Linux Audio

Check our new training course

Loading...
v5.9
   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#include <linux/perf_event.h>
  75#include <linux/ptrace.h>
  76#include <linux/vmalloc.h>
  77
  78#include <trace/events/kmem.h>
  79
  80#include <asm/io.h>
  81#include <asm/mmu_context.h>
  82#include <asm/pgalloc.h>
  83#include <linux/uaccess.h>
  84#include <asm/tlb.h>
  85#include <asm/tlbflush.h>
 
  86
  87#include "pgalloc-track.h"
  88#include "internal.h"
  89
  90#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
  91#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
  92#endif
  93
  94#ifndef CONFIG_NEED_MULTIPLE_NODES
  95/* use the per-pgdat data instead for discontigmem - mbligh */
  96unsigned long max_mapnr;
  97EXPORT_SYMBOL(max_mapnr);
  98
  99struct page *mem_map;
 100EXPORT_SYMBOL(mem_map);
 101#endif
 102
 103/*
 104 * A number of key systems in x86 including ioremap() rely on the assumption
 105 * that high_memory defines the upper bound on direct map memory, then end
 106 * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
 107 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
 108 * and ZONE_HIGHMEM.
 109 */
 110void *high_memory;
 111EXPORT_SYMBOL(high_memory);
 112
 113/*
 114 * Randomize the address space (stacks, mmaps, brk, etc.).
 115 *
 116 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
 117 *   as ancient (libc5 based) binaries can segfault. )
 118 */
 119int randomize_va_space __read_mostly =
 120#ifdef CONFIG_COMPAT_BRK
 121					1;
 122#else
 123					2;
 124#endif
 125
 126#ifndef arch_faults_on_old_pte
 127static inline bool arch_faults_on_old_pte(void)
 128{
 129	/*
 130	 * Those arches which don't have hw access flag feature need to
 131	 * implement their own helper. By default, "true" means pagefault
 132	 * will be hit on old pte.
 133	 */
 134	return true;
 135}
 136#endif
 137
 138static int __init disable_randmaps(char *s)
 139{
 140	randomize_va_space = 0;
 141	return 1;
 142}
 143__setup("norandmaps", disable_randmaps);
 144
 145unsigned long zero_pfn __read_mostly;
 146EXPORT_SYMBOL(zero_pfn);
 147
 148unsigned long highest_memmap_pfn __read_mostly;
 149
 150/*
 151 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
 152 */
 153static int __init init_zero_pfn(void)
 154{
 155	zero_pfn = page_to_pfn(ZERO_PAGE(0));
 156	return 0;
 157}
 158core_initcall(init_zero_pfn);
 159
 160void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
 161{
 162	trace_rss_stat(mm, member, count);
 163}
 164
 165#if defined(SPLIT_RSS_COUNTING)
 166
 167void sync_mm_rss(struct mm_struct *mm)
 168{
 169	int i;
 170
 171	for (i = 0; i < NR_MM_COUNTERS; i++) {
 172		if (current->rss_stat.count[i]) {
 173			add_mm_counter(mm, i, current->rss_stat.count[i]);
 174			current->rss_stat.count[i] = 0;
 175		}
 176	}
 177	current->rss_stat.events = 0;
 178}
 179
 180static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
 181{
 182	struct task_struct *task = current;
 183
 184	if (likely(task->mm == mm))
 185		task->rss_stat.count[member] += val;
 186	else
 187		add_mm_counter(mm, member, val);
 188}
 189#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
 190#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
 191
 192/* sync counter once per 64 page faults */
 193#define TASK_RSS_EVENTS_THRESH	(64)
 194static void check_sync_rss_stat(struct task_struct *task)
 195{
 196	if (unlikely(task != current))
 197		return;
 198	if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
 199		sync_mm_rss(task->mm);
 200}
 201#else /* SPLIT_RSS_COUNTING */
 202
 203#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
 204#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
 205
 206static void check_sync_rss_stat(struct task_struct *task)
 207{
 208}
 209
 210#endif /* SPLIT_RSS_COUNTING */
 211
 212/*
 213 * Note: this doesn't free the actual pages themselves. That
 214 * has been handled earlier when unmapping all the memory regions.
 215 */
 216static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 217			   unsigned long addr)
 218{
 219	pgtable_t token = pmd_pgtable(*pmd);
 220	pmd_clear(pmd);
 221	pte_free_tlb(tlb, token, addr);
 222	mm_dec_nr_ptes(tlb->mm);
 223}
 224
 225static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 226				unsigned long addr, unsigned long end,
 227				unsigned long floor, unsigned long ceiling)
 228{
 229	pmd_t *pmd;
 230	unsigned long next;
 231	unsigned long start;
 232
 233	start = addr;
 234	pmd = pmd_offset(pud, addr);
 235	do {
 236		next = pmd_addr_end(addr, end);
 237		if (pmd_none_or_clear_bad(pmd))
 238			continue;
 239		free_pte_range(tlb, pmd, addr);
 240	} while (pmd++, addr = next, addr != end);
 241
 242	start &= PUD_MASK;
 243	if (start < floor)
 244		return;
 245	if (ceiling) {
 246		ceiling &= PUD_MASK;
 247		if (!ceiling)
 248			return;
 249	}
 250	if (end - 1 > ceiling - 1)
 251		return;
 252
 253	pmd = pmd_offset(pud, start);
 254	pud_clear(pud);
 255	pmd_free_tlb(tlb, pmd, start);
 256	mm_dec_nr_pmds(tlb->mm);
 257}
 258
 259static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
 260				unsigned long addr, unsigned long end,
 261				unsigned long floor, unsigned long ceiling)
 262{
 263	pud_t *pud;
 264	unsigned long next;
 265	unsigned long start;
 266
 267	start = addr;
 268	pud = pud_offset(p4d, addr);
 269	do {
 270		next = pud_addr_end(addr, end);
 271		if (pud_none_or_clear_bad(pud))
 272			continue;
 273		free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 274	} while (pud++, addr = next, addr != end);
 275
 276	start &= P4D_MASK;
 277	if (start < floor)
 278		return;
 279	if (ceiling) {
 280		ceiling &= P4D_MASK;
 281		if (!ceiling)
 282			return;
 283	}
 284	if (end - 1 > ceiling - 1)
 285		return;
 286
 287	pud = pud_offset(p4d, start);
 288	p4d_clear(p4d);
 289	pud_free_tlb(tlb, pud, start);
 290	mm_dec_nr_puds(tlb->mm);
 291}
 292
 293static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
 294				unsigned long addr, unsigned long end,
 295				unsigned long floor, unsigned long ceiling)
 296{
 297	p4d_t *p4d;
 298	unsigned long next;
 299	unsigned long start;
 300
 301	start = addr;
 302	p4d = p4d_offset(pgd, addr);
 303	do {
 304		next = p4d_addr_end(addr, end);
 305		if (p4d_none_or_clear_bad(p4d))
 306			continue;
 307		free_pud_range(tlb, p4d, addr, next, floor, ceiling);
 308	} while (p4d++, addr = next, addr != end);
 309
 310	start &= PGDIR_MASK;
 311	if (start < floor)
 312		return;
 313	if (ceiling) {
 314		ceiling &= PGDIR_MASK;
 315		if (!ceiling)
 316			return;
 317	}
 318	if (end - 1 > ceiling - 1)
 319		return;
 320
 321	p4d = p4d_offset(pgd, start);
 322	pgd_clear(pgd);
 323	p4d_free_tlb(tlb, p4d, start);
 324}
 325
 326/*
 327 * This function frees user-level page tables of a process.
 328 */
 329void free_pgd_range(struct mmu_gather *tlb,
 330			unsigned long addr, unsigned long end,
 331			unsigned long floor, unsigned long ceiling)
 332{
 333	pgd_t *pgd;
 334	unsigned long next;
 335
 336	/*
 337	 * The next few lines have given us lots of grief...
 338	 *
 339	 * Why are we testing PMD* at this top level?  Because often
 340	 * there will be no work to do at all, and we'd prefer not to
 341	 * go all the way down to the bottom just to discover that.
 342	 *
 343	 * Why all these "- 1"s?  Because 0 represents both the bottom
 344	 * of the address space and the top of it (using -1 for the
 345	 * top wouldn't help much: the masks would do the wrong thing).
 346	 * The rule is that addr 0 and floor 0 refer to the bottom of
 347	 * the address space, but end 0 and ceiling 0 refer to the top
 348	 * Comparisons need to use "end - 1" and "ceiling - 1" (though
 349	 * that end 0 case should be mythical).
 350	 *
 351	 * Wherever addr is brought up or ceiling brought down, we must
 352	 * be careful to reject "the opposite 0" before it confuses the
 353	 * subsequent tests.  But what about where end is brought down
 354	 * by PMD_SIZE below? no, end can't go down to 0 there.
 355	 *
 356	 * Whereas we round start (addr) and ceiling down, by different
 357	 * masks at different levels, in order to test whether a table
 358	 * now has no other vmas using it, so can be freed, we don't
 359	 * bother to round floor or end up - the tests don't need that.
 360	 */
 361
 362	addr &= PMD_MASK;
 363	if (addr < floor) {
 364		addr += PMD_SIZE;
 365		if (!addr)
 366			return;
 367	}
 368	if (ceiling) {
 369		ceiling &= PMD_MASK;
 370		if (!ceiling)
 371			return;
 372	}
 373	if (end - 1 > ceiling - 1)
 374		end -= PMD_SIZE;
 375	if (addr > end - 1)
 376		return;
 377	/*
 378	 * We add page table cache pages with PAGE_SIZE,
 379	 * (see pte_free_tlb()), flush the tlb if we need
 380	 */
 381	tlb_change_page_size(tlb, PAGE_SIZE);
 382	pgd = pgd_offset(tlb->mm, addr);
 383	do {
 384		next = pgd_addr_end(addr, end);
 385		if (pgd_none_or_clear_bad(pgd))
 386			continue;
 387		free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
 388	} while (pgd++, addr = next, addr != end);
 389}
 390
 391void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
 392		unsigned long floor, unsigned long ceiling)
 393{
 394	while (vma) {
 395		struct vm_area_struct *next = vma->vm_next;
 396		unsigned long addr = vma->vm_start;
 397
 398		/*
 399		 * Hide vma from rmap and truncate_pagecache before freeing
 400		 * pgtables
 401		 */
 402		unlink_anon_vmas(vma);
 403		unlink_file_vma(vma);
 404
 405		if (is_vm_hugetlb_page(vma)) {
 406			hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 407				floor, next ? next->vm_start : ceiling);
 408		} else {
 409			/*
 410			 * Optimization: gather nearby vmas into one call down
 411			 */
 412			while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 413			       && !is_vm_hugetlb_page(next)) {
 414				vma = next;
 415				next = vma->vm_next;
 416				unlink_anon_vmas(vma);
 417				unlink_file_vma(vma);
 418			}
 419			free_pgd_range(tlb, addr, vma->vm_end,
 420				floor, next ? next->vm_start : ceiling);
 421		}
 422		vma = next;
 423	}
 424}
 425
 426int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
 427{
 428	spinlock_t *ptl;
 429	pgtable_t new = pte_alloc_one(mm);
 430	if (!new)
 431		return -ENOMEM;
 432
 433	/*
 434	 * Ensure all pte setup (eg. pte page lock and page clearing) are
 435	 * visible before the pte is made visible to other CPUs by being
 436	 * put into page tables.
 437	 *
 438	 * The other side of the story is the pointer chasing in the page
 439	 * table walking code (when walking the page table without locking;
 440	 * ie. most of the time). Fortunately, these data accesses consist
 441	 * of a chain of data-dependent loads, meaning most CPUs (alpha
 442	 * being the notable exception) will already guarantee loads are
 443	 * seen in-order. See the alpha page table accessors for the
 444	 * smp_rmb() barriers in page table walking code.
 445	 */
 446	smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
 447
 448	ptl = pmd_lock(mm, pmd);
 449	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
 450		mm_inc_nr_ptes(mm);
 451		pmd_populate(mm, pmd, new);
 452		new = NULL;
 453	}
 454	spin_unlock(ptl);
 455	if (new)
 456		pte_free(mm, new);
 457	return 0;
 458}
 459
 460int __pte_alloc_kernel(pmd_t *pmd)
 461{
 462	pte_t *new = pte_alloc_one_kernel(&init_mm);
 463	if (!new)
 464		return -ENOMEM;
 465
 466	smp_wmb(); /* See comment in __pte_alloc */
 467
 468	spin_lock(&init_mm.page_table_lock);
 469	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
 470		pmd_populate_kernel(&init_mm, pmd, new);
 471		new = NULL;
 472	}
 473	spin_unlock(&init_mm.page_table_lock);
 474	if (new)
 475		pte_free_kernel(&init_mm, new);
 476	return 0;
 477}
 478
 479static inline void init_rss_vec(int *rss)
 480{
 481	memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
 482}
 483
 484static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
 485{
 486	int i;
 487
 488	if (current->mm == mm)
 489		sync_mm_rss(mm);
 490	for (i = 0; i < NR_MM_COUNTERS; i++)
 491		if (rss[i])
 492			add_mm_counter(mm, i, rss[i]);
 493}
 494
 495/*
 496 * This function is called to print an error when a bad pte
 497 * is found. For example, we might have a PFN-mapped pte in
 498 * a region that doesn't allow it.
 499 *
 500 * The calling function must still handle the error.
 501 */
 502static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
 503			  pte_t pte, struct page *page)
 504{
 505	pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
 506	p4d_t *p4d = p4d_offset(pgd, addr);
 507	pud_t *pud = pud_offset(p4d, addr);
 508	pmd_t *pmd = pmd_offset(pud, addr);
 509	struct address_space *mapping;
 510	pgoff_t index;
 511	static unsigned long resume;
 512	static unsigned long nr_shown;
 513	static unsigned long nr_unshown;
 514
 515	/*
 516	 * Allow a burst of 60 reports, then keep quiet for that minute;
 517	 * or allow a steady drip of one report per second.
 518	 */
 519	if (nr_shown == 60) {
 520		if (time_before(jiffies, resume)) {
 521			nr_unshown++;
 522			return;
 523		}
 524		if (nr_unshown) {
 525			pr_alert("BUG: Bad page map: %lu messages suppressed\n",
 526				 nr_unshown);
 527			nr_unshown = 0;
 528		}
 529		nr_shown = 0;
 530	}
 531	if (nr_shown++ == 0)
 532		resume = jiffies + 60 * HZ;
 533
 534	mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
 535	index = linear_page_index(vma, addr);
 536
 537	pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
 538		 current->comm,
 539		 (long long)pte_val(pte), (long long)pmd_val(*pmd));
 540	if (page)
 541		dump_page(page, "bad pte");
 542	pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
 543		 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
 544	pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
 545		 vma->vm_file,
 546		 vma->vm_ops ? vma->vm_ops->fault : NULL,
 547		 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
 548		 mapping ? mapping->a_ops->readpage : NULL);
 549	dump_stack();
 550	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 551}
 552
 553/*
 554 * vm_normal_page -- This function gets the "struct page" associated with a pte.
 555 *
 556 * "Special" mappings do not wish to be associated with a "struct page" (either
 557 * it doesn't exist, or it exists but they don't want to touch it). In this
 558 * case, NULL is returned here. "Normal" mappings do have a struct page.
 559 *
 560 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
 561 * pte bit, in which case this function is trivial. Secondly, an architecture
 562 * may not have a spare pte bit, which requires a more complicated scheme,
 563 * described below.
 564 *
 565 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
 566 * special mapping (even if there are underlying and valid "struct pages").
 567 * COWed pages of a VM_PFNMAP are always normal.
 568 *
 569 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
 570 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
 571 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
 572 * mapping will always honor the rule
 573 *
 574 *	pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 575 *
 576 * And for normal mappings this is false.
 577 *
 578 * This restricts such mappings to be a linear translation from virtual address
 579 * to pfn. To get around this restriction, we allow arbitrary mappings so long
 580 * as the vma is not a COW mapping; in that case, we know that all ptes are
 581 * special (because none can have been COWed).
 582 *
 583 *
 584 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
 585 *
 586 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
 587 * page" backing, however the difference is that _all_ pages with a struct
 588 * page (that is, those where pfn_valid is true) are refcounted and considered
 589 * normal pages by the VM. The disadvantage is that pages are refcounted
 590 * (which can be slower and simply not an option for some PFNMAP users). The
 591 * advantage is that we don't have to follow the strict linearity rule of
 592 * PFNMAP mappings in order to support COWable mappings.
 593 *
 594 */
 595struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 596			    pte_t pte)
 597{
 598	unsigned long pfn = pte_pfn(pte);
 599
 600	if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
 601		if (likely(!pte_special(pte)))
 602			goto check_pfn;
 603		if (vma->vm_ops && vma->vm_ops->find_special_page)
 604			return vma->vm_ops->find_special_page(vma, addr);
 605		if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
 606			return NULL;
 607		if (is_zero_pfn(pfn))
 608			return NULL;
 609		if (pte_devmap(pte))
 610			return NULL;
 611
 612		print_bad_pte(vma, addr, pte, NULL);
 613		return NULL;
 614	}
 615
 616	/* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
 617
 618	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 619		if (vma->vm_flags & VM_MIXEDMAP) {
 620			if (!pfn_valid(pfn))
 621				return NULL;
 622			goto out;
 623		} else {
 624			unsigned long off;
 625			off = (addr - vma->vm_start) >> PAGE_SHIFT;
 626			if (pfn == vma->vm_pgoff + off)
 627				return NULL;
 628			if (!is_cow_mapping(vma->vm_flags))
 629				return NULL;
 630		}
 631	}
 632
 633	if (is_zero_pfn(pfn))
 634		return NULL;
 635
 636check_pfn:
 637	if (unlikely(pfn > highest_memmap_pfn)) {
 638		print_bad_pte(vma, addr, pte, NULL);
 639		return NULL;
 640	}
 641
 642	/*
 643	 * NOTE! We still have PageReserved() pages in the page tables.
 644	 * eg. VDSO mappings can cause them to exist.
 645	 */
 646out:
 647	return pfn_to_page(pfn);
 648}
 649
 650#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 651struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
 652				pmd_t pmd)
 653{
 654	unsigned long pfn = pmd_pfn(pmd);
 655
 656	/*
 657	 * There is no pmd_special() but there may be special pmds, e.g.
 658	 * in a direct-access (dax) mapping, so let's just replicate the
 659	 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
 660	 */
 661	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 662		if (vma->vm_flags & VM_MIXEDMAP) {
 663			if (!pfn_valid(pfn))
 664				return NULL;
 665			goto out;
 666		} else {
 667			unsigned long off;
 668			off = (addr - vma->vm_start) >> PAGE_SHIFT;
 669			if (pfn == vma->vm_pgoff + off)
 670				return NULL;
 671			if (!is_cow_mapping(vma->vm_flags))
 672				return NULL;
 673		}
 674	}
 675
 676	if (pmd_devmap(pmd))
 677		return NULL;
 678	if (is_huge_zero_pmd(pmd))
 679		return NULL;
 680	if (unlikely(pfn > highest_memmap_pfn))
 681		return NULL;
 682
 683	/*
 684	 * NOTE! We still have PageReserved() pages in the page tables.
 685	 * eg. VDSO mappings can cause them to exist.
 686	 */
 687out:
 688	return pfn_to_page(pfn);
 689}
 690#endif
 691
 692/*
 693 * copy one vm_area from one task to the other. Assumes the page tables
 694 * already present in the new task to be cleared in the whole range
 695 * covered by this vma.
 696 */
 697
 698static unsigned long
 699copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 700		pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 701		unsigned long addr, int *rss)
 702{
 703	unsigned long vm_flags = vma->vm_flags;
 704	pte_t pte = *src_pte;
 705	struct page *page;
 706	swp_entry_t entry = pte_to_swp_entry(pte);
 707
 708	if (likely(!non_swap_entry(entry))) {
 709		if (swap_duplicate(entry) < 0)
 710			return entry.val;
 711
 712		/* make sure dst_mm is on swapoff's mmlist. */
 713		if (unlikely(list_empty(&dst_mm->mmlist))) {
 714			spin_lock(&mmlist_lock);
 715			if (list_empty(&dst_mm->mmlist))
 716				list_add(&dst_mm->mmlist,
 717						&src_mm->mmlist);
 718			spin_unlock(&mmlist_lock);
 719		}
 720		rss[MM_SWAPENTS]++;
 721	} else if (is_migration_entry(entry)) {
 722		page = migration_entry_to_page(entry);
 
 
 
 
 723
 724		rss[mm_counter(page)]++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 725
 726		if (is_write_migration_entry(entry) &&
 727				is_cow_mapping(vm_flags)) {
 728			/*
 729			 * COW mappings require pages in both
 730			 * parent and child to be set to read.
 
 
 
 
 
 731			 */
 732			make_migration_entry_read(&entry);
 733			pte = swp_entry_to_pte(entry);
 734			if (pte_swp_soft_dirty(*src_pte))
 735				pte = pte_swp_mksoft_dirty(pte);
 736			if (pte_swp_uffd_wp(*src_pte))
 737				pte = pte_swp_mkuffd_wp(pte);
 738			set_pte_at(src_mm, addr, src_pte, pte);
 739		}
 740	} else if (is_device_private_entry(entry)) {
 741		page = device_private_entry_to_page(entry);
 742
 743		/*
 744		 * Update rss count even for unaddressable pages, as
 745		 * they should treated just like normal pages in this
 746		 * respect.
 747		 *
 748		 * We will likely want to have some new rss counters
 749		 * for unaddressable pages, at some point. But for now
 750		 * keep things as they are.
 751		 */
 752		get_page(page);
 753		rss[mm_counter(page)]++;
 754		page_dup_rmap(page, false);
 755
 756		/*
 757		 * We do not preserve soft-dirty information, because so
 758		 * far, checkpoint/restore is the only feature that
 759		 * requires that. And checkpoint/restore does not work
 760		 * when a device driver is involved (you cannot easily
 761		 * save and restore device driver state).
 762		 */
 763		if (is_write_device_private_entry(entry) &&
 764		    is_cow_mapping(vm_flags)) {
 765			make_device_private_entry_read(&entry);
 766			pte = swp_entry_to_pte(entry);
 767			if (pte_swp_uffd_wp(*src_pte))
 768				pte = pte_swp_mkuffd_wp(pte);
 769			set_pte_at(src_mm, addr, src_pte, pte);
 770		}
 771	}
 772	set_pte_at(dst_mm, addr, dst_pte, pte);
 773	return 0;
 774}
 775
 776/*
 777 * Copy a present and normal page if necessary.
 778 *
 779 * NOTE! The usual case is that this doesn't need to do
 780 * anything, and can just return a positive value. That
 781 * will let the caller know that it can just increase
 782 * the page refcount and re-use the pte the traditional
 783 * way.
 784 *
 785 * But _if_ we need to copy it because it needs to be
 786 * pinned in the parent (and the child should get its own
 787 * copy rather than just a reference to the same page),
 788 * we'll do that here and return zero to let the caller
 789 * know we're done.
 790 *
 791 * And if we need a pre-allocated page but don't yet have
 792 * one, return a negative error to let the preallocation
 793 * code know so that it can do so outside the page table
 794 * lock.
 795 */
 796static inline int
 797copy_present_page(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 798		pte_t *dst_pte, pte_t *src_pte,
 799		struct vm_area_struct *vma, struct vm_area_struct *new,
 800		unsigned long addr, int *rss, struct page **prealloc,
 801		pte_t pte, struct page *page)
 802{
 803	struct page *new_page;
 804
 805	if (!is_cow_mapping(vma->vm_flags))
 806		return 1;
 807
 808	/*
 809	 * What we want to do is to check whether this page may
 810	 * have been pinned by the parent process.  If so,
 811	 * instead of wrprotect the pte on both sides, we copy
 812	 * the page immediately so that we'll always guarantee
 813	 * the pinned page won't be randomly replaced in the
 814	 * future.
 815	 *
 816	 * The page pinning checks are just "has this mm ever
 817	 * seen pinning", along with the (inexact) check of
 818	 * the page count. That might give false positives for
 819	 * for pinning, but it will work correctly.
 820	 */
 821	if (likely(!atomic_read(&src_mm->has_pinned)))
 822		return 1;
 823	if (likely(!page_maybe_dma_pinned(page)))
 824		return 1;
 825
 826	new_page = *prealloc;
 827	if (!new_page)
 828		return -EAGAIN;
 829
 830	/*
 831	 * We have a prealloc page, all good!  Take it
 832	 * over and copy the page & arm it.
 833	 */
 834	*prealloc = NULL;
 835	copy_user_highpage(new_page, page, addr, vma);
 836	__SetPageUptodate(new_page);
 837	page_add_new_anon_rmap(new_page, new, addr, false);
 838	lru_cache_add_inactive_or_unevictable(new_page, new);
 839	rss[mm_counter(new_page)]++;
 840
 841	/* All done, just insert the new page copy in the child */
 842	pte = mk_pte(new_page, new->vm_page_prot);
 843	pte = maybe_mkwrite(pte_mkdirty(pte), new);
 844	set_pte_at(dst_mm, addr, dst_pte, pte);
 845	return 0;
 846}
 847
 848/*
 849 * Copy one pte.  Returns 0 if succeeded, or -EAGAIN if one preallocated page
 850 * is required to copy this pte.
 851 */
 852static inline int
 853copy_present_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 854		pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 855		struct vm_area_struct *new,
 856		unsigned long addr, int *rss, struct page **prealloc)
 857{
 858	unsigned long vm_flags = vma->vm_flags;
 859	pte_t pte = *src_pte;
 860	struct page *page;
 861
 862	page = vm_normal_page(vma, addr, pte);
 863	if (page) {
 864		int retval;
 865
 866		retval = copy_present_page(dst_mm, src_mm,
 867			dst_pte, src_pte,
 868			vma, new,
 869			addr, rss, prealloc,
 870			pte, page);
 871		if (retval <= 0)
 872			return retval;
 873
 874		get_page(page);
 875		page_dup_rmap(page, false);
 876		rss[mm_counter(page)]++;
 877	}
 878
 879	/*
 880	 * If it's a COW mapping, write protect it both
 881	 * in the parent and the child
 882	 */
 883	if (is_cow_mapping(vm_flags) && pte_write(pte)) {
 884		ptep_set_wrprotect(src_mm, addr, src_pte);
 885		pte = pte_wrprotect(pte);
 886	}
 887
 888	/*
 889	 * If it's a shared mapping, mark it clean in
 890	 * the child
 891	 */
 892	if (vm_flags & VM_SHARED)
 893		pte = pte_mkclean(pte);
 894	pte = pte_mkold(pte);
 895
 896	/*
 897	 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
 898	 * does not have the VM_UFFD_WP, which means that the uffd
 899	 * fork event is not enabled.
 900	 */
 901	if (!(vm_flags & VM_UFFD_WP))
 902		pte = pte_clear_uffd_wp(pte);
 
 903
 
 904	set_pte_at(dst_mm, addr, dst_pte, pte);
 905	return 0;
 906}
 907
 908static inline struct page *
 909page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
 910		   unsigned long addr)
 911{
 912	struct page *new_page;
 913
 914	new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
 915	if (!new_page)
 916		return NULL;
 917
 918	if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) {
 919		put_page(new_page);
 920		return NULL;
 921	}
 922	cgroup_throttle_swaprate(new_page, GFP_KERNEL);
 923
 924	return new_page;
 925}
 926
 927static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 928		   pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 929		   struct vm_area_struct *new,
 930		   unsigned long addr, unsigned long end)
 931{
 932	pte_t *orig_src_pte, *orig_dst_pte;
 933	pte_t *src_pte, *dst_pte;
 934	spinlock_t *src_ptl, *dst_ptl;
 935	int progress, ret = 0;
 936	int rss[NR_MM_COUNTERS];
 937	swp_entry_t entry = (swp_entry_t){0};
 938	struct page *prealloc = NULL;
 939
 940again:
 941	progress = 0;
 942	init_rss_vec(rss);
 943
 944	dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
 945	if (!dst_pte) {
 946		ret = -ENOMEM;
 947		goto out;
 948	}
 949	src_pte = pte_offset_map(src_pmd, addr);
 950	src_ptl = pte_lockptr(src_mm, src_pmd);
 951	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 952	orig_src_pte = src_pte;
 953	orig_dst_pte = dst_pte;
 954	arch_enter_lazy_mmu_mode();
 955
 956	do {
 957		/*
 958		 * We are holding two locks at this point - either of them
 959		 * could generate latencies in another task on another CPU.
 960		 */
 961		if (progress >= 32) {
 962			progress = 0;
 963			if (need_resched() ||
 964			    spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
 965				break;
 966		}
 967		if (pte_none(*src_pte)) {
 968			progress++;
 969			continue;
 970		}
 971		if (unlikely(!pte_present(*src_pte))) {
 972			entry.val = copy_nonpresent_pte(dst_mm, src_mm,
 973							dst_pte, src_pte,
 974							vma, addr, rss);
 975			if (entry.val)
 976				break;
 977			progress += 8;
 978			continue;
 979		}
 980		/* copy_present_pte() will clear `*prealloc' if consumed */
 981		ret = copy_present_pte(dst_mm, src_mm, dst_pte, src_pte,
 982				       vma, new, addr, rss, &prealloc);
 983		/*
 984		 * If we need a pre-allocated page for this pte, drop the
 985		 * locks, allocate, and try again.
 986		 */
 987		if (unlikely(ret == -EAGAIN))
 988			break;
 989		if (unlikely(prealloc)) {
 990			/*
 991			 * pre-alloc page cannot be reused by next time so as
 992			 * to strictly follow mempolicy (e.g., alloc_page_vma()
 993			 * will allocate page according to address).  This
 994			 * could only happen if one pinned pte changed.
 995			 */
 996			put_page(prealloc);
 997			prealloc = NULL;
 998		}
 999		progress += 8;
1000	} while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1001
1002	arch_leave_lazy_mmu_mode();
1003	spin_unlock(src_ptl);
1004	pte_unmap(orig_src_pte);
1005	add_mm_rss_vec(dst_mm, rss);
1006	pte_unmap_unlock(orig_dst_pte, dst_ptl);
1007	cond_resched();
1008
1009	if (entry.val) {
1010		if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1011			ret = -ENOMEM;
1012			goto out;
1013		}
1014		entry.val = 0;
1015	} else if (ret) {
1016		WARN_ON_ONCE(ret != -EAGAIN);
1017		prealloc = page_copy_prealloc(src_mm, vma, addr);
1018		if (!prealloc)
1019			return -ENOMEM;
1020		/* We've captured and resolved the error. Reset, try again. */
1021		ret = 0;
1022	}
1023	if (addr != end)
1024		goto again;
1025out:
1026	if (unlikely(prealloc))
1027		put_page(prealloc);
1028	return ret;
1029}
1030
1031static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1032		pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1033		struct vm_area_struct *new,
1034		unsigned long addr, unsigned long end)
1035{
1036	pmd_t *src_pmd, *dst_pmd;
1037	unsigned long next;
1038
1039	dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1040	if (!dst_pmd)
1041		return -ENOMEM;
1042	src_pmd = pmd_offset(src_pud, addr);
1043	do {
1044		next = pmd_addr_end(addr, end);
1045		if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1046			|| pmd_devmap(*src_pmd)) {
1047			int err;
1048			VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1049			err = copy_huge_pmd(dst_mm, src_mm,
1050					    dst_pmd, src_pmd, addr, vma);
1051			if (err == -ENOMEM)
1052				return -ENOMEM;
1053			if (!err)
1054				continue;
1055			/* fall through */
1056		}
1057		if (pmd_none_or_clear_bad(src_pmd))
1058			continue;
1059		if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1060				   vma, new, addr, next))
1061			return -ENOMEM;
1062	} while (dst_pmd++, src_pmd++, addr = next, addr != end);
1063	return 0;
1064}
1065
1066static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1067		p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1068		struct vm_area_struct *new,
1069		unsigned long addr, unsigned long end)
1070{
1071	pud_t *src_pud, *dst_pud;
1072	unsigned long next;
1073
1074	dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1075	if (!dst_pud)
1076		return -ENOMEM;
1077	src_pud = pud_offset(src_p4d, addr);
1078	do {
1079		next = pud_addr_end(addr, end);
1080		if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1081			int err;
1082
1083			VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1084			err = copy_huge_pud(dst_mm, src_mm,
1085					    dst_pud, src_pud, addr, vma);
1086			if (err == -ENOMEM)
1087				return -ENOMEM;
1088			if (!err)
1089				continue;
1090			/* fall through */
1091		}
1092		if (pud_none_or_clear_bad(src_pud))
1093			continue;
1094		if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1095				   vma, new, addr, next))
1096			return -ENOMEM;
1097	} while (dst_pud++, src_pud++, addr = next, addr != end);
1098	return 0;
1099}
1100
1101static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1102		pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1103		struct vm_area_struct *new,
1104		unsigned long addr, unsigned long end)
1105{
1106	p4d_t *src_p4d, *dst_p4d;
1107	unsigned long next;
1108
1109	dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1110	if (!dst_p4d)
1111		return -ENOMEM;
1112	src_p4d = p4d_offset(src_pgd, addr);
1113	do {
1114		next = p4d_addr_end(addr, end);
1115		if (p4d_none_or_clear_bad(src_p4d))
1116			continue;
1117		if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1118				   vma, new, addr, next))
1119			return -ENOMEM;
1120	} while (dst_p4d++, src_p4d++, addr = next, addr != end);
1121	return 0;
1122}
1123
1124int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1125		    struct vm_area_struct *vma, struct vm_area_struct *new)
1126{
1127	pgd_t *src_pgd, *dst_pgd;
1128	unsigned long next;
1129	unsigned long addr = vma->vm_start;
1130	unsigned long end = vma->vm_end;
1131	struct mmu_notifier_range range;
1132	bool is_cow;
1133	int ret;
1134
1135	/*
1136	 * Don't copy ptes where a page fault will fill them correctly.
1137	 * Fork becomes much lighter when there are big shared or private
1138	 * readonly mappings. The tradeoff is that copy_page_range is more
1139	 * efficient than faulting.
1140	 */
1141	if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1142			!vma->anon_vma)
1143		return 0;
1144
1145	if (is_vm_hugetlb_page(vma))
1146		return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1147
1148	if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1149		/*
1150		 * We do not free on error cases below as remove_vma
1151		 * gets called on error from higher level routine
1152		 */
1153		ret = track_pfn_copy(vma);
1154		if (ret)
1155			return ret;
1156	}
1157
1158	/*
1159	 * We need to invalidate the secondary MMU mappings only when
1160	 * there could be a permission downgrade on the ptes of the
1161	 * parent mm. And a permission downgrade will only happen if
1162	 * is_cow_mapping() returns true.
1163	 */
1164	is_cow = is_cow_mapping(vma->vm_flags);
1165
1166	if (is_cow) {
1167		mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1168					0, vma, src_mm, addr, end);
1169		mmu_notifier_invalidate_range_start(&range);
1170	}
1171
1172	ret = 0;
1173	dst_pgd = pgd_offset(dst_mm, addr);
1174	src_pgd = pgd_offset(src_mm, addr);
1175	do {
1176		next = pgd_addr_end(addr, end);
1177		if (pgd_none_or_clear_bad(src_pgd))
1178			continue;
1179		if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1180					    vma, new, addr, next))) {
1181			ret = -ENOMEM;
1182			break;
1183		}
1184	} while (dst_pgd++, src_pgd++, addr = next, addr != end);
1185
1186	if (is_cow)
1187		mmu_notifier_invalidate_range_end(&range);
1188	return ret;
1189}
1190
1191static unsigned long zap_pte_range(struct mmu_gather *tlb,
1192				struct vm_area_struct *vma, pmd_t *pmd,
1193				unsigned long addr, unsigned long end,
1194				struct zap_details *details)
1195{
1196	struct mm_struct *mm = tlb->mm;
1197	int force_flush = 0;
1198	int rss[NR_MM_COUNTERS];
1199	spinlock_t *ptl;
1200	pte_t *start_pte;
1201	pte_t *pte;
1202	swp_entry_t entry;
1203
1204	tlb_change_page_size(tlb, PAGE_SIZE);
1205again:
1206	init_rss_vec(rss);
1207	start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1208	pte = start_pte;
1209	flush_tlb_batched_pending(mm);
1210	arch_enter_lazy_mmu_mode();
1211	do {
1212		pte_t ptent = *pte;
1213		if (pte_none(ptent))
1214			continue;
1215
1216		if (need_resched())
1217			break;
1218
1219		if (pte_present(ptent)) {
1220			struct page *page;
1221
1222			page = vm_normal_page(vma, addr, ptent);
1223			if (unlikely(details) && page) {
1224				/*
1225				 * unmap_shared_mapping_pages() wants to
1226				 * invalidate cache without truncating:
1227				 * unmap shared but keep private pages.
1228				 */
1229				if (details->check_mapping &&
1230				    details->check_mapping != page_rmapping(page))
1231					continue;
1232			}
1233			ptent = ptep_get_and_clear_full(mm, addr, pte,
1234							tlb->fullmm);
1235			tlb_remove_tlb_entry(tlb, pte, addr);
1236			if (unlikely(!page))
1237				continue;
1238
1239			if (!PageAnon(page)) {
1240				if (pte_dirty(ptent)) {
1241					force_flush = 1;
1242					set_page_dirty(page);
1243				}
1244				if (pte_young(ptent) &&
1245				    likely(!(vma->vm_flags & VM_SEQ_READ)))
1246					mark_page_accessed(page);
1247			}
1248			rss[mm_counter(page)]--;
1249			page_remove_rmap(page, false);
1250			if (unlikely(page_mapcount(page) < 0))
1251				print_bad_pte(vma, addr, ptent, page);
1252			if (unlikely(__tlb_remove_page(tlb, page))) {
1253				force_flush = 1;
1254				addr += PAGE_SIZE;
1255				break;
1256			}
1257			continue;
1258		}
1259
1260		entry = pte_to_swp_entry(ptent);
1261		if (is_device_private_entry(entry)) {
1262			struct page *page = device_private_entry_to_page(entry);
1263
1264			if (unlikely(details && details->check_mapping)) {
1265				/*
1266				 * unmap_shared_mapping_pages() wants to
1267				 * invalidate cache without truncating:
1268				 * unmap shared but keep private pages.
1269				 */
1270				if (details->check_mapping !=
1271				    page_rmapping(page))
1272					continue;
1273			}
1274
1275			pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1276			rss[mm_counter(page)]--;
1277			page_remove_rmap(page, false);
1278			put_page(page);
1279			continue;
1280		}
1281
1282		/* If details->check_mapping, we leave swap entries. */
1283		if (unlikely(details))
1284			continue;
1285
1286		if (!non_swap_entry(entry))
1287			rss[MM_SWAPENTS]--;
1288		else if (is_migration_entry(entry)) {
1289			struct page *page;
1290
1291			page = migration_entry_to_page(entry);
1292			rss[mm_counter(page)]--;
1293		}
1294		if (unlikely(!free_swap_and_cache(entry)))
1295			print_bad_pte(vma, addr, ptent, NULL);
1296		pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1297	} while (pte++, addr += PAGE_SIZE, addr != end);
1298
1299	add_mm_rss_vec(mm, rss);
1300	arch_leave_lazy_mmu_mode();
1301
1302	/* Do the actual TLB flush before dropping ptl */
1303	if (force_flush)
1304		tlb_flush_mmu_tlbonly(tlb);
1305	pte_unmap_unlock(start_pte, ptl);
1306
1307	/*
1308	 * If we forced a TLB flush (either due to running out of
1309	 * batch buffers or because we needed to flush dirty TLB
1310	 * entries before releasing the ptl), free the batched
1311	 * memory too. Restart if we didn't do everything.
1312	 */
1313	if (force_flush) {
1314		force_flush = 0;
1315		tlb_flush_mmu(tlb);
1316	}
1317
1318	if (addr != end) {
1319		cond_resched();
1320		goto again;
1321	}
1322
1323	return addr;
1324}
1325
1326static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1327				struct vm_area_struct *vma, pud_t *pud,
1328				unsigned long addr, unsigned long end,
1329				struct zap_details *details)
1330{
1331	pmd_t *pmd;
1332	unsigned long next;
1333
1334	pmd = pmd_offset(pud, addr);
1335	do {
1336		next = pmd_addr_end(addr, end);
1337		if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1338			if (next - addr != HPAGE_PMD_SIZE)
1339				__split_huge_pmd(vma, pmd, addr, false, NULL);
1340			else if (zap_huge_pmd(tlb, vma, pmd, addr))
1341				goto next;
1342			/* fall through */
1343		}
1344		/*
1345		 * Here there can be other concurrent MADV_DONTNEED or
1346		 * trans huge page faults running, and if the pmd is
1347		 * none or trans huge it can change under us. This is
1348		 * because MADV_DONTNEED holds the mmap_lock in read
1349		 * mode.
1350		 */
1351		if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1352			goto next;
1353		next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1354next:
1355		cond_resched();
1356	} while (pmd++, addr = next, addr != end);
1357
1358	return addr;
1359}
1360
1361static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1362				struct vm_area_struct *vma, p4d_t *p4d,
1363				unsigned long addr, unsigned long end,
1364				struct zap_details *details)
1365{
1366	pud_t *pud;
1367	unsigned long next;
1368
1369	pud = pud_offset(p4d, addr);
1370	do {
1371		next = pud_addr_end(addr, end);
1372		if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1373			if (next - addr != HPAGE_PUD_SIZE) {
1374				mmap_assert_locked(tlb->mm);
1375				split_huge_pud(vma, pud, addr);
1376			} else if (zap_huge_pud(tlb, vma, pud, addr))
1377				goto next;
1378			/* fall through */
1379		}
1380		if (pud_none_or_clear_bad(pud))
1381			continue;
1382		next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1383next:
1384		cond_resched();
1385	} while (pud++, addr = next, addr != end);
1386
1387	return addr;
1388}
1389
1390static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1391				struct vm_area_struct *vma, pgd_t *pgd,
1392				unsigned long addr, unsigned long end,
1393				struct zap_details *details)
1394{
1395	p4d_t *p4d;
1396	unsigned long next;
1397
1398	p4d = p4d_offset(pgd, addr);
1399	do {
1400		next = p4d_addr_end(addr, end);
1401		if (p4d_none_or_clear_bad(p4d))
1402			continue;
1403		next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1404	} while (p4d++, addr = next, addr != end);
1405
1406	return addr;
1407}
1408
1409void unmap_page_range(struct mmu_gather *tlb,
1410			     struct vm_area_struct *vma,
1411			     unsigned long addr, unsigned long end,
1412			     struct zap_details *details)
1413{
1414	pgd_t *pgd;
1415	unsigned long next;
1416
1417	BUG_ON(addr >= end);
1418	tlb_start_vma(tlb, vma);
1419	pgd = pgd_offset(vma->vm_mm, addr);
1420	do {
1421		next = pgd_addr_end(addr, end);
1422		if (pgd_none_or_clear_bad(pgd))
1423			continue;
1424		next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1425	} while (pgd++, addr = next, addr != end);
1426	tlb_end_vma(tlb, vma);
1427}
1428
1429
1430static void unmap_single_vma(struct mmu_gather *tlb,
1431		struct vm_area_struct *vma, unsigned long start_addr,
1432		unsigned long end_addr,
1433		struct zap_details *details)
1434{
1435	unsigned long start = max(vma->vm_start, start_addr);
1436	unsigned long end;
1437
1438	if (start >= vma->vm_end)
1439		return;
1440	end = min(vma->vm_end, end_addr);
1441	if (end <= vma->vm_start)
1442		return;
1443
1444	if (vma->vm_file)
1445		uprobe_munmap(vma, start, end);
1446
1447	if (unlikely(vma->vm_flags & VM_PFNMAP))
1448		untrack_pfn(vma, 0, 0);
1449
1450	if (start != end) {
1451		if (unlikely(is_vm_hugetlb_page(vma))) {
1452			/*
1453			 * It is undesirable to test vma->vm_file as it
1454			 * should be non-null for valid hugetlb area.
1455			 * However, vm_file will be NULL in the error
1456			 * cleanup path of mmap_region. When
1457			 * hugetlbfs ->mmap method fails,
1458			 * mmap_region() nullifies vma->vm_file
1459			 * before calling this function to clean up.
1460			 * Since no pte has actually been setup, it is
1461			 * safe to do nothing in this case.
1462			 */
1463			if (vma->vm_file) {
1464				i_mmap_lock_write(vma->vm_file->f_mapping);
1465				__unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1466				i_mmap_unlock_write(vma->vm_file->f_mapping);
1467			}
1468		} else
1469			unmap_page_range(tlb, vma, start, end, details);
1470	}
1471}
1472
1473/**
1474 * unmap_vmas - unmap a range of memory covered by a list of vma's
1475 * @tlb: address of the caller's struct mmu_gather
1476 * @vma: the starting vma
1477 * @start_addr: virtual address at which to start unmapping
1478 * @end_addr: virtual address at which to end unmapping
1479 *
1480 * Unmap all pages in the vma list.
1481 *
1482 * Only addresses between `start' and `end' will be unmapped.
1483 *
1484 * The VMA list must be sorted in ascending virtual address order.
1485 *
1486 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1487 * range after unmap_vmas() returns.  So the only responsibility here is to
1488 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1489 * drops the lock and schedules.
1490 */
1491void unmap_vmas(struct mmu_gather *tlb,
1492		struct vm_area_struct *vma, unsigned long start_addr,
1493		unsigned long end_addr)
1494{
1495	struct mmu_notifier_range range;
1496
1497	mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1498				start_addr, end_addr);
1499	mmu_notifier_invalidate_range_start(&range);
1500	for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1501		unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1502	mmu_notifier_invalidate_range_end(&range);
1503}
1504
1505/**
1506 * zap_page_range - remove user pages in a given range
1507 * @vma: vm_area_struct holding the applicable pages
1508 * @start: starting address of pages to zap
1509 * @size: number of bytes to zap
1510 *
1511 * Caller must protect the VMA list
1512 */
1513void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1514		unsigned long size)
1515{
1516	struct mmu_notifier_range range;
1517	struct mmu_gather tlb;
1518
1519	lru_add_drain();
1520	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1521				start, start + size);
1522	tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1523	update_hiwater_rss(vma->vm_mm);
1524	mmu_notifier_invalidate_range_start(&range);
1525	for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1526		unmap_single_vma(&tlb, vma, start, range.end, NULL);
1527	mmu_notifier_invalidate_range_end(&range);
1528	tlb_finish_mmu(&tlb, start, range.end);
1529}
1530
1531/**
1532 * zap_page_range_single - remove user pages in a given range
1533 * @vma: vm_area_struct holding the applicable pages
1534 * @address: starting address of pages to zap
1535 * @size: number of bytes to zap
1536 * @details: details of shared cache invalidation
1537 *
1538 * The range must fit into one VMA.
1539 */
1540static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1541		unsigned long size, struct zap_details *details)
1542{
1543	struct mmu_notifier_range range;
1544	struct mmu_gather tlb;
1545
1546	lru_add_drain();
1547	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1548				address, address + size);
1549	tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1550	update_hiwater_rss(vma->vm_mm);
1551	mmu_notifier_invalidate_range_start(&range);
1552	unmap_single_vma(&tlb, vma, address, range.end, details);
1553	mmu_notifier_invalidate_range_end(&range);
1554	tlb_finish_mmu(&tlb, address, range.end);
1555}
1556
1557/**
1558 * zap_vma_ptes - remove ptes mapping the vma
1559 * @vma: vm_area_struct holding ptes to be zapped
1560 * @address: starting address of pages to zap
1561 * @size: number of bytes to zap
1562 *
1563 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1564 *
1565 * The entire address range must be fully contained within the vma.
1566 *
1567 */
1568void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1569		unsigned long size)
1570{
1571	if (address < vma->vm_start || address + size > vma->vm_end ||
1572	    		!(vma->vm_flags & VM_PFNMAP))
1573		return;
1574
1575	zap_page_range_single(vma, address, size, NULL);
1576}
1577EXPORT_SYMBOL_GPL(zap_vma_ptes);
1578
1579static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
 
1580{
1581	pgd_t *pgd;
1582	p4d_t *p4d;
1583	pud_t *pud;
1584	pmd_t *pmd;
1585
1586	pgd = pgd_offset(mm, addr);
1587	p4d = p4d_alloc(mm, pgd, addr);
1588	if (!p4d)
1589		return NULL;
1590	pud = pud_alloc(mm, p4d, addr);
1591	if (!pud)
1592		return NULL;
1593	pmd = pmd_alloc(mm, pud, addr);
1594	if (!pmd)
1595		return NULL;
1596
1597	VM_BUG_ON(pmd_trans_huge(*pmd));
1598	return pmd;
1599}
1600
1601pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1602			spinlock_t **ptl)
1603{
1604	pmd_t *pmd = walk_to_pmd(mm, addr);
1605
1606	if (!pmd)
1607		return NULL;
1608	return pte_alloc_map_lock(mm, pmd, addr, ptl);
1609}
1610
1611static int validate_page_before_insert(struct page *page)
1612{
1613	if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1614		return -EINVAL;
1615	flush_dcache_page(page);
1616	return 0;
1617}
1618
1619static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1620			unsigned long addr, struct page *page, pgprot_t prot)
1621{
1622	if (!pte_none(*pte))
1623		return -EBUSY;
1624	/* Ok, finally just insert the thing.. */
1625	get_page(page);
1626	inc_mm_counter_fast(mm, mm_counter_file(page));
1627	page_add_file_rmap(page, false);
1628	set_pte_at(mm, addr, pte, mk_pte(page, prot));
1629	return 0;
1630}
1631
1632/*
1633 * This is the old fallback for page remapping.
1634 *
1635 * For historical reasons, it only allows reserved pages. Only
1636 * old drivers should use this, and they needed to mark their
1637 * pages reserved for the old functions anyway.
1638 */
1639static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1640			struct page *page, pgprot_t prot)
1641{
1642	struct mm_struct *mm = vma->vm_mm;
1643	int retval;
1644	pte_t *pte;
1645	spinlock_t *ptl;
1646
1647	retval = validate_page_before_insert(page);
1648	if (retval)
1649		goto out;
1650	retval = -ENOMEM;
 
1651	pte = get_locked_pte(mm, addr, &ptl);
1652	if (!pte)
1653		goto out;
1654	retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1655	pte_unmap_unlock(pte, ptl);
1656out:
1657	return retval;
1658}
1659
1660#ifdef pte_index
1661static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1662			unsigned long addr, struct page *page, pgprot_t prot)
1663{
1664	int err;
1665
1666	if (!page_count(page))
1667		return -EINVAL;
1668	err = validate_page_before_insert(page);
1669	if (err)
1670		return err;
1671	return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1672}
1673
1674/* insert_pages() amortizes the cost of spinlock operations
1675 * when inserting pages in a loop. Arch *must* define pte_index.
1676 */
1677static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1678			struct page **pages, unsigned long *num, pgprot_t prot)
1679{
1680	pmd_t *pmd = NULL;
1681	pte_t *start_pte, *pte;
1682	spinlock_t *pte_lock;
1683	struct mm_struct *const mm = vma->vm_mm;
1684	unsigned long curr_page_idx = 0;
1685	unsigned long remaining_pages_total = *num;
1686	unsigned long pages_to_write_in_pmd;
1687	int ret;
1688more:
1689	ret = -EFAULT;
1690	pmd = walk_to_pmd(mm, addr);
1691	if (!pmd)
1692		goto out;
1693
1694	pages_to_write_in_pmd = min_t(unsigned long,
1695		remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1696
1697	/* Allocate the PTE if necessary; takes PMD lock once only. */
1698	ret = -ENOMEM;
1699	if (pte_alloc(mm, pmd))
1700		goto out;
 
1701
1702	while (pages_to_write_in_pmd) {
1703		int pte_idx = 0;
1704		const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1705
1706		start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1707		for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1708			int err = insert_page_in_batch_locked(mm, pte,
1709				addr, pages[curr_page_idx], prot);
1710			if (unlikely(err)) {
1711				pte_unmap_unlock(start_pte, pte_lock);
1712				ret = err;
1713				remaining_pages_total -= pte_idx;
1714				goto out;
1715			}
1716			addr += PAGE_SIZE;
1717			++curr_page_idx;
1718		}
1719		pte_unmap_unlock(start_pte, pte_lock);
1720		pages_to_write_in_pmd -= batch_size;
1721		remaining_pages_total -= batch_size;
1722	}
1723	if (remaining_pages_total)
1724		goto more;
1725	ret = 0;
1726out:
1727	*num = remaining_pages_total;
1728	return ret;
1729}
1730#endif  /* ifdef pte_index */
1731
1732/**
1733 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1734 * @vma: user vma to map to
1735 * @addr: target start user address of these pages
1736 * @pages: source kernel pages
1737 * @num: in: number of pages to map. out: number of pages that were *not*
1738 * mapped. (0 means all pages were successfully mapped).
1739 *
1740 * Preferred over vm_insert_page() when inserting multiple pages.
1741 *
1742 * In case of error, we may have mapped a subset of the provided
1743 * pages. It is the caller's responsibility to account for this case.
1744 *
1745 * The same restrictions apply as in vm_insert_page().
1746 */
1747int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1748			struct page **pages, unsigned long *num)
1749{
1750#ifdef pte_index
1751	const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1752
1753	if (addr < vma->vm_start || end_addr >= vma->vm_end)
1754		return -EFAULT;
1755	if (!(vma->vm_flags & VM_MIXEDMAP)) {
1756		BUG_ON(mmap_read_trylock(vma->vm_mm));
1757		BUG_ON(vma->vm_flags & VM_PFNMAP);
1758		vma->vm_flags |= VM_MIXEDMAP;
1759	}
1760	/* Defer page refcount checking till we're about to map that page. */
1761	return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1762#else
1763	unsigned long idx = 0, pgcount = *num;
1764	int err = -EINVAL;
1765
1766	for (; idx < pgcount; ++idx) {
1767		err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1768		if (err)
1769			break;
1770	}
1771	*num = pgcount - idx;
1772	return err;
1773#endif  /* ifdef pte_index */
1774}
1775EXPORT_SYMBOL(vm_insert_pages);
1776
1777/**
1778 * vm_insert_page - insert single page into user vma
1779 * @vma: user vma to map to
1780 * @addr: target user address of this page
1781 * @page: source kernel page
1782 *
1783 * This allows drivers to insert individual pages they've allocated
1784 * into a user vma.
1785 *
1786 * The page has to be a nice clean _individual_ kernel allocation.
1787 * If you allocate a compound page, you need to have marked it as
1788 * such (__GFP_COMP), or manually just split the page up yourself
1789 * (see split_page()).
1790 *
1791 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1792 * took an arbitrary page protection parameter. This doesn't allow
1793 * that. Your vma protection will have to be set up correctly, which
1794 * means that if you want a shared writable mapping, you'd better
1795 * ask for a shared writable mapping!
1796 *
1797 * The page does not need to be reserved.
1798 *
1799 * Usually this function is called from f_op->mmap() handler
1800 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1801 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1802 * function from other places, for example from page-fault handler.
1803 *
1804 * Return: %0 on success, negative error code otherwise.
1805 */
1806int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1807			struct page *page)
1808{
1809	if (addr < vma->vm_start || addr >= vma->vm_end)
1810		return -EFAULT;
1811	if (!page_count(page))
1812		return -EINVAL;
1813	if (!(vma->vm_flags & VM_MIXEDMAP)) {
1814		BUG_ON(mmap_read_trylock(vma->vm_mm));
1815		BUG_ON(vma->vm_flags & VM_PFNMAP);
1816		vma->vm_flags |= VM_MIXEDMAP;
1817	}
1818	return insert_page(vma, addr, page, vma->vm_page_prot);
1819}
1820EXPORT_SYMBOL(vm_insert_page);
1821
1822/*
1823 * __vm_map_pages - maps range of kernel pages into user vma
1824 * @vma: user vma to map to
1825 * @pages: pointer to array of source kernel pages
1826 * @num: number of pages in page array
1827 * @offset: user's requested vm_pgoff
1828 *
1829 * This allows drivers to map range of kernel pages into a user vma.
1830 *
1831 * Return: 0 on success and error code otherwise.
1832 */
1833static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1834				unsigned long num, unsigned long offset)
1835{
1836	unsigned long count = vma_pages(vma);
1837	unsigned long uaddr = vma->vm_start;
1838	int ret, i;
1839
1840	/* Fail if the user requested offset is beyond the end of the object */
1841	if (offset >= num)
1842		return -ENXIO;
1843
1844	/* Fail if the user requested size exceeds available object size */
1845	if (count > num - offset)
1846		return -ENXIO;
1847
1848	for (i = 0; i < count; i++) {
1849		ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1850		if (ret < 0)
1851			return ret;
1852		uaddr += PAGE_SIZE;
1853	}
1854
1855	return 0;
1856}
1857
1858/**
1859 * vm_map_pages - maps range of kernel pages starts with non zero offset
1860 * @vma: user vma to map to
1861 * @pages: pointer to array of source kernel pages
1862 * @num: number of pages in page array
1863 *
1864 * Maps an object consisting of @num pages, catering for the user's
1865 * requested vm_pgoff
1866 *
1867 * If we fail to insert any page into the vma, the function will return
1868 * immediately leaving any previously inserted pages present.  Callers
1869 * from the mmap handler may immediately return the error as their caller
1870 * will destroy the vma, removing any successfully inserted pages. Other
1871 * callers should make their own arrangements for calling unmap_region().
1872 *
1873 * Context: Process context. Called by mmap handlers.
1874 * Return: 0 on success and error code otherwise.
1875 */
1876int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1877				unsigned long num)
1878{
1879	return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1880}
1881EXPORT_SYMBOL(vm_map_pages);
1882
1883/**
1884 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1885 * @vma: user vma to map to
1886 * @pages: pointer to array of source kernel pages
1887 * @num: number of pages in page array
1888 *
1889 * Similar to vm_map_pages(), except that it explicitly sets the offset
1890 * to 0. This function is intended for the drivers that did not consider
1891 * vm_pgoff.
1892 *
1893 * Context: Process context. Called by mmap handlers.
1894 * Return: 0 on success and error code otherwise.
1895 */
1896int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1897				unsigned long num)
1898{
1899	return __vm_map_pages(vma, pages, num, 0);
1900}
1901EXPORT_SYMBOL(vm_map_pages_zero);
1902
1903static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1904			pfn_t pfn, pgprot_t prot, bool mkwrite)
1905{
1906	struct mm_struct *mm = vma->vm_mm;
1907	pte_t *pte, entry;
1908	spinlock_t *ptl;
1909
1910	pte = get_locked_pte(mm, addr, &ptl);
1911	if (!pte)
1912		return VM_FAULT_OOM;
1913	if (!pte_none(*pte)) {
1914		if (mkwrite) {
1915			/*
1916			 * For read faults on private mappings the PFN passed
1917			 * in may not match the PFN we have mapped if the
1918			 * mapped PFN is a writeable COW page.  In the mkwrite
1919			 * case we are creating a writable PTE for a shared
1920			 * mapping and we expect the PFNs to match. If they
1921			 * don't match, we are likely racing with block
1922			 * allocation and mapping invalidation so just skip the
1923			 * update.
1924			 */
1925			if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1926				WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1927				goto out_unlock;
1928			}
1929			entry = pte_mkyoung(*pte);
1930			entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1931			if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1932				update_mmu_cache(vma, addr, pte);
1933		}
1934		goto out_unlock;
1935	}
1936
1937	/* Ok, finally just insert the thing.. */
1938	if (pfn_t_devmap(pfn))
1939		entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1940	else
1941		entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1942
1943	if (mkwrite) {
1944		entry = pte_mkyoung(entry);
1945		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1946	}
1947
1948	set_pte_at(mm, addr, pte, entry);
1949	update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1950
1951out_unlock:
1952	pte_unmap_unlock(pte, ptl);
1953	return VM_FAULT_NOPAGE;
1954}
1955
1956/**
1957 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1958 * @vma: user vma to map to
1959 * @addr: target user address of this page
1960 * @pfn: source kernel pfn
1961 * @pgprot: pgprot flags for the inserted page
1962 *
1963 * This is exactly like vmf_insert_pfn(), except that it allows drivers
1964 * to override pgprot on a per-page basis.
1965 *
1966 * This only makes sense for IO mappings, and it makes no sense for
1967 * COW mappings.  In general, using multiple vmas is preferable;
1968 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1969 * impractical.
1970 *
1971 * See vmf_insert_mixed_prot() for a discussion of the implication of using
1972 * a value of @pgprot different from that of @vma->vm_page_prot.
1973 *
1974 * Context: Process context.  May allocate using %GFP_KERNEL.
1975 * Return: vm_fault_t value.
1976 */
1977vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1978			unsigned long pfn, pgprot_t pgprot)
1979{
1980	/*
1981	 * Technically, architectures with pte_special can avoid all these
1982	 * restrictions (same for remap_pfn_range).  However we would like
1983	 * consistency in testing and feature parity among all, so we should
1984	 * try to keep these invariants in place for everybody.
1985	 */
1986	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1987	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1988						(VM_PFNMAP|VM_MIXEDMAP));
1989	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1990	BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1991
1992	if (addr < vma->vm_start || addr >= vma->vm_end)
1993		return VM_FAULT_SIGBUS;
1994
1995	if (!pfn_modify_allowed(pfn, pgprot))
1996		return VM_FAULT_SIGBUS;
1997
1998	track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1999
2000	return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2001			false);
2002}
2003EXPORT_SYMBOL(vmf_insert_pfn_prot);
2004
2005/**
2006 * vmf_insert_pfn - insert single pfn into user vma
2007 * @vma: user vma to map to
2008 * @addr: target user address of this page
2009 * @pfn: source kernel pfn
2010 *
2011 * Similar to vm_insert_page, this allows drivers to insert individual pages
2012 * they've allocated into a user vma. Same comments apply.
2013 *
2014 * This function should only be called from a vm_ops->fault handler, and
2015 * in that case the handler should return the result of this function.
2016 *
2017 * vma cannot be a COW mapping.
2018 *
2019 * As this is called only for pages that do not currently exist, we
2020 * do not need to flush old virtual caches or the TLB.
2021 *
2022 * Context: Process context.  May allocate using %GFP_KERNEL.
2023 * Return: vm_fault_t value.
2024 */
2025vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2026			unsigned long pfn)
2027{
2028	return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2029}
2030EXPORT_SYMBOL(vmf_insert_pfn);
2031
2032static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2033{
2034	/* these checks mirror the abort conditions in vm_normal_page */
2035	if (vma->vm_flags & VM_MIXEDMAP)
2036		return true;
2037	if (pfn_t_devmap(pfn))
2038		return true;
2039	if (pfn_t_special(pfn))
2040		return true;
2041	if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2042		return true;
2043	return false;
2044}
2045
2046static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2047		unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2048		bool mkwrite)
2049{
 
2050	int err;
2051
2052	BUG_ON(!vm_mixed_ok(vma, pfn));
2053
2054	if (addr < vma->vm_start || addr >= vma->vm_end)
2055		return VM_FAULT_SIGBUS;
2056
2057	track_pfn_insert(vma, &pgprot, pfn);
2058
2059	if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2060		return VM_FAULT_SIGBUS;
2061
2062	/*
2063	 * If we don't have pte special, then we have to use the pfn_valid()
2064	 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2065	 * refcount the page if pfn_valid is true (hence insert_page rather
2066	 * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2067	 * without pte special, it would there be refcounted as a normal page.
2068	 */
2069	if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2070	    !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2071		struct page *page;
2072
2073		/*
2074		 * At this point we are committed to insert_page()
2075		 * regardless of whether the caller specified flags that
2076		 * result in pfn_t_has_page() == false.
2077		 */
2078		page = pfn_to_page(pfn_t_to_pfn(pfn));
2079		err = insert_page(vma, addr, page, pgprot);
2080	} else {
2081		return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2082	}
2083
2084	if (err == -ENOMEM)
2085		return VM_FAULT_OOM;
2086	if (err < 0 && err != -EBUSY)
2087		return VM_FAULT_SIGBUS;
2088
2089	return VM_FAULT_NOPAGE;
2090}
2091
2092/**
2093 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2094 * @vma: user vma to map to
2095 * @addr: target user address of this page
2096 * @pfn: source kernel pfn
2097 * @pgprot: pgprot flags for the inserted page
2098 *
2099 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2100 * to override pgprot on a per-page basis.
2101 *
2102 * Typically this function should be used by drivers to set caching- and
2103 * encryption bits different than those of @vma->vm_page_prot, because
2104 * the caching- or encryption mode may not be known at mmap() time.
2105 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2106 * to set caching and encryption bits for those vmas (except for COW pages).
2107 * This is ensured by core vm only modifying these page table entries using
2108 * functions that don't touch caching- or encryption bits, using pte_modify()
2109 * if needed. (See for example mprotect()).
2110 * Also when new page-table entries are created, this is only done using the
2111 * fault() callback, and never using the value of vma->vm_page_prot,
2112 * except for page-table entries that point to anonymous pages as the result
2113 * of COW.
2114 *
2115 * Context: Process context.  May allocate using %GFP_KERNEL.
2116 * Return: vm_fault_t value.
2117 */
2118vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2119				 pfn_t pfn, pgprot_t pgprot)
2120{
2121	return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2122}
2123EXPORT_SYMBOL(vmf_insert_mixed_prot);
2124
2125vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2126		pfn_t pfn)
2127{
2128	return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2129}
2130EXPORT_SYMBOL(vmf_insert_mixed);
2131
2132/*
2133 *  If the insertion of PTE failed because someone else already added a
2134 *  different entry in the mean time, we treat that as success as we assume
2135 *  the same entry was actually inserted.
2136 */
2137vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2138		unsigned long addr, pfn_t pfn)
2139{
2140	return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2141}
2142EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2143
2144/*
2145 * maps a range of physical memory into the requested pages. the old
2146 * mappings are removed. any references to nonexistent pages results
2147 * in null mappings (currently treated as "copy-on-access")
2148 */
2149static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2150			unsigned long addr, unsigned long end,
2151			unsigned long pfn, pgprot_t prot)
2152{
2153	pte_t *pte;
2154	spinlock_t *ptl;
2155	int err = 0;
2156
2157	pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2158	if (!pte)
2159		return -ENOMEM;
2160	arch_enter_lazy_mmu_mode();
2161	do {
2162		BUG_ON(!pte_none(*pte));
2163		if (!pfn_modify_allowed(pfn, prot)) {
2164			err = -EACCES;
2165			break;
2166		}
2167		set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2168		pfn++;
2169	} while (pte++, addr += PAGE_SIZE, addr != end);
2170	arch_leave_lazy_mmu_mode();
2171	pte_unmap_unlock(pte - 1, ptl);
2172	return err;
2173}
2174
2175static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2176			unsigned long addr, unsigned long end,
2177			unsigned long pfn, pgprot_t prot)
2178{
2179	pmd_t *pmd;
2180	unsigned long next;
2181	int err;
2182
2183	pfn -= addr >> PAGE_SHIFT;
2184	pmd = pmd_alloc(mm, pud, addr);
2185	if (!pmd)
2186		return -ENOMEM;
2187	VM_BUG_ON(pmd_trans_huge(*pmd));
2188	do {
2189		next = pmd_addr_end(addr, end);
2190		err = remap_pte_range(mm, pmd, addr, next,
2191				pfn + (addr >> PAGE_SHIFT), prot);
2192		if (err)
2193			return err;
2194	} while (pmd++, addr = next, addr != end);
2195	return 0;
2196}
2197
2198static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2199			unsigned long addr, unsigned long end,
2200			unsigned long pfn, pgprot_t prot)
2201{
2202	pud_t *pud;
2203	unsigned long next;
2204	int err;
2205
2206	pfn -= addr >> PAGE_SHIFT;
2207	pud = pud_alloc(mm, p4d, addr);
2208	if (!pud)
2209		return -ENOMEM;
2210	do {
2211		next = pud_addr_end(addr, end);
2212		err = remap_pmd_range(mm, pud, addr, next,
2213				pfn + (addr >> PAGE_SHIFT), prot);
2214		if (err)
2215			return err;
2216	} while (pud++, addr = next, addr != end);
2217	return 0;
2218}
2219
2220static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2221			unsigned long addr, unsigned long end,
2222			unsigned long pfn, pgprot_t prot)
2223{
2224	p4d_t *p4d;
2225	unsigned long next;
2226	int err;
2227
2228	pfn -= addr >> PAGE_SHIFT;
2229	p4d = p4d_alloc(mm, pgd, addr);
2230	if (!p4d)
2231		return -ENOMEM;
2232	do {
2233		next = p4d_addr_end(addr, end);
2234		err = remap_pud_range(mm, p4d, addr, next,
2235				pfn + (addr >> PAGE_SHIFT), prot);
2236		if (err)
2237			return err;
2238	} while (p4d++, addr = next, addr != end);
2239	return 0;
2240}
2241
2242/**
2243 * remap_pfn_range - remap kernel memory to userspace
2244 * @vma: user vma to map to
2245 * @addr: target page aligned user address to start at
2246 * @pfn: page frame number of kernel physical memory address
2247 * @size: size of mapping area
2248 * @prot: page protection flags for this mapping
2249 *
2250 * Note: this is only safe if the mm semaphore is held when called.
2251 *
2252 * Return: %0 on success, negative error code otherwise.
2253 */
2254int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2255		    unsigned long pfn, unsigned long size, pgprot_t prot)
2256{
2257	pgd_t *pgd;
2258	unsigned long next;
2259	unsigned long end = addr + PAGE_ALIGN(size);
2260	struct mm_struct *mm = vma->vm_mm;
2261	unsigned long remap_pfn = pfn;
2262	int err;
2263
2264	if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2265		return -EINVAL;
2266
2267	/*
2268	 * Physically remapped pages are special. Tell the
2269	 * rest of the world about it:
2270	 *   VM_IO tells people not to look at these pages
2271	 *	(accesses can have side effects).
2272	 *   VM_PFNMAP tells the core MM that the base pages are just
2273	 *	raw PFN mappings, and do not have a "struct page" associated
2274	 *	with them.
2275	 *   VM_DONTEXPAND
2276	 *      Disable vma merging and expanding with mremap().
2277	 *   VM_DONTDUMP
2278	 *      Omit vma from core dump, even when VM_IO turned off.
2279	 *
2280	 * There's a horrible special case to handle copy-on-write
2281	 * behaviour that some programs depend on. We mark the "original"
2282	 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2283	 * See vm_normal_page() for details.
2284	 */
2285	if (is_cow_mapping(vma->vm_flags)) {
2286		if (addr != vma->vm_start || end != vma->vm_end)
2287			return -EINVAL;
2288		vma->vm_pgoff = pfn;
2289	}
2290
2291	err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2292	if (err)
2293		return -EINVAL;
2294
2295	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2296
2297	BUG_ON(addr >= end);
2298	pfn -= addr >> PAGE_SHIFT;
2299	pgd = pgd_offset(mm, addr);
2300	flush_cache_range(vma, addr, end);
2301	do {
2302		next = pgd_addr_end(addr, end);
2303		err = remap_p4d_range(mm, pgd, addr, next,
2304				pfn + (addr >> PAGE_SHIFT), prot);
2305		if (err)
2306			break;
2307	} while (pgd++, addr = next, addr != end);
2308
2309	if (err)
2310		untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2311
2312	return err;
2313}
2314EXPORT_SYMBOL(remap_pfn_range);
2315
2316/**
2317 * vm_iomap_memory - remap memory to userspace
2318 * @vma: user vma to map to
2319 * @start: start of the physical memory to be mapped
2320 * @len: size of area
2321 *
2322 * This is a simplified io_remap_pfn_range() for common driver use. The
2323 * driver just needs to give us the physical memory range to be mapped,
2324 * we'll figure out the rest from the vma information.
2325 *
2326 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2327 * whatever write-combining details or similar.
2328 *
2329 * Return: %0 on success, negative error code otherwise.
2330 */
2331int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2332{
2333	unsigned long vm_len, pfn, pages;
2334
2335	/* Check that the physical memory area passed in looks valid */
2336	if (start + len < start)
2337		return -EINVAL;
2338	/*
2339	 * You *really* shouldn't map things that aren't page-aligned,
2340	 * but we've historically allowed it because IO memory might
2341	 * just have smaller alignment.
2342	 */
2343	len += start & ~PAGE_MASK;
2344	pfn = start >> PAGE_SHIFT;
2345	pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2346	if (pfn + pages < pfn)
2347		return -EINVAL;
2348
2349	/* We start the mapping 'vm_pgoff' pages into the area */
2350	if (vma->vm_pgoff > pages)
2351		return -EINVAL;
2352	pfn += vma->vm_pgoff;
2353	pages -= vma->vm_pgoff;
2354
2355	/* Can we fit all of the mapping? */
2356	vm_len = vma->vm_end - vma->vm_start;
2357	if (vm_len >> PAGE_SHIFT > pages)
2358		return -EINVAL;
2359
2360	/* Ok, let it rip */
2361	return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2362}
2363EXPORT_SYMBOL(vm_iomap_memory);
2364
2365static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2366				     unsigned long addr, unsigned long end,
2367				     pte_fn_t fn, void *data, bool create,
2368				     pgtbl_mod_mask *mask)
2369{
2370	pte_t *pte;
2371	int err = 0;
2372	spinlock_t *ptl;
2373
2374	if (create) {
2375		pte = (mm == &init_mm) ?
2376			pte_alloc_kernel_track(pmd, addr, mask) :
2377			pte_alloc_map_lock(mm, pmd, addr, &ptl);
2378		if (!pte)
2379			return -ENOMEM;
2380	} else {
2381		pte = (mm == &init_mm) ?
2382			pte_offset_kernel(pmd, addr) :
2383			pte_offset_map_lock(mm, pmd, addr, &ptl);
2384	}
2385
2386	BUG_ON(pmd_huge(*pmd));
2387
2388	arch_enter_lazy_mmu_mode();
2389
2390	do {
2391		if (create || !pte_none(*pte)) {
2392			err = fn(pte++, addr, data);
2393			if (err)
2394				break;
2395		}
2396	} while (addr += PAGE_SIZE, addr != end);
2397	*mask |= PGTBL_PTE_MODIFIED;
2398
2399	arch_leave_lazy_mmu_mode();
2400
2401	if (mm != &init_mm)
2402		pte_unmap_unlock(pte-1, ptl);
2403	return err;
2404}
2405
2406static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2407				     unsigned long addr, unsigned long end,
2408				     pte_fn_t fn, void *data, bool create,
2409				     pgtbl_mod_mask *mask)
2410{
2411	pmd_t *pmd;
2412	unsigned long next;
2413	int err = 0;
2414
2415	BUG_ON(pud_huge(*pud));
2416
2417	if (create) {
2418		pmd = pmd_alloc_track(mm, pud, addr, mask);
2419		if (!pmd)
2420			return -ENOMEM;
2421	} else {
2422		pmd = pmd_offset(pud, addr);
2423	}
2424	do {
2425		next = pmd_addr_end(addr, end);
2426		if (create || !pmd_none_or_clear_bad(pmd)) {
2427			err = apply_to_pte_range(mm, pmd, addr, next, fn, data,
2428						 create, mask);
2429			if (err)
2430				break;
2431		}
2432	} while (pmd++, addr = next, addr != end);
2433	return err;
2434}
2435
2436static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2437				     unsigned long addr, unsigned long end,
2438				     pte_fn_t fn, void *data, bool create,
2439				     pgtbl_mod_mask *mask)
2440{
2441	pud_t *pud;
2442	unsigned long next;
2443	int err = 0;
2444
2445	if (create) {
2446		pud = pud_alloc_track(mm, p4d, addr, mask);
2447		if (!pud)
2448			return -ENOMEM;
2449	} else {
2450		pud = pud_offset(p4d, addr);
2451	}
2452	do {
2453		next = pud_addr_end(addr, end);
2454		if (create || !pud_none_or_clear_bad(pud)) {
2455			err = apply_to_pmd_range(mm, pud, addr, next, fn, data,
2456						 create, mask);
2457			if (err)
2458				break;
2459		}
2460	} while (pud++, addr = next, addr != end);
2461	return err;
2462}
2463
2464static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2465				     unsigned long addr, unsigned long end,
2466				     pte_fn_t fn, void *data, bool create,
2467				     pgtbl_mod_mask *mask)
2468{
2469	p4d_t *p4d;
2470	unsigned long next;
2471	int err = 0;
2472
2473	if (create) {
2474		p4d = p4d_alloc_track(mm, pgd, addr, mask);
2475		if (!p4d)
2476			return -ENOMEM;
2477	} else {
2478		p4d = p4d_offset(pgd, addr);
2479	}
2480	do {
2481		next = p4d_addr_end(addr, end);
2482		if (create || !p4d_none_or_clear_bad(p4d)) {
2483			err = apply_to_pud_range(mm, p4d, addr, next, fn, data,
2484						 create, mask);
2485			if (err)
2486				break;
2487		}
2488	} while (p4d++, addr = next, addr != end);
2489	return err;
2490}
2491
2492static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2493				 unsigned long size, pte_fn_t fn,
2494				 void *data, bool create)
 
 
 
2495{
2496	pgd_t *pgd;
2497	unsigned long start = addr, next;
2498	unsigned long end = addr + size;
2499	pgtbl_mod_mask mask = 0;
2500	int err = 0;
2501
2502	if (WARN_ON(addr >= end))
2503		return -EINVAL;
2504
2505	pgd = pgd_offset(mm, addr);
2506	do {
2507		next = pgd_addr_end(addr, end);
2508		if (!create && pgd_none_or_clear_bad(pgd))
2509			continue;
2510		err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create, &mask);
2511		if (err)
2512			break;
2513	} while (pgd++, addr = next, addr != end);
2514
2515	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2516		arch_sync_kernel_mappings(start, start + size);
2517
2518	return err;
2519}
2520
2521/*
2522 * Scan a region of virtual memory, filling in page tables as necessary
2523 * and calling a provided function on each leaf page table.
2524 */
2525int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2526			unsigned long size, pte_fn_t fn, void *data)
2527{
2528	return __apply_to_page_range(mm, addr, size, fn, data, true);
2529}
2530EXPORT_SYMBOL_GPL(apply_to_page_range);
2531
2532/*
2533 * Scan a region of virtual memory, calling a provided function on
2534 * each leaf page table where it exists.
2535 *
2536 * Unlike apply_to_page_range, this does _not_ fill in page tables
2537 * where they are absent.
2538 */
2539int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2540				 unsigned long size, pte_fn_t fn, void *data)
2541{
2542	return __apply_to_page_range(mm, addr, size, fn, data, false);
2543}
2544EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2545
2546/*
2547 * handle_pte_fault chooses page fault handler according to an entry which was
2548 * read non-atomically.  Before making any commitment, on those architectures
2549 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2550 * parts, do_swap_page must check under lock before unmapping the pte and
2551 * proceeding (but do_wp_page is only called after already making such a check;
2552 * and do_anonymous_page can safely check later on).
2553 */
2554static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2555				pte_t *page_table, pte_t orig_pte)
2556{
2557	int same = 1;
2558#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2559	if (sizeof(pte_t) > sizeof(unsigned long)) {
2560		spinlock_t *ptl = pte_lockptr(mm, pmd);
2561		spin_lock(ptl);
2562		same = pte_same(*page_table, orig_pte);
2563		spin_unlock(ptl);
2564	}
2565#endif
2566	pte_unmap(page_table);
2567	return same;
2568}
2569
2570static inline bool cow_user_page(struct page *dst, struct page *src,
2571				 struct vm_fault *vmf)
2572{
2573	bool ret;
2574	void *kaddr;
2575	void __user *uaddr;
2576	bool locked = false;
2577	struct vm_area_struct *vma = vmf->vma;
2578	struct mm_struct *mm = vma->vm_mm;
2579	unsigned long addr = vmf->address;
2580
2581	if (likely(src)) {
2582		copy_user_highpage(dst, src, addr, vma);
2583		return true;
2584	}
2585
2586	/*
2587	 * If the source page was a PFN mapping, we don't have
2588	 * a "struct page" for it. We do a best-effort copy by
2589	 * just copying from the original user address. If that
2590	 * fails, we just zero-fill it. Live with it.
2591	 */
2592	kaddr = kmap_atomic(dst);
2593	uaddr = (void __user *)(addr & PAGE_MASK);
2594
2595	/*
2596	 * On architectures with software "accessed" bits, we would
2597	 * take a double page fault, so mark it accessed here.
2598	 */
2599	if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2600		pte_t entry;
2601
2602		vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2603		locked = true;
2604		if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2605			/*
2606			 * Other thread has already handled the fault
2607			 * and update local tlb only
2608			 */
2609			update_mmu_tlb(vma, addr, vmf->pte);
2610			ret = false;
2611			goto pte_unlock;
2612		}
2613
2614		entry = pte_mkyoung(vmf->orig_pte);
2615		if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2616			update_mmu_cache(vma, addr, vmf->pte);
2617	}
2618
2619	/*
2620	 * This really shouldn't fail, because the page is there
2621	 * in the page tables. But it might just be unreadable,
2622	 * in which case we just give up and fill the result with
2623	 * zeroes.
2624	 */
2625	if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2626		if (locked)
2627			goto warn;
2628
2629		/* Re-validate under PTL if the page is still mapped */
2630		vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2631		locked = true;
2632		if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2633			/* The PTE changed under us, update local tlb */
2634			update_mmu_tlb(vma, addr, vmf->pte);
2635			ret = false;
2636			goto pte_unlock;
2637		}
2638
2639		/*
2640		 * The same page can be mapped back since last copy attempt.
2641		 * Try to copy again under PTL.
 
 
2642		 */
2643		if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2644			/*
2645			 * Give a warn in case there can be some obscure
2646			 * use-case
2647			 */
2648warn:
2649			WARN_ON_ONCE(1);
2650			clear_page(kaddr);
2651		}
2652	}
2653
2654	ret = true;
2655
2656pte_unlock:
2657	if (locked)
2658		pte_unmap_unlock(vmf->pte, vmf->ptl);
2659	kunmap_atomic(kaddr);
2660	flush_dcache_page(dst);
2661
2662	return ret;
2663}
2664
2665static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2666{
2667	struct file *vm_file = vma->vm_file;
2668
2669	if (vm_file)
2670		return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2671
2672	/*
2673	 * Special mappings (e.g. VDSO) do not have any file so fake
2674	 * a default GFP_KERNEL for them.
2675	 */
2676	return GFP_KERNEL;
2677}
2678
2679/*
2680 * Notify the address space that the page is about to become writable so that
2681 * it can prohibit this or wait for the page to get into an appropriate state.
2682 *
2683 * We do this without the lock held, so that it can sleep if it needs to.
2684 */
2685static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2686{
2687	vm_fault_t ret;
2688	struct page *page = vmf->page;
2689	unsigned int old_flags = vmf->flags;
2690
2691	vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2692
2693	if (vmf->vma->vm_file &&
2694	    IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2695		return VM_FAULT_SIGBUS;
2696
2697	ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2698	/* Restore original flags so that caller is not surprised */
2699	vmf->flags = old_flags;
2700	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2701		return ret;
2702	if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2703		lock_page(page);
2704		if (!page->mapping) {
2705			unlock_page(page);
2706			return 0; /* retry */
2707		}
2708		ret |= VM_FAULT_LOCKED;
2709	} else
2710		VM_BUG_ON_PAGE(!PageLocked(page), page);
2711	return ret;
2712}
2713
2714/*
2715 * Handle dirtying of a page in shared file mapping on a write fault.
2716 *
2717 * The function expects the page to be locked and unlocks it.
2718 */
2719static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
 
2720{
2721	struct vm_area_struct *vma = vmf->vma;
2722	struct address_space *mapping;
2723	struct page *page = vmf->page;
2724	bool dirtied;
2725	bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2726
2727	dirtied = set_page_dirty(page);
2728	VM_BUG_ON_PAGE(PageAnon(page), page);
2729	/*
2730	 * Take a local copy of the address_space - page.mapping may be zeroed
2731	 * by truncate after unlock_page().   The address_space itself remains
2732	 * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2733	 * release semantics to prevent the compiler from undoing this copying.
2734	 */
2735	mapping = page_rmapping(page);
2736	unlock_page(page);
2737
2738	if (!page_mkwrite)
2739		file_update_time(vma->vm_file);
2740
2741	/*
2742	 * Throttle page dirtying rate down to writeback speed.
2743	 *
2744	 * mapping may be NULL here because some device drivers do not
2745	 * set page.mapping but still dirty their pages
2746	 *
2747	 * Drop the mmap_lock before waiting on IO, if we can. The file
2748	 * is pinning the mapping, as per above.
2749	 */
2750	if ((dirtied || page_mkwrite) && mapping) {
2751		struct file *fpin;
2752
2753		fpin = maybe_unlock_mmap_for_io(vmf, NULL);
 
2754		balance_dirty_pages_ratelimited(mapping);
2755		if (fpin) {
2756			fput(fpin);
2757			return VM_FAULT_RETRY;
2758		}
2759	}
2760
2761	return 0;
 
2762}
2763
2764/*
2765 * Handle write page faults for pages that can be reused in the current vma
2766 *
2767 * This can happen either due to the mapping being with the VM_SHARED flag,
2768 * or due to us being the last reference standing to the page. In either
2769 * case, all we need to do here is to mark the page as writable and update
2770 * any related book-keeping.
2771 */
2772static inline void wp_page_reuse(struct vm_fault *vmf)
2773	__releases(vmf->ptl)
2774{
2775	struct vm_area_struct *vma = vmf->vma;
2776	struct page *page = vmf->page;
2777	pte_t entry;
2778	/*
2779	 * Clear the pages cpupid information as the existing
2780	 * information potentially belongs to a now completely
2781	 * unrelated process.
2782	 */
2783	if (page)
2784		page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2785
2786	flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2787	entry = pte_mkyoung(vmf->orig_pte);
2788	entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2789	if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2790		update_mmu_cache(vma, vmf->address, vmf->pte);
2791	pte_unmap_unlock(vmf->pte, vmf->ptl);
2792	count_vm_event(PGREUSE);
2793}
2794
2795/*
2796 * Handle the case of a page which we actually need to copy to a new page.
2797 *
2798 * Called with mmap_lock locked and the old page referenced, but
2799 * without the ptl held.
2800 *
2801 * High level logic flow:
2802 *
2803 * - Allocate a page, copy the content of the old page to the new one.
2804 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2805 * - Take the PTL. If the pte changed, bail out and release the allocated page
2806 * - If the pte is still the way we remember it, update the page table and all
2807 *   relevant references. This includes dropping the reference the page-table
2808 *   held to the old page, as well as updating the rmap.
2809 * - In any case, unlock the PTL and drop the reference we took to the old page.
2810 */
2811static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2812{
2813	struct vm_area_struct *vma = vmf->vma;
2814	struct mm_struct *mm = vma->vm_mm;
2815	struct page *old_page = vmf->page;
2816	struct page *new_page = NULL;
2817	pte_t entry;
2818	int page_copied = 0;
 
2819	struct mmu_notifier_range range;
2820
2821	if (unlikely(anon_vma_prepare(vma)))
2822		goto oom;
2823
2824	if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2825		new_page = alloc_zeroed_user_highpage_movable(vma,
2826							      vmf->address);
2827		if (!new_page)
2828			goto oom;
2829	} else {
2830		new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2831				vmf->address);
2832		if (!new_page)
2833			goto oom;
2834
2835		if (!cow_user_page(new_page, old_page, vmf)) {
2836			/*
2837			 * COW failed, if the fault was solved by other,
2838			 * it's fine. If not, userspace would re-fault on
2839			 * the same address and we will handle the fault
2840			 * from the second attempt.
2841			 */
2842			put_page(new_page);
2843			if (old_page)
2844				put_page(old_page);
2845			return 0;
2846		}
2847	}
2848
2849	if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
2850		goto oom_free_new;
2851	cgroup_throttle_swaprate(new_page, GFP_KERNEL);
2852
2853	__SetPageUptodate(new_page);
2854
2855	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2856				vmf->address & PAGE_MASK,
2857				(vmf->address & PAGE_MASK) + PAGE_SIZE);
2858	mmu_notifier_invalidate_range_start(&range);
2859
2860	/*
2861	 * Re-check the pte - we dropped the lock
2862	 */
2863	vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2864	if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2865		if (old_page) {
2866			if (!PageAnon(old_page)) {
2867				dec_mm_counter_fast(mm,
2868						mm_counter_file(old_page));
2869				inc_mm_counter_fast(mm, MM_ANONPAGES);
2870			}
2871		} else {
2872			inc_mm_counter_fast(mm, MM_ANONPAGES);
2873		}
2874		flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2875		entry = mk_pte(new_page, vma->vm_page_prot);
2876		entry = pte_sw_mkyoung(entry);
2877		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2878		/*
2879		 * Clear the pte entry and flush it first, before updating the
2880		 * pte with the new entry. This will avoid a race condition
2881		 * seen in the presence of one thread doing SMC and another
2882		 * thread doing COW.
2883		 */
2884		ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2885		page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2886		lru_cache_add_inactive_or_unevictable(new_page, vma);
 
2887		/*
2888		 * We call the notify macro here because, when using secondary
2889		 * mmu page tables (such as kvm shadow page tables), we want the
2890		 * new page to be mapped directly into the secondary page table.
2891		 */
2892		set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2893		update_mmu_cache(vma, vmf->address, vmf->pte);
2894		if (old_page) {
2895			/*
2896			 * Only after switching the pte to the new page may
2897			 * we remove the mapcount here. Otherwise another
2898			 * process may come and find the rmap count decremented
2899			 * before the pte is switched to the new page, and
2900			 * "reuse" the old page writing into it while our pte
2901			 * here still points into it and can be read by other
2902			 * threads.
2903			 *
2904			 * The critical issue is to order this
2905			 * page_remove_rmap with the ptp_clear_flush above.
2906			 * Those stores are ordered by (if nothing else,)
2907			 * the barrier present in the atomic_add_negative
2908			 * in page_remove_rmap.
2909			 *
2910			 * Then the TLB flush in ptep_clear_flush ensures that
2911			 * no process can access the old page before the
2912			 * decremented mapcount is visible. And the old page
2913			 * cannot be reused until after the decremented
2914			 * mapcount is visible. So transitively, TLBs to
2915			 * old page will be flushed before it can be reused.
2916			 */
2917			page_remove_rmap(old_page, false);
2918		}
2919
2920		/* Free the old page.. */
2921		new_page = old_page;
2922		page_copied = 1;
2923	} else {
2924		update_mmu_tlb(vma, vmf->address, vmf->pte);
2925	}
2926
2927	if (new_page)
2928		put_page(new_page);
2929
2930	pte_unmap_unlock(vmf->pte, vmf->ptl);
2931	/*
2932	 * No need to double call mmu_notifier->invalidate_range() callback as
2933	 * the above ptep_clear_flush_notify() did already call it.
2934	 */
2935	mmu_notifier_invalidate_range_only_end(&range);
2936	if (old_page) {
2937		/*
2938		 * Don't let another task, with possibly unlocked vma,
2939		 * keep the mlocked page.
2940		 */
2941		if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2942			lock_page(old_page);	/* LRU manipulation */
2943			if (PageMlocked(old_page))
2944				munlock_vma_page(old_page);
2945			unlock_page(old_page);
2946		}
2947		put_page(old_page);
2948	}
2949	return page_copied ? VM_FAULT_WRITE : 0;
2950oom_free_new:
2951	put_page(new_page);
2952oom:
2953	if (old_page)
2954		put_page(old_page);
2955	return VM_FAULT_OOM;
2956}
2957
2958/**
2959 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2960 *			  writeable once the page is prepared
2961 *
2962 * @vmf: structure describing the fault
2963 *
2964 * This function handles all that is needed to finish a write page fault in a
2965 * shared mapping due to PTE being read-only once the mapped page is prepared.
2966 * It handles locking of PTE and modifying it.
2967 *
2968 * The function expects the page to be locked or other protection against
2969 * concurrent faults / writeback (such as DAX radix tree locks).
2970 *
2971 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2972 * we acquired PTE lock.
2973 */
2974vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2975{
2976	WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2977	vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2978				       &vmf->ptl);
2979	/*
2980	 * We might have raced with another page fault while we released the
2981	 * pte_offset_map_lock.
2982	 */
2983	if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2984		update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
2985		pte_unmap_unlock(vmf->pte, vmf->ptl);
2986		return VM_FAULT_NOPAGE;
2987	}
2988	wp_page_reuse(vmf);
2989	return 0;
2990}
2991
2992/*
2993 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2994 * mapping
2995 */
2996static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2997{
2998	struct vm_area_struct *vma = vmf->vma;
2999
3000	if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3001		vm_fault_t ret;
3002
3003		pte_unmap_unlock(vmf->pte, vmf->ptl);
3004		vmf->flags |= FAULT_FLAG_MKWRITE;
3005		ret = vma->vm_ops->pfn_mkwrite(vmf);
3006		if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3007			return ret;
3008		return finish_mkwrite_fault(vmf);
3009	}
3010	wp_page_reuse(vmf);
3011	return VM_FAULT_WRITE;
3012}
3013
3014static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3015	__releases(vmf->ptl)
3016{
3017	struct vm_area_struct *vma = vmf->vma;
3018	vm_fault_t ret = VM_FAULT_WRITE;
3019
3020	get_page(vmf->page);
3021
3022	if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3023		vm_fault_t tmp;
3024
3025		pte_unmap_unlock(vmf->pte, vmf->ptl);
3026		tmp = do_page_mkwrite(vmf);
3027		if (unlikely(!tmp || (tmp &
3028				      (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3029			put_page(vmf->page);
3030			return tmp;
3031		}
3032		tmp = finish_mkwrite_fault(vmf);
3033		if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3034			unlock_page(vmf->page);
3035			put_page(vmf->page);
3036			return tmp;
3037		}
3038	} else {
3039		wp_page_reuse(vmf);
3040		lock_page(vmf->page);
3041	}
3042	ret |= fault_dirty_shared_page(vmf);
3043	put_page(vmf->page);
3044
3045	return ret;
3046}
3047
3048/*
3049 * This routine handles present pages, when users try to write
3050 * to a shared page. It is done by copying the page to a new address
3051 * and decrementing the shared-page counter for the old page.
3052 *
3053 * Note that this routine assumes that the protection checks have been
3054 * done by the caller (the low-level page fault routine in most cases).
3055 * Thus we can safely just mark it writable once we've done any necessary
3056 * COW.
3057 *
3058 * We also mark the page dirty at this point even though the page will
3059 * change only once the write actually happens. This avoids a few races,
3060 * and potentially makes it more efficient.
3061 *
3062 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3063 * but allow concurrent faults), with pte both mapped and locked.
3064 * We return with mmap_lock still held, but pte unmapped and unlocked.
3065 */
3066static vm_fault_t do_wp_page(struct vm_fault *vmf)
3067	__releases(vmf->ptl)
3068{
3069	struct vm_area_struct *vma = vmf->vma;
3070
3071	if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3072		pte_unmap_unlock(vmf->pte, vmf->ptl);
3073		return handle_userfault(vmf, VM_UFFD_WP);
3074	}
3075
3076	vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3077	if (!vmf->page) {
3078		/*
3079		 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3080		 * VM_PFNMAP VMA.
3081		 *
3082		 * We should not cow pages in a shared writeable mapping.
3083		 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3084		 */
3085		if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3086				     (VM_WRITE|VM_SHARED))
3087			return wp_pfn_shared(vmf);
3088
3089		pte_unmap_unlock(vmf->pte, vmf->ptl);
3090		return wp_page_copy(vmf);
3091	}
3092
3093	/*
3094	 * Take out anonymous pages first, anonymous shared vmas are
3095	 * not dirty accountable.
3096	 */
3097	if (PageAnon(vmf->page)) {
3098		struct page *page = vmf->page;
3099
3100		/* PageKsm() doesn't necessarily raise the page refcount */
3101		if (PageKsm(page) || page_count(page) != 1)
3102			goto copy;
3103		if (!trylock_page(page))
3104			goto copy;
3105		if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3106			unlock_page(page);
3107			goto copy;
 
 
 
 
 
 
 
 
 
 
 
 
 
3108		}
3109		/*
3110		 * Ok, we've got the only map reference, and the only
3111		 * page count reference, and the page is locked,
3112		 * it's dark out, and we're wearing sunglasses. Hit it.
3113		 */
3114		unlock_page(page);
3115		wp_page_reuse(vmf);
3116		return VM_FAULT_WRITE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3117	} else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3118					(VM_WRITE|VM_SHARED))) {
3119		return wp_page_shared(vmf);
3120	}
3121copy:
3122	/*
3123	 * Ok, we need to copy. Oh, well..
3124	 */
3125	get_page(vmf->page);
3126
3127	pte_unmap_unlock(vmf->pte, vmf->ptl);
3128	return wp_page_copy(vmf);
3129}
3130
3131static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3132		unsigned long start_addr, unsigned long end_addr,
3133		struct zap_details *details)
3134{
3135	zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3136}
3137
3138static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3139					    struct zap_details *details)
3140{
3141	struct vm_area_struct *vma;
3142	pgoff_t vba, vea, zba, zea;
3143
3144	vma_interval_tree_foreach(vma, root,
3145			details->first_index, details->last_index) {
3146
3147		vba = vma->vm_pgoff;
3148		vea = vba + vma_pages(vma) - 1;
3149		zba = details->first_index;
3150		if (zba < vba)
3151			zba = vba;
3152		zea = details->last_index;
3153		if (zea > vea)
3154			zea = vea;
3155
3156		unmap_mapping_range_vma(vma,
3157			((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3158			((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3159				details);
3160	}
3161}
3162
3163/**
3164 * unmap_mapping_pages() - Unmap pages from processes.
3165 * @mapping: The address space containing pages to be unmapped.
3166 * @start: Index of first page to be unmapped.
3167 * @nr: Number of pages to be unmapped.  0 to unmap to end of file.
3168 * @even_cows: Whether to unmap even private COWed pages.
3169 *
3170 * Unmap the pages in this address space from any userspace process which
3171 * has them mmaped.  Generally, you want to remove COWed pages as well when
3172 * a file is being truncated, but not when invalidating pages from the page
3173 * cache.
3174 */
3175void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3176		pgoff_t nr, bool even_cows)
3177{
3178	struct zap_details details = { };
3179
3180	details.check_mapping = even_cows ? NULL : mapping;
3181	details.first_index = start;
3182	details.last_index = start + nr - 1;
3183	if (details.last_index < details.first_index)
3184		details.last_index = ULONG_MAX;
3185
3186	i_mmap_lock_write(mapping);
3187	if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3188		unmap_mapping_range_tree(&mapping->i_mmap, &details);
3189	i_mmap_unlock_write(mapping);
3190}
3191
3192/**
3193 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3194 * address_space corresponding to the specified byte range in the underlying
3195 * file.
3196 *
3197 * @mapping: the address space containing mmaps to be unmapped.
3198 * @holebegin: byte in first page to unmap, relative to the start of
3199 * the underlying file.  This will be rounded down to a PAGE_SIZE
3200 * boundary.  Note that this is different from truncate_pagecache(), which
3201 * must keep the partial page.  In contrast, we must get rid of
3202 * partial pages.
3203 * @holelen: size of prospective hole in bytes.  This will be rounded
3204 * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
3205 * end of the file.
3206 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3207 * but 0 when invalidating pagecache, don't throw away private data.
3208 */
3209void unmap_mapping_range(struct address_space *mapping,
3210		loff_t const holebegin, loff_t const holelen, int even_cows)
3211{
3212	pgoff_t hba = holebegin >> PAGE_SHIFT;
3213	pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3214
3215	/* Check for overflow. */
3216	if (sizeof(holelen) > sizeof(hlen)) {
3217		long long holeend =
3218			(holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3219		if (holeend & ~(long long)ULONG_MAX)
3220			hlen = ULONG_MAX - hba + 1;
3221	}
3222
3223	unmap_mapping_pages(mapping, hba, hlen, even_cows);
3224}
3225EXPORT_SYMBOL(unmap_mapping_range);
3226
3227/*
3228 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3229 * but allow concurrent faults), and pte mapped but not yet locked.
3230 * We return with pte unmapped and unlocked.
3231 *
3232 * We return with the mmap_lock locked or unlocked in the same cases
3233 * as does filemap_fault().
3234 */
3235vm_fault_t do_swap_page(struct vm_fault *vmf)
3236{
3237	struct vm_area_struct *vma = vmf->vma;
3238	struct page *page = NULL, *swapcache;
 
3239	swp_entry_t entry;
3240	pte_t pte;
3241	int locked;
3242	int exclusive = 0;
3243	vm_fault_t ret = 0;
3244	void *shadow = NULL;
3245
3246	if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3247		goto out;
3248
3249	entry = pte_to_swp_entry(vmf->orig_pte);
3250	if (unlikely(non_swap_entry(entry))) {
3251		if (is_migration_entry(entry)) {
3252			migration_entry_wait(vma->vm_mm, vmf->pmd,
3253					     vmf->address);
3254		} else if (is_device_private_entry(entry)) {
3255			vmf->page = device_private_entry_to_page(entry);
3256			ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3257		} else if (is_hwpoison_entry(entry)) {
3258			ret = VM_FAULT_HWPOISON;
3259		} else {
3260			print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3261			ret = VM_FAULT_SIGBUS;
3262		}
3263		goto out;
3264	}
3265
3266
3267	delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3268	page = lookup_swap_cache(entry, vma, vmf->address);
3269	swapcache = page;
3270
3271	if (!page) {
3272		struct swap_info_struct *si = swp_swap_info(entry);
3273
3274		if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3275		    __swap_count(entry) == 1) {
3276			/* skip swapcache */
3277			page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3278							vmf->address);
3279			if (page) {
3280				int err;
3281
3282				__SetPageLocked(page);
3283				__SetPageSwapBacked(page);
3284				set_page_private(page, entry.val);
3285
3286				/* Tell memcg to use swap ownership records */
3287				SetPageSwapCache(page);
3288				err = mem_cgroup_charge(page, vma->vm_mm,
3289							GFP_KERNEL);
3290				ClearPageSwapCache(page);
3291				if (err) {
3292					ret = VM_FAULT_OOM;
3293					goto out_page;
3294				}
3295
3296				shadow = get_shadow_from_swap_cache(entry);
3297				if (shadow)
3298					workingset_refault(page, shadow);
3299
3300				lru_cache_add(page);
3301				swap_readpage(page, true);
3302			}
3303		} else {
3304			page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3305						vmf);
3306			swapcache = page;
3307		}
3308
3309		if (!page) {
3310			/*
3311			 * Back out if somebody else faulted in this pte
3312			 * while we released the pte lock.
3313			 */
3314			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3315					vmf->address, &vmf->ptl);
3316			if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3317				ret = VM_FAULT_OOM;
3318			delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3319			goto unlock;
3320		}
3321
3322		/* Had to read the page from swap area: Major fault */
3323		ret = VM_FAULT_MAJOR;
3324		count_vm_event(PGMAJFAULT);
3325		count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3326	} else if (PageHWPoison(page)) {
3327		/*
3328		 * hwpoisoned dirty swapcache pages are kept for killing
3329		 * owner processes (which may be unknown at hwpoison time)
3330		 */
3331		ret = VM_FAULT_HWPOISON;
3332		delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3333		goto out_release;
3334	}
3335
3336	locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3337
3338	delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3339	if (!locked) {
3340		ret |= VM_FAULT_RETRY;
3341		goto out_release;
3342	}
3343
3344	/*
3345	 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3346	 * release the swapcache from under us.  The page pin, and pte_same
3347	 * test below, are not enough to exclude that.  Even if it is still
3348	 * swapcache, we need to check that the page's swap has not changed.
3349	 */
3350	if (unlikely((!PageSwapCache(page) ||
3351			page_private(page) != entry.val)) && swapcache)
3352		goto out_page;
3353
3354	page = ksm_might_need_to_copy(page, vma, vmf->address);
3355	if (unlikely(!page)) {
3356		ret = VM_FAULT_OOM;
3357		page = swapcache;
3358		goto out_page;
3359	}
3360
3361	cgroup_throttle_swaprate(page, GFP_KERNEL);
 
 
 
 
3362
3363	/*
3364	 * Back out if somebody else already faulted in this pte.
3365	 */
3366	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3367			&vmf->ptl);
3368	if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3369		goto out_nomap;
3370
3371	if (unlikely(!PageUptodate(page))) {
3372		ret = VM_FAULT_SIGBUS;
3373		goto out_nomap;
3374	}
3375
3376	/*
3377	 * The page isn't present yet, go ahead with the fault.
3378	 *
3379	 * Be careful about the sequence of operations here.
3380	 * To get its accounting right, reuse_swap_page() must be called
3381	 * while the page is counted on swap but not yet in mapcount i.e.
3382	 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3383	 * must be called after the swap_free(), or it will never succeed.
3384	 */
3385
3386	inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3387	dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3388	pte = mk_pte(page, vma->vm_page_prot);
3389	if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3390		pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3391		vmf->flags &= ~FAULT_FLAG_WRITE;
3392		ret |= VM_FAULT_WRITE;
3393		exclusive = RMAP_EXCLUSIVE;
3394	}
3395	flush_icache_page(vma, page);
3396	if (pte_swp_soft_dirty(vmf->orig_pte))
3397		pte = pte_mksoft_dirty(pte);
3398	if (pte_swp_uffd_wp(vmf->orig_pte)) {
3399		pte = pte_mkuffd_wp(pte);
3400		pte = pte_wrprotect(pte);
3401	}
3402	set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3403	arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3404	vmf->orig_pte = pte;
3405
3406	/* ksm created a completely new copy */
3407	if (unlikely(page != swapcache && swapcache)) {
3408		page_add_new_anon_rmap(page, vma, vmf->address, false);
3409		lru_cache_add_inactive_or_unevictable(page, vma);
 
3410	} else {
3411		do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
 
 
3412	}
3413
3414	swap_free(entry);
3415	if (mem_cgroup_swap_full(page) ||
3416	    (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3417		try_to_free_swap(page);
3418	unlock_page(page);
3419	if (page != swapcache && swapcache) {
3420		/*
3421		 * Hold the lock to avoid the swap entry to be reused
3422		 * until we take the PT lock for the pte_same() check
3423		 * (to avoid false positives from pte_same). For
3424		 * further safety release the lock after the swap_free
3425		 * so that the swap count won't change under a
3426		 * parallel locked swapcache.
3427		 */
3428		unlock_page(swapcache);
3429		put_page(swapcache);
3430	}
3431
3432	if (vmf->flags & FAULT_FLAG_WRITE) {
3433		ret |= do_wp_page(vmf);
3434		if (ret & VM_FAULT_ERROR)
3435			ret &= VM_FAULT_ERROR;
3436		goto out;
3437	}
3438
3439	/* No need to invalidate - it was non-present before */
3440	update_mmu_cache(vma, vmf->address, vmf->pte);
3441unlock:
3442	pte_unmap_unlock(vmf->pte, vmf->ptl);
3443out:
3444	return ret;
3445out_nomap:
 
3446	pte_unmap_unlock(vmf->pte, vmf->ptl);
3447out_page:
3448	unlock_page(page);
3449out_release:
3450	put_page(page);
3451	if (page != swapcache && swapcache) {
3452		unlock_page(swapcache);
3453		put_page(swapcache);
3454	}
3455	return ret;
3456}
3457
3458/*
3459 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3460 * but allow concurrent faults), and pte mapped but not yet locked.
3461 * We return with mmap_lock still held, but pte unmapped and unlocked.
3462 */
3463static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3464{
3465	struct vm_area_struct *vma = vmf->vma;
 
3466	struct page *page;
3467	vm_fault_t ret = 0;
3468	pte_t entry;
3469
3470	/* File mapping without ->vm_ops ? */
3471	if (vma->vm_flags & VM_SHARED)
3472		return VM_FAULT_SIGBUS;
3473
3474	/*
3475	 * Use pte_alloc() instead of pte_alloc_map().  We can't run
3476	 * pte_offset_map() on pmds where a huge pmd might be created
3477	 * from a different thread.
3478	 *
3479	 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3480	 * parallel threads are excluded by other means.
3481	 *
3482	 * Here we only have mmap_read_lock(mm).
3483	 */
3484	if (pte_alloc(vma->vm_mm, vmf->pmd))
3485		return VM_FAULT_OOM;
3486
3487	/* See the comment in pte_alloc_one_map() */
3488	if (unlikely(pmd_trans_unstable(vmf->pmd)))
3489		return 0;
3490
3491	/* Use the zero-page for reads */
3492	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3493			!mm_forbids_zeropage(vma->vm_mm)) {
3494		entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3495						vma->vm_page_prot));
3496		vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3497				vmf->address, &vmf->ptl);
3498		if (!pte_none(*vmf->pte)) {
3499			update_mmu_tlb(vma, vmf->address, vmf->pte);
3500			goto unlock;
3501		}
3502		ret = check_stable_address_space(vma->vm_mm);
3503		if (ret)
3504			goto unlock;
3505		/* Deliver the page fault to userland, check inside PT lock */
3506		if (userfaultfd_missing(vma)) {
3507			pte_unmap_unlock(vmf->pte, vmf->ptl);
3508			return handle_userfault(vmf, VM_UFFD_MISSING);
3509		}
3510		goto setpte;
3511	}
3512
3513	/* Allocate our own private page. */
3514	if (unlikely(anon_vma_prepare(vma)))
3515		goto oom;
3516	page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3517	if (!page)
3518		goto oom;
3519
3520	if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
 
3521		goto oom_free_page;
3522	cgroup_throttle_swaprate(page, GFP_KERNEL);
3523
3524	/*
3525	 * The memory barrier inside __SetPageUptodate makes sure that
3526	 * preceding stores to the page contents become visible before
3527	 * the set_pte_at() write.
3528	 */
3529	__SetPageUptodate(page);
3530
3531	entry = mk_pte(page, vma->vm_page_prot);
3532	entry = pte_sw_mkyoung(entry);
3533	if (vma->vm_flags & VM_WRITE)
3534		entry = pte_mkwrite(pte_mkdirty(entry));
3535
3536	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3537			&vmf->ptl);
3538	if (!pte_none(*vmf->pte)) {
3539		update_mmu_cache(vma, vmf->address, vmf->pte);
3540		goto release;
3541	}
3542
3543	ret = check_stable_address_space(vma->vm_mm);
3544	if (ret)
3545		goto release;
3546
3547	/* Deliver the page fault to userland, check inside PT lock */
3548	if (userfaultfd_missing(vma)) {
3549		pte_unmap_unlock(vmf->pte, vmf->ptl);
 
3550		put_page(page);
3551		return handle_userfault(vmf, VM_UFFD_MISSING);
3552	}
3553
3554	inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3555	page_add_new_anon_rmap(page, vma, vmf->address, false);
3556	lru_cache_add_inactive_or_unevictable(page, vma);
 
3557setpte:
3558	set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3559
3560	/* No need to invalidate - it was non-present before */
3561	update_mmu_cache(vma, vmf->address, vmf->pte);
3562unlock:
3563	pte_unmap_unlock(vmf->pte, vmf->ptl);
3564	return ret;
3565release:
 
3566	put_page(page);
3567	goto unlock;
3568oom_free_page:
3569	put_page(page);
3570oom:
3571	return VM_FAULT_OOM;
3572}
3573
3574/*
3575 * The mmap_lock must have been held on entry, and may have been
3576 * released depending on flags and vma->vm_ops->fault() return value.
3577 * See filemap_fault() and __lock_page_retry().
3578 */
3579static vm_fault_t __do_fault(struct vm_fault *vmf)
3580{
3581	struct vm_area_struct *vma = vmf->vma;
3582	vm_fault_t ret;
3583
3584	/*
3585	 * Preallocate pte before we take page_lock because this might lead to
3586	 * deadlocks for memcg reclaim which waits for pages under writeback:
3587	 *				lock_page(A)
3588	 *				SetPageWriteback(A)
3589	 *				unlock_page(A)
3590	 * lock_page(B)
3591	 *				lock_page(B)
3592	 * pte_alloc_pne
3593	 *   shrink_page_list
3594	 *     wait_on_page_writeback(A)
3595	 *				SetPageWriteback(B)
3596	 *				unlock_page(B)
3597	 *				# flush A, B to clear the writeback
3598	 */
3599	if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3600		vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3601		if (!vmf->prealloc_pte)
3602			return VM_FAULT_OOM;
3603		smp_wmb(); /* See comment in __pte_alloc() */
3604	}
3605
3606	ret = vma->vm_ops->fault(vmf);
3607	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3608			    VM_FAULT_DONE_COW)))
3609		return ret;
3610
3611	if (unlikely(PageHWPoison(vmf->page))) {
3612		if (ret & VM_FAULT_LOCKED)
3613			unlock_page(vmf->page);
3614		put_page(vmf->page);
3615		vmf->page = NULL;
3616		return VM_FAULT_HWPOISON;
3617	}
3618
3619	if (unlikely(!(ret & VM_FAULT_LOCKED)))
3620		lock_page(vmf->page);
3621	else
3622		VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3623
3624	return ret;
3625}
3626
3627/*
3628 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3629 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3630 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3631 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3632 */
3633static int pmd_devmap_trans_unstable(pmd_t *pmd)
3634{
3635	return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3636}
3637
3638static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3639{
3640	struct vm_area_struct *vma = vmf->vma;
3641
3642	if (!pmd_none(*vmf->pmd))
3643		goto map_pte;
3644	if (vmf->prealloc_pte) {
3645		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3646		if (unlikely(!pmd_none(*vmf->pmd))) {
3647			spin_unlock(vmf->ptl);
3648			goto map_pte;
3649		}
3650
3651		mm_inc_nr_ptes(vma->vm_mm);
3652		pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3653		spin_unlock(vmf->ptl);
3654		vmf->prealloc_pte = NULL;
3655	} else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3656		return VM_FAULT_OOM;
3657	}
3658map_pte:
3659	/*
3660	 * If a huge pmd materialized under us just retry later.  Use
3661	 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3662	 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3663	 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3664	 * running immediately after a huge pmd fault in a different thread of
3665	 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3666	 * All we have to ensure is that it is a regular pmd that we can walk
3667	 * with pte_offset_map() and we can do that through an atomic read in
3668	 * C, which is what pmd_trans_unstable() provides.
3669	 */
3670	if (pmd_devmap_trans_unstable(vmf->pmd))
3671		return VM_FAULT_NOPAGE;
3672
3673	/*
3674	 * At this point we know that our vmf->pmd points to a page of ptes
3675	 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3676	 * for the duration of the fault.  If a racing MADV_DONTNEED runs and
3677	 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3678	 * be valid and we will re-check to make sure the vmf->pte isn't
3679	 * pte_none() under vmf->ptl protection when we return to
3680	 * alloc_set_pte().
3681	 */
3682	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3683			&vmf->ptl);
3684	return 0;
3685}
3686
3687#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3688static void deposit_prealloc_pte(struct vm_fault *vmf)
3689{
3690	struct vm_area_struct *vma = vmf->vma;
3691
3692	pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3693	/*
3694	 * We are going to consume the prealloc table,
3695	 * count that as nr_ptes.
3696	 */
3697	mm_inc_nr_ptes(vma->vm_mm);
3698	vmf->prealloc_pte = NULL;
3699}
3700
3701static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3702{
3703	struct vm_area_struct *vma = vmf->vma;
3704	bool write = vmf->flags & FAULT_FLAG_WRITE;
3705	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3706	pmd_t entry;
3707	int i;
3708	vm_fault_t ret;
3709
3710	if (!transhuge_vma_suitable(vma, haddr))
3711		return VM_FAULT_FALLBACK;
3712
3713	ret = VM_FAULT_FALLBACK;
3714	page = compound_head(page);
3715
3716	/*
3717	 * Archs like ppc64 need additonal space to store information
3718	 * related to pte entry. Use the preallocated table for that.
3719	 */
3720	if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3721		vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3722		if (!vmf->prealloc_pte)
3723			return VM_FAULT_OOM;
3724		smp_wmb(); /* See comment in __pte_alloc() */
3725	}
3726
3727	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3728	if (unlikely(!pmd_none(*vmf->pmd)))
3729		goto out;
3730
3731	for (i = 0; i < HPAGE_PMD_NR; i++)
3732		flush_icache_page(vma, page + i);
3733
3734	entry = mk_huge_pmd(page, vma->vm_page_prot);
3735	if (write)
3736		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3737
3738	add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3739	page_add_file_rmap(page, true);
3740	/*
3741	 * deposit and withdraw with pmd lock held
3742	 */
3743	if (arch_needs_pgtable_deposit())
3744		deposit_prealloc_pte(vmf);
3745
3746	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3747
3748	update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3749
3750	/* fault is handled */
3751	ret = 0;
3752	count_vm_event(THP_FILE_MAPPED);
3753out:
3754	spin_unlock(vmf->ptl);
3755	return ret;
3756}
3757#else
3758static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3759{
3760	BUILD_BUG();
3761	return 0;
3762}
3763#endif
3764
3765/**
3766 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3767 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3768 *
3769 * @vmf: fault environment
 
3770 * @page: page to map
3771 *
3772 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3773 * return.
3774 *
3775 * Target users are page handler itself and implementations of
3776 * vm_ops->map_pages.
3777 *
3778 * Return: %0 on success, %VM_FAULT_ code in case of error.
3779 */
3780vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct page *page)
 
3781{
3782	struct vm_area_struct *vma = vmf->vma;
3783	bool write = vmf->flags & FAULT_FLAG_WRITE;
3784	pte_t entry;
3785	vm_fault_t ret;
3786
3787	if (pmd_none(*vmf->pmd) && PageTransCompound(page)) {
 
 
 
 
3788		ret = do_set_pmd(vmf, page);
3789		if (ret != VM_FAULT_FALLBACK)
3790			return ret;
3791	}
3792
3793	if (!vmf->pte) {
3794		ret = pte_alloc_one_map(vmf);
3795		if (ret)
3796			return ret;
3797	}
3798
3799	/* Re-check under ptl */
3800	if (unlikely(!pte_none(*vmf->pte))) {
3801		update_mmu_tlb(vma, vmf->address, vmf->pte);
3802		return VM_FAULT_NOPAGE;
3803	}
3804
3805	flush_icache_page(vma, page);
3806	entry = mk_pte(page, vma->vm_page_prot);
3807	entry = pte_sw_mkyoung(entry);
3808	if (write)
3809		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3810	/* copy-on-write page */
3811	if (write && !(vma->vm_flags & VM_SHARED)) {
3812		inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3813		page_add_new_anon_rmap(page, vma, vmf->address, false);
3814		lru_cache_add_inactive_or_unevictable(page, vma);
 
3815	} else {
3816		inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3817		page_add_file_rmap(page, false);
3818	}
3819	set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3820
3821	/* no need to invalidate: a not-present page won't be cached */
3822	update_mmu_cache(vma, vmf->address, vmf->pte);
3823
3824	return 0;
3825}
3826
3827
3828/**
3829 * finish_fault - finish page fault once we have prepared the page to fault
3830 *
3831 * @vmf: structure describing the fault
3832 *
3833 * This function handles all that is needed to finish a page fault once the
3834 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3835 * given page, adds reverse page mapping, handles memcg charges and LRU
3836 * addition.
3837 *
3838 * The function expects the page to be locked and on success it consumes a
3839 * reference of a page being mapped (for the PTE which maps it).
3840 *
3841 * Return: %0 on success, %VM_FAULT_ code in case of error.
3842 */
3843vm_fault_t finish_fault(struct vm_fault *vmf)
3844{
3845	struct page *page;
3846	vm_fault_t ret = 0;
3847
3848	/* Did we COW the page? */
3849	if ((vmf->flags & FAULT_FLAG_WRITE) &&
3850	    !(vmf->vma->vm_flags & VM_SHARED))
3851		page = vmf->cow_page;
3852	else
3853		page = vmf->page;
3854
3855	/*
3856	 * check even for read faults because we might have lost our CoWed
3857	 * page
3858	 */
3859	if (!(vmf->vma->vm_flags & VM_SHARED))
3860		ret = check_stable_address_space(vmf->vma->vm_mm);
3861	if (!ret)
3862		ret = alloc_set_pte(vmf, page);
3863	if (vmf->pte)
3864		pte_unmap_unlock(vmf->pte, vmf->ptl);
3865	return ret;
3866}
3867
3868static unsigned long fault_around_bytes __read_mostly =
3869	rounddown_pow_of_two(65536);
3870
3871#ifdef CONFIG_DEBUG_FS
3872static int fault_around_bytes_get(void *data, u64 *val)
3873{
3874	*val = fault_around_bytes;
3875	return 0;
3876}
3877
3878/*
3879 * fault_around_bytes must be rounded down to the nearest page order as it's
3880 * what do_fault_around() expects to see.
3881 */
3882static int fault_around_bytes_set(void *data, u64 val)
3883{
3884	if (val / PAGE_SIZE > PTRS_PER_PTE)
3885		return -EINVAL;
3886	if (val > PAGE_SIZE)
3887		fault_around_bytes = rounddown_pow_of_two(val);
3888	else
3889		fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3890	return 0;
3891}
3892DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3893		fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3894
3895static int __init fault_around_debugfs(void)
3896{
3897	debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3898				   &fault_around_bytes_fops);
3899	return 0;
3900}
3901late_initcall(fault_around_debugfs);
3902#endif
3903
3904/*
3905 * do_fault_around() tries to map few pages around the fault address. The hope
3906 * is that the pages will be needed soon and this will lower the number of
3907 * faults to handle.
3908 *
3909 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3910 * not ready to be mapped: not up-to-date, locked, etc.
3911 *
3912 * This function is called with the page table lock taken. In the split ptlock
3913 * case the page table lock only protects only those entries which belong to
3914 * the page table corresponding to the fault address.
3915 *
3916 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3917 * only once.
3918 *
3919 * fault_around_bytes defines how many bytes we'll try to map.
3920 * do_fault_around() expects it to be set to a power of two less than or equal
3921 * to PTRS_PER_PTE.
3922 *
3923 * The virtual address of the area that we map is naturally aligned to
3924 * fault_around_bytes rounded down to the machine page size
3925 * (and therefore to page order).  This way it's easier to guarantee
3926 * that we don't cross page table boundaries.
3927 */
3928static vm_fault_t do_fault_around(struct vm_fault *vmf)
3929{
3930	unsigned long address = vmf->address, nr_pages, mask;
3931	pgoff_t start_pgoff = vmf->pgoff;
3932	pgoff_t end_pgoff;
3933	int off;
3934	vm_fault_t ret = 0;
3935
3936	nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3937	mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3938
3939	vmf->address = max(address & mask, vmf->vma->vm_start);
3940	off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3941	start_pgoff -= off;
3942
3943	/*
3944	 *  end_pgoff is either the end of the page table, the end of
3945	 *  the vma or nr_pages from start_pgoff, depending what is nearest.
3946	 */
3947	end_pgoff = start_pgoff -
3948		((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3949		PTRS_PER_PTE - 1;
3950	end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3951			start_pgoff + nr_pages - 1);
3952
3953	if (pmd_none(*vmf->pmd)) {
3954		vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3955		if (!vmf->prealloc_pte)
3956			goto out;
3957		smp_wmb(); /* See comment in __pte_alloc() */
3958	}
3959
3960	vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3961
3962	/* Huge page is mapped? Page fault is solved */
3963	if (pmd_trans_huge(*vmf->pmd)) {
3964		ret = VM_FAULT_NOPAGE;
3965		goto out;
3966	}
3967
3968	/* ->map_pages() haven't done anything useful. Cold page cache? */
3969	if (!vmf->pte)
3970		goto out;
3971
3972	/* check if the page fault is solved */
3973	vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3974	if (!pte_none(*vmf->pte))
3975		ret = VM_FAULT_NOPAGE;
3976	pte_unmap_unlock(vmf->pte, vmf->ptl);
3977out:
3978	vmf->address = address;
3979	vmf->pte = NULL;
3980	return ret;
3981}
3982
3983static vm_fault_t do_read_fault(struct vm_fault *vmf)
3984{
3985	struct vm_area_struct *vma = vmf->vma;
3986	vm_fault_t ret = 0;
3987
3988	/*
3989	 * Let's call ->map_pages() first and use ->fault() as fallback
3990	 * if page by the offset is not ready to be mapped (cold cache or
3991	 * something).
3992	 */
3993	if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3994		ret = do_fault_around(vmf);
3995		if (ret)
3996			return ret;
3997	}
3998
3999	ret = __do_fault(vmf);
4000	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4001		return ret;
4002
4003	ret |= finish_fault(vmf);
4004	unlock_page(vmf->page);
4005	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4006		put_page(vmf->page);
4007	return ret;
4008}
4009
4010static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4011{
4012	struct vm_area_struct *vma = vmf->vma;
4013	vm_fault_t ret;
4014
4015	if (unlikely(anon_vma_prepare(vma)))
4016		return VM_FAULT_OOM;
4017
4018	vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4019	if (!vmf->cow_page)
4020		return VM_FAULT_OOM;
4021
4022	if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
 
4023		put_page(vmf->cow_page);
4024		return VM_FAULT_OOM;
4025	}
4026	cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4027
4028	ret = __do_fault(vmf);
4029	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4030		goto uncharge_out;
4031	if (ret & VM_FAULT_DONE_COW)
4032		return ret;
4033
4034	copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4035	__SetPageUptodate(vmf->cow_page);
4036
4037	ret |= finish_fault(vmf);
4038	unlock_page(vmf->page);
4039	put_page(vmf->page);
4040	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4041		goto uncharge_out;
4042	return ret;
4043uncharge_out:
 
4044	put_page(vmf->cow_page);
4045	return ret;
4046}
4047
4048static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4049{
4050	struct vm_area_struct *vma = vmf->vma;
4051	vm_fault_t ret, tmp;
4052
4053	ret = __do_fault(vmf);
4054	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4055		return ret;
4056
4057	/*
4058	 * Check if the backing address space wants to know that the page is
4059	 * about to become writable
4060	 */
4061	if (vma->vm_ops->page_mkwrite) {
4062		unlock_page(vmf->page);
4063		tmp = do_page_mkwrite(vmf);
4064		if (unlikely(!tmp ||
4065				(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4066			put_page(vmf->page);
4067			return tmp;
4068		}
4069	}
4070
4071	ret |= finish_fault(vmf);
4072	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4073					VM_FAULT_RETRY))) {
4074		unlock_page(vmf->page);
4075		put_page(vmf->page);
4076		return ret;
4077	}
4078
4079	ret |= fault_dirty_shared_page(vmf);
4080	return ret;
4081}
4082
4083/*
4084 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4085 * but allow concurrent faults).
4086 * The mmap_lock may have been released depending on flags and our
4087 * return value.  See filemap_fault() and __lock_page_or_retry().
4088 * If mmap_lock is released, vma may become invalid (for example
4089 * by other thread calling munmap()).
4090 */
4091static vm_fault_t do_fault(struct vm_fault *vmf)
4092{
4093	struct vm_area_struct *vma = vmf->vma;
4094	struct mm_struct *vm_mm = vma->vm_mm;
4095	vm_fault_t ret;
4096
4097	/*
4098	 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4099	 */
4100	if (!vma->vm_ops->fault) {
4101		/*
4102		 * If we find a migration pmd entry or a none pmd entry, which
4103		 * should never happen, return SIGBUS
4104		 */
4105		if (unlikely(!pmd_present(*vmf->pmd)))
4106			ret = VM_FAULT_SIGBUS;
4107		else {
4108			vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4109						       vmf->pmd,
4110						       vmf->address,
4111						       &vmf->ptl);
4112			/*
4113			 * Make sure this is not a temporary clearing of pte
4114			 * by holding ptl and checking again. A R/M/W update
4115			 * of pte involves: take ptl, clearing the pte so that
4116			 * we don't have concurrent modification by hardware
4117			 * followed by an update.
4118			 */
4119			if (unlikely(pte_none(*vmf->pte)))
4120				ret = VM_FAULT_SIGBUS;
4121			else
4122				ret = VM_FAULT_NOPAGE;
4123
4124			pte_unmap_unlock(vmf->pte, vmf->ptl);
4125		}
4126	} else if (!(vmf->flags & FAULT_FLAG_WRITE))
4127		ret = do_read_fault(vmf);
4128	else if (!(vma->vm_flags & VM_SHARED))
4129		ret = do_cow_fault(vmf);
4130	else
4131		ret = do_shared_fault(vmf);
4132
4133	/* preallocated pagetable is unused: free it */
4134	if (vmf->prealloc_pte) {
4135		pte_free(vm_mm, vmf->prealloc_pte);
4136		vmf->prealloc_pte = NULL;
4137	}
4138	return ret;
4139}
4140
4141static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4142				unsigned long addr, int page_nid,
4143				int *flags)
4144{
4145	get_page(page);
4146
4147	count_vm_numa_event(NUMA_HINT_FAULTS);
4148	if (page_nid == numa_node_id()) {
4149		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4150		*flags |= TNF_FAULT_LOCAL;
4151	}
4152
4153	return mpol_misplaced(page, vma, addr);
4154}
4155
4156static vm_fault_t do_numa_page(struct vm_fault *vmf)
4157{
4158	struct vm_area_struct *vma = vmf->vma;
4159	struct page *page = NULL;
4160	int page_nid = NUMA_NO_NODE;
4161	int last_cpupid;
4162	int target_nid;
4163	bool migrated = false;
4164	pte_t pte, old_pte;
4165	bool was_writable = pte_savedwrite(vmf->orig_pte);
4166	int flags = 0;
4167
4168	/*
4169	 * The "pte" at this point cannot be used safely without
4170	 * validation through pte_unmap_same(). It's of NUMA type but
4171	 * the pfn may be screwed if the read is non atomic.
4172	 */
4173	vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4174	spin_lock(vmf->ptl);
4175	if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4176		pte_unmap_unlock(vmf->pte, vmf->ptl);
4177		goto out;
4178	}
4179
4180	/*
4181	 * Make it present again, Depending on how arch implementes non
4182	 * accessible ptes, some can allow access by kernel mode.
4183	 */
4184	old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4185	pte = pte_modify(old_pte, vma->vm_page_prot);
4186	pte = pte_mkyoung(pte);
4187	if (was_writable)
4188		pte = pte_mkwrite(pte);
4189	ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4190	update_mmu_cache(vma, vmf->address, vmf->pte);
4191
4192	page = vm_normal_page(vma, vmf->address, pte);
4193	if (!page) {
4194		pte_unmap_unlock(vmf->pte, vmf->ptl);
4195		return 0;
4196	}
4197
4198	/* TODO: handle PTE-mapped THP */
4199	if (PageCompound(page)) {
4200		pte_unmap_unlock(vmf->pte, vmf->ptl);
4201		return 0;
4202	}
4203
4204	/*
4205	 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4206	 * much anyway since they can be in shared cache state. This misses
4207	 * the case where a mapping is writable but the process never writes
4208	 * to it but pte_write gets cleared during protection updates and
4209	 * pte_dirty has unpredictable behaviour between PTE scan updates,
4210	 * background writeback, dirty balancing and application behaviour.
4211	 */
4212	if (!pte_write(pte))
4213		flags |= TNF_NO_GROUP;
4214
4215	/*
4216	 * Flag if the page is shared between multiple address spaces. This
4217	 * is later used when determining whether to group tasks together
4218	 */
4219	if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4220		flags |= TNF_SHARED;
4221
4222	last_cpupid = page_cpupid_last(page);
4223	page_nid = page_to_nid(page);
4224	target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4225			&flags);
4226	pte_unmap_unlock(vmf->pte, vmf->ptl);
4227	if (target_nid == NUMA_NO_NODE) {
4228		put_page(page);
4229		goto out;
4230	}
4231
4232	/* Migrate to the requested node */
4233	migrated = migrate_misplaced_page(page, vma, target_nid);
4234	if (migrated) {
4235		page_nid = target_nid;
4236		flags |= TNF_MIGRATED;
4237	} else
4238		flags |= TNF_MIGRATE_FAIL;
4239
4240out:
4241	if (page_nid != NUMA_NO_NODE)
4242		task_numa_fault(last_cpupid, page_nid, 1, flags);
4243	return 0;
4244}
4245
4246static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4247{
4248	if (vma_is_anonymous(vmf->vma))
4249		return do_huge_pmd_anonymous_page(vmf);
4250	if (vmf->vma->vm_ops->huge_fault)
4251		return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4252	return VM_FAULT_FALLBACK;
4253}
4254
4255/* `inline' is required to avoid gcc 4.1.2 build error */
4256static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4257{
4258	if (vma_is_anonymous(vmf->vma)) {
4259		if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
4260			return handle_userfault(vmf, VM_UFFD_WP);
4261		return do_huge_pmd_wp_page(vmf, orig_pmd);
4262	}
4263	if (vmf->vma->vm_ops->huge_fault) {
4264		vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4265
4266		if (!(ret & VM_FAULT_FALLBACK))
4267			return ret;
4268	}
4269
4270	/* COW or write-notify handled on pte level: split pmd. */
 
4271	__split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4272
4273	return VM_FAULT_FALLBACK;
4274}
4275
 
 
 
 
 
4276static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4277{
4278#if defined(CONFIG_TRANSPARENT_HUGEPAGE) &&			\
4279	defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4280	/* No support for anonymous transparent PUD pages yet */
4281	if (vma_is_anonymous(vmf->vma))
4282		goto split;
4283	if (vmf->vma->vm_ops->huge_fault) {
4284		vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4285
4286		if (!(ret & VM_FAULT_FALLBACK))
4287			return ret;
4288	}
4289split:
4290	/* COW or write-notify not handled on PUD level: split pud.*/
4291	__split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4292#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4293	return VM_FAULT_FALLBACK;
4294}
4295
4296static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4297{
4298#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4299	/* No support for anonymous transparent PUD pages yet */
4300	if (vma_is_anonymous(vmf->vma))
4301		return VM_FAULT_FALLBACK;
4302	if (vmf->vma->vm_ops->huge_fault)
4303		return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4304#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4305	return VM_FAULT_FALLBACK;
4306}
4307
4308/*
4309 * These routines also need to handle stuff like marking pages dirty
4310 * and/or accessed for architectures that don't do it in hardware (most
4311 * RISC architectures).  The early dirtying is also good on the i386.
4312 *
4313 * There is also a hook called "update_mmu_cache()" that architectures
4314 * with external mmu caches can use to update those (ie the Sparc or
4315 * PowerPC hashed page tables that act as extended TLBs).
4316 *
4317 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4318 * concurrent faults).
4319 *
4320 * The mmap_lock may have been released depending on flags and our return value.
4321 * See filemap_fault() and __lock_page_or_retry().
4322 */
4323static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4324{
4325	pte_t entry;
4326
4327	if (unlikely(pmd_none(*vmf->pmd))) {
4328		/*
4329		 * Leave __pte_alloc() until later: because vm_ops->fault may
4330		 * want to allocate huge page, and if we expose page table
4331		 * for an instant, it will be difficult to retract from
4332		 * concurrent faults and from rmap lookups.
4333		 */
4334		vmf->pte = NULL;
4335	} else {
4336		/* See comment in pte_alloc_one_map() */
4337		if (pmd_devmap_trans_unstable(vmf->pmd))
4338			return 0;
4339		/*
4340		 * A regular pmd is established and it can't morph into a huge
4341		 * pmd from under us anymore at this point because we hold the
4342		 * mmap_lock read mode and khugepaged takes it in write mode.
4343		 * So now it's safe to run pte_offset_map().
4344		 */
4345		vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4346		vmf->orig_pte = *vmf->pte;
4347
4348		/*
4349		 * some architectures can have larger ptes than wordsize,
4350		 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4351		 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4352		 * accesses.  The code below just needs a consistent view
4353		 * for the ifs and we later double check anyway with the
4354		 * ptl lock held. So here a barrier will do.
4355		 */
4356		barrier();
4357		if (pte_none(vmf->orig_pte)) {
4358			pte_unmap(vmf->pte);
4359			vmf->pte = NULL;
4360		}
4361	}
4362
4363	if (!vmf->pte) {
4364		if (vma_is_anonymous(vmf->vma))
4365			return do_anonymous_page(vmf);
4366		else
4367			return do_fault(vmf);
4368	}
4369
4370	if (!pte_present(vmf->orig_pte))
4371		return do_swap_page(vmf);
4372
4373	if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4374		return do_numa_page(vmf);
4375
4376	vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4377	spin_lock(vmf->ptl);
4378	entry = vmf->orig_pte;
4379	if (unlikely(!pte_same(*vmf->pte, entry))) {
4380		update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4381		goto unlock;
4382	}
4383	if (vmf->flags & FAULT_FLAG_WRITE) {
4384		if (!pte_write(entry))
4385			return do_wp_page(vmf);
4386		entry = pte_mkdirty(entry);
4387	}
4388	entry = pte_mkyoung(entry);
4389	if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4390				vmf->flags & FAULT_FLAG_WRITE)) {
4391		update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4392	} else {
4393		/* Skip spurious TLB flush for retried page fault */
4394		if (vmf->flags & FAULT_FLAG_TRIED)
4395			goto unlock;
4396		/*
4397		 * This is needed only for protection faults but the arch code
4398		 * is not yet telling us if this is a protection fault or not.
4399		 * This still avoids useless tlb flushes for .text page faults
4400		 * with threads.
4401		 */
4402		if (vmf->flags & FAULT_FLAG_WRITE)
4403			flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4404	}
4405unlock:
4406	pte_unmap_unlock(vmf->pte, vmf->ptl);
4407	return 0;
4408}
4409
4410/*
4411 * By the time we get here, we already hold the mm semaphore
4412 *
4413 * The mmap_lock may have been released depending on flags and our
4414 * return value.  See filemap_fault() and __lock_page_or_retry().
4415 */
4416static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4417		unsigned long address, unsigned int flags)
4418{
4419	struct vm_fault vmf = {
4420		.vma = vma,
4421		.address = address & PAGE_MASK,
4422		.flags = flags,
4423		.pgoff = linear_page_index(vma, address),
4424		.gfp_mask = __get_fault_gfp_mask(vma),
4425	};
4426	unsigned int dirty = flags & FAULT_FLAG_WRITE;
4427	struct mm_struct *mm = vma->vm_mm;
4428	pgd_t *pgd;
4429	p4d_t *p4d;
4430	vm_fault_t ret;
4431
4432	pgd = pgd_offset(mm, address);
4433	p4d = p4d_alloc(mm, pgd, address);
4434	if (!p4d)
4435		return VM_FAULT_OOM;
4436
4437	vmf.pud = pud_alloc(mm, p4d, address);
4438	if (!vmf.pud)
4439		return VM_FAULT_OOM;
4440retry_pud:
4441	if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4442		ret = create_huge_pud(&vmf);
4443		if (!(ret & VM_FAULT_FALLBACK))
4444			return ret;
4445	} else {
4446		pud_t orig_pud = *vmf.pud;
4447
4448		barrier();
4449		if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4450
4451			/* NUMA case for anonymous PUDs would go here */
4452
4453			if (dirty && !pud_write(orig_pud)) {
4454				ret = wp_huge_pud(&vmf, orig_pud);
4455				if (!(ret & VM_FAULT_FALLBACK))
4456					return ret;
4457			} else {
4458				huge_pud_set_accessed(&vmf, orig_pud);
4459				return 0;
4460			}
4461		}
4462	}
4463
4464	vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4465	if (!vmf.pmd)
4466		return VM_FAULT_OOM;
4467
4468	/* Huge pud page fault raced with pmd_alloc? */
4469	if (pud_trans_unstable(vmf.pud))
4470		goto retry_pud;
4471
4472	if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4473		ret = create_huge_pmd(&vmf);
4474		if (!(ret & VM_FAULT_FALLBACK))
4475			return ret;
4476	} else {
4477		pmd_t orig_pmd = *vmf.pmd;
4478
4479		barrier();
4480		if (unlikely(is_swap_pmd(orig_pmd))) {
4481			VM_BUG_ON(thp_migration_supported() &&
4482					  !is_pmd_migration_entry(orig_pmd));
4483			if (is_pmd_migration_entry(orig_pmd))
4484				pmd_migration_entry_wait(mm, vmf.pmd);
4485			return 0;
4486		}
4487		if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4488			if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4489				return do_huge_pmd_numa_page(&vmf, orig_pmd);
4490
4491			if (dirty && !pmd_write(orig_pmd)) {
4492				ret = wp_huge_pmd(&vmf, orig_pmd);
4493				if (!(ret & VM_FAULT_FALLBACK))
4494					return ret;
4495			} else {
4496				huge_pmd_set_accessed(&vmf, orig_pmd);
4497				return 0;
4498			}
4499		}
4500	}
4501
4502	return handle_pte_fault(&vmf);
4503}
4504
4505/**
4506 * mm_account_fault - Do page fault accountings
4507 *
4508 * @regs: the pt_regs struct pointer.  When set to NULL, will skip accounting
4509 *        of perf event counters, but we'll still do the per-task accounting to
4510 *        the task who triggered this page fault.
4511 * @address: the faulted address.
4512 * @flags: the fault flags.
4513 * @ret: the fault retcode.
4514 *
4515 * This will take care of most of the page fault accountings.  Meanwhile, it
4516 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4517 * updates.  However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4518 * still be in per-arch page fault handlers at the entry of page fault.
4519 */
4520static inline void mm_account_fault(struct pt_regs *regs,
4521				    unsigned long address, unsigned int flags,
4522				    vm_fault_t ret)
4523{
4524	bool major;
4525
4526	/*
4527	 * We don't do accounting for some specific faults:
4528	 *
4529	 * - Unsuccessful faults (e.g. when the address wasn't valid).  That
4530	 *   includes arch_vma_access_permitted() failing before reaching here.
4531	 *   So this is not a "this many hardware page faults" counter.  We
4532	 *   should use the hw profiling for that.
4533	 *
4534	 * - Incomplete faults (VM_FAULT_RETRY).  They will only be counted
4535	 *   once they're completed.
4536	 */
4537	if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4538		return;
4539
4540	/*
4541	 * We define the fault as a major fault when the final successful fault
4542	 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4543	 * handle it immediately previously).
4544	 */
4545	major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4546
4547	if (major)
4548		current->maj_flt++;
4549	else
4550		current->min_flt++;
4551
4552	/*
4553	 * If the fault is done for GUP, regs will be NULL.  We only do the
4554	 * accounting for the per thread fault counters who triggered the
4555	 * fault, and we skip the perf event updates.
4556	 */
4557	if (!regs)
4558		return;
4559
4560	if (major)
4561		perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4562	else
4563		perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4564}
4565
4566/*
4567 * By the time we get here, we already hold the mm semaphore
4568 *
4569 * The mmap_lock may have been released depending on flags and our
4570 * return value.  See filemap_fault() and __lock_page_or_retry().
4571 */
4572vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4573			   unsigned int flags, struct pt_regs *regs)
4574{
4575	vm_fault_t ret;
4576
4577	__set_current_state(TASK_RUNNING);
4578
4579	count_vm_event(PGFAULT);
4580	count_memcg_event_mm(vma->vm_mm, PGFAULT);
4581
4582	/* do counter updates before entering really critical section. */
4583	check_sync_rss_stat(current);
4584
4585	if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4586					    flags & FAULT_FLAG_INSTRUCTION,
4587					    flags & FAULT_FLAG_REMOTE))
4588		return VM_FAULT_SIGSEGV;
4589
4590	/*
4591	 * Enable the memcg OOM handling for faults triggered in user
4592	 * space.  Kernel faults are handled more gracefully.
4593	 */
4594	if (flags & FAULT_FLAG_USER)
4595		mem_cgroup_enter_user_fault();
4596
4597	if (unlikely(is_vm_hugetlb_page(vma)))
4598		ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4599	else
4600		ret = __handle_mm_fault(vma, address, flags);
4601
4602	if (flags & FAULT_FLAG_USER) {
4603		mem_cgroup_exit_user_fault();
4604		/*
4605		 * The task may have entered a memcg OOM situation but
4606		 * if the allocation error was handled gracefully (no
4607		 * VM_FAULT_OOM), there is no need to kill anything.
4608		 * Just clean up the OOM state peacefully.
4609		 */
4610		if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4611			mem_cgroup_oom_synchronize(false);
4612	}
4613
4614	mm_account_fault(regs, address, flags, ret);
4615
4616	return ret;
4617}
4618EXPORT_SYMBOL_GPL(handle_mm_fault);
4619
4620#ifndef __PAGETABLE_P4D_FOLDED
4621/*
4622 * Allocate p4d page table.
4623 * We've already handled the fast-path in-line.
4624 */
4625int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4626{
4627	p4d_t *new = p4d_alloc_one(mm, address);
4628	if (!new)
4629		return -ENOMEM;
4630
4631	smp_wmb(); /* See comment in __pte_alloc */
4632
4633	spin_lock(&mm->page_table_lock);
4634	if (pgd_present(*pgd))		/* Another has populated it */
4635		p4d_free(mm, new);
4636	else
4637		pgd_populate(mm, pgd, new);
4638	spin_unlock(&mm->page_table_lock);
4639	return 0;
4640}
4641#endif /* __PAGETABLE_P4D_FOLDED */
4642
4643#ifndef __PAGETABLE_PUD_FOLDED
4644/*
4645 * Allocate page upper directory.
4646 * We've already handled the fast-path in-line.
4647 */
4648int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4649{
4650	pud_t *new = pud_alloc_one(mm, address);
4651	if (!new)
4652		return -ENOMEM;
4653
4654	smp_wmb(); /* See comment in __pte_alloc */
4655
4656	spin_lock(&mm->page_table_lock);
 
4657	if (!p4d_present(*p4d)) {
4658		mm_inc_nr_puds(mm);
4659		p4d_populate(mm, p4d, new);
4660	} else	/* Another has populated it */
4661		pud_free(mm, new);
 
 
 
 
 
 
 
4662	spin_unlock(&mm->page_table_lock);
4663	return 0;
4664}
4665#endif /* __PAGETABLE_PUD_FOLDED */
4666
4667#ifndef __PAGETABLE_PMD_FOLDED
4668/*
4669 * Allocate page middle directory.
4670 * We've already handled the fast-path in-line.
4671 */
4672int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4673{
4674	spinlock_t *ptl;
4675	pmd_t *new = pmd_alloc_one(mm, address);
4676	if (!new)
4677		return -ENOMEM;
4678
4679	smp_wmb(); /* See comment in __pte_alloc */
4680
4681	ptl = pud_lock(mm, pud);
 
4682	if (!pud_present(*pud)) {
4683		mm_inc_nr_pmds(mm);
4684		pud_populate(mm, pud, new);
4685	} else	/* Another has populated it */
4686		pmd_free(mm, new);
 
 
 
 
 
 
 
4687	spin_unlock(ptl);
4688	return 0;
4689}
4690#endif /* __PAGETABLE_PMD_FOLDED */
4691
4692static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4693			    struct mmu_notifier_range *range,
4694			    pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4695{
4696	pgd_t *pgd;
4697	p4d_t *p4d;
4698	pud_t *pud;
4699	pmd_t *pmd;
4700	pte_t *ptep;
4701
4702	pgd = pgd_offset(mm, address);
4703	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4704		goto out;
4705
4706	p4d = p4d_offset(pgd, address);
4707	if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4708		goto out;
4709
4710	pud = pud_offset(p4d, address);
4711	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4712		goto out;
4713
4714	pmd = pmd_offset(pud, address);
4715	VM_BUG_ON(pmd_trans_huge(*pmd));
4716
4717	if (pmd_huge(*pmd)) {
4718		if (!pmdpp)
4719			goto out;
4720
4721		if (range) {
4722			mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4723						NULL, mm, address & PMD_MASK,
4724						(address & PMD_MASK) + PMD_SIZE);
4725			mmu_notifier_invalidate_range_start(range);
4726		}
4727		*ptlp = pmd_lock(mm, pmd);
4728		if (pmd_huge(*pmd)) {
4729			*pmdpp = pmd;
4730			return 0;
4731		}
4732		spin_unlock(*ptlp);
4733		if (range)
4734			mmu_notifier_invalidate_range_end(range);
4735	}
4736
4737	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4738		goto out;
4739
4740	if (range) {
4741		mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4742					address & PAGE_MASK,
4743					(address & PAGE_MASK) + PAGE_SIZE);
4744		mmu_notifier_invalidate_range_start(range);
4745	}
4746	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4747	if (!pte_present(*ptep))
4748		goto unlock;
4749	*ptepp = ptep;
4750	return 0;
4751unlock:
4752	pte_unmap_unlock(ptep, *ptlp);
4753	if (range)
4754		mmu_notifier_invalidate_range_end(range);
4755out:
4756	return -EINVAL;
4757}
4758
4759static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4760			     pte_t **ptepp, spinlock_t **ptlp)
4761{
4762	int res;
4763
4764	/* (void) is needed to make gcc happy */
4765	(void) __cond_lock(*ptlp,
4766			   !(res = __follow_pte_pmd(mm, address, NULL,
4767						    ptepp, NULL, ptlp)));
4768	return res;
4769}
4770
4771int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4772		   struct mmu_notifier_range *range,
4773		   pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4774{
4775	int res;
4776
4777	/* (void) is needed to make gcc happy */
4778	(void) __cond_lock(*ptlp,
4779			   !(res = __follow_pte_pmd(mm, address, range,
4780						    ptepp, pmdpp, ptlp)));
4781	return res;
4782}
4783EXPORT_SYMBOL(follow_pte_pmd);
4784
4785/**
4786 * follow_pfn - look up PFN at a user virtual address
4787 * @vma: memory mapping
4788 * @address: user virtual address
4789 * @pfn: location to store found PFN
4790 *
4791 * Only IO mappings and raw PFN mappings are allowed.
4792 *
4793 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4794 */
4795int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4796	unsigned long *pfn)
4797{
4798	int ret = -EINVAL;
4799	spinlock_t *ptl;
4800	pte_t *ptep;
4801
4802	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4803		return ret;
4804
4805	ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4806	if (ret)
4807		return ret;
4808	*pfn = pte_pfn(*ptep);
4809	pte_unmap_unlock(ptep, ptl);
4810	return 0;
4811}
4812EXPORT_SYMBOL(follow_pfn);
4813
4814#ifdef CONFIG_HAVE_IOREMAP_PROT
4815int follow_phys(struct vm_area_struct *vma,
4816		unsigned long address, unsigned int flags,
4817		unsigned long *prot, resource_size_t *phys)
4818{
4819	int ret = -EINVAL;
4820	pte_t *ptep, pte;
4821	spinlock_t *ptl;
4822
4823	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4824		goto out;
4825
4826	if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4827		goto out;
4828	pte = *ptep;
4829
4830	if ((flags & FOLL_WRITE) && !pte_write(pte))
4831		goto unlock;
4832
4833	*prot = pgprot_val(pte_pgprot(pte));
4834	*phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4835
4836	ret = 0;
4837unlock:
4838	pte_unmap_unlock(ptep, ptl);
4839out:
4840	return ret;
4841}
4842
4843int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4844			void *buf, int len, int write)
4845{
4846	resource_size_t phys_addr;
4847	unsigned long prot = 0;
4848	void __iomem *maddr;
4849	int offset = addr & (PAGE_SIZE-1);
4850
4851	if (follow_phys(vma, addr, write, &prot, &phys_addr))
4852		return -EINVAL;
4853
4854	maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4855	if (!maddr)
4856		return -ENOMEM;
4857
4858	if (write)
4859		memcpy_toio(maddr + offset, buf, len);
4860	else
4861		memcpy_fromio(buf, maddr + offset, len);
4862	iounmap(maddr);
4863
4864	return len;
4865}
4866EXPORT_SYMBOL_GPL(generic_access_phys);
4867#endif
4868
4869/*
4870 * Access another process' address space as given in mm.  If non-NULL, use the
4871 * given task for page fault accounting.
4872 */
4873int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4874		unsigned long addr, void *buf, int len, unsigned int gup_flags)
4875{
4876	struct vm_area_struct *vma;
4877	void *old_buf = buf;
4878	int write = gup_flags & FOLL_WRITE;
4879
4880	if (mmap_read_lock_killable(mm))
4881		return 0;
4882
4883	/* ignore errors, just check how much was successfully transferred */
4884	while (len) {
4885		int bytes, ret, offset;
4886		void *maddr;
4887		struct page *page = NULL;
4888
4889		ret = get_user_pages_remote(mm, addr, 1,
4890				gup_flags, &page, &vma, NULL);
4891		if (ret <= 0) {
4892#ifndef CONFIG_HAVE_IOREMAP_PROT
4893			break;
4894#else
4895			/*
4896			 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4897			 * we can access using slightly different code.
4898			 */
4899			vma = find_vma(mm, addr);
4900			if (!vma || vma->vm_start > addr)
4901				break;
4902			if (vma->vm_ops && vma->vm_ops->access)
4903				ret = vma->vm_ops->access(vma, addr, buf,
4904							  len, write);
4905			if (ret <= 0)
4906				break;
4907			bytes = ret;
4908#endif
4909		} else {
4910			bytes = len;
4911			offset = addr & (PAGE_SIZE-1);
4912			if (bytes > PAGE_SIZE-offset)
4913				bytes = PAGE_SIZE-offset;
4914
4915			maddr = kmap(page);
4916			if (write) {
4917				copy_to_user_page(vma, page, addr,
4918						  maddr + offset, buf, bytes);
4919				set_page_dirty_lock(page);
4920			} else {
4921				copy_from_user_page(vma, page, addr,
4922						    buf, maddr + offset, bytes);
4923			}
4924			kunmap(page);
4925			put_page(page);
4926		}
4927		len -= bytes;
4928		buf += bytes;
4929		addr += bytes;
4930	}
4931	mmap_read_unlock(mm);
4932
4933	return buf - old_buf;
4934}
4935
4936/**
4937 * access_remote_vm - access another process' address space
4938 * @mm:		the mm_struct of the target address space
4939 * @addr:	start address to access
4940 * @buf:	source or destination buffer
4941 * @len:	number of bytes to transfer
4942 * @gup_flags:	flags modifying lookup behaviour
4943 *
4944 * The caller must hold a reference on @mm.
4945 *
4946 * Return: number of bytes copied from source to destination.
4947 */
4948int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4949		void *buf, int len, unsigned int gup_flags)
4950{
4951	return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4952}
4953
4954/*
4955 * Access another process' address space.
4956 * Source/target buffer must be kernel space,
4957 * Do not walk the page table directly, use get_user_pages
4958 */
4959int access_process_vm(struct task_struct *tsk, unsigned long addr,
4960		void *buf, int len, unsigned int gup_flags)
4961{
4962	struct mm_struct *mm;
4963	int ret;
4964
4965	mm = get_task_mm(tsk);
4966	if (!mm)
4967		return 0;
4968
4969	ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4970
4971	mmput(mm);
4972
4973	return ret;
4974}
4975EXPORT_SYMBOL_GPL(access_process_vm);
4976
4977/*
4978 * Print the name of a VMA.
4979 */
4980void print_vma_addr(char *prefix, unsigned long ip)
4981{
4982	struct mm_struct *mm = current->mm;
4983	struct vm_area_struct *vma;
4984
4985	/*
4986	 * we might be running from an atomic context so we cannot sleep
4987	 */
4988	if (!mmap_read_trylock(mm))
4989		return;
4990
4991	vma = find_vma(mm, ip);
4992	if (vma && vma->vm_file) {
4993		struct file *f = vma->vm_file;
4994		char *buf = (char *)__get_free_page(GFP_NOWAIT);
4995		if (buf) {
4996			char *p;
4997
4998			p = file_path(f, buf, PAGE_SIZE);
4999			if (IS_ERR(p))
5000				p = "?";
5001			printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5002					vma->vm_start,
5003					vma->vm_end - vma->vm_start);
5004			free_page((unsigned long)buf);
5005		}
5006	}
5007	mmap_read_unlock(mm);
5008}
5009
5010#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5011void __might_fault(const char *file, int line)
5012{
5013	/*
5014	 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5015	 * holding the mmap_lock, this is safe because kernel memory doesn't
5016	 * get paged out, therefore we'll never actually fault, and the
5017	 * below annotations will generate false positives.
5018	 */
5019	if (uaccess_kernel())
5020		return;
5021	if (pagefault_disabled())
5022		return;
5023	__might_sleep(file, line, 0);
5024#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5025	if (current->mm)
5026		might_lock_read(&current->mm->mmap_lock);
5027#endif
5028}
5029EXPORT_SYMBOL(__might_fault);
5030#endif
5031
5032#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5033/*
5034 * Process all subpages of the specified huge page with the specified
5035 * operation.  The target subpage will be processed last to keep its
5036 * cache lines hot.
5037 */
5038static inline void process_huge_page(
5039	unsigned long addr_hint, unsigned int pages_per_huge_page,
5040	void (*process_subpage)(unsigned long addr, int idx, void *arg),
5041	void *arg)
5042{
5043	int i, n, base, l;
5044	unsigned long addr = addr_hint &
5045		~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5046
5047	/* Process target subpage last to keep its cache lines hot */
5048	might_sleep();
5049	n = (addr_hint - addr) / PAGE_SIZE;
5050	if (2 * n <= pages_per_huge_page) {
5051		/* If target subpage in first half of huge page */
5052		base = 0;
5053		l = n;
5054		/* Process subpages at the end of huge page */
5055		for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5056			cond_resched();
5057			process_subpage(addr + i * PAGE_SIZE, i, arg);
5058		}
5059	} else {
5060		/* If target subpage in second half of huge page */
5061		base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5062		l = pages_per_huge_page - n;
5063		/* Process subpages at the begin of huge page */
5064		for (i = 0; i < base; i++) {
5065			cond_resched();
5066			process_subpage(addr + i * PAGE_SIZE, i, arg);
5067		}
5068	}
5069	/*
5070	 * Process remaining subpages in left-right-left-right pattern
5071	 * towards the target subpage
5072	 */
5073	for (i = 0; i < l; i++) {
5074		int left_idx = base + i;
5075		int right_idx = base + 2 * l - 1 - i;
5076
5077		cond_resched();
5078		process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5079		cond_resched();
5080		process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5081	}
5082}
5083
5084static void clear_gigantic_page(struct page *page,
5085				unsigned long addr,
5086				unsigned int pages_per_huge_page)
5087{
5088	int i;
5089	struct page *p = page;
5090
5091	might_sleep();
5092	for (i = 0; i < pages_per_huge_page;
5093	     i++, p = mem_map_next(p, page, i)) {
5094		cond_resched();
5095		clear_user_highpage(p, addr + i * PAGE_SIZE);
5096	}
5097}
5098
5099static void clear_subpage(unsigned long addr, int idx, void *arg)
5100{
5101	struct page *page = arg;
5102
5103	clear_user_highpage(page + idx, addr);
5104}
5105
5106void clear_huge_page(struct page *page,
5107		     unsigned long addr_hint, unsigned int pages_per_huge_page)
5108{
5109	unsigned long addr = addr_hint &
5110		~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5111
5112	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5113		clear_gigantic_page(page, addr, pages_per_huge_page);
5114		return;
5115	}
5116
5117	process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5118}
5119
5120static void copy_user_gigantic_page(struct page *dst, struct page *src,
5121				    unsigned long addr,
5122				    struct vm_area_struct *vma,
5123				    unsigned int pages_per_huge_page)
5124{
5125	int i;
5126	struct page *dst_base = dst;
5127	struct page *src_base = src;
5128
5129	for (i = 0; i < pages_per_huge_page; ) {
5130		cond_resched();
5131		copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5132
5133		i++;
5134		dst = mem_map_next(dst, dst_base, i);
5135		src = mem_map_next(src, src_base, i);
5136	}
5137}
5138
5139struct copy_subpage_arg {
5140	struct page *dst;
5141	struct page *src;
5142	struct vm_area_struct *vma;
5143};
5144
5145static void copy_subpage(unsigned long addr, int idx, void *arg)
5146{
5147	struct copy_subpage_arg *copy_arg = arg;
5148
5149	copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5150			   addr, copy_arg->vma);
5151}
5152
5153void copy_user_huge_page(struct page *dst, struct page *src,
5154			 unsigned long addr_hint, struct vm_area_struct *vma,
5155			 unsigned int pages_per_huge_page)
5156{
5157	unsigned long addr = addr_hint &
5158		~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5159	struct copy_subpage_arg arg = {
5160		.dst = dst,
5161		.src = src,
5162		.vma = vma,
5163	};
5164
5165	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5166		copy_user_gigantic_page(dst, src, addr, vma,
5167					pages_per_huge_page);
5168		return;
5169	}
5170
5171	process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5172}
5173
5174long copy_huge_page_from_user(struct page *dst_page,
5175				const void __user *usr_src,
5176				unsigned int pages_per_huge_page,
5177				bool allow_pagefault)
5178{
5179	void *src = (void *)usr_src;
5180	void *page_kaddr;
5181	unsigned long i, rc = 0;
5182	unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5183
5184	for (i = 0; i < pages_per_huge_page; i++) {
5185		if (allow_pagefault)
5186			page_kaddr = kmap(dst_page + i);
5187		else
5188			page_kaddr = kmap_atomic(dst_page + i);
5189		rc = copy_from_user(page_kaddr,
5190				(const void __user *)(src + i * PAGE_SIZE),
5191				PAGE_SIZE);
5192		if (allow_pagefault)
5193			kunmap(dst_page + i);
5194		else
5195			kunmap_atomic(page_kaddr);
5196
5197		ret_val -= (PAGE_SIZE - rc);
5198		if (rc)
5199			break;
5200
5201		cond_resched();
5202	}
5203	return ret_val;
5204}
5205#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5206
5207#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5208
5209static struct kmem_cache *page_ptl_cachep;
5210
5211void __init ptlock_cache_init(void)
5212{
5213	page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5214			SLAB_PANIC, NULL);
5215}
5216
5217bool ptlock_alloc(struct page *page)
5218{
5219	spinlock_t *ptl;
5220
5221	ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5222	if (!ptl)
5223		return false;
5224	page->ptl = ptl;
5225	return true;
5226}
5227
5228void ptlock_free(struct page *page)
5229{
5230	kmem_cache_free(page_ptl_cachep, page->ptl);
5231}
5232#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