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
v4.17
 
   1/*
   2 *  linux/mm/memory.c
   3 *
   4 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5 */
   6
   7/*
   8 * demand-loading started 01.12.91 - seems it is high on the list of
   9 * things wanted, and it should be easy to implement. - Linus
  10 */
  11
  12/*
  13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
  14 * pages started 02.12.91, seems to work. - Linus.
  15 *
  16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
  17 * would have taken more than the 6M I have free, but it worked well as
  18 * far as I could see.
  19 *
  20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
  21 */
  22
  23/*
  24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
  25 * thought has to go into this. Oh, well..
  26 * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
  27 *		Found it. Everything seems to work now.
  28 * 20.12.91  -  Ok, making the swap-device changeable like the root.
  29 */
  30
  31/*
  32 * 05.04.94  -  Multi-page memory management added for v1.1.
  33 *              Idea by Alex Bligh (alex@cconcepts.co.uk)
  34 *
  35 * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
  36 *		(Gerhard.Wichert@pdb.siemens.de)
  37 *
  38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
  39 */
  40
  41#include <linux/kernel_stat.h>
  42#include <linux/mm.h>
  43#include <linux/sched/mm.h>
  44#include <linux/sched/coredump.h>
  45#include <linux/sched/numa_balancing.h>
  46#include <linux/sched/task.h>
  47#include <linux/hugetlb.h>
  48#include <linux/mman.h>
  49#include <linux/swap.h>
  50#include <linux/highmem.h>
  51#include <linux/pagemap.h>
  52#include <linux/memremap.h>
  53#include <linux/ksm.h>
  54#include <linux/rmap.h>
  55#include <linux/export.h>
  56#include <linux/delayacct.h>
  57#include <linux/init.h>
  58#include <linux/pfn_t.h>
  59#include <linux/writeback.h>
  60#include <linux/memcontrol.h>
  61#include <linux/mmu_notifier.h>
  62#include <linux/swapops.h>
  63#include <linux/elf.h>
  64#include <linux/gfp.h>
  65#include <linux/migrate.h>
  66#include <linux/string.h>
  67#include <linux/dma-debug.h>
  68#include <linux/debugfs.h>
  69#include <linux/userfaultfd_k.h>
  70#include <linux/dax.h>
  71#include <linux/oom.h>
 
 
 
 
 
 
  72
  73#include <asm/io.h>
  74#include <asm/mmu_context.h>
  75#include <asm/pgalloc.h>
  76#include <linux/uaccess.h>
  77#include <asm/tlb.h>
  78#include <asm/tlbflush.h>
  79#include <asm/pgtable.h>
  80
 
  81#include "internal.h"
  82
  83#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
  84#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
  85#endif
  86
  87#ifndef CONFIG_NEED_MULTIPLE_NODES
  88/* use the per-pgdat data instead for discontigmem - mbligh */
  89unsigned long max_mapnr;
  90EXPORT_SYMBOL(max_mapnr);
  91
  92struct page *mem_map;
  93EXPORT_SYMBOL(mem_map);
  94#endif
  95
  96/*
  97 * A number of key systems in x86 including ioremap() rely on the assumption
  98 * that high_memory defines the upper bound on direct map memory, then end
  99 * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
 100 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
 101 * and ZONE_HIGHMEM.
 102 */
 103void *high_memory;
 104EXPORT_SYMBOL(high_memory);
 105
 106/*
 107 * Randomize the address space (stacks, mmaps, brk, etc.).
 108 *
 109 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
 110 *   as ancient (libc5 based) binaries can segfault. )
 111 */
 112int randomize_va_space __read_mostly =
 113#ifdef CONFIG_COMPAT_BRK
 114					1;
 115#else
 116					2;
 117#endif
 118
 
 
 
 
 
 
 
 
 
 
 
 
 119static int __init disable_randmaps(char *s)
 120{
 121	randomize_va_space = 0;
 122	return 1;
 123}
 124__setup("norandmaps", disable_randmaps);
 125
 126unsigned long zero_pfn __read_mostly;
 127EXPORT_SYMBOL(zero_pfn);
 128
 129unsigned long highest_memmap_pfn __read_mostly;
 130
 131/*
 132 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
 133 */
 134static int __init init_zero_pfn(void)
 135{
 136	zero_pfn = page_to_pfn(ZERO_PAGE(0));
 137	return 0;
 138}
 139core_initcall(init_zero_pfn);
 140
 
 
 
 
 141
 142#if defined(SPLIT_RSS_COUNTING)
 143
 144void sync_mm_rss(struct mm_struct *mm)
 145{
 146	int i;
 147
 148	for (i = 0; i < NR_MM_COUNTERS; i++) {
 149		if (current->rss_stat.count[i]) {
 150			add_mm_counter(mm, i, current->rss_stat.count[i]);
 151			current->rss_stat.count[i] = 0;
 152		}
 153	}
 154	current->rss_stat.events = 0;
 155}
 156
 157static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
 158{
 159	struct task_struct *task = current;
 160
 161	if (likely(task->mm == mm))
 162		task->rss_stat.count[member] += val;
 163	else
 164		add_mm_counter(mm, member, val);
 165}
 166#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
 167#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
 168
 169/* sync counter once per 64 page faults */
 170#define TASK_RSS_EVENTS_THRESH	(64)
 171static void check_sync_rss_stat(struct task_struct *task)
 172{
 173	if (unlikely(task != current))
 174		return;
 175	if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
 176		sync_mm_rss(task->mm);
 177}
 178#else /* SPLIT_RSS_COUNTING */
 179
 180#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
 181#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
 182
 183static void check_sync_rss_stat(struct task_struct *task)
 184{
 185}
 186
 187#endif /* SPLIT_RSS_COUNTING */
 188
 189#ifdef HAVE_GENERIC_MMU_GATHER
 190
 191static bool tlb_next_batch(struct mmu_gather *tlb)
 192{
 193	struct mmu_gather_batch *batch;
 194
 195	batch = tlb->active;
 196	if (batch->next) {
 197		tlb->active = batch->next;
 198		return true;
 199	}
 200
 201	if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
 202		return false;
 203
 204	batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
 205	if (!batch)
 206		return false;
 207
 208	tlb->batch_count++;
 209	batch->next = NULL;
 210	batch->nr   = 0;
 211	batch->max  = MAX_GATHER_BATCH;
 212
 213	tlb->active->next = batch;
 214	tlb->active = batch;
 215
 216	return true;
 217}
 218
 219void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
 220				unsigned long start, unsigned long end)
 221{
 222	tlb->mm = mm;
 223
 224	/* Is it from 0 to ~0? */
 225	tlb->fullmm     = !(start | (end+1));
 226	tlb->need_flush_all = 0;
 227	tlb->local.next = NULL;
 228	tlb->local.nr   = 0;
 229	tlb->local.max  = ARRAY_SIZE(tlb->__pages);
 230	tlb->active     = &tlb->local;
 231	tlb->batch_count = 0;
 232
 233#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 234	tlb->batch = NULL;
 235#endif
 236	tlb->page_size = 0;
 237
 238	__tlb_reset_range(tlb);
 239}
 240
 241static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
 242{
 243	if (!tlb->end)
 244		return;
 245
 246	tlb_flush(tlb);
 247	mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
 248#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 249	tlb_table_flush(tlb);
 250#endif
 251	__tlb_reset_range(tlb);
 252}
 253
 254static void tlb_flush_mmu_free(struct mmu_gather *tlb)
 255{
 256	struct mmu_gather_batch *batch;
 257
 258	for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
 259		free_pages_and_swap_cache(batch->pages, batch->nr);
 260		batch->nr = 0;
 261	}
 262	tlb->active = &tlb->local;
 263}
 264
 265void tlb_flush_mmu(struct mmu_gather *tlb)
 266{
 267	tlb_flush_mmu_tlbonly(tlb);
 268	tlb_flush_mmu_free(tlb);
 269}
 270
 271/* tlb_finish_mmu
 272 *	Called at the end of the shootdown operation to free up any resources
 273 *	that were required.
 274 */
 275void arch_tlb_finish_mmu(struct mmu_gather *tlb,
 276		unsigned long start, unsigned long end, bool force)
 277{
 278	struct mmu_gather_batch *batch, *next;
 279
 280	if (force)
 281		__tlb_adjust_range(tlb, start, end - start);
 282
 283	tlb_flush_mmu(tlb);
 284
 285	/* keep the page table cache within bounds */
 286	check_pgt_cache();
 287
 288	for (batch = tlb->local.next; batch; batch = next) {
 289		next = batch->next;
 290		free_pages((unsigned long)batch, 0);
 291	}
 292	tlb->local.next = NULL;
 293}
 294
 295/* __tlb_remove_page
 296 *	Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
 297 *	handling the additional races in SMP caused by other CPUs caching valid
 298 *	mappings in their TLBs. Returns the number of free page slots left.
 299 *	When out of page slots we must call tlb_flush_mmu().
 300 *returns true if the caller should flush.
 301 */
 302bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
 303{
 304	struct mmu_gather_batch *batch;
 305
 306	VM_BUG_ON(!tlb->end);
 307	VM_WARN_ON(tlb->page_size != page_size);
 308
 309	batch = tlb->active;
 310	/*
 311	 * Add the page and check if we are full. If so
 312	 * force a flush.
 313	 */
 314	batch->pages[batch->nr++] = page;
 315	if (batch->nr == batch->max) {
 316		if (!tlb_next_batch(tlb))
 317			return true;
 318		batch = tlb->active;
 319	}
 320	VM_BUG_ON_PAGE(batch->nr > batch->max, page);
 321
 322	return false;
 323}
 324
 325#endif /* HAVE_GENERIC_MMU_GATHER */
 326
 327#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 328
 329/*
 330 * See the comment near struct mmu_table_batch.
 331 */
 332
 333static void tlb_remove_table_smp_sync(void *arg)
 334{
 335	/* Simply deliver the interrupt */
 336}
 337
 338static void tlb_remove_table_one(void *table)
 339{
 340	/*
 341	 * This isn't an RCU grace period and hence the page-tables cannot be
 342	 * assumed to be actually RCU-freed.
 343	 *
 344	 * It is however sufficient for software page-table walkers that rely on
 345	 * IRQ disabling. See the comment near struct mmu_table_batch.
 346	 */
 347	smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
 348	__tlb_remove_table(table);
 349}
 350
 351static void tlb_remove_table_rcu(struct rcu_head *head)
 352{
 353	struct mmu_table_batch *batch;
 354	int i;
 355
 356	batch = container_of(head, struct mmu_table_batch, rcu);
 357
 358	for (i = 0; i < batch->nr; i++)
 359		__tlb_remove_table(batch->tables[i]);
 360
 361	free_page((unsigned long)batch);
 362}
 363
 364void tlb_table_flush(struct mmu_gather *tlb)
 365{
 366	struct mmu_table_batch **batch = &tlb->batch;
 367
 368	if (*batch) {
 369		call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
 370		*batch = NULL;
 371	}
 372}
 373
 374void tlb_remove_table(struct mmu_gather *tlb, void *table)
 375{
 376	struct mmu_table_batch **batch = &tlb->batch;
 377
 378	/*
 379	 * When there's less then two users of this mm there cannot be a
 380	 * concurrent page-table walk.
 381	 */
 382	if (atomic_read(&tlb->mm->mm_users) < 2) {
 383		__tlb_remove_table(table);
 384		return;
 385	}
 386
 387	if (*batch == NULL) {
 388		*batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
 389		if (*batch == NULL) {
 390			tlb_remove_table_one(table);
 391			return;
 392		}
 393		(*batch)->nr = 0;
 394	}
 395	(*batch)->tables[(*batch)->nr++] = table;
 396	if ((*batch)->nr == MAX_TABLE_BATCH)
 397		tlb_table_flush(tlb);
 398}
 399
 400#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
 401
 402/**
 403 * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
 404 * @tlb: the mmu_gather structure to initialize
 405 * @mm: the mm_struct of the target address space
 406 * @start: start of the region that will be removed from the page-table
 407 * @end: end of the region that will be removed from the page-table
 408 *
 409 * Called to initialize an (on-stack) mmu_gather structure for page-table
 410 * tear-down from @mm. The @start and @end are set to 0 and -1
 411 * respectively when @mm is without users and we're going to destroy
 412 * the full address space (exit/execve).
 413 */
 414void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
 415			unsigned long start, unsigned long end)
 416{
 417	arch_tlb_gather_mmu(tlb, mm, start, end);
 418	inc_tlb_flush_pending(tlb->mm);
 419}
 420
 421void tlb_finish_mmu(struct mmu_gather *tlb,
 422		unsigned long start, unsigned long end)
 423{
 424	/*
 425	 * If there are parallel threads are doing PTE changes on same range
 426	 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
 427	 * flush by batching, a thread has stable TLB entry can fail to flush
 428	 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
 429	 * forcefully if we detect parallel PTE batching threads.
 430	 */
 431	bool force = mm_tlb_flush_nested(tlb->mm);
 432
 433	arch_tlb_finish_mmu(tlb, start, end, force);
 434	dec_tlb_flush_pending(tlb->mm);
 435}
 436
 437/*
 438 * Note: this doesn't free the actual pages themselves. That
 439 * has been handled earlier when unmapping all the memory regions.
 440 */
 441static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 442			   unsigned long addr)
 443{
 444	pgtable_t token = pmd_pgtable(*pmd);
 445	pmd_clear(pmd);
 446	pte_free_tlb(tlb, token, addr);
 447	mm_dec_nr_ptes(tlb->mm);
 448}
 449
 450static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 451				unsigned long addr, unsigned long end,
 452				unsigned long floor, unsigned long ceiling)
 453{
 454	pmd_t *pmd;
 455	unsigned long next;
 456	unsigned long start;
 457
 458	start = addr;
 459	pmd = pmd_offset(pud, addr);
 460	do {
 461		next = pmd_addr_end(addr, end);
 462		if (pmd_none_or_clear_bad(pmd))
 463			continue;
 464		free_pte_range(tlb, pmd, addr);
 465	} while (pmd++, addr = next, addr != end);
 466
 467	start &= PUD_MASK;
 468	if (start < floor)
 469		return;
 470	if (ceiling) {
 471		ceiling &= PUD_MASK;
 472		if (!ceiling)
 473			return;
 474	}
 475	if (end - 1 > ceiling - 1)
 476		return;
 477
 478	pmd = pmd_offset(pud, start);
 479	pud_clear(pud);
 480	pmd_free_tlb(tlb, pmd, start);
 481	mm_dec_nr_pmds(tlb->mm);
 482}
 483
 484static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
 485				unsigned long addr, unsigned long end,
 486				unsigned long floor, unsigned long ceiling)
 487{
 488	pud_t *pud;
 489	unsigned long next;
 490	unsigned long start;
 491
 492	start = addr;
 493	pud = pud_offset(p4d, addr);
 494	do {
 495		next = pud_addr_end(addr, end);
 496		if (pud_none_or_clear_bad(pud))
 497			continue;
 498		free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 499	} while (pud++, addr = next, addr != end);
 500
 501	start &= P4D_MASK;
 502	if (start < floor)
 503		return;
 504	if (ceiling) {
 505		ceiling &= P4D_MASK;
 506		if (!ceiling)
 507			return;
 508	}
 509	if (end - 1 > ceiling - 1)
 510		return;
 511
 512	pud = pud_offset(p4d, start);
 513	p4d_clear(p4d);
 514	pud_free_tlb(tlb, pud, start);
 515	mm_dec_nr_puds(tlb->mm);
 516}
 517
 518static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
 519				unsigned long addr, unsigned long end,
 520				unsigned long floor, unsigned long ceiling)
 521{
 522	p4d_t *p4d;
 523	unsigned long next;
 524	unsigned long start;
 525
 526	start = addr;
 527	p4d = p4d_offset(pgd, addr);
 528	do {
 529		next = p4d_addr_end(addr, end);
 530		if (p4d_none_or_clear_bad(p4d))
 531			continue;
 532		free_pud_range(tlb, p4d, addr, next, floor, ceiling);
 533	} while (p4d++, addr = next, addr != end);
 534
 535	start &= PGDIR_MASK;
 536	if (start < floor)
 537		return;
 538	if (ceiling) {
 539		ceiling &= PGDIR_MASK;
 540		if (!ceiling)
 541			return;
 542	}
 543	if (end - 1 > ceiling - 1)
 544		return;
 545
 546	p4d = p4d_offset(pgd, start);
 547	pgd_clear(pgd);
 548	p4d_free_tlb(tlb, p4d, start);
 549}
 550
 551/*
 552 * This function frees user-level page tables of a process.
 553 */
 554void free_pgd_range(struct mmu_gather *tlb,
 555			unsigned long addr, unsigned long end,
 556			unsigned long floor, unsigned long ceiling)
 557{
 558	pgd_t *pgd;
 559	unsigned long next;
 560
 561	/*
 562	 * The next few lines have given us lots of grief...
 563	 *
 564	 * Why are we testing PMD* at this top level?  Because often
 565	 * there will be no work to do at all, and we'd prefer not to
 566	 * go all the way down to the bottom just to discover that.
 567	 *
 568	 * Why all these "- 1"s?  Because 0 represents both the bottom
 569	 * of the address space and the top of it (using -1 for the
 570	 * top wouldn't help much: the masks would do the wrong thing).
 571	 * The rule is that addr 0 and floor 0 refer to the bottom of
 572	 * the address space, but end 0 and ceiling 0 refer to the top
 573	 * Comparisons need to use "end - 1" and "ceiling - 1" (though
 574	 * that end 0 case should be mythical).
 575	 *
 576	 * Wherever addr is brought up or ceiling brought down, we must
 577	 * be careful to reject "the opposite 0" before it confuses the
 578	 * subsequent tests.  But what about where end is brought down
 579	 * by PMD_SIZE below? no, end can't go down to 0 there.
 580	 *
 581	 * Whereas we round start (addr) and ceiling down, by different
 582	 * masks at different levels, in order to test whether a table
 583	 * now has no other vmas using it, so can be freed, we don't
 584	 * bother to round floor or end up - the tests don't need that.
 585	 */
 586
 587	addr &= PMD_MASK;
 588	if (addr < floor) {
 589		addr += PMD_SIZE;
 590		if (!addr)
 591			return;
 592	}
 593	if (ceiling) {
 594		ceiling &= PMD_MASK;
 595		if (!ceiling)
 596			return;
 597	}
 598	if (end - 1 > ceiling - 1)
 599		end -= PMD_SIZE;
 600	if (addr > end - 1)
 601		return;
 602	/*
 603	 * We add page table cache pages with PAGE_SIZE,
 604	 * (see pte_free_tlb()), flush the tlb if we need
 605	 */
 606	tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
 607	pgd = pgd_offset(tlb->mm, addr);
 608	do {
 609		next = pgd_addr_end(addr, end);
 610		if (pgd_none_or_clear_bad(pgd))
 611			continue;
 612		free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
 613	} while (pgd++, addr = next, addr != end);
 614}
 615
 616void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
 617		unsigned long floor, unsigned long ceiling)
 618{
 619	while (vma) {
 620		struct vm_area_struct *next = vma->vm_next;
 621		unsigned long addr = vma->vm_start;
 622
 623		/*
 624		 * Hide vma from rmap and truncate_pagecache before freeing
 625		 * pgtables
 626		 */
 627		unlink_anon_vmas(vma);
 628		unlink_file_vma(vma);
 629
 630		if (is_vm_hugetlb_page(vma)) {
 631			hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 632				floor, next ? next->vm_start : ceiling);
 633		} else {
 634			/*
 635			 * Optimization: gather nearby vmas into one call down
 636			 */
 637			while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 638			       && !is_vm_hugetlb_page(next)) {
 639				vma = next;
 640				next = vma->vm_next;
 641				unlink_anon_vmas(vma);
 642				unlink_file_vma(vma);
 643			}
 644			free_pgd_range(tlb, addr, vma->vm_end,
 645				floor, next ? next->vm_start : ceiling);
 646		}
 647		vma = next;
 648	}
 649}
 650
 651int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 652{
 653	spinlock_t *ptl;
 654	pgtable_t new = pte_alloc_one(mm, address);
 655	if (!new)
 656		return -ENOMEM;
 657
 658	/*
 659	 * Ensure all pte setup (eg. pte page lock and page clearing) are
 660	 * visible before the pte is made visible to other CPUs by being
 661	 * put into page tables.
 662	 *
 663	 * The other side of the story is the pointer chasing in the page
 664	 * table walking code (when walking the page table without locking;
 665	 * ie. most of the time). Fortunately, these data accesses consist
 666	 * of a chain of data-dependent loads, meaning most CPUs (alpha
 667	 * being the notable exception) will already guarantee loads are
 668	 * seen in-order. See the alpha page table accessors for the
 669	 * smp_read_barrier_depends() barriers in page table walking code.
 670	 */
 671	smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
 672
 673	ptl = pmd_lock(mm, pmd);
 674	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
 675		mm_inc_nr_ptes(mm);
 676		pmd_populate(mm, pmd, new);
 677		new = NULL;
 678	}
 679	spin_unlock(ptl);
 680	if (new)
 681		pte_free(mm, new);
 682	return 0;
 683}
 684
 685int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
 686{
 687	pte_t *new = pte_alloc_one_kernel(&init_mm, address);
 688	if (!new)
 689		return -ENOMEM;
 690
 691	smp_wmb(); /* See comment in __pte_alloc */
 692
 693	spin_lock(&init_mm.page_table_lock);
 694	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
 695		pmd_populate_kernel(&init_mm, pmd, new);
 696		new = NULL;
 697	}
 698	spin_unlock(&init_mm.page_table_lock);
 699	if (new)
 700		pte_free_kernel(&init_mm, new);
 701	return 0;
 702}
 703
 704static inline void init_rss_vec(int *rss)
 705{
 706	memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
 707}
 708
 709static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
 710{
 711	int i;
 712
 713	if (current->mm == mm)
 714		sync_mm_rss(mm);
 715	for (i = 0; i < NR_MM_COUNTERS; i++)
 716		if (rss[i])
 717			add_mm_counter(mm, i, rss[i]);
 718}
 719
 720/*
 721 * This function is called to print an error when a bad pte
 722 * is found. For example, we might have a PFN-mapped pte in
 723 * a region that doesn't allow it.
 724 *
 725 * The calling function must still handle the error.
 726 */
 727static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
 728			  pte_t pte, struct page *page)
 729{
 730	pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
 731	p4d_t *p4d = p4d_offset(pgd, addr);
 732	pud_t *pud = pud_offset(p4d, addr);
 733	pmd_t *pmd = pmd_offset(pud, addr);
 734	struct address_space *mapping;
 735	pgoff_t index;
 736	static unsigned long resume;
 737	static unsigned long nr_shown;
 738	static unsigned long nr_unshown;
 739
 740	/*
 741	 * Allow a burst of 60 reports, then keep quiet for that minute;
 742	 * or allow a steady drip of one report per second.
 743	 */
 744	if (nr_shown == 60) {
 745		if (time_before(jiffies, resume)) {
 746			nr_unshown++;
 747			return;
 748		}
 749		if (nr_unshown) {
 750			pr_alert("BUG: Bad page map: %lu messages suppressed\n",
 751				 nr_unshown);
 752			nr_unshown = 0;
 753		}
 754		nr_shown = 0;
 755	}
 756	if (nr_shown++ == 0)
 757		resume = jiffies + 60 * HZ;
 758
 759	mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
 760	index = linear_page_index(vma, addr);
 761
 762	pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
 763		 current->comm,
 764		 (long long)pte_val(pte), (long long)pmd_val(*pmd));
 765	if (page)
 766		dump_page(page, "bad pte");
 767	pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
 768		 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
 769	pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
 770		 vma->vm_file,
 771		 vma->vm_ops ? vma->vm_ops->fault : NULL,
 772		 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
 773		 mapping ? mapping->a_ops->readpage : NULL);
 774	dump_stack();
 775	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 776}
 777
 778/*
 779 * vm_normal_page -- This function gets the "struct page" associated with a pte.
 780 *
 781 * "Special" mappings do not wish to be associated with a "struct page" (either
 782 * it doesn't exist, or it exists but they don't want to touch it). In this
 783 * case, NULL is returned here. "Normal" mappings do have a struct page.
 784 *
 785 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
 786 * pte bit, in which case this function is trivial. Secondly, an architecture
 787 * may not have a spare pte bit, which requires a more complicated scheme,
 788 * described below.
 789 *
 790 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
 791 * special mapping (even if there are underlying and valid "struct pages").
 792 * COWed pages of a VM_PFNMAP are always normal.
 793 *
 794 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
 795 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
 796 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
 797 * mapping will always honor the rule
 798 *
 799 *	pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 800 *
 801 * And for normal mappings this is false.
 802 *
 803 * This restricts such mappings to be a linear translation from virtual address
 804 * to pfn. To get around this restriction, we allow arbitrary mappings so long
 805 * as the vma is not a COW mapping; in that case, we know that all ptes are
 806 * special (because none can have been COWed).
 807 *
 808 *
 809 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
 810 *
 811 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
 812 * page" backing, however the difference is that _all_ pages with a struct
 813 * page (that is, those where pfn_valid is true) are refcounted and considered
 814 * normal pages by the VM. The disadvantage is that pages are refcounted
 815 * (which can be slower and simply not an option for some PFNMAP users). The
 816 * advantage is that we don't have to follow the strict linearity rule of
 817 * PFNMAP mappings in order to support COWable mappings.
 818 *
 819 */
 820#ifdef __HAVE_ARCH_PTE_SPECIAL
 821# define HAVE_PTE_SPECIAL 1
 822#else
 823# define HAVE_PTE_SPECIAL 0
 824#endif
 825struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 826			     pte_t pte, bool with_public_device)
 827{
 828	unsigned long pfn = pte_pfn(pte);
 829
 830	if (HAVE_PTE_SPECIAL) {
 831		if (likely(!pte_special(pte)))
 832			goto check_pfn;
 833		if (vma->vm_ops && vma->vm_ops->find_special_page)
 834			return vma->vm_ops->find_special_page(vma, addr);
 835		if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
 836			return NULL;
 837		if (is_zero_pfn(pfn))
 838			return NULL;
 
 
 839
 840		/*
 841		 * Device public pages are special pages (they are ZONE_DEVICE
 842		 * pages but different from persistent memory). They behave
 843		 * allmost like normal pages. The difference is that they are
 844		 * not on the lru and thus should never be involve with any-
 845		 * thing that involve lru manipulation (mlock, numa balancing,
 846		 * ...).
 847		 *
 848		 * This is why we still want to return NULL for such page from
 849		 * vm_normal_page() so that we do not have to special case all
 850		 * call site of vm_normal_page().
 851		 */
 852		if (likely(pfn <= highest_memmap_pfn)) {
 853			struct page *page = pfn_to_page(pfn);
 854
 855			if (is_device_public_page(page)) {
 856				if (with_public_device)
 857					return page;
 858				return NULL;
 859			}
 860		}
 861		print_bad_pte(vma, addr, pte, NULL);
 862		return NULL;
 863	}
 864
 865	/* !HAVE_PTE_SPECIAL case follows: */
 866
 867	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 868		if (vma->vm_flags & VM_MIXEDMAP) {
 869			if (!pfn_valid(pfn))
 870				return NULL;
 871			goto out;
 872		} else {
 873			unsigned long off;
 874			off = (addr - vma->vm_start) >> PAGE_SHIFT;
 875			if (pfn == vma->vm_pgoff + off)
 876				return NULL;
 877			if (!is_cow_mapping(vma->vm_flags))
 878				return NULL;
 879		}
 880	}
 881
 882	if (is_zero_pfn(pfn))
 883		return NULL;
 
 884check_pfn:
 885	if (unlikely(pfn > highest_memmap_pfn)) {
 886		print_bad_pte(vma, addr, pte, NULL);
 887		return NULL;
 888	}
 889
 890	/*
 891	 * NOTE! We still have PageReserved() pages in the page tables.
 892	 * eg. VDSO mappings can cause them to exist.
 893	 */
 894out:
 895	return pfn_to_page(pfn);
 896}
 897
 898#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 899struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
 900				pmd_t pmd)
 901{
 902	unsigned long pfn = pmd_pfn(pmd);
 903
 904	/*
 905	 * There is no pmd_special() but there may be special pmds, e.g.
 906	 * in a direct-access (dax) mapping, so let's just replicate the
 907	 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
 908	 */
 909	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 910		if (vma->vm_flags & VM_MIXEDMAP) {
 911			if (!pfn_valid(pfn))
 912				return NULL;
 913			goto out;
 914		} else {
 915			unsigned long off;
 916			off = (addr - vma->vm_start) >> PAGE_SHIFT;
 917			if (pfn == vma->vm_pgoff + off)
 918				return NULL;
 919			if (!is_cow_mapping(vma->vm_flags))
 920				return NULL;
 921		}
 922	}
 923
 924	if (is_zero_pfn(pfn))
 
 
 925		return NULL;
 926	if (unlikely(pfn > highest_memmap_pfn))
 927		return NULL;
 928
 929	/*
 930	 * NOTE! We still have PageReserved() pages in the page tables.
 931	 * eg. VDSO mappings can cause them to exist.
 932	 */
 933out:
 934	return pfn_to_page(pfn);
 935}
 936#endif
 937
 938/*
 939 * copy one vm_area from one task to the other. Assumes the page tables
 940 * already present in the new task to be cleared in the whole range
 941 * covered by this vma.
 942 */
 943
 944static inline unsigned long
 945copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 946		pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 947		unsigned long addr, int *rss)
 948{
 949	unsigned long vm_flags = vma->vm_flags;
 950	pte_t pte = *src_pte;
 951	struct page *page;
 
 952
 953	/* pte contains position in swap or file, so copy. */
 954	if (unlikely(!pte_present(pte))) {
 955		swp_entry_t entry = pte_to_swp_entry(pte);
 956
 957		if (likely(!non_swap_entry(entry))) {
 958			if (swap_duplicate(entry) < 0)
 959				return entry.val;
 960
 961			/* make sure dst_mm is on swapoff's mmlist. */
 962			if (unlikely(list_empty(&dst_mm->mmlist))) {
 963				spin_lock(&mmlist_lock);
 964				if (list_empty(&dst_mm->mmlist))
 965					list_add(&dst_mm->mmlist,
 966							&src_mm->mmlist);
 967				spin_unlock(&mmlist_lock);
 968			}
 969			rss[MM_SWAPENTS]++;
 970		} else if (is_migration_entry(entry)) {
 971			page = migration_entry_to_page(entry);
 972
 973			rss[mm_counter(page)]++;
 974
 975			if (is_write_migration_entry(entry) &&
 976					is_cow_mapping(vm_flags)) {
 977				/*
 978				 * COW mappings require pages in both
 979				 * parent and child to be set to read.
 980				 */
 981				make_migration_entry_read(&entry);
 982				pte = swp_entry_to_pte(entry);
 983				if (pte_swp_soft_dirty(*src_pte))
 984					pte = pte_swp_mksoft_dirty(pte);
 985				set_pte_at(src_mm, addr, src_pte, pte);
 986			}
 987		} else if (is_device_private_entry(entry)) {
 988			page = device_private_entry_to_page(entry);
 989
 
 
 990			/*
 991			 * Update rss count even for unaddressable pages, as
 992			 * they should treated just like normal pages in this
 993			 * respect.
 994			 *
 995			 * We will likely want to have some new rss counters
 996			 * for unaddressable pages, at some point. But for now
 997			 * keep things as they are.
 998			 */
 999			get_page(page);
1000			rss[mm_counter(page)]++;
1001			page_dup_rmap(page, false);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1002
1003			/*
1004			 * We do not preserve soft-dirty information, because so
1005			 * far, checkpoint/restore is the only feature that
1006			 * requires that. And checkpoint/restore does not work
1007			 * when a device driver is involved (you cannot easily
1008			 * save and restore device driver state).
1009			 */
1010			if (is_write_device_private_entry(entry) &&
1011			    is_cow_mapping(vm_flags)) {
1012				make_device_private_entry_read(&entry);
1013				pte = swp_entry_to_pte(entry);
1014				set_pte_at(src_mm, addr, src_pte, pte);
1015			}
 
1016		}
1017		goto out_set_pte;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1018	}
1019
1020	/*
1021	 * If it's a COW mapping, write protect it both
1022	 * in the parent and the child
1023	 */
1024	if (is_cow_mapping(vm_flags)) {
1025		ptep_set_wrprotect(src_mm, addr, src_pte);
1026		pte = pte_wrprotect(pte);
1027	}
1028
1029	/*
1030	 * If it's a shared mapping, mark it clean in
1031	 * the child
1032	 */
1033	if (vm_flags & VM_SHARED)
1034		pte = pte_mkclean(pte);
1035	pte = pte_mkold(pte);
1036
1037	page = vm_normal_page(vma, addr, pte);
1038	if (page) {
1039		get_page(page);
1040		page_dup_rmap(page, false);
1041		rss[mm_counter(page)]++;
1042	} else if (pte_devmap(pte)) {
1043		page = pte_page(pte);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1044
1045		/*
1046		 * Cache coherent device memory behave like regular page and
1047		 * not like persistent memory page. For more informations see
1048		 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1049		 */
1050		if (is_device_public_page(page)) {
1051			get_page(page);
1052			page_dup_rmap(page, false);
1053			rss[mm_counter(page)]++;
1054		}
1055	}
 
1056
1057out_set_pte:
1058	set_pte_at(dst_mm, addr, dst_pte, pte);
1059	return 0;
1060}
1061
1062static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1063		   pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 
1064		   unsigned long addr, unsigned long end)
1065{
1066	pte_t *orig_src_pte, *orig_dst_pte;
1067	pte_t *src_pte, *dst_pte;
1068	spinlock_t *src_ptl, *dst_ptl;
1069	int progress = 0;
1070	int rss[NR_MM_COUNTERS];
1071	swp_entry_t entry = (swp_entry_t){0};
 
1072
1073again:
 
1074	init_rss_vec(rss);
1075
1076	dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1077	if (!dst_pte)
1078		return -ENOMEM;
 
 
1079	src_pte = pte_offset_map(src_pmd, addr);
1080	src_ptl = pte_lockptr(src_mm, src_pmd);
1081	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1082	orig_src_pte = src_pte;
1083	orig_dst_pte = dst_pte;
1084	arch_enter_lazy_mmu_mode();
1085
1086	do {
1087		/*
1088		 * We are holding two locks at this point - either of them
1089		 * could generate latencies in another task on another CPU.
1090		 */
1091		if (progress >= 32) {
1092			progress = 0;
1093			if (need_resched() ||
1094			    spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1095				break;
1096		}
1097		if (pte_none(*src_pte)) {
1098			progress++;
1099			continue;
1100		}
1101		entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
 
 
1102							vma, addr, rss);
1103		if (entry.val)
 
 
 
 
 
 
 
 
 
 
 
 
1104			break;
 
 
 
 
 
 
 
 
 
 
1105		progress += 8;
1106	} while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1107
1108	arch_leave_lazy_mmu_mode();
1109	spin_unlock(src_ptl);
1110	pte_unmap(orig_src_pte);
1111	add_mm_rss_vec(dst_mm, rss);
1112	pte_unmap_unlock(orig_dst_pte, dst_ptl);
1113	cond_resched();
1114
1115	if (entry.val) {
1116		if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
 
 
 
 
 
 
 
 
1117			return -ENOMEM;
1118		progress = 0;
 
1119	}
1120	if (addr != end)
1121		goto again;
1122	return 0;
 
 
 
1123}
1124
1125static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1126		pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
 
1127		unsigned long addr, unsigned long end)
1128{
1129	pmd_t *src_pmd, *dst_pmd;
1130	unsigned long next;
1131
1132	dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1133	if (!dst_pmd)
1134		return -ENOMEM;
1135	src_pmd = pmd_offset(src_pud, addr);
1136	do {
1137		next = pmd_addr_end(addr, end);
1138		if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1139			|| pmd_devmap(*src_pmd)) {
1140			int err;
1141			VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1142			err = copy_huge_pmd(dst_mm, src_mm,
1143					    dst_pmd, src_pmd, addr, vma);
1144			if (err == -ENOMEM)
1145				return -ENOMEM;
1146			if (!err)
1147				continue;
1148			/* fall through */
1149		}
1150		if (pmd_none_or_clear_bad(src_pmd))
1151			continue;
1152		if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1153						vma, addr, next))
1154			return -ENOMEM;
1155	} while (dst_pmd++, src_pmd++, addr = next, addr != end);
1156	return 0;
1157}
1158
1159static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1160		p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
 
1161		unsigned long addr, unsigned long end)
1162{
1163	pud_t *src_pud, *dst_pud;
1164	unsigned long next;
1165
1166	dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1167	if (!dst_pud)
1168		return -ENOMEM;
1169	src_pud = pud_offset(src_p4d, addr);
1170	do {
1171		next = pud_addr_end(addr, end);
1172		if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1173			int err;
1174
1175			VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1176			err = copy_huge_pud(dst_mm, src_mm,
1177					    dst_pud, src_pud, addr, vma);
1178			if (err == -ENOMEM)
1179				return -ENOMEM;
1180			if (!err)
1181				continue;
1182			/* fall through */
1183		}
1184		if (pud_none_or_clear_bad(src_pud))
1185			continue;
1186		if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1187						vma, addr, next))
1188			return -ENOMEM;
1189	} while (dst_pud++, src_pud++, addr = next, addr != end);
1190	return 0;
1191}
1192
1193static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1194		pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
 
1195		unsigned long addr, unsigned long end)
1196{
1197	p4d_t *src_p4d, *dst_p4d;
1198	unsigned long next;
1199
1200	dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1201	if (!dst_p4d)
1202		return -ENOMEM;
1203	src_p4d = p4d_offset(src_pgd, addr);
1204	do {
1205		next = p4d_addr_end(addr, end);
1206		if (p4d_none_or_clear_bad(src_p4d))
1207			continue;
1208		if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1209						vma, addr, next))
1210			return -ENOMEM;
1211	} while (dst_p4d++, src_p4d++, addr = next, addr != end);
1212	return 0;
1213}
1214
1215int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1216		struct vm_area_struct *vma)
1217{
1218	pgd_t *src_pgd, *dst_pgd;
1219	unsigned long next;
1220	unsigned long addr = vma->vm_start;
1221	unsigned long end = vma->vm_end;
1222	unsigned long mmun_start;	/* For mmu_notifiers */
1223	unsigned long mmun_end;		/* For mmu_notifiers */
1224	bool is_cow;
1225	int ret;
1226
1227	/*
1228	 * Don't copy ptes where a page fault will fill them correctly.
1229	 * Fork becomes much lighter when there are big shared or private
1230	 * readonly mappings. The tradeoff is that copy_page_range is more
1231	 * efficient than faulting.
1232	 */
1233	if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1234			!vma->anon_vma)
1235		return 0;
1236
1237	if (is_vm_hugetlb_page(vma))
1238		return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1239
1240	if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1241		/*
1242		 * We do not free on error cases below as remove_vma
1243		 * gets called on error from higher level routine
1244		 */
1245		ret = track_pfn_copy(vma);
1246		if (ret)
1247			return ret;
1248	}
1249
1250	/*
1251	 * We need to invalidate the secondary MMU mappings only when
1252	 * there could be a permission downgrade on the ptes of the
1253	 * parent mm. And a permission downgrade will only happen if
1254	 * is_cow_mapping() returns true.
1255	 */
1256	is_cow = is_cow_mapping(vma->vm_flags);
1257	mmun_start = addr;
1258	mmun_end   = end;
1259	if (is_cow)
1260		mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1261						    mmun_end);
 
1262
1263	ret = 0;
1264	dst_pgd = pgd_offset(dst_mm, addr);
1265	src_pgd = pgd_offset(src_mm, addr);
1266	do {
1267		next = pgd_addr_end(addr, end);
1268		if (pgd_none_or_clear_bad(src_pgd))
1269			continue;
1270		if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1271					    vma, addr, next))) {
1272			ret = -ENOMEM;
1273			break;
1274		}
1275	} while (dst_pgd++, src_pgd++, addr = next, addr != end);
1276
1277	if (is_cow)
1278		mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1279	return ret;
1280}
1281
1282static unsigned long zap_pte_range(struct mmu_gather *tlb,
1283				struct vm_area_struct *vma, pmd_t *pmd,
1284				unsigned long addr, unsigned long end,
1285				struct zap_details *details)
1286{
1287	struct mm_struct *mm = tlb->mm;
1288	int force_flush = 0;
1289	int rss[NR_MM_COUNTERS];
1290	spinlock_t *ptl;
1291	pte_t *start_pte;
1292	pte_t *pte;
1293	swp_entry_t entry;
1294
1295	tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1296again:
1297	init_rss_vec(rss);
1298	start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1299	pte = start_pte;
1300	flush_tlb_batched_pending(mm);
1301	arch_enter_lazy_mmu_mode();
1302	do {
1303		pte_t ptent = *pte;
1304		if (pte_none(ptent))
1305			continue;
1306
 
 
 
1307		if (pte_present(ptent)) {
1308			struct page *page;
1309
1310			page = _vm_normal_page(vma, addr, ptent, true);
1311			if (unlikely(details) && page) {
1312				/*
1313				 * unmap_shared_mapping_pages() wants to
1314				 * invalidate cache without truncating:
1315				 * unmap shared but keep private pages.
1316				 */
1317				if (details->check_mapping &&
1318				    details->check_mapping != page_rmapping(page))
1319					continue;
1320			}
1321			ptent = ptep_get_and_clear_full(mm, addr, pte,
1322							tlb->fullmm);
1323			tlb_remove_tlb_entry(tlb, pte, addr);
1324			if (unlikely(!page))
1325				continue;
1326
1327			if (!PageAnon(page)) {
1328				if (pte_dirty(ptent)) {
1329					force_flush = 1;
1330					set_page_dirty(page);
1331				}
1332				if (pte_young(ptent) &&
1333				    likely(!(vma->vm_flags & VM_SEQ_READ)))
1334					mark_page_accessed(page);
1335			}
1336			rss[mm_counter(page)]--;
1337			page_remove_rmap(page, false);
1338			if (unlikely(page_mapcount(page) < 0))
1339				print_bad_pte(vma, addr, ptent, page);
1340			if (unlikely(__tlb_remove_page(tlb, page))) {
1341				force_flush = 1;
1342				addr += PAGE_SIZE;
1343				break;
1344			}
1345			continue;
1346		}
1347
1348		entry = pte_to_swp_entry(ptent);
1349		if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1350			struct page *page = device_private_entry_to_page(entry);
1351
1352			if (unlikely(details && details->check_mapping)) {
1353				/*
1354				 * unmap_shared_mapping_pages() wants to
1355				 * invalidate cache without truncating:
1356				 * unmap shared but keep private pages.
1357				 */
1358				if (details->check_mapping !=
1359				    page_rmapping(page))
1360					continue;
1361			}
1362
1363			pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1364			rss[mm_counter(page)]--;
1365			page_remove_rmap(page, false);
1366			put_page(page);
1367			continue;
1368		}
1369
1370		/* If details->check_mapping, we leave swap entries. */
1371		if (unlikely(details))
1372			continue;
1373
1374		entry = pte_to_swp_entry(ptent);
1375		if (!non_swap_entry(entry))
1376			rss[MM_SWAPENTS]--;
1377		else if (is_migration_entry(entry)) {
1378			struct page *page;
1379
1380			page = migration_entry_to_page(entry);
1381			rss[mm_counter(page)]--;
1382		}
1383		if (unlikely(!free_swap_and_cache(entry)))
1384			print_bad_pte(vma, addr, ptent, NULL);
1385		pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1386	} while (pte++, addr += PAGE_SIZE, addr != end);
1387
1388	add_mm_rss_vec(mm, rss);
1389	arch_leave_lazy_mmu_mode();
1390
1391	/* Do the actual TLB flush before dropping ptl */
1392	if (force_flush)
1393		tlb_flush_mmu_tlbonly(tlb);
1394	pte_unmap_unlock(start_pte, ptl);
1395
1396	/*
1397	 * If we forced a TLB flush (either due to running out of
1398	 * batch buffers or because we needed to flush dirty TLB
1399	 * entries before releasing the ptl), free the batched
1400	 * memory too. Restart if we didn't do everything.
1401	 */
1402	if (force_flush) {
1403		force_flush = 0;
1404		tlb_flush_mmu_free(tlb);
1405		if (addr != end)
1406			goto again;
 
 
 
1407	}
1408
1409	return addr;
1410}
1411
1412static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1413				struct vm_area_struct *vma, pud_t *pud,
1414				unsigned long addr, unsigned long end,
1415				struct zap_details *details)
1416{
1417	pmd_t *pmd;
1418	unsigned long next;
1419
1420	pmd = pmd_offset(pud, addr);
1421	do {
1422		next = pmd_addr_end(addr, end);
1423		if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1424			if (next - addr != HPAGE_PMD_SIZE) {
1425				VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1426				    !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1427				__split_huge_pmd(vma, pmd, addr, false, NULL);
1428			} else if (zap_huge_pmd(tlb, vma, pmd, addr))
1429				goto next;
1430			/* fall through */
1431		}
1432		/*
1433		 * Here there can be other concurrent MADV_DONTNEED or
1434		 * trans huge page faults running, and if the pmd is
1435		 * none or trans huge it can change under us. This is
1436		 * because MADV_DONTNEED holds the mmap_sem in read
1437		 * mode.
1438		 */
1439		if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1440			goto next;
1441		next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1442next:
1443		cond_resched();
1444	} while (pmd++, addr = next, addr != end);
1445
1446	return addr;
1447}
1448
1449static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1450				struct vm_area_struct *vma, p4d_t *p4d,
1451				unsigned long addr, unsigned long end,
1452				struct zap_details *details)
1453{
1454	pud_t *pud;
1455	unsigned long next;
1456
1457	pud = pud_offset(p4d, addr);
1458	do {
1459		next = pud_addr_end(addr, end);
1460		if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1461			if (next - addr != HPAGE_PUD_SIZE) {
1462				VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1463				split_huge_pud(vma, pud, addr);
1464			} else if (zap_huge_pud(tlb, vma, pud, addr))
1465				goto next;
1466			/* fall through */
1467		}
1468		if (pud_none_or_clear_bad(pud))
1469			continue;
1470		next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1471next:
1472		cond_resched();
1473	} while (pud++, addr = next, addr != end);
1474
1475	return addr;
1476}
1477
1478static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1479				struct vm_area_struct *vma, pgd_t *pgd,
1480				unsigned long addr, unsigned long end,
1481				struct zap_details *details)
1482{
1483	p4d_t *p4d;
1484	unsigned long next;
1485
1486	p4d = p4d_offset(pgd, addr);
1487	do {
1488		next = p4d_addr_end(addr, end);
1489		if (p4d_none_or_clear_bad(p4d))
1490			continue;
1491		next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1492	} while (p4d++, addr = next, addr != end);
1493
1494	return addr;
1495}
1496
1497void unmap_page_range(struct mmu_gather *tlb,
1498			     struct vm_area_struct *vma,
1499			     unsigned long addr, unsigned long end,
1500			     struct zap_details *details)
1501{
1502	pgd_t *pgd;
1503	unsigned long next;
1504
1505	BUG_ON(addr >= end);
1506	tlb_start_vma(tlb, vma);
1507	pgd = pgd_offset(vma->vm_mm, addr);
1508	do {
1509		next = pgd_addr_end(addr, end);
1510		if (pgd_none_or_clear_bad(pgd))
1511			continue;
1512		next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1513	} while (pgd++, addr = next, addr != end);
1514	tlb_end_vma(tlb, vma);
1515}
1516
1517
1518static void unmap_single_vma(struct mmu_gather *tlb,
1519		struct vm_area_struct *vma, unsigned long start_addr,
1520		unsigned long end_addr,
1521		struct zap_details *details)
1522{
1523	unsigned long start = max(vma->vm_start, start_addr);
1524	unsigned long end;
1525
1526	if (start >= vma->vm_end)
1527		return;
1528	end = min(vma->vm_end, end_addr);
1529	if (end <= vma->vm_start)
1530		return;
1531
1532	if (vma->vm_file)
1533		uprobe_munmap(vma, start, end);
1534
1535	if (unlikely(vma->vm_flags & VM_PFNMAP))
1536		untrack_pfn(vma, 0, 0);
1537
1538	if (start != end) {
1539		if (unlikely(is_vm_hugetlb_page(vma))) {
1540			/*
1541			 * It is undesirable to test vma->vm_file as it
1542			 * should be non-null for valid hugetlb area.
1543			 * However, vm_file will be NULL in the error
1544			 * cleanup path of mmap_region. When
1545			 * hugetlbfs ->mmap method fails,
1546			 * mmap_region() nullifies vma->vm_file
1547			 * before calling this function to clean up.
1548			 * Since no pte has actually been setup, it is
1549			 * safe to do nothing in this case.
1550			 */
1551			if (vma->vm_file) {
1552				i_mmap_lock_write(vma->vm_file->f_mapping);
1553				__unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1554				i_mmap_unlock_write(vma->vm_file->f_mapping);
1555			}
1556		} else
1557			unmap_page_range(tlb, vma, start, end, details);
1558	}
1559}
1560
1561/**
1562 * unmap_vmas - unmap a range of memory covered by a list of vma's
1563 * @tlb: address of the caller's struct mmu_gather
1564 * @vma: the starting vma
1565 * @start_addr: virtual address at which to start unmapping
1566 * @end_addr: virtual address at which to end unmapping
1567 *
1568 * Unmap all pages in the vma list.
1569 *
1570 * Only addresses between `start' and `end' will be unmapped.
1571 *
1572 * The VMA list must be sorted in ascending virtual address order.
1573 *
1574 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1575 * range after unmap_vmas() returns.  So the only responsibility here is to
1576 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1577 * drops the lock and schedules.
1578 */
1579void unmap_vmas(struct mmu_gather *tlb,
1580		struct vm_area_struct *vma, unsigned long start_addr,
1581		unsigned long end_addr)
1582{
1583	struct mm_struct *mm = vma->vm_mm;
1584
1585	mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
 
 
1586	for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1587		unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1588	mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1589}
1590
1591/**
1592 * zap_page_range - remove user pages in a given range
1593 * @vma: vm_area_struct holding the applicable pages
1594 * @start: starting address of pages to zap
1595 * @size: number of bytes to zap
1596 *
1597 * Caller must protect the VMA list
1598 */
1599void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1600		unsigned long size)
1601{
1602	struct mm_struct *mm = vma->vm_mm;
1603	struct mmu_gather tlb;
1604	unsigned long end = start + size;
1605
1606	lru_add_drain();
1607	tlb_gather_mmu(&tlb, mm, start, end);
1608	update_hiwater_rss(mm);
1609	mmu_notifier_invalidate_range_start(mm, start, end);
1610	for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
1611		unmap_single_vma(&tlb, vma, start, end, NULL);
1612
1613		/*
1614		 * zap_page_range does not specify whether mmap_sem should be
1615		 * held for read or write. That allows parallel zap_page_range
1616		 * operations to unmap a PTE and defer a flush meaning that
1617		 * this call observes pte_none and fails to flush the TLB.
1618		 * Rather than adding a complex API, ensure that no stale
1619		 * TLB entries exist when this call returns.
1620		 */
1621		flush_tlb_range(vma, start, end);
1622	}
1623
1624	mmu_notifier_invalidate_range_end(mm, start, end);
1625	tlb_finish_mmu(&tlb, start, end);
1626}
1627
1628/**
1629 * zap_page_range_single - remove user pages in a given range
1630 * @vma: vm_area_struct holding the applicable pages
1631 * @address: starting address of pages to zap
1632 * @size: number of bytes to zap
1633 * @details: details of shared cache invalidation
1634 *
1635 * The range must fit into one VMA.
1636 */
1637static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1638		unsigned long size, struct zap_details *details)
1639{
1640	struct mm_struct *mm = vma->vm_mm;
1641	struct mmu_gather tlb;
1642	unsigned long end = address + size;
1643
1644	lru_add_drain();
1645	tlb_gather_mmu(&tlb, mm, address, end);
1646	update_hiwater_rss(mm);
1647	mmu_notifier_invalidate_range_start(mm, address, end);
1648	unmap_single_vma(&tlb, vma, address, end, details);
1649	mmu_notifier_invalidate_range_end(mm, address, end);
1650	tlb_finish_mmu(&tlb, address, end);
 
 
1651}
1652
1653/**
1654 * zap_vma_ptes - remove ptes mapping the vma
1655 * @vma: vm_area_struct holding ptes to be zapped
1656 * @address: starting address of pages to zap
1657 * @size: number of bytes to zap
1658 *
1659 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1660 *
1661 * The entire address range must be fully contained within the vma.
1662 *
1663 * Returns 0 if successful.
1664 */
1665int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1666		unsigned long size)
1667{
1668	if (address < vma->vm_start || address + size > vma->vm_end ||
1669	    		!(vma->vm_flags & VM_PFNMAP))
1670		return -1;
 
1671	zap_page_range_single(vma, address, size, NULL);
1672	return 0;
1673}
1674EXPORT_SYMBOL_GPL(zap_vma_ptes);
1675
1676pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1677			spinlock_t **ptl)
1678{
1679	pgd_t *pgd;
1680	p4d_t *p4d;
1681	pud_t *pud;
1682	pmd_t *pmd;
1683
1684	pgd = pgd_offset(mm, addr);
1685	p4d = p4d_alloc(mm, pgd, addr);
1686	if (!p4d)
1687		return NULL;
1688	pud = pud_alloc(mm, p4d, addr);
1689	if (!pud)
1690		return NULL;
1691	pmd = pmd_alloc(mm, pud, addr);
1692	if (!pmd)
1693		return NULL;
1694
1695	VM_BUG_ON(pmd_trans_huge(*pmd));
 
 
 
 
 
 
 
 
 
 
1696	return pte_alloc_map_lock(mm, pmd, addr, ptl);
1697}
1698
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1699/*
1700 * This is the old fallback for page remapping.
1701 *
1702 * For historical reasons, it only allows reserved pages. Only
1703 * old drivers should use this, and they needed to mark their
1704 * pages reserved for the old functions anyway.
1705 */
1706static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1707			struct page *page, pgprot_t prot)
1708{
1709	struct mm_struct *mm = vma->vm_mm;
1710	int retval;
1711	pte_t *pte;
1712	spinlock_t *ptl;
1713
1714	retval = -EINVAL;
1715	if (PageAnon(page))
1716		goto out;
1717	retval = -ENOMEM;
1718	flush_dcache_page(page);
1719	pte = get_locked_pte(mm, addr, &ptl);
1720	if (!pte)
1721		goto out;
1722	retval = -EBUSY;
1723	if (!pte_none(*pte))
1724		goto out_unlock;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1725
1726	/* Ok, finally just insert the thing.. */
1727	get_page(page);
1728	inc_mm_counter_fast(mm, mm_counter_file(page));
1729	page_add_file_rmap(page, false);
1730	set_pte_at(mm, addr, pte, mk_pte(page, prot));
1731
1732	retval = 0;
1733	pte_unmap_unlock(pte, ptl);
1734	return retval;
1735out_unlock:
1736	pte_unmap_unlock(pte, ptl);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1737out:
1738	return retval;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1739}
 
1740
1741/**
1742 * vm_insert_page - insert single page into user vma
1743 * @vma: user vma to map to
1744 * @addr: target user address of this page
1745 * @page: source kernel page
1746 *
1747 * This allows drivers to insert individual pages they've allocated
1748 * into a user vma.
1749 *
1750 * The page has to be a nice clean _individual_ kernel allocation.
1751 * If you allocate a compound page, you need to have marked it as
1752 * such (__GFP_COMP), or manually just split the page up yourself
1753 * (see split_page()).
1754 *
1755 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1756 * took an arbitrary page protection parameter. This doesn't allow
1757 * that. Your vma protection will have to be set up correctly, which
1758 * means that if you want a shared writable mapping, you'd better
1759 * ask for a shared writable mapping!
1760 *
1761 * The page does not need to be reserved.
1762 *
1763 * Usually this function is called from f_op->mmap() handler
1764 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1765 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1766 * function from other places, for example from page-fault handler.
 
 
1767 */
1768int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1769			struct page *page)
1770{
1771	if (addr < vma->vm_start || addr >= vma->vm_end)
1772		return -EFAULT;
1773	if (!page_count(page))
1774		return -EINVAL;
1775	if (!(vma->vm_flags & VM_MIXEDMAP)) {
1776		BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1777		BUG_ON(vma->vm_flags & VM_PFNMAP);
1778		vma->vm_flags |= VM_MIXEDMAP;
1779	}
1780	return insert_page(vma, addr, page, vma->vm_page_prot);
1781}
1782EXPORT_SYMBOL(vm_insert_page);
1783
1784static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1785			pfn_t pfn, pgprot_t prot, bool mkwrite)
1786{
1787	struct mm_struct *mm = vma->vm_mm;
1788	int retval;
1789	pte_t *pte, entry;
1790	spinlock_t *ptl;
1791
1792	retval = -ENOMEM;
1793	pte = get_locked_pte(mm, addr, &ptl);
1794	if (!pte)
1795		goto out;
1796	retval = -EBUSY;
1797	if (!pte_none(*pte)) {
1798		if (mkwrite) {
1799			/*
1800			 * For read faults on private mappings the PFN passed
1801			 * in may not match the PFN we have mapped if the
1802			 * mapped PFN is a writeable COW page.  In the mkwrite
1803			 * case we are creating a writable PTE for a shared
1804			 * mapping and we expect the PFNs to match.
 
 
 
1805			 */
1806			if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
 
1807				goto out_unlock;
1808			entry = *pte;
1809			goto out_mkwrite;
1810		} else
1811			goto out_unlock;
 
 
 
1812	}
1813
1814	/* Ok, finally just insert the thing.. */
1815	if (pfn_t_devmap(pfn))
1816		entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1817	else
1818		entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1819
1820out_mkwrite:
1821	if (mkwrite) {
1822		entry = pte_mkyoung(entry);
1823		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1824	}
1825
1826	set_pte_at(mm, addr, pte, entry);
1827	update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1828
1829	retval = 0;
1830out_unlock:
1831	pte_unmap_unlock(pte, ptl);
1832out:
1833	return retval;
1834}
1835
1836/**
1837 * vm_insert_pfn - insert single pfn into user vma
1838 * @vma: user vma to map to
1839 * @addr: target user address of this page
1840 * @pfn: source kernel pfn
1841 *
1842 * Similar to vm_insert_page, this allows drivers to insert individual pages
1843 * they've allocated into a user vma. Same comments apply.
1844 *
1845 * This function should only be called from a vm_ops->fault handler, and
1846 * in that case the handler should return NULL.
1847 *
1848 * vma cannot be a COW mapping.
1849 *
1850 * As this is called only for pages that do not currently exist, we
1851 * do not need to flush old virtual caches or the TLB.
1852 */
1853int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1854			unsigned long pfn)
1855{
1856	return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1857}
1858EXPORT_SYMBOL(vm_insert_pfn);
1859
1860/**
1861 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1862 * @vma: user vma to map to
1863 * @addr: target user address of this page
1864 * @pfn: source kernel pfn
1865 * @pgprot: pgprot flags for the inserted page
1866 *
1867 * This is exactly like vm_insert_pfn, except that it allows drivers to
1868 * to override pgprot on a per-page basis.
1869 *
1870 * This only makes sense for IO mappings, and it makes no sense for
1871 * cow mappings.  In general, using multiple vmas is preferable;
1872 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1873 * impractical.
 
 
 
 
 
 
1874 */
1875int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1876			unsigned long pfn, pgprot_t pgprot)
1877{
1878	int ret;
1879	/*
1880	 * Technically, architectures with pte_special can avoid all these
1881	 * restrictions (same for remap_pfn_range).  However we would like
1882	 * consistency in testing and feature parity among all, so we should
1883	 * try to keep these invariants in place for everybody.
1884	 */
1885	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1886	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1887						(VM_PFNMAP|VM_MIXEDMAP));
1888	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1889	BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1890
1891	if (addr < vma->vm_start || addr >= vma->vm_end)
1892		return -EFAULT;
 
 
 
1893
1894	track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1895
1896	ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1897			false);
 
 
1898
1899	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1900}
1901EXPORT_SYMBOL(vm_insert_pfn_prot);
1902
1903static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1904{
1905	/* these checks mirror the abort conditions in vm_normal_page */
1906	if (vma->vm_flags & VM_MIXEDMAP)
1907		return true;
1908	if (pfn_t_devmap(pfn))
1909		return true;
1910	if (pfn_t_special(pfn))
1911		return true;
1912	if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1913		return true;
1914	return false;
1915}
1916
1917static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1918			pfn_t pfn, bool mkwrite)
 
1919{
1920	pgprot_t pgprot = vma->vm_page_prot;
1921
1922	BUG_ON(!vm_mixed_ok(vma, pfn));
1923
1924	if (addr < vma->vm_start || addr >= vma->vm_end)
1925		return -EFAULT;
1926
1927	track_pfn_insert(vma, &pgprot, pfn);
1928
 
 
 
1929	/*
1930	 * If we don't have pte special, then we have to use the pfn_valid()
1931	 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1932	 * refcount the page if pfn_valid is true (hence insert_page rather
1933	 * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1934	 * without pte special, it would there be refcounted as a normal page.
1935	 */
1936	if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
 
1937		struct page *page;
1938
1939		/*
1940		 * At this point we are committed to insert_page()
1941		 * regardless of whether the caller specified flags that
1942		 * result in pfn_t_has_page() == false.
1943		 */
1944		page = pfn_to_page(pfn_t_to_pfn(pfn));
1945		return insert_page(vma, addr, page, pgprot);
 
 
1946	}
1947	return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1948}
 
1949
1950int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1951			pfn_t pfn)
1952{
1953	return __vm_insert_mixed(vma, addr, pfn, false);
1954
1955}
1956EXPORT_SYMBOL(vm_insert_mixed);
1957
1958int vm_insert_mixed_mkwrite(struct vm_area_struct *vma, unsigned long addr,
1959			pfn_t pfn)
 
 
 
 
 
1960{
1961	return __vm_insert_mixed(vma, addr, pfn, true);
1962}
1963EXPORT_SYMBOL(vm_insert_mixed_mkwrite);
1964
1965/*
1966 * maps a range of physical memory into the requested pages. the old
1967 * mappings are removed. any references to nonexistent pages results
1968 * in null mappings (currently treated as "copy-on-access")
1969 */
1970static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1971			unsigned long addr, unsigned long end,
1972			unsigned long pfn, pgprot_t prot)
1973{
1974	pte_t *pte;
1975	spinlock_t *ptl;
 
1976
1977	pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1978	if (!pte)
1979		return -ENOMEM;
1980	arch_enter_lazy_mmu_mode();
1981	do {
1982		BUG_ON(!pte_none(*pte));
 
 
 
 
1983		set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1984		pfn++;
1985	} while (pte++, addr += PAGE_SIZE, addr != end);
1986	arch_leave_lazy_mmu_mode();
1987	pte_unmap_unlock(pte - 1, ptl);
1988	return 0;
1989}
1990
1991static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1992			unsigned long addr, unsigned long end,
1993			unsigned long pfn, pgprot_t prot)
1994{
1995	pmd_t *pmd;
1996	unsigned long next;
 
1997
1998	pfn -= addr >> PAGE_SHIFT;
1999	pmd = pmd_alloc(mm, pud, addr);
2000	if (!pmd)
2001		return -ENOMEM;
2002	VM_BUG_ON(pmd_trans_huge(*pmd));
2003	do {
2004		next = pmd_addr_end(addr, end);
2005		if (remap_pte_range(mm, pmd, addr, next,
2006				pfn + (addr >> PAGE_SHIFT), prot))
2007			return -ENOMEM;
 
2008	} while (pmd++, addr = next, addr != end);
2009	return 0;
2010}
2011
2012static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2013			unsigned long addr, unsigned long end,
2014			unsigned long pfn, pgprot_t prot)
2015{
2016	pud_t *pud;
2017	unsigned long next;
 
2018
2019	pfn -= addr >> PAGE_SHIFT;
2020	pud = pud_alloc(mm, p4d, addr);
2021	if (!pud)
2022		return -ENOMEM;
2023	do {
2024		next = pud_addr_end(addr, end);
2025		if (remap_pmd_range(mm, pud, addr, next,
2026				pfn + (addr >> PAGE_SHIFT), prot))
2027			return -ENOMEM;
 
2028	} while (pud++, addr = next, addr != end);
2029	return 0;
2030}
2031
2032static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2033			unsigned long addr, unsigned long end,
2034			unsigned long pfn, pgprot_t prot)
2035{
2036	p4d_t *p4d;
2037	unsigned long next;
 
2038
2039	pfn -= addr >> PAGE_SHIFT;
2040	p4d = p4d_alloc(mm, pgd, addr);
2041	if (!p4d)
2042		return -ENOMEM;
2043	do {
2044		next = p4d_addr_end(addr, end);
2045		if (remap_pud_range(mm, p4d, addr, next,
2046				pfn + (addr >> PAGE_SHIFT), prot))
2047			return -ENOMEM;
 
2048	} while (p4d++, addr = next, addr != end);
2049	return 0;
2050}
2051
2052/**
2053 * remap_pfn_range - remap kernel memory to userspace
2054 * @vma: user vma to map to
2055 * @addr: target user address to start at
2056 * @pfn: physical address of kernel memory
2057 * @size: size of map area
2058 * @prot: page protection flags for this mapping
2059 *
2060 *  Note: this is only safe if the mm semaphore is held when called.
 
 
2061 */
2062int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2063		    unsigned long pfn, unsigned long size, pgprot_t prot)
2064{
2065	pgd_t *pgd;
2066	unsigned long next;
2067	unsigned long end = addr + PAGE_ALIGN(size);
2068	struct mm_struct *mm = vma->vm_mm;
2069	unsigned long remap_pfn = pfn;
2070	int err;
2071
 
 
 
2072	/*
2073	 * Physically remapped pages are special. Tell the
2074	 * rest of the world about it:
2075	 *   VM_IO tells people not to look at these pages
2076	 *	(accesses can have side effects).
2077	 *   VM_PFNMAP tells the core MM that the base pages are just
2078	 *	raw PFN mappings, and do not have a "struct page" associated
2079	 *	with them.
2080	 *   VM_DONTEXPAND
2081	 *      Disable vma merging and expanding with mremap().
2082	 *   VM_DONTDUMP
2083	 *      Omit vma from core dump, even when VM_IO turned off.
2084	 *
2085	 * There's a horrible special case to handle copy-on-write
2086	 * behaviour that some programs depend on. We mark the "original"
2087	 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2088	 * See vm_normal_page() for details.
2089	 */
2090	if (is_cow_mapping(vma->vm_flags)) {
2091		if (addr != vma->vm_start || end != vma->vm_end)
2092			return -EINVAL;
2093		vma->vm_pgoff = pfn;
2094	}
2095
2096	err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2097	if (err)
2098		return -EINVAL;
2099
2100	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2101
2102	BUG_ON(addr >= end);
2103	pfn -= addr >> PAGE_SHIFT;
2104	pgd = pgd_offset(mm, addr);
2105	flush_cache_range(vma, addr, end);
2106	do {
2107		next = pgd_addr_end(addr, end);
2108		err = remap_p4d_range(mm, pgd, addr, next,
2109				pfn + (addr >> PAGE_SHIFT), prot);
2110		if (err)
2111			break;
2112	} while (pgd++, addr = next, addr != end);
2113
2114	if (err)
2115		untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2116
2117	return err;
2118}
2119EXPORT_SYMBOL(remap_pfn_range);
2120
2121/**
2122 * vm_iomap_memory - remap memory to userspace
2123 * @vma: user vma to map to
2124 * @start: start of area
2125 * @len: size of area
2126 *
2127 * This is a simplified io_remap_pfn_range() for common driver use. The
2128 * driver just needs to give us the physical memory range to be mapped,
2129 * we'll figure out the rest from the vma information.
2130 *
2131 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2132 * whatever write-combining details or similar.
 
 
2133 */
2134int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2135{
2136	unsigned long vm_len, pfn, pages;
2137
2138	/* Check that the physical memory area passed in looks valid */
2139	if (start + len < start)
2140		return -EINVAL;
2141	/*
2142	 * You *really* shouldn't map things that aren't page-aligned,
2143	 * but we've historically allowed it because IO memory might
2144	 * just have smaller alignment.
2145	 */
2146	len += start & ~PAGE_MASK;
2147	pfn = start >> PAGE_SHIFT;
2148	pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2149	if (pfn + pages < pfn)
2150		return -EINVAL;
2151
2152	/* We start the mapping 'vm_pgoff' pages into the area */
2153	if (vma->vm_pgoff > pages)
2154		return -EINVAL;
2155	pfn += vma->vm_pgoff;
2156	pages -= vma->vm_pgoff;
2157
2158	/* Can we fit all of the mapping? */
2159	vm_len = vma->vm_end - vma->vm_start;
2160	if (vm_len >> PAGE_SHIFT > pages)
2161		return -EINVAL;
2162
2163	/* Ok, let it rip */
2164	return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2165}
2166EXPORT_SYMBOL(vm_iomap_memory);
2167
2168static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2169				     unsigned long addr, unsigned long end,
2170				     pte_fn_t fn, void *data)
 
2171{
2172	pte_t *pte;
2173	int err;
2174	pgtable_t token;
2175	spinlock_t *uninitialized_var(ptl);
2176
2177	pte = (mm == &init_mm) ?
2178		pte_alloc_kernel(pmd, addr) :
2179		pte_alloc_map_lock(mm, pmd, addr, &ptl);
2180	if (!pte)
2181		return -ENOMEM;
 
 
 
 
 
 
2182
2183	BUG_ON(pmd_huge(*pmd));
2184
2185	arch_enter_lazy_mmu_mode();
2186
2187	token = pmd_pgtable(*pmd);
2188
2189	do {
2190		err = fn(pte++, token, addr, data);
2191		if (err)
2192			break;
 
 
2193	} while (addr += PAGE_SIZE, addr != end);
 
2194
2195	arch_leave_lazy_mmu_mode();
2196
2197	if (mm != &init_mm)
2198		pte_unmap_unlock(pte-1, ptl);
2199	return err;
2200}
2201
2202static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2203				     unsigned long addr, unsigned long end,
2204				     pte_fn_t fn, void *data)
 
2205{
2206	pmd_t *pmd;
2207	unsigned long next;
2208	int err;
2209
2210	BUG_ON(pud_huge(*pud));
2211
2212	pmd = pmd_alloc(mm, pud, addr);
2213	if (!pmd)
2214		return -ENOMEM;
 
 
 
 
2215	do {
2216		next = pmd_addr_end(addr, end);
2217		err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2218		if (err)
2219			break;
 
 
 
2220	} while (pmd++, addr = next, addr != end);
2221	return err;
2222}
2223
2224static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2225				     unsigned long addr, unsigned long end,
2226				     pte_fn_t fn, void *data)
 
2227{
2228	pud_t *pud;
2229	unsigned long next;
2230	int err;
2231
2232	pud = pud_alloc(mm, p4d, addr);
2233	if (!pud)
2234		return -ENOMEM;
 
 
 
 
2235	do {
2236		next = pud_addr_end(addr, end);
2237		err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2238		if (err)
2239			break;
 
 
 
2240	} while (pud++, addr = next, addr != end);
2241	return err;
2242}
2243
2244static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2245				     unsigned long addr, unsigned long end,
2246				     pte_fn_t fn, void *data)
 
2247{
2248	p4d_t *p4d;
2249	unsigned long next;
2250	int err;
2251
2252	p4d = p4d_alloc(mm, pgd, addr);
2253	if (!p4d)
2254		return -ENOMEM;
 
 
 
 
2255	do {
2256		next = p4d_addr_end(addr, end);
2257		err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2258		if (err)
2259			break;
 
 
 
2260	} while (p4d++, addr = next, addr != end);
2261	return err;
2262}
2263
2264/*
2265 * Scan a region of virtual memory, filling in page tables as necessary
2266 * and calling a provided function on each leaf page table.
2267 */
2268int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2269			unsigned long size, pte_fn_t fn, void *data)
2270{
2271	pgd_t *pgd;
2272	unsigned long next;
2273	unsigned long end = addr + size;
2274	int err;
 
2275
2276	if (WARN_ON(addr >= end))
2277		return -EINVAL;
2278
2279	pgd = pgd_offset(mm, addr);
2280	do {
2281		next = pgd_addr_end(addr, end);
2282		err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
 
 
2283		if (err)
2284			break;
2285	} while (pgd++, addr = next, addr != end);
2286
 
 
 
2287	return err;
2288}
 
 
 
 
 
 
 
 
 
 
2289EXPORT_SYMBOL_GPL(apply_to_page_range);
2290
2291/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2292 * handle_pte_fault chooses page fault handler according to an entry which was
2293 * read non-atomically.  Before making any commitment, on those architectures
2294 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2295 * parts, do_swap_page must check under lock before unmapping the pte and
2296 * proceeding (but do_wp_page is only called after already making such a check;
2297 * and do_anonymous_page can safely check later on).
2298 */
2299static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2300				pte_t *page_table, pte_t orig_pte)
2301{
2302	int same = 1;
2303#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2304	if (sizeof(pte_t) > sizeof(unsigned long)) {
2305		spinlock_t *ptl = pte_lockptr(mm, pmd);
2306		spin_lock(ptl);
2307		same = pte_same(*page_table, orig_pte);
2308		spin_unlock(ptl);
2309	}
2310#endif
2311	pte_unmap(page_table);
2312	return same;
2313}
2314
2315static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
 
2316{
2317	debug_dma_assert_idle(src);
 
 
 
 
 
 
 
 
 
 
 
2318
2319	/*
2320	 * If the source page was a PFN mapping, we don't have
2321	 * a "struct page" for it. We do a best-effort copy by
2322	 * just copying from the original user address. If that
2323	 * fails, we just zero-fill it. Live with it.
2324	 */
2325	if (unlikely(!src)) {
2326		void *kaddr = kmap_atomic(dst);
2327		void __user *uaddr = (void __user *)(va & PAGE_MASK);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2328
2329		/*
2330		 * This really shouldn't fail, because the page is there
2331		 * in the page tables. But it might just be unreadable,
2332		 * in which case we just give up and fill the result with
2333		 * zeroes.
2334		 */
2335		if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
 
 
 
 
 
 
2336			clear_page(kaddr);
2337		kunmap_atomic(kaddr);
2338		flush_dcache_page(dst);
2339	} else
2340		copy_user_highpage(dst, src, va, vma);
 
 
 
 
 
 
 
 
2341}
2342
2343static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2344{
2345	struct file *vm_file = vma->vm_file;
2346
2347	if (vm_file)
2348		return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2349
2350	/*
2351	 * Special mappings (e.g. VDSO) do not have any file so fake
2352	 * a default GFP_KERNEL for them.
2353	 */
2354	return GFP_KERNEL;
2355}
2356
2357/*
2358 * Notify the address space that the page is about to become writable so that
2359 * it can prohibit this or wait for the page to get into an appropriate state.
2360 *
2361 * We do this without the lock held, so that it can sleep if it needs to.
2362 */
2363static int do_page_mkwrite(struct vm_fault *vmf)
2364{
2365	int ret;
2366	struct page *page = vmf->page;
2367	unsigned int old_flags = vmf->flags;
2368
2369	vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2370
 
 
 
 
2371	ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2372	/* Restore original flags so that caller is not surprised */
2373	vmf->flags = old_flags;
2374	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2375		return ret;
2376	if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2377		lock_page(page);
2378		if (!page->mapping) {
2379			unlock_page(page);
2380			return 0; /* retry */
2381		}
2382		ret |= VM_FAULT_LOCKED;
2383	} else
2384		VM_BUG_ON_PAGE(!PageLocked(page), page);
2385	return ret;
2386}
2387
2388/*
2389 * Handle dirtying of a page in shared file mapping on a write fault.
2390 *
2391 * The function expects the page to be locked and unlocks it.
2392 */
2393static void fault_dirty_shared_page(struct vm_area_struct *vma,
2394				    struct page *page)
2395{
 
2396	struct address_space *mapping;
 
2397	bool dirtied;
2398	bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2399
2400	dirtied = set_page_dirty(page);
2401	VM_BUG_ON_PAGE(PageAnon(page), page);
2402	/*
2403	 * Take a local copy of the address_space - page.mapping may be zeroed
2404	 * by truncate after unlock_page().   The address_space itself remains
2405	 * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2406	 * release semantics to prevent the compiler from undoing this copying.
2407	 */
2408	mapping = page_rmapping(page);
2409	unlock_page(page);
2410
 
 
 
 
 
 
 
 
 
 
 
 
2411	if ((dirtied || page_mkwrite) && mapping) {
2412		/*
2413		 * Some device drivers do not set page.mapping
2414		 * but still dirty their pages
2415		 */
2416		balance_dirty_pages_ratelimited(mapping);
 
 
 
 
2417	}
2418
2419	if (!page_mkwrite)
2420		file_update_time(vma->vm_file);
2421}
2422
2423/*
2424 * Handle write page faults for pages that can be reused in the current vma
2425 *
2426 * This can happen either due to the mapping being with the VM_SHARED flag,
2427 * or due to us being the last reference standing to the page. In either
2428 * case, all we need to do here is to mark the page as writable and update
2429 * any related book-keeping.
2430 */
2431static inline void wp_page_reuse(struct vm_fault *vmf)
2432	__releases(vmf->ptl)
2433{
2434	struct vm_area_struct *vma = vmf->vma;
2435	struct page *page = vmf->page;
2436	pte_t entry;
2437	/*
2438	 * Clear the pages cpupid information as the existing
2439	 * information potentially belongs to a now completely
2440	 * unrelated process.
2441	 */
2442	if (page)
2443		page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2444
2445	flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2446	entry = pte_mkyoung(vmf->orig_pte);
2447	entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2448	if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2449		update_mmu_cache(vma, vmf->address, vmf->pte);
2450	pte_unmap_unlock(vmf->pte, vmf->ptl);
 
2451}
2452
2453/*
2454 * Handle the case of a page which we actually need to copy to a new page.
2455 *
2456 * Called with mmap_sem locked and the old page referenced, but
2457 * without the ptl held.
2458 *
2459 * High level logic flow:
2460 *
2461 * - Allocate a page, copy the content of the old page to the new one.
2462 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2463 * - Take the PTL. If the pte changed, bail out and release the allocated page
2464 * - If the pte is still the way we remember it, update the page table and all
2465 *   relevant references. This includes dropping the reference the page-table
2466 *   held to the old page, as well as updating the rmap.
2467 * - In any case, unlock the PTL and drop the reference we took to the old page.
2468 */
2469static int wp_page_copy(struct vm_fault *vmf)
2470{
2471	struct vm_area_struct *vma = vmf->vma;
2472	struct mm_struct *mm = vma->vm_mm;
2473	struct page *old_page = vmf->page;
2474	struct page *new_page = NULL;
2475	pte_t entry;
2476	int page_copied = 0;
2477	const unsigned long mmun_start = vmf->address & PAGE_MASK;
2478	const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2479	struct mem_cgroup *memcg;
2480
2481	if (unlikely(anon_vma_prepare(vma)))
2482		goto oom;
2483
2484	if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2485		new_page = alloc_zeroed_user_highpage_movable(vma,
2486							      vmf->address);
2487		if (!new_page)
2488			goto oom;
2489	} else {
2490		new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2491				vmf->address);
2492		if (!new_page)
2493			goto oom;
2494		cow_user_page(new_page, old_page, vmf->address, vma);
 
 
 
 
 
 
 
 
 
 
 
 
2495	}
2496
2497	if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2498		goto oom_free_new;
 
2499
2500	__SetPageUptodate(new_page);
2501
2502	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 
 
 
2503
2504	/*
2505	 * Re-check the pte - we dropped the lock
2506	 */
2507	vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2508	if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2509		if (old_page) {
2510			if (!PageAnon(old_page)) {
2511				dec_mm_counter_fast(mm,
2512						mm_counter_file(old_page));
2513				inc_mm_counter_fast(mm, MM_ANONPAGES);
2514			}
2515		} else {
2516			inc_mm_counter_fast(mm, MM_ANONPAGES);
2517		}
2518		flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2519		entry = mk_pte(new_page, vma->vm_page_prot);
 
2520		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2521		/*
2522		 * Clear the pte entry and flush it first, before updating the
2523		 * pte with the new entry. This will avoid a race condition
2524		 * seen in the presence of one thread doing SMC and another
2525		 * thread doing COW.
2526		 */
2527		ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2528		page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2529		mem_cgroup_commit_charge(new_page, memcg, false, false);
2530		lru_cache_add_active_or_unevictable(new_page, vma);
2531		/*
2532		 * We call the notify macro here because, when using secondary
2533		 * mmu page tables (such as kvm shadow page tables), we want the
2534		 * new page to be mapped directly into the secondary page table.
2535		 */
2536		set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2537		update_mmu_cache(vma, vmf->address, vmf->pte);
2538		if (old_page) {
2539			/*
2540			 * Only after switching the pte to the new page may
2541			 * we remove the mapcount here. Otherwise another
2542			 * process may come and find the rmap count decremented
2543			 * before the pte is switched to the new page, and
2544			 * "reuse" the old page writing into it while our pte
2545			 * here still points into it and can be read by other
2546			 * threads.
2547			 *
2548			 * The critical issue is to order this
2549			 * page_remove_rmap with the ptp_clear_flush above.
2550			 * Those stores are ordered by (if nothing else,)
2551			 * the barrier present in the atomic_add_negative
2552			 * in page_remove_rmap.
2553			 *
2554			 * Then the TLB flush in ptep_clear_flush ensures that
2555			 * no process can access the old page before the
2556			 * decremented mapcount is visible. And the old page
2557			 * cannot be reused until after the decremented
2558			 * mapcount is visible. So transitively, TLBs to
2559			 * old page will be flushed before it can be reused.
2560			 */
2561			page_remove_rmap(old_page, false);
2562		}
2563
2564		/* Free the old page.. */
2565		new_page = old_page;
2566		page_copied = 1;
2567	} else {
2568		mem_cgroup_cancel_charge(new_page, memcg, false);
2569	}
2570
2571	if (new_page)
2572		put_page(new_page);
2573
2574	pte_unmap_unlock(vmf->pte, vmf->ptl);
2575	/*
2576	 * No need to double call mmu_notifier->invalidate_range() callback as
2577	 * the above ptep_clear_flush_notify() did already call it.
2578	 */
2579	mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2580	if (old_page) {
2581		/*
2582		 * Don't let another task, with possibly unlocked vma,
2583		 * keep the mlocked page.
2584		 */
2585		if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2586			lock_page(old_page);	/* LRU manipulation */
2587			if (PageMlocked(old_page))
2588				munlock_vma_page(old_page);
2589			unlock_page(old_page);
2590		}
2591		put_page(old_page);
2592	}
2593	return page_copied ? VM_FAULT_WRITE : 0;
2594oom_free_new:
2595	put_page(new_page);
2596oom:
2597	if (old_page)
2598		put_page(old_page);
2599	return VM_FAULT_OOM;
2600}
2601
2602/**
2603 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2604 *			  writeable once the page is prepared
2605 *
2606 * @vmf: structure describing the fault
2607 *
2608 * This function handles all that is needed to finish a write page fault in a
2609 * shared mapping due to PTE being read-only once the mapped page is prepared.
2610 * It handles locking of PTE and modifying it. The function returns
2611 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2612 * lock.
2613 *
2614 * The function expects the page to be locked or other protection against
2615 * concurrent faults / writeback (such as DAX radix tree locks).
 
 
 
2616 */
2617int finish_mkwrite_fault(struct vm_fault *vmf)
2618{
2619	WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2620	vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2621				       &vmf->ptl);
2622	/*
2623	 * We might have raced with another page fault while we released the
2624	 * pte_offset_map_lock.
2625	 */
2626	if (!pte_same(*vmf->pte, vmf->orig_pte)) {
 
2627		pte_unmap_unlock(vmf->pte, vmf->ptl);
2628		return VM_FAULT_NOPAGE;
2629	}
2630	wp_page_reuse(vmf);
2631	return 0;
2632}
2633
2634/*
2635 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2636 * mapping
2637 */
2638static int wp_pfn_shared(struct vm_fault *vmf)
2639{
2640	struct vm_area_struct *vma = vmf->vma;
2641
2642	if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2643		int ret;
2644
2645		pte_unmap_unlock(vmf->pte, vmf->ptl);
2646		vmf->flags |= FAULT_FLAG_MKWRITE;
2647		ret = vma->vm_ops->pfn_mkwrite(vmf);
2648		if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2649			return ret;
2650		return finish_mkwrite_fault(vmf);
2651	}
2652	wp_page_reuse(vmf);
2653	return VM_FAULT_WRITE;
2654}
2655
2656static int wp_page_shared(struct vm_fault *vmf)
2657	__releases(vmf->ptl)
2658{
2659	struct vm_area_struct *vma = vmf->vma;
 
2660
2661	get_page(vmf->page);
2662
2663	if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2664		int tmp;
2665
2666		pte_unmap_unlock(vmf->pte, vmf->ptl);
2667		tmp = do_page_mkwrite(vmf);
2668		if (unlikely(!tmp || (tmp &
2669				      (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2670			put_page(vmf->page);
2671			return tmp;
2672		}
2673		tmp = finish_mkwrite_fault(vmf);
2674		if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2675			unlock_page(vmf->page);
2676			put_page(vmf->page);
2677			return tmp;
2678		}
2679	} else {
2680		wp_page_reuse(vmf);
2681		lock_page(vmf->page);
2682	}
2683	fault_dirty_shared_page(vma, vmf->page);
2684	put_page(vmf->page);
2685
2686	return VM_FAULT_WRITE;
2687}
2688
2689/*
2690 * This routine handles present pages, when users try to write
2691 * to a shared page. It is done by copying the page to a new address
2692 * and decrementing the shared-page counter for the old page.
2693 *
2694 * Note that this routine assumes that the protection checks have been
2695 * done by the caller (the low-level page fault routine in most cases).
2696 * Thus we can safely just mark it writable once we've done any necessary
2697 * COW.
2698 *
2699 * We also mark the page dirty at this point even though the page will
2700 * change only once the write actually happens. This avoids a few races,
2701 * and potentially makes it more efficient.
2702 *
2703 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2704 * but allow concurrent faults), with pte both mapped and locked.
2705 * We return with mmap_sem still held, but pte unmapped and unlocked.
2706 */
2707static int do_wp_page(struct vm_fault *vmf)
2708	__releases(vmf->ptl)
2709{
2710	struct vm_area_struct *vma = vmf->vma;
2711
 
 
 
 
 
2712	vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2713	if (!vmf->page) {
2714		/*
2715		 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2716		 * VM_PFNMAP VMA.
2717		 *
2718		 * We should not cow pages in a shared writeable mapping.
2719		 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2720		 */
2721		if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2722				     (VM_WRITE|VM_SHARED))
2723			return wp_pfn_shared(vmf);
2724
2725		pte_unmap_unlock(vmf->pte, vmf->ptl);
2726		return wp_page_copy(vmf);
2727	}
2728
2729	/*
2730	 * Take out anonymous pages first, anonymous shared vmas are
2731	 * not dirty accountable.
2732	 */
2733	if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2734		int total_map_swapcount;
2735		if (!trylock_page(vmf->page)) {
2736			get_page(vmf->page);
2737			pte_unmap_unlock(vmf->pte, vmf->ptl);
2738			lock_page(vmf->page);
2739			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2740					vmf->address, &vmf->ptl);
2741			if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2742				unlock_page(vmf->page);
2743				pte_unmap_unlock(vmf->pte, vmf->ptl);
2744				put_page(vmf->page);
2745				return 0;
2746			}
2747			put_page(vmf->page);
2748		}
2749		if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2750			if (total_map_swapcount == 1) {
2751				/*
2752				 * The page is all ours. Move it to
2753				 * our anon_vma so the rmap code will
2754				 * not search our parent or siblings.
2755				 * Protected against the rmap code by
2756				 * the page lock.
2757				 */
2758				page_move_anon_rmap(vmf->page, vma);
2759			}
2760			unlock_page(vmf->page);
2761			wp_page_reuse(vmf);
2762			return VM_FAULT_WRITE;
2763		}
2764		unlock_page(vmf->page);
2765	} else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2766					(VM_WRITE|VM_SHARED))) {
2767		return wp_page_shared(vmf);
2768	}
2769
2770	/*
2771	 * Ok, we need to copy. Oh, well..
2772	 */
2773	get_page(vmf->page);
2774
2775	pte_unmap_unlock(vmf->pte, vmf->ptl);
2776	return wp_page_copy(vmf);
2777}
2778
2779static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2780		unsigned long start_addr, unsigned long end_addr,
2781		struct zap_details *details)
2782{
2783	zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2784}
2785
2786static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2787					    struct zap_details *details)
2788{
2789	struct vm_area_struct *vma;
2790	pgoff_t vba, vea, zba, zea;
2791
2792	vma_interval_tree_foreach(vma, root,
2793			details->first_index, details->last_index) {
2794
2795		vba = vma->vm_pgoff;
2796		vea = vba + vma_pages(vma) - 1;
2797		zba = details->first_index;
2798		if (zba < vba)
2799			zba = vba;
2800		zea = details->last_index;
2801		if (zea > vea)
2802			zea = vea;
2803
2804		unmap_mapping_range_vma(vma,
2805			((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2806			((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2807				details);
2808	}
2809}
2810
2811/**
2812 * unmap_mapping_pages() - Unmap pages from processes.
2813 * @mapping: The address space containing pages to be unmapped.
2814 * @start: Index of first page to be unmapped.
2815 * @nr: Number of pages to be unmapped.  0 to unmap to end of file.
2816 * @even_cows: Whether to unmap even private COWed pages.
2817 *
2818 * Unmap the pages in this address space from any userspace process which
2819 * has them mmaped.  Generally, you want to remove COWed pages as well when
2820 * a file is being truncated, but not when invalidating pages from the page
2821 * cache.
2822 */
2823void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2824		pgoff_t nr, bool even_cows)
2825{
2826	struct zap_details details = { };
2827
2828	details.check_mapping = even_cows ? NULL : mapping;
2829	details.first_index = start;
2830	details.last_index = start + nr - 1;
2831	if (details.last_index < details.first_index)
2832		details.last_index = ULONG_MAX;
2833
2834	i_mmap_lock_write(mapping);
2835	if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2836		unmap_mapping_range_tree(&mapping->i_mmap, &details);
2837	i_mmap_unlock_write(mapping);
2838}
2839
2840/**
2841 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2842 * address_space corresponding to the specified byte range in the underlying
2843 * file.
2844 *
2845 * @mapping: the address space containing mmaps to be unmapped.
2846 * @holebegin: byte in first page to unmap, relative to the start of
2847 * the underlying file.  This will be rounded down to a PAGE_SIZE
2848 * boundary.  Note that this is different from truncate_pagecache(), which
2849 * must keep the partial page.  In contrast, we must get rid of
2850 * partial pages.
2851 * @holelen: size of prospective hole in bytes.  This will be rounded
2852 * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2853 * end of the file.
2854 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2855 * but 0 when invalidating pagecache, don't throw away private data.
2856 */
2857void unmap_mapping_range(struct address_space *mapping,
2858		loff_t const holebegin, loff_t const holelen, int even_cows)
2859{
2860	pgoff_t hba = holebegin >> PAGE_SHIFT;
2861	pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2862
2863	/* Check for overflow. */
2864	if (sizeof(holelen) > sizeof(hlen)) {
2865		long long holeend =
2866			(holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2867		if (holeend & ~(long long)ULONG_MAX)
2868			hlen = ULONG_MAX - hba + 1;
2869	}
2870
2871	unmap_mapping_pages(mapping, hba, hlen, even_cows);
2872}
2873EXPORT_SYMBOL(unmap_mapping_range);
2874
2875/*
2876 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2877 * but allow concurrent faults), and pte mapped but not yet locked.
2878 * We return with pte unmapped and unlocked.
2879 *
2880 * We return with the mmap_sem locked or unlocked in the same cases
2881 * as does filemap_fault().
2882 */
2883int do_swap_page(struct vm_fault *vmf)
2884{
2885	struct vm_area_struct *vma = vmf->vma;
2886	struct page *page = NULL, *swapcache;
2887	struct mem_cgroup *memcg;
2888	swp_entry_t entry;
2889	pte_t pte;
2890	int locked;
2891	int exclusive = 0;
2892	int ret = 0;
 
2893
2894	if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2895		goto out;
2896
2897	entry = pte_to_swp_entry(vmf->orig_pte);
2898	if (unlikely(non_swap_entry(entry))) {
2899		if (is_migration_entry(entry)) {
2900			migration_entry_wait(vma->vm_mm, vmf->pmd,
2901					     vmf->address);
2902		} else if (is_device_private_entry(entry)) {
2903			/*
2904			 * For un-addressable device memory we call the pgmap
2905			 * fault handler callback. The callback must migrate
2906			 * the page back to some CPU accessible page.
2907			 */
2908			ret = device_private_entry_fault(vma, vmf->address, entry,
2909						 vmf->flags, vmf->pmd);
2910		} else if (is_hwpoison_entry(entry)) {
2911			ret = VM_FAULT_HWPOISON;
2912		} else {
2913			print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2914			ret = VM_FAULT_SIGBUS;
2915		}
2916		goto out;
2917	}
2918
2919
2920	delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2921	page = lookup_swap_cache(entry, vma, vmf->address);
2922	swapcache = page;
2923
2924	if (!page) {
2925		struct swap_info_struct *si = swp_swap_info(entry);
2926
2927		if (si->flags & SWP_SYNCHRONOUS_IO &&
2928				__swap_count(si, entry) == 1) {
2929			/* skip swapcache */
2930			page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2931							vmf->address);
2932			if (page) {
 
 
2933				__SetPageLocked(page);
2934				__SetPageSwapBacked(page);
2935				set_page_private(page, entry.val);
2936				lru_cache_add_anon(page);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2937				swap_readpage(page, true);
2938			}
2939		} else {
2940			page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2941						vmf);
2942			swapcache = page;
2943		}
2944
2945		if (!page) {
2946			/*
2947			 * Back out if somebody else faulted in this pte
2948			 * while we released the pte lock.
2949			 */
2950			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2951					vmf->address, &vmf->ptl);
2952			if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2953				ret = VM_FAULT_OOM;
2954			delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2955			goto unlock;
2956		}
2957
2958		/* Had to read the page from swap area: Major fault */
2959		ret = VM_FAULT_MAJOR;
2960		count_vm_event(PGMAJFAULT);
2961		count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2962	} else if (PageHWPoison(page)) {
2963		/*
2964		 * hwpoisoned dirty swapcache pages are kept for killing
2965		 * owner processes (which may be unknown at hwpoison time)
2966		 */
2967		ret = VM_FAULT_HWPOISON;
2968		delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2969		goto out_release;
2970	}
2971
2972	locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2973
2974	delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2975	if (!locked) {
2976		ret |= VM_FAULT_RETRY;
2977		goto out_release;
2978	}
2979
2980	/*
2981	 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2982	 * release the swapcache from under us.  The page pin, and pte_same
2983	 * test below, are not enough to exclude that.  Even if it is still
2984	 * swapcache, we need to check that the page's swap has not changed.
2985	 */
2986	if (unlikely((!PageSwapCache(page) ||
2987			page_private(page) != entry.val)) && swapcache)
2988		goto out_page;
2989
2990	page = ksm_might_need_to_copy(page, vma, vmf->address);
2991	if (unlikely(!page)) {
2992		ret = VM_FAULT_OOM;
2993		page = swapcache;
2994		goto out_page;
2995	}
2996
2997	if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2998				&memcg, false)) {
2999		ret = VM_FAULT_OOM;
3000		goto out_page;
3001	}
3002
3003	/*
3004	 * Back out if somebody else already faulted in this pte.
3005	 */
3006	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3007			&vmf->ptl);
3008	if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3009		goto out_nomap;
3010
3011	if (unlikely(!PageUptodate(page))) {
3012		ret = VM_FAULT_SIGBUS;
3013		goto out_nomap;
3014	}
3015
3016	/*
3017	 * The page isn't present yet, go ahead with the fault.
3018	 *
3019	 * Be careful about the sequence of operations here.
3020	 * To get its accounting right, reuse_swap_page() must be called
3021	 * while the page is counted on swap but not yet in mapcount i.e.
3022	 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3023	 * must be called after the swap_free(), or it will never succeed.
3024	 */
3025
3026	inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3027	dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3028	pte = mk_pte(page, vma->vm_page_prot);
3029	if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3030		pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3031		vmf->flags &= ~FAULT_FLAG_WRITE;
3032		ret |= VM_FAULT_WRITE;
3033		exclusive = RMAP_EXCLUSIVE;
3034	}
3035	flush_icache_page(vma, page);
3036	if (pte_swp_soft_dirty(vmf->orig_pte))
3037		pte = pte_mksoft_dirty(pte);
 
 
 
 
3038	set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3039	arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3040	vmf->orig_pte = pte;
3041
3042	/* ksm created a completely new copy */
3043	if (unlikely(page != swapcache && swapcache)) {
3044		page_add_new_anon_rmap(page, vma, vmf->address, false);
3045		mem_cgroup_commit_charge(page, memcg, false, false);
3046		lru_cache_add_active_or_unevictable(page, vma);
3047	} else {
3048		do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3049		mem_cgroup_commit_charge(page, memcg, true, false);
3050		activate_page(page);
3051	}
3052
3053	swap_free(entry);
3054	if (mem_cgroup_swap_full(page) ||
3055	    (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3056		try_to_free_swap(page);
3057	unlock_page(page);
3058	if (page != swapcache && swapcache) {
3059		/*
3060		 * Hold the lock to avoid the swap entry to be reused
3061		 * until we take the PT lock for the pte_same() check
3062		 * (to avoid false positives from pte_same). For
3063		 * further safety release the lock after the swap_free
3064		 * so that the swap count won't change under a
3065		 * parallel locked swapcache.
3066		 */
3067		unlock_page(swapcache);
3068		put_page(swapcache);
3069	}
3070
3071	if (vmf->flags & FAULT_FLAG_WRITE) {
3072		ret |= do_wp_page(vmf);
3073		if (ret & VM_FAULT_ERROR)
3074			ret &= VM_FAULT_ERROR;
3075		goto out;
3076	}
3077
3078	/* No need to invalidate - it was non-present before */
3079	update_mmu_cache(vma, vmf->address, vmf->pte);
3080unlock:
3081	pte_unmap_unlock(vmf->pte, vmf->ptl);
3082out:
3083	return ret;
3084out_nomap:
3085	mem_cgroup_cancel_charge(page, memcg, false);
3086	pte_unmap_unlock(vmf->pte, vmf->ptl);
3087out_page:
3088	unlock_page(page);
3089out_release:
3090	put_page(page);
3091	if (page != swapcache && swapcache) {
3092		unlock_page(swapcache);
3093		put_page(swapcache);
3094	}
3095	return ret;
3096}
3097
3098/*
3099 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3100 * but allow concurrent faults), and pte mapped but not yet locked.
3101 * We return with mmap_sem still held, but pte unmapped and unlocked.
3102 */
3103static int do_anonymous_page(struct vm_fault *vmf)
3104{
3105	struct vm_area_struct *vma = vmf->vma;
3106	struct mem_cgroup *memcg;
3107	struct page *page;
3108	int ret = 0;
3109	pte_t entry;
3110
3111	/* File mapping without ->vm_ops ? */
3112	if (vma->vm_flags & VM_SHARED)
3113		return VM_FAULT_SIGBUS;
3114
3115	/*
3116	 * Use pte_alloc() instead of pte_alloc_map().  We can't run
3117	 * pte_offset_map() on pmds where a huge pmd might be created
3118	 * from a different thread.
3119	 *
3120	 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3121	 * parallel threads are excluded by other means.
3122	 *
3123	 * Here we only have down_read(mmap_sem).
3124	 */
3125	if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3126		return VM_FAULT_OOM;
3127
3128	/* See the comment in pte_alloc_one_map() */
3129	if (unlikely(pmd_trans_unstable(vmf->pmd)))
3130		return 0;
3131
3132	/* Use the zero-page for reads */
3133	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3134			!mm_forbids_zeropage(vma->vm_mm)) {
3135		entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3136						vma->vm_page_prot));
3137		vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3138				vmf->address, &vmf->ptl);
3139		if (!pte_none(*vmf->pte))
 
3140			goto unlock;
 
3141		ret = check_stable_address_space(vma->vm_mm);
3142		if (ret)
3143			goto unlock;
3144		/* Deliver the page fault to userland, check inside PT lock */
3145		if (userfaultfd_missing(vma)) {
3146			pte_unmap_unlock(vmf->pte, vmf->ptl);
3147			return handle_userfault(vmf, VM_UFFD_MISSING);
3148		}
3149		goto setpte;
3150	}
3151
3152	/* Allocate our own private page. */
3153	if (unlikely(anon_vma_prepare(vma)))
3154		goto oom;
3155	page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3156	if (!page)
3157		goto oom;
3158
3159	if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
3160		goto oom_free_page;
 
3161
3162	/*
3163	 * The memory barrier inside __SetPageUptodate makes sure that
3164	 * preceeding stores to the page contents become visible before
3165	 * the set_pte_at() write.
3166	 */
3167	__SetPageUptodate(page);
3168
3169	entry = mk_pte(page, vma->vm_page_prot);
 
3170	if (vma->vm_flags & VM_WRITE)
3171		entry = pte_mkwrite(pte_mkdirty(entry));
3172
3173	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3174			&vmf->ptl);
3175	if (!pte_none(*vmf->pte))
 
3176		goto release;
 
3177
3178	ret = check_stable_address_space(vma->vm_mm);
3179	if (ret)
3180		goto release;
3181
3182	/* Deliver the page fault to userland, check inside PT lock */
3183	if (userfaultfd_missing(vma)) {
3184		pte_unmap_unlock(vmf->pte, vmf->ptl);
3185		mem_cgroup_cancel_charge(page, memcg, false);
3186		put_page(page);
3187		return handle_userfault(vmf, VM_UFFD_MISSING);
3188	}
3189
3190	inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3191	page_add_new_anon_rmap(page, vma, vmf->address, false);
3192	mem_cgroup_commit_charge(page, memcg, false, false);
3193	lru_cache_add_active_or_unevictable(page, vma);
3194setpte:
3195	set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3196
3197	/* No need to invalidate - it was non-present before */
3198	update_mmu_cache(vma, vmf->address, vmf->pte);
3199unlock:
3200	pte_unmap_unlock(vmf->pte, vmf->ptl);
3201	return ret;
3202release:
3203	mem_cgroup_cancel_charge(page, memcg, false);
3204	put_page(page);
3205	goto unlock;
3206oom_free_page:
3207	put_page(page);
3208oom:
3209	return VM_FAULT_OOM;
3210}
3211
3212/*
3213 * The mmap_sem must have been held on entry, and may have been
3214 * released depending on flags and vma->vm_ops->fault() return value.
3215 * See filemap_fault() and __lock_page_retry().
3216 */
3217static int __do_fault(struct vm_fault *vmf)
3218{
3219	struct vm_area_struct *vma = vmf->vma;
3220	int ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3221
3222	ret = vma->vm_ops->fault(vmf);
3223	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3224			    VM_FAULT_DONE_COW)))
3225		return ret;
3226
3227	if (unlikely(PageHWPoison(vmf->page))) {
3228		if (ret & VM_FAULT_LOCKED)
3229			unlock_page(vmf->page);
3230		put_page(vmf->page);
3231		vmf->page = NULL;
3232		return VM_FAULT_HWPOISON;
3233	}
3234
3235	if (unlikely(!(ret & VM_FAULT_LOCKED)))
3236		lock_page(vmf->page);
3237	else
3238		VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3239
3240	return ret;
3241}
3242
3243/*
3244 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3245 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3246 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3247 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3248 */
3249static int pmd_devmap_trans_unstable(pmd_t *pmd)
3250{
3251	return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3252}
3253
3254static int pte_alloc_one_map(struct vm_fault *vmf)
3255{
3256	struct vm_area_struct *vma = vmf->vma;
3257
3258	if (!pmd_none(*vmf->pmd))
3259		goto map_pte;
3260	if (vmf->prealloc_pte) {
3261		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3262		if (unlikely(!pmd_none(*vmf->pmd))) {
3263			spin_unlock(vmf->ptl);
3264			goto map_pte;
3265		}
3266
3267		mm_inc_nr_ptes(vma->vm_mm);
3268		pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3269		spin_unlock(vmf->ptl);
3270		vmf->prealloc_pte = NULL;
3271	} else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3272		return VM_FAULT_OOM;
3273	}
3274map_pte:
3275	/*
3276	 * If a huge pmd materialized under us just retry later.  Use
3277	 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3278	 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3279	 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3280	 * running immediately after a huge pmd fault in a different thread of
3281	 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3282	 * All we have to ensure is that it is a regular pmd that we can walk
3283	 * with pte_offset_map() and we can do that through an atomic read in
3284	 * C, which is what pmd_trans_unstable() provides.
3285	 */
3286	if (pmd_devmap_trans_unstable(vmf->pmd))
3287		return VM_FAULT_NOPAGE;
3288
3289	/*
3290	 * At this point we know that our vmf->pmd points to a page of ptes
3291	 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3292	 * for the duration of the fault.  If a racing MADV_DONTNEED runs and
3293	 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3294	 * be valid and we will re-check to make sure the vmf->pte isn't
3295	 * pte_none() under vmf->ptl protection when we return to
3296	 * alloc_set_pte().
3297	 */
3298	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3299			&vmf->ptl);
3300	return 0;
3301}
3302
3303#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3304
3305#define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3306static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3307		unsigned long haddr)
3308{
3309	if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3310			(vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3311		return false;
3312	if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3313		return false;
3314	return true;
3315}
3316
3317static void deposit_prealloc_pte(struct vm_fault *vmf)
3318{
3319	struct vm_area_struct *vma = vmf->vma;
3320
3321	pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3322	/*
3323	 * We are going to consume the prealloc table,
3324	 * count that as nr_ptes.
3325	 */
3326	mm_inc_nr_ptes(vma->vm_mm);
3327	vmf->prealloc_pte = NULL;
3328}
3329
3330static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3331{
3332	struct vm_area_struct *vma = vmf->vma;
3333	bool write = vmf->flags & FAULT_FLAG_WRITE;
3334	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3335	pmd_t entry;
3336	int i, ret;
 
3337
3338	if (!transhuge_vma_suitable(vma, haddr))
3339		return VM_FAULT_FALLBACK;
3340
3341	ret = VM_FAULT_FALLBACK;
3342	page = compound_head(page);
3343
3344	/*
3345	 * Archs like ppc64 need additonal space to store information
3346	 * related to pte entry. Use the preallocated table for that.
3347	 */
3348	if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3349		vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3350		if (!vmf->prealloc_pte)
3351			return VM_FAULT_OOM;
3352		smp_wmb(); /* See comment in __pte_alloc() */
3353	}
3354
3355	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3356	if (unlikely(!pmd_none(*vmf->pmd)))
3357		goto out;
3358
3359	for (i = 0; i < HPAGE_PMD_NR; i++)
3360		flush_icache_page(vma, page + i);
3361
3362	entry = mk_huge_pmd(page, vma->vm_page_prot);
3363	if (write)
3364		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3365
3366	add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3367	page_add_file_rmap(page, true);
3368	/*
3369	 * deposit and withdraw with pmd lock held
3370	 */
3371	if (arch_needs_pgtable_deposit())
3372		deposit_prealloc_pte(vmf);
3373
3374	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3375
3376	update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3377
3378	/* fault is handled */
3379	ret = 0;
3380	count_vm_event(THP_FILE_MAPPED);
3381out:
3382	spin_unlock(vmf->ptl);
3383	return ret;
3384}
3385#else
3386static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3387{
3388	BUILD_BUG();
3389	return 0;
3390}
3391#endif
3392
3393/**
3394 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3395 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3396 *
3397 * @vmf: fault environment
3398 * @memcg: memcg to charge page (only for private mappings)
3399 * @page: page to map
3400 *
3401 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3402 * return.
3403 *
3404 * Target users are page handler itself and implementations of
3405 * vm_ops->map_pages.
 
 
3406 */
3407int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3408		struct page *page)
3409{
3410	struct vm_area_struct *vma = vmf->vma;
3411	bool write = vmf->flags & FAULT_FLAG_WRITE;
3412	pte_t entry;
3413	int ret;
3414
3415	if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3416			IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3417		/* THP on COW? */
3418		VM_BUG_ON_PAGE(memcg, page);
3419
 
3420		ret = do_set_pmd(vmf, page);
3421		if (ret != VM_FAULT_FALLBACK)
3422			return ret;
3423	}
3424
3425	if (!vmf->pte) {
3426		ret = pte_alloc_one_map(vmf);
3427		if (ret)
3428			return ret;
3429	}
3430
3431	/* Re-check under ptl */
3432	if (unlikely(!pte_none(*vmf->pte)))
 
3433		return VM_FAULT_NOPAGE;
 
3434
3435	flush_icache_page(vma, page);
3436	entry = mk_pte(page, vma->vm_page_prot);
 
3437	if (write)
3438		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3439	/* copy-on-write page */
3440	if (write && !(vma->vm_flags & VM_SHARED)) {
3441		inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3442		page_add_new_anon_rmap(page, vma, vmf->address, false);
3443		mem_cgroup_commit_charge(page, memcg, false, false);
3444		lru_cache_add_active_or_unevictable(page, vma);
3445	} else {
3446		inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3447		page_add_file_rmap(page, false);
3448	}
3449	set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3450
3451	/* no need to invalidate: a not-present page won't be cached */
3452	update_mmu_cache(vma, vmf->address, vmf->pte);
3453
3454	return 0;
3455}
3456
3457
3458/**
3459 * finish_fault - finish page fault once we have prepared the page to fault
3460 *
3461 * @vmf: structure describing the fault
3462 *
3463 * This function handles all that is needed to finish a page fault once the
3464 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3465 * given page, adds reverse page mapping, handles memcg charges and LRU
3466 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3467 * error.
3468 *
3469 * The function expects the page to be locked and on success it consumes a
3470 * reference of a page being mapped (for the PTE which maps it).
 
 
3471 */
3472int finish_fault(struct vm_fault *vmf)
3473{
3474	struct page *page;
3475	int ret = 0;
3476
3477	/* Did we COW the page? */
3478	if ((vmf->flags & FAULT_FLAG_WRITE) &&
3479	    !(vmf->vma->vm_flags & VM_SHARED))
3480		page = vmf->cow_page;
3481	else
3482		page = vmf->page;
3483
3484	/*
3485	 * check even for read faults because we might have lost our CoWed
3486	 * page
3487	 */
3488	if (!(vmf->vma->vm_flags & VM_SHARED))
3489		ret = check_stable_address_space(vmf->vma->vm_mm);
3490	if (!ret)
3491		ret = alloc_set_pte(vmf, vmf->memcg, page);
3492	if (vmf->pte)
3493		pte_unmap_unlock(vmf->pte, vmf->ptl);
3494	return ret;
3495}
3496
3497static unsigned long fault_around_bytes __read_mostly =
3498	rounddown_pow_of_two(65536);
3499
3500#ifdef CONFIG_DEBUG_FS
3501static int fault_around_bytes_get(void *data, u64 *val)
3502{
3503	*val = fault_around_bytes;
3504	return 0;
3505}
3506
3507/*
3508 * fault_around_bytes must be rounded down to the nearest page order as it's
3509 * what do_fault_around() expects to see.
3510 */
3511static int fault_around_bytes_set(void *data, u64 val)
3512{
3513	if (val / PAGE_SIZE > PTRS_PER_PTE)
3514		return -EINVAL;
3515	if (val > PAGE_SIZE)
3516		fault_around_bytes = rounddown_pow_of_two(val);
3517	else
3518		fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3519	return 0;
3520}
3521DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3522		fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3523
3524static int __init fault_around_debugfs(void)
3525{
3526	void *ret;
3527
3528	ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3529			&fault_around_bytes_fops);
3530	if (!ret)
3531		pr_warn("Failed to create fault_around_bytes in debugfs");
3532	return 0;
3533}
3534late_initcall(fault_around_debugfs);
3535#endif
3536
3537/*
3538 * do_fault_around() tries to map few pages around the fault address. The hope
3539 * is that the pages will be needed soon and this will lower the number of
3540 * faults to handle.
3541 *
3542 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3543 * not ready to be mapped: not up-to-date, locked, etc.
3544 *
3545 * This function is called with the page table lock taken. In the split ptlock
3546 * case the page table lock only protects only those entries which belong to
3547 * the page table corresponding to the fault address.
3548 *
3549 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3550 * only once.
3551 *
3552 * fault_around_bytes defines how many bytes we'll try to map.
3553 * do_fault_around() expects it to be set to a power of two less than or equal
3554 * to PTRS_PER_PTE.
3555 *
3556 * The virtual address of the area that we map is naturally aligned to
3557 * fault_around_bytes rounded down to the machine page size
3558 * (and therefore to page order).  This way it's easier to guarantee
3559 * that we don't cross page table boundaries.
3560 */
3561static int do_fault_around(struct vm_fault *vmf)
3562{
3563	unsigned long address = vmf->address, nr_pages, mask;
3564	pgoff_t start_pgoff = vmf->pgoff;
3565	pgoff_t end_pgoff;
3566	int off, ret = 0;
 
3567
3568	nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3569	mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3570
3571	vmf->address = max(address & mask, vmf->vma->vm_start);
3572	off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3573	start_pgoff -= off;
3574
3575	/*
3576	 *  end_pgoff is either the end of the page table, the end of
3577	 *  the vma or nr_pages from start_pgoff, depending what is nearest.
3578	 */
3579	end_pgoff = start_pgoff -
3580		((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3581		PTRS_PER_PTE - 1;
3582	end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3583			start_pgoff + nr_pages - 1);
3584
3585	if (pmd_none(*vmf->pmd)) {
3586		vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3587						  vmf->address);
3588		if (!vmf->prealloc_pte)
3589			goto out;
3590		smp_wmb(); /* See comment in __pte_alloc() */
3591	}
3592
3593	vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3594
3595	/* Huge page is mapped? Page fault is solved */
3596	if (pmd_trans_huge(*vmf->pmd)) {
3597		ret = VM_FAULT_NOPAGE;
3598		goto out;
3599	}
3600
3601	/* ->map_pages() haven't done anything useful. Cold page cache? */
3602	if (!vmf->pte)
3603		goto out;
3604
3605	/* check if the page fault is solved */
3606	vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3607	if (!pte_none(*vmf->pte))
3608		ret = VM_FAULT_NOPAGE;
3609	pte_unmap_unlock(vmf->pte, vmf->ptl);
3610out:
3611	vmf->address = address;
3612	vmf->pte = NULL;
3613	return ret;
3614}
3615
3616static int do_read_fault(struct vm_fault *vmf)
3617{
3618	struct vm_area_struct *vma = vmf->vma;
3619	int ret = 0;
3620
3621	/*
3622	 * Let's call ->map_pages() first and use ->fault() as fallback
3623	 * if page by the offset is not ready to be mapped (cold cache or
3624	 * something).
3625	 */
3626	if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3627		ret = do_fault_around(vmf);
3628		if (ret)
3629			return ret;
3630	}
3631
3632	ret = __do_fault(vmf);
3633	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3634		return ret;
3635
3636	ret |= finish_fault(vmf);
3637	unlock_page(vmf->page);
3638	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3639		put_page(vmf->page);
3640	return ret;
3641}
3642
3643static int do_cow_fault(struct vm_fault *vmf)
3644{
3645	struct vm_area_struct *vma = vmf->vma;
3646	int ret;
3647
3648	if (unlikely(anon_vma_prepare(vma)))
3649		return VM_FAULT_OOM;
3650
3651	vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3652	if (!vmf->cow_page)
3653		return VM_FAULT_OOM;
3654
3655	if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3656				&vmf->memcg, false)) {
3657		put_page(vmf->cow_page);
3658		return VM_FAULT_OOM;
3659	}
 
3660
3661	ret = __do_fault(vmf);
3662	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3663		goto uncharge_out;
3664	if (ret & VM_FAULT_DONE_COW)
3665		return ret;
3666
3667	copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3668	__SetPageUptodate(vmf->cow_page);
3669
3670	ret |= finish_fault(vmf);
3671	unlock_page(vmf->page);
3672	put_page(vmf->page);
3673	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3674		goto uncharge_out;
3675	return ret;
3676uncharge_out:
3677	mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3678	put_page(vmf->cow_page);
3679	return ret;
3680}
3681
3682static int do_shared_fault(struct vm_fault *vmf)
3683{
3684	struct vm_area_struct *vma = vmf->vma;
3685	int ret, tmp;
3686
3687	ret = __do_fault(vmf);
3688	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3689		return ret;
3690
3691	/*
3692	 * Check if the backing address space wants to know that the page is
3693	 * about to become writable
3694	 */
3695	if (vma->vm_ops->page_mkwrite) {
3696		unlock_page(vmf->page);
3697		tmp = do_page_mkwrite(vmf);
3698		if (unlikely(!tmp ||
3699				(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3700			put_page(vmf->page);
3701			return tmp;
3702		}
3703	}
3704
3705	ret |= finish_fault(vmf);
3706	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3707					VM_FAULT_RETRY))) {
3708		unlock_page(vmf->page);
3709		put_page(vmf->page);
3710		return ret;
3711	}
3712
3713	fault_dirty_shared_page(vma, vmf->page);
3714	return ret;
3715}
3716
3717/*
3718 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3719 * but allow concurrent faults).
3720 * The mmap_sem may have been released depending on flags and our
3721 * return value.  See filemap_fault() and __lock_page_or_retry().
 
 
3722 */
3723static int do_fault(struct vm_fault *vmf)
3724{
3725	struct vm_area_struct *vma = vmf->vma;
3726	int ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3727
3728	/* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3729	if (!vma->vm_ops->fault)
3730		ret = VM_FAULT_SIGBUS;
3731	else if (!(vmf->flags & FAULT_FLAG_WRITE))
3732		ret = do_read_fault(vmf);
3733	else if (!(vma->vm_flags & VM_SHARED))
3734		ret = do_cow_fault(vmf);
3735	else
3736		ret = do_shared_fault(vmf);
3737
3738	/* preallocated pagetable is unused: free it */
3739	if (vmf->prealloc_pte) {
3740		pte_free(vma->vm_mm, vmf->prealloc_pte);
3741		vmf->prealloc_pte = NULL;
3742	}
3743	return ret;
3744}
3745
3746static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3747				unsigned long addr, int page_nid,
3748				int *flags)
3749{
3750	get_page(page);
3751
3752	count_vm_numa_event(NUMA_HINT_FAULTS);
3753	if (page_nid == numa_node_id()) {
3754		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3755		*flags |= TNF_FAULT_LOCAL;
3756	}
3757
3758	return mpol_misplaced(page, vma, addr);
3759}
3760
3761static int do_numa_page(struct vm_fault *vmf)
3762{
3763	struct vm_area_struct *vma = vmf->vma;
3764	struct page *page = NULL;
3765	int page_nid = -1;
3766	int last_cpupid;
3767	int target_nid;
3768	bool migrated = false;
3769	pte_t pte;
3770	bool was_writable = pte_savedwrite(vmf->orig_pte);
3771	int flags = 0;
3772
3773	/*
3774	 * The "pte" at this point cannot be used safely without
3775	 * validation through pte_unmap_same(). It's of NUMA type but
3776	 * the pfn may be screwed if the read is non atomic.
3777	 */
3778	vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3779	spin_lock(vmf->ptl);
3780	if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3781		pte_unmap_unlock(vmf->pte, vmf->ptl);
3782		goto out;
3783	}
3784
3785	/*
3786	 * Make it present again, Depending on how arch implementes non
3787	 * accessible ptes, some can allow access by kernel mode.
3788	 */
3789	pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3790	pte = pte_modify(pte, vma->vm_page_prot);
3791	pte = pte_mkyoung(pte);
3792	if (was_writable)
3793		pte = pte_mkwrite(pte);
3794	ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3795	update_mmu_cache(vma, vmf->address, vmf->pte);
3796
3797	page = vm_normal_page(vma, vmf->address, pte);
3798	if (!page) {
3799		pte_unmap_unlock(vmf->pte, vmf->ptl);
3800		return 0;
3801	}
3802
3803	/* TODO: handle PTE-mapped THP */
3804	if (PageCompound(page)) {
3805		pte_unmap_unlock(vmf->pte, vmf->ptl);
3806		return 0;
3807	}
3808
3809	/*
3810	 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3811	 * much anyway since they can be in shared cache state. This misses
3812	 * the case where a mapping is writable but the process never writes
3813	 * to it but pte_write gets cleared during protection updates and
3814	 * pte_dirty has unpredictable behaviour between PTE scan updates,
3815	 * background writeback, dirty balancing and application behaviour.
3816	 */
3817	if (!pte_write(pte))
3818		flags |= TNF_NO_GROUP;
3819
3820	/*
3821	 * Flag if the page is shared between multiple address spaces. This
3822	 * is later used when determining whether to group tasks together
3823	 */
3824	if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3825		flags |= TNF_SHARED;
3826
3827	last_cpupid = page_cpupid_last(page);
3828	page_nid = page_to_nid(page);
3829	target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3830			&flags);
3831	pte_unmap_unlock(vmf->pte, vmf->ptl);
3832	if (target_nid == -1) {
3833		put_page(page);
3834		goto out;
3835	}
3836
3837	/* Migrate to the requested node */
3838	migrated = migrate_misplaced_page(page, vma, target_nid);
3839	if (migrated) {
3840		page_nid = target_nid;
3841		flags |= TNF_MIGRATED;
3842	} else
3843		flags |= TNF_MIGRATE_FAIL;
3844
3845out:
3846	if (page_nid != -1)
3847		task_numa_fault(last_cpupid, page_nid, 1, flags);
3848	return 0;
3849}
3850
3851static inline int create_huge_pmd(struct vm_fault *vmf)
3852{
3853	if (vma_is_anonymous(vmf->vma))
3854		return do_huge_pmd_anonymous_page(vmf);
3855	if (vmf->vma->vm_ops->huge_fault)
3856		return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3857	return VM_FAULT_FALLBACK;
3858}
3859
3860/* `inline' is required to avoid gcc 4.1.2 build error */
3861static inline int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3862{
3863	if (vma_is_anonymous(vmf->vma))
 
 
3864		return do_huge_pmd_wp_page(vmf, orig_pmd);
3865	if (vmf->vma->vm_ops->huge_fault)
3866		return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
 
 
 
 
 
3867
3868	/* COW handled on pte level: split pmd */
3869	VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3870	__split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3871
3872	return VM_FAULT_FALLBACK;
3873}
3874
3875static inline bool vma_is_accessible(struct vm_area_struct *vma)
3876{
3877	return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3878}
3879
3880static int create_huge_pud(struct vm_fault *vmf)
3881{
3882#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3883	/* No support for anonymous transparent PUD pages yet */
3884	if (vma_is_anonymous(vmf->vma))
3885		return VM_FAULT_FALLBACK;
3886	if (vmf->vma->vm_ops->huge_fault)
3887		return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
 
 
 
 
 
 
 
3888#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3889	return VM_FAULT_FALLBACK;
3890}
3891
3892static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3893{
3894#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3895	/* No support for anonymous transparent PUD pages yet */
3896	if (vma_is_anonymous(vmf->vma))
3897		return VM_FAULT_FALLBACK;
3898	if (vmf->vma->vm_ops->huge_fault)
3899		return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3900#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3901	return VM_FAULT_FALLBACK;
3902}
3903
3904/*
3905 * These routines also need to handle stuff like marking pages dirty
3906 * and/or accessed for architectures that don't do it in hardware (most
3907 * RISC architectures).  The early dirtying is also good on the i386.
3908 *
3909 * There is also a hook called "update_mmu_cache()" that architectures
3910 * with external mmu caches can use to update those (ie the Sparc or
3911 * PowerPC hashed page tables that act as extended TLBs).
3912 *
3913 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3914 * concurrent faults).
3915 *
3916 * The mmap_sem may have been released depending on flags and our return value.
3917 * See filemap_fault() and __lock_page_or_retry().
3918 */
3919static int handle_pte_fault(struct vm_fault *vmf)
3920{
3921	pte_t entry;
3922
3923	if (unlikely(pmd_none(*vmf->pmd))) {
3924		/*
3925		 * Leave __pte_alloc() until later: because vm_ops->fault may
3926		 * want to allocate huge page, and if we expose page table
3927		 * for an instant, it will be difficult to retract from
3928		 * concurrent faults and from rmap lookups.
3929		 */
3930		vmf->pte = NULL;
3931	} else {
3932		/* See comment in pte_alloc_one_map() */
3933		if (pmd_devmap_trans_unstable(vmf->pmd))
3934			return 0;
3935		/*
3936		 * A regular pmd is established and it can't morph into a huge
3937		 * pmd from under us anymore at this point because we hold the
3938		 * mmap_sem read mode and khugepaged takes it in write mode.
3939		 * So now it's safe to run pte_offset_map().
3940		 */
3941		vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3942		vmf->orig_pte = *vmf->pte;
3943
3944		/*
3945		 * some architectures can have larger ptes than wordsize,
3946		 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3947		 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3948		 * accesses.  The code below just needs a consistent view
3949		 * for the ifs and we later double check anyway with the
3950		 * ptl lock held. So here a barrier will do.
3951		 */
3952		barrier();
3953		if (pte_none(vmf->orig_pte)) {
3954			pte_unmap(vmf->pte);
3955			vmf->pte = NULL;
3956		}
3957	}
3958
3959	if (!vmf->pte) {
3960		if (vma_is_anonymous(vmf->vma))
3961			return do_anonymous_page(vmf);
3962		else
3963			return do_fault(vmf);
3964	}
3965
3966	if (!pte_present(vmf->orig_pte))
3967		return do_swap_page(vmf);
3968
3969	if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3970		return do_numa_page(vmf);
3971
3972	vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3973	spin_lock(vmf->ptl);
3974	entry = vmf->orig_pte;
3975	if (unlikely(!pte_same(*vmf->pte, entry)))
 
3976		goto unlock;
 
3977	if (vmf->flags & FAULT_FLAG_WRITE) {
3978		if (!pte_write(entry))
3979			return do_wp_page(vmf);
3980		entry = pte_mkdirty(entry);
3981	}
3982	entry = pte_mkyoung(entry);
3983	if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3984				vmf->flags & FAULT_FLAG_WRITE)) {
3985		update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3986	} else {
 
 
 
3987		/*
3988		 * This is needed only for protection faults but the arch code
3989		 * is not yet telling us if this is a protection fault or not.
3990		 * This still avoids useless tlb flushes for .text page faults
3991		 * with threads.
3992		 */
3993		if (vmf->flags & FAULT_FLAG_WRITE)
3994			flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3995	}
3996unlock:
3997	pte_unmap_unlock(vmf->pte, vmf->ptl);
3998	return 0;
3999}
4000
4001/*
4002 * By the time we get here, we already hold the mm semaphore
4003 *
4004 * The mmap_sem may have been released depending on flags and our
4005 * return value.  See filemap_fault() and __lock_page_or_retry().
4006 */
4007static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4008		unsigned int flags)
4009{
4010	struct vm_fault vmf = {
4011		.vma = vma,
4012		.address = address & PAGE_MASK,
4013		.flags = flags,
4014		.pgoff = linear_page_index(vma, address),
4015		.gfp_mask = __get_fault_gfp_mask(vma),
4016	};
4017	unsigned int dirty = flags & FAULT_FLAG_WRITE;
4018	struct mm_struct *mm = vma->vm_mm;
4019	pgd_t *pgd;
4020	p4d_t *p4d;
4021	int ret;
4022
4023	pgd = pgd_offset(mm, address);
4024	p4d = p4d_alloc(mm, pgd, address);
4025	if (!p4d)
4026		return VM_FAULT_OOM;
4027
4028	vmf.pud = pud_alloc(mm, p4d, address);
4029	if (!vmf.pud)
4030		return VM_FAULT_OOM;
4031	if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
 
4032		ret = create_huge_pud(&vmf);
4033		if (!(ret & VM_FAULT_FALLBACK))
4034			return ret;
4035	} else {
4036		pud_t orig_pud = *vmf.pud;
4037
4038		barrier();
4039		if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4040
4041			/* NUMA case for anonymous PUDs would go here */
4042
4043			if (dirty && !pud_write(orig_pud)) {
4044				ret = wp_huge_pud(&vmf, orig_pud);
4045				if (!(ret & VM_FAULT_FALLBACK))
4046					return ret;
4047			} else {
4048				huge_pud_set_accessed(&vmf, orig_pud);
4049				return 0;
4050			}
4051		}
4052	}
4053
4054	vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4055	if (!vmf.pmd)
4056		return VM_FAULT_OOM;
4057	if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
 
 
 
 
 
4058		ret = create_huge_pmd(&vmf);
4059		if (!(ret & VM_FAULT_FALLBACK))
4060			return ret;
4061	} else {
4062		pmd_t orig_pmd = *vmf.pmd;
4063
4064		barrier();
4065		if (unlikely(is_swap_pmd(orig_pmd))) {
4066			VM_BUG_ON(thp_migration_supported() &&
4067					  !is_pmd_migration_entry(orig_pmd));
4068			if (is_pmd_migration_entry(orig_pmd))
4069				pmd_migration_entry_wait(mm, vmf.pmd);
4070			return 0;
4071		}
4072		if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4073			if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4074				return do_huge_pmd_numa_page(&vmf, orig_pmd);
4075
4076			if (dirty && !pmd_write(orig_pmd)) {
4077				ret = wp_huge_pmd(&vmf, orig_pmd);
4078				if (!(ret & VM_FAULT_FALLBACK))
4079					return ret;
4080			} else {
4081				huge_pmd_set_accessed(&vmf, orig_pmd);
4082				return 0;
4083			}
4084		}
4085	}
4086
4087	return handle_pte_fault(&vmf);
4088}
4089
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4090/*
4091 * By the time we get here, we already hold the mm semaphore
4092 *
4093 * The mmap_sem may have been released depending on flags and our
4094 * return value.  See filemap_fault() and __lock_page_or_retry().
4095 */
4096int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4097		unsigned int flags)
4098{
4099	int ret;
4100
4101	__set_current_state(TASK_RUNNING);
4102
4103	count_vm_event(PGFAULT);
4104	count_memcg_event_mm(vma->vm_mm, PGFAULT);
4105
4106	/* do counter updates before entering really critical section. */
4107	check_sync_rss_stat(current);
4108
4109	if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4110					    flags & FAULT_FLAG_INSTRUCTION,
4111					    flags & FAULT_FLAG_REMOTE))
4112		return VM_FAULT_SIGSEGV;
4113
4114	/*
4115	 * Enable the memcg OOM handling for faults triggered in user
4116	 * space.  Kernel faults are handled more gracefully.
4117	 */
4118	if (flags & FAULT_FLAG_USER)
4119		mem_cgroup_oom_enable();
4120
4121	if (unlikely(is_vm_hugetlb_page(vma)))
4122		ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4123	else
4124		ret = __handle_mm_fault(vma, address, flags);
4125
4126	if (flags & FAULT_FLAG_USER) {
4127		mem_cgroup_oom_disable();
4128		/*
4129		 * The task may have entered a memcg OOM situation but
4130		 * if the allocation error was handled gracefully (no
4131		 * VM_FAULT_OOM), there is no need to kill anything.
4132		 * Just clean up the OOM state peacefully.
4133		 */
4134		if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4135			mem_cgroup_oom_synchronize(false);
4136	}
4137
 
 
4138	return ret;
4139}
4140EXPORT_SYMBOL_GPL(handle_mm_fault);
4141
4142#ifndef __PAGETABLE_P4D_FOLDED
4143/*
4144 * Allocate p4d page table.
4145 * We've already handled the fast-path in-line.
4146 */
4147int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4148{
4149	p4d_t *new = p4d_alloc_one(mm, address);
4150	if (!new)
4151		return -ENOMEM;
4152
4153	smp_wmb(); /* See comment in __pte_alloc */
4154
4155	spin_lock(&mm->page_table_lock);
4156	if (pgd_present(*pgd))		/* Another has populated it */
4157		p4d_free(mm, new);
4158	else
4159		pgd_populate(mm, pgd, new);
4160	spin_unlock(&mm->page_table_lock);
4161	return 0;
4162}
4163#endif /* __PAGETABLE_P4D_FOLDED */
4164
4165#ifndef __PAGETABLE_PUD_FOLDED
4166/*
4167 * Allocate page upper directory.
4168 * We've already handled the fast-path in-line.
4169 */
4170int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4171{
4172	pud_t *new = pud_alloc_one(mm, address);
4173	if (!new)
4174		return -ENOMEM;
4175
4176	smp_wmb(); /* See comment in __pte_alloc */
4177
4178	spin_lock(&mm->page_table_lock);
4179#ifndef __ARCH_HAS_5LEVEL_HACK
4180	if (!p4d_present(*p4d)) {
4181		mm_inc_nr_puds(mm);
4182		p4d_populate(mm, p4d, new);
4183	} else	/* Another has populated it */
4184		pud_free(mm, new);
4185#else
4186	if (!pgd_present(*p4d)) {
4187		mm_inc_nr_puds(mm);
4188		pgd_populate(mm, p4d, new);
4189	} else	/* Another has populated it */
4190		pud_free(mm, new);
4191#endif /* __ARCH_HAS_5LEVEL_HACK */
4192	spin_unlock(&mm->page_table_lock);
4193	return 0;
4194}
4195#endif /* __PAGETABLE_PUD_FOLDED */
4196
4197#ifndef __PAGETABLE_PMD_FOLDED
4198/*
4199 * Allocate page middle directory.
4200 * We've already handled the fast-path in-line.
4201 */
4202int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4203{
4204	spinlock_t *ptl;
4205	pmd_t *new = pmd_alloc_one(mm, address);
4206	if (!new)
4207		return -ENOMEM;
4208
4209	smp_wmb(); /* See comment in __pte_alloc */
4210
4211	ptl = pud_lock(mm, pud);
4212#ifndef __ARCH_HAS_4LEVEL_HACK
4213	if (!pud_present(*pud)) {
4214		mm_inc_nr_pmds(mm);
4215		pud_populate(mm, pud, new);
4216	} else	/* Another has populated it */
4217		pmd_free(mm, new);
4218#else
4219	if (!pgd_present(*pud)) {
4220		mm_inc_nr_pmds(mm);
4221		pgd_populate(mm, pud, new);
4222	} else /* Another has populated it */
4223		pmd_free(mm, new);
4224#endif /* __ARCH_HAS_4LEVEL_HACK */
4225	spin_unlock(ptl);
4226	return 0;
4227}
4228#endif /* __PAGETABLE_PMD_FOLDED */
4229
4230static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4231			    unsigned long *start, unsigned long *end,
4232			    pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4233{
4234	pgd_t *pgd;
4235	p4d_t *p4d;
4236	pud_t *pud;
4237	pmd_t *pmd;
4238	pte_t *ptep;
4239
4240	pgd = pgd_offset(mm, address);
4241	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4242		goto out;
4243
4244	p4d = p4d_offset(pgd, address);
4245	if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4246		goto out;
4247
4248	pud = pud_offset(p4d, address);
4249	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4250		goto out;
4251
4252	pmd = pmd_offset(pud, address);
4253	VM_BUG_ON(pmd_trans_huge(*pmd));
4254
4255	if (pmd_huge(*pmd)) {
4256		if (!pmdpp)
4257			goto out;
4258
4259		if (start && end) {
4260			*start = address & PMD_MASK;
4261			*end = *start + PMD_SIZE;
4262			mmu_notifier_invalidate_range_start(mm, *start, *end);
 
4263		}
4264		*ptlp = pmd_lock(mm, pmd);
4265		if (pmd_huge(*pmd)) {
4266			*pmdpp = pmd;
4267			return 0;
4268		}
4269		spin_unlock(*ptlp);
4270		if (start && end)
4271			mmu_notifier_invalidate_range_end(mm, *start, *end);
4272	}
4273
4274	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4275		goto out;
4276
4277	if (start && end) {
4278		*start = address & PAGE_MASK;
4279		*end = *start + PAGE_SIZE;
4280		mmu_notifier_invalidate_range_start(mm, *start, *end);
 
4281	}
4282	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4283	if (!pte_present(*ptep))
4284		goto unlock;
4285	*ptepp = ptep;
4286	return 0;
4287unlock:
4288	pte_unmap_unlock(ptep, *ptlp);
4289	if (start && end)
4290		mmu_notifier_invalidate_range_end(mm, *start, *end);
4291out:
4292	return -EINVAL;
4293}
4294
4295static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4296			     pte_t **ptepp, spinlock_t **ptlp)
4297{
4298	int res;
4299
4300	/* (void) is needed to make gcc happy */
4301	(void) __cond_lock(*ptlp,
4302			   !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4303						    ptepp, NULL, ptlp)));
4304	return res;
4305}
4306
4307int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4308			     unsigned long *start, unsigned long *end,
4309			     pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4310{
4311	int res;
4312
4313	/* (void) is needed to make gcc happy */
4314	(void) __cond_lock(*ptlp,
4315			   !(res = __follow_pte_pmd(mm, address, start, end,
4316						    ptepp, pmdpp, ptlp)));
4317	return res;
4318}
4319EXPORT_SYMBOL(follow_pte_pmd);
4320
4321/**
4322 * follow_pfn - look up PFN at a user virtual address
4323 * @vma: memory mapping
4324 * @address: user virtual address
4325 * @pfn: location to store found PFN
4326 *
4327 * Only IO mappings and raw PFN mappings are allowed.
4328 *
4329 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4330 */
4331int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4332	unsigned long *pfn)
4333{
4334	int ret = -EINVAL;
4335	spinlock_t *ptl;
4336	pte_t *ptep;
4337
4338	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4339		return ret;
4340
4341	ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4342	if (ret)
4343		return ret;
4344	*pfn = pte_pfn(*ptep);
4345	pte_unmap_unlock(ptep, ptl);
4346	return 0;
4347}
4348EXPORT_SYMBOL(follow_pfn);
4349
4350#ifdef CONFIG_HAVE_IOREMAP_PROT
4351int follow_phys(struct vm_area_struct *vma,
4352		unsigned long address, unsigned int flags,
4353		unsigned long *prot, resource_size_t *phys)
4354{
4355	int ret = -EINVAL;
4356	pte_t *ptep, pte;
4357	spinlock_t *ptl;
4358
4359	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4360		goto out;
4361
4362	if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4363		goto out;
4364	pte = *ptep;
4365
4366	if ((flags & FOLL_WRITE) && !pte_write(pte))
4367		goto unlock;
4368
4369	*prot = pgprot_val(pte_pgprot(pte));
4370	*phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4371
4372	ret = 0;
4373unlock:
4374	pte_unmap_unlock(ptep, ptl);
4375out:
4376	return ret;
4377}
4378
4379int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4380			void *buf, int len, int write)
4381{
4382	resource_size_t phys_addr;
4383	unsigned long prot = 0;
4384	void __iomem *maddr;
4385	int offset = addr & (PAGE_SIZE-1);
4386
4387	if (follow_phys(vma, addr, write, &prot, &phys_addr))
4388		return -EINVAL;
4389
4390	maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
 
 
 
4391	if (write)
4392		memcpy_toio(maddr + offset, buf, len);
4393	else
4394		memcpy_fromio(buf, maddr + offset, len);
4395	iounmap(maddr);
4396
4397	return len;
4398}
4399EXPORT_SYMBOL_GPL(generic_access_phys);
4400#endif
4401
4402/*
4403 * Access another process' address space as given in mm.  If non-NULL, use the
4404 * given task for page fault accounting.
4405 */
4406int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4407		unsigned long addr, void *buf, int len, unsigned int gup_flags)
4408{
4409	struct vm_area_struct *vma;
4410	void *old_buf = buf;
4411	int write = gup_flags & FOLL_WRITE;
4412
4413	down_read(&mm->mmap_sem);
 
 
4414	/* ignore errors, just check how much was successfully transferred */
4415	while (len) {
4416		int bytes, ret, offset;
4417		void *maddr;
4418		struct page *page = NULL;
4419
4420		ret = get_user_pages_remote(tsk, mm, addr, 1,
4421				gup_flags, &page, &vma, NULL);
4422		if (ret <= 0) {
4423#ifndef CONFIG_HAVE_IOREMAP_PROT
4424			break;
4425#else
4426			/*
4427			 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4428			 * we can access using slightly different code.
4429			 */
4430			vma = find_vma(mm, addr);
4431			if (!vma || vma->vm_start > addr)
4432				break;
4433			if (vma->vm_ops && vma->vm_ops->access)
4434				ret = vma->vm_ops->access(vma, addr, buf,
4435							  len, write);
4436			if (ret <= 0)
4437				break;
4438			bytes = ret;
4439#endif
4440		} else {
4441			bytes = len;
4442			offset = addr & (PAGE_SIZE-1);
4443			if (bytes > PAGE_SIZE-offset)
4444				bytes = PAGE_SIZE-offset;
4445
4446			maddr = kmap(page);
4447			if (write) {
4448				copy_to_user_page(vma, page, addr,
4449						  maddr + offset, buf, bytes);
4450				set_page_dirty_lock(page);
4451			} else {
4452				copy_from_user_page(vma, page, addr,
4453						    buf, maddr + offset, bytes);
4454			}
4455			kunmap(page);
4456			put_page(page);
4457		}
4458		len -= bytes;
4459		buf += bytes;
4460		addr += bytes;
4461	}
4462	up_read(&mm->mmap_sem);
4463
4464	return buf - old_buf;
4465}
4466
4467/**
4468 * access_remote_vm - access another process' address space
4469 * @mm:		the mm_struct of the target address space
4470 * @addr:	start address to access
4471 * @buf:	source or destination buffer
4472 * @len:	number of bytes to transfer
4473 * @gup_flags:	flags modifying lookup behaviour
4474 *
4475 * The caller must hold a reference on @mm.
 
 
4476 */
4477int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4478		void *buf, int len, unsigned int gup_flags)
4479{
4480	return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4481}
4482
4483/*
4484 * Access another process' address space.
4485 * Source/target buffer must be kernel space,
4486 * Do not walk the page table directly, use get_user_pages
4487 */
4488int access_process_vm(struct task_struct *tsk, unsigned long addr,
4489		void *buf, int len, unsigned int gup_flags)
4490{
4491	struct mm_struct *mm;
4492	int ret;
4493
4494	mm = get_task_mm(tsk);
4495	if (!mm)
4496		return 0;
4497
4498	ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4499
4500	mmput(mm);
4501
4502	return ret;
4503}
4504EXPORT_SYMBOL_GPL(access_process_vm);
4505
4506/*
4507 * Print the name of a VMA.
4508 */
4509void print_vma_addr(char *prefix, unsigned long ip)
4510{
4511	struct mm_struct *mm = current->mm;
4512	struct vm_area_struct *vma;
4513
4514	/*
4515	 * we might be running from an atomic context so we cannot sleep
4516	 */
4517	if (!down_read_trylock(&mm->mmap_sem))
4518		return;
4519
4520	vma = find_vma(mm, ip);
4521	if (vma && vma->vm_file) {
4522		struct file *f = vma->vm_file;
4523		char *buf = (char *)__get_free_page(GFP_NOWAIT);
4524		if (buf) {
4525			char *p;
4526
4527			p = file_path(f, buf, PAGE_SIZE);
4528			if (IS_ERR(p))
4529				p = "?";
4530			printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4531					vma->vm_start,
4532					vma->vm_end - vma->vm_start);
4533			free_page((unsigned long)buf);
4534		}
4535	}
4536	up_read(&mm->mmap_sem);
4537}
4538
4539#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4540void __might_fault(const char *file, int line)
4541{
4542	/*
4543	 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4544	 * holding the mmap_sem, this is safe because kernel memory doesn't
4545	 * get paged out, therefore we'll never actually fault, and the
4546	 * below annotations will generate false positives.
4547	 */
4548	if (uaccess_kernel())
4549		return;
4550	if (pagefault_disabled())
4551		return;
4552	__might_sleep(file, line, 0);
4553#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4554	if (current->mm)
4555		might_lock_read(&current->mm->mmap_sem);
4556#endif
4557}
4558EXPORT_SYMBOL(__might_fault);
4559#endif
4560
4561#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4562static void clear_gigantic_page(struct page *page,
4563				unsigned long addr,
4564				unsigned int pages_per_huge_page)
4565{
4566	int i;
4567	struct page *p = page;
4568
4569	might_sleep();
4570	for (i = 0; i < pages_per_huge_page;
4571	     i++, p = mem_map_next(p, page, i)) {
4572		cond_resched();
4573		clear_user_highpage(p, addr + i * PAGE_SIZE);
4574	}
4575}
4576void clear_huge_page(struct page *page,
4577		     unsigned long addr_hint, unsigned int pages_per_huge_page)
4578{
4579	int i, n, base, l;
4580	unsigned long addr = addr_hint &
4581		~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4582
4583	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4584		clear_gigantic_page(page, addr, pages_per_huge_page);
4585		return;
4586	}
4587
4588	/* Clear sub-page to access last to keep its cache lines hot */
4589	might_sleep();
4590	n = (addr_hint - addr) / PAGE_SIZE;
4591	if (2 * n <= pages_per_huge_page) {
4592		/* If sub-page to access in first half of huge page */
4593		base = 0;
4594		l = n;
4595		/* Clear sub-pages at the end of huge page */
4596		for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4597			cond_resched();
4598			clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4599		}
4600	} else {
4601		/* If sub-page to access in second half of huge page */
4602		base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4603		l = pages_per_huge_page - n;
4604		/* Clear sub-pages at the begin of huge page */
4605		for (i = 0; i < base; i++) {
4606			cond_resched();
4607			clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4608		}
4609	}
4610	/*
4611	 * Clear remaining sub-pages in left-right-left-right pattern
4612	 * towards the sub-page to access
4613	 */
4614	for (i = 0; i < l; i++) {
4615		int left_idx = base + i;
4616		int right_idx = base + 2 * l - 1 - i;
4617
4618		cond_resched();
4619		clear_user_highpage(page + left_idx,
4620				    addr + left_idx * PAGE_SIZE);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4621		cond_resched();
4622		clear_user_highpage(page + right_idx,
4623				    addr + right_idx * PAGE_SIZE);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4624	}
 
 
4625}
4626
4627static void copy_user_gigantic_page(struct page *dst, struct page *src,
4628				    unsigned long addr,
4629				    struct vm_area_struct *vma,
4630				    unsigned int pages_per_huge_page)
4631{
4632	int i;
4633	struct page *dst_base = dst;
4634	struct page *src_base = src;
4635
4636	for (i = 0; i < pages_per_huge_page; ) {
4637		cond_resched();
4638		copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4639
4640		i++;
4641		dst = mem_map_next(dst, dst_base, i);
4642		src = mem_map_next(src, src_base, i);
4643	}
4644}
4645
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4646void copy_user_huge_page(struct page *dst, struct page *src,
4647			 unsigned long addr, struct vm_area_struct *vma,
4648			 unsigned int pages_per_huge_page)
4649{
4650	int i;
 
 
 
 
 
 
4651
4652	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4653		copy_user_gigantic_page(dst, src, addr, vma,
4654					pages_per_huge_page);
4655		return;
4656	}
4657
4658	might_sleep();
4659	for (i = 0; i < pages_per_huge_page; i++) {
4660		cond_resched();
4661		copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4662	}
4663}
4664
4665long copy_huge_page_from_user(struct page *dst_page,
4666				const void __user *usr_src,
4667				unsigned int pages_per_huge_page,
4668				bool allow_pagefault)
4669{
4670	void *src = (void *)usr_src;
4671	void *page_kaddr;
4672	unsigned long i, rc = 0;
4673	unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4674
4675	for (i = 0; i < pages_per_huge_page; i++) {
4676		if (allow_pagefault)
4677			page_kaddr = kmap(dst_page + i);
4678		else
4679			page_kaddr = kmap_atomic(dst_page + i);
4680		rc = copy_from_user(page_kaddr,
4681				(const void __user *)(src + i * PAGE_SIZE),
4682				PAGE_SIZE);
4683		if (allow_pagefault)
4684			kunmap(dst_page + i);
4685		else
4686			kunmap_atomic(page_kaddr);
4687
4688		ret_val -= (PAGE_SIZE - rc);
4689		if (rc)
4690			break;
4691
4692		cond_resched();
4693	}
4694	return ret_val;
4695}
4696#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4697
4698#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4699
4700static struct kmem_cache *page_ptl_cachep;
4701
4702void __init ptlock_cache_init(void)
4703{
4704	page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4705			SLAB_PANIC, NULL);
4706}
4707
4708bool ptlock_alloc(struct page *page)
4709{
4710	spinlock_t *ptl;
4711
4712	ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4713	if (!ptl)
4714		return false;
4715	page->ptl = ptl;
4716	return true;
4717}
4718
4719void ptlock_free(struct page *page)
4720{
4721	kmem_cache_free(page_ptl_cachep, page->ptl);
4722}
4723#endif