Linux Audio

Check our new training course

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