had to read the page from swap area: Major fault */
3934		ret = VM_FAULT_MAJOR;
3935		count_vm_event(PGMAJFAULT);
3936		count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3937	} else if (PageHWPoison(page)) {
3938		/*
3939		 * hwpoisoned dirty swapcache pages are kept for killing
3940		 * owner processes (which may be unknown at hwpoison time)
3941		 */
3942		ret = VM_FAULT_HWPOISON;
 
3943		goto out_release;
3944	}
3945
3946	ret |= folio_lock_or_retry(folio, vmf);
3947	if (ret & VM_FAULT_RETRY)
3948		goto out_release;
3949
3950	if (swapcache) {
3951		/*
3952		 * Make sure folio_free_swap() or swapoff did not release the
3953		 * swapcache from under us.  The page pin, and pte_same test
3954		 * below, are not enough to exclude that.  Even if it is still
3955		 * swapcache, we need to check that the page's swap has not
3956		 * changed.
3957		 */
3958		if (unlikely(!folio_test_swapcache(folio) ||
3959			     page_swap_entry(page).val != entry.val))
3960			goto out_page;
3961
3962		/*
3963		 * KSM sometimes has to copy on read faults, for example, if
3964		 * page->index of !PageKSM() pages would be nonlinear inside the
3965		 * anon VMA -- PageKSM() is lost on actual swapout.
3966		 */
3967		folio = ksm_might_need_to_copy(folio, vma, vmf->address);
3968		if (unlikely(!folio)) {
3969			ret = VM_FAULT_OOM;
3970			folio = swapcache;
3971			goto out_page;
3972		} else if (unlikely(folio == ERR_PTR(-EHWPOISON))) {
3973			ret = VM_FAULT_HWPOISON;
3974			folio = swapcache;
3975			goto out_page;
3976		}
3977		if (folio != swapcache)
3978			page = folio_page(folio, 0);
3979
3980		/*
3981		 * If we want to map a page that's in the swapcache writable, we
3982		 * have to detect via the refcount if we're really the exclusive
3983		 * owner. Try removing the extra reference from the local LRU
3984		 * caches if required.
3985		 */
3986		if ((vmf->flags & FAULT_FLAG_WRITE) && folio == swapcache &&
3987		    !folio_test_ksm(folio) && !folio_test_lru(folio))
3988			lru_add_drain();
3989	}
3990
3991	folio_throttle_swaprate(folio, GFP_KERNEL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3992
3993	/*
3994	 * Back out if somebody else already faulted in this pte.
3995	 */
3996	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3997			&vmf->ptl);
3998	if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3999		goto out_nomap;
4000
4001	if (unlikely(!folio_test_uptodate(folio))) {
4002		ret = VM_FAULT_SIGBUS;
4003		goto out_nomap;
4004	}
4005
4006	/*
4007	 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
4008	 * must never point at an anonymous page in the swapcache that is
4009	 * PG_anon_exclusive. Sanity check that this holds and especially, that
4010	 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
4011	 * check after taking the PT lock and making sure that nobody
4012	 * concurrently faulted in this page and set PG_anon_exclusive.
4013	 */
4014	BUG_ON(!folio_test_anon(folio) && folio_test_mappedtodisk(folio));
4015	BUG_ON(folio_test_anon(folio) && PageAnonExclusive(page));
4016
4017	/*
4018	 * Check under PT lock (to protect against concurrent fork() sharing
4019	 * the swap entry concurrently) for certainly exclusive pages.
4020	 */
4021	if (!folio_test_ksm(folio)) {
4022		exclusive = pte_swp_exclusive(vmf->orig_pte);
4023		if (folio != swapcache) {
4024			/*
4025			 * We have a fresh page that is not exposed to the
4026			 * swapcache -> certainly exclusive.
4027			 */
4028			exclusive = true;
4029		} else if (exclusive && folio_test_writeback(folio) &&
4030			  data_race(si->flags & SWP_STABLE_WRITES)) {
4031			/*
4032			 * This is tricky: not all swap backends support
4033			 * concurrent page modifications while under writeback.
4034			 *
4035			 * So if we stumble over such a page in the swapcache
4036			 * we must not set the page exclusive, otherwise we can
4037			 * map it writable without further checks and modify it
4038			 * while still under writeback.
4039			 *
4040			 * For these problematic swap backends, simply drop the
4041			 * exclusive marker: this is perfectly fine as we start
4042			 * writeback only if we fully unmapped the page and
4043			 * there are no unexpected references on the page after
4044			 * unmapping succeeded. After fully unmapped, no
4045			 * further GUP references (FOLL_GET and FOLL_PIN) can
4046			 * appear, so dropping the exclusive marker and mapping
4047			 * it only R/O is fine.
4048			 */
4049			exclusive = false;
4050		}
4051	}
4052
4053	/*
4054	 * Some architectures may have to restore extra metadata to the page
4055	 * when reading from swap. This metadata may be indexed by swap entry
4056	 * so this must be called before swap_free().
4057	 */
4058	arch_swap_restore(entry, folio);
4059
4060	/*
4061	 * Remove the swap entry and conditionally try to free up the swapcache.
4062	 * We're already holding a reference on the page but haven't mapped it
4063	 * yet.
4064	 */
4065	swap_free(entry);
4066	if (should_try_to_free_swap(folio, vma, vmf->flags))
4067		folio_free_swap(folio);
4068
4069	inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
4070	dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
4071	pte = mk_pte(page, vma->vm_page_prot);
4072
4073	/*
4074	 * Same logic as in do_wp_page(); however, optimize for pages that are
4075	 * certainly not shared either because we just allocated them without
4076	 * exposing them to the swapcache or because the swap entry indicates
4077	 * exclusivity.
4078	 */
4079	if (!folio_test_ksm(folio) &&
4080	    (exclusive || folio_ref_count(folio) == 1)) {
4081		if (vmf->flags & FAULT_FLAG_WRITE) {
4082			pte = maybe_mkwrite(pte_mkdirty(pte), vma);
4083			vmf->flags &= ~FAULT_FLAG_WRITE;
4084		}
4085		rmap_flags |= RMAP_EXCLUSIVE;
4086	}
4087	flush_icache_page(vma, page);
4088	if (pte_swp_soft_dirty(vmf->orig_pte))
4089		pte = pte_mksoft_dirty(pte);
4090	if (pte_swp_uffd_wp(vmf->orig_pte))
4091		pte = pte_mkuffd_wp(pte);
4092	vmf->orig_pte = pte;
4093
4094	/* ksm created a completely new copy */
4095	if (unlikely(folio != swapcache && swapcache)) {
4096		folio_add_new_anon_rmap(folio, vma, vmf->address);
4097		folio_add_lru_vma(folio, vma);
 
4098	} else {
4099		folio_add_anon_rmap_pte(folio, page, vma, vmf->address,
4100					rmap_flags);
 
4101	}
4102
4103	VM_BUG_ON(!folio_test_anon(folio) ||
4104			(pte_write(pte) && !PageAnonExclusive(page)));
4105	set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
4106	arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
4107
4108	folio_unlock(folio);
4109	if (folio != swapcache && swapcache) {
4110		/*
4111		 * Hold the lock to avoid the swap entry to be reused
4112		 * until we take the PT lock for the pte_same() check
4113		 * (to avoid false positives from pte_same). For
4114		 * further safety release the lock after the swap_free
4115		 * so that the swap count won't change under a
4116		 * parallel locked swapcache.
4117		 */
4118		folio_unlock(swapcache);
4119		folio_put(swapcache);
4120	}
4121
4122	if (vmf->flags & FAULT_FLAG_WRITE) {
4123		ret |= do_wp_page(vmf);
4124		if (ret & VM_FAULT_ERROR)
4125			ret &= VM_FAULT_ERROR;
4126		goto out;
4127	}
4128
4129	/* No need to invalidate - it was non-present before */
4130	update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
4131unlock:
4132	if (vmf->pte)
4133		pte_unmap_unlock(vmf->pte, vmf->ptl);
4134out:
4135	/* Clear the swap cache pin for direct swapin after PTL unlock */
4136	if (need_clear_cache)
4137		swapcache_clear(si, entry);
4138	if (si)
4139		put_swap_device(si);
4140	return ret;
4141out_nomap:
4142	if (vmf->pte)
4143		pte_unmap_unlock(vmf->pte, vmf->ptl);
4144out_page:
4145	folio_unlock(folio);
4146out_release:
4147	folio_put(folio);
4148	if (folio != swapcache && swapcache) {
4149		folio_unlock(swapcache);
4150		folio_put(swapcache);
4151	}
4152	if (need_clear_cache)
4153		swapcache_clear(si, entry);
4154	if (si)
4155		put_swap_device(si);
4156	return ret;
4157}
4158
4159static bool pte_range_none(pte_t *pte, int nr_pages)
4160{
4161	int i;
4162
4163	for (i = 0; i < nr_pages; i++) {
4164		if (!pte_none(ptep_get_lockless(pte + i)))
4165			return false;
4166	}
4167
4168	return true;
4169}
4170
4171static struct folio *alloc_anon_folio(struct vm_fault *vmf)
4172{
4173#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4174	struct vm_area_struct *vma = vmf->vma;
4175	unsigned long orders;
4176	struct folio *folio;
4177	unsigned long addr;
4178	pte_t *pte;
4179	gfp_t gfp;
4180	int order;
4181
4182	/*
4183	 * If uffd is active for the vma we need per-page fault fidelity to
4184	 * maintain the uffd semantics.
4185	 */
4186	if (unlikely(userfaultfd_armed(vma)))
4187		goto fallback;
4188
4189	/*
4190	 * Get a list of all the (large) orders below PMD_ORDER that are enabled
4191	 * for this vma. Then filter out the orders that can't be allocated over
4192	 * the faulting address and still be fully contained in the vma.
4193	 */
4194	orders = thp_vma_allowable_orders(vma, vma->vm_flags, false, true, true,
4195					  BIT(PMD_ORDER) - 1);
4196	orders = thp_vma_suitable_orders(vma, vmf->address, orders);
4197
4198	if (!orders)
4199		goto fallback;
4200
4201	pte = pte_offset_map(vmf->pmd, vmf->address & PMD_MASK);
4202	if (!pte)
4203		return ERR_PTR(-EAGAIN);
4204
4205	/*
4206	 * Find the highest order where the aligned range is completely
4207	 * pte_none(). Note that all remaining orders will be completely
4208	 * pte_none().
4209	 */
4210	order = highest_order(orders);
4211	while (orders) {
4212		addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
4213		if (pte_range_none(pte + pte_index(addr), 1 << order))
4214			break;
4215		order = next_order(&orders, order);
4216	}
4217
4218	pte_unmap(pte);
4219
4220	/* Try allocating the highest of the remaining orders. */
4221	gfp = vma_thp_gfp_mask(vma);
4222	while (orders) {
4223		addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
4224		folio = vma_alloc_folio(gfp, order, vma, addr, true);
4225		if (folio) {
4226			clear_huge_page(&folio->page, vmf->address, 1 << order);
4227			return folio;
4228		}
4229		order = next_order(&orders, order);
4230	}
4231
4232fallback:
4233#endif
4234	return vma_alloc_zeroed_movable_folio(vmf->vma, vmf->address);
4235}
4236
4237/*
4238 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4239 * but allow concurrent faults), and pte mapped but not yet locked.
4240 * We return with mmap_lock still held, but pte unmapped and unlocked.
4241 */
4242static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
4243{
4244	bool uffd_wp = vmf_orig_pte_uffd_wp(vmf);
4245	struct vm_area_struct *vma = vmf->vma;
4246	unsigned long addr = vmf->address;
4247	struct folio *folio;
4248	vm_fault_t ret = 0;
4249	int nr_pages = 1;
4250	pte_t entry;
4251	int i;
4252
4253	/* File mapping without ->vm_ops ? */
4254	if (vma->vm_flags & VM_SHARED)
4255		return VM_FAULT_SIGBUS;
4256
4257	/*
4258	 * Use pte_alloc() instead of pte_alloc_map(), so that OOM can
4259	 * be distinguished from a transient failure of pte_offset_map().
 
 
 
 
 
 
4260	 */
4261	if (pte_alloc(vma->vm_mm, vmf->pmd))
4262		return VM_FAULT_OOM;
4263
 
 
 
 
4264	/* Use the zero-page for reads */
4265	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
4266			!mm_forbids_zeropage(vma->vm_mm)) {
4267		entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
4268						vma->vm_page_prot));
4269		vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4270				vmf->address, &vmf->ptl);
4271		if (!vmf->pte)
4272			goto unlock;
4273		if (vmf_pte_changed(vmf)) {
4274			update_mmu_tlb(vma, vmf->address, vmf->pte);
4275			goto unlock;
4276		}
4277		ret = check_stable_address_space(vma->vm_mm);
4278		if (ret)
4279			goto unlock;
4280		/* Deliver the page fault to userland, check inside PT lock */
4281		if (userfaultfd_missing(vma)) {
4282			pte_unmap_unlock(vmf->pte, vmf->ptl);
4283			return handle_userfault(vmf, VM_UFFD_MISSING);
4284		}
4285		goto setpte;
4286	}
4287
4288	/* Allocate our own private page. */
4289	if (unlikely(anon_vma_prepare(vma)))
4290		goto oom;
4291	/* Returns NULL on OOM or ERR_PTR(-EAGAIN) if we must retry the fault */
4292	folio = alloc_anon_folio(vmf);
4293	if (IS_ERR(folio))
4294		return 0;
4295	if (!folio)
4296		goto oom;
4297
4298	nr_pages = folio_nr_pages(folio);
4299	addr = ALIGN_DOWN(vmf->address, nr_pages * PAGE_SIZE);
4300
4301	if (mem_cgroup_charge(folio, vma->vm_mm, GFP_KERNEL))
4302		goto oom_free_page;
4303	folio_throttle_swaprate(folio, GFP_KERNEL);
4304
4305	/*
4306	 * The memory barrier inside __folio_mark_uptodate makes sure that
4307	 * preceding stores to the page contents become visible before
4308	 * the set_pte_at() write.
4309	 */
4310	__folio_mark_uptodate(folio);
4311
4312	entry = mk_pte(&folio->page, vma->vm_page_prot);
4313	entry = pte_sw_mkyoung(entry);
4314	if (vma->vm_flags & VM_WRITE)
4315		entry = pte_mkwrite(pte_mkdirty(entry), vma);
4316
4317	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
4318	if (!vmf->pte)
4319		goto release;
4320	if (nr_pages == 1 && vmf_pte_changed(vmf)) {
4321		update_mmu_tlb(vma, addr, vmf->pte);
4322		goto release;
4323	} else if (nr_pages > 1 && !pte_range_none(vmf->pte, nr_pages)) {
4324		for (i = 0; i < nr_pages; i++)
4325			update_mmu_tlb(vma, addr + PAGE_SIZE * i, vmf->pte + i);
4326		goto release;
4327	}
4328
4329	ret = check_stable_address_space(vma->vm_mm);
4330	if (ret)
4331		goto release;
4332
4333	/* Deliver the page fault to userland, check inside PT lock */
4334	if (userfaultfd_missing(vma)) {
4335		pte_unmap_unlock(vmf->pte, vmf->ptl);
4336		folio_put(folio);
 
4337		return handle_userfault(vmf, VM_UFFD_MISSING);
4338	}
4339
4340	folio_ref_add(folio, nr_pages - 1);
4341	add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr_pages);
4342	folio_add_new_anon_rmap(folio, vma, addr);
4343	folio_add_lru_vma(folio, vma);
4344setpte:
4345	if (uffd_wp)
4346		entry = pte_mkuffd_wp(entry);
4347	set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr_pages);
4348
4349	/* No need to invalidate - it was non-present before */
4350	update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr_pages);
4351unlock:
4352	if (vmf->pte)
4353		pte_unmap_unlock(vmf->pte, vmf->ptl);
4354	return ret;
4355release:
4356	folio_put(folio);
 
4357	goto unlock;
4358oom_free_page:
4359	folio_put(folio);
4360oom:
4361	return VM_FAULT_OOM;
4362}
4363
4364/*
4365 * The mmap_lock must have been held on entry, and may have been
4366 * released depending on flags and vma->vm_ops->fault() return value.
4367 * See filemap_fault() and __lock_page_retry().
4368 */
4369static vm_fault_t __do_fault(struct vm_fault *vmf)
4370{
4371	struct vm_area_struct *vma = vmf->vma;
4372	struct folio *folio;
4373	vm_fault_t ret;
4374
4375	/*
4376	 * Preallocate pte before we take page_lock because this might lead to
4377	 * deadlocks for memcg reclaim which waits for pages under writeback:
4378	 *				lock_page(A)
4379	 *				SetPageWriteback(A)
4380	 *				unlock_page(A)
4381	 * lock_page(B)
4382	 *				lock_page(B)
4383	 * pte_alloc_one
4384	 *   shrink_page_list
4385	 *     wait_on_page_writeback(A)
4386	 *				SetPageWriteback(B)
4387	 *				unlock_page(B)
4388	 *				# flush A, B to clear the writeback
4389	 */
4390	if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
4391		vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4392		if (!vmf->prealloc_pte)
4393			return VM_FAULT_OOM;
 
4394	}
4395
4396	ret = vma->vm_ops->fault(vmf);
4397	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
4398			    VM_FAULT_DONE_COW)))
4399		return ret;
4400
4401	folio = page_folio(vmf->page);
4402	if (unlikely(PageHWPoison(vmf->page))) {
4403		vm_fault_t poisonret = VM_FAULT_HWPOISON;
4404		if (ret & VM_FAULT_LOCKED) {
4405			if (page_mapped(vmf->page))
4406				unmap_mapping_folio(folio);
4407			/* Retry if a clean folio was removed from the cache. */
4408			if (mapping_evict_folio(folio->mapping, folio))
4409				poisonret = VM_FAULT_NOPAGE;
4410			folio_unlock(folio);
4411		}
4412		folio_put(folio);
4413		vmf->page = NULL;
4414		return poisonret;
4415	}
4416
4417	if (unlikely(!(ret & VM_FAULT_LOCKED)))
4418		folio_lock(folio);
4419	else
4420		VM_BUG_ON_PAGE(!folio_test_locked(folio), vmf->page);
4421
4422	return ret;
4423}
4424
4425#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4426static void deposit_prealloc_pte(struct vm_fault *vmf)
4427{
4428	struct vm_area_struct *vma = vmf->vma;
4429
4430	pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
4431	/*
4432	 * We are going to consume the prealloc table,
4433	 * count that as nr_ptes.
4434	 */
4435	mm_inc_nr_ptes(vma->vm_mm);
4436	vmf->prealloc_pte = NULL;
4437}
4438
4439vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4440{
4441	struct folio *folio = page_folio(page);
4442	struct vm_area_struct *vma = vmf->vma;
4443	bool write = vmf->flags & FAULT_FLAG_WRITE;
4444	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
4445	pmd_t entry;
4446	vm_fault_t ret = VM_FAULT_FALLBACK;
 
4447
4448	if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER))
4449		return ret;
4450
4451	if (page != &folio->page || folio_order(folio) != HPAGE_PMD_ORDER)
4452		return ret;
4453
4454	/*
4455	 * Just backoff if any subpage of a THP is corrupted otherwise
4456	 * the corrupted page may mapped by PMD silently to escape the
4457	 * check.  This kind of THP just can be PTE mapped.  Access to
4458	 * the corrupted subpage should trigger SIGBUS as expected.
4459	 */
4460	if (unlikely(folio_test_has_hwpoisoned(folio)))
4461		return ret;
4462
4463	/*
4464	 * Archs like ppc64 need additional space to store information
4465	 * related to pte entry. Use the preallocated table for that.
4466	 */
4467	if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
4468		vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4469		if (!vmf->prealloc_pte)
4470			return VM_FAULT_OOM;
 
4471	}
4472
4473	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
4474	if (unlikely(!pmd_none(*vmf->pmd)))
4475		goto out;
4476
4477	flush_icache_pages(vma, page, HPAGE_PMD_NR);
 
4478
4479	entry = mk_huge_pmd(page, vma->vm_page_prot);
4480	if (write)
4481		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
4482
4483	add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
4484	folio_add_file_rmap_pmd(folio, page, vma);
4485
4486	/*
4487	 * deposit and withdraw with pmd lock held
4488	 */
4489	if (arch_needs_pgtable_deposit())
4490		deposit_prealloc_pte(vmf);
4491
4492	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
4493
4494	update_mmu_cache_pmd(vma, haddr, vmf->pmd);
4495
4496	/* fault is handled */
4497	ret = 0;
4498	count_vm_event(THP_FILE_MAPPED);
4499out:
4500	spin_unlock(vmf->ptl);
4501	return ret;
4502}
4503#else
4504vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4505{
4506	return VM_FAULT_FALLBACK;
 
4507}
4508#endif
4509
4510/**
4511 * set_pte_range - Set a range of PTEs to point to pages in a folio.
4512 * @vmf: Fault decription.
4513 * @folio: The folio that contains @page.
4514 * @page: The first page to create a PTE for.
4515 * @nr: The number of PTEs to create.
4516 * @addr: The first address to create a PTE for.
 
 
 
 
 
 
 
 
4517 */
4518void set_pte_range(struct vm_fault *vmf, struct folio *folio,
4519		struct page *page, unsigned int nr, unsigned long addr)
4520{
4521	struct vm_area_struct *vma = vmf->vma;
4522	bool uffd_wp = vmf_orig_pte_uffd_wp(vmf);
4523	bool write = vmf->flags & FAULT_FLAG_WRITE;
4524	bool prefault = in_range(vmf->address, addr, nr * PAGE_SIZE);
4525	pte_t entry;
 
4526
4527	flush_icache_pages(vma, page, nr);
4528	entry = mk_pte(page, vma->vm_page_prot);
 
 
4529
4530	if (prefault && arch_wants_old_prefaulted_pte())
4531		entry = pte_mkold(entry);
4532	else
4533		entry = pte_sw_mkyoung(entry);
4534
 
 
 
 
 
 
 
 
 
 
 
 
4535	if (write)
4536		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
4537	if (unlikely(uffd_wp))
4538		entry = pte_mkuffd_wp(entry);
4539	/* copy-on-write page */
4540	if (write && !(vma->vm_flags & VM_SHARED)) {
4541		add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr);
4542		VM_BUG_ON_FOLIO(nr != 1, folio);
4543		folio_add_new_anon_rmap(folio, vma, addr);
4544		folio_add_lru_vma(folio, vma);
4545	} else {
4546		add_mm_counter(vma->vm_mm, mm_counter_file(page), nr);
4547		folio_add_file_rmap_ptes(folio, page, nr, vma);
4548	}
4549	set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr);
4550
4551	/* no need to invalidate: a not-present page won't be cached */
4552	update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr);
4553}
4554
4555static bool vmf_pte_changed(struct vm_fault *vmf)
4556{
4557	if (vmf->flags & FAULT_FLAG_ORIG_PTE_VALID)
4558		return !pte_same(ptep_get(vmf->pte), vmf->orig_pte);
4559
4560	return !pte_none(ptep_get(vmf->pte));
4561}
4562
 
4563/**
4564 * finish_fault - finish page fault once we have prepared the page to fault
4565 *
4566 * @vmf: structure describing the fault
4567 *
4568 * This function handles all that is needed to finish a page fault once the
4569 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4570 * given page, adds reverse page mapping, handles memcg charges and LRU
4571 * addition.
4572 *
4573 * The function expects the page to be locked and on success it consumes a
4574 * reference of a page being mapped (for the PTE which maps it).
4575 *
4576 * Return: %0 on success, %VM_FAULT_ code in case of error.
4577 */
4578vm_fault_t finish_fault(struct vm_fault *vmf)
4579{
4580	struct vm_area_struct *vma = vmf->vma;
4581	struct page *page;
4582	vm_fault_t ret;
4583
4584	/* Did we COW the page? */
4585	if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
 
4586		page = vmf->cow_page;
4587	else
4588		page = vmf->page;
4589
4590	/*
4591	 * check even for read faults because we might have lost our CoWed
4592	 * page
4593	 */
4594	if (!(vma->vm_flags & VM_SHARED)) {
4595		ret = check_stable_address_space(vma->vm_mm);
4596		if (ret)
4597			return ret;
4598	}
4599
4600	if (pmd_none(*vmf->pmd)) {
4601		if (PageTransCompound(page)) {
4602			ret = do_set_pmd(vmf, page);
4603			if (ret != VM_FAULT_FALLBACK)
4604				return ret;
4605		}
4606
4607		if (vmf->prealloc_pte)
4608			pmd_install(vma->vm_mm, vmf->pmd, &vmf->prealloc_pte);
4609		else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
4610			return VM_FAULT_OOM;
4611	}
4612
4613	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4614				      vmf->address, &vmf->ptl);
4615	if (!vmf->pte)
4616		return VM_FAULT_NOPAGE;
4617
4618	/* Re-check under ptl */
4619	if (likely(!vmf_pte_changed(vmf))) {
4620		struct folio *folio = page_folio(page);
4621
4622		set_pte_range(vmf, folio, page, 1, vmf->address);
4623		ret = 0;
4624	} else {
4625		update_mmu_tlb(vma, vmf->address, vmf->pte);
4626		ret = VM_FAULT_NOPAGE;
4627	}
4628
4629	pte_unmap_unlock(vmf->pte, vmf->ptl);
4630	return ret;
4631}
4632
4633static unsigned long fault_around_pages __read_mostly =
4634	65536 >> PAGE_SHIFT;
4635
4636#ifdef CONFIG_DEBUG_FS
4637static int fault_around_bytes_get(void *data, u64 *val)
4638{
4639	*val = fault_around_pages << PAGE_SHIFT;
4640	return 0;
4641}
4642
4643/*
4644 * fault_around_bytes must be rounded down to the nearest page order as it's
4645 * what do_fault_around() expects to see.
4646 */
4647static int fault_around_bytes_set(void *data, u64 val)
4648{
4649	if (val / PAGE_SIZE > PTRS_PER_PTE)
4650		return -EINVAL;
4651
4652	/*
4653	 * The minimum value is 1 page, however this results in no fault-around
4654	 * at all. See should_fault_around().
4655	 */
4656	fault_around_pages = max(rounddown_pow_of_two(val) >> PAGE_SHIFT, 1UL);
4657
4658	return 0;
4659}
4660DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4661		fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4662
4663static int __init fault_around_debugfs(void)
4664{
4665	debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4666				   &fault_around_bytes_fops);
4667	return 0;
4668}
4669late_initcall(fault_around_debugfs);
4670#endif
4671
4672/*
4673 * do_fault_around() tries to map few pages around the fault address. The hope
4674 * is that the pages will be needed soon and this will lower the number of
4675 * faults to handle.
4676 *
4677 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4678 * not ready to be mapped: not up-to-date, locked, etc.
4679 *
4680 * This function doesn't cross VMA or page table boundaries, in order to call
4681 * map_pages() and acquire a PTE lock only once.
 
 
 
 
4682 *
4683 * fault_around_pages defines how many pages we'll try to map.
4684 * do_fault_around() expects it to be set to a power of two less than or equal
4685 * to PTRS_PER_PTE.
4686 *
4687 * The virtual address of the area that we map is naturally aligned to
4688 * fault_around_pages * PAGE_SIZE rounded down to the machine page size
4689 * (and therefore to page order).  This way it's easier to guarantee
4690 * that we don't cross page table boundaries.
4691 */
4692static vm_fault_t do_fault_around(struct vm_fault *vmf)
4693{
4694	pgoff_t nr_pages = READ_ONCE(fault_around_pages);
4695	pgoff_t pte_off = pte_index(vmf->address);
4696	/* The page offset of vmf->address within the VMA. */
4697	pgoff_t vma_off = vmf->pgoff - vmf->vma->vm_pgoff;
4698	pgoff_t from_pte, to_pte;
4699	vm_fault_t ret;
 
 
4700
4701	/* The PTE offset of the start address, clamped to the VMA. */
4702	from_pte = max(ALIGN_DOWN(pte_off, nr_pages),
4703		       pte_off - min(pte_off, vma_off));
4704
4705	/* The PTE offset of the end address, clamped to the VMA and PTE. */
4706	to_pte = min3(from_pte + nr_pages, (pgoff_t)PTRS_PER_PTE,
4707		      pte_off + vma_pages(vmf->vma) - vma_off) - 1;
 
 
 
 
 
 
4708
4709	if (pmd_none(*vmf->pmd)) {
4710		vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4711		if (!vmf->prealloc_pte)
4712			return VM_FAULT_OOM;
 
4713	}
4714
4715	rcu_read_lock();
4716	ret = vmf->vma->vm_ops->map_pages(vmf,
4717			vmf->pgoff + from_pte - pte_off,
4718			vmf->pgoff + to_pte - pte_off);
4719	rcu_read_unlock();
4720
4721	return ret;
4722}
4723
4724/* Return true if we should do read fault-around, false otherwise */
4725static inline bool should_fault_around(struct vm_fault *vmf)
4726{
4727	/* No ->map_pages?  No way to fault around... */
4728	if (!vmf->vma->vm_ops->map_pages)
4729		return false;
4730
4731	if (uffd_disable_fault_around(vmf->vma))
4732		return false;
 
4733
4734	/* A single page implies no faulting 'around' at all. */
4735	return fault_around_pages > 1;
 
 
 
 
 
 
 
4736}
4737
4738static vm_fault_t do_read_fault(struct vm_fault *vmf)
4739{
 
4740	vm_fault_t ret = 0;
4741	struct folio *folio;
4742
4743	/*
4744	 * Let's call ->map_pages() first and use ->fault() as fallback
4745	 * if page by the offset is not ready to be mapped (cold cache or
4746	 * something).
4747	 */
4748	if (should_fault_around(vmf)) {
4749		ret = do_fault_around(vmf);
4750		if (ret)
4751			return ret;
4752	}
4753
4754	ret = vmf_can_call_fault(vmf);
4755	if (ret)
4756		return ret;
4757
4758	ret = __do_fault(vmf);
4759	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4760		return ret;
4761
4762	ret |= finish_fault(vmf);
4763	folio = page_folio(vmf->page);
4764	folio_unlock(folio);
4765	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4766		folio_put(folio);
4767	return ret;
4768}
4769
4770static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4771{
4772	struct vm_area_struct *vma = vmf->vma;
4773	struct folio *folio;
4774	vm_fault_t ret;
4775
4776	ret = vmf_can_call_fault(vmf);
4777	if (!ret)
4778		ret = vmf_anon_prepare(vmf);
4779	if (ret)
4780		return ret;
4781
4782	folio = folio_prealloc(vma->vm_mm, vma, vmf->address, false);
4783	if (!folio)
4784		return VM_FAULT_OOM;
4785
4786	vmf->cow_page = &folio->page;
 
 
 
 
4787
4788	ret = __do_fault(vmf);
4789	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4790		goto uncharge_out;
4791	if (ret & VM_FAULT_DONE_COW)
4792		return ret;
4793
4794	copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4795	__folio_mark_uptodate(folio);
4796
4797	ret |= finish_fault(vmf);
4798	unlock_page(vmf->page);
4799	put_page(vmf->page);
4800	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4801		goto uncharge_out;
4802	return ret;
4803uncharge_out:
4804	folio_put(folio);
 
4805	return ret;
4806}
4807
4808static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4809{
4810	struct vm_area_struct *vma = vmf->vma;
4811	vm_fault_t ret, tmp;
4812	struct folio *folio;
4813
4814	ret = vmf_can_call_fault(vmf);
4815	if (ret)
4816		return ret;
4817
4818	ret = __do_fault(vmf);
4819	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4820		return ret;
4821
4822	folio = page_folio(vmf->page);
4823
4824	/*
4825	 * Check if the backing address space wants to know that the page is
4826	 * about to become writable
4827	 */
4828	if (vma->vm_ops->page_mkwrite) {
4829		folio_unlock(folio);
4830		tmp = do_page_mkwrite(vmf, folio);
4831		if (unlikely(!tmp ||
4832				(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4833			folio_put(folio);
4834			return tmp;
4835		}
4836	}
4837
4838	ret |= finish_fault(vmf);
4839	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4840					VM_FAULT_RETRY))) {
4841		folio_unlock(folio);
4842		folio_put(folio);
4843		return ret;
4844	}
4845
4846	ret |= fault_dirty_shared_page(vmf);
4847	return ret;
4848}
4849
4850/*
4851 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4852 * but allow concurrent faults).
4853 * The mmap_lock may have been released depending on flags and our
4854 * return value.  See filemap_fault() and __folio_lock_or_retry().
4855 * If mmap_lock is released, vma may become invalid (for example
4856 * by other thread calling munmap()).
4857 */
4858static vm_fault_t do_fault(struct vm_fault *vmf)
4859{
4860	struct vm_area_struct *vma = vmf->vma;
4861	struct mm_struct *vm_mm = vma->vm_mm;
4862	vm_fault_t ret;
4863
4864	/*
4865	 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4866	 */
4867	if (!vma->vm_ops->fault) {
4868		vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
4869					       vmf->address, &vmf->ptl);
4870		if (unlikely(!vmf->pte))
 
 
4871			ret = VM_FAULT_SIGBUS;
4872		else {
 
 
 
 
4873			/*
4874			 * Make sure this is not a temporary clearing of pte
4875			 * by holding ptl and checking again. A R/M/W update
4876			 * of pte involves: take ptl, clearing the pte so that
4877			 * we don't have concurrent modification by hardware
4878			 * followed by an update.
4879			 */
4880			if (unlikely(pte_none(ptep_get(vmf->pte))))
4881				ret = VM_FAULT_SIGBUS;
4882			else
4883				ret = VM_FAULT_NOPAGE;
4884
4885			pte_unmap_unlock(vmf->pte, vmf->ptl);
4886		}
4887	} else if (!(vmf->flags & FAULT_FLAG_WRITE))
4888		ret = do_read_fault(vmf);
4889	else if (!(vma->vm_flags & VM_SHARED))
4890		ret = do_cow_fault(vmf);
4891	else
4892		ret = do_shared_fault(vmf);
4893
4894	/* preallocated pagetable is unused: free it */
4895	if (vmf->prealloc_pte) {
4896		pte_free(vm_mm, vmf->prealloc_pte);
4897		vmf->prealloc_pte = NULL;
4898	}
4899	return ret;
4900}
4901
4902int numa_migrate_prep(struct folio *folio, struct vm_area_struct *vma,
4903		      unsigned long addr, int page_nid, int *flags)
 
4904{
4905	folio_get(folio);
4906
4907	/* Record the current PID acceesing VMA */
4908	vma_set_access_pid_bit(vma);
4909
4910	count_vm_numa_event(NUMA_HINT_FAULTS);
4911	if (page_nid == numa_node_id()) {
4912		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4913		*flags |= TNF_FAULT_LOCAL;
4914	}
4915
4916	return mpol_misplaced(folio, vma, addr);
4917}
4918
4919static vm_fault_t do_numa_page(struct vm_fault *vmf)
4920{
4921	struct vm_area_struct *vma = vmf->vma;
4922	struct folio *folio = NULL;
4923	int nid = NUMA_NO_NODE;
4924	bool writable = false;
4925	int last_cpupid;
4926	int target_nid;
 
4927	pte_t pte, old_pte;
 
4928	int flags = 0;
4929
4930	/*
4931	 * The "pte" at this point cannot be used safely without
4932	 * validation through pte_unmap_same(). It's of NUMA type but
4933	 * the pfn may be screwed if the read is non atomic.
4934	 */
 
4935	spin_lock(vmf->ptl);
4936	if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
4937		pte_unmap_unlock(vmf->pte, vmf->ptl);
4938		goto out;
4939	}
4940
4941	/* Get the normal PTE  */
4942	old_pte = ptep_get(vmf->pte);
4943	pte = pte_modify(old_pte, vma->vm_page_prot);
4944
4945	/*
4946	 * Detect now whether the PTE could be writable; this information
4947	 * is only valid while holding the PT lock.
4948	 */
4949	writable = pte_write(pte);
4950	if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
4951	    can_change_pte_writable(vma, vmf->address, pte))
4952		writable = true;
4953
4954	folio = vm_normal_folio(vma, vmf->address, pte);
4955	if (!folio || folio_is_zone_device(folio))
4956		goto out_map;
 
 
 
 
 
4957
4958	/* TODO: handle PTE-mapped THP */
4959	if (folio_test_large(folio))
4960		goto out_map;
 
 
4961
4962	/*
4963	 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4964	 * much anyway since they can be in shared cache state. This misses
4965	 * the case where a mapping is writable but the process never writes
4966	 * to it but pte_write gets cleared during protection updates and
4967	 * pte_dirty has unpredictable behaviour between PTE scan updates,
4968	 * background writeback, dirty balancing and application behaviour.
4969	 */
4970	if (!writable)
4971		flags |= TNF_NO_GROUP;
4972
4973	/*
4974	 * Flag if the folio is shared between multiple address spaces. This
4975	 * is later used when determining whether to group tasks together
4976	 */
4977	if (folio_estimated_sharers(folio) > 1 && (vma->vm_flags & VM_SHARED))
4978		flags |= TNF_SHARED;
4979
4980	nid = folio_nid(folio);
4981	/*
4982	 * For memory tiering mode, cpupid of slow memory page is used
4983	 * to record page access time.  So use default value.
4984	 */
4985	if ((sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4986	    !node_is_toptier(nid))
4987		last_cpupid = (-1 & LAST_CPUPID_MASK);
4988	else
4989		last_cpupid = folio_last_cpupid(folio);
4990	target_nid = numa_migrate_prep(folio, vma, vmf->address, nid, &flags);
4991	if (target_nid == NUMA_NO_NODE) {
4992		folio_put(folio);
4993		goto out_map;
4994	}
4995	pte_unmap_unlock(vmf->pte, vmf->ptl);
4996	writable = false;
4997
4998	/* Migrate to the requested node */
4999	if (migrate_misplaced_folio(folio, vma, target_nid)) {
5000		nid = target_nid;
 
5001		flags |= TNF_MIGRATED;
5002	} else {
5003		flags |= TNF_MIGRATE_FAIL;
5004		vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
5005					       vmf->address, &vmf->ptl);
5006		if (unlikely(!vmf->pte))
5007			goto out;
5008		if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
5009			pte_unmap_unlock(vmf->pte, vmf->ptl);
5010			goto out;
5011		}
5012		goto out_map;
5013	}
5014
5015out:
5016	if (nid != NUMA_NO_NODE)
5017		task_numa_fault(last_cpupid, nid, 1, flags);
5018	return 0;
5019out_map:
5020	/*
5021	 * Make it present again, depending on how arch implements
5022	 * non-accessible ptes, some can allow access by kernel mode.
5023	 */
5024	old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
5025	pte = pte_modify(old_pte, vma->vm_page_prot);
5026	pte = pte_mkyoung(pte);
5027	if (writable)
5028		pte = pte_mkwrite(pte, vma);
5029	ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
5030	update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
5031	pte_unmap_unlock(vmf->pte, vmf->ptl);
5032	goto out;
5033}
5034
5035static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
5036{
5037	struct vm_area_struct *vma = vmf->vma;
5038	if (vma_is_anonymous(vma))
5039		return do_huge_pmd_anonymous_page(vmf);
5040	if (vma->vm_ops->huge_fault)
5041		return vma->vm_ops->huge_fault(vmf, PMD_ORDER);
5042	return VM_FAULT_FALLBACK;
5043}
5044
5045/* `inline' is required to avoid gcc 4.1.2 build error */
5046static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
5047{
5048	struct vm_area_struct *vma = vmf->vma;
5049	const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
5050	vm_fault_t ret;
5051
5052	if (vma_is_anonymous(vma)) {
5053		if (likely(!unshare) &&
5054		    userfaultfd_huge_pmd_wp(vma, vmf->orig_pmd)) {
5055			if (userfaultfd_wp_async(vmf->vma))
5056				goto split;
5057			return handle_userfault(vmf, VM_UFFD_WP);
5058		}
5059		return do_huge_pmd_wp_page(vmf);
5060	}
5061
5062	if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
5063		if (vma->vm_ops->huge_fault) {
5064			ret = vma->vm_ops->huge_fault(vmf, PMD_ORDER);
5065			if (!(ret & VM_FAULT_FALLBACK))
5066				return ret;
5067		}
5068	}
5069
5070split:
5071	/* COW or write-notify handled on pte level: split pmd. */
5072	__split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
5073
5074	return VM_FAULT_FALLBACK;
5075}
5076
 
 
 
 
 
5077static vm_fault_t create_huge_pud(struct vm_fault *vmf)
5078{
5079#if defined(CONFIG_TRANSPARENT_HUGEPAGE) &&			\
5080	defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5081	struct vm_area_struct *vma = vmf->vma;
5082	/* No support for anonymous transparent PUD pages yet */
5083	if (vma_is_anonymous(vma))
5084		return VM_FAULT_FALLBACK;
5085	if (vma->vm_ops->huge_fault)
5086		return vma->vm_ops->huge_fault(vmf, PUD_ORDER);
5087#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5088	return VM_FAULT_FALLBACK;
5089}
5090
5091static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
5092{
5093#if defined(CONFIG_TRANSPARENT_HUGEPAGE) &&			\
5094	defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5095	struct vm_area_struct *vma = vmf->vma;
5096	vm_fault_t ret;
5097
5098	/* No support for anonymous transparent PUD pages yet */
5099	if (vma_is_anonymous(vma))
5100		goto split;
5101	if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
5102		if (vma->vm_ops->huge_fault) {
5103			ret = vma->vm_ops->huge_fault(vmf, PUD_ORDER);
5104			if (!(ret & VM_FAULT_FALLBACK))
5105				return ret;
5106		}
5107	}
5108split:
5109	/* COW or write-notify not handled on PUD level: split pud.*/
5110	__split_huge_pud(vma, vmf->pud, vmf->address);
5111#endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
5112	return VM_FAULT_FALLBACK;
5113}
5114
5115/*
5116 * These routines also need to handle stuff like marking pages dirty
5117 * and/or accessed for architectures that don't do it in hardware (most
5118 * RISC architectures).  The early dirtying is also good on the i386.
5119 *
5120 * There is also a hook called "update_mmu_cache()" that architectures
5121 * with external mmu caches can use to update those (ie the Sparc or
5122 * PowerPC hashed page tables that act as extended TLBs).
5123 *
5124 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
5125 * concurrent faults).
5126 *
5127 * The mmap_lock may have been released depending on flags and our return value.
5128 * See filemap_fault() and __folio_lock_or_retry().
5129 */
5130static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
5131{
5132	pte_t entry;
5133
5134	if (unlikely(pmd_none(*vmf->pmd))) {
5135		/*
5136		 * Leave __pte_alloc() until later: because vm_ops->fault may
5137		 * want to allocate huge page, and if we expose page table
5138		 * for an instant, it will be difficult to retract from
5139		 * concurrent faults and from rmap lookups.
5140		 */
5141		vmf->pte = NULL;
5142		vmf->flags &= ~FAULT_FLAG_ORIG_PTE_VALID;
5143	} else {
 
 
 
5144		/*
5145		 * A regular pmd is established and it can't morph into a huge
5146		 * pmd by anon khugepaged, since that takes mmap_lock in write
5147		 * mode; but shmem or file collapse to THP could still morph
5148		 * it into a huge pmd: just retry later if so.
5149		 */
5150		vmf->pte = pte_offset_map_nolock(vmf->vma->vm_mm, vmf->pmd,
5151						 vmf->address, &vmf->ptl);
5152		if (unlikely(!vmf->pte))
5153			return 0;
5154		vmf->orig_pte = ptep_get_lockless(vmf->pte);
5155		vmf->flags |= FAULT_FLAG_ORIG_PTE_VALID;
5156
 
 
 
 
 
 
 
 
 
5157		if (pte_none(vmf->orig_pte)) {
5158			pte_unmap(vmf->pte);
5159			vmf->pte = NULL;
5160		}
5161	}
5162
5163	if (!vmf->pte)
5164		return do_pte_missing(vmf);
 
 
 
 
5165
5166	if (!pte_present(vmf->orig_pte))
5167		return do_swap_page(vmf);
5168
5169	if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
5170		return do_numa_page(vmf);
5171
 
5172	spin_lock(vmf->ptl);
5173	entry = vmf->orig_pte;
5174	if (unlikely(!pte_same(ptep_get(vmf->pte), entry))) {
5175		update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
5176		goto unlock;
5177	}
5178	if (vmf->flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
5179		if (!pte_write(entry))
5180			return do_wp_page(vmf);
5181		else if (likely(vmf->flags & FAULT_FLAG_WRITE))
5182			entry = pte_mkdirty(entry);
5183	}
5184	entry = pte_mkyoung(entry);
5185	if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
5186				vmf->flags & FAULT_FLAG_WRITE)) {
5187		update_mmu_cache_range(vmf, vmf->vma, vmf->address,
5188				vmf->pte, 1);
5189	} else {
5190		/* Skip spurious TLB flush for retried page fault */
5191		if (vmf->flags & FAULT_FLAG_TRIED)
5192			goto unlock;
5193		/*
5194		 * This is needed only for protection faults but the arch code
5195		 * is not yet telling us if this is a protection fault or not.
5196		 * This still avoids useless tlb flushes for .text page faults
5197		 * with threads.
5198		 */
5199		if (vmf->flags & FAULT_FLAG_WRITE)
5200			flush_tlb_fix_spurious_fault(vmf->vma, vmf->address,
5201						     vmf->pte);
5202	}
5203unlock:
5204	pte_unmap_unlock(vmf->pte, vmf->ptl);
5205	return 0;
5206}
5207
5208/*
5209 * On entry, we hold either the VMA lock or the mmap_lock
5210 * (FAULT_FLAG_VMA_LOCK tells you which).  If VM_FAULT_RETRY is set in
5211 * the result, the mmap_lock is not held on exit.  See filemap_fault()
5212 * and __folio_lock_or_retry().
5213 */
5214static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
5215		unsigned long address, unsigned int flags)
5216{
5217	struct vm_fault vmf = {
5218		.vma = vma,
5219		.address = address & PAGE_MASK,
5220		.real_address = address,
5221		.flags = flags,
5222		.pgoff = linear_page_index(vma, address),
5223		.gfp_mask = __get_fault_gfp_mask(vma),
5224	};
 
5225	struct mm_struct *mm = vma->vm_mm;
5226	unsigned long vm_flags = vma->vm_flags;
5227	pgd_t *pgd;
5228	p4d_t *p4d;
5229	vm_fault_t ret;
5230
5231	pgd = pgd_offset(mm, address);
5232	p4d = p4d_alloc(mm, pgd, address);
5233	if (!p4d)
5234		return VM_FAULT_OOM;
5235
5236	vmf.pud = pud_alloc(mm, p4d, address);
5237	if (!vmf.pud)
5238		return VM_FAULT_OOM;
5239retry_pud:
5240	if (pud_none(*vmf.pud) &&
5241	    thp_vma_allowable_order(vma, vm_flags, false, true, true, PUD_ORDER)) {
5242		ret = create_huge_pud(&vmf);
5243		if (!(ret & VM_FAULT_FALLBACK))
5244			return ret;
5245	} else {
5246		pud_t orig_pud = *vmf.pud;
5247
5248		barrier();
5249		if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
5250
5251			/*
5252			 * TODO once we support anonymous PUDs: NUMA case and
5253			 * FAULT_FLAG_UNSHARE handling.
5254			 */
5255			if ((flags & FAULT_FLAG_WRITE) && !pud_write(orig_pud)) {
5256				ret = wp_huge_pud(&vmf, orig_pud);
5257				if (!(ret & VM_FAULT_FALLBACK))
5258					return ret;
5259			} else {
5260				huge_pud_set_accessed(&vmf, orig_pud);
5261				return 0;
5262			}
5263		}
5264	}
5265
5266	vmf.pmd = pmd_alloc(mm, vmf.pud, address);
5267	if (!vmf.pmd)
5268		return VM_FAULT_OOM;
5269
5270	/* Huge pud page fault raced with pmd_alloc? */
5271	if (pud_trans_unstable(vmf.pud))
5272		goto retry_pud;
5273
5274	if (pmd_none(*vmf.pmd) &&
5275	    thp_vma_allowable_order(vma, vm_flags, false, true, true, PMD_ORDER)) {
5276		ret = create_huge_pmd(&vmf);
5277		if (!(ret & VM_FAULT_FALLBACK))
5278			return ret;
5279	} else {
5280		vmf.orig_pmd = pmdp_get_lockless(vmf.pmd);
5281
5282		if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
 
5283			VM_BUG_ON(thp_migration_supported() &&
5284					  !is_pmd_migration_entry(vmf.orig_pmd));
5285			if (is_pmd_migration_entry(vmf.orig_pmd))
5286				pmd_migration_entry_wait(mm, vmf.pmd);
5287			return 0;
5288		}
5289		if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) {
5290			if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma))
5291				return do_huge_pmd_numa_page(&vmf);
5292
5293			if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
5294			    !pmd_write(vmf.orig_pmd)) {
5295				ret = wp_huge_pmd(&vmf);
5296				if (!(ret & VM_FAULT_FALLBACK))
5297					return ret;
5298			} else {
5299				huge_pmd_set_accessed(&vmf);
5300				return 0;
5301			}
5302		}
5303	}
5304
5305	return handle_pte_fault(&vmf);
5306}
5307
5308/**
5309 * mm_account_fault - Do page fault accounting
5310 * @mm: mm from which memcg should be extracted. It can be NULL.
5311 * @regs: the pt_regs struct pointer.  When set to NULL, will skip accounting
5312 *        of perf event counters, but we'll still do the per-task accounting to
5313 *        the task who triggered this page fault.
5314 * @address: the faulted address.
5315 * @flags: the fault flags.
5316 * @ret: the fault retcode.
5317 *
5318 * This will take care of most of the page fault accounting.  Meanwhile, it
5319 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5320 * updates.  However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5321 * still be in per-arch page fault handlers at the entry of page fault.
5322 */
5323static inline void mm_account_fault(struct mm_struct *mm, struct pt_regs *regs,
5324				    unsigned long address, unsigned int flags,
5325				    vm_fault_t ret)
5326{
5327	bool major;
5328
5329	/* Incomplete faults will be accounted upon completion. */
5330	if (ret & VM_FAULT_RETRY)
5331		return;
5332
5333	/*
5334	 * To preserve the behavior of older kernels, PGFAULT counters record
5335	 * both successful and failed faults, as opposed to perf counters,
5336	 * which ignore failed cases.
5337	 */
5338	count_vm_event(PGFAULT);
5339	count_memcg_event_mm(mm, PGFAULT);
5340
5341	/*
5342	 * Do not account for unsuccessful faults (e.g. when the address wasn't
5343	 * valid).  That includes arch_vma_access_permitted() failing before
5344	 * reaching here. So this is not a "this many hardware page faults"
5345	 * counter.  We should use the hw profiling for that.
5346	 */
5347	if (ret & VM_FAULT_ERROR)
5348		return;
5349
5350	/*
5351	 * We define the fault as a major fault when the final successful fault
5352	 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5353	 * handle it immediately previously).
5354	 */
5355	major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
5356
5357	if (major)
5358		current->maj_flt++;
5359	else
5360		current->min_flt++;
5361
5362	/*
5363	 * If the fault is done for GUP, regs will be NULL.  We only do the
5364	 * accounting for the per thread fault counters who triggered the
5365	 * fault, and we skip the perf event updates.
5366	 */
5367	if (!regs)
5368		return;
5369
5370	if (major)
5371		perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
5372	else
5373		perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
5374}
5375
5376#ifdef CONFIG_LRU_GEN
5377static void lru_gen_enter_fault(struct vm_area_struct *vma)
5378{
5379	/* the LRU algorithm only applies to accesses with recency */
5380	current->in_lru_fault = vma_has_recency(vma);
5381}
5382
5383static void lru_gen_exit_fault(void)
5384{
5385	current->in_lru_fault = false;
5386}
5387#else
5388static void lru_gen_enter_fault(struct vm_area_struct *vma)
5389{
5390}
5391
5392static void lru_gen_exit_fault(void)
5393{
5394}
5395#endif /* CONFIG_LRU_GEN */
5396
5397static vm_fault_t sanitize_fault_flags(struct vm_area_struct *vma,
5398				       unsigned int *flags)
5399{
5400	if (unlikely(*flags & FAULT_FLAG_UNSHARE)) {
5401		if (WARN_ON_ONCE(*flags & FAULT_FLAG_WRITE))
5402			return VM_FAULT_SIGSEGV;
5403		/*
5404		 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
5405		 * just treat it like an ordinary read-fault otherwise.
5406		 */
5407		if (!is_cow_mapping(vma->vm_flags))
5408			*flags &= ~FAULT_FLAG_UNSHARE;
5409	} else if (*flags & FAULT_FLAG_WRITE) {
5410		/* Write faults on read-only mappings are impossible ... */
5411		if (WARN_ON_ONCE(!(vma->vm_flags & VM_MAYWRITE)))
5412			return VM_FAULT_SIGSEGV;
5413		/* ... and FOLL_FORCE only applies to COW mappings. */
5414		if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE) &&
5415				 !is_cow_mapping(vma->vm_flags)))
5416			return VM_FAULT_SIGSEGV;
5417	}
5418#ifdef CONFIG_PER_VMA_LOCK
5419	/*
5420	 * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of
5421	 * the assumption that lock is dropped on VM_FAULT_RETRY.
5422	 */
5423	if (WARN_ON_ONCE((*flags &
5424			(FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)) ==
5425			(FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)))
5426		return VM_FAULT_SIGSEGV;
5427#endif
5428
5429	return 0;
5430}
5431
5432/*
5433 * By the time we get here, we already hold the mm semaphore
5434 *
5435 * The mmap_lock may have been released depending on flags and our
5436 * return value.  See filemap_fault() and __folio_lock_or_retry().
5437 */
5438vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
5439			   unsigned int flags, struct pt_regs *regs)
5440{
5441	/* If the fault handler drops the mmap_lock, vma may be freed */
5442	struct mm_struct *mm = vma->vm_mm;
5443	vm_fault_t ret;
5444
5445	__set_current_state(TASK_RUNNING);
5446
5447	ret = sanitize_fault_flags(vma, &flags);
5448	if (ret)
5449		goto out;
 
 
5450
5451	if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
5452					    flags & FAULT_FLAG_INSTRUCTION,
5453					    flags & FAULT_FLAG_REMOTE)) {
5454		ret = VM_FAULT_SIGSEGV;
5455		goto out;
5456	}
5457
5458	/*
5459	 * Enable the memcg OOM handling for faults triggered in user
5460	 * space.  Kernel faults are handled more gracefully.
5461	 */
5462	if (flags & FAULT_FLAG_USER)
5463		mem_cgroup_enter_user_fault();
5464
5465	lru_gen_enter_fault(vma);
5466
5467	if (unlikely(is_vm_hugetlb_page(vma)))
5468		ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
5469	else
5470		ret = __handle_mm_fault(vma, address, flags);
5471
5472	lru_gen_exit_fault();
5473
5474	if (flags & FAULT_FLAG_USER) {
5475		mem_cgroup_exit_user_fault();
5476		/*
5477		 * The task may have entered a memcg OOM situation but
5478		 * if the allocation error was handled gracefully (no
5479		 * VM_FAULT_OOM), there is no need to kill anything.
5480		 * Just clean up the OOM state peacefully.
5481		 */
5482		if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
5483			mem_cgroup_oom_synchronize(false);
5484	}
5485out:
5486	mm_account_fault(mm, regs, address, flags, ret);
5487
5488	return ret;
5489}
5490EXPORT_SYMBOL_GPL(handle_mm_fault);
5491
5492#ifdef CONFIG_LOCK_MM_AND_FIND_VMA
5493#include <linux/extable.h>
5494
5495static inline bool get_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5496{
5497	if (likely(mmap_read_trylock(mm)))
5498		return true;
5499
5500	if (regs && !user_mode(regs)) {
5501		unsigned long ip = exception_ip(regs);
5502		if (!search_exception_tables(ip))
5503			return false;
5504	}
5505
5506	return !mmap_read_lock_killable(mm);
5507}
5508
5509static inline bool mmap_upgrade_trylock(struct mm_struct *mm)
5510{
5511	/*
5512	 * We don't have this operation yet.
5513	 *
5514	 * It should be easy enough to do: it's basically a
5515	 *    atomic_long_try_cmpxchg_acquire()
5516	 * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but
5517	 * it also needs the proper lockdep magic etc.
5518	 */
5519	return false;
5520}
5521
5522static inline bool upgrade_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5523{
5524	mmap_read_unlock(mm);
5525	if (regs && !user_mode(regs)) {
5526		unsigned long ip = exception_ip(regs);
5527		if (!search_exception_tables(ip))
5528			return false;
5529	}
5530	return !mmap_write_lock_killable(mm);
5531}
5532
5533/*
5534 * Helper for page fault handling.
5535 *
5536 * This is kind of equivalend to "mmap_read_lock()" followed
5537 * by "find_extend_vma()", except it's a lot more careful about
5538 * the locking (and will drop the lock on failure).
5539 *
5540 * For example, if we have a kernel bug that causes a page
5541 * fault, we don't want to just use mmap_read_lock() to get
5542 * the mm lock, because that would deadlock if the bug were
5543 * to happen while we're holding the mm lock for writing.
5544 *
5545 * So this checks the exception tables on kernel faults in
5546 * order to only do this all for instructions that are actually
5547 * expected to fault.
5548 *
5549 * We can also actually take the mm lock for writing if we
5550 * need to extend the vma, which helps the VM layer a lot.
5551 */
5552struct vm_area_struct *lock_mm_and_find_vma(struct mm_struct *mm,
5553			unsigned long addr, struct pt_regs *regs)
5554{
5555	struct vm_area_struct *vma;
5556
5557	if (!get_mmap_lock_carefully(mm, regs))
5558		return NULL;
5559
5560	vma = find_vma(mm, addr);
5561	if (likely(vma && (vma->vm_start <= addr)))
5562		return vma;
5563
5564	/*
5565	 * Well, dang. We might still be successful, but only
5566	 * if we can extend a vma to do so.
5567	 */
5568	if (!vma || !(vma->vm_flags & VM_GROWSDOWN)) {
5569		mmap_read_unlock(mm);
5570		return NULL;
5571	}
5572
5573	/*
5574	 * We can try to upgrade the mmap lock atomically,
5575	 * in which case we can continue to use the vma
5576	 * we already looked up.
5577	 *
5578	 * Otherwise we'll have to drop the mmap lock and
5579	 * re-take it, and also look up the vma again,
5580	 * re-checking it.
5581	 */
5582	if (!mmap_upgrade_trylock(mm)) {
5583		if (!upgrade_mmap_lock_carefully(mm, regs))
5584			return NULL;
5585
5586		vma = find_vma(mm, addr);
5587		if (!vma)
5588			goto fail;
5589		if (vma->vm_start <= addr)
5590			goto success;
5591		if (!(vma->vm_flags & VM_GROWSDOWN))
5592			goto fail;
5593	}
5594
5595	if (expand_stack_locked(vma, addr))
5596		goto fail;
5597
5598success:
5599	mmap_write_downgrade(mm);
5600	return vma;
5601
5602fail:
5603	mmap_write_unlock(mm);
5604	return NULL;
5605}
5606#endif
5607
5608#ifdef CONFIG_PER_VMA_LOCK
5609/*
5610 * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be
5611 * stable and not isolated. If the VMA is not found or is being modified the
5612 * function returns NULL.
5613 */
5614struct vm_area_struct *lock_vma_under_rcu(struct mm_struct *mm,
5615					  unsigned long address)
5616{
5617	MA_STATE(mas, &mm->mm_mt, address, address);
5618	struct vm_area_struct *vma;
5619
5620	rcu_read_lock();
5621retry:
5622	vma = mas_walk(&mas);
5623	if (!vma)
5624		goto inval;
5625
5626	if (!vma_start_read(vma))
5627		goto inval;
5628
5629	/*
5630	 * find_mergeable_anon_vma uses adjacent vmas which are not locked.
5631	 * This check must happen after vma_start_read(); otherwise, a
5632	 * concurrent mremap() with MREMAP_DONTUNMAP could dissociate the VMA
5633	 * from its anon_vma.
5634	 */
5635	if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma))
5636		goto inval_end_read;
5637
5638	/* Check since vm_start/vm_end might change before we lock the VMA */
5639	if (unlikely(address < vma->vm_start || address >= vma->vm_end))
5640		goto inval_end_read;
5641
5642	/* Check if the VMA got isolated after we found it */
5643	if (vma->detached) {
5644		vma_end_read(vma);
5645		count_vm_vma_lock_event(VMA_LOCK_MISS);
5646		/* The area was replaced with another one */
5647		goto retry;
5648	}
5649
5650	rcu_read_unlock();
5651	return vma;
5652
5653inval_end_read:
5654	vma_end_read(vma);
5655inval:
5656	rcu_read_unlock();
5657	count_vm_vma_lock_event(VMA_LOCK_ABORT);
5658	return NULL;
5659}
5660#endif /* CONFIG_PER_VMA_LOCK */
5661
5662#ifndef __PAGETABLE_P4D_FOLDED
5663/*
5664 * Allocate p4d page table.
5665 * We've already handled the fast-path in-line.
5666 */
5667int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
5668{
5669	p4d_t *new = p4d_alloc_one(mm, address);
5670	if (!new)
5671		return -ENOMEM;
5672
 
 
5673	spin_lock(&mm->page_table_lock);
5674	if (pgd_present(*pgd)) {	/* Another has populated it */
5675		p4d_free(mm, new);
5676	} else {
5677		smp_wmb(); /* See comment in pmd_install() */
5678		pgd_populate(mm, pgd, new);
5679	}
5680	spin_unlock(&mm->page_table_lock);
5681	return 0;
5682}
5683#endif /* __PAGETABLE_P4D_FOLDED */
5684
5685#ifndef __PAGETABLE_PUD_FOLDED
5686/*
5687 * Allocate page upper directory.
5688 * We've already handled the fast-path in-line.
5689 */
5690int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
5691{
5692	pud_t *new = pud_alloc_one(mm, address);
5693	if (!new)
5694		return -ENOMEM;
5695
 
 
5696	spin_lock(&mm->page_table_lock);
 
5697	if (!p4d_present(*p4d)) {
5698		mm_inc_nr_puds(mm);
5699		smp_wmb(); /* See comment in pmd_install() */
5700		p4d_populate(mm, p4d, new);
5701	} else	/* Another has populated it */
5702		pud_free(mm, new);
 
 
 
 
 
 
 
5703	spin_unlock(&mm->page_table_lock);
5704	return 0;
5705}
5706#endif /* __PAGETABLE_PUD_FOLDED */
5707
5708#ifndef __PAGETABLE_PMD_FOLDED
5709/*
5710 * Allocate page middle directory.
5711 * We've already handled the fast-path in-line.
5712 */
5713int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
5714{
5715	spinlock_t *ptl;
5716	pmd_t *new = pmd_alloc_one(mm, address);
5717	if (!new)
5718		return -ENOMEM;
5719
 
 
5720	ptl = pud_lock(mm, pud);
 
5721	if (!pud_present(*pud)) {
5722		mm_inc_nr_pmds(mm);
5723		smp_wmb(); /* See comment in pmd_install() */
5724		pud_populate(mm, pud, new);
5725	} else {	/* Another has populated it */
5726		pmd_free(mm, new);
5727	}
 
 
 
 
 
 
5728	spin_unlock(ptl);
5729	return 0;
5730}
5731#endif /* __PAGETABLE_PMD_FOLDED */
5732
5733/**
5734 * follow_pte - look up PTE at a user virtual address
5735 * @mm: the mm_struct of the target address space
5736 * @address: user virtual address
5737 * @ptepp: location to store found PTE
5738 * @ptlp: location to store the lock for the PTE
5739 *
5740 * On a successful return, the pointer to the PTE is stored in @ptepp;
5741 * the corresponding lock is taken and its location is stored in @ptlp.
5742 * The contents of the PTE are only stable until @ptlp is released;
5743 * any further use, if any, must be protected against invalidation
5744 * with MMU notifiers.
5745 *
5746 * Only IO mappings and raw PFN mappings are allowed.  The mmap semaphore
5747 * should be taken for read.
5748 *
5749 * KVM uses this function.  While it is arguably less bad than ``follow_pfn``,
5750 * it is not a good general-purpose API.
5751 *
5752 * Return: zero on success, -ve otherwise.
5753 */
5754int follow_pte(struct mm_struct *mm, unsigned long address,
5755	       pte_t **ptepp, spinlock_t **ptlp)
5756{
5757	pgd_t *pgd;
5758	p4d_t *p4d;
5759	pud_t *pud;
5760	pmd_t *pmd;
5761	pte_t *ptep;
5762
5763	pgd = pgd_offset(mm, address);
5764	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
5765		goto out;
5766
5767	p4d = p4d_offset(pgd, address);
5768	if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
5769		goto out;
5770
5771	pud = pud_offset(p4d, address);
5772	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
5773		goto out;
5774
5775	pmd = pmd_offset(pud, address);
5776	VM_BUG_ON(pmd_trans_huge(*pmd));
5777
5778	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
5779	if (!ptep)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5780		goto out;
5781	if (!pte_present(ptep_get(ptep)))
 
 
 
 
 
 
 
 
5782		goto unlock;
5783	*ptepp = ptep;
5784	return 0;
5785unlock:
5786	pte_unmap_unlock(ptep, *ptlp);
 
 
5787out:
5788	return -EINVAL;
5789}
5790EXPORT_SYMBOL_GPL(follow_pte);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5791
5792/**
5793 * follow_pfn - look up PFN at a user virtual address
5794 * @vma: memory mapping
5795 * @address: user virtual address
5796 * @pfn: location to store found PFN
5797 *
5798 * Only IO mappings and raw PFN mappings are allowed.
5799 *
5800 * This function does not allow the caller to read the permissions
5801 * of the PTE.  Do not use it.
5802 *
5803 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5804 */
5805int follow_pfn(struct vm_area_struct *vma, unsigned long address,
5806	unsigned long *pfn)
5807{
5808	int ret = -EINVAL;
5809	spinlock_t *ptl;
5810	pte_t *ptep;
5811
5812	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5813		return ret;
5814
5815	ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
5816	if (ret)
5817		return ret;
5818	*pfn = pte_pfn(ptep_get(ptep));
5819	pte_unmap_unlock(ptep, ptl);
5820	return 0;
5821}
5822EXPORT_SYMBOL(follow_pfn);
5823
5824#ifdef CONFIG_HAVE_IOREMAP_PROT
5825int follow_phys(struct vm_area_struct *vma,
5826		unsigned long address, unsigned int flags,
5827		unsigned long *prot, resource_size_t *phys)
5828{
5829	int ret = -EINVAL;
5830	pte_t *ptep, pte;
5831	spinlock_t *ptl;
5832
5833	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5834		goto out;
5835
5836	if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
5837		goto out;
5838	pte = ptep_get(ptep);
5839
5840	if ((flags & FOLL_WRITE) && !pte_write(pte))
5841		goto unlock;
5842
5843	*prot = pgprot_val(pte_pgprot(pte));
5844	*phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5845
5846	ret = 0;
5847unlock:
5848	pte_unmap_unlock(ptep, ptl);
5849out:
5850	return ret;
5851}
5852
5853/**
5854 * generic_access_phys - generic implementation for iomem mmap access
5855 * @vma: the vma to access
5856 * @addr: userspace address, not relative offset within @vma
5857 * @buf: buffer to read/write
5858 * @len: length of transfer
5859 * @write: set to FOLL_WRITE when writing, otherwise reading
5860 *
5861 * This is a generic implementation for &vm_operations_struct.access for an
5862 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5863 * not page based.
5864 */
5865int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
5866			void *buf, int len, int write)
5867{
5868	resource_size_t phys_addr;
5869	unsigned long prot = 0;
5870	void __iomem *maddr;
5871	pte_t *ptep, pte;
5872	spinlock_t *ptl;
5873	int offset = offset_in_page(addr);
5874	int ret = -EINVAL;
5875
5876	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5877		return -EINVAL;
5878
5879retry:
5880	if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5881		return -EINVAL;
5882	pte = ptep_get(ptep);
5883	pte_unmap_unlock(ptep, ptl);
5884
5885	prot = pgprot_val(pte_pgprot(pte));
5886	phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5887
5888	if ((write & FOLL_WRITE) && !pte_write(pte))
5889		return -EINVAL;
5890
5891	maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
5892	if (!maddr)
5893		return -ENOMEM;
5894
5895	if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5896		goto out_unmap;
5897
5898	if (!pte_same(pte, ptep_get(ptep))) {
5899		pte_unmap_unlock(ptep, ptl);
5900		iounmap(maddr);
5901
5902		goto retry;
5903	}
5904
5905	if (write)
5906		memcpy_toio(maddr + offset, buf, len);
5907	else
5908		memcpy_fromio(buf, maddr + offset, len);
5909	ret = len;
5910	pte_unmap_unlock(ptep, ptl);
5911out_unmap:
5912	iounmap(maddr);
5913
5914	return ret;
5915}
5916EXPORT_SYMBOL_GPL(generic_access_phys);
5917#endif
5918
5919/*
5920 * Access another process' address space as given in mm.
 
5921 */
5922static int __access_remote_vm(struct mm_struct *mm, unsigned long addr,
5923			      void *buf, int len, unsigned int gup_flags)
5924{
 
5925	void *old_buf = buf;
5926	int write = gup_flags & FOLL_WRITE;
5927
5928	if (mmap_read_lock_killable(mm))
5929		return 0;
5930
5931	/* Untag the address before looking up the VMA */
5932	addr = untagged_addr_remote(mm, addr);
5933
5934	/* Avoid triggering the temporary warning in __get_user_pages */
5935	if (!vma_lookup(mm, addr) && !expand_stack(mm, addr))
5936		return 0;
5937
5938	/* ignore errors, just check how much was successfully transferred */
5939	while (len) {
5940		int bytes, offset;
5941		void *maddr;
5942		struct vm_area_struct *vma = NULL;
5943		struct page *page = get_user_page_vma_remote(mm, addr,
5944							     gup_flags, &vma);
5945
5946		if (IS_ERR(page)) {
5947			/* We might need to expand the stack to access it */
5948			vma = vma_lookup(mm, addr);
5949			if (!vma) {
5950				vma = expand_stack(mm, addr);
5951
5952				/* mmap_lock was dropped on failure */
5953				if (!vma)
5954					return buf - old_buf;
5955
5956				/* Try again if stack expansion worked */
5957				continue;
5958			}
5959
 
 
 
 
 
 
5960			/*
5961			 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5962			 * we can access using slightly different code.
5963			 */
5964			bytes = 0;
5965#ifdef CONFIG_HAVE_IOREMAP_PROT
 
5966			if (vma->vm_ops && vma->vm_ops->access)
5967				bytes = vma->vm_ops->access(vma, addr, buf,
5968							    len, write);
5969#endif
5970			if (bytes <= 0)
5971				break;
 
 
5972		} else {
5973			bytes = len;
5974			offset = addr & (PAGE_SIZE-1);
5975			if (bytes > PAGE_SIZE-offset)
5976				bytes = PAGE_SIZE-offset;
5977
5978			maddr = kmap_local_page(page);
5979			if (write) {
5980				copy_to_user_page(vma, page, addr,
5981						  maddr + offset, buf, bytes);
5982				set_page_dirty_lock(page);
5983			} else {
5984				copy_from_user_page(vma, page, addr,
5985						    buf, maddr + offset, bytes);
5986			}
5987			unmap_and_put_page(page, maddr);
 
5988		}
5989		len -= bytes;
5990		buf += bytes;
5991		addr += bytes;
5992	}
5993	mmap_read_unlock(mm);
5994
5995	return buf - old_buf;
5996}
5997
5998/**
5999 * access_remote_vm - access another process' address space
6000 * @mm:		the mm_struct of the target address space
6001 * @addr:	start address to access
6002 * @buf:	source or destination buffer
6003 * @len:	number of bytes to transfer
6004 * @gup_flags:	flags modifying lookup behaviour
6005 *
6006 * The caller must hold a reference on @mm.
6007 *
6008 * Return: number of bytes copied from source to destination.
6009 */
6010int access_remote_vm(struct mm_struct *mm, unsigned long addr,
6011		void *buf, int len, unsigned int gup_flags)
6012{
6013	return __access_remote_vm(mm, addr, buf, len, gup_flags);
6014}
6015
6016/*
6017 * Access another process' address space.
6018 * Source/target buffer must be kernel space,
6019 * Do not walk the page table directly, use get_user_pages
6020 */
6021int access_process_vm(struct task_struct *tsk, unsigned long addr,
6022		void *buf, int len, unsigned int gup_flags)
6023{
6024	struct mm_struct *mm;
6025	int ret;
6026
6027	mm = get_task_mm(tsk);
6028	if (!mm)
6029		return 0;
6030
6031	ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
6032
6033	mmput(mm);
6034
6035	return ret;
6036}
6037EXPORT_SYMBOL_GPL(access_process_vm);
6038
6039/*
6040 * Print the name of a VMA.
6041 */
6042void print_vma_addr(char *prefix, unsigned long ip)
6043{
6044	struct mm_struct *mm = current->mm;
6045	struct vm_area_struct *vma;
6046
6047	/*
6048	 * we might be running from an atomic context so we cannot sleep
6049	 */
6050	if (!mmap_read_trylock(mm))
6051		return;
6052
6053	vma = find_vma(mm, ip);
6054	if (vma && vma->vm_file) {
6055		struct file *f = vma->vm_file;
6056		char *buf = (char *)__get_free_page(GFP_NOWAIT);
6057		if (buf) {
6058			char *p;
6059
6060			p = file_path(f, buf, PAGE_SIZE);
6061			if (IS_ERR(p))
6062				p = "?";
6063			printk("%s%s[%lx+%lx]", prefix, kbasename(p),
6064					vma->vm_start,
6065					vma->vm_end - vma->vm_start);
6066			free_page((unsigned long)buf);
6067		}
6068	}
6069	mmap_read_unlock(mm);
6070}
6071
6072#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6073void __might_fault(const char *file, int line)
6074{
 
 
 
 
 
 
 
 
6075	if (pagefault_disabled())
6076		return;
6077	__might_sleep(file, line);
6078#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6079	if (current->mm)
6080		might_lock_read(&current->mm->mmap_lock);
6081#endif
6082}
6083EXPORT_SYMBOL(__might_fault);
6084#endif
6085
6086#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
6087/*
6088 * Process all subpages of the specified huge page with the specified
6089 * operation.  The target subpage will be processed last to keep its
6090 * cache lines hot.
6091 */
6092static inline int process_huge_page(
6093	unsigned long addr_hint, unsigned int pages_per_huge_page,
6094	int (*process_subpage)(unsigned long addr, int idx, void *arg),
6095	void *arg)
6096{
6097	int i, n, base, l, ret;
6098	unsigned long addr = addr_hint &
6099		~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6100
6101	/* Process target subpage last to keep its cache lines hot */
6102	might_sleep();
6103	n = (addr_hint - addr) / PAGE_SIZE;
6104	if (2 * n <= pages_per_huge_page) {
6105		/* If target subpage in first half of huge page */
6106		base = 0;
6107		l = n;
6108		/* Process subpages at the end of huge page */
6109		for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
6110			cond_resched();
6111			ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
6112			if (ret)
6113				return ret;
6114		}
6115	} else {
6116		/* If target subpage in second half of huge page */
6117		base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
6118		l = pages_per_huge_page - n;
6119		/* Process subpages at the begin of huge page */
6120		for (i = 0; i < base; i++) {
6121			cond_resched();
6122			ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
6123			if (ret)
6124				return ret;
6125		}
6126	}
6127	/*
6128	 * Process remaining subpages in left-right-left-right pattern
6129	 * towards the target subpage
6130	 */
6131	for (i = 0; i < l; i++) {
6132		int left_idx = base + i;
6133		int right_idx = base + 2 * l - 1 - i;
6134
6135		cond_resched();
6136		ret = process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
6137		if (ret)
6138			return ret;
6139		cond_resched();
6140		ret = process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
6141		if (ret)
6142			return ret;
6143	}
6144	return 0;
6145}
6146
6147static void clear_gigantic_page(struct page *page,
6148				unsigned long addr,
6149				unsigned int pages_per_huge_page)
6150{
6151	int i;
6152	struct page *p;
6153
6154	might_sleep();
6155	for (i = 0; i < pages_per_huge_page; i++) {
6156		p = nth_page(page, i);
6157		cond_resched();
6158		clear_user_highpage(p, addr + i * PAGE_SIZE);
6159	}
6160}
6161
6162static int clear_subpage(unsigned long addr, int idx, void *arg)
6163{
6164	struct page *page = arg;
6165
6166	clear_user_highpage(page + idx, addr);
6167	return 0;
6168}
6169
6170void clear_huge_page(struct page *page,
6171		     unsigned long addr_hint, unsigned int pages_per_huge_page)
6172{
6173	unsigned long addr = addr_hint &
6174		~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6175
6176	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
6177		clear_gigantic_page(page, addr, pages_per_huge_page);
6178		return;
6179	}
6180
6181	process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
6182}
6183
6184static int copy_user_gigantic_page(struct folio *dst, struct folio *src,
6185				     unsigned long addr,
6186				     struct vm_area_struct *vma,
6187				     unsigned int pages_per_huge_page)
6188{
6189	int i;
6190	struct page *dst_page;
6191	struct page *src_page;
6192
6193	for (i = 0; i < pages_per_huge_page; i++) {
6194		dst_page = folio_page(dst, i);
6195		src_page = folio_page(src, i);
6196
 
6197		cond_resched();
6198		if (copy_mc_user_highpage(dst_page, src_page,
6199					  addr + i*PAGE_SIZE, vma)) {
6200			memory_failure_queue(page_to_pfn(src_page), 0);
6201			return -EHWPOISON;
6202		}
6203	}
6204	return 0;
6205}
6206
6207struct copy_subpage_arg {
6208	struct page *dst;
6209	struct page *src;
6210	struct vm_area_struct *vma;
6211};
6212
6213static int copy_subpage(unsigned long addr, int idx, void *arg)
6214{
6215	struct copy_subpage_arg *copy_arg = arg;
6216
6217	if (copy_mc_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
6218				  addr, copy_arg->vma)) {
6219		memory_failure_queue(page_to_pfn(copy_arg->src + idx), 0);
6220		return -EHWPOISON;
6221	}
6222	return 0;
6223}
6224
6225int copy_user_large_folio(struct folio *dst, struct folio *src,
6226			  unsigned long addr_hint, struct vm_area_struct *vma)
 
6227{
6228	unsigned int pages_per_huge_page = folio_nr_pages(dst);
6229	unsigned long addr = addr_hint &
6230		~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6231	struct copy_subpage_arg arg = {
6232		.dst = &dst->page,
6233		.src = &src->page,
6234		.vma = vma,
6235	};
6236
6237	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES))
6238		return copy_user_gigantic_page(dst, src, addr, vma,
6239					       pages_per_huge_page);
 
 
6240
6241	return process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
6242}
6243
6244long copy_folio_from_user(struct folio *dst_folio,
6245			   const void __user *usr_src,
6246			   bool allow_pagefault)
 
6247{
6248	void *kaddr;
 
6249	unsigned long i, rc = 0;
6250	unsigned int nr_pages = folio_nr_pages(dst_folio);
6251	unsigned long ret_val = nr_pages * PAGE_SIZE;
6252	struct page *subpage;
6253
6254	for (i = 0; i < nr_pages; i++) {
6255		subpage = folio_page(dst_folio, i);
6256		kaddr = kmap_local_page(subpage);
6257		if (!allow_pagefault)
6258			pagefault_disable();
6259		rc = copy_from_user(kaddr, usr_src + i * PAGE_SIZE, PAGE_SIZE);
6260		if (!allow_pagefault)
6261			pagefault_enable();
6262		kunmap_local(kaddr);
 
6263
6264		ret_val -= (PAGE_SIZE - rc);
6265		if (rc)
6266			break;
6267
6268		flush_dcache_page(subpage);
6269
6270		cond_resched();
6271	}
6272	return ret_val;
6273}
6274#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
6275
6276#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
6277
6278static struct kmem_cache *page_ptl_cachep;
6279
6280void __init ptlock_cache_init(void)
6281{
6282	page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
6283			SLAB_PANIC, NULL);
6284}
6285
6286bool ptlock_alloc(struct ptdesc *ptdesc)
6287{
6288	spinlock_t *ptl;
6289
6290	ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
6291	if (!ptl)
6292		return false;
6293	ptdesc->ptl = ptl;
6294	return true;
6295}
6296
6297void ptlock_free(struct ptdesc *ptdesc)
6298{
6299	kmem_cache_free(page_ptl_cachep, ptdesc->ptl);
6300}
6301#endif
v5.4
 
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *  linux/mm/memory.c
   4 *
   5 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   6 */
   7
   8/*
   9 * demand-loading started 01.12.91 - seems it is high on the list of
  10 * things wanted, and it should be easy to implement. - Linus
  11 */
  12
  13/*
  14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
  15 * pages started 02.12.91, seems to work. - Linus.
  16 *
  17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
  18 * would have taken more than the 6M I have free, but it worked well as
  19 * far as I could see.
  20 *
  21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
  22 */
  23
  24/*
  25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
  26 * thought has to go into this. Oh, well..
  27 * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
  28 *		Found it. Everything seems to work now.
  29 * 20.12.91  -  Ok, making the swap-device changeable like the root.
  30 */
  31
  32/*
  33 * 05.04.94  -  Multi-page memory management added for v1.1.
  34 *              Idea by Alex Bligh (alex@cconcepts.co.uk)
  35 *
  36 * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
  37 *		(Gerhard.Wichert@pdb.siemens.de)
  38 *
  39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
  40 */
  41
  42#include <linux/kernel_stat.h>
  43#include <linux/mm.h>
 
  44#include <linux/sched/mm.h>
  45#include <linux/sched/coredump.h>
  46#include <linux/sched/numa_balancing.h>
  47#include <linux/sched/task.h>
  48#include <linux/hugetlb.h>
  49#include <linux/mman.h>
  50#include <linux/swap.h>
  51#include <linux/highmem.h>
  52#include <linux/pagemap.h>
  53#include <linux/memremap.h>
 
  54#include <linux/ksm.h>
  55#include <linux/rmap.h>
  56#include <linux/export.h>
  57#include <linux/delayacct.h>
  58#include <linux/init.h>
  59#include <linux/pfn_t.h>
  60#include <linux/writeback.h>
  61#include <linux/memcontrol.h>
  62#include <linux/mmu_notifier.h>
  63#include <linux/swapops.h>
  64#include <linux/elf.h>
  65#include <linux/gfp.h>
  66#include <linux/migrate.h>
  67#include <linux/string.h>
  68#include <linux/dma-debug.h>
  69#include <linux/debugfs.h>
  70#include <linux/userfaultfd_k.h>
  71#include <linux/dax.h>
  72#include <linux/oom.h>
  73#include <linux/numa.h>
 
 
 
 
 
 
  74
  75#include <asm/io.h>
  76#include <asm/mmu_context.h>
  77#include <asm/pgalloc.h>
  78#include <linux/uaccess.h>
  79#include <asm/tlb.h>
  80#include <asm/tlbflush.h>
  81#include <asm/pgtable.h>
  82
 
  83#include "internal.h"
 
  84
  85#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
  86#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
  87#endif
  88
  89#ifndef CONFIG_NEED_MULTIPLE_NODES
  90/* use the per-pgdat data instead for discontigmem - mbligh */
  91unsigned long max_mapnr;
  92EXPORT_SYMBOL(max_mapnr);
  93
  94struct page *mem_map;
  95EXPORT_SYMBOL(mem_map);
  96#endif
  97
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  98/*
  99 * A number of key systems in x86 including ioremap() rely on the assumption
 100 * that high_memory defines the upper bound on direct map memory, then end
 101 * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
 102 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
 103 * and ZONE_HIGHMEM.
 104 */
 105void *high_memory;
 106EXPORT_SYMBOL(high_memory);
 107
 108/*
 109 * Randomize the address space (stacks, mmaps, brk, etc.).
 110 *
 111 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
 112 *   as ancient (libc5 based) binaries can segfault. )
 113 */
 114int randomize_va_space __read_mostly =
 115#ifdef CONFIG_COMPAT_BRK
 116					1;
 117#else
 118					2;
 119#endif
 120
 
 
 
 
 
 
 
 
 
 
 
 
 121static int __init disable_randmaps(char *s)
 122{
 123	randomize_va_space = 0;
 124	return 1;
 125}
 126__setup("norandmaps", disable_randmaps);
 127
 128unsigned long zero_pfn __read_mostly;
 129EXPORT_SYMBOL(zero_pfn);
 130
 131unsigned long highest_memmap_pfn __read_mostly;
 132
 133/*
 134 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
 135 */
 136static int __init init_zero_pfn(void)
 137{
 138	zero_pfn = page_to_pfn(ZERO_PAGE(0));
 139	return 0;
 140}
 141core_initcall(init_zero_pfn);
 142
 143
 144#if defined(SPLIT_RSS_COUNTING)
 145
 146void sync_mm_rss(struct mm_struct *mm)
 147{
 148	int i;
 149
 150	for (i = 0; i < NR_MM_COUNTERS; i++) {
 151		if (current->rss_stat.count[i]) {
 152			add_mm_counter(mm, i, current->rss_stat.count[i]);
 153			current->rss_stat.count[i] = 0;
 154		}
 155	}
 156	current->rss_stat.events = 0;
 157}
 158
 159static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
 160{
 161	struct task_struct *task = current;
 162
 163	if (likely(task->mm == mm))
 164		task->rss_stat.count[member] += val;
 165	else
 166		add_mm_counter(mm, member, val);
 167}
 168#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
 169#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
 170
 171/* sync counter once per 64 page faults */
 172#define TASK_RSS_EVENTS_THRESH	(64)
 173static void check_sync_rss_stat(struct task_struct *task)
 174{
 175	if (unlikely(task != current))
 176		return;
 177	if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
 178		sync_mm_rss(task->mm);
 179}
 180#else /* SPLIT_RSS_COUNTING */
 181
 182#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
 183#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
 184
 185static void check_sync_rss_stat(struct task_struct *task)
 186{
 
 187}
 188
 189#endif /* SPLIT_RSS_COUNTING */
 190
 191/*
 192 * Note: this doesn't free the actual pages themselves. That
 193 * has been handled earlier when unmapping all the memory regions.
 194 */
 195static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 196			   unsigned long addr)
 197{
 198	pgtable_t token = pmd_pgtable(*pmd);
 199	pmd_clear(pmd);
 200	pte_free_tlb(tlb, token, addr);
 201	mm_dec_nr_ptes(tlb->mm);
 202}
 203
 204static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 205				unsigned long addr, unsigned long end,
 206				unsigned long floor, unsigned long ceiling)
 207{
 208	pmd_t *pmd;
 209	unsigned long next;
 210	unsigned long start;
 211
 212	start = addr;
 213	pmd = pmd_offset(pud, addr);
 214	do {
 215		next = pmd_addr_end(addr, end);
 216		if (pmd_none_or_clear_bad(pmd))
 217			continue;
 218		free_pte_range(tlb, pmd, addr);
 219	} while (pmd++, addr = next, addr != end);
 220
 221	start &= PUD_MASK;
 222	if (start < floor)
 223		return;
 224	if (ceiling) {
 225		ceiling &= PUD_MASK;
 226		if (!ceiling)
 227			return;
 228	}
 229	if (end - 1 > ceiling - 1)
 230		return;
 231
 232	pmd = pmd_offset(pud, start);
 233	pud_clear(pud);
 234	pmd_free_tlb(tlb, pmd, start);
 235	mm_dec_nr_pmds(tlb->mm);
 236}
 237
 238static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
 239				unsigned long addr, unsigned long end,
 240				unsigned long floor, unsigned long ceiling)
 241{
 242	pud_t *pud;
 243	unsigned long next;
 244	unsigned long start;
 245
 246	start = addr;
 247	pud = pud_offset(p4d, addr);
 248	do {
 249		next = pud_addr_end(addr, end);
 250		if (pud_none_or_clear_bad(pud))
 251			continue;
 252		free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 253	} while (pud++, addr = next, addr != end);
 254
 255	start &= P4D_MASK;
 256	if (start < floor)
 257		return;
 258	if (ceiling) {
 259		ceiling &= P4D_MASK;
 260		if (!ceiling)
 261			return;
 262	}
 263	if (end - 1 > ceiling - 1)
 264		return;
 265
 266	pud = pud_offset(p4d, start);
 267	p4d_clear(p4d);
 268	pud_free_tlb(tlb, pud, start);
 269	mm_dec_nr_puds(tlb->mm);
 270}
 271
 272static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
 273				unsigned long addr, unsigned long end,
 274				unsigned long floor, unsigned long ceiling)
 275{
 276	p4d_t *p4d;
 277	unsigned long next;
 278	unsigned long start;
 279
 280	start = addr;
 281	p4d = p4d_offset(pgd, addr);
 282	do {
 283		next = p4d_addr_end(addr, end);
 284		if (p4d_none_or_clear_bad(p4d))
 285			continue;
 286		free_pud_range(tlb, p4d, addr, next, floor, ceiling);
 287	} while (p4d++, addr = next, addr != end);
 288
 289	start &= PGDIR_MASK;
 290	if (start < floor)
 291		return;
 292	if (ceiling) {
 293		ceiling &= PGDIR_MASK;
 294		if (!ceiling)
 295			return;
 296	}
 297	if (end - 1 > ceiling - 1)
 298		return;
 299
 300	p4d = p4d_offset(pgd, start);
 301	pgd_clear(pgd);
 302	p4d_free_tlb(tlb, p4d, start);
 303}
 304
 305/*
 306 * This function frees user-level page tables of a process.
 307 */
 308void free_pgd_range(struct mmu_gather *tlb,
 309			unsigned long addr, unsigned long end,
 310			unsigned long floor, unsigned long ceiling)
 311{
 312	pgd_t *pgd;
 313	unsigned long next;
 314
 315	/*
 316	 * The next few lines have given us lots of grief...
 317	 *
 318	 * Why are we testing PMD* at this top level?  Because often
 319	 * there will be no work to do at all, and we'd prefer not to
 320	 * go all the way down to the bottom just to discover that.
 321	 *
 322	 * Why all these "- 1"s?  Because 0 represents both the bottom
 323	 * of the address space and the top of it (using -1 for the
 324	 * top wouldn't help much: the masks would do the wrong thing).
 325	 * The rule is that addr 0 and floor 0 refer to the bottom of
 326	 * the address space, but end 0 and ceiling 0 refer to the top
 327	 * Comparisons need to use "end - 1" and "ceiling - 1" (though
 328	 * that end 0 case should be mythical).
 329	 *
 330	 * Wherever addr is brought up or ceiling brought down, we must
 331	 * be careful to reject "the opposite 0" before it confuses the
 332	 * subsequent tests.  But what about where end is brought down
 333	 * by PMD_SIZE below? no, end can't go down to 0 there.
 334	 *
 335	 * Whereas we round start (addr) and ceiling down, by different
 336	 * masks at different levels, in order to test whether a table
 337	 * now has no other vmas using it, so can be freed, we don't
 338	 * bother to round floor or end up - the tests don't need that.
 339	 */
 340
 341	addr &= PMD_MASK;
 342	if (addr < floor) {
 343		addr += PMD_SIZE;
 344		if (!addr)
 345			return;
 346	}
 347	if (ceiling) {
 348		ceiling &= PMD_MASK;
 349		if (!ceiling)
 350			return;
 351	}
 352	if (end - 1 > ceiling - 1)
 353		end -= PMD_SIZE;
 354	if (addr > end - 1)
 355		return;
 356	/*
 357	 * We add page table cache pages with PAGE_SIZE,
 358	 * (see pte_free_tlb()), flush the tlb if we need
 359	 */
 360	tlb_change_page_size(tlb, PAGE_SIZE);
 361	pgd = pgd_offset(tlb->mm, addr);
 362	do {
 363		next = pgd_addr_end(addr, end);
 364		if (pgd_none_or_clear_bad(pgd))
 365			continue;
 366		free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
 367	} while (pgd++, addr = next, addr != end);
 368}
 369
 370void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
 371		unsigned long floor, unsigned long ceiling)
 
 372{
 373	while (vma) {
 374		struct vm_area_struct *next = vma->vm_next;
 375		unsigned long addr = vma->vm_start;
 
 
 
 
 
 
 
 
 
 376
 377		/*
 378		 * Hide vma from rmap and truncate_pagecache before freeing
 379		 * pgtables
 380		 */
 
 
 381		unlink_anon_vmas(vma);
 382		unlink_file_vma(vma);
 383
 384		if (is_vm_hugetlb_page(vma)) {
 385			hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 386				floor, next ? next->vm_start : ceiling);
 387		} else {
 388			/*
 389			 * Optimization: gather nearby vmas into one call down
 390			 */
 391			while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 392			       && !is_vm_hugetlb_page(next)) {
 393				vma = next;
 394				next = vma->vm_next;
 
 
 
 
 395				unlink_anon_vmas(vma);
 396				unlink_file_vma(vma);
 397			}
 398			free_pgd_range(tlb, addr, vma->vm_end,
 399				floor, next ? next->vm_start : ceiling);
 400		}
 401		vma = next;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 402	}
 
 403}
 404
 405int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
 406{
 407	spinlock_t *ptl;
 408	pgtable_t new = pte_alloc_one(mm);
 409	if (!new)
 410		return -ENOMEM;
 411
 412	/*
 413	 * Ensure all pte setup (eg. pte page lock and page clearing) are
 414	 * visible before the pte is made visible to other CPUs by being
 415	 * put into page tables.
 416	 *
 417	 * The other side of the story is the pointer chasing in the page
 418	 * table walking code (when walking the page table without locking;
 419	 * ie. most of the time). Fortunately, these data accesses consist
 420	 * of a chain of data-dependent loads, meaning most CPUs (alpha
 421	 * being the notable exception) will already guarantee loads are
 422	 * seen in-order. See the alpha page table accessors for the
 423	 * smp_read_barrier_depends() barriers in page table walking code.
 424	 */
 425	smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
 426
 427	ptl = pmd_lock(mm, pmd);
 428	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
 429		mm_inc_nr_ptes(mm);
 430		pmd_populate(mm, pmd, new);
 431		new = NULL;
 432	}
 433	spin_unlock(ptl);
 434	if (new)
 435		pte_free(mm, new);
 436	return 0;
 437}
 438
 439int __pte_alloc_kernel(pmd_t *pmd)
 440{
 441	pte_t *new = pte_alloc_one_kernel(&init_mm);
 442	if (!new)
 443		return -ENOMEM;
 444
 445	smp_wmb(); /* See comment in __pte_alloc */
 446
 447	spin_lock(&init_mm.page_table_lock);
 448	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
 
 449		pmd_populate_kernel(&init_mm, pmd, new);
 450		new = NULL;
 451	}
 452	spin_unlock(&init_mm.page_table_lock);
 453	if (new)
 454		pte_free_kernel(&init_mm, new);
 455	return 0;
 456}
 457
 458static inline void init_rss_vec(int *rss)
 459{
 460	memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
 461}
 462
 463static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
 464{
 465	int i;
 466
 467	if (current->mm == mm)
 468		sync_mm_rss(mm);
 469	for (i = 0; i < NR_MM_COUNTERS; i++)
 470		if (rss[i])
 471			add_mm_counter(mm, i, rss[i]);
 472}
 473
 474/*
 475 * This function is called to print an error when a bad pte
 476 * is found. For example, we might have a PFN-mapped pte in
 477 * a region that doesn't allow it.
 478 *
 479 * The calling function must still handle the error.
 480 */
 481static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
 482			  pte_t pte, struct page *page)
 483{
 484	pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
 485	p4d_t *p4d = p4d_offset(pgd, addr);
 486	pud_t *pud = pud_offset(p4d, addr);
 487	pmd_t *pmd = pmd_offset(pud, addr);
 488	struct address_space *mapping;
 489	pgoff_t index;
 490	static unsigned long resume;
 491	static unsigned long nr_shown;
 492	static unsigned long nr_unshown;
 493
 494	/*
 495	 * Allow a burst of 60 reports, then keep quiet for that minute;
 496	 * or allow a steady drip of one report per second.
 497	 */
 498	if (nr_shown == 60) {
 499		if (time_before(jiffies, resume)) {
 500			nr_unshown++;
 501			return;
 502		}
 503		if (nr_unshown) {
 504			pr_alert("BUG: Bad page map: %lu messages suppressed\n",
 505				 nr_unshown);
 506			nr_unshown = 0;
 507		}
 508		nr_shown = 0;
 509	}
 510	if (nr_shown++ == 0)
 511		resume = jiffies + 60 * HZ;
 512
 513	mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
 514	index = linear_page_index(vma, addr);
 515
 516	pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
 517		 current->comm,
 518		 (long long)pte_val(pte), (long long)pmd_val(*pmd));
 519	if (page)
 520		dump_page(page, "bad pte");
 521	pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
 522		 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
 523	pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
 524		 vma->vm_file,
 525		 vma->vm_ops ? vma->vm_ops->fault : NULL,
 526		 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
 527		 mapping ? mapping->a_ops->readpage : NULL);
 528	dump_stack();
 529	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 530}
 531
 532/*
 533 * vm_normal_page -- This function gets the "struct page" associated with a pte.
 534 *
 535 * "Special" mappings do not wish to be associated with a "struct page" (either
 536 * it doesn't exist, or it exists but they don't want to touch it). In this
 537 * case, NULL is returned here. "Normal" mappings do have a struct page.
 538 *
 539 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
 540 * pte bit, in which case this function is trivial. Secondly, an architecture
 541 * may not have a spare pte bit, which requires a more complicated scheme,
 542 * described below.
 543 *
 544 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
 545 * special mapping (even if there are underlying and valid "struct pages").
 546 * COWed pages of a VM_PFNMAP are always normal.
 547 *
 548 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
 549 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
 550 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
 551 * mapping will always honor the rule
 552 *
 553 *	pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 554 *
 555 * And for normal mappings this is false.
 556 *
 557 * This restricts such mappings to be a linear translation from virtual address
 558 * to pfn. To get around this restriction, we allow arbitrary mappings so long
 559 * as the vma is not a COW mapping; in that case, we know that all ptes are
 560 * special (because none can have been COWed).
 561 *
 562 *
 563 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
 564 *
 565 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
 566 * page" backing, however the difference is that _all_ pages with a struct
 567 * page (that is, those where pfn_valid is true) are refcounted and considered
 568 * normal pages by the VM. The disadvantage is that pages are refcounted
 569 * (which can be slower and simply not an option for some PFNMAP users). The
 570 * advantage is that we don't have to follow the strict linearity rule of
 571 * PFNMAP mappings in order to support COWable mappings.
 572 *
 573 */
 574struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 575			    pte_t pte)
 576{
 577	unsigned long pfn = pte_pfn(pte);
 578
 579	if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
 580		if (likely(!pte_special(pte)))
 581			goto check_pfn;
 582		if (vma->vm_ops && vma->vm_ops->find_special_page)
 583			return vma->vm_ops->find_special_page(vma, addr);
 584		if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
 585			return NULL;
 586		if (is_zero_pfn(pfn))
 587			return NULL;
 588		if (pte_devmap(pte))
 
 
 
 
 
 
 
 
 589			return NULL;
 590
 591		print_bad_pte(vma, addr, pte, NULL);
 592		return NULL;
 593	}
 594
 595	/* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
 596
 597	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 598		if (vma->vm_flags & VM_MIXEDMAP) {
 599			if (!pfn_valid(pfn))
 600				return NULL;
 601			goto out;
 602		} else {
 603			unsigned long off;
 604			off = (addr - vma->vm_start) >> PAGE_SHIFT;
 605			if (pfn == vma->vm_pgoff + off)
 606				return NULL;
 607			if (!is_cow_mapping(vma->vm_flags))
 608				return NULL;
 609		}
 610	}
 611
 612	if (is_zero_pfn(pfn))
 613		return NULL;
 614
 615check_pfn:
 616	if (unlikely(pfn > highest_memmap_pfn)) {
 617		print_bad_pte(vma, addr, pte, NULL);
 618		return NULL;
 619	}
 620
 621	/*
 622	 * NOTE! We still have PageReserved() pages in the page tables.
 623	 * eg. VDSO mappings can cause them to exist.
 624	 */
 625out:
 626	return pfn_to_page(pfn);
 627}
 628
 
 
 
 
 
 
 
 
 
 
 629#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 630struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
 631				pmd_t pmd)
 632{
 633	unsigned long pfn = pmd_pfn(pmd);
 634
 635	/*
 636	 * There is no pmd_special() but there may be special pmds, e.g.
 637	 * in a direct-access (dax) mapping, so let's just replicate the
 638	 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
 639	 */
 640	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 641		if (vma->vm_flags & VM_MIXEDMAP) {
 642			if (!pfn_valid(pfn))
 643				return NULL;
 644			goto out;
 645		} else {
 646			unsigned long off;
 647			off = (addr - vma->vm_start) >> PAGE_SHIFT;
 648			if (pfn == vma->vm_pgoff + off)
 649				return NULL;
 650			if (!is_cow_mapping(vma->vm_flags))
 651				return NULL;
 652		}
 653	}
 654
 655	if (pmd_devmap(pmd))
 656		return NULL;
 657	if (is_zero_pfn(pfn))
 658		return NULL;
 659	if (unlikely(pfn > highest_memmap_pfn))
 660		return NULL;
 661
 662	/*
 663	 * NOTE! We still have PageReserved() pages in the page tables.
 664	 * eg. VDSO mappings can cause them to exist.
 665	 */
 666out:
 667	return pfn_to_page(pfn);
 668}
 
 
 
 
 
 
 
 
 
 
 669#endif
 670
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 671/*
 672 * copy one vm_area from one task to the other. Assumes the page tables
 673 * already present in the new task to be cleared in the whole range
 674 * covered by this vma.
 675 */
 676
 677static inline unsigned long
 678copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 679		pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 680		unsigned long addr, int *rss)
 681{
 682	unsigned long vm_flags = vma->vm_flags;
 683	pte_t pte = *src_pte;
 
 
 684	struct page *page;
 
 685
 686	/* pte contains position in swap or file, so copy. */
 687	if (unlikely(!pte_present(pte))) {
 688		swp_entry_t entry = pte_to_swp_entry(pte);
 689
 690		if (likely(!non_swap_entry(entry))) {
 691			if (swap_duplicate(entry) < 0)
 692				return entry.val;
 693
 694			/* make sure dst_mm is on swapoff's mmlist. */
 695			if (unlikely(list_empty(&dst_mm->mmlist))) {
 696				spin_lock(&mmlist_lock);
 697				if (list_empty(&dst_mm->mmlist))
 698					list_add(&dst_mm->mmlist,
 699							&src_mm->mmlist);
 700				spin_unlock(&mmlist_lock);
 701			}
 702			rss[MM_SWAPENTS]++;
 703		} else if (is_migration_entry(entry)) {
 704			page = migration_entry_to_page(entry);
 705
 706			rss[mm_counter(page)]++;
 707
 708			if (is_write_migration_entry(entry) &&
 709					is_cow_mapping(vm_flags)) {
 710				/*
 711				 * COW mappings require pages in both
 712				 * parent and child to be set to read.
 713				 */
 714				make_migration_entry_read(&entry);
 715				pte = swp_entry_to_pte(entry);
 716				if (pte_swp_soft_dirty(*src_pte))
 717					pte = pte_swp_mksoft_dirty(pte);
 718				set_pte_at(src_mm, addr, src_pte, pte);
 719			}
 720		} else if (is_device_private_entry(entry)) {
 721			page = device_private_entry_to_page(entry);
 722
 
 
 723			/*
 724			 * Update rss count even for unaddressable pages, as
 725			 * they should treated just like normal pages in this
 726			 * respect.
 727			 *
 728			 * We will likely want to have some new rss counters
 729			 * for unaddressable pages, at some point. But for now
 730			 * keep things as they are.
 731			 */
 732			get_page(page);
 733			rss[mm_counter(page)]++;
 734			page_dup_rmap(page, false);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 735
 736			/*
 737			 * We do not preserve soft-dirty information, because so
 738			 * far, checkpoint/restore is the only feature that
 739			 * requires that. And checkpoint/restore does not work
 740			 * when a device driver is involved (you cannot easily
 741			 * save and restore device driver state).
 742			 */
 743			if (is_write_device_private_entry(entry) &&
 744			    is_cow_mapping(vm_flags)) {
 745				make_device_private_entry_read(&entry);
 746				pte = swp_entry_to_pte(entry);
 747				set_pte_at(src_mm, addr, src_pte, pte);
 748			}
 
 
 749		}
 750		goto out_set_pte;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 751	}
 752
 753	/*
 754	 * If it's a COW mapping, write protect it both
 755	 * in the parent and the child
 756	 */
 757	if (is_cow_mapping(vm_flags) && pte_write(pte)) {
 758		ptep_set_wrprotect(src_mm, addr, src_pte);
 759		pte = pte_wrprotect(pte);
 760	}
 
 761
 762	/*
 763	 * If it's a shared mapping, mark it clean in
 764	 * the child
 765	 */
 766	if (vm_flags & VM_SHARED)
 767		pte = pte_mkclean(pte);
 768	pte = pte_mkold(pte);
 769
 770	page = vm_normal_page(vma, addr, pte);
 771	if (page) {
 772		get_page(page);
 773		page_dup_rmap(page, false);
 774		rss[mm_counter(page)]++;
 775	} else if (pte_devmap(pte)) {
 776		page = pte_page(pte);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 777	}
 
 778
 779out_set_pte:
 780	set_pte_at(dst_mm, addr, dst_pte, pte);
 781	return 0;
 782}
 783
 784static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 785		   pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 786		   unsigned long addr, unsigned long end)
 
 787{
 
 
 788	pte_t *orig_src_pte, *orig_dst_pte;
 789	pte_t *src_pte, *dst_pte;
 
 790	spinlock_t *src_ptl, *dst_ptl;
 791	int progress = 0;
 792	int rss[NR_MM_COUNTERS];
 793	swp_entry_t entry = (swp_entry_t){0};
 
 794
 795again:
 
 796	init_rss_vec(rss);
 797
 
 
 
 
 
 
 
 
 798	dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
 799	if (!dst_pte)
 800		return -ENOMEM;
 801	src_pte = pte_offset_map(src_pmd, addr);
 802	src_ptl = pte_lockptr(src_mm, src_pmd);
 
 
 
 
 
 
 803	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 804	orig_src_pte = src_pte;
 805	orig_dst_pte = dst_pte;
 806	arch_enter_lazy_mmu_mode();
 807
 808	do {
 809		/*
 810		 * We are holding two locks at this point - either of them
 811		 * could generate latencies in another task on another CPU.
 812		 */
 813		if (progress >= 32) {
 814			progress = 0;
 815			if (need_resched() ||
 816			    spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
 817				break;
 818		}
 819		if (pte_none(*src_pte)) {
 
 820			progress++;
 821			continue;
 822		}
 823		entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
 824							vma, addr, rss);
 825		if (entry.val)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 826			break;
 
 
 
 
 
 
 
 
 
 
 827		progress += 8;
 828	} while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
 829
 830	arch_leave_lazy_mmu_mode();
 831	spin_unlock(src_ptl);
 832	pte_unmap(orig_src_pte);
 833	add_mm_rss_vec(dst_mm, rss);
 834	pte_unmap_unlock(orig_dst_pte, dst_ptl);
 835	cond_resched();
 836
 837	if (entry.val) {
 838		if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
 
 
 
 
 
 
 
 
 
 
 839			return -ENOMEM;
 840		progress = 0;
 
 841	}
 
 
 
 
 842	if (addr != end)
 843		goto again;
 844	return 0;
 
 
 
 845}
 846
 847static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 848		pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
 849		unsigned long addr, unsigned long end)
 
 850{
 
 
 851	pmd_t *src_pmd, *dst_pmd;
 852	unsigned long next;
 853
 854	dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
 855	if (!dst_pmd)
 856		return -ENOMEM;
 857	src_pmd = pmd_offset(src_pud, addr);
 858	do {
 859		next = pmd_addr_end(addr, end);
 860		if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
 861			|| pmd_devmap(*src_pmd)) {
 862			int err;
 863			VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
 864			err = copy_huge_pmd(dst_mm, src_mm,
 865					    dst_pmd, src_pmd, addr, vma);
 866			if (err == -ENOMEM)
 867				return -ENOMEM;
 868			if (!err)
 869				continue;
 870			/* fall through */
 871		}
 872		if (pmd_none_or_clear_bad(src_pmd))
 873			continue;
 874		if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
 875						vma, addr, next))
 876			return -ENOMEM;
 877	} while (dst_pmd++, src_pmd++, addr = next, addr != end);
 878	return 0;
 879}
 880
 881static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 882		p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
 883		unsigned long addr, unsigned long end)
 
 884{
 
 
 885	pud_t *src_pud, *dst_pud;
 886	unsigned long next;
 887
 888	dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
 889	if (!dst_pud)
 890		return -ENOMEM;
 891	src_pud = pud_offset(src_p4d, addr);
 892	do {
 893		next = pud_addr_end(addr, end);
 894		if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
 895			int err;
 896
 897			VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
 898			err = copy_huge_pud(dst_mm, src_mm,
 899					    dst_pud, src_pud, addr, vma);
 900			if (err == -ENOMEM)
 901				return -ENOMEM;
 902			if (!err)
 903				continue;
 904			/* fall through */
 905		}
 906		if (pud_none_or_clear_bad(src_pud))
 907			continue;
 908		if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
 909						vma, addr, next))
 910			return -ENOMEM;
 911	} while (dst_pud++, src_pud++, addr = next, addr != end);
 912	return 0;
 913}
 914
 915static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 916		pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
 917		unsigned long addr, unsigned long end)
 
 918{
 
 919	p4d_t *src_p4d, *dst_p4d;
 920	unsigned long next;
 921
 922	dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
 923	if (!dst_p4d)
 924		return -ENOMEM;
 925	src_p4d = p4d_offset(src_pgd, addr);
 926	do {
 927		next = p4d_addr_end(addr, end);
 928		if (p4d_none_or_clear_bad(src_p4d))
 929			continue;
 930		if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
 931						vma, addr, next))
 932			return -ENOMEM;
 933	} while (dst_p4d++, src_p4d++, addr = next, addr != end);
 934	return 0;
 935}
 936
 937int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 938		struct vm_area_struct *vma)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 939{
 940	pgd_t *src_pgd, *dst_pgd;
 941	unsigned long next;
 942	unsigned long addr = vma->vm_start;
 943	unsigned long end = vma->vm_end;
 
 
 944	struct mmu_notifier_range range;
 945	bool is_cow;
 946	int ret;
 947
 948	/*
 949	 * Don't copy ptes where a page fault will fill them correctly.
 950	 * Fork becomes much lighter when there are big shared or private
 951	 * readonly mappings. The tradeoff is that copy_page_range is more
 952	 * efficient than faulting.
 953	 */
 954	if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
 955			!vma->anon_vma)
 956		return 0;
 957
 958	if (is_vm_hugetlb_page(vma))
 959		return copy_hugetlb_page_range(dst_mm, src_mm, vma);
 960
 961	if (unlikely(vma->vm_flags & VM_PFNMAP)) {
 962		/*
 963		 * We do not free on error cases below as remove_vma
 964		 * gets called on error from higher level routine
 965		 */
 966		ret = track_pfn_copy(vma);
 967		if (ret)
 968			return ret;
 969	}
 970
 971	/*
 972	 * We need to invalidate the secondary MMU mappings only when
 973	 * there could be a permission downgrade on the ptes of the
 974	 * parent mm. And a permission downgrade will only happen if
 975	 * is_cow_mapping() returns true.
 976	 */
 977	is_cow = is_cow_mapping(vma->vm_flags);
 978
 979	if (is_cow) {
 980		mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
 981					0, vma, src_mm, addr, end);
 982		mmu_notifier_invalidate_range_start(&range);
 
 
 
 
 
 
 
 
 
 983	}
 984
 985	ret = 0;
 986	dst_pgd = pgd_offset(dst_mm, addr);
 987	src_pgd = pgd_offset(src_mm, addr);
 988	do {
 989		next = pgd_addr_end(addr, end);
 990		if (pgd_none_or_clear_bad(src_pgd))
 991			continue;
 992		if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
 993					    vma, addr, next))) {
 
 994			ret = -ENOMEM;
 995			break;
 996		}
 997	} while (dst_pgd++, src_pgd++, addr = next, addr != end);
 998
 999	if (is_cow)
 
1000		mmu_notifier_invalidate_range_end(&range);
 
1001	return ret;
1002}
1003
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1004static unsigned long zap_pte_range(struct mmu_gather *tlb,
1005				struct vm_area_struct *vma, pmd_t *pmd,
1006				unsigned long addr, unsigned long end,
1007				struct zap_details *details)
1008{
1009	struct mm_struct *mm = tlb->mm;
1010	int force_flush = 0;
1011	int rss[NR_MM_COUNTERS];
1012	spinlock_t *ptl;
1013	pte_t *start_pte;
1014	pte_t *pte;
1015	swp_entry_t entry;
1016
1017	tlb_change_page_size(tlb, PAGE_SIZE);
1018again:
1019	init_rss_vec(rss);
1020	start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1021	pte = start_pte;
 
 
1022	flush_tlb_batched_pending(mm);
1023	arch_enter_lazy_mmu_mode();
1024	do {
1025		pte_t ptent = *pte;
 
 
 
1026		if (pte_none(ptent))
1027			continue;
1028
1029		if (need_resched())
1030			break;
1031
1032		if (pte_present(ptent)) {
1033			struct page *page;
1034
1035			page = vm_normal_page(vma, addr, ptent);
1036			if (unlikely(details) && page) {
1037				/*
1038				 * unmap_shared_mapping_pages() wants to
1039				 * invalidate cache without truncating:
1040				 * unmap shared but keep private pages.
1041				 */
1042				if (details->check_mapping &&
1043				    details->check_mapping != page_rmapping(page))
1044					continue;
1045			}
1046			ptent = ptep_get_and_clear_full(mm, addr, pte,
1047							tlb->fullmm);
 
1048			tlb_remove_tlb_entry(tlb, pte, addr);
1049			if (unlikely(!page))
 
 
 
1050				continue;
 
1051
1052			if (!PageAnon(page)) {
 
 
1053				if (pte_dirty(ptent)) {
1054					force_flush = 1;
1055					set_page_dirty(page);
 
 
 
1056				}
1057				if (pte_young(ptent) &&
1058				    likely(!(vma->vm_flags & VM_SEQ_READ)))
1059					mark_page_accessed(page);
1060			}
1061			rss[mm_counter(page)]--;
1062			page_remove_rmap(page, false);
1063			if (unlikely(page_mapcount(page) < 0))
1064				print_bad_pte(vma, addr, ptent, page);
1065			if (unlikely(__tlb_remove_page(tlb, page))) {
 
 
1066				force_flush = 1;
1067				addr += PAGE_SIZE;
1068				break;
1069			}
1070			continue;
1071		}
1072
1073		entry = pte_to_swp_entry(ptent);
1074		if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1075			struct page *page = device_private_entry_to_page(entry);
1076
1077			if (unlikely(details && details->check_mapping)) {
1078				/*
1079				 * unmap_shared_mapping_pages() wants to
1080				 * invalidate cache without truncating:
1081				 * unmap shared but keep private pages.
1082				 */
1083				if (details->check_mapping !=
1084				    page_rmapping(page))
1085					continue;
1086			}
1087
1088			pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1089			rss[mm_counter(page)]--;
1090			page_remove_rmap(page, false);
1091			put_page(page);
1092			continue;
1093		}
1094
1095		/* If details->check_mapping, we leave swap entries. */
1096		if (unlikely(details))
1097			continue;
1098
1099		if (!non_swap_entry(entry))
1100			rss[MM_SWAPENTS]--;
1101		else if (is_migration_entry(entry)) {
1102			struct page *page;
1103
1104			page = migration_entry_to_page(entry);
 
 
1105			rss[mm_counter(page)]--;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1106		}
1107		if (unlikely(!free_swap_and_cache(entry)))
1108			print_bad_pte(vma, addr, ptent, NULL);
1109		pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 
1110	} while (pte++, addr += PAGE_SIZE, addr != end);
1111
1112	add_mm_rss_vec(mm, rss);
1113	arch_leave_lazy_mmu_mode();
1114
1115	/* Do the actual TLB flush before dropping ptl */
1116	if (force_flush)
1117		tlb_flush_mmu_tlbonly(tlb);
 
 
1118	pte_unmap_unlock(start_pte, ptl);
1119
1120	/*
1121	 * If we forced a TLB flush (either due to running out of
1122	 * batch buffers or because we needed to flush dirty TLB
1123	 * entries before releasing the ptl), free the batched
1124	 * memory too. Restart if we didn't do everything.
1125	 */
1126	if (force_flush) {
1127		force_flush = 0;
1128		tlb_flush_mmu(tlb);
1129	}
1130
1131	if (addr != end) {
1132		cond_resched();
1133		goto again;
1134	}
1135
1136	return addr;
1137}
1138
1139static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1140				struct vm_area_struct *vma, pud_t *pud,
1141				unsigned long addr, unsigned long end,
1142				struct zap_details *details)
1143{
1144	pmd_t *pmd;
1145	unsigned long next;
1146
1147	pmd = pmd_offset(pud, addr);
1148	do {
1149		next = pmd_addr_end(addr, end);
1150		if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1151			if (next - addr != HPAGE_PMD_SIZE)
1152				__split_huge_pmd(vma, pmd, addr, false, NULL);
1153			else if (zap_huge_pmd(tlb, vma, pmd, addr))
1154				goto next;
 
 
1155			/* fall through */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1156		}
1157		/*
1158		 * Here there can be other concurrent MADV_DONTNEED or
1159		 * trans huge page faults running, and if the pmd is
1160		 * none or trans huge it can change under us. This is
1161		 * because MADV_DONTNEED holds the mmap_sem in read
1162		 * mode.
1163		 */
1164		if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1165			goto next;
1166		next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1167next:
1168		cond_resched();
1169	} while (pmd++, addr = next, addr != end);
1170
1171	return addr;
1172}
1173
1174static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1175				struct vm_area_struct *vma, p4d_t *p4d,
1176				unsigned long addr, unsigned long end,
1177				struct zap_details *details)
1178{
1179	pud_t *pud;
1180	unsigned long next;
1181
1182	pud = pud_offset(p4d, addr);
1183	do {
1184		next = pud_addr_end(addr, end);
1185		if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1186			if (next - addr != HPAGE_PUD_SIZE) {
1187				VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1188				split_huge_pud(vma, pud, addr);
1189			} else if (zap_huge_pud(tlb, vma, pud, addr))
1190				goto next;
1191			/* fall through */
1192		}
1193		if (pud_none_or_clear_bad(pud))
1194			continue;
1195		next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1196next:
1197		cond_resched();
1198	} while (pud++, addr = next, addr != end);
1199
1200	return addr;
1201}
1202
1203static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1204				struct vm_area_struct *vma, pgd_t *pgd,
1205				unsigned long addr, unsigned long end,
1206				struct zap_details *details)
1207{
1208	p4d_t *p4d;
1209	unsigned long next;
1210
1211	p4d = p4d_offset(pgd, addr);
1212	do {
1213		next = p4d_addr_end(addr, end);
1214		if (p4d_none_or_clear_bad(p4d))
1215			continue;
1216		next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1217	} while (p4d++, addr = next, addr != end);
1218
1219	return addr;
1220}
1221
1222void unmap_page_range(struct mmu_gather *tlb,
1223			     struct vm_area_struct *vma,
1224			     unsigned long addr, unsigned long end,
1225			     struct zap_details *details)
1226{
1227	pgd_t *pgd;
1228	unsigned long next;
1229
1230	BUG_ON(addr >= end);
1231	tlb_start_vma(tlb, vma);
1232	pgd = pgd_offset(vma->vm_mm, addr);
1233	do {
1234		next = pgd_addr_end(addr, end);
1235		if (pgd_none_or_clear_bad(pgd))
1236			continue;
1237		next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1238	} while (pgd++, addr = next, addr != end);
1239	tlb_end_vma(tlb, vma);
1240}
1241
1242
1243static void unmap_single_vma(struct mmu_gather *tlb,
1244		struct vm_area_struct *vma, unsigned long start_addr,
1245		unsigned long end_addr,
1246		struct zap_details *details)
1247{
1248	unsigned long start = max(vma->vm_start, start_addr);
1249	unsigned long end;
1250
1251	if (start >= vma->vm_end)
1252		return;
1253	end = min(vma->vm_end, end_addr);
1254	if (end <= vma->vm_start)
1255		return;
1256
1257	if (vma->vm_file)
1258		uprobe_munmap(vma, start, end);
1259
1260	if (unlikely(vma->vm_flags & VM_PFNMAP))
1261		untrack_pfn(vma, 0, 0);
1262
1263	if (start != end) {
1264		if (unlikely(is_vm_hugetlb_page(vma))) {
1265			/*
1266			 * It is undesirable to test vma->vm_file as it
1267			 * should be non-null for valid hugetlb area.
1268			 * However, vm_file will be NULL in the error
1269			 * cleanup path of mmap_region. When
1270			 * hugetlbfs ->mmap method fails,
1271			 * mmap_region() nullifies vma->vm_file
1272			 * before calling this function to clean up.
1273			 * Since no pte has actually been setup, it is
1274			 * safe to do nothing in this case.
1275			 */
1276			if (vma->vm_file) {
1277				i_mmap_lock_write(vma->vm_file->f_mapping);
1278				__unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1279				i_mmap_unlock_write(vma->vm_file->f_mapping);
 
1280			}
1281		} else
1282			unmap_page_range(tlb, vma, start, end, details);
1283	}
1284}
1285
1286/**
1287 * unmap_vmas - unmap a range of memory covered by a list of vma's
1288 * @tlb: address of the caller's struct mmu_gather
 
1289 * @vma: the starting vma
1290 * @start_addr: virtual address at which to start unmapping
1291 * @end_addr: virtual address at which to end unmapping
 
 
1292 *
1293 * Unmap all pages in the vma list.
1294 *
1295 * Only addresses between `start' and `end' will be unmapped.
1296 *
1297 * The VMA list must be sorted in ascending virtual address order.
1298 *
1299 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1300 * range after unmap_vmas() returns.  So the only responsibility here is to
1301 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1302 * drops the lock and schedules.
1303 */
1304void unmap_vmas(struct mmu_gather *tlb,
1305		struct vm_area_struct *vma, unsigned long start_addr,
1306		unsigned long end_addr)
 
1307{
1308	struct mmu_notifier_range range;
 
 
 
 
 
1309
1310	mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1311				start_addr, end_addr);
1312	mmu_notifier_invalidate_range_start(&range);
1313	for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1314		unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
 
 
 
 
 
 
 
1315	mmu_notifier_invalidate_range_end(&range);
1316}
1317
1318/**
1319 * zap_page_range - remove user pages in a given range
1320 * @vma: vm_area_struct holding the applicable pages
1321 * @start: starting address of pages to zap
1322 * @size: number of bytes to zap
1323 *
1324 * Caller must protect the VMA list
1325 */
1326void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1327		unsigned long size)
1328{
1329	struct mmu_notifier_range range;
1330	struct mmu_gather tlb;
1331
1332	lru_add_drain();
1333	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1334				start, start + size);
1335	tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1336	update_hiwater_rss(vma->vm_mm);
1337	mmu_notifier_invalidate_range_start(&range);
1338	for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1339		unmap_single_vma(&tlb, vma, start, range.end, NULL);
1340	mmu_notifier_invalidate_range_end(&range);
1341	tlb_finish_mmu(&tlb, start, range.end);
1342}
1343
1344/**
1345 * zap_page_range_single - remove user pages in a given range
1346 * @vma: vm_area_struct holding the applicable pages
1347 * @address: starting address of pages to zap
1348 * @size: number of bytes to zap
1349 * @details: details of shared cache invalidation
1350 *
1351 * The range must fit into one VMA.
1352 */
1353static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1354		unsigned long size, struct zap_details *details)
1355{
 
1356	struct mmu_notifier_range range;
1357	struct mmu_gather tlb;
1358
1359	lru_add_drain();
1360	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1361				address, address + size);
1362	tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
 
1363	update_hiwater_rss(vma->vm_mm);
1364	mmu_notifier_invalidate_range_start(&range);
1365	unmap_single_vma(&tlb, vma, address, range.end, details);
 
 
 
 
1366	mmu_notifier_invalidate_range_end(&range);
1367	tlb_finish_mmu(&tlb, address, range.end);
 
1368}
1369
1370/**
1371 * zap_vma_ptes - remove ptes mapping the vma
1372 * @vma: vm_area_struct holding ptes to be zapped
1373 * @address: starting address of pages to zap
1374 * @size: number of bytes to zap
1375 *
1376 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1377 *
1378 * The entire address range must be fully contained within the vma.
1379 *
1380 */
1381void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1382		unsigned long size)
1383{
1384	if (address < vma->vm_start || address + size > vma->vm_end ||
1385	    		!(vma->vm_flags & VM_PFNMAP))
1386		return;
1387
1388	zap_page_range_single(vma, address, size, NULL);
1389}
1390EXPORT_SYMBOL_GPL(zap_vma_ptes);
1391
1392pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1393			spinlock_t **ptl)
1394{
1395	pgd_t *pgd;
1396	p4d_t *p4d;
1397	pud_t *pud;
1398	pmd_t *pmd;
1399
1400	pgd = pgd_offset(mm, addr);
1401	p4d = p4d_alloc(mm, pgd, addr);
1402	if (!p4d)
1403		return NULL;
1404	pud = pud_alloc(mm, p4d, addr);
1405	if (!pud)
1406		return NULL;
1407	pmd = pmd_alloc(mm, pud, addr);
1408	if (!pmd)
1409		return NULL;
1410
1411	VM_BUG_ON(pmd_trans_huge(*pmd));
 
 
 
 
 
 
 
 
 
 
1412	return pte_alloc_map_lock(mm, pmd, addr, ptl);
1413}
1414
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1415/*
1416 * This is the old fallback for page remapping.
1417 *
1418 * For historical reasons, it only allows reserved pages. Only
1419 * old drivers should use this, and they needed to mark their
1420 * pages reserved for the old functions anyway.
1421 */
1422static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1423			struct page *page, pgprot_t prot)
1424{
1425	struct mm_struct *mm = vma->vm_mm;
1426	int retval;
1427	pte_t *pte;
1428	spinlock_t *ptl;
1429
1430	retval = -EINVAL;
1431	if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1432		goto out;
1433	retval = -ENOMEM;
1434	flush_dcache_page(page);
1435	pte = get_locked_pte(mm, addr, &ptl);
1436	if (!pte)
1437		goto out;
1438	retval = -EBUSY;
1439	if (!pte_none(*pte))
1440		goto out_unlock;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1441
1442	/* Ok, finally just insert the thing.. */
1443	get_page(page);
1444	inc_mm_counter_fast(mm, mm_counter_file(page));
1445	page_add_file_rmap(page, false);
1446	set_pte_at(mm, addr, pte, mk_pte(page, prot));
1447
1448	retval = 0;
1449out_unlock:
1450	pte_unmap_unlock(pte, ptl);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1451out:
1452	return retval;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1453}
 
1454
1455/**
1456 * vm_insert_page - insert single page into user vma
1457 * @vma: user vma to map to
1458 * @addr: target user address of this page
1459 * @page: source kernel page
1460 *
1461 * This allows drivers to insert individual pages they've allocated
1462 * into a user vma.
1463 *
1464 * The page has to be a nice clean _individual_ kernel allocation.
1465 * If you allocate a compound page, you need to have marked it as
1466 * such (__GFP_COMP), or manually just split the page up yourself
1467 * (see split_page()).
1468 *
1469 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1470 * took an arbitrary page protection parameter. This doesn't allow
1471 * that. Your vma protection will have to be set up correctly, which
1472 * means that if you want a shared writable mapping, you'd better
1473 * ask for a shared writable mapping!
1474 *
1475 * The page does not need to be reserved.
1476 *
1477 * Usually this function is called from f_op->mmap() handler
1478 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1479 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1480 * function from other places, for example from page-fault handler.
1481 *
1482 * Return: %0 on success, negative error code otherwise.
1483 */
1484int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1485			struct page *page)
1486{
1487	if (addr < vma->vm_start || addr >= vma->vm_end)
1488		return -EFAULT;
1489	if (!page_count(page))
1490		return -EINVAL;
1491	if (!(vma->vm_flags & VM_MIXEDMAP)) {
1492		BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1493		BUG_ON(vma->vm_flags & VM_PFNMAP);
1494		vma->vm_flags |= VM_MIXEDMAP;
1495	}
1496	return insert_page(vma, addr, page, vma->vm_page_prot);
1497}
1498EXPORT_SYMBOL(vm_insert_page);
1499
1500/*
1501 * __vm_map_pages - maps range of kernel pages into user vma
1502 * @vma: user vma to map to
1503 * @pages: pointer to array of source kernel pages
1504 * @num: number of pages in page array
1505 * @offset: user's requested vm_pgoff
1506 *
1507 * This allows drivers to map range of kernel pages into a user vma.
1508 *
1509 * Return: 0 on success and error code otherwise.
1510 */
1511static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1512				unsigned long num, unsigned long offset)
1513{
1514	unsigned long count = vma_pages(vma);
1515	unsigned long uaddr = vma->vm_start;
1516	int ret, i;
1517
1518	/* Fail if the user requested offset is beyond the end of the object */
1519	if (offset >= num)
1520		return -ENXIO;
1521
1522	/* Fail if the user requested size exceeds available object size */
1523	if (count > num - offset)
1524		return -ENXIO;
1525
1526	for (i = 0; i < count; i++) {
1527		ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1528		if (ret < 0)
1529			return ret;
1530		uaddr += PAGE_SIZE;
1531	}
1532
1533	return 0;
1534}
1535
1536/**
1537 * vm_map_pages - maps range of kernel pages starts with non zero offset
1538 * @vma: user vma to map to
1539 * @pages: pointer to array of source kernel pages
1540 * @num: number of pages in page array
1541 *
1542 * Maps an object consisting of @num pages, catering for the user's
1543 * requested vm_pgoff
1544 *
1545 * If we fail to insert any page into the vma, the function will return
1546 * immediately leaving any previously inserted pages present.  Callers
1547 * from the mmap handler may immediately return the error as their caller
1548 * will destroy the vma, removing any successfully inserted pages. Other
1549 * callers should make their own arrangements for calling unmap_region().
1550 *
1551 * Context: Process context. Called by mmap handlers.
1552 * Return: 0 on success and error code otherwise.
1553 */
1554int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1555				unsigned long num)
1556{
1557	return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1558}
1559EXPORT_SYMBOL(vm_map_pages);
1560
1561/**
1562 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1563 * @vma: user vma to map to
1564 * @pages: pointer to array of source kernel pages
1565 * @num: number of pages in page array
1566 *
1567 * Similar to vm_map_pages(), except that it explicitly sets the offset
1568 * to 0. This function is intended for the drivers that did not consider
1569 * vm_pgoff.
1570 *
1571 * Context: Process context. Called by mmap handlers.
1572 * Return: 0 on success and error code otherwise.
1573 */
1574int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1575				unsigned long num)
1576{
1577	return __vm_map_pages(vma, pages, num, 0);
1578}
1579EXPORT_SYMBOL(vm_map_pages_zero);
1580
1581static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1582			pfn_t pfn, pgprot_t prot, bool mkwrite)
1583{
1584	struct mm_struct *mm = vma->vm_mm;
1585	pte_t *pte, entry;
1586	spinlock_t *ptl;
1587
1588	pte = get_locked_pte(mm, addr, &ptl);
1589	if (!pte)
1590		return VM_FAULT_OOM;
1591	if (!pte_none(*pte)) {
 
1592		if (mkwrite) {
1593			/*
1594			 * For read faults on private mappings the PFN passed
1595			 * in may not match the PFN we have mapped if the
1596			 * mapped PFN is a writeable COW page.  In the mkwrite
1597			 * case we are creating a writable PTE for a shared
1598			 * mapping and we expect the PFNs to match. If they
1599			 * don't match, we are likely racing with block
1600			 * allocation and mapping invalidation so just skip the
1601			 * update.
1602			 */
1603			if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1604				WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1605				goto out_unlock;
1606			}
1607			entry = pte_mkyoung(*pte);
1608			entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1609			if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1610				update_mmu_cache(vma, addr, pte);
1611		}
1612		goto out_unlock;
1613	}
1614
1615	/* Ok, finally just insert the thing.. */
1616	if (pfn_t_devmap(pfn))
1617		entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1618	else
1619		entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1620
1621	if (mkwrite) {
1622		entry = pte_mkyoung(entry);
1623		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1624	}
1625
1626	set_pte_at(mm, addr, pte, entry);
1627	update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1628
1629out_unlock:
1630	pte_unmap_unlock(pte, ptl);
1631	return VM_FAULT_NOPAGE;
1632}
1633
1634/**
1635 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1636 * @vma: user vma to map to
1637 * @addr: target user address of this page
1638 * @pfn: source kernel pfn
1639 * @pgprot: pgprot flags for the inserted page
1640 *
1641 * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1642 * to override pgprot on a per-page basis.
1643 *
1644 * This only makes sense for IO mappings, and it makes no sense for
1645 * COW mappings.  In general, using multiple vmas is preferable;
1646 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1647 * impractical.
1648 *
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1649 * Context: Process context.  May allocate using %GFP_KERNEL.
1650 * Return: vm_fault_t value.
1651 */
1652vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1653			unsigned long pfn, pgprot_t pgprot)
1654{
1655	/*
1656	 * Technically, architectures with pte_special can avoid all these
1657	 * restrictions (same for remap_pfn_range).  However we would like
1658	 * consistency in testing and feature parity among all, so we should
1659	 * try to keep these invariants in place for everybody.
1660	 */
1661	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1662	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1663						(VM_PFNMAP|VM_MIXEDMAP));
1664	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1665	BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1666
1667	if (addr < vma->vm_start || addr >= vma->vm_end)
1668		return VM_FAULT_SIGBUS;
1669
1670	if (!pfn_modify_allowed(pfn, pgprot))
1671		return VM_FAULT_SIGBUS;
1672
1673	track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1674
1675	return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1676			false);
1677}
1678EXPORT_SYMBOL(vmf_insert_pfn_prot);
1679
1680/**
1681 * vmf_insert_pfn - insert single pfn into user vma
1682 * @vma: user vma to map to
1683 * @addr: target user address of this page
1684 * @pfn: source kernel pfn
1685 *
1686 * Similar to vm_insert_page, this allows drivers to insert individual pages
1687 * they've allocated into a user vma. Same comments apply.
1688 *
1689 * This function should only be called from a vm_ops->fault handler, and
1690 * in that case the handler should return the result of this function.
1691 *
1692 * vma cannot be a COW mapping.
1693 *
1694 * As this is called only for pages that do not currently exist, we
1695 * do not need to flush old virtual caches or the TLB.
1696 *
1697 * Context: Process context.  May allocate using %GFP_KERNEL.
1698 * Return: vm_fault_t value.
1699 */
1700vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1701			unsigned long pfn)
1702{
1703	return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1704}
1705EXPORT_SYMBOL(vmf_insert_pfn);
1706
1707static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1708{
1709	/* these checks mirror the abort conditions in vm_normal_page */
1710	if (vma->vm_flags & VM_MIXEDMAP)
1711		return true;
1712	if (pfn_t_devmap(pfn))
1713		return true;
1714	if (pfn_t_special(pfn))
1715		return true;
1716	if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1717		return true;
1718	return false;
1719}
1720
1721static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1722		unsigned long addr, pfn_t pfn, bool mkwrite)
1723{
1724	pgprot_t pgprot = vma->vm_page_prot;
1725	int err;
1726
1727	BUG_ON(!vm_mixed_ok(vma, pfn));
1728
1729	if (addr < vma->vm_start || addr >= vma->vm_end)
1730		return VM_FAULT_SIGBUS;
1731
1732	track_pfn_insert(vma, &pgprot, pfn);
1733
1734	if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1735		return VM_FAULT_SIGBUS;
1736
1737	/*
1738	 * If we don't have pte special, then we have to use the pfn_valid()
1739	 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1740	 * refcount the page if pfn_valid is true (hence insert_page rather
1741	 * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1742	 * without pte special, it would there be refcounted as a normal page.
1743	 */
1744	if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1745	    !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1746		struct page *page;
1747
1748		/*
1749		 * At this point we are committed to insert_page()
1750		 * regardless of whether the caller specified flags that
1751		 * result in pfn_t_has_page() == false.
1752		 */
1753		page = pfn_to_page(pfn_t_to_pfn(pfn));
1754		err = insert_page(vma, addr, page, pgprot);
1755	} else {
1756		return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1757	}
1758
1759	if (err == -ENOMEM)
1760		return VM_FAULT_OOM;
1761	if (err < 0 && err != -EBUSY)
1762		return VM_FAULT_SIGBUS;
1763
1764	return VM_FAULT_NOPAGE;
1765}
1766
1767vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1768		pfn_t pfn)
1769{
1770	return __vm_insert_mixed(vma, addr, pfn, false);
1771}
1772EXPORT_SYMBOL(vmf_insert_mixed);
1773
1774/*
1775 *  If the insertion of PTE failed because someone else already added a
1776 *  different entry in the mean time, we treat that as success as we assume
1777 *  the same entry was actually inserted.
1778 */
1779vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1780		unsigned long addr, pfn_t pfn)
1781{
1782	return __vm_insert_mixed(vma, addr, pfn, true);
1783}
1784EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1785
1786/*
1787 * maps a range of physical memory into the requested pages. the old
1788 * mappings are removed. any references to nonexistent pages results
1789 * in null mappings (currently treated as "copy-on-access")
1790 */
1791static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1792			unsigned long addr, unsigned long end,
1793			unsigned long pfn, pgprot_t prot)
1794{
1795	pte_t *pte;
1796	spinlock_t *ptl;
1797	int err = 0;
1798
1799	pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1800	if (!pte)
1801		return -ENOMEM;
1802	arch_enter_lazy_mmu_mode();
1803	do {
1804		BUG_ON(!pte_none(*pte));
1805		if (!pfn_modify_allowed(pfn, prot)) {
1806			err = -EACCES;
1807			break;
1808		}
1809		set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1810		pfn++;
1811	} while (pte++, addr += PAGE_SIZE, addr != end);
1812	arch_leave_lazy_mmu_mode();
1813	pte_unmap_unlock(pte - 1, ptl);
1814	return err;
1815}
1816
1817static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1818			unsigned long addr, unsigned long end,
1819			unsigned long pfn, pgprot_t prot)
1820{
1821	pmd_t *pmd;
1822	unsigned long next;
1823	int err;
1824
1825	pfn -= addr >> PAGE_SHIFT;
1826	pmd = pmd_alloc(mm, pud, addr);
1827	if (!pmd)
1828		return -ENOMEM;
1829	VM_BUG_ON(pmd_trans_huge(*pmd));
1830	do {
1831		next = pmd_addr_end(addr, end);
1832		err = remap_pte_range(mm, pmd, addr, next,
1833				pfn + (addr >> PAGE_SHIFT), prot);
1834		if (err)
1835			return err;
1836	} while (pmd++, addr = next, addr != end);
1837	return 0;
1838}
1839
1840static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1841			unsigned long addr, unsigned long end,
1842			unsigned long pfn, pgprot_t prot)
1843{
1844	pud_t *pud;
1845	unsigned long next;
1846	int err;
1847
1848	pfn -= addr >> PAGE_SHIFT;
1849	pud = pud_alloc(mm, p4d, addr);
1850	if (!pud)
1851		return -ENOMEM;
1852	do {
1853		next = pud_addr_end(addr, end);
1854		err = remap_pmd_range(mm, pud, addr, next,
1855				pfn + (addr >> PAGE_SHIFT), prot);
1856		if (err)
1857			return err;
1858	} while (pud++, addr = next, addr != end);
1859	return 0;
1860}
1861
1862static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1863			unsigned long addr, unsigned long end,
1864			unsigned long pfn, pgprot_t prot)
1865{
1866	p4d_t *p4d;
1867	unsigned long next;
1868	int err;
1869
1870	pfn -= addr >> PAGE_SHIFT;
1871	p4d = p4d_alloc(mm, pgd, addr);
1872	if (!p4d)
1873		return -ENOMEM;
1874	do {
1875		next = p4d_addr_end(addr, end);
1876		err = remap_pud_range(mm, p4d, addr, next,
1877				pfn + (addr >> PAGE_SHIFT), prot);
1878		if (err)
1879			return err;
1880	} while (p4d++, addr = next, addr != end);
1881	return 0;
1882}
1883
1884/**
1885 * remap_pfn_range - remap kernel memory to userspace
1886 * @vma: user vma to map to
1887 * @addr: target user address to start at
1888 * @pfn: physical address of kernel memory
1889 * @size: size of map area
1890 * @prot: page protection flags for this mapping
1891 *
1892 * Note: this is only safe if the mm semaphore is held when called.
1893 *
1894 * Return: %0 on success, negative error code otherwise.
1895 */
1896int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1897		    unsigned long pfn, unsigned long size, pgprot_t prot)
1898{
1899	pgd_t *pgd;
1900	unsigned long next;
1901	unsigned long end = addr + PAGE_ALIGN(size);
1902	struct mm_struct *mm = vma->vm_mm;
1903	unsigned long remap_pfn = pfn;
1904	int err;
1905
 
 
 
1906	/*
1907	 * Physically remapped pages are special. Tell the
1908	 * rest of the world about it:
1909	 *   VM_IO tells people not to look at these pages
1910	 *	(accesses can have side effects).
1911	 *   VM_PFNMAP tells the core MM that the base pages are just
1912	 *	raw PFN mappings, and do not have a "struct page" associated
1913	 *	with them.
1914	 *   VM_DONTEXPAND
1915	 *      Disable vma merging and expanding with mremap().
1916	 *   VM_DONTDUMP
1917	 *      Omit vma from core dump, even when VM_IO turned off.
1918	 *
1919	 * There's a horrible special case to handle copy-on-write
1920	 * behaviour that some programs depend on. We mark the "original"
1921	 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1922	 * See vm_normal_page() for details.
1923	 */
1924	if (is_cow_mapping(vma->vm_flags)) {
1925		if (addr != vma->vm_start || end != vma->vm_end)
1926			return -EINVAL;
1927		vma->vm_pgoff = pfn;
1928	}
1929
1930	err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1931	if (err)
1932		return -EINVAL;
1933
1934	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1935
1936	BUG_ON(addr >= end);
1937	pfn -= addr >> PAGE_SHIFT;
1938	pgd = pgd_offset(mm, addr);
1939	flush_cache_range(vma, addr, end);
1940	do {
1941		next = pgd_addr_end(addr, end);
1942		err = remap_p4d_range(mm, pgd, addr, next,
1943				pfn + (addr >> PAGE_SHIFT), prot);
1944		if (err)
1945			break;
1946	} while (pgd++, addr = next, addr != end);
1947
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1948	if (err)
1949		untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1950
 
 
 
1951	return err;
1952}
1953EXPORT_SYMBOL(remap_pfn_range);
1954
1955/**
1956 * vm_iomap_memory - remap memory to userspace
1957 * @vma: user vma to map to
1958 * @start: start of area
1959 * @len: size of area
1960 *
1961 * This is a simplified io_remap_pfn_range() for common driver use. The
1962 * driver just needs to give us the physical memory range to be mapped,
1963 * we'll figure out the rest from the vma information.
1964 *
1965 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1966 * whatever write-combining details or similar.
1967 *
1968 * Return: %0 on success, negative error code otherwise.
1969 */
1970int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1971{
1972	unsigned long vm_len, pfn, pages;
1973
1974	/* Check that the physical memory area passed in looks valid */
1975	if (start + len < start)
1976		return -EINVAL;
1977	/*
1978	 * You *really* shouldn't map things that aren't page-aligned,
1979	 * but we've historically allowed it because IO memory might
1980	 * just have smaller alignment.
1981	 */
1982	len += start & ~PAGE_MASK;
1983	pfn = start >> PAGE_SHIFT;
1984	pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1985	if (pfn + pages < pfn)
1986		return -EINVAL;
1987
1988	/* We start the mapping 'vm_pgoff' pages into the area */
1989	if (vma->vm_pgoff > pages)
1990		return -EINVAL;
1991	pfn += vma->vm_pgoff;
1992	pages -= vma->vm_pgoff;
1993
1994	/* Can we fit all of the mapping? */
1995	vm_len = vma->vm_end - vma->vm_start;
1996	if (vm_len >> PAGE_SHIFT > pages)
1997		return -EINVAL;
1998
1999	/* Ok, let it rip */
2000	return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2001}
2002EXPORT_SYMBOL(vm_iomap_memory);
2003
2004static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2005				     unsigned long addr, unsigned long end,
2006				     pte_fn_t fn, void *data)
 
2007{
2008	pte_t *pte;
2009	int err;
2010	spinlock_t *uninitialized_var(ptl);
2011
2012	pte = (mm == &init_mm) ?
2013		pte_alloc_kernel(pmd, addr) :
2014		pte_alloc_map_lock(mm, pmd, addr, &ptl);
2015	if (!pte)
2016		return -ENOMEM;
2017
2018	BUG_ON(pmd_huge(*pmd));
 
 
 
 
 
 
2019
2020	arch_enter_lazy_mmu_mode();
2021
2022	do {
2023		err = fn(pte++, addr, data);
2024		if (err)
2025			break;
2026	} while (addr += PAGE_SIZE, addr != end);
 
 
 
 
 
2027
2028	arch_leave_lazy_mmu_mode();
2029
2030	if (mm != &init_mm)
2031		pte_unmap_unlock(pte-1, ptl);
2032	return err;
2033}
2034
2035static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2036				     unsigned long addr, unsigned long end,
2037				     pte_fn_t fn, void *data)
 
2038{
2039	pmd_t *pmd;
2040	unsigned long next;
2041	int err;
2042
2043	BUG_ON(pud_huge(*pud));
2044
2045	pmd = pmd_alloc(mm, pud, addr);
2046	if (!pmd)
2047		return -ENOMEM;
 
 
 
 
2048	do {
2049		next = pmd_addr_end(addr, end);
2050		err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
 
 
 
 
 
 
 
 
 
 
2051		if (err)
2052			break;
2053	} while (pmd++, addr = next, addr != end);
 
2054	return err;
2055}
2056
2057static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2058				     unsigned long addr, unsigned long end,
2059				     pte_fn_t fn, void *data)
 
2060{
2061	pud_t *pud;
2062	unsigned long next;
2063	int err;
2064
2065	pud = pud_alloc(mm, p4d, addr);
2066	if (!pud)
2067		return -ENOMEM;
 
 
 
 
2068	do {
2069		next = pud_addr_end(addr, end);
2070		err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
 
 
 
 
 
 
 
 
 
 
2071		if (err)
2072			break;
2073	} while (pud++, addr = next, addr != end);
 
2074	return err;
2075}
2076
2077static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2078				     unsigned long addr, unsigned long end,
2079				     pte_fn_t fn, void *data)
 
2080{
2081	p4d_t *p4d;
2082	unsigned long next;
2083	int err;
2084
2085	p4d = p4d_alloc(mm, pgd, addr);
2086	if (!p4d)
2087		return -ENOMEM;
 
 
 
 
2088	do {
2089		next = p4d_addr_end(addr, end);
2090		err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
 
 
 
 
 
 
 
 
 
 
2091		if (err)
2092			break;
2093	} while (p4d++, addr = next, addr != end);
 
2094	return err;
2095}
2096
2097/*
2098 * Scan a region of virtual memory, filling in page tables as necessary
2099 * and calling a provided function on each leaf page table.
2100 */
2101int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2102			unsigned long size, pte_fn_t fn, void *data)
2103{
2104	pgd_t *pgd;
2105	unsigned long next;
2106	unsigned long end = addr + size;
2107	int err;
 
2108
2109	if (WARN_ON(addr >= end))
2110		return -EINVAL;
2111
2112	pgd = pgd_offset(mm, addr);
2113	do {
2114		next = pgd_addr_end(addr, end);
2115		err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
 
 
 
 
 
 
 
 
 
 
2116		if (err)
2117			break;
2118	} while (pgd++, addr = next, addr != end);
2119
 
 
 
2120	return err;
2121}
 
 
 
 
 
 
 
 
 
 
2122EXPORT_SYMBOL_GPL(apply_to_page_range);
2123
2124/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2125 * handle_pte_fault chooses page fault handler according to an entry which was
2126 * read non-atomically.  Before making any commitment, on those architectures
2127 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2128 * parts, do_swap_page must check under lock before unmapping the pte and
2129 * proceeding (but do_wp_page is only called after already making such a check;
2130 * and do_anonymous_page can safely check later on).
2131 */
2132static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2133				pte_t *page_table, pte_t orig_pte)
2134{
2135	int same = 1;
2136#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2137	if (sizeof(pte_t) > sizeof(unsigned long)) {
2138		spinlock_t *ptl = pte_lockptr(mm, pmd);
2139		spin_lock(ptl);
2140		same = pte_same(*page_table, orig_pte);
2141		spin_unlock(ptl);
2142	}
2143#endif
2144	pte_unmap(page_table);
 
2145	return same;
2146}
2147
2148static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
 
 
 
 
 
 
 
2149{
2150	debug_dma_assert_idle(src);
 
 
 
 
 
 
 
 
 
 
 
 
 
2151
2152	/*
2153	 * If the source page was a PFN mapping, we don't have
2154	 * a "struct page" for it. We do a best-effort copy by
2155	 * just copying from the original user address. If that
2156	 * fails, we just zero-fill it. Live with it.
2157	 */
2158	if (unlikely(!src)) {
2159		void *kaddr = kmap_atomic(dst);
2160		void __user *uaddr = (void __user *)(va & PAGE_MASK);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2161
2162		/*
2163		 * This really shouldn't fail, because the page is there
2164		 * in the page tables. But it might just be unreadable,
2165		 * in which case we just give up and fill the result with
2166		 * zeroes.
2167		 */
2168		if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
 
 
 
 
 
 
2169			clear_page(kaddr);
2170		kunmap_atomic(kaddr);
2171		flush_dcache_page(dst);
2172	} else
2173		copy_user_highpage(dst, src, va, vma);
 
 
 
 
 
 
 
 
 
2174}
2175
2176static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2177{
2178	struct file *vm_file = vma->vm_file;
2179
2180	if (vm_file)
2181		return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2182
2183	/*
2184	 * Special mappings (e.g. VDSO) do not have any file so fake
2185	 * a default GFP_KERNEL for them.
2186	 */
2187	return GFP_KERNEL;
2188}
2189
2190/*
2191 * Notify the address space that the page is about to become writable so that
2192 * it can prohibit this or wait for the page to get into an appropriate state.
2193 *
2194 * We do this without the lock held, so that it can sleep if it needs to.
2195 */
2196static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2197{
2198	vm_fault_t ret;
2199	struct page *page = vmf->page;
2200	unsigned int old_flags = vmf->flags;
2201
2202	vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2203
2204	if (vmf->vma->vm_file &&
2205	    IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2206		return VM_FAULT_SIGBUS;
2207
2208	ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2209	/* Restore original flags so that caller is not surprised */
2210	vmf->flags = old_flags;
2211	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2212		return ret;
2213	if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2214		lock_page(page);
2215		if (!page->mapping) {
2216			unlock_page(page);
2217			return 0; /* retry */
2218		}
2219		ret |= VM_FAULT_LOCKED;
2220	} else
2221		VM_BUG_ON_PAGE(!PageLocked(page), page);
2222	return ret;
2223}
2224
2225/*
2226 * Handle dirtying of a page in shared file mapping on a write fault.
2227 *
2228 * The function expects the page to be locked and unlocks it.
2229 */
2230static void fault_dirty_shared_page(struct vm_area_struct *vma,
2231				    struct page *page)
2232{
 
2233	struct address_space *mapping;
 
2234	bool dirtied;
2235	bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2236
2237	dirtied = set_page_dirty(page);
2238	VM_BUG_ON_PAGE(PageAnon(page), page);
2239	/*
2240	 * Take a local copy of the address_space - page.mapping may be zeroed
2241	 * by truncate after unlock_page().   The address_space itself remains
2242	 * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2243	 * release semantics to prevent the compiler from undoing this copying.
2244	 */
2245	mapping = page_rmapping(page);
2246	unlock_page(page);
 
 
 
2247
 
 
 
 
 
 
 
 
 
2248	if ((dirtied || page_mkwrite) && mapping) {
2249		/*
2250		 * Some device drivers do not set page.mapping
2251		 * but still dirty their pages
2252		 */
2253		balance_dirty_pages_ratelimited(mapping);
 
 
 
 
2254	}
2255
2256	if (!page_mkwrite)
2257		file_update_time(vma->vm_file);
2258}
2259
2260/*
2261 * Handle write page faults for pages that can be reused in the current vma
2262 *
2263 * This can happen either due to the mapping being with the VM_SHARED flag,
2264 * or due to us being the last reference standing to the page. In either
2265 * case, all we need to do here is to mark the page as writable and update
2266 * any related book-keeping.
2267 */
2268static inline void wp_page_reuse(struct vm_fault *vmf)
2269	__releases(vmf->ptl)
2270{
2271	struct vm_area_struct *vma = vmf->vma;
2272	struct page *page = vmf->page;
2273	pte_t entry;
2274	/*
2275	 * Clear the pages cpupid information as the existing
2276	 * information potentially belongs to a now completely
2277	 * unrelated process.
2278	 */
2279	if (page)
2280		page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
 
 
 
 
 
 
2281
2282	flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2283	entry = pte_mkyoung(vmf->orig_pte);
2284	entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2285	if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2286		update_mmu_cache(vma, vmf->address, vmf->pte);
2287	pte_unmap_unlock(vmf->pte, vmf->ptl);
 
2288}
2289
2290/*
2291 * Handle the case of a page which we actually need to copy to a new page.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2292 *
2293 * Called with mmap_sem locked and the old page referenced, but
2294 * without the ptl held.
2295 *
2296 * High level logic flow:
2297 *
2298 * - Allocate a page, copy the content of the old page to the new one.
2299 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2300 * - Take the PTL. If the pte changed, bail out and release the allocated page
2301 * - If the pte is still the way we remember it, update the page table and all
2302 *   relevant references. This includes dropping the reference the page-table
2303 *   held to the old page, as well as updating the rmap.
2304 * - In any case, unlock the PTL and drop the reference we took to the old page.
2305 */
2306static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2307{
 
2308	struct vm_area_struct *vma = vmf->vma;
2309	struct mm_struct *mm = vma->vm_mm;
2310	struct page *old_page = vmf->page;
2311	struct page *new_page = NULL;
2312	pte_t entry;
2313	int page_copied = 0;
2314	struct mem_cgroup *memcg;
2315	struct mmu_notifier_range range;
 
 
 
 
2316
2317	if (unlikely(anon_vma_prepare(vma)))
 
 
 
 
 
 
 
 
2318		goto oom;
2319
2320	if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2321		new_page = alloc_zeroed_user_highpage_movable(vma,
2322							      vmf->address);
2323		if (!new_page)
2324			goto oom;
2325	} else {
2326		new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2327				vmf->address);
2328		if (!new_page)
2329			goto oom;
2330		cow_user_page(new_page, old_page, vmf->address, vma);
 
 
 
 
 
 
 
 
 
2331	}
2332
2333	if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2334		goto oom_free_new;
2335
2336	__SetPageUptodate(new_page);
2337
2338	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2339				vmf->address & PAGE_MASK,
2340				(vmf->address & PAGE_MASK) + PAGE_SIZE);
2341	mmu_notifier_invalidate_range_start(&range);
2342
2343	/*
2344	 * Re-check the pte - we dropped the lock
2345	 */
2346	vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2347	if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2348		if (old_page) {
2349			if (!PageAnon(old_page)) {
2350				dec_mm_counter_fast(mm,
2351						mm_counter_file(old_page));
2352				inc_mm_counter_fast(mm, MM_ANONPAGES);
2353			}
2354		} else {
2355			inc_mm_counter_fast(mm, MM_ANONPAGES);
 
2356		}
2357		flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2358		entry = mk_pte(new_page, vma->vm_page_prot);
2359		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 
 
 
 
 
 
 
 
 
2360		/*
2361		 * Clear the pte entry and flush it first, before updating the
2362		 * pte with the new entry. This will avoid a race condition
2363		 * seen in the presence of one thread doing SMC and another
2364		 * thread doing COW.
 
2365		 */
2366		ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2367		page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2368		mem_cgroup_commit_charge(new_page, memcg, false, false);
2369		lru_cache_add_active_or_unevictable(new_page, vma);
2370		/*
2371		 * We call the notify macro here because, when using secondary
2372		 * mmu page tables (such as kvm shadow page tables), we want the
2373		 * new page to be mapped directly into the secondary page table.
2374		 */
 
2375		set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2376		update_mmu_cache(vma, vmf->address, vmf->pte);
2377		if (old_page) {
2378			/*
2379			 * Only after switching the pte to the new page may
2380			 * we remove the mapcount here. Otherwise another
2381			 * process may come and find the rmap count decremented
2382			 * before the pte is switched to the new page, and
2383			 * "reuse" the old page writing into it while our pte
2384			 * here still points into it and can be read by other
2385			 * threads.
2386			 *
2387			 * The critical issue is to order this
2388			 * page_remove_rmap with the ptp_clear_flush above.
2389			 * Those stores are ordered by (if nothing else,)
2390			 * the barrier present in the atomic_add_negative
2391			 * in page_remove_rmap.
2392			 *
2393			 * Then the TLB flush in ptep_clear_flush ensures that
2394			 * no process can access the old page before the
2395			 * decremented mapcount is visible. And the old page
2396			 * cannot be reused until after the decremented
2397			 * mapcount is visible. So transitively, TLBs to
2398			 * old page will be flushed before it can be reused.
2399			 */
2400			page_remove_rmap(old_page, false);
2401		}
2402
2403		/* Free the old page.. */
2404		new_page = old_page;
2405		page_copied = 1;
2406	} else {
2407		mem_cgroup_cancel_charge(new_page, memcg, false);
 
 
2408	}
2409
2410	if (new_page)
2411		put_page(new_page);
 
 
 
 
 
 
 
2412
2413	pte_unmap_unlock(vmf->pte, vmf->ptl);
2414	/*
2415	 * No need to double call mmu_notifier->invalidate_range() callback as
2416	 * the above ptep_clear_flush_notify() did already call it.
2417	 */
2418	mmu_notifier_invalidate_range_only_end(&range);
2419	if (old_page) {
2420		/*
2421		 * Don't let another task, with possibly unlocked vma,
2422		 * keep the mlocked page.
2423		 */
2424		if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2425			lock_page(old_page);	/* LRU manipulation */
2426			if (PageMlocked(old_page))
2427				munlock_vma_page(old_page);
2428			unlock_page(old_page);
2429		}
2430		put_page(old_page);
2431	}
2432	return page_copied ? VM_FAULT_WRITE : 0;
2433oom_free_new:
2434	put_page(new_page);
2435oom:
2436	if (old_page)
2437		put_page(old_page);
2438	return VM_FAULT_OOM;
 
 
 
 
2439}
2440
2441/**
2442 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2443 *			  writeable once the page is prepared
2444 *
2445 * @vmf: structure describing the fault
 
2446 *
2447 * This function handles all that is needed to finish a write page fault in a
2448 * shared mapping due to PTE being read-only once the mapped page is prepared.
2449 * It handles locking of PTE and modifying it.
2450 *
2451 * The function expects the page to be locked or other protection against
2452 * concurrent faults / writeback (such as DAX radix tree locks).
2453 *
2454 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2455 * we acquired PTE lock.
2456 */
2457vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2458{
2459	WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2460	vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2461				       &vmf->ptl);
 
 
2462	/*
2463	 * We might have raced with another page fault while we released the
2464	 * pte_offset_map_lock.
2465	 */
2466	if (!pte_same(*vmf->pte, vmf->orig_pte)) {
 
2467		pte_unmap_unlock(vmf->pte, vmf->ptl);
2468		return VM_FAULT_NOPAGE;
2469	}
2470	wp_page_reuse(vmf);
2471	return 0;
2472}
2473
2474/*
2475 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2476 * mapping
2477 */
2478static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2479{
2480	struct vm_area_struct *vma = vmf->vma;
2481
2482	if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2483		vm_fault_t ret;
2484
2485		pte_unmap_unlock(vmf->pte, vmf->ptl);
 
 
 
 
2486		vmf->flags |= FAULT_FLAG_MKWRITE;
2487		ret = vma->vm_ops->pfn_mkwrite(vmf);
2488		if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2489			return ret;
2490		return finish_mkwrite_fault(vmf);
2491	}
2492	wp_page_reuse(vmf);
2493	return VM_FAULT_WRITE;
2494}
2495
2496static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2497	__releases(vmf->ptl)
2498{
2499	struct vm_area_struct *vma = vmf->vma;
 
2500
2501	get_page(vmf->page);
2502
2503	if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2504		vm_fault_t tmp;
2505
2506		pte_unmap_unlock(vmf->pte, vmf->ptl);
2507		tmp = do_page_mkwrite(vmf);
 
 
 
 
 
 
2508		if (unlikely(!tmp || (tmp &
2509				      (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2510			put_page(vmf->page);
2511			return tmp;
2512		}
2513		tmp = finish_mkwrite_fault(vmf);
2514		if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2515			unlock_page(vmf->page);
2516			put_page(vmf->page);
2517			return tmp;
2518		}
2519	} else {
2520		wp_page_reuse(vmf);
2521		lock_page(vmf->page);
2522	}
2523	fault_dirty_shared_page(vma, vmf->page);
2524	put_page(vmf->page);
 
 
 
2525
2526	return VM_FAULT_WRITE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2527}
2528
2529/*
2530 * This routine handles present pages, when users try to write
2531 * to a shared page. It is done by copying the page to a new address
2532 * and decrementing the shared-page counter for the old page.
 
 
 
 
2533 *
2534 * Note that this routine assumes that the protection checks have been
2535 * done by the caller (the low-level page fault routine in most cases).
2536 * Thus we can safely just mark it writable once we've done any necessary
2537 * COW.
2538 *
2539 * We also mark the page dirty at this point even though the page will
2540 * change only once the write actually happens. This avoids a few races,
2541 * and potentially makes it more efficient.
2542 *
2543 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2544 * but allow concurrent faults), with pte both mapped and locked.
2545 * We return with mmap_sem still held, but pte unmapped and unlocked.
2546 */
2547static vm_fault_t do_wp_page(struct vm_fault *vmf)
2548	__releases(vmf->ptl)
2549{
 
2550	struct vm_area_struct *vma = vmf->vma;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2551
2552	vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2553	if (!vmf->page) {
 
 
 
 
 
 
 
 
2554		/*
2555		 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2556		 * VM_PFNMAP VMA.
2557		 *
2558		 * We should not cow pages in a shared writeable mapping.
2559		 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2560		 */
2561		if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2562				     (VM_WRITE|VM_SHARED))
2563			return wp_pfn_shared(vmf);
2564
2565		pte_unmap_unlock(vmf->pte, vmf->ptl);
2566		return wp_page_copy(vmf);
2567	}
2568
2569	/*
2570	 * Take out anonymous pages first, anonymous shared vmas are
2571	 * not dirty accountable.
 
 
 
2572	 */
2573	if (PageAnon(vmf->page)) {
2574		int total_map_swapcount;
2575		if (PageKsm(vmf->page) && (PageSwapCache(vmf->page) ||
2576					   page_count(vmf->page) != 1))
2577			goto copy;
2578		if (!trylock_page(vmf->page)) {
2579			get_page(vmf->page);
2580			pte_unmap_unlock(vmf->pte, vmf->ptl);
2581			lock_page(vmf->page);
2582			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2583					vmf->address, &vmf->ptl);
2584			if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2585				unlock_page(vmf->page);
2586				pte_unmap_unlock(vmf->pte, vmf->ptl);
2587				put_page(vmf->page);
2588				return 0;
2589			}
2590			put_page(vmf->page);
2591		}
2592		if (PageKsm(vmf->page)) {
2593			bool reused = reuse_ksm_page(vmf->page, vmf->vma,
2594						     vmf->address);
2595			unlock_page(vmf->page);
2596			if (!reused)
2597				goto copy;
2598			wp_page_reuse(vmf);
2599			return VM_FAULT_WRITE;
2600		}
2601		if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2602			if (total_map_swapcount == 1) {
2603				/*
2604				 * The page is all ours. Move it to
2605				 * our anon_vma so the rmap code will
2606				 * not search our parent or siblings.
2607				 * Protected against the rmap code by
2608				 * the page lock.
2609				 */
2610				page_move_anon_rmap(vmf->page, vma);
2611			}
2612			unlock_page(vmf->page);
2613			wp_page_reuse(vmf);
2614			return VM_FAULT_WRITE;
2615		}
2616		unlock_page(vmf->page);
2617	} else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2618					(VM_WRITE|VM_SHARED))) {
2619		return wp_page_shared(vmf);
2620	}
2621copy:
2622	/*
2623	 * Ok, we need to copy. Oh, well..
2624	 */
2625	get_page(vmf->page);
 
2626
2627	pte_unmap_unlock(vmf->pte, vmf->ptl);
 
 
 
 
2628	return wp_page_copy(vmf);
2629}
2630
2631static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2632		unsigned long start_addr, unsigned long end_addr,
2633		struct zap_details *details)
2634{
2635	zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2636}
2637
2638static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
 
 
2639					    struct zap_details *details)
2640{
2641	struct vm_area_struct *vma;
2642	pgoff_t vba, vea, zba, zea;
2643
2644	vma_interval_tree_foreach(vma, root,
2645			details->first_index, details->last_index) {
2646
2647		vba = vma->vm_pgoff;
2648		vea = vba + vma_pages(vma) - 1;
2649		zba = details->first_index;
2650		if (zba < vba)
2651			zba = vba;
2652		zea = details->last_index;
2653		if (zea > vea)
2654			zea = vea;
2655
2656		unmap_mapping_range_vma(vma,
2657			((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2658			((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2659				details);
2660	}
2661}
2662
2663/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2664 * unmap_mapping_pages() - Unmap pages from processes.
2665 * @mapping: The address space containing pages to be unmapped.
2666 * @start: Index of first page to be unmapped.
2667 * @nr: Number of pages to be unmapped.  0 to unmap to end of file.
2668 * @even_cows: Whether to unmap even private COWed pages.
2669 *
2670 * Unmap the pages in this address space from any userspace process which
2671 * has them mmaped.  Generally, you want to remove COWed pages as well when
2672 * a file is being truncated, but not when invalidating pages from the page
2673 * cache.
2674 */
2675void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2676		pgoff_t nr, bool even_cows)
2677{
2678	struct zap_details details = { };
 
 
2679
2680	details.check_mapping = even_cows ? NULL : mapping;
2681	details.first_index = start;
2682	details.last_index = start + nr - 1;
2683	if (details.last_index < details.first_index)
2684		details.last_index = ULONG_MAX;
2685
2686	i_mmap_lock_write(mapping);
2687	if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2688		unmap_mapping_range_tree(&mapping->i_mmap, &details);
2689	i_mmap_unlock_write(mapping);
 
2690}
 
2691
2692/**
2693 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2694 * address_space corresponding to the specified byte range in the underlying
2695 * file.
2696 *
2697 * @mapping: the address space containing mmaps to be unmapped.
2698 * @holebegin: byte in first page to unmap, relative to the start of
2699 * the underlying file.  This will be rounded down to a PAGE_SIZE
2700 * boundary.  Note that this is different from truncate_pagecache(), which
2701 * must keep the partial page.  In contrast, we must get rid of
2702 * partial pages.
2703 * @holelen: size of prospective hole in bytes.  This will be rounded
2704 * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2705 * end of the file.
2706 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2707 * but 0 when invalidating pagecache, don't throw away private data.
2708 */
2709void unmap_mapping_range(struct address_space *mapping,
2710		loff_t const holebegin, loff_t const holelen, int even_cows)
2711{
2712	pgoff_t hba = holebegin >> PAGE_SHIFT;
2713	pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2714
2715	/* Check for overflow. */
2716	if (sizeof(holelen) > sizeof(hlen)) {
2717		long long holeend =
2718			(holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2719		if (holeend & ~(long long)ULONG_MAX)
2720			hlen = ULONG_MAX - hba + 1;
2721	}
2722
2723	unmap_mapping_pages(mapping, hba, hlen, even_cows);
2724}
2725EXPORT_SYMBOL(unmap_mapping_range);
2726
2727/*
2728 * We enter with non-exclusive mmap_sem (to exclude vma changes,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2729 * but allow concurrent faults), and pte mapped but not yet locked.
2730 * We return with pte unmapped and unlocked.
2731 *
2732 * We return with the mmap_sem locked or unlocked in the same cases
2733 * as does filemap_fault().
2734 */
2735vm_fault_t do_swap_page(struct vm_fault *vmf)
2736{
2737	struct vm_area_struct *vma = vmf->vma;
2738	struct page *page = NULL, *swapcache;
2739	struct mem_cgroup *memcg;
 
 
 
 
2740	swp_entry_t entry;
2741	pte_t pte;
2742	int locked;
2743	int exclusive = 0;
2744	vm_fault_t ret = 0;
 
2745
2746	if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2747		goto out;
2748
2749	entry = pte_to_swp_entry(vmf->orig_pte);
2750	if (unlikely(non_swap_entry(entry))) {
2751		if (is_migration_entry(entry)) {
2752			migration_entry_wait(vma->vm_mm, vmf->pmd,
2753					     vmf->address);
 
 
 
2754		} else if (is_device_private_entry(entry)) {
2755			vmf->page = device_private_entry_to_page(entry);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2756			ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
 
2757		} else if (is_hwpoison_entry(entry)) {
2758			ret = VM_FAULT_HWPOISON;
 
 
2759		} else {
2760			print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2761			ret = VM_FAULT_SIGBUS;
2762		}
2763		goto out;
2764	}
2765
 
 
 
 
2766
2767	delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2768	page = lookup_swap_cache(entry, vma, vmf->address);
2769	swapcache = page;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2770
2771	if (!page) {
2772		struct swap_info_struct *si = swp_swap_info(entry);
 
 
 
 
 
 
 
 
 
 
 
 
 
2773
2774		if (si->flags & SWP_SYNCHRONOUS_IO &&
2775				__swap_count(entry) == 1) {
2776			/* skip swapcache */
2777			page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2778							vmf->address);
2779			if (page) {
2780				__SetPageLocked(page);
2781				__SetPageSwapBacked(page);
2782				set_page_private(page, entry.val);
2783				lru_cache_add_anon(page);
2784				swap_readpage(page, true);
2785			}
2786		} else {
2787			page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2788						vmf);
2789			swapcache = page;
 
 
2790		}
2791
2792		if (!page) {
2793			/*
2794			 * Back out if somebody else faulted in this pte
2795			 * while we released the pte lock.
2796			 */
2797			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2798					vmf->address, &vmf->ptl);
2799			if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
 
2800				ret = VM_FAULT_OOM;
2801			delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2802			goto unlock;
2803		}
2804
2805		/* Had to read the page from swap area: Major fault */
2806		ret = VM_FAULT_MAJOR;
2807		count_vm_event(PGMAJFAULT);
2808		count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2809	} else if (PageHWPoison(page)) {
2810		/*
2811		 * hwpoisoned dirty swapcache pages are kept for killing
2812		 * owner processes (which may be unknown at hwpoison time)
2813		 */
2814		ret = VM_FAULT_HWPOISON;
2815		delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2816		goto out_release;
2817	}
2818
2819	locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
 
 
2820
2821	delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2822	if (!locked) {
2823		ret |= VM_FAULT_RETRY;
2824		goto out_release;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2825	}
2826
2827	/*
2828	 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2829	 * release the swapcache from under us.  The page pin, and pte_same
2830	 * test below, are not enough to exclude that.  Even if it is still
2831	 * swapcache, we need to check that the page's swap has not changed.
2832	 */
2833	if (unlikely((!PageSwapCache(page) ||
2834			page_private(page) != entry.val)) && swapcache)
2835		goto out_page;
2836
2837	page = ksm_might_need_to_copy(page, vma, vmf->address);
2838	if (unlikely(!page)) {
2839		ret = VM_FAULT_OOM;
2840		page = swapcache;
2841		goto out_page;
2842	}
2843
2844	if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
2845					&memcg, false)) {
2846		ret = VM_FAULT_OOM;
2847		goto out_page;
2848	}
2849
2850	/*
2851	 * Back out if somebody else already faulted in this pte.
2852	 */
2853	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2854			&vmf->ptl);
2855	if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2856		goto out_nomap;
2857
2858	if (unlikely(!PageUptodate(page))) {
2859		ret = VM_FAULT_SIGBUS;
2860		goto out_nomap;
2861	}
2862
2863	/*
2864	 * The page isn't present yet, go ahead with the fault.
2865	 *
2866	 * Be careful about the sequence of operations here.
2867	 * To get its accounting right, reuse_swap_page() must be called
2868	 * while the page is counted on swap but not yet in mapcount i.e.
2869	 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2870	 * must be called after the swap_free(), or it will never succeed.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2871	 */
 
 
 
2872
2873	inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2874	dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2875	pte = mk_pte(page, vma->vm_page_prot);
2876	if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2877		pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2878		vmf->flags &= ~FAULT_FLAG_WRITE;
2879		ret |= VM_FAULT_WRITE;
2880		exclusive = RMAP_EXCLUSIVE;
 
 
 
 
 
 
 
 
 
2881	}
2882	flush_icache_page(vma, page);
2883	if (pte_swp_soft_dirty(vmf->orig_pte))
2884		pte = pte_mksoft_dirty(pte);
2885	set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2886	arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
2887	vmf->orig_pte = pte;
2888
2889	/* ksm created a completely new copy */
2890	if (unlikely(page != swapcache && swapcache)) {
2891		page_add_new_anon_rmap(page, vma, vmf->address, false);
2892		mem_cgroup_commit_charge(page, memcg, false, false);
2893		lru_cache_add_active_or_unevictable(page, vma);
2894	} else {
2895		do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2896		mem_cgroup_commit_charge(page, memcg, true, false);
2897		activate_page(page);
2898	}
2899
2900	swap_free(entry);
2901	if (mem_cgroup_swap_full(page) ||
2902	    (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2903		try_to_free_swap(page);
2904	unlock_page(page);
2905	if (page != swapcache && swapcache) {
 
2906		/*
2907		 * Hold the lock to avoid the swap entry to be reused
2908		 * until we take the PT lock for the pte_same() check
2909		 * (to avoid false positives from pte_same). For
2910		 * further safety release the lock after the swap_free
2911		 * so that the swap count won't change under a
2912		 * parallel locked swapcache.
2913		 */
2914		unlock_page(swapcache);
2915		put_page(swapcache);
2916	}
2917
2918	if (vmf->flags & FAULT_FLAG_WRITE) {
2919		ret |= do_wp_page(vmf);
2920		if (ret & VM_FAULT_ERROR)
2921			ret &= VM_FAULT_ERROR;
2922		goto out;
2923	}
2924
2925	/* No need to invalidate - it was non-present before */
2926	update_mmu_cache(vma, vmf->address, vmf->pte);
2927unlock:
2928	pte_unmap_unlock(vmf->pte, vmf->ptl);
 
2929out:
 
 
 
 
 
2930	return ret;
2931out_nomap:
2932	mem_cgroup_cancel_charge(page, memcg, false);
2933	pte_unmap_unlock(vmf->pte, vmf->ptl);
2934out_page:
2935	unlock_page(page);
2936out_release:
2937	put_page(page);
2938	if (page != swapcache && swapcache) {
2939		unlock_page(swapcache);
2940		put_page(swapcache);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2941	}
2942	return ret;
 
 
 
2943}
2944
2945/*
2946 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2947 * but allow concurrent faults), and pte mapped but not yet locked.
2948 * We return with mmap_sem still held, but pte unmapped and unlocked.
2949 */
2950static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
2951{
 
2952	struct vm_area_struct *vma = vmf->vma;
2953	struct mem_cgroup *memcg;
2954	struct page *page;
2955	vm_fault_t ret = 0;
 
2956	pte_t entry;
 
2957
2958	/* File mapping without ->vm_ops ? */
2959	if (vma->vm_flags & VM_SHARED)
2960		return VM_FAULT_SIGBUS;
2961
2962	/*
2963	 * Use pte_alloc() instead of pte_alloc_map().  We can't run
2964	 * pte_offset_map() on pmds where a huge pmd might be created
2965	 * from a different thread.
2966	 *
2967	 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2968	 * parallel threads are excluded by other means.
2969	 *
2970	 * Here we only have down_read(mmap_sem).
2971	 */
2972	if (pte_alloc(vma->vm_mm, vmf->pmd))
2973		return VM_FAULT_OOM;
2974
2975	/* See the comment in pte_alloc_one_map() */
2976	if (unlikely(pmd_trans_unstable(vmf->pmd)))
2977		return 0;
2978
2979	/* Use the zero-page for reads */
2980	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2981			!mm_forbids_zeropage(vma->vm_mm)) {
2982		entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2983						vma->vm_page_prot));
2984		vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2985				vmf->address, &vmf->ptl);
2986		if (!pte_none(*vmf->pte))
2987			goto unlock;
 
 
 
 
2988		ret = check_stable_address_space(vma->vm_mm);
2989		if (ret)
2990			goto unlock;
2991		/* Deliver the page fault to userland, check inside PT lock */
2992		if (userfaultfd_missing(vma)) {
2993			pte_unmap_unlock(vmf->pte, vmf->ptl);
2994			return handle_userfault(vmf, VM_UFFD_MISSING);
2995		}
2996		goto setpte;
2997	}
2998
2999	/* Allocate our own private page. */
3000	if (unlikely(anon_vma_prepare(vma)))
3001		goto oom;
3002	page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3003	if (!page)
 
 
 
3004		goto oom;
3005
3006	if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
3007					false))
 
 
3008		goto oom_free_page;
 
3009
3010	/*
3011	 * The memory barrier inside __SetPageUptodate makes sure that
3012	 * preceeding stores to the page contents become visible before
3013	 * the set_pte_at() write.
3014	 */
3015	__SetPageUptodate(page);
3016
3017	entry = mk_pte(page, vma->vm_page_prot);
 
3018	if (vma->vm_flags & VM_WRITE)
3019		entry = pte_mkwrite(pte_mkdirty(entry));
3020
3021	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3022			&vmf->ptl);
3023	if (!pte_none(*vmf->pte))
 
 
 
 
 
 
3024		goto release;
 
3025
3026	ret = check_stable_address_space(vma->vm_mm);
3027	if (ret)
3028		goto release;
3029
3030	/* Deliver the page fault to userland, check inside PT lock */
3031	if (userfaultfd_missing(vma)) {
3032		pte_unmap_unlock(vmf->pte, vmf->ptl);
3033		mem_cgroup_cancel_charge(page, memcg, false);
3034		put_page(page);
3035		return handle_userfault(vmf, VM_UFFD_MISSING);
3036	}
3037
3038	inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3039	page_add_new_anon_rmap(page, vma, vmf->address, false);
3040	mem_cgroup_commit_charge(page, memcg, false, false);
3041	lru_cache_add_active_or_unevictable(page, vma);
3042setpte:
3043	set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
 
 
3044
3045	/* No need to invalidate - it was non-present before */
3046	update_mmu_cache(vma, vmf->address, vmf->pte);
3047unlock:
3048	pte_unmap_unlock(vmf->pte, vmf->ptl);
 
3049	return ret;
3050release:
3051	mem_cgroup_cancel_charge(page, memcg, false);
3052	put_page(page);
3053	goto unlock;
3054oom_free_page:
3055	put_page(page);
3056oom:
3057	return VM_FAULT_OOM;
3058}
3059
3060/*
3061 * The mmap_sem must have been held on entry, and may have been
3062 * released depending on flags and vma->vm_ops->fault() return value.
3063 * See filemap_fault() and __lock_page_retry().
3064 */
3065static vm_fault_t __do_fault(struct vm_fault *vmf)
3066{
3067	struct vm_area_struct *vma = vmf->vma;
 
3068	vm_fault_t ret;
3069
3070	/*
3071	 * Preallocate pte before we take page_lock because this might lead to
3072	 * deadlocks for memcg reclaim which waits for pages under writeback:
3073	 *				lock_page(A)
3074	 *				SetPageWriteback(A)
3075	 *				unlock_page(A)
3076	 * lock_page(B)
3077	 *				lock_page(B)
3078	 * pte_alloc_pne
3079	 *   shrink_page_list
3080	 *     wait_on_page_writeback(A)
3081	 *				SetPageWriteback(B)
3082	 *				unlock_page(B)
3083	 *				# flush A, B to clear the writeback
3084	 */
3085	if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3086		vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3087		if (!vmf->prealloc_pte)
3088			return VM_FAULT_OOM;
3089		smp_wmb(); /* See comment in __pte_alloc() */
3090	}
3091
3092	ret = vma->vm_ops->fault(vmf);
3093	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3094			    VM_FAULT_DONE_COW)))
3095		return ret;
3096
 
3097	if (unlikely(PageHWPoison(vmf->page))) {
3098		if (ret & VM_FAULT_LOCKED)
3099			unlock_page(vmf->page);
3100		put_page(vmf->page);
 
 
 
 
 
 
 
3101		vmf->page = NULL;
3102		return VM_FAULT_HWPOISON;
3103	}
3104
3105	if (unlikely(!(ret & VM_FAULT_LOCKED)))
3106		lock_page(vmf->page);
3107	else
3108		VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3109
3110	return ret;
3111}
3112
3113/*
3114 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3115 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3116 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3117 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3118 */
3119static int pmd_devmap_trans_unstable(pmd_t *pmd)
3120{
3121	return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3122}
3123
3124static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3125{
3126	struct vm_area_struct *vma = vmf->vma;
3127
3128	if (!pmd_none(*vmf->pmd))
3129		goto map_pte;
3130	if (vmf->prealloc_pte) {
3131		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3132		if (unlikely(!pmd_none(*vmf->pmd))) {
3133			spin_unlock(vmf->ptl);
3134			goto map_pte;
3135		}
3136
3137		mm_inc_nr_ptes(vma->vm_mm);
3138		pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3139		spin_unlock(vmf->ptl);
3140		vmf->prealloc_pte = NULL;
3141	} else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3142		return VM_FAULT_OOM;
3143	}
3144map_pte:
3145	/*
3146	 * If a huge pmd materialized under us just retry later.  Use
3147	 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3148	 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3149	 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3150	 * running immediately after a huge pmd fault in a different thread of
3151	 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3152	 * All we have to ensure is that it is a regular pmd that we can walk
3153	 * with pte_offset_map() and we can do that through an atomic read in
3154	 * C, which is what pmd_trans_unstable() provides.
3155	 */
3156	if (pmd_devmap_trans_unstable(vmf->pmd))
3157		return VM_FAULT_NOPAGE;
3158
3159	/*
3160	 * At this point we know that our vmf->pmd points to a page of ptes
3161	 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3162	 * for the duration of the fault.  If a racing MADV_DONTNEED runs and
3163	 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3164	 * be valid and we will re-check to make sure the vmf->pte isn't
3165	 * pte_none() under vmf->ptl protection when we return to
3166	 * alloc_set_pte().
3167	 */
3168	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3169			&vmf->ptl);
3170	return 0;
3171}
3172
3173#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3174static void deposit_prealloc_pte(struct vm_fault *vmf)
3175{
3176	struct vm_area_struct *vma = vmf->vma;
3177
3178	pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3179	/*
3180	 * We are going to consume the prealloc table,
3181	 * count that as nr_ptes.
3182	 */
3183	mm_inc_nr_ptes(vma->vm_mm);
3184	vmf->prealloc_pte = NULL;
3185}
3186
3187static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3188{
 
3189	struct vm_area_struct *vma = vmf->vma;
3190	bool write = vmf->flags & FAULT_FLAG_WRITE;
3191	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3192	pmd_t entry;
3193	int i;
3194	vm_fault_t ret;
3195
3196	if (!transhuge_vma_suitable(vma, haddr))
3197		return VM_FAULT_FALLBACK;
3198
3199	ret = VM_FAULT_FALLBACK;
3200	page = compound_head(page);
3201
3202	/*
3203	 * Archs like ppc64 need additonal space to store information
 
 
 
 
 
 
 
 
 
3204	 * related to pte entry. Use the preallocated table for that.
3205	 */
3206	if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3207		vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3208		if (!vmf->prealloc_pte)
3209			return VM_FAULT_OOM;
3210		smp_wmb(); /* See comment in __pte_alloc() */
3211	}
3212
3213	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3214	if (unlikely(!pmd_none(*vmf->pmd)))
3215		goto out;
3216
3217	for (i = 0; i < HPAGE_PMD_NR; i++)
3218		flush_icache_page(vma, page + i);
3219
3220	entry = mk_huge_pmd(page, vma->vm_page_prot);
3221	if (write)
3222		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3223
3224	add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3225	page_add_file_rmap(page, true);
 
3226	/*
3227	 * deposit and withdraw with pmd lock held
3228	 */
3229	if (arch_needs_pgtable_deposit())
3230		deposit_prealloc_pte(vmf);
3231
3232	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3233
3234	update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3235
3236	/* fault is handled */
3237	ret = 0;
3238	count_vm_event(THP_FILE_MAPPED);
3239out:
3240	spin_unlock(vmf->ptl);
3241	return ret;
3242}
3243#else
3244static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3245{
3246	BUILD_BUG();
3247	return 0;
3248}
3249#endif
3250
3251/**
3252 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3253 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3254 *
3255 * @vmf: fault environment
3256 * @memcg: memcg to charge page (only for private mappings)
3257 * @page: page to map
3258 *
3259 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3260 * return.
3261 *
3262 * Target users are page handler itself and implementations of
3263 * vm_ops->map_pages.
3264 *
3265 * Return: %0 on success, %VM_FAULT_ code in case of error.
3266 */
3267vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3268		struct page *page)
3269{
3270	struct vm_area_struct *vma = vmf->vma;
 
3271	bool write = vmf->flags & FAULT_FLAG_WRITE;
 
3272	pte_t entry;
3273	vm_fault_t ret;
3274
3275	if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3276			IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3277		/* THP on COW? */
3278		VM_BUG_ON_PAGE(memcg, page);
3279
3280		ret = do_set_pmd(vmf, page);
3281		if (ret != VM_FAULT_FALLBACK)
3282			return ret;
3283	}
3284
3285	if (!vmf->pte) {
3286		ret = pte_alloc_one_map(vmf);
3287		if (ret)
3288			return ret;
3289	}
3290
3291	/* Re-check under ptl */
3292	if (unlikely(!pte_none(*vmf->pte)))
3293		return VM_FAULT_NOPAGE;
3294
3295	flush_icache_page(vma, page);
3296	entry = mk_pte(page, vma->vm_page_prot);
3297	if (write)
3298		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 
 
3299	/* copy-on-write page */
3300	if (write && !(vma->vm_flags & VM_SHARED)) {
3301		inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3302		page_add_new_anon_rmap(page, vma, vmf->address, false);
3303		mem_cgroup_commit_charge(page, memcg, false, false);
3304		lru_cache_add_active_or_unevictable(page, vma);
3305	} else {
3306		inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3307		page_add_file_rmap(page, false);
3308	}
3309	set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3310
3311	/* no need to invalidate: a not-present page won't be cached */
3312	update_mmu_cache(vma, vmf->address, vmf->pte);
 
3313
3314	return 0;
 
 
 
 
 
3315}
3316
3317
3318/**
3319 * finish_fault - finish page fault once we have prepared the page to fault
3320 *
3321 * @vmf: structure describing the fault
3322 *
3323 * This function handles all that is needed to finish a page fault once the
3324 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3325 * given page, adds reverse page mapping, handles memcg charges and LRU
3326 * addition.
3327 *
3328 * The function expects the page to be locked and on success it consumes a
3329 * reference of a page being mapped (for the PTE which maps it).
3330 *
3331 * Return: %0 on success, %VM_FAULT_ code in case of error.
3332 */
3333vm_fault_t finish_fault(struct vm_fault *vmf)
3334{
 
3335	struct page *page;
3336	vm_fault_t ret = 0;
3337
3338	/* Did we COW the page? */
3339	if ((vmf->flags & FAULT_FLAG_WRITE) &&
3340	    !(vmf->vma->vm_flags & VM_SHARED))
3341		page = vmf->cow_page;
3342	else
3343		page = vmf->page;
3344
3345	/*
3346	 * check even for read faults because we might have lost our CoWed
3347	 * page
3348	 */
3349	if (!(vmf->vma->vm_flags & VM_SHARED))
3350		ret = check_stable_address_space(vmf->vma->vm_mm);
3351	if (!ret)
3352		ret = alloc_set_pte(vmf, vmf->memcg, page);
3353	if (vmf->pte)
3354		pte_unmap_unlock(vmf->pte, vmf->ptl);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3355	return ret;
3356}
3357
3358static unsigned long fault_around_bytes __read_mostly =
3359	rounddown_pow_of_two(65536);
3360
3361#ifdef CONFIG_DEBUG_FS
3362static int fault_around_bytes_get(void *data, u64 *val)
3363{
3364	*val = fault_around_bytes;
3365	return 0;
3366}
3367
3368/*
3369 * fault_around_bytes must be rounded down to the nearest page order as it's
3370 * what do_fault_around() expects to see.
3371 */
3372static int fault_around_bytes_set(void *data, u64 val)
3373{
3374	if (val / PAGE_SIZE > PTRS_PER_PTE)
3375		return -EINVAL;
3376	if (val > PAGE_SIZE)
3377		fault_around_bytes = rounddown_pow_of_two(val);
3378	else
3379		fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
 
 
 
3380	return 0;
3381}
3382DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3383		fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3384
3385static int __init fault_around_debugfs(void)
3386{
3387	debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3388				   &fault_around_bytes_fops);
3389	return 0;
3390}
3391late_initcall(fault_around_debugfs);
3392#endif
3393
3394/*
3395 * do_fault_around() tries to map few pages around the fault address. The hope
3396 * is that the pages will be needed soon and this will lower the number of
3397 * faults to handle.
3398 *
3399 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3400 * not ready to be mapped: not up-to-date, locked, etc.
3401 *
3402 * This function is called with the page table lock taken. In the split ptlock
3403 * case the page table lock only protects only those entries which belong to
3404 * the page table corresponding to the fault address.
3405 *
3406 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3407 * only once.
3408 *
3409 * fault_around_bytes defines how many bytes we'll try to map.
3410 * do_fault_around() expects it to be set to a power of two less than or equal
3411 * to PTRS_PER_PTE.
3412 *
3413 * The virtual address of the area that we map is naturally aligned to
3414 * fault_around_bytes rounded down to the machine page size
3415 * (and therefore to page order).  This way it's easier to guarantee
3416 * that we don't cross page table boundaries.
3417 */
3418static vm_fault_t do_fault_around(struct vm_fault *vmf)
3419{
3420	unsigned long address = vmf->address, nr_pages, mask;
3421	pgoff_t start_pgoff = vmf->pgoff;
3422	pgoff_t end_pgoff;
3423	int off;
3424	vm_fault_t ret = 0;
3425
3426	nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3427	mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3428
3429	vmf->address = max(address & mask, vmf->vma->vm_start);
3430	off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3431	start_pgoff -= off;
3432
3433	/*
3434	 *  end_pgoff is either the end of the page table, the end of
3435	 *  the vma or nr_pages from start_pgoff, depending what is nearest.
3436	 */
3437	end_pgoff = start_pgoff -
3438		((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3439		PTRS_PER_PTE - 1;
3440	end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3441			start_pgoff + nr_pages - 1);
3442
3443	if (pmd_none(*vmf->pmd)) {
3444		vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3445		if (!vmf->prealloc_pte)
3446			goto out;
3447		smp_wmb(); /* See comment in __pte_alloc() */
3448	}
3449
3450	vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
 
 
 
 
3451
3452	/* Huge page is mapped? Page fault is solved */
3453	if (pmd_trans_huge(*vmf->pmd)) {
3454		ret = VM_FAULT_NOPAGE;
3455		goto out;
3456	}
 
 
 
 
3457
3458	/* ->map_pages() haven't done anything useful. Cold page cache? */
3459	if (!vmf->pte)
3460		goto out;
3461
3462	/* check if the page fault is solved */
3463	vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3464	if (!pte_none(*vmf->pte))
3465		ret = VM_FAULT_NOPAGE;
3466	pte_unmap_unlock(vmf->pte, vmf->ptl);
3467out:
3468	vmf->address = address;
3469	vmf->pte = NULL;
3470	return ret;
3471}
3472
3473static vm_fault_t do_read_fault(struct vm_fault *vmf)
3474{
3475	struct vm_area_struct *vma = vmf->vma;
3476	vm_fault_t ret = 0;
 
3477
3478	/*
3479	 * Let's call ->map_pages() first and use ->fault() as fallback
3480	 * if page by the offset is not ready to be mapped (cold cache or
3481	 * something).
3482	 */
3483	if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3484		ret = do_fault_around(vmf);
3485		if (ret)
3486			return ret;
3487	}
3488
 
 
 
 
3489	ret = __do_fault(vmf);
3490	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3491		return ret;
3492
3493	ret |= finish_fault(vmf);
3494	unlock_page(vmf->page);
 
3495	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3496		put_page(vmf->page);
3497	return ret;
3498}
3499
3500static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3501{
3502	struct vm_area_struct *vma = vmf->vma;
 
3503	vm_fault_t ret;
3504
3505	if (unlikely(anon_vma_prepare(vma)))
3506		return VM_FAULT_OOM;
 
 
 
3507
3508	vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3509	if (!vmf->cow_page)
3510		return VM_FAULT_OOM;
3511
3512	if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3513				&vmf->memcg, false)) {
3514		put_page(vmf->cow_page);
3515		return VM_FAULT_OOM;
3516	}
3517
3518	ret = __do_fault(vmf);
3519	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3520		goto uncharge_out;
3521	if (ret & VM_FAULT_DONE_COW)
3522		return ret;
3523
3524	copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3525	__SetPageUptodate(vmf->cow_page);
3526
3527	ret |= finish_fault(vmf);
3528	unlock_page(vmf->page);
3529	put_page(vmf->page);
3530	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3531		goto uncharge_out;
3532	return ret;
3533uncharge_out:
3534	mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3535	put_page(vmf->cow_page);
3536	return ret;
3537}
3538
3539static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3540{
3541	struct vm_area_struct *vma = vmf->vma;
3542	vm_fault_t ret, tmp;
 
 
 
 
 
3543
3544	ret = __do_fault(vmf);
3545	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3546		return ret;
3547
 
 
3548	/*
3549	 * Check if the backing address space wants to know that the page is
3550	 * about to become writable
3551	 */
3552	if (vma->vm_ops->page_mkwrite) {
3553		unlock_page(vmf->page);
3554		tmp = do_page_mkwrite(vmf);
3555		if (unlikely(!tmp ||
3556				(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3557			put_page(vmf->page);
3558			return tmp;
3559		}
3560	}
3561
3562	ret |= finish_fault(vmf);
3563	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3564					VM_FAULT_RETRY))) {
3565		unlock_page(vmf->page);
3566		put_page(vmf->page);
3567		return ret;
3568	}
3569
3570	fault_dirty_shared_page(vma, vmf->page);
3571	return ret;
3572}
3573
3574/*
3575 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3576 * but allow concurrent faults).
3577 * The mmap_sem may have been released depending on flags and our
3578 * return value.  See filemap_fault() and __lock_page_or_retry().
3579 * If mmap_sem is released, vma may become invalid (for example
3580 * by other thread calling munmap()).
3581 */
3582static vm_fault_t do_fault(struct vm_fault *vmf)
3583{
3584	struct vm_area_struct *vma = vmf->vma;
3585	struct mm_struct *vm_mm = vma->vm_mm;
3586	vm_fault_t ret;
3587
3588	/*
3589	 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3590	 */
3591	if (!vma->vm_ops->fault) {
3592		/*
3593		 * If we find a migration pmd entry or a none pmd entry, which
3594		 * should never happen, return SIGBUS
3595		 */
3596		if (unlikely(!pmd_present(*vmf->pmd)))
3597			ret = VM_FAULT_SIGBUS;
3598		else {
3599			vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3600						       vmf->pmd,
3601						       vmf->address,
3602						       &vmf->ptl);
3603			/*
3604			 * Make sure this is not a temporary clearing of pte
3605			 * by holding ptl and checking again. A R/M/W update
3606			 * of pte involves: take ptl, clearing the pte so that
3607			 * we don't have concurrent modification by hardware
3608			 * followed by an update.
3609			 */
3610			if (unlikely(pte_none(*vmf->pte)))
3611				ret = VM_FAULT_SIGBUS;
3612			else
3613				ret = VM_FAULT_NOPAGE;
3614
3615			pte_unmap_unlock(vmf->pte, vmf->ptl);
3616		}
3617	} else if (!(vmf->flags & FAULT_FLAG_WRITE))
3618		ret = do_read_fault(vmf);
3619	else if (!(vma->vm_flags & VM_SHARED))
3620		ret = do_cow_fault(vmf);
3621	else
3622		ret = do_shared_fault(vmf);
3623
3624	/* preallocated pagetable is unused: free it */
3625	if (vmf->prealloc_pte) {
3626		pte_free(vm_mm, vmf->prealloc_pte);
3627		vmf->prealloc_pte = NULL;
3628	}
3629	return ret;
3630}
3631
3632static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3633				unsigned long addr, int page_nid,
3634				int *flags)
3635{
3636	get_page(page);
 
 
 
3637
3638	count_vm_numa_event(NUMA_HINT_FAULTS);
3639	if (page_nid == numa_node_id()) {
3640		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3641		*flags |= TNF_FAULT_LOCAL;
3642	}
3643
3644	return mpol_misplaced(page, vma, addr);
3645}
3646
3647static vm_fault_t do_numa_page(struct vm_fault *vmf)
3648{
3649	struct vm_area_struct *vma = vmf->vma;
3650	struct page *page = NULL;
3651	int page_nid = NUMA_NO_NODE;
 
3652	int last_cpupid;
3653	int target_nid;
3654	bool migrated = false;
3655	pte_t pte, old_pte;
3656	bool was_writable = pte_savedwrite(vmf->orig_pte);
3657	int flags = 0;
3658
3659	/*
3660	 * The "pte" at this point cannot be used safely without
3661	 * validation through pte_unmap_same(). It's of NUMA type but
3662	 * the pfn may be screwed if the read is non atomic.
3663	 */
3664	vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3665	spin_lock(vmf->ptl);
3666	if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3667		pte_unmap_unlock(vmf->pte, vmf->ptl);
3668		goto out;
3669	}
3670
 
 
 
 
3671	/*
3672	 * Make it present again, Depending on how arch implementes non
3673	 * accessible ptes, some can allow access by kernel mode.
3674	 */
3675	old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
3676	pte = pte_modify(old_pte, vma->vm_page_prot);
3677	pte = pte_mkyoung(pte);
3678	if (was_writable)
3679		pte = pte_mkwrite(pte);
3680	ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
3681	update_mmu_cache(vma, vmf->address, vmf->pte);
3682
3683	page = vm_normal_page(vma, vmf->address, pte);
3684	if (!page) {
3685		pte_unmap_unlock(vmf->pte, vmf->ptl);
3686		return 0;
3687	}
3688
3689	/* TODO: handle PTE-mapped THP */
3690	if (PageCompound(page)) {
3691		pte_unmap_unlock(vmf->pte, vmf->ptl);
3692		return 0;
3693	}
3694
3695	/*
3696	 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3697	 * much anyway since they can be in shared cache state. This misses
3698	 * the case where a mapping is writable but the process never writes
3699	 * to it but pte_write gets cleared during protection updates and
3700	 * pte_dirty has unpredictable behaviour between PTE scan updates,
3701	 * background writeback, dirty balancing and application behaviour.
3702	 */
3703	if (!pte_write(pte))
3704		flags |= TNF_NO_GROUP;
3705
3706	/*
3707	 * Flag if the page is shared between multiple address spaces. This
3708	 * is later used when determining whether to group tasks together
3709	 */
3710	if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3711		flags |= TNF_SHARED;
3712
3713	last_cpupid = page_cpupid_last(page);
3714	page_nid = page_to_nid(page);
3715	target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3716			&flags);
3717	pte_unmap_unlock(vmf->pte, vmf->ptl);
 
 
 
 
 
 
3718	if (target_nid == NUMA_NO_NODE) {
3719		put_page(page);
3720		goto out;
3721	}
 
 
3722
3723	/* Migrate to the requested node */
3724	migrated = migrate_misplaced_page(page, vma, target_nid);
3725	if (migrated) {
3726		page_nid = target_nid;
3727		flags |= TNF_MIGRATED;
3728	} else
3729		flags |= TNF_MIGRATE_FAIL;
 
 
 
 
 
 
 
 
 
 
3730
3731out:
3732	if (page_nid != NUMA_NO_NODE)
3733		task_numa_fault(last_cpupid, page_nid, 1, flags);
3734	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3735}
3736
3737static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
3738{
3739	if (vma_is_anonymous(vmf->vma))
 
3740		return do_huge_pmd_anonymous_page(vmf);
3741	if (vmf->vma->vm_ops->huge_fault)
3742		return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3743	return VM_FAULT_FALLBACK;
3744}
3745
3746/* `inline' is required to avoid gcc 4.1.2 build error */
3747static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3748{
3749	if (vma_is_anonymous(vmf->vma))
3750		return do_huge_pmd_wp_page(vmf, orig_pmd);
3751	if (vmf->vma->vm_ops->huge_fault)
3752		return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3753
3754	/* COW handled on pte level: split pmd */
3755	VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3756	__split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3757
3758	return VM_FAULT_FALLBACK;
3759}
3760
3761static inline bool vma_is_accessible(struct vm_area_struct *vma)
3762{
3763	return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3764}
3765
3766static vm_fault_t create_huge_pud(struct vm_fault *vmf)
3767{
3768#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 
 
3769	/* No support for anonymous transparent PUD pages yet */
3770	if (vma_is_anonymous(vmf->vma))
3771		return VM_FAULT_FALLBACK;
3772	if (vmf->vma->vm_ops->huge_fault)
3773		return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3774#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3775	return VM_FAULT_FALLBACK;
3776}
3777
3778static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3779{
3780#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 
 
 
 
3781	/* No support for anonymous transparent PUD pages yet */
3782	if (vma_is_anonymous(vmf->vma))
3783		return VM_FAULT_FALLBACK;
3784	if (vmf->vma->vm_ops->huge_fault)
3785		return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3786#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 
 
 
 
 
 
 
 
3787	return VM_FAULT_FALLBACK;
3788}
3789
3790/*
3791 * These routines also need to handle stuff like marking pages dirty
3792 * and/or accessed for architectures that don't do it in hardware (most
3793 * RISC architectures).  The early dirtying is also good on the i386.
3794 *
3795 * There is also a hook called "update_mmu_cache()" that architectures
3796 * with external mmu caches can use to update those (ie the Sparc or
3797 * PowerPC hashed page tables that act as extended TLBs).
3798 *
3799 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3800 * concurrent faults).
3801 *
3802 * The mmap_sem may have been released depending on flags and our return value.
3803 * See filemap_fault() and __lock_page_or_retry().
3804 */
3805static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
3806{
3807	pte_t entry;
3808
3809	if (unlikely(pmd_none(*vmf->pmd))) {
3810		/*
3811		 * Leave __pte_alloc() until later: because vm_ops->fault may
3812		 * want to allocate huge page, and if we expose page table
3813		 * for an instant, it will be difficult to retract from
3814		 * concurrent faults and from rmap lookups.
3815		 */
3816		vmf->pte = NULL;
 
3817	} else {
3818		/* See comment in pte_alloc_one_map() */
3819		if (pmd_devmap_trans_unstable(vmf->pmd))
3820			return 0;
3821		/*
3822		 * A regular pmd is established and it can't morph into a huge
3823		 * pmd from under us anymore at this point because we hold the
3824		 * mmap_sem read mode and khugepaged takes it in write mode.
3825		 * So now it's safe to run pte_offset_map().
3826		 */
3827		vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3828		vmf->orig_pte = *vmf->pte;
 
 
 
 
3829
3830		/*
3831		 * some architectures can have larger ptes than wordsize,
3832		 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3833		 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3834		 * accesses.  The code below just needs a consistent view
3835		 * for the ifs and we later double check anyway with the
3836		 * ptl lock held. So here a barrier will do.
3837		 */
3838		barrier();
3839		if (pte_none(vmf->orig_pte)) {
3840			pte_unmap(vmf->pte);
3841			vmf->pte = NULL;
3842		}
3843	}
3844
3845	if (!vmf->pte) {
3846		if (vma_is_anonymous(vmf->vma))
3847			return do_anonymous_page(vmf);
3848		else
3849			return do_fault(vmf);
3850	}
3851
3852	if (!pte_present(vmf->orig_pte))
3853		return do_swap_page(vmf);
3854
3855	if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3856		return do_numa_page(vmf);
3857
3858	vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3859	spin_lock(vmf->ptl);
3860	entry = vmf->orig_pte;
3861	if (unlikely(!pte_same(*vmf->pte, entry)))
 
3862		goto unlock;
3863	if (vmf->flags & FAULT_FLAG_WRITE) {
 
3864		if (!pte_write(entry))
3865			return do_wp_page(vmf);
3866		entry = pte_mkdirty(entry);
 
3867	}
3868	entry = pte_mkyoung(entry);
3869	if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3870				vmf->flags & FAULT_FLAG_WRITE)) {
3871		update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
 
3872	} else {
 
 
 
3873		/*
3874		 * This is needed only for protection faults but the arch code
3875		 * is not yet telling us if this is a protection fault or not.
3876		 * This still avoids useless tlb flushes for .text page faults
3877		 * with threads.
3878		 */
3879		if (vmf->flags & FAULT_FLAG_WRITE)
3880			flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
 
3881	}
3882unlock:
3883	pte_unmap_unlock(vmf->pte, vmf->ptl);
3884	return 0;
3885}
3886
3887/*
3888 * By the time we get here, we already hold the mm semaphore
3889 *
3890 * The mmap_sem may have been released depending on flags and our
3891 * return value.  See filemap_fault() and __lock_page_or_retry().
3892 */
3893static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
3894		unsigned long address, unsigned int flags)
3895{
3896	struct vm_fault vmf = {
3897		.vma = vma,
3898		.address = address & PAGE_MASK,
 
3899		.flags = flags,
3900		.pgoff = linear_page_index(vma, address),
3901		.gfp_mask = __get_fault_gfp_mask(vma),
3902	};
3903	unsigned int dirty = flags & FAULT_FLAG_WRITE;
3904	struct mm_struct *mm = vma->vm_mm;
 
3905	pgd_t *pgd;
3906	p4d_t *p4d;
3907	vm_fault_t ret;
3908
3909	pgd = pgd_offset(mm, address);
3910	p4d = p4d_alloc(mm, pgd, address);
3911	if (!p4d)
3912		return VM_FAULT_OOM;
3913
3914	vmf.pud = pud_alloc(mm, p4d, address);
3915	if (!vmf.pud)
3916		return VM_FAULT_OOM;
3917	if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
 
 
3918		ret = create_huge_pud(&vmf);
3919		if (!(ret & VM_FAULT_FALLBACK))
3920			return ret;
3921	} else {
3922		pud_t orig_pud = *vmf.pud;
3923
3924		barrier();
3925		if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
3926
3927			/* NUMA case for anonymous PUDs would go here */
3928
3929			if (dirty && !pud_write(orig_pud)) {
 
 
3930				ret = wp_huge_pud(&vmf, orig_pud);
3931				if (!(ret & VM_FAULT_FALLBACK))
3932					return ret;
3933			} else {
3934				huge_pud_set_accessed(&vmf, orig_pud);
3935				return 0;
3936			}
3937		}
3938	}
3939
3940	vmf.pmd = pmd_alloc(mm, vmf.pud, address);
3941	if (!vmf.pmd)
3942		return VM_FAULT_OOM;
3943	if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
 
 
 
 
 
 
3944		ret = create_huge_pmd(&vmf);
3945		if (!(ret & VM_FAULT_FALLBACK))
3946			return ret;
3947	} else {
3948		pmd_t orig_pmd = *vmf.pmd;
3949
3950		barrier();
3951		if (unlikely(is_swap_pmd(orig_pmd))) {
3952			VM_BUG_ON(thp_migration_supported() &&
3953					  !is_pmd_migration_entry(orig_pmd));
3954			if (is_pmd_migration_entry(orig_pmd))
3955				pmd_migration_entry_wait(mm, vmf.pmd);
3956			return 0;
3957		}
3958		if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3959			if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3960				return do_huge_pmd_numa_page(&vmf, orig_pmd);
3961
3962			if (dirty && !pmd_write(orig_pmd)) {
3963				ret = wp_huge_pmd(&vmf, orig_pmd);
 
3964				if (!(ret & VM_FAULT_FALLBACK))
3965					return ret;
3966			} else {
3967				huge_pmd_set_accessed(&vmf, orig_pmd);
3968				return 0;
3969			}
3970		}
3971	}
3972
3973	return handle_pte_fault(&vmf);
3974}
3975
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3976/*
3977 * By the time we get here, we already hold the mm semaphore
3978 *
3979 * The mmap_sem may have been released depending on flags and our
3980 * return value.  See filemap_fault() and __lock_page_or_retry().
3981 */
3982vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3983		unsigned int flags)
3984{
 
 
3985	vm_fault_t ret;
3986
3987	__set_current_state(TASK_RUNNING);
3988
3989	count_vm_event(PGFAULT);
3990	count_memcg_event_mm(vma->vm_mm, PGFAULT);
3991
3992	/* do counter updates before entering really critical section. */
3993	check_sync_rss_stat(current);
3994
3995	if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3996					    flags & FAULT_FLAG_INSTRUCTION,
3997					    flags & FAULT_FLAG_REMOTE))
3998		return VM_FAULT_SIGSEGV;
 
 
3999
4000	/*
4001	 * Enable the memcg OOM handling for faults triggered in user
4002	 * space.  Kernel faults are handled more gracefully.
4003	 */
4004	if (flags & FAULT_FLAG_USER)
4005		mem_cgroup_enter_user_fault();
4006
 
 
4007	if (unlikely(is_vm_hugetlb_page(vma)))
4008		ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4009	else
4010		ret = __handle_mm_fault(vma, address, flags);
4011
 
 
4012	if (flags & FAULT_FLAG_USER) {
4013		mem_cgroup_exit_user_fault();
4014		/*
4015		 * The task may have entered a memcg OOM situation but
4016		 * if the allocation error was handled gracefully (no
4017		 * VM_FAULT_OOM), there is no need to kill anything.
4018		 * Just clean up the OOM state peacefully.
4019		 */
4020		if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4021			mem_cgroup_oom_synchronize(false);
4022	}
 
 
4023
4024	return ret;
4025}
4026EXPORT_SYMBOL_GPL(handle_mm_fault);
4027
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4028#ifndef __PAGETABLE_P4D_FOLDED
4029/*
4030 * Allocate p4d page table.
4031 * We've already handled the fast-path in-line.
4032 */
4033int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4034{
4035	p4d_t *new = p4d_alloc_one(mm, address);
4036	if (!new)
4037		return -ENOMEM;
4038
4039	smp_wmb(); /* See comment in __pte_alloc */
4040
4041	spin_lock(&mm->page_table_lock);
4042	if (pgd_present(*pgd))		/* Another has populated it */
4043		p4d_free(mm, new);
4044	else
 
4045		pgd_populate(mm, pgd, new);
 
4046	spin_unlock(&mm->page_table_lock);
4047	return 0;
4048}
4049#endif /* __PAGETABLE_P4D_FOLDED */
4050
4051#ifndef __PAGETABLE_PUD_FOLDED
4052/*
4053 * Allocate page upper directory.
4054 * We've already handled the fast-path in-line.
4055 */
4056int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4057{
4058	pud_t *new = pud_alloc_one(mm, address);
4059	if (!new)
4060		return -ENOMEM;
4061
4062	smp_wmb(); /* See comment in __pte_alloc */
4063
4064	spin_lock(&mm->page_table_lock);
4065#ifndef __ARCH_HAS_5LEVEL_HACK
4066	if (!p4d_present(*p4d)) {
4067		mm_inc_nr_puds(mm);
 
4068		p4d_populate(mm, p4d, new);
4069	} else	/* Another has populated it */
4070		pud_free(mm, new);
4071#else
4072	if (!pgd_present(*p4d)) {
4073		mm_inc_nr_puds(mm);
4074		pgd_populate(mm, p4d, new);
4075	} else	/* Another has populated it */
4076		pud_free(mm, new);
4077#endif /* __ARCH_HAS_5LEVEL_HACK */
4078	spin_unlock(&mm->page_table_lock);
4079	return 0;
4080}
4081#endif /* __PAGETABLE_PUD_FOLDED */
4082
4083#ifndef __PAGETABLE_PMD_FOLDED
4084/*
4085 * Allocate page middle directory.
4086 * We've already handled the fast-path in-line.
4087 */
4088int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4089{
4090	spinlock_t *ptl;
4091	pmd_t *new = pmd_alloc_one(mm, address);
4092	if (!new)
4093		return -ENOMEM;
4094
4095	smp_wmb(); /* See comment in __pte_alloc */
4096
4097	ptl = pud_lock(mm, pud);
4098#ifndef __ARCH_HAS_4LEVEL_HACK
4099	if (!pud_present(*pud)) {
4100		mm_inc_nr_pmds(mm);
 
4101		pud_populate(mm, pud, new);
4102	} else	/* Another has populated it */
4103		pmd_free(mm, new);
4104#else
4105	if (!pgd_present(*pud)) {
4106		mm_inc_nr_pmds(mm);
4107		pgd_populate(mm, pud, new);
4108	} else /* Another has populated it */
4109		pmd_free(mm, new);
4110#endif /* __ARCH_HAS_4LEVEL_HACK */
4111	spin_unlock(ptl);
4112	return 0;
4113}
4114#endif /* __PAGETABLE_PMD_FOLDED */
4115
4116static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4117			    struct mmu_notifier_range *range,
4118			    pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4119{
4120	pgd_t *pgd;
4121	p4d_t *p4d;
4122	pud_t *pud;
4123	pmd_t *pmd;
4124	pte_t *ptep;
4125
4126	pgd = pgd_offset(mm, address);
4127	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4128		goto out;
4129
4130	p4d = p4d_offset(pgd, address);
4131	if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4132		goto out;
4133
4134	pud = pud_offset(p4d, address);
4135	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4136		goto out;
4137
4138	pmd = pmd_offset(pud, address);
4139	VM_BUG_ON(pmd_trans_huge(*pmd));
4140
4141	if (pmd_huge(*pmd)) {
4142		if (!pmdpp)
4143			goto out;
4144
4145		if (range) {
4146			mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4147						NULL, mm, address & PMD_MASK,
4148						(address & PMD_MASK) + PMD_SIZE);
4149			mmu_notifier_invalidate_range_start(range);
4150		}
4151		*ptlp = pmd_lock(mm, pmd);
4152		if (pmd_huge(*pmd)) {
4153			*pmdpp = pmd;
4154			return 0;
4155		}
4156		spin_unlock(*ptlp);
4157		if (range)
4158			mmu_notifier_invalidate_range_end(range);
4159	}
4160
4161	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4162		goto out;
4163
4164	if (range) {
4165		mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4166					address & PAGE_MASK,
4167					(address & PAGE_MASK) + PAGE_SIZE);
4168		mmu_notifier_invalidate_range_start(range);
4169	}
4170	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4171	if (!pte_present(*ptep))
4172		goto unlock;
4173	*ptepp = ptep;
4174	return 0;
4175unlock:
4176	pte_unmap_unlock(ptep, *ptlp);
4177	if (range)
4178		mmu_notifier_invalidate_range_end(range);
4179out:
4180	return -EINVAL;
4181}
4182
4183static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4184			     pte_t **ptepp, spinlock_t **ptlp)
4185{
4186	int res;
4187
4188	/* (void) is needed to make gcc happy */
4189	(void) __cond_lock(*ptlp,
4190			   !(res = __follow_pte_pmd(mm, address, NULL,
4191						    ptepp, NULL, ptlp)));
4192	return res;
4193}
4194
4195int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4196		   struct mmu_notifier_range *range,
4197		   pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4198{
4199	int res;
4200
4201	/* (void) is needed to make gcc happy */
4202	(void) __cond_lock(*ptlp,
4203			   !(res = __follow_pte_pmd(mm, address, range,
4204						    ptepp, pmdpp, ptlp)));
4205	return res;
4206}
4207EXPORT_SYMBOL(follow_pte_pmd);
4208
4209/**
4210 * follow_pfn - look up PFN at a user virtual address
4211 * @vma: memory mapping
4212 * @address: user virtual address
4213 * @pfn: location to store found PFN
4214 *
4215 * Only IO mappings and raw PFN mappings are allowed.
4216 *
 
 
 
4217 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4218 */
4219int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4220	unsigned long *pfn)
4221{
4222	int ret = -EINVAL;
4223	spinlock_t *ptl;
4224	pte_t *ptep;
4225
4226	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4227		return ret;
4228
4229	ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4230	if (ret)
4231		return ret;
4232	*pfn = pte_pfn(*ptep);
4233	pte_unmap_unlock(ptep, ptl);
4234	return 0;
4235}
4236EXPORT_SYMBOL(follow_pfn);
4237
4238#ifdef CONFIG_HAVE_IOREMAP_PROT
4239int follow_phys(struct vm_area_struct *vma,
4240		unsigned long address, unsigned int flags,
4241		unsigned long *prot, resource_size_t *phys)
4242{
4243	int ret = -EINVAL;
4244	pte_t *ptep, pte;
4245	spinlock_t *ptl;
4246
4247	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4248		goto out;
4249
4250	if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4251		goto out;
4252	pte = *ptep;
4253
4254	if ((flags & FOLL_WRITE) && !pte_write(pte))
4255		goto unlock;
4256
4257	*prot = pgprot_val(pte_pgprot(pte));
4258	*phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4259
4260	ret = 0;
4261unlock:
4262	pte_unmap_unlock(ptep, ptl);
4263out:
4264	return ret;
4265}
4266
 
 
 
 
 
 
 
 
 
 
 
 
4267int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4268			void *buf, int len, int write)
4269{
4270	resource_size_t phys_addr;
4271	unsigned long prot = 0;
4272	void __iomem *maddr;
4273	int offset = addr & (PAGE_SIZE-1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4274
4275	if (follow_phys(vma, addr, write, &prot, &phys_addr))
4276		return -EINVAL;
4277
4278	maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4279	if (!maddr)
4280		return -ENOMEM;
4281
 
 
 
 
 
 
 
 
 
 
4282	if (write)
4283		memcpy_toio(maddr + offset, buf, len);
4284	else
4285		memcpy_fromio(buf, maddr + offset, len);
 
 
 
4286	iounmap(maddr);
4287
4288	return len;
4289}
4290EXPORT_SYMBOL_GPL(generic_access_phys);
4291#endif
4292
4293/*
4294 * Access another process' address space as given in mm.  If non-NULL, use the
4295 * given task for page fault accounting.
4296 */
4297int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4298		unsigned long addr, void *buf, int len, unsigned int gup_flags)
4299{
4300	struct vm_area_struct *vma;
4301	void *old_buf = buf;
4302	int write = gup_flags & FOLL_WRITE;
4303
4304	if (down_read_killable(&mm->mmap_sem))
 
 
 
 
 
 
 
4305		return 0;
4306
4307	/* ignore errors, just check how much was successfully transferred */
4308	while (len) {
4309		int bytes, ret, offset;
4310		void *maddr;
4311		struct page *page = NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4312
4313		ret = get_user_pages_remote(tsk, mm, addr, 1,
4314				gup_flags, &page, &vma, NULL);
4315		if (ret <= 0) {
4316#ifndef CONFIG_HAVE_IOREMAP_PROT
4317			break;
4318#else
4319			/*
4320			 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4321			 * we can access using slightly different code.
4322			 */
4323			vma = find_vma(mm, addr);
4324			if (!vma || vma->vm_start > addr)
4325				break;
4326			if (vma->vm_ops && vma->vm_ops->access)
4327				ret = vma->vm_ops->access(vma, addr, buf,
4328							  len, write);
4329			if (ret <= 0)
 
4330				break;
4331			bytes = ret;
4332#endif
4333		} else {
4334			bytes = len;
4335			offset = addr & (PAGE_SIZE-1);
4336			if (bytes > PAGE_SIZE-offset)
4337				bytes = PAGE_SIZE-offset;
4338
4339			maddr = kmap(page);
4340			if (write) {
4341				copy_to_user_page(vma, page, addr,
4342						  maddr + offset, buf, bytes);
4343				set_page_dirty_lock(page);
4344			} else {
4345				copy_from_user_page(vma, page, addr,
4346						    buf, maddr + offset, bytes);
4347			}
4348			kunmap(page);
4349			put_page(page);
4350		}
4351		len -= bytes;
4352		buf += bytes;
4353		addr += bytes;
4354	}
4355	up_read(&mm->mmap_sem);
4356
4357	return buf - old_buf;
4358}
4359
4360/**
4361 * access_remote_vm - access another process' address space
4362 * @mm:		the mm_struct of the target address space
4363 * @addr:	start address to access
4364 * @buf:	source or destination buffer
4365 * @len:	number of bytes to transfer
4366 * @gup_flags:	flags modifying lookup behaviour
4367 *
4368 * The caller must hold a reference on @mm.
4369 *
4370 * Return: number of bytes copied from source to destination.
4371 */
4372int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4373		void *buf, int len, unsigned int gup_flags)
4374{
4375	return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4376}
4377
4378/*
4379 * Access another process' address space.
4380 * Source/target buffer must be kernel space,
4381 * Do not walk the page table directly, use get_user_pages
4382 */
4383int access_process_vm(struct task_struct *tsk, unsigned long addr,
4384		void *buf, int len, unsigned int gup_flags)
4385{
4386	struct mm_struct *mm;
4387	int ret;
4388
4389	mm = get_task_mm(tsk);
4390	if (!mm)
4391		return 0;
4392
4393	ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4394
4395	mmput(mm);
4396
4397	return ret;
4398}
4399EXPORT_SYMBOL_GPL(access_process_vm);
4400
4401/*
4402 * Print the name of a VMA.
4403 */
4404void print_vma_addr(char *prefix, unsigned long ip)
4405{
4406	struct mm_struct *mm = current->mm;
4407	struct vm_area_struct *vma;
4408
4409	/*
4410	 * we might be running from an atomic context so we cannot sleep
4411	 */
4412	if (!down_read_trylock(&mm->mmap_sem))
4413		return;
4414
4415	vma = find_vma(mm, ip);
4416	if (vma && vma->vm_file) {
4417		struct file *f = vma->vm_file;
4418		char *buf = (char *)__get_free_page(GFP_NOWAIT);
4419		if (buf) {
4420			char *p;
4421
4422			p = file_path(f, buf, PAGE_SIZE);
4423			if (IS_ERR(p))
4424				p = "?";
4425			printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4426					vma->vm_start,
4427					vma->vm_end - vma->vm_start);
4428			free_page((unsigned long)buf);
4429		}
4430	}
4431	up_read(&mm->mmap_sem);
4432}
4433
4434#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4435void __might_fault(const char *file, int line)
4436{
4437	/*
4438	 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4439	 * holding the mmap_sem, this is safe because kernel memory doesn't
4440	 * get paged out, therefore we'll never actually fault, and the
4441	 * below annotations will generate false positives.
4442	 */
4443	if (uaccess_kernel())
4444		return;
4445	if (pagefault_disabled())
4446		return;
4447	__might_sleep(file, line, 0);
4448#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4449	if (current->mm)
4450		might_lock_read(&current->mm->mmap_sem);
4451#endif
4452}
4453EXPORT_SYMBOL(__might_fault);
4454#endif
4455
4456#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4457/*
4458 * Process all subpages of the specified huge page with the specified
4459 * operation.  The target subpage will be processed last to keep its
4460 * cache lines hot.
4461 */
4462static inline void process_huge_page(
4463	unsigned long addr_hint, unsigned int pages_per_huge_page,
4464	void (*process_subpage)(unsigned long addr, int idx, void *arg),
4465	void *arg)
4466{
4467	int i, n, base, l;
4468	unsigned long addr = addr_hint &
4469		~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4470
4471	/* Process target subpage last to keep its cache lines hot */
4472	might_sleep();
4473	n = (addr_hint - addr) / PAGE_SIZE;
4474	if (2 * n <= pages_per_huge_page) {
4475		/* If target subpage in first half of huge page */
4476		base = 0;
4477		l = n;
4478		/* Process subpages at the end of huge page */
4479		for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4480			cond_resched();
4481			process_subpage(addr + i * PAGE_SIZE, i, arg);
 
 
4482		}
4483	} else {
4484		/* If target subpage in second half of huge page */
4485		base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4486		l = pages_per_huge_page - n;
4487		/* Process subpages at the begin of huge page */
4488		for (i = 0; i < base; i++) {
4489			cond_resched();
4490			process_subpage(addr + i * PAGE_SIZE, i, arg);
 
 
4491		}
4492	}
4493	/*
4494	 * Process remaining subpages in left-right-left-right pattern
4495	 * towards the target subpage
4496	 */
4497	for (i = 0; i < l; i++) {
4498		int left_idx = base + i;
4499		int right_idx = base + 2 * l - 1 - i;
4500
4501		cond_resched();
4502		process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
 
 
4503		cond_resched();
4504		process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
 
 
4505	}
 
4506}
4507
4508static void clear_gigantic_page(struct page *page,
4509				unsigned long addr,
4510				unsigned int pages_per_huge_page)
4511{
4512	int i;
4513	struct page *p = page;
4514
4515	might_sleep();
4516	for (i = 0; i < pages_per_huge_page;
4517	     i++, p = mem_map_next(p, page, i)) {
4518		cond_resched();
4519		clear_user_highpage(p, addr + i * PAGE_SIZE);
4520	}
4521}
4522
4523static void clear_subpage(unsigned long addr, int idx, void *arg)
4524{
4525	struct page *page = arg;
4526
4527	clear_user_highpage(page + idx, addr);
 
4528}
4529
4530void clear_huge_page(struct page *page,
4531		     unsigned long addr_hint, unsigned int pages_per_huge_page)
4532{
4533	unsigned long addr = addr_hint &
4534		~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4535
4536	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4537		clear_gigantic_page(page, addr, pages_per_huge_page);
4538		return;
4539	}
4540
4541	process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4542}
4543
4544static void copy_user_gigantic_page(struct page *dst, struct page *src,
4545				    unsigned long addr,
4546				    struct vm_area_struct *vma,
4547				    unsigned int pages_per_huge_page)
4548{
4549	int i;
4550	struct page *dst_base = dst;
4551	struct page *src_base = src;
 
 
 
 
4552
4553	for (i = 0; i < pages_per_huge_page; ) {
4554		cond_resched();
4555		copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4556
4557		i++;
4558		dst = mem_map_next(dst, dst_base, i);
4559		src = mem_map_next(src, src_base, i);
4560	}
 
4561}
4562
4563struct copy_subpage_arg {
4564	struct page *dst;
4565	struct page *src;
4566	struct vm_area_struct *vma;
4567};
4568
4569static void copy_subpage(unsigned long addr, int idx, void *arg)
4570{
4571	struct copy_subpage_arg *copy_arg = arg;
4572
4573	copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
4574			   addr, copy_arg->vma);
 
 
 
 
4575}
4576
4577void copy_user_huge_page(struct page *dst, struct page *src,
4578			 unsigned long addr_hint, struct vm_area_struct *vma,
4579			 unsigned int pages_per_huge_page)
4580{
 
4581	unsigned long addr = addr_hint &
4582		~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4583	struct copy_subpage_arg arg = {
4584		.dst = dst,
4585		.src = src,
4586		.vma = vma,
4587	};
4588
4589	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4590		copy_user_gigantic_page(dst, src, addr, vma,
4591					pages_per_huge_page);
4592		return;
4593	}
4594
4595	process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
4596}
4597
4598long copy_huge_page_from_user(struct page *dst_page,
4599				const void __user *usr_src,
4600				unsigned int pages_per_huge_page,
4601				bool allow_pagefault)
4602{
4603	void *src = (void *)usr_src;
4604	void *page_kaddr;
4605	unsigned long i, rc = 0;
4606	unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4607
4608	for (i = 0; i < pages_per_huge_page; i++) {
4609		if (allow_pagefault)
4610			page_kaddr = kmap(dst_page + i);
4611		else
4612			page_kaddr = kmap_atomic(dst_page + i);
4613		rc = copy_from_user(page_kaddr,
4614				(const void __user *)(src + i * PAGE_SIZE),
4615				PAGE_SIZE);
4616		if (allow_pagefault)
4617			kunmap(dst_page + i);
4618		else
4619			kunmap_atomic(page_kaddr);
4620
4621		ret_val -= (PAGE_SIZE - rc);
4622		if (rc)
4623			break;
4624
 
 
4625		cond_resched();
4626	}
4627	return ret_val;
4628}
4629#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4630
4631#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4632
4633static struct kmem_cache *page_ptl_cachep;
4634
4635void __init ptlock_cache_init(void)
4636{
4637	page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4638			SLAB_PANIC, NULL);
4639}
4640
4641bool ptlock_alloc(struct page *page)
4642{
4643	spinlock_t *ptl;
4644
4645	ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4646	if (!ptl)
4647		return false;
4648	page->ptl = ptl;
4649	return true;
4650}
4651
4652void ptlock_free(struct page *page)
4653{
4654	kmem_cache_free(page_ptl_cachep, page->ptl);
4655}
4656#endif