1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Memory merging support.
   4 *
   5 * This code enables dynamic sharing of identical pages found in different
   6 * memory areas, even if they are not shared by fork()
   7 *
   8 * Copyright (C) 2008-2009 Red Hat, Inc.
   9 * Authors:
  10 *	Izik Eidus
  11 *	Andrea Arcangeli
  12 *	Chris Wright
  13 *	Hugh Dickins
 
 
  14 */
  15
  16#include <linux/errno.h>
  17#include <linux/mm.h>
  18#include <linux/fs.h>
  19#include <linux/mman.h>
  20#include <linux/sched.h>
  21#include <linux/sched/mm.h>
  22#include <linux/sched/coredump.h>
  23#include <linux/rwsem.h>
  24#include <linux/pagemap.h>
  25#include <linux/rmap.h>
  26#include <linux/spinlock.h>
  27#include <linux/xxhash.h>
  28#include <linux/delay.h>
  29#include <linux/kthread.h>
  30#include <linux/wait.h>
  31#include <linux/slab.h>
  32#include <linux/rbtree.h>
  33#include <linux/memory.h>
  34#include <linux/mmu_notifier.h>
  35#include <linux/swap.h>
  36#include <linux/ksm.h>
  37#include <linux/hashtable.h>
  38#include <linux/freezer.h>
  39#include <linux/oom.h>
  40#include <linux/numa.h>
  41
  42#include <asm/tlbflush.h>
  43#include "internal.h"
  44
  45#ifdef CONFIG_NUMA
  46#define NUMA(x)		(x)
  47#define DO_NUMA(x)	do { (x); } while (0)
  48#else
  49#define NUMA(x)		(0)
  50#define DO_NUMA(x)	do { } while (0)
  51#endif
  52
  53/**
  54 * DOC: Overview
  55 *
  56 * A few notes about the KSM scanning process,
  57 * to make it easier to understand the data structures below:
  58 *
  59 * In order to reduce excessive scanning, KSM sorts the memory pages by their
  60 * contents into a data structure that holds pointers to the pages' locations.
  61 *
  62 * Since the contents of the pages may change at any moment, KSM cannot just
  63 * insert the pages into a normal sorted tree and expect it to find anything.
  64 * Therefore KSM uses two data structures - the stable and the unstable tree.
  65 *
  66 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
  67 * by their contents.  Because each such page is write-protected, searching on
  68 * this tree is fully assured to be working (except when pages are unmapped),
  69 * and therefore this tree is called the stable tree.
  70 *
  71 * The stable tree node includes information required for reverse
  72 * mapping from a KSM page to virtual addresses that map this page.
  73 *
  74 * In order to avoid large latencies of the rmap walks on KSM pages,
  75 * KSM maintains two types of nodes in the stable tree:
  76 *
  77 * * the regular nodes that keep the reverse mapping structures in a
  78 *   linked list
  79 * * the "chains" that link nodes ("dups") that represent the same
  80 *   write protected memory content, but each "dup" corresponds to a
  81 *   different KSM page copy of that content
  82 *
  83 * Internally, the regular nodes, "dups" and "chains" are represented
  84 * using the same :c:type:`struct stable_node` structure.
  85 *
  86 * In addition to the stable tree, KSM uses a second data structure called the
  87 * unstable tree: this tree holds pointers to pages which have been found to
  88 * be "unchanged for a period of time".  The unstable tree sorts these pages
  89 * by their contents, but since they are not write-protected, KSM cannot rely
  90 * upon the unstable tree to work correctly - the unstable tree is liable to
  91 * be corrupted as its contents are modified, and so it is called unstable.
  92 *
  93 * KSM solves this problem by several techniques:
  94 *
  95 * 1) The unstable tree is flushed every time KSM completes scanning all
  96 *    memory areas, and then the tree is rebuilt again from the beginning.
  97 * 2) KSM will only insert into the unstable tree, pages whose hash value
  98 *    has not changed since the previous scan of all memory areas.
  99 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
 100 *    colors of the nodes and not on their contents, assuring that even when
 101 *    the tree gets "corrupted" it won't get out of balance, so scanning time
 102 *    remains the same (also, searching and inserting nodes in an rbtree uses
 103 *    the same algorithm, so we have no overhead when we flush and rebuild).
 104 * 4) KSM never flushes the stable tree, which means that even if it were to
 105 *    take 10 attempts to find a page in the unstable tree, once it is found,
 106 *    it is secured in the stable tree.  (When we scan a new page, we first
 107 *    compare it against the stable tree, and then against the unstable tree.)
 108 *
 109 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
 110 * stable trees and multiple unstable trees: one of each for each NUMA node.
 111 */
 112
 113/**
 114 * struct mm_slot - ksm information per mm that is being scanned
 115 * @link: link to the mm_slots hash list
 116 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
 117 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
 118 * @mm: the mm that this information is valid for
 119 */
 120struct mm_slot {
 121	struct hlist_node link;
 122	struct list_head mm_list;
 123	struct rmap_item *rmap_list;
 124	struct mm_struct *mm;
 125};
 126
 127/**
 128 * struct ksm_scan - cursor for scanning
 129 * @mm_slot: the current mm_slot we are scanning
 130 * @address: the next address inside that to be scanned
 131 * @rmap_list: link to the next rmap to be scanned in the rmap_list
 132 * @seqnr: count of completed full scans (needed when removing unstable node)
 133 *
 134 * There is only the one ksm_scan instance of this cursor structure.
 135 */
 136struct ksm_scan {
 137	struct mm_slot *mm_slot;
 138	unsigned long address;
 139	struct rmap_item **rmap_list;
 140	unsigned long seqnr;
 141};
 142
 143/**
 144 * struct stable_node - node of the stable rbtree
 145 * @node: rb node of this ksm page in the stable tree
 146 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
 147 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
 148 * @list: linked into migrate_nodes, pending placement in the proper node tree
 149 * @hlist: hlist head of rmap_items using this ksm page
 150 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
 151 * @chain_prune_time: time of the last full garbage collection
 152 * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
 153 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
 154 */
 155struct stable_node {
 156	union {
 157		struct rb_node node;	/* when node of stable tree */
 158		struct {		/* when listed for migration */
 159			struct list_head *head;
 160			struct {
 161				struct hlist_node hlist_dup;
 162				struct list_head list;
 163			};
 164		};
 165	};
 166	struct hlist_head hlist;
 167	union {
 168		unsigned long kpfn;
 169		unsigned long chain_prune_time;
 170	};
 171	/*
 172	 * STABLE_NODE_CHAIN can be any negative number in
 173	 * rmap_hlist_len negative range, but better not -1 to be able
 174	 * to reliably detect underflows.
 175	 */
 176#define STABLE_NODE_CHAIN -1024
 177	int rmap_hlist_len;
 178#ifdef CONFIG_NUMA
 179	int nid;
 180#endif
 181};
 182
 183/**
 184 * struct rmap_item - reverse mapping item for virtual addresses
 185 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
 186 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
 187 * @nid: NUMA node id of unstable tree in which linked (may not match page)
 188 * @mm: the memory structure this rmap_item is pointing into
 189 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
 190 * @oldchecksum: previous checksum of the page at that virtual address
 191 * @node: rb node of this rmap_item in the unstable tree
 192 * @head: pointer to stable_node heading this list in the stable tree
 193 * @hlist: link into hlist of rmap_items hanging off that stable_node
 194 */
 195struct rmap_item {
 196	struct rmap_item *rmap_list;
 197	union {
 198		struct anon_vma *anon_vma;	/* when stable */
 199#ifdef CONFIG_NUMA
 200		int nid;		/* when node of unstable tree */
 201#endif
 202	};
 203	struct mm_struct *mm;
 204	unsigned long address;		/* + low bits used for flags below */
 205	unsigned int oldchecksum;	/* when unstable */
 206	union {
 207		struct rb_node node;	/* when node of unstable tree */
 208		struct {		/* when listed from stable tree */
 209			struct stable_node *head;
 210			struct hlist_node hlist;
 211		};
 212	};
 213};
 214
 215#define SEQNR_MASK	0x0ff	/* low bits of unstable tree seqnr */
 216#define UNSTABLE_FLAG	0x100	/* is a node of the unstable tree */
 217#define STABLE_FLAG	0x200	/* is listed from the stable tree */
 218#define KSM_FLAG_MASK	(SEQNR_MASK|UNSTABLE_FLAG|STABLE_FLAG)
 219				/* to mask all the flags */
 220
 221/* The stable and unstable tree heads */
 222static struct rb_root one_stable_tree[1] = { RB_ROOT };
 223static struct rb_root one_unstable_tree[1] = { RB_ROOT };
 224static struct rb_root *root_stable_tree = one_stable_tree;
 225static struct rb_root *root_unstable_tree = one_unstable_tree;
 226
 227/* Recently migrated nodes of stable tree, pending proper placement */
 228static LIST_HEAD(migrate_nodes);
 229#define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
 230
 231#define MM_SLOTS_HASH_BITS 10
 232static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
 
 233
 234static struct mm_slot ksm_mm_head = {
 235	.mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
 236};
 237static struct ksm_scan ksm_scan = {
 238	.mm_slot = &ksm_mm_head,
 239};
 240
 241static struct kmem_cache *rmap_item_cache;
 242static struct kmem_cache *stable_node_cache;
 243static struct kmem_cache *mm_slot_cache;
 244
 245/* The number of nodes in the stable tree */
 246static unsigned long ksm_pages_shared;
 247
 248/* The number of page slots additionally sharing those nodes */
 249static unsigned long ksm_pages_sharing;
 250
 251/* The number of nodes in the unstable tree */
 252static unsigned long ksm_pages_unshared;
 253
 254/* The number of rmap_items in use: to calculate pages_volatile */
 255static unsigned long ksm_rmap_items;
 256
 257/* The number of stable_node chains */
 258static unsigned long ksm_stable_node_chains;
 259
 260/* The number of stable_node dups linked to the stable_node chains */
 261static unsigned long ksm_stable_node_dups;
 262
 263/* Delay in pruning stale stable_node_dups in the stable_node_chains */
 264static int ksm_stable_node_chains_prune_millisecs = 2000;
 265
 266/* Maximum number of page slots sharing a stable node */
 267static int ksm_max_page_sharing = 256;
 268
 269/* Number of pages ksmd should scan in one batch */
 270static unsigned int ksm_thread_pages_to_scan = 100;
 271
 272/* Milliseconds ksmd should sleep between batches */
 273static unsigned int ksm_thread_sleep_millisecs = 20;
 274
 275/* Checksum of an empty (zeroed) page */
 276static unsigned int zero_checksum __read_mostly;
 277
 278/* Whether to merge empty (zeroed) pages with actual zero pages */
 279static bool ksm_use_zero_pages __read_mostly;
 280
 281#ifdef CONFIG_NUMA
 282/* Zeroed when merging across nodes is not allowed */
 283static unsigned int ksm_merge_across_nodes = 1;
 284static int ksm_nr_node_ids = 1;
 285#else
 286#define ksm_merge_across_nodes	1U
 287#define ksm_nr_node_ids		1
 288#endif
 289
 290#define KSM_RUN_STOP	0
 291#define KSM_RUN_MERGE	1
 292#define KSM_RUN_UNMERGE	2
 293#define KSM_RUN_OFFLINE	4
 294static unsigned long ksm_run = KSM_RUN_STOP;
 295static void wait_while_offlining(void);
 296
 297static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
 298static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait);
 299static DEFINE_MUTEX(ksm_thread_mutex);
 300static DEFINE_SPINLOCK(ksm_mmlist_lock);
 301
 302#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
 303		sizeof(struct __struct), __alignof__(struct __struct),\
 304		(__flags), NULL)
 305
 306static int __init ksm_slab_init(void)
 307{
 308	rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
 309	if (!rmap_item_cache)
 310		goto out;
 311
 312	stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
 313	if (!stable_node_cache)
 314		goto out_free1;
 315
 316	mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
 317	if (!mm_slot_cache)
 318		goto out_free2;
 319
 320	return 0;
 321
 322out_free2:
 323	kmem_cache_destroy(stable_node_cache);
 324out_free1:
 325	kmem_cache_destroy(rmap_item_cache);
 326out:
 327	return -ENOMEM;
 328}
 329
 330static void __init ksm_slab_free(void)
 331{
 332	kmem_cache_destroy(mm_slot_cache);
 333	kmem_cache_destroy(stable_node_cache);
 334	kmem_cache_destroy(rmap_item_cache);
 335	mm_slot_cache = NULL;
 336}
 337
 338static __always_inline bool is_stable_node_chain(struct stable_node *chain)
 339{
 340	return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
 341}
 342
 343static __always_inline bool is_stable_node_dup(struct stable_node *dup)
 344{
 345	return dup->head == STABLE_NODE_DUP_HEAD;
 346}
 347
 348static inline void stable_node_chain_add_dup(struct stable_node *dup,
 349					     struct stable_node *chain)
 350{
 351	VM_BUG_ON(is_stable_node_dup(dup));
 352	dup->head = STABLE_NODE_DUP_HEAD;
 353	VM_BUG_ON(!is_stable_node_chain(chain));
 354	hlist_add_head(&dup->hlist_dup, &chain->hlist);
 355	ksm_stable_node_dups++;
 356}
 357
 358static inline void __stable_node_dup_del(struct stable_node *dup)
 359{
 360	VM_BUG_ON(!is_stable_node_dup(dup));
 361	hlist_del(&dup->hlist_dup);
 362	ksm_stable_node_dups--;
 363}
 364
 365static inline void stable_node_dup_del(struct stable_node *dup)
 366{
 367	VM_BUG_ON(is_stable_node_chain(dup));
 368	if (is_stable_node_dup(dup))
 369		__stable_node_dup_del(dup);
 370	else
 371		rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
 372#ifdef CONFIG_DEBUG_VM
 373	dup->head = NULL;
 374#endif
 375}
 376
 377static inline struct rmap_item *alloc_rmap_item(void)
 378{
 379	struct rmap_item *rmap_item;
 380
 381	rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
 382						__GFP_NORETRY | __GFP_NOWARN);
 383	if (rmap_item)
 384		ksm_rmap_items++;
 385	return rmap_item;
 386}
 387
 388static inline void free_rmap_item(struct rmap_item *rmap_item)
 389{
 390	ksm_rmap_items--;
 391	rmap_item->mm = NULL;	/* debug safety */
 392	kmem_cache_free(rmap_item_cache, rmap_item);
 393}
 394
 395static inline struct stable_node *alloc_stable_node(void)
 396{
 397	/*
 398	 * The allocation can take too long with GFP_KERNEL when memory is under
 399	 * pressure, which may lead to hung task warnings.  Adding __GFP_HIGH
 400	 * grants access to memory reserves, helping to avoid this problem.
 401	 */
 402	return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
 403}
 404
 405static inline void free_stable_node(struct stable_node *stable_node)
 406{
 407	VM_BUG_ON(stable_node->rmap_hlist_len &&
 408		  !is_stable_node_chain(stable_node));
 409	kmem_cache_free(stable_node_cache, stable_node);
 410}
 411
 412static inline struct mm_slot *alloc_mm_slot(void)
 413{
 414	if (!mm_slot_cache)	/* initialization failed */
 415		return NULL;
 416	return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 417}
 418
 419static inline void free_mm_slot(struct mm_slot *mm_slot)
 420{
 421	kmem_cache_free(mm_slot_cache, mm_slot);
 422}
 423
 424static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 425{
 426	struct mm_slot *slot;
 427
 428	hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
 429		if (slot->mm == mm)
 430			return slot;
 431
 
 
 
 
 
 432	return NULL;
 433}
 434
 435static void insert_to_mm_slots_hash(struct mm_struct *mm,
 436				    struct mm_slot *mm_slot)
 437{
 
 
 
 438	mm_slot->mm = mm;
 439	hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
 
 
 
 
 
 440}
 441
 442/*
 443 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
 444 * page tables after it has passed through ksm_exit() - which, if necessary,
 445 * takes mmap_lock briefly to serialize against them.  ksm_exit() does not set
 446 * a special flag: they can just back out as soon as mm_users goes to zero.
 447 * ksm_test_exit() is used throughout to make this test for exit: in some
 448 * places for correctness, in some places just to avoid unnecessary work.
 449 */
 450static inline bool ksm_test_exit(struct mm_struct *mm)
 451{
 452	return atomic_read(&mm->mm_users) == 0;
 453}
 454
 455/*
 456 * We use break_ksm to break COW on a ksm page: it's a stripped down
 457 *
 458 *	if (get_user_pages(addr, 1, FOLL_WRITE, &page, NULL) == 1)
 459 *		put_page(page);
 460 *
 461 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
 462 * in case the application has unmapped and remapped mm,addr meanwhile.
 463 * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
 464 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
 465 *
 466 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
 467 * of the process that owns 'vma'.  We also do not want to enforce
 468 * protection keys here anyway.
 469 */
 470static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
 471{
 472	struct page *page;
 473	vm_fault_t ret = 0;
 474
 475	do {
 476		cond_resched();
 477		page = follow_page(vma, addr,
 478				FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
 479		if (IS_ERR_OR_NULL(page))
 480			break;
 481		if (PageKsm(page))
 482			ret = handle_mm_fault(vma, addr,
 483					      FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE,
 484					      NULL);
 485		else
 486			ret = VM_FAULT_WRITE;
 487		put_page(page);
 488	} while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
 489	/*
 490	 * We must loop because handle_mm_fault() may back out if there's
 491	 * any difficulty e.g. if pte accessed bit gets updated concurrently.
 492	 *
 493	 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
 494	 * COW has been broken, even if the vma does not permit VM_WRITE;
 495	 * but note that a concurrent fault might break PageKsm for us.
 496	 *
 497	 * VM_FAULT_SIGBUS could occur if we race with truncation of the
 498	 * backing file, which also invalidates anonymous pages: that's
 499	 * okay, that truncation will have unmapped the PageKsm for us.
 500	 *
 501	 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
 502	 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
 503	 * current task has TIF_MEMDIE set, and will be OOM killed on return
 504	 * to user; and ksmd, having no mm, would never be chosen for that.
 505	 *
 506	 * But if the mm is in a limited mem_cgroup, then the fault may fail
 507	 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
 508	 * even ksmd can fail in this way - though it's usually breaking ksm
 509	 * just to undo a merge it made a moment before, so unlikely to oom.
 510	 *
 511	 * That's a pity: we might therefore have more kernel pages allocated
 512	 * than we're counting as nodes in the stable tree; but ksm_do_scan
 513	 * will retry to break_cow on each pass, so should recover the page
 514	 * in due course.  The important thing is to not let VM_MERGEABLE
 515	 * be cleared while any such pages might remain in the area.
 516	 */
 517	return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
 518}
 519
 520static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
 521		unsigned long addr)
 522{
 523	struct vm_area_struct *vma;
 524	if (ksm_test_exit(mm))
 525		return NULL;
 526	vma = find_vma(mm, addr);
 527	if (!vma || vma->vm_start > addr)
 528		return NULL;
 529	if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
 530		return NULL;
 531	return vma;
 532}
 533
 534static void break_cow(struct rmap_item *rmap_item)
 535{
 536	struct mm_struct *mm = rmap_item->mm;
 537	unsigned long addr = rmap_item->address;
 538	struct vm_area_struct *vma;
 539
 540	/*
 541	 * It is not an accident that whenever we want to break COW
 542	 * to undo, we also need to drop a reference to the anon_vma.
 543	 */
 544	put_anon_vma(rmap_item->anon_vma);
 545
 546	mmap_read_lock(mm);
 547	vma = find_mergeable_vma(mm, addr);
 548	if (vma)
 549		break_ksm(vma, addr);
 550	mmap_read_unlock(mm);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 551}
 552
 553static struct page *get_mergeable_page(struct rmap_item *rmap_item)
 554{
 555	struct mm_struct *mm = rmap_item->mm;
 556	unsigned long addr = rmap_item->address;
 557	struct vm_area_struct *vma;
 558	struct page *page;
 559
 560	mmap_read_lock(mm);
 561	vma = find_mergeable_vma(mm, addr);
 562	if (!vma)
 563		goto out;
 564
 565	page = follow_page(vma, addr, FOLL_GET);
 566	if (IS_ERR_OR_NULL(page))
 567		goto out;
 568	if (PageAnon(page)) {
 569		flush_anon_page(vma, page, addr);
 570		flush_dcache_page(page);
 571	} else {
 572		put_page(page);
 573out:
 574		page = NULL;
 575	}
 576	mmap_read_unlock(mm);
 577	return page;
 578}
 579
 580/*
 581 * This helper is used for getting right index into array of tree roots.
 582 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
 583 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
 584 * every node has its own stable and unstable tree.
 585 */
 586static inline int get_kpfn_nid(unsigned long kpfn)
 587{
 588	return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
 589}
 590
 591static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
 592						   struct rb_root *root)
 593{
 594	struct stable_node *chain = alloc_stable_node();
 595	VM_BUG_ON(is_stable_node_chain(dup));
 596	if (likely(chain)) {
 597		INIT_HLIST_HEAD(&chain->hlist);
 598		chain->chain_prune_time = jiffies;
 599		chain->rmap_hlist_len = STABLE_NODE_CHAIN;
 600#if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
 601		chain->nid = NUMA_NO_NODE; /* debug */
 602#endif
 603		ksm_stable_node_chains++;
 604
 605		/*
 606		 * Put the stable node chain in the first dimension of
 607		 * the stable tree and at the same time remove the old
 608		 * stable node.
 609		 */
 610		rb_replace_node(&dup->node, &chain->node, root);
 611
 612		/*
 613		 * Move the old stable node to the second dimension
 614		 * queued in the hlist_dup. The invariant is that all
 615		 * dup stable_nodes in the chain->hlist point to pages
 616		 * that are write protected and have the exact same
 617		 * content.
 618		 */
 619		stable_node_chain_add_dup(dup, chain);
 620	}
 621	return chain;
 622}
 623
 624static inline void free_stable_node_chain(struct stable_node *chain,
 625					  struct rb_root *root)
 626{
 627	rb_erase(&chain->node, root);
 628	free_stable_node(chain);
 629	ksm_stable_node_chains--;
 630}
 631
 632static void remove_node_from_stable_tree(struct stable_node *stable_node)
 633{
 634	struct rmap_item *rmap_item;
 
 635
 636	/* check it's not STABLE_NODE_CHAIN or negative */
 637	BUG_ON(stable_node->rmap_hlist_len < 0);
 638
 639	hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
 640		if (rmap_item->hlist.next)
 641			ksm_pages_sharing--;
 642		else
 643			ksm_pages_shared--;
 644		VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
 645		stable_node->rmap_hlist_len--;
 646		put_anon_vma(rmap_item->anon_vma);
 647		rmap_item->address &= PAGE_MASK;
 648		cond_resched();
 649	}
 650
 651	/*
 652	 * We need the second aligned pointer of the migrate_nodes
 653	 * list_head to stay clear from the rb_parent_color union
 654	 * (aligned and different than any node) and also different
 655	 * from &migrate_nodes. This will verify that future list.h changes
 656	 * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it.
 657	 */
 658#if defined(GCC_VERSION) && GCC_VERSION >= 40903
 659	BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
 660	BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
 661#endif
 662
 663	if (stable_node->head == &migrate_nodes)
 664		list_del(&stable_node->list);
 665	else
 666		stable_node_dup_del(stable_node);
 667	free_stable_node(stable_node);
 668}
 669
 670enum get_ksm_page_flags {
 671	GET_KSM_PAGE_NOLOCK,
 672	GET_KSM_PAGE_LOCK,
 673	GET_KSM_PAGE_TRYLOCK
 674};
 675
 676/*
 677 * get_ksm_page: checks if the page indicated by the stable node
 678 * is still its ksm page, despite having held no reference to it.
 679 * In which case we can trust the content of the page, and it
 680 * returns the gotten page; but if the page has now been zapped,
 681 * remove the stale node from the stable tree and return NULL.
 682 * But beware, the stable node's page might be being migrated.
 683 *
 684 * You would expect the stable_node to hold a reference to the ksm page.
 685 * But if it increments the page's count, swapping out has to wait for
 686 * ksmd to come around again before it can free the page, which may take
 687 * seconds or even minutes: much too unresponsive.  So instead we use a
 688 * "keyhole reference": access to the ksm page from the stable node peeps
 689 * out through its keyhole to see if that page still holds the right key,
 690 * pointing back to this stable node.  This relies on freeing a PageAnon
 691 * page to reset its page->mapping to NULL, and relies on no other use of
 692 * a page to put something that might look like our key in page->mapping.
 
 
 
 
 
 
 
 
 
 
 
 693 * is on its way to being freed; but it is an anomaly to bear in mind.
 694 */
 695static struct page *get_ksm_page(struct stable_node *stable_node,
 696				 enum get_ksm_page_flags flags)
 697{
 698	struct page *page;
 699	void *expected_mapping;
 700	unsigned long kpfn;
 701
 702	expected_mapping = (void *)((unsigned long)stable_node |
 703					PAGE_MAPPING_KSM);
 704again:
 705	kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
 706	page = pfn_to_page(kpfn);
 707	if (READ_ONCE(page->mapping) != expected_mapping)
 708		goto stale;
 709
 710	/*
 711	 * We cannot do anything with the page while its refcount is 0.
 712	 * Usually 0 means free, or tail of a higher-order page: in which
 713	 * case this node is no longer referenced, and should be freed;
 714	 * however, it might mean that the page is under page_ref_freeze().
 715	 * The __remove_mapping() case is easy, again the node is now stale;
 716	 * the same is in reuse_ksm_page() case; but if page is swapcache
 717	 * in migrate_page_move_mapping(), it might still be our page,
 718	 * in which case it's essential to keep the node.
 719	 */
 720	while (!get_page_unless_zero(page)) {
 721		/*
 722		 * Another check for page->mapping != expected_mapping would
 723		 * work here too.  We have chosen the !PageSwapCache test to
 724		 * optimize the common case, when the page is or is about to
 725		 * be freed: PageSwapCache is cleared (under spin_lock_irq)
 726		 * in the ref_freeze section of __remove_mapping(); but Anon
 727		 * page->mapping reset to NULL later, in free_pages_prepare().
 728		 */
 729		if (!PageSwapCache(page))
 730			goto stale;
 731		cpu_relax();
 732	}
 733
 734	if (READ_ONCE(page->mapping) != expected_mapping) {
 735		put_page(page);
 736		goto stale;
 737	}
 738
 739	if (flags == GET_KSM_PAGE_TRYLOCK) {
 740		if (!trylock_page(page)) {
 741			put_page(page);
 742			return ERR_PTR(-EBUSY);
 743		}
 744	} else if (flags == GET_KSM_PAGE_LOCK)
 745		lock_page(page);
 746
 747	if (flags != GET_KSM_PAGE_NOLOCK) {
 748		if (READ_ONCE(page->mapping) != expected_mapping) {
 749			unlock_page(page);
 750			put_page(page);
 751			goto stale;
 752		}
 753	}
 754	return page;
 755
 756stale:
 757	/*
 758	 * We come here from above when page->mapping or !PageSwapCache
 759	 * suggests that the node is stale; but it might be under migration.
 760	 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
 761	 * before checking whether node->kpfn has been changed.
 762	 */
 763	smp_rmb();
 764	if (READ_ONCE(stable_node->kpfn) != kpfn)
 765		goto again;
 766	remove_node_from_stable_tree(stable_node);
 767	return NULL;
 768}
 769
 770/*
 771 * Removing rmap_item from stable or unstable tree.
 772 * This function will clean the information from the stable/unstable tree.
 773 */
 774static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
 775{
 776	if (rmap_item->address & STABLE_FLAG) {
 777		struct stable_node *stable_node;
 778		struct page *page;
 779
 780		stable_node = rmap_item->head;
 781		page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
 782		if (!page)
 783			goto out;
 784
 
 785		hlist_del(&rmap_item->hlist);
 786		unlock_page(page);
 787		put_page(page);
 788
 789		if (!hlist_empty(&stable_node->hlist))
 790			ksm_pages_sharing--;
 791		else
 792			ksm_pages_shared--;
 793		VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
 794		stable_node->rmap_hlist_len--;
 795
 796		put_anon_vma(rmap_item->anon_vma);
 797		rmap_item->address &= PAGE_MASK;
 798
 799	} else if (rmap_item->address & UNSTABLE_FLAG) {
 800		unsigned char age;
 801		/*
 802		 * Usually ksmd can and must skip the rb_erase, because
 803		 * root_unstable_tree was already reset to RB_ROOT.
 804		 * But be careful when an mm is exiting: do the rb_erase
 805		 * if this rmap_item was inserted by this scan, rather
 806		 * than left over from before.
 807		 */
 808		age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
 809		BUG_ON(age > 1);
 810		if (!age)
 811			rb_erase(&rmap_item->node,
 812				 root_unstable_tree + NUMA(rmap_item->nid));
 813		ksm_pages_unshared--;
 814		rmap_item->address &= PAGE_MASK;
 815	}
 816out:
 817	cond_resched();		/* we're called from many long loops */
 818}
 819
 820static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
 821				       struct rmap_item **rmap_list)
 822{
 823	while (*rmap_list) {
 824		struct rmap_item *rmap_item = *rmap_list;
 825		*rmap_list = rmap_item->rmap_list;
 826		remove_rmap_item_from_tree(rmap_item);
 827		free_rmap_item(rmap_item);
 828	}
 829}
 830
 831/*
 832 * Though it's very tempting to unmerge rmap_items from stable tree rather
 833 * than check every pte of a given vma, the locking doesn't quite work for
 834 * that - an rmap_item is assigned to the stable tree after inserting ksm
 835 * page and upping mmap_lock.  Nor does it fit with the way we skip dup'ing
 836 * rmap_items from parent to child at fork time (so as not to waste time
 837 * if exit comes before the next scan reaches it).
 838 *
 839 * Similarly, although we'd like to remove rmap_items (so updating counts
 840 * and freeing memory) when unmerging an area, it's easier to leave that
 841 * to the next pass of ksmd - consider, for example, how ksmd might be
 842 * in cmp_and_merge_page on one of the rmap_items we would be removing.
 843 */
 844static int unmerge_ksm_pages(struct vm_area_struct *vma,
 845			     unsigned long start, unsigned long end)
 846{
 847	unsigned long addr;
 848	int err = 0;
 849
 850	for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
 851		if (ksm_test_exit(vma->vm_mm))
 852			break;
 853		if (signal_pending(current))
 854			err = -ERESTARTSYS;
 855		else
 856			err = break_ksm(vma, addr);
 857	}
 858	return err;
 859}
 860
 861static inline struct stable_node *page_stable_node(struct page *page)
 862{
 863	return PageKsm(page) ? page_rmapping(page) : NULL;
 864}
 865
 866static inline void set_page_stable_node(struct page *page,
 867					struct stable_node *stable_node)
 868{
 869	page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
 870}
 871
 872#ifdef CONFIG_SYSFS
 873/*
 874 * Only called through the sysfs control interface:
 875 */
 876static int remove_stable_node(struct stable_node *stable_node)
 877{
 878	struct page *page;
 879	int err;
 880
 881	page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
 882	if (!page) {
 883		/*
 884		 * get_ksm_page did remove_node_from_stable_tree itself.
 885		 */
 886		return 0;
 887	}
 888
 889	/*
 890	 * Page could be still mapped if this races with __mmput() running in
 891	 * between ksm_exit() and exit_mmap(). Just refuse to let
 892	 * merge_across_nodes/max_page_sharing be switched.
 893	 */
 894	err = -EBUSY;
 895	if (!page_mapped(page)) {
 896		/*
 897		 * The stable node did not yet appear stale to get_ksm_page(),
 898		 * since that allows for an unmapped ksm page to be recognized
 899		 * right up until it is freed; but the node is safe to remove.
 900		 * This page might be in a pagevec waiting to be freed,
 901		 * or it might be PageSwapCache (perhaps under writeback),
 902		 * or it might have been removed from swapcache a moment ago.
 903		 */
 904		set_page_stable_node(page, NULL);
 905		remove_node_from_stable_tree(stable_node);
 906		err = 0;
 907	}
 908
 909	unlock_page(page);
 910	put_page(page);
 911	return err;
 912}
 913
 914static int remove_stable_node_chain(struct stable_node *stable_node,
 915				    struct rb_root *root)
 916{
 917	struct stable_node *dup;
 918	struct hlist_node *hlist_safe;
 919
 920	if (!is_stable_node_chain(stable_node)) {
 921		VM_BUG_ON(is_stable_node_dup(stable_node));
 922		if (remove_stable_node(stable_node))
 923			return true;
 924		else
 925			return false;
 926	}
 927
 928	hlist_for_each_entry_safe(dup, hlist_safe,
 929				  &stable_node->hlist, hlist_dup) {
 930		VM_BUG_ON(!is_stable_node_dup(dup));
 931		if (remove_stable_node(dup))
 932			return true;
 933	}
 934	BUG_ON(!hlist_empty(&stable_node->hlist));
 935	free_stable_node_chain(stable_node, root);
 936	return false;
 937}
 938
 939static int remove_all_stable_nodes(void)
 940{
 941	struct stable_node *stable_node, *next;
 942	int nid;
 943	int err = 0;
 944
 945	for (nid = 0; nid < ksm_nr_node_ids; nid++) {
 946		while (root_stable_tree[nid].rb_node) {
 947			stable_node = rb_entry(root_stable_tree[nid].rb_node,
 948						struct stable_node, node);
 949			if (remove_stable_node_chain(stable_node,
 950						     root_stable_tree + nid)) {
 951				err = -EBUSY;
 952				break;	/* proceed to next nid */
 953			}
 954			cond_resched();
 955		}
 956	}
 957	list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
 958		if (remove_stable_node(stable_node))
 959			err = -EBUSY;
 960		cond_resched();
 961	}
 962	return err;
 963}
 964
 965static int unmerge_and_remove_all_rmap_items(void)
 966{
 967	struct mm_slot *mm_slot;
 968	struct mm_struct *mm;
 969	struct vm_area_struct *vma;
 970	int err = 0;
 971
 972	spin_lock(&ksm_mmlist_lock);
 973	ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
 974						struct mm_slot, mm_list);
 975	spin_unlock(&ksm_mmlist_lock);
 976
 977	for (mm_slot = ksm_scan.mm_slot;
 978			mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
 979		mm = mm_slot->mm;
 980		mmap_read_lock(mm);
 981		for (vma = mm->mmap; vma; vma = vma->vm_next) {
 982			if (ksm_test_exit(mm))
 983				break;
 984			if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
 985				continue;
 986			err = unmerge_ksm_pages(vma,
 987						vma->vm_start, vma->vm_end);
 988			if (err)
 989				goto error;
 990		}
 991
 992		remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
 993		mmap_read_unlock(mm);
 994
 995		spin_lock(&ksm_mmlist_lock);
 996		ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
 997						struct mm_slot, mm_list);
 998		if (ksm_test_exit(mm)) {
 999			hash_del(&mm_slot->link);
1000			list_del(&mm_slot->mm_list);
1001			spin_unlock(&ksm_mmlist_lock);
1002
1003			free_mm_slot(mm_slot);
1004			clear_bit(MMF_VM_MERGEABLE, &mm->flags);
 
1005			mmdrop(mm);
1006		} else
1007			spin_unlock(&ksm_mmlist_lock);
 
 
1008	}
1009
1010	/* Clean up stable nodes, but don't worry if some are still busy */
1011	remove_all_stable_nodes();
1012	ksm_scan.seqnr = 0;
1013	return 0;
1014
1015error:
1016	mmap_read_unlock(mm);
1017	spin_lock(&ksm_mmlist_lock);
1018	ksm_scan.mm_slot = &ksm_mm_head;
1019	spin_unlock(&ksm_mmlist_lock);
1020	return err;
1021}
1022#endif /* CONFIG_SYSFS */
1023
1024static u32 calc_checksum(struct page *page)
1025{
1026	u32 checksum;
1027	void *addr = kmap_atomic(page);
1028	checksum = xxhash(addr, PAGE_SIZE, 0);
1029	kunmap_atomic(addr);
1030	return checksum;
1031}
1032
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1033static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1034			      pte_t *orig_pte)
1035{
1036	struct mm_struct *mm = vma->vm_mm;
1037	struct page_vma_mapped_walk pvmw = {
1038		.page = page,
1039		.vma = vma,
1040	};
1041	int swapped;
1042	int err = -EFAULT;
1043	struct mmu_notifier_range range;
1044
1045	pvmw.address = page_address_in_vma(page, vma);
1046	if (pvmw.address == -EFAULT)
1047		goto out;
1048
1049	BUG_ON(PageTransCompound(page));
 
 
 
1050
1051	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
1052				pvmw.address,
1053				pvmw.address + PAGE_SIZE);
1054	mmu_notifier_invalidate_range_start(&range);
1055
1056	if (!page_vma_mapped_walk(&pvmw))
1057		goto out_mn;
1058	if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1059		goto out_unlock;
1060
1061	if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1062	    (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1063						mm_tlb_flush_pending(mm)) {
1064		pte_t entry;
1065
1066		swapped = PageSwapCache(page);
1067		flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1068		/*
1069		 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1070		 * take any lock, therefore the check that we are going to make
1071		 * with the pagecount against the mapcount is racey and
1072		 * O_DIRECT can happen right after the check.
1073		 * So we clear the pte and flush the tlb before the check
1074		 * this assure us that no O_DIRECT can happen after the check
1075		 * or in the middle of the check.
1076		 *
1077		 * No need to notify as we are downgrading page table to read
1078		 * only not changing it to point to a new page.
1079		 *
1080		 * See Documentation/vm/mmu_notifier.rst
1081		 */
1082		entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1083		/*
1084		 * Check that no O_DIRECT or similar I/O is in progress on the
1085		 * page
1086		 */
1087		if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1088			set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1089			goto out_unlock;
1090		}
1091		if (pte_dirty(entry))
1092			set_page_dirty(page);
1093
1094		if (pte_protnone(entry))
1095			entry = pte_mkclean(pte_clear_savedwrite(entry));
1096		else
1097			entry = pte_mkclean(pte_wrprotect(entry));
1098		set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1099	}
1100	*orig_pte = *pvmw.pte;
1101	err = 0;
1102
1103out_unlock:
1104	page_vma_mapped_walk_done(&pvmw);
1105out_mn:
1106	mmu_notifier_invalidate_range_end(&range);
1107out:
1108	return err;
1109}
1110
1111/**
1112 * replace_page - replace page in vma by new ksm page
1113 * @vma:      vma that holds the pte pointing to page
1114 * @page:     the page we are replacing by kpage
1115 * @kpage:    the ksm page we replace page by
1116 * @orig_pte: the original value of the pte
1117 *
1118 * Returns 0 on success, -EFAULT on failure.
1119 */
1120static int replace_page(struct vm_area_struct *vma, struct page *page,
1121			struct page *kpage, pte_t orig_pte)
1122{
1123	struct mm_struct *mm = vma->vm_mm;
 
 
1124	pmd_t *pmd;
1125	pte_t *ptep;
1126	pte_t newpte;
1127	spinlock_t *ptl;
1128	unsigned long addr;
1129	int err = -EFAULT;
1130	struct mmu_notifier_range range;
1131
1132	addr = page_address_in_vma(page, vma);
1133	if (addr == -EFAULT)
1134		goto out;
1135
1136	pmd = mm_find_pmd(mm, addr);
1137	if (!pmd)
1138		goto out;
1139
1140	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr,
1141				addr + PAGE_SIZE);
1142	mmu_notifier_invalidate_range_start(&range);
 
 
 
 
 
1143
1144	ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1145	if (!pte_same(*ptep, orig_pte)) {
1146		pte_unmap_unlock(ptep, ptl);
1147		goto out_mn;
1148	}
1149
1150	/*
1151	 * No need to check ksm_use_zero_pages here: we can only have a
1152	 * zero_page here if ksm_use_zero_pages was enabled already.
1153	 */
1154	if (!is_zero_pfn(page_to_pfn(kpage))) {
1155		get_page(kpage);
1156		page_add_anon_rmap(kpage, vma, addr, false);
1157		newpte = mk_pte(kpage, vma->vm_page_prot);
1158	} else {
1159		newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1160					       vma->vm_page_prot));
1161		/*
1162		 * We're replacing an anonymous page with a zero page, which is
1163		 * not anonymous. We need to do proper accounting otherwise we
1164		 * will get wrong values in /proc, and a BUG message in dmesg
1165		 * when tearing down the mm.
1166		 */
1167		dec_mm_counter(mm, MM_ANONPAGES);
1168	}
1169
1170	flush_cache_page(vma, addr, pte_pfn(*ptep));
1171	/*
1172	 * No need to notify as we are replacing a read only page with another
1173	 * read only page with the same content.
1174	 *
1175	 * See Documentation/vm/mmu_notifier.rst
1176	 */
1177	ptep_clear_flush(vma, addr, ptep);
1178	set_pte_at_notify(mm, addr, ptep, newpte);
1179
1180	page_remove_rmap(page, false);
1181	if (!page_mapped(page))
1182		try_to_free_swap(page);
1183	put_page(page);
1184
1185	pte_unmap_unlock(ptep, ptl);
1186	err = 0;
1187out_mn:
1188	mmu_notifier_invalidate_range_end(&range);
1189out:
1190	return err;
1191}
1192
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1193/*
1194 * try_to_merge_one_page - take two pages and merge them into one
1195 * @vma: the vma that holds the pte pointing to page
1196 * @page: the PageAnon page that we want to replace with kpage
1197 * @kpage: the PageKsm page that we want to map instead of page,
1198 *         or NULL the first time when we want to use page as kpage.
1199 *
1200 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1201 */
1202static int try_to_merge_one_page(struct vm_area_struct *vma,
1203				 struct page *page, struct page *kpage)
1204{
1205	pte_t orig_pte = __pte(0);
1206	int err = -EFAULT;
1207
1208	if (page == kpage)			/* ksm page forked */
1209		return 0;
1210
 
 
 
 
 
1211	if (!PageAnon(page))
1212		goto out;
1213
1214	/*
1215	 * We need the page lock to read a stable PageSwapCache in
1216	 * write_protect_page().  We use trylock_page() instead of
1217	 * lock_page() because we don't want to wait here - we
1218	 * prefer to continue scanning and merging different pages,
1219	 * then come back to this page when it is unlocked.
1220	 */
1221	if (!trylock_page(page))
1222		goto out;
1223
1224	if (PageTransCompound(page)) {
1225		if (split_huge_page(page))
1226			goto out_unlock;
1227	}
1228
1229	/*
1230	 * If this anonymous page is mapped only here, its pte may need
1231	 * to be write-protected.  If it's mapped elsewhere, all of its
1232	 * ptes are necessarily already write-protected.  But in either
1233	 * case, we need to lock and check page_count is not raised.
1234	 */
1235	if (write_protect_page(vma, page, &orig_pte) == 0) {
1236		if (!kpage) {
1237			/*
1238			 * While we hold page lock, upgrade page from
1239			 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1240			 * stable_tree_insert() will update stable_node.
1241			 */
1242			set_page_stable_node(page, NULL);
1243			mark_page_accessed(page);
1244			/*
1245			 * Page reclaim just frees a clean page with no dirty
1246			 * ptes: make sure that the ksm page would be swapped.
1247			 */
1248			if (!PageDirty(page))
1249				SetPageDirty(page);
1250			err = 0;
1251		} else if (pages_identical(page, kpage))
1252			err = replace_page(vma, page, kpage, orig_pte);
1253	}
1254
1255	if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1256		munlock_vma_page(page);
1257		if (!PageMlocked(kpage)) {
1258			unlock_page(page);
1259			lock_page(kpage);
1260			mlock_vma_page(kpage);
1261			page = kpage;		/* for final unlock */
1262		}
1263	}
1264
1265out_unlock:
1266	unlock_page(page);
1267out:
1268	return err;
1269}
1270
1271/*
1272 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1273 * but no new kernel page is allocated: kpage must already be a ksm page.
1274 *
1275 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1276 */
1277static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1278				      struct page *page, struct page *kpage)
1279{
1280	struct mm_struct *mm = rmap_item->mm;
1281	struct vm_area_struct *vma;
1282	int err = -EFAULT;
1283
1284	mmap_read_lock(mm);
1285	vma = find_mergeable_vma(mm, rmap_item->address);
1286	if (!vma)
 
 
1287		goto out;
1288
1289	err = try_to_merge_one_page(vma, page, kpage);
1290	if (err)
1291		goto out;
1292
1293	/* Unstable nid is in union with stable anon_vma: remove first */
1294	remove_rmap_item_from_tree(rmap_item);
1295
1296	/* Must get reference to anon_vma while still holding mmap_lock */
1297	rmap_item->anon_vma = vma->anon_vma;
1298	get_anon_vma(vma->anon_vma);
1299out:
1300	mmap_read_unlock(mm);
1301	return err;
1302}
1303
1304/*
1305 * try_to_merge_two_pages - take two identical pages and prepare them
1306 * to be merged into one page.
1307 *
1308 * This function returns the kpage if we successfully merged two identical
1309 * pages into one ksm page, NULL otherwise.
1310 *
1311 * Note that this function upgrades page to ksm page: if one of the pages
1312 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1313 */
1314static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1315					   struct page *page,
1316					   struct rmap_item *tree_rmap_item,
1317					   struct page *tree_page)
1318{
1319	int err;
1320
1321	err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1322	if (!err) {
1323		err = try_to_merge_with_ksm_page(tree_rmap_item,
1324							tree_page, page);
1325		/*
1326		 * If that fails, we have a ksm page with only one pte
1327		 * pointing to it: so break it.
1328		 */
1329		if (err)
1330			break_cow(rmap_item);
1331	}
1332	return err ? NULL : page;
1333}
1334
1335static __always_inline
1336bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1337{
1338	VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1339	/*
1340	 * Check that at least one mapping still exists, otherwise
1341	 * there's no much point to merge and share with this
1342	 * stable_node, as the underlying tree_page of the other
1343	 * sharer is going to be freed soon.
1344	 */
1345	return stable_node->rmap_hlist_len &&
1346		stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1347}
1348
1349static __always_inline
1350bool is_page_sharing_candidate(struct stable_node *stable_node)
1351{
1352	return __is_page_sharing_candidate(stable_node, 0);
1353}
1354
1355static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1356				    struct stable_node **_stable_node,
1357				    struct rb_root *root,
1358				    bool prune_stale_stable_nodes)
1359{
1360	struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1361	struct hlist_node *hlist_safe;
1362	struct page *_tree_page, *tree_page = NULL;
1363	int nr = 0;
1364	int found_rmap_hlist_len;
1365
1366	if (!prune_stale_stable_nodes ||
1367	    time_before(jiffies, stable_node->chain_prune_time +
1368			msecs_to_jiffies(
1369				ksm_stable_node_chains_prune_millisecs)))
1370		prune_stale_stable_nodes = false;
1371	else
1372		stable_node->chain_prune_time = jiffies;
1373
1374	hlist_for_each_entry_safe(dup, hlist_safe,
1375				  &stable_node->hlist, hlist_dup) {
1376		cond_resched();
1377		/*
1378		 * We must walk all stable_node_dup to prune the stale
1379		 * stable nodes during lookup.
1380		 *
1381		 * get_ksm_page can drop the nodes from the
1382		 * stable_node->hlist if they point to freed pages
1383		 * (that's why we do a _safe walk). The "dup"
1384		 * stable_node parameter itself will be freed from
1385		 * under us if it returns NULL.
1386		 */
1387		_tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK);
1388		if (!_tree_page)
1389			continue;
1390		nr += 1;
1391		if (is_page_sharing_candidate(dup)) {
1392			if (!found ||
1393			    dup->rmap_hlist_len > found_rmap_hlist_len) {
1394				if (found)
1395					put_page(tree_page);
1396				found = dup;
1397				found_rmap_hlist_len = found->rmap_hlist_len;
1398				tree_page = _tree_page;
1399
1400				/* skip put_page for found dup */
1401				if (!prune_stale_stable_nodes)
1402					break;
1403				continue;
1404			}
1405		}
1406		put_page(_tree_page);
1407	}
1408
1409	if (found) {
1410		/*
1411		 * nr is counting all dups in the chain only if
1412		 * prune_stale_stable_nodes is true, otherwise we may
1413		 * break the loop at nr == 1 even if there are
1414		 * multiple entries.
1415		 */
1416		if (prune_stale_stable_nodes && nr == 1) {
1417			/*
1418			 * If there's not just one entry it would
1419			 * corrupt memory, better BUG_ON. In KSM
1420			 * context with no lock held it's not even
1421			 * fatal.
1422			 */
1423			BUG_ON(stable_node->hlist.first->next);
1424
1425			/*
1426			 * There's just one entry and it is below the
1427			 * deduplication limit so drop the chain.
1428			 */
1429			rb_replace_node(&stable_node->node, &found->node,
1430					root);
1431			free_stable_node(stable_node);
1432			ksm_stable_node_chains--;
1433			ksm_stable_node_dups--;
1434			/*
1435			 * NOTE: the caller depends on the stable_node
1436			 * to be equal to stable_node_dup if the chain
1437			 * was collapsed.
1438			 */
1439			*_stable_node = found;
1440			/*
1441			 * Just for robustneess as stable_node is
1442			 * otherwise left as a stable pointer, the
1443			 * compiler shall optimize it away at build
1444			 * time.
1445			 */
1446			stable_node = NULL;
1447		} else if (stable_node->hlist.first != &found->hlist_dup &&
1448			   __is_page_sharing_candidate(found, 1)) {
1449			/*
1450			 * If the found stable_node dup can accept one
1451			 * more future merge (in addition to the one
1452			 * that is underway) and is not at the head of
1453			 * the chain, put it there so next search will
1454			 * be quicker in the !prune_stale_stable_nodes
1455			 * case.
1456			 *
1457			 * NOTE: it would be inaccurate to use nr > 1
1458			 * instead of checking the hlist.first pointer
1459			 * directly, because in the
1460			 * prune_stale_stable_nodes case "nr" isn't
1461			 * the position of the found dup in the chain,
1462			 * but the total number of dups in the chain.
1463			 */
1464			hlist_del(&found->hlist_dup);
1465			hlist_add_head(&found->hlist_dup,
1466				       &stable_node->hlist);
1467		}
1468	}
1469
1470	*_stable_node_dup = found;
1471	return tree_page;
1472}
1473
1474static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1475					       struct rb_root *root)
1476{
1477	if (!is_stable_node_chain(stable_node))
1478		return stable_node;
1479	if (hlist_empty(&stable_node->hlist)) {
1480		free_stable_node_chain(stable_node, root);
1481		return NULL;
1482	}
1483	return hlist_entry(stable_node->hlist.first,
1484			   typeof(*stable_node), hlist_dup);
1485}
1486
1487/*
1488 * Like for get_ksm_page, this function can free the *_stable_node and
1489 * *_stable_node_dup if the returned tree_page is NULL.
1490 *
1491 * It can also free and overwrite *_stable_node with the found
1492 * stable_node_dup if the chain is collapsed (in which case
1493 * *_stable_node will be equal to *_stable_node_dup like if the chain
1494 * never existed). It's up to the caller to verify tree_page is not
1495 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1496 *
1497 * *_stable_node_dup is really a second output parameter of this
1498 * function and will be overwritten in all cases, the caller doesn't
1499 * need to initialize it.
1500 */
1501static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1502					struct stable_node **_stable_node,
1503					struct rb_root *root,
1504					bool prune_stale_stable_nodes)
1505{
1506	struct stable_node *stable_node = *_stable_node;
1507	if (!is_stable_node_chain(stable_node)) {
1508		if (is_page_sharing_candidate(stable_node)) {
1509			*_stable_node_dup = stable_node;
1510			return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK);
1511		}
1512		/*
1513		 * _stable_node_dup set to NULL means the stable_node
1514		 * reached the ksm_max_page_sharing limit.
1515		 */
1516		*_stable_node_dup = NULL;
1517		return NULL;
1518	}
1519	return stable_node_dup(_stable_node_dup, _stable_node, root,
1520			       prune_stale_stable_nodes);
1521}
1522
1523static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1524						struct stable_node **s_n,
1525						struct rb_root *root)
1526{
1527	return __stable_node_chain(s_n_d, s_n, root, true);
1528}
1529
1530static __always_inline struct page *chain(struct stable_node **s_n_d,
1531					  struct stable_node *s_n,
1532					  struct rb_root *root)
1533{
1534	struct stable_node *old_stable_node = s_n;
1535	struct page *tree_page;
1536
1537	tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1538	/* not pruning dups so s_n cannot have changed */
1539	VM_BUG_ON(s_n != old_stable_node);
1540	return tree_page;
1541}
1542
1543/*
1544 * stable_tree_search - search for page inside the stable tree
1545 *
1546 * This function checks if there is a page inside the stable tree
1547 * with identical content to the page that we are scanning right now.
1548 *
1549 * This function returns the stable tree node of identical content if found,
1550 * NULL otherwise.
1551 */
1552static struct page *stable_tree_search(struct page *page)
1553{
1554	int nid;
1555	struct rb_root *root;
1556	struct rb_node **new;
1557	struct rb_node *parent;
1558	struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1559	struct stable_node *page_node;
1560
1561	page_node = page_stable_node(page);
1562	if (page_node && page_node->head != &migrate_nodes) {
1563		/* ksm page forked */
1564		get_page(page);
1565		return page;
1566	}
1567
1568	nid = get_kpfn_nid(page_to_pfn(page));
1569	root = root_stable_tree + nid;
1570again:
1571	new = &root->rb_node;
1572	parent = NULL;
1573
1574	while (*new) {
1575		struct page *tree_page;
1576		int ret;
1577
1578		cond_resched();
1579		stable_node = rb_entry(*new, struct stable_node, node);
1580		stable_node_any = NULL;
1581		tree_page = chain_prune(&stable_node_dup, &stable_node,	root);
1582		/*
1583		 * NOTE: stable_node may have been freed by
1584		 * chain_prune() if the returned stable_node_dup is
1585		 * not NULL. stable_node_dup may have been inserted in
1586		 * the rbtree instead as a regular stable_node (in
1587		 * order to collapse the stable_node chain if a single
1588		 * stable_node dup was found in it). In such case the
1589		 * stable_node is overwritten by the calleee to point
1590		 * to the stable_node_dup that was collapsed in the
1591		 * stable rbtree and stable_node will be equal to
1592		 * stable_node_dup like if the chain never existed.
1593		 */
1594		if (!stable_node_dup) {
1595			/*
1596			 * Either all stable_node dups were full in
1597			 * this stable_node chain, or this chain was
1598			 * empty and should be rb_erased.
1599			 */
1600			stable_node_any = stable_node_dup_any(stable_node,
1601							      root);
1602			if (!stable_node_any) {
1603				/* rb_erase just run */
1604				goto again;
1605			}
1606			/*
1607			 * Take any of the stable_node dups page of
1608			 * this stable_node chain to let the tree walk
1609			 * continue. All KSM pages belonging to the
1610			 * stable_node dups in a stable_node chain
1611			 * have the same content and they're
1612			 * write protected at all times. Any will work
1613			 * fine to continue the walk.
1614			 */
1615			tree_page = get_ksm_page(stable_node_any,
1616						 GET_KSM_PAGE_NOLOCK);
1617		}
1618		VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1619		if (!tree_page) {
1620			/*
1621			 * If we walked over a stale stable_node,
1622			 * get_ksm_page() will call rb_erase() and it
1623			 * may rebalance the tree from under us. So
1624			 * restart the search from scratch. Returning
1625			 * NULL would be safe too, but we'd generate
1626			 * false negative insertions just because some
1627			 * stable_node was stale.
1628			 */
1629			goto again;
1630		}
1631
1632		ret = memcmp_pages(page, tree_page);
1633		put_page(tree_page);
1634
1635		parent = *new;
1636		if (ret < 0)
1637			new = &parent->rb_left;
1638		else if (ret > 0)
1639			new = &parent->rb_right;
1640		else {
1641			if (page_node) {
1642				VM_BUG_ON(page_node->head != &migrate_nodes);
1643				/*
1644				 * Test if the migrated page should be merged
1645				 * into a stable node dup. If the mapcount is
1646				 * 1 we can migrate it with another KSM page
1647				 * without adding it to the chain.
1648				 */
1649				if (page_mapcount(page) > 1)
1650					goto chain_append;
1651			}
1652
1653			if (!stable_node_dup) {
1654				/*
1655				 * If the stable_node is a chain and
1656				 * we got a payload match in memcmp
1657				 * but we cannot merge the scanned
1658				 * page in any of the existing
1659				 * stable_node dups because they're
1660				 * all full, we need to wait the
1661				 * scanned page to find itself a match
1662				 * in the unstable tree to create a
1663				 * brand new KSM page to add later to
1664				 * the dups of this stable_node.
1665				 */
1666				return NULL;
1667			}
1668
1669			/*
1670			 * Lock and unlock the stable_node's page (which
1671			 * might already have been migrated) so that page
1672			 * migration is sure to notice its raised count.
1673			 * It would be more elegant to return stable_node
1674			 * than kpage, but that involves more changes.
1675			 */
1676			tree_page = get_ksm_page(stable_node_dup,
1677						 GET_KSM_PAGE_TRYLOCK);
1678
1679			if (PTR_ERR(tree_page) == -EBUSY)
1680				return ERR_PTR(-EBUSY);
1681
1682			if (unlikely(!tree_page))
1683				/*
1684				 * The tree may have been rebalanced,
1685				 * so re-evaluate parent and new.
1686				 */
1687				goto again;
1688			unlock_page(tree_page);
1689
1690			if (get_kpfn_nid(stable_node_dup->kpfn) !=
1691			    NUMA(stable_node_dup->nid)) {
1692				put_page(tree_page);
1693				goto replace;
1694			}
1695			return tree_page;
1696		}
1697	}
1698
1699	if (!page_node)
1700		return NULL;
1701
1702	list_del(&page_node->list);
1703	DO_NUMA(page_node->nid = nid);
1704	rb_link_node(&page_node->node, parent, new);
1705	rb_insert_color(&page_node->node, root);
1706out:
1707	if (is_page_sharing_candidate(page_node)) {
1708		get_page(page);
1709		return page;
1710	} else
1711		return NULL;
1712
1713replace:
1714	/*
1715	 * If stable_node was a chain and chain_prune collapsed it,
1716	 * stable_node has been updated to be the new regular
1717	 * stable_node. A collapse of the chain is indistinguishable
1718	 * from the case there was no chain in the stable
1719	 * rbtree. Otherwise stable_node is the chain and
1720	 * stable_node_dup is the dup to replace.
1721	 */
1722	if (stable_node_dup == stable_node) {
1723		VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1724		VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1725		/* there is no chain */
1726		if (page_node) {
1727			VM_BUG_ON(page_node->head != &migrate_nodes);
1728			list_del(&page_node->list);
1729			DO_NUMA(page_node->nid = nid);
1730			rb_replace_node(&stable_node_dup->node,
1731					&page_node->node,
1732					root);
1733			if (is_page_sharing_candidate(page_node))
1734				get_page(page);
1735			else
1736				page = NULL;
1737		} else {
1738			rb_erase(&stable_node_dup->node, root);
1739			page = NULL;
1740		}
1741	} else {
1742		VM_BUG_ON(!is_stable_node_chain(stable_node));
1743		__stable_node_dup_del(stable_node_dup);
1744		if (page_node) {
1745			VM_BUG_ON(page_node->head != &migrate_nodes);
1746			list_del(&page_node->list);
1747			DO_NUMA(page_node->nid = nid);
1748			stable_node_chain_add_dup(page_node, stable_node);
1749			if (is_page_sharing_candidate(page_node))
1750				get_page(page);
1751			else
1752				page = NULL;
1753		} else {
1754			page = NULL;
1755		}
1756	}
1757	stable_node_dup->head = &migrate_nodes;
1758	list_add(&stable_node_dup->list, stable_node_dup->head);
1759	return page;
1760
1761chain_append:
1762	/* stable_node_dup could be null if it reached the limit */
1763	if (!stable_node_dup)
1764		stable_node_dup = stable_node_any;
1765	/*
1766	 * If stable_node was a chain and chain_prune collapsed it,
1767	 * stable_node has been updated to be the new regular
1768	 * stable_node. A collapse of the chain is indistinguishable
1769	 * from the case there was no chain in the stable
1770	 * rbtree. Otherwise stable_node is the chain and
1771	 * stable_node_dup is the dup to replace.
1772	 */
1773	if (stable_node_dup == stable_node) {
1774		VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1775		VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1776		/* chain is missing so create it */
1777		stable_node = alloc_stable_node_chain(stable_node_dup,
1778						      root);
1779		if (!stable_node)
1780			return NULL;
1781	}
1782	/*
1783	 * Add this stable_node dup that was
1784	 * migrated to the stable_node chain
1785	 * of the current nid for this page
1786	 * content.
1787	 */
1788	VM_BUG_ON(!is_stable_node_chain(stable_node));
1789	VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1790	VM_BUG_ON(page_node->head != &migrate_nodes);
1791	list_del(&page_node->list);
1792	DO_NUMA(page_node->nid = nid);
1793	stable_node_chain_add_dup(page_node, stable_node);
1794	goto out;
1795}
1796
1797/*
1798 * stable_tree_insert - insert stable tree node pointing to new ksm page
1799 * into the stable tree.
1800 *
1801 * This function returns the stable tree node just allocated on success,
1802 * NULL otherwise.
1803 */
1804static struct stable_node *stable_tree_insert(struct page *kpage)
1805{
1806	int nid;
1807	unsigned long kpfn;
1808	struct rb_root *root;
1809	struct rb_node **new;
1810	struct rb_node *parent;
1811	struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1812	bool need_chain = false;
1813
1814	kpfn = page_to_pfn(kpage);
1815	nid = get_kpfn_nid(kpfn);
1816	root = root_stable_tree + nid;
1817again:
1818	parent = NULL;
1819	new = &root->rb_node;
1820
1821	while (*new) {
1822		struct page *tree_page;
1823		int ret;
1824
1825		cond_resched();
1826		stable_node = rb_entry(*new, struct stable_node, node);
1827		stable_node_any = NULL;
1828		tree_page = chain(&stable_node_dup, stable_node, root);
1829		if (!stable_node_dup) {
1830			/*
1831			 * Either all stable_node dups were full in
1832			 * this stable_node chain, or this chain was
1833			 * empty and should be rb_erased.
1834			 */
1835			stable_node_any = stable_node_dup_any(stable_node,
1836							      root);
1837			if (!stable_node_any) {
1838				/* rb_erase just run */
1839				goto again;
1840			}
1841			/*
1842			 * Take any of the stable_node dups page of
1843			 * this stable_node chain to let the tree walk
1844			 * continue. All KSM pages belonging to the
1845			 * stable_node dups in a stable_node chain
1846			 * have the same content and they're
1847			 * write protected at all times. Any will work
1848			 * fine to continue the walk.
1849			 */
1850			tree_page = get_ksm_page(stable_node_any,
1851						 GET_KSM_PAGE_NOLOCK);
1852		}
1853		VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1854		if (!tree_page) {
1855			/*
1856			 * If we walked over a stale stable_node,
1857			 * get_ksm_page() will call rb_erase() and it
1858			 * may rebalance the tree from under us. So
1859			 * restart the search from scratch. Returning
1860			 * NULL would be safe too, but we'd generate
1861			 * false negative insertions just because some
1862			 * stable_node was stale.
1863			 */
1864			goto again;
1865		}
1866
1867		ret = memcmp_pages(kpage, tree_page);
1868		put_page(tree_page);
1869
1870		parent = *new;
1871		if (ret < 0)
1872			new = &parent->rb_left;
1873		else if (ret > 0)
1874			new = &parent->rb_right;
1875		else {
1876			need_chain = true;
1877			break;
 
 
 
 
1878		}
1879	}
1880
1881	stable_node_dup = alloc_stable_node();
1882	if (!stable_node_dup)
1883		return NULL;
1884
1885	INIT_HLIST_HEAD(&stable_node_dup->hlist);
1886	stable_node_dup->kpfn = kpfn;
1887	set_page_stable_node(kpage, stable_node_dup);
1888	stable_node_dup->rmap_hlist_len = 0;
1889	DO_NUMA(stable_node_dup->nid = nid);
1890	if (!need_chain) {
1891		rb_link_node(&stable_node_dup->node, parent, new);
1892		rb_insert_color(&stable_node_dup->node, root);
1893	} else {
1894		if (!is_stable_node_chain(stable_node)) {
1895			struct stable_node *orig = stable_node;
1896			/* chain is missing so create it */
1897			stable_node = alloc_stable_node_chain(orig, root);
1898			if (!stable_node) {
1899				free_stable_node(stable_node_dup);
1900				return NULL;
1901			}
1902		}
1903		stable_node_chain_add_dup(stable_node_dup, stable_node);
1904	}
1905
1906	return stable_node_dup;
1907}
1908
1909/*
1910 * unstable_tree_search_insert - search for identical page,
1911 * else insert rmap_item into the unstable tree.
1912 *
1913 * This function searches for a page in the unstable tree identical to the
1914 * page currently being scanned; and if no identical page is found in the
1915 * tree, we insert rmap_item as a new object into the unstable tree.
1916 *
1917 * This function returns pointer to rmap_item found to be identical
1918 * to the currently scanned page, NULL otherwise.
1919 *
1920 * This function does both searching and inserting, because they share
1921 * the same walking algorithm in an rbtree.
1922 */
1923static
1924struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1925					      struct page *page,
1926					      struct page **tree_pagep)
 
1927{
1928	struct rb_node **new;
1929	struct rb_root *root;
1930	struct rb_node *parent = NULL;
1931	int nid;
1932
1933	nid = get_kpfn_nid(page_to_pfn(page));
1934	root = root_unstable_tree + nid;
1935	new = &root->rb_node;
1936
1937	while (*new) {
1938		struct rmap_item *tree_rmap_item;
1939		struct page *tree_page;
1940		int ret;
1941
1942		cond_resched();
1943		tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1944		tree_page = get_mergeable_page(tree_rmap_item);
1945		if (!tree_page)
1946			return NULL;
1947
1948		/*
1949		 * Don't substitute a ksm page for a forked page.
1950		 */
1951		if (page == tree_page) {
1952			put_page(tree_page);
1953			return NULL;
1954		}
1955
1956		ret = memcmp_pages(page, tree_page);
1957
1958		parent = *new;
1959		if (ret < 0) {
1960			put_page(tree_page);
1961			new = &parent->rb_left;
1962		} else if (ret > 0) {
1963			put_page(tree_page);
1964			new = &parent->rb_right;
1965		} else if (!ksm_merge_across_nodes &&
1966			   page_to_nid(tree_page) != nid) {
1967			/*
1968			 * If tree_page has been migrated to another NUMA node,
1969			 * it will be flushed out and put in the right unstable
1970			 * tree next time: only merge with it when across_nodes.
1971			 */
1972			put_page(tree_page);
1973			return NULL;
1974		} else {
1975			*tree_pagep = tree_page;
1976			return tree_rmap_item;
1977		}
1978	}
1979
1980	rmap_item->address |= UNSTABLE_FLAG;
1981	rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1982	DO_NUMA(rmap_item->nid = nid);
1983	rb_link_node(&rmap_item->node, parent, new);
1984	rb_insert_color(&rmap_item->node, root);
1985
1986	ksm_pages_unshared++;
1987	return NULL;
1988}
1989
1990/*
1991 * stable_tree_append - add another rmap_item to the linked list of
1992 * rmap_items hanging off a given node of the stable tree, all sharing
1993 * the same ksm page.
1994 */
1995static void stable_tree_append(struct rmap_item *rmap_item,
1996			       struct stable_node *stable_node,
1997			       bool max_page_sharing_bypass)
1998{
1999	/*
2000	 * rmap won't find this mapping if we don't insert the
2001	 * rmap_item in the right stable_node
2002	 * duplicate. page_migration could break later if rmap breaks,
2003	 * so we can as well crash here. We really need to check for
2004	 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
2005	 * for other negative values as an underflow if detected here
2006	 * for the first time (and not when decreasing rmap_hlist_len)
2007	 * would be sign of memory corruption in the stable_node.
2008	 */
2009	BUG_ON(stable_node->rmap_hlist_len < 0);
2010
2011	stable_node->rmap_hlist_len++;
2012	if (!max_page_sharing_bypass)
2013		/* possibly non fatal but unexpected overflow, only warn */
2014		WARN_ON_ONCE(stable_node->rmap_hlist_len >
2015			     ksm_max_page_sharing);
2016
2017	rmap_item->head = stable_node;
2018	rmap_item->address |= STABLE_FLAG;
2019	hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2020
2021	if (rmap_item->hlist.next)
2022		ksm_pages_sharing++;
2023	else
2024		ksm_pages_shared++;
2025}
2026
2027/*
2028 * cmp_and_merge_page - first see if page can be merged into the stable tree;
2029 * if not, compare checksum to previous and if it's the same, see if page can
2030 * be inserted into the unstable tree, or merged with a page already there and
2031 * both transferred to the stable tree.
2032 *
2033 * @page: the page that we are searching identical page to.
2034 * @rmap_item: the reverse mapping into the virtual address of this page
2035 */
2036static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2037{
2038	struct mm_struct *mm = rmap_item->mm;
2039	struct rmap_item *tree_rmap_item;
2040	struct page *tree_page = NULL;
2041	struct stable_node *stable_node;
2042	struct page *kpage;
2043	unsigned int checksum;
2044	int err;
2045	bool max_page_sharing_bypass = false;
2046
2047	stable_node = page_stable_node(page);
2048	if (stable_node) {
2049		if (stable_node->head != &migrate_nodes &&
2050		    get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2051		    NUMA(stable_node->nid)) {
2052			stable_node_dup_del(stable_node);
2053			stable_node->head = &migrate_nodes;
2054			list_add(&stable_node->list, stable_node->head);
2055		}
2056		if (stable_node->head != &migrate_nodes &&
2057		    rmap_item->head == stable_node)
2058			return;
2059		/*
2060		 * If it's a KSM fork, allow it to go over the sharing limit
2061		 * without warnings.
2062		 */
2063		if (!is_page_sharing_candidate(stable_node))
2064			max_page_sharing_bypass = true;
2065	}
2066
2067	/* We first start with searching the page inside the stable tree */
2068	kpage = stable_tree_search(page);
2069	if (kpage == page && rmap_item->head == stable_node) {
2070		put_page(kpage);
2071		return;
2072	}
2073
2074	remove_rmap_item_from_tree(rmap_item);
2075
2076	if (kpage) {
2077		if (PTR_ERR(kpage) == -EBUSY)
2078			return;
2079
2080		err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2081		if (!err) {
2082			/*
2083			 * The page was successfully merged:
2084			 * add its rmap_item to the stable tree.
2085			 */
2086			lock_page(kpage);
2087			stable_tree_append(rmap_item, page_stable_node(kpage),
2088					   max_page_sharing_bypass);
2089			unlock_page(kpage);
2090		}
2091		put_page(kpage);
2092		return;
2093	}
2094
2095	/*
2096	 * If the hash value of the page has changed from the last time
2097	 * we calculated it, this page is changing frequently: therefore we
2098	 * don't want to insert it in the unstable tree, and we don't want
2099	 * to waste our time searching for something identical to it there.
2100	 */
2101	checksum = calc_checksum(page);
2102	if (rmap_item->oldchecksum != checksum) {
2103		rmap_item->oldchecksum = checksum;
2104		return;
2105	}
2106
2107	/*
2108	 * Same checksum as an empty page. We attempt to merge it with the
2109	 * appropriate zero page if the user enabled this via sysfs.
2110	 */
2111	if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2112		struct vm_area_struct *vma;
2113
2114		mmap_read_lock(mm);
2115		vma = find_mergeable_vma(mm, rmap_item->address);
2116		if (vma) {
2117			err = try_to_merge_one_page(vma, page,
2118					ZERO_PAGE(rmap_item->address));
2119		} else {
2120			/*
2121			 * If the vma is out of date, we do not need to
2122			 * continue.
2123			 */
2124			err = 0;
2125		}
2126		mmap_read_unlock(mm);
2127		/*
2128		 * In case of failure, the page was not really empty, so we
2129		 * need to continue. Otherwise we're done.
2130		 */
2131		if (!err)
2132			return;
2133	}
2134	tree_rmap_item =
2135		unstable_tree_search_insert(rmap_item, page, &tree_page);
2136	if (tree_rmap_item) {
2137		bool split;
2138
2139		kpage = try_to_merge_two_pages(rmap_item, page,
2140						tree_rmap_item, tree_page);
 
2141		/*
2142		 * If both pages we tried to merge belong to the same compound
2143		 * page, then we actually ended up increasing the reference
2144		 * count of the same compound page twice, and split_huge_page
2145		 * failed.
2146		 * Here we set a flag if that happened, and we use it later to
2147		 * try split_huge_page again. Since we call put_page right
2148		 * afterwards, the reference count will be correct and
2149		 * split_huge_page should succeed.
2150		 */
2151		split = PageTransCompound(page)
2152			&& compound_head(page) == compound_head(tree_page);
2153		put_page(tree_page);
2154		if (kpage) {
2155			/*
2156			 * The pages were successfully merged: insert new
2157			 * node in the stable tree and add both rmap_items.
2158			 */
2159			lock_page(kpage);
2160			stable_node = stable_tree_insert(kpage);
2161			if (stable_node) {
2162				stable_tree_append(tree_rmap_item, stable_node,
2163						   false);
2164				stable_tree_append(rmap_item, stable_node,
2165						   false);
2166			}
2167			unlock_page(kpage);
2168
2169			/*
2170			 * If we fail to insert the page into the stable tree,
2171			 * we will have 2 virtual addresses that are pointing
2172			 * to a ksm page left outside the stable tree,
2173			 * in which case we need to break_cow on both.
2174			 */
2175			if (!stable_node) {
2176				break_cow(tree_rmap_item);
2177				break_cow(rmap_item);
2178			}
2179		} else if (split) {
2180			/*
2181			 * We are here if we tried to merge two pages and
2182			 * failed because they both belonged to the same
2183			 * compound page. We will split the page now, but no
2184			 * merging will take place.
2185			 * We do not want to add the cost of a full lock; if
2186			 * the page is locked, it is better to skip it and
2187			 * perhaps try again later.
2188			 */
2189			if (!trylock_page(page))
2190				return;
2191			split_huge_page(page);
2192			unlock_page(page);
2193		}
2194	}
2195}
2196
2197static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2198					    struct rmap_item **rmap_list,
2199					    unsigned long addr)
2200{
2201	struct rmap_item *rmap_item;
2202
2203	while (*rmap_list) {
2204		rmap_item = *rmap_list;
2205		if ((rmap_item->address & PAGE_MASK) == addr)
2206			return rmap_item;
2207		if (rmap_item->address > addr)
2208			break;
2209		*rmap_list = rmap_item->rmap_list;
2210		remove_rmap_item_from_tree(rmap_item);
2211		free_rmap_item(rmap_item);
2212	}
2213
2214	rmap_item = alloc_rmap_item();
2215	if (rmap_item) {
2216		/* It has already been zeroed */
2217		rmap_item->mm = mm_slot->mm;
2218		rmap_item->address = addr;
2219		rmap_item->rmap_list = *rmap_list;
2220		*rmap_list = rmap_item;
2221	}
2222	return rmap_item;
2223}
2224
2225static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2226{
2227	struct mm_struct *mm;
2228	struct mm_slot *slot;
2229	struct vm_area_struct *vma;
2230	struct rmap_item *rmap_item;
2231	int nid;
2232
2233	if (list_empty(&ksm_mm_head.mm_list))
2234		return NULL;
2235
2236	slot = ksm_scan.mm_slot;
2237	if (slot == &ksm_mm_head) {
2238		/*
2239		 * A number of pages can hang around indefinitely on per-cpu
2240		 * pagevecs, raised page count preventing write_protect_page
2241		 * from merging them.  Though it doesn't really matter much,
2242		 * it is puzzling to see some stuck in pages_volatile until
2243		 * other activity jostles them out, and they also prevented
2244		 * LTP's KSM test from succeeding deterministically; so drain
2245		 * them here (here rather than on entry to ksm_do_scan(),
2246		 * so we don't IPI too often when pages_to_scan is set low).
2247		 */
2248		lru_add_drain_all();
2249
2250		/*
2251		 * Whereas stale stable_nodes on the stable_tree itself
2252		 * get pruned in the regular course of stable_tree_search(),
2253		 * those moved out to the migrate_nodes list can accumulate:
2254		 * so prune them once before each full scan.
2255		 */
2256		if (!ksm_merge_across_nodes) {
2257			struct stable_node *stable_node, *next;
2258			struct page *page;
2259
2260			list_for_each_entry_safe(stable_node, next,
2261						 &migrate_nodes, list) {
2262				page = get_ksm_page(stable_node,
2263						    GET_KSM_PAGE_NOLOCK);
2264				if (page)
2265					put_page(page);
2266				cond_resched();
2267			}
2268		}
2269
2270		for (nid = 0; nid < ksm_nr_node_ids; nid++)
2271			root_unstable_tree[nid] = RB_ROOT;
2272
2273		spin_lock(&ksm_mmlist_lock);
2274		slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2275		ksm_scan.mm_slot = slot;
2276		spin_unlock(&ksm_mmlist_lock);
2277		/*
2278		 * Although we tested list_empty() above, a racing __ksm_exit
2279		 * of the last mm on the list may have removed it since then.
2280		 */
2281		if (slot == &ksm_mm_head)
2282			return NULL;
2283next_mm:
2284		ksm_scan.address = 0;
2285		ksm_scan.rmap_list = &slot->rmap_list;
2286	}
2287
2288	mm = slot->mm;
2289	mmap_read_lock(mm);
2290	if (ksm_test_exit(mm))
2291		vma = NULL;
2292	else
2293		vma = find_vma(mm, ksm_scan.address);
2294
2295	for (; vma; vma = vma->vm_next) {
2296		if (!(vma->vm_flags & VM_MERGEABLE))
2297			continue;
2298		if (ksm_scan.address < vma->vm_start)
2299			ksm_scan.address = vma->vm_start;
2300		if (!vma->anon_vma)
2301			ksm_scan.address = vma->vm_end;
2302
2303		while (ksm_scan.address < vma->vm_end) {
2304			if (ksm_test_exit(mm))
2305				break;
2306			*page = follow_page(vma, ksm_scan.address, FOLL_GET);
2307			if (IS_ERR_OR_NULL(*page)) {
2308				ksm_scan.address += PAGE_SIZE;
2309				cond_resched();
2310				continue;
2311			}
2312			if (PageAnon(*page)) {
 
2313				flush_anon_page(vma, *page, ksm_scan.address);
2314				flush_dcache_page(*page);
2315				rmap_item = get_next_rmap_item(slot,
2316					ksm_scan.rmap_list, ksm_scan.address);
2317				if (rmap_item) {
2318					ksm_scan.rmap_list =
2319							&rmap_item->rmap_list;
2320					ksm_scan.address += PAGE_SIZE;
2321				} else
2322					put_page(*page);
2323				mmap_read_unlock(mm);
2324				return rmap_item;
2325			}
2326			put_page(*page);
2327			ksm_scan.address += PAGE_SIZE;
2328			cond_resched();
2329		}
2330	}
2331
2332	if (ksm_test_exit(mm)) {
2333		ksm_scan.address = 0;
2334		ksm_scan.rmap_list = &slot->rmap_list;
2335	}
2336	/*
2337	 * Nuke all the rmap_items that are above this current rmap:
2338	 * because there were no VM_MERGEABLE vmas with such addresses.
2339	 */
2340	remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
2341
2342	spin_lock(&ksm_mmlist_lock);
2343	ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2344						struct mm_slot, mm_list);
2345	if (ksm_scan.address == 0) {
2346		/*
2347		 * We've completed a full scan of all vmas, holding mmap_lock
2348		 * throughout, and found no VM_MERGEABLE: so do the same as
2349		 * __ksm_exit does to remove this mm from all our lists now.
2350		 * This applies either when cleaning up after __ksm_exit
2351		 * (but beware: we can reach here even before __ksm_exit),
2352		 * or when all VM_MERGEABLE areas have been unmapped (and
2353		 * mmap_lock then protects against race with MADV_MERGEABLE).
2354		 */
2355		hash_del(&slot->link);
2356		list_del(&slot->mm_list);
2357		spin_unlock(&ksm_mmlist_lock);
2358
2359		free_mm_slot(slot);
2360		clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2361		mmap_read_unlock(mm);
2362		mmdrop(mm);
2363	} else {
2364		mmap_read_unlock(mm);
2365		/*
2366		 * mmap_read_unlock(mm) first because after
2367		 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2368		 * already have been freed under us by __ksm_exit()
2369		 * because the "mm_slot" is still hashed and
2370		 * ksm_scan.mm_slot doesn't point to it anymore.
2371		 */
2372		spin_unlock(&ksm_mmlist_lock);
 
2373	}
2374
2375	/* Repeat until we've completed scanning the whole list */
2376	slot = ksm_scan.mm_slot;
2377	if (slot != &ksm_mm_head)
2378		goto next_mm;
2379
2380	ksm_scan.seqnr++;
2381	return NULL;
2382}
2383
2384/**
2385 * ksm_do_scan  - the ksm scanner main worker function.
2386 * @scan_npages:  number of pages we want to scan before we return.
2387 */
2388static void ksm_do_scan(unsigned int scan_npages)
2389{
2390	struct rmap_item *rmap_item;
2391	struct page *page;
2392
2393	while (scan_npages-- && likely(!freezing(current))) {
2394		cond_resched();
2395		rmap_item = scan_get_next_rmap_item(&page);
2396		if (!rmap_item)
2397			return;
2398		cmp_and_merge_page(page, rmap_item);
 
2399		put_page(page);
2400	}
2401}
2402
2403static int ksmd_should_run(void)
2404{
2405	return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2406}
2407
2408static int ksm_scan_thread(void *nothing)
2409{
2410	unsigned int sleep_ms;
2411
2412	set_freezable();
2413	set_user_nice(current, 5);
2414
2415	while (!kthread_should_stop()) {
2416		mutex_lock(&ksm_thread_mutex);
2417		wait_while_offlining();
2418		if (ksmd_should_run())
2419			ksm_do_scan(ksm_thread_pages_to_scan);
2420		mutex_unlock(&ksm_thread_mutex);
2421
2422		try_to_freeze();
2423
2424		if (ksmd_should_run()) {
2425			sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs);
2426			wait_event_interruptible_timeout(ksm_iter_wait,
2427				sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs),
2428				msecs_to_jiffies(sleep_ms));
2429		} else {
2430			wait_event_freezable(ksm_thread_wait,
2431				ksmd_should_run() || kthread_should_stop());
2432		}
2433	}
2434	return 0;
2435}
2436
2437int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2438		unsigned long end, int advice, unsigned long *vm_flags)
2439{
2440	struct mm_struct *mm = vma->vm_mm;
2441	int err;
2442
2443	switch (advice) {
2444	case MADV_MERGEABLE:
2445		/*
2446		 * Be somewhat over-protective for now!
2447		 */
2448		if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
2449				 VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
2450				 VM_HUGETLB | VM_MIXEDMAP))
 
2451			return 0;		/* just ignore the advice */
2452
2453		if (vma_is_dax(vma))
2454			return 0;
2455
2456#ifdef VM_SAO
2457		if (*vm_flags & VM_SAO)
2458			return 0;
2459#endif
2460#ifdef VM_SPARC_ADI
2461		if (*vm_flags & VM_SPARC_ADI)
2462			return 0;
2463#endif
2464
2465		if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2466			err = __ksm_enter(mm);
2467			if (err)
2468				return err;
2469		}
2470
2471		*vm_flags |= VM_MERGEABLE;
2472		break;
2473
2474	case MADV_UNMERGEABLE:
2475		if (!(*vm_flags & VM_MERGEABLE))
2476			return 0;		/* just ignore the advice */
2477
2478		if (vma->anon_vma) {
2479			err = unmerge_ksm_pages(vma, start, end);
2480			if (err)
2481				return err;
2482		}
2483
2484		*vm_flags &= ~VM_MERGEABLE;
2485		break;
2486	}
2487
2488	return 0;
2489}
2490EXPORT_SYMBOL_GPL(ksm_madvise);
2491
2492int __ksm_enter(struct mm_struct *mm)
2493{
2494	struct mm_slot *mm_slot;
2495	int needs_wakeup;
2496
2497	mm_slot = alloc_mm_slot();
2498	if (!mm_slot)
2499		return -ENOMEM;
2500
2501	/* Check ksm_run too?  Would need tighter locking */
2502	needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2503
2504	spin_lock(&ksm_mmlist_lock);
2505	insert_to_mm_slots_hash(mm, mm_slot);
2506	/*
2507	 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2508	 * insert just behind the scanning cursor, to let the area settle
2509	 * down a little; when fork is followed by immediate exec, we don't
2510	 * want ksmd to waste time setting up and tearing down an rmap_list.
2511	 *
2512	 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2513	 * scanning cursor, otherwise KSM pages in newly forked mms will be
2514	 * missed: then we might as well insert at the end of the list.
2515	 */
2516	if (ksm_run & KSM_RUN_UNMERGE)
2517		list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2518	else
2519		list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2520	spin_unlock(&ksm_mmlist_lock);
2521
2522	set_bit(MMF_VM_MERGEABLE, &mm->flags);
2523	mmgrab(mm);
2524
2525	if (needs_wakeup)
2526		wake_up_interruptible(&ksm_thread_wait);
2527
2528	return 0;
2529}
2530
2531void __ksm_exit(struct mm_struct *mm)
2532{
2533	struct mm_slot *mm_slot;
2534	int easy_to_free = 0;
2535
2536	/*
2537	 * This process is exiting: if it's straightforward (as is the
2538	 * case when ksmd was never running), free mm_slot immediately.
2539	 * But if it's at the cursor or has rmap_items linked to it, use
2540	 * mmap_lock to synchronize with any break_cows before pagetables
2541	 * are freed, and leave the mm_slot on the list for ksmd to free.
2542	 * Beware: ksm may already have noticed it exiting and freed the slot.
2543	 */
2544
2545	spin_lock(&ksm_mmlist_lock);
2546	mm_slot = get_mm_slot(mm);
2547	if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2548		if (!mm_slot->rmap_list) {
2549			hash_del(&mm_slot->link);
2550			list_del(&mm_slot->mm_list);
2551			easy_to_free = 1;
2552		} else {
2553			list_move(&mm_slot->mm_list,
2554				  &ksm_scan.mm_slot->mm_list);
2555		}
2556	}
2557	spin_unlock(&ksm_mmlist_lock);
2558
2559	if (easy_to_free) {
2560		free_mm_slot(mm_slot);
2561		clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2562		mmdrop(mm);
2563	} else if (mm_slot) {
2564		mmap_write_lock(mm);
2565		mmap_write_unlock(mm);
2566	}
2567}
2568
2569struct page *ksm_might_need_to_copy(struct page *page,
2570			struct vm_area_struct *vma, unsigned long address)
2571{
2572	struct anon_vma *anon_vma = page_anon_vma(page);
2573	struct page *new_page;
2574
2575	if (PageKsm(page)) {
2576		if (page_stable_node(page) &&
2577		    !(ksm_run & KSM_RUN_UNMERGE))
2578			return page;	/* no need to copy it */
2579	} else if (!anon_vma) {
2580		return page;		/* no need to copy it */
2581	} else if (anon_vma->root == vma->anon_vma->root &&
2582		 page->index == linear_page_index(vma, address)) {
2583		return page;		/* still no need to copy it */
2584	}
2585	if (!PageUptodate(page))
2586		return page;		/* let do_swap_page report the error */
2587
2588	new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2589	if (new_page && mem_cgroup_charge(new_page, vma->vm_mm, GFP_KERNEL)) {
2590		put_page(new_page);
2591		new_page = NULL;
2592	}
2593	if (new_page) {
2594		copy_user_highpage(new_page, page, address, vma);
2595
2596		SetPageDirty(new_page);
2597		__SetPageUptodate(new_page);
2598		__SetPageLocked(new_page);
 
 
 
 
 
 
2599	}
2600
2601	return new_page;
2602}
2603
2604void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
 
2605{
2606	struct stable_node *stable_node;
2607	struct rmap_item *rmap_item;
 
 
 
2608	int search_new_forks = 0;
2609
2610	VM_BUG_ON_PAGE(!PageKsm(page), page);
2611
2612	/*
2613	 * Rely on the page lock to protect against concurrent modifications
2614	 * to that page's node of the stable tree.
2615	 */
2616	VM_BUG_ON_PAGE(!PageLocked(page), page);
2617
2618	stable_node = page_stable_node(page);
2619	if (!stable_node)
2620		return;
2621again:
2622	hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2623		struct anon_vma *anon_vma = rmap_item->anon_vma;
2624		struct anon_vma_chain *vmac;
2625		struct vm_area_struct *vma;
2626
2627		cond_resched();
2628		anon_vma_lock_read(anon_vma);
2629		anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2630					       0, ULONG_MAX) {
2631			unsigned long addr;
2632
2633			cond_resched();
2634			vma = vmac->vma;
2635
2636			/* Ignore the stable/unstable/sqnr flags */
2637			addr = rmap_item->address & ~KSM_FLAG_MASK;
2638
2639			if (addr < vma->vm_start || addr >= vma->vm_end)
2640				continue;
2641			/*
2642			 * Initially we examine only the vma which covers this
2643			 * rmap_item; but later, if there is still work to do,
2644			 * we examine covering vmas in other mms: in case they
2645			 * were forked from the original since ksmd passed.
2646			 */
2647			if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2648				continue;
2649
2650			if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2651				continue;
2652
2653			if (!rwc->rmap_one(page, vma, addr, rwc->arg)) {
2654				anon_vma_unlock_read(anon_vma);
2655				return;
2656			}
2657			if (rwc->done && rwc->done(page)) {
2658				anon_vma_unlock_read(anon_vma);
2659				return;
2660			}
2661		}
2662		anon_vma_unlock_read(anon_vma);
 
 
2663	}
2664	if (!search_new_forks++)
2665		goto again;
 
 
2666}
2667
2668#ifdef CONFIG_MIGRATION
2669void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2670{
2671	struct stable_node *stable_node;
 
 
 
 
2672
2673	VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2674	VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2675	VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2676
2677	stable_node = page_stable_node(newpage);
2678	if (stable_node) {
2679		VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2680		stable_node->kpfn = page_to_pfn(newpage);
2681		/*
2682		 * newpage->mapping was set in advance; now we need smp_wmb()
2683		 * to make sure that the new stable_node->kpfn is visible
2684		 * to get_ksm_page() before it can see that oldpage->mapping
2685		 * has gone stale (or that PageSwapCache has been cleared).
2686		 */
2687		smp_wmb();
2688		set_page_stable_node(oldpage, NULL);
2689	}
2690}
2691#endif /* CONFIG_MIGRATION */
2692
2693#ifdef CONFIG_MEMORY_HOTREMOVE
2694static void wait_while_offlining(void)
2695{
2696	while (ksm_run & KSM_RUN_OFFLINE) {
2697		mutex_unlock(&ksm_thread_mutex);
2698		wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2699			    TASK_UNINTERRUPTIBLE);
2700		mutex_lock(&ksm_thread_mutex);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2701	}
 
 
 
 
2702}
2703
2704static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2705					 unsigned long start_pfn,
2706					 unsigned long end_pfn)
2707{
2708	if (stable_node->kpfn >= start_pfn &&
2709	    stable_node->kpfn < end_pfn) {
2710		/*
2711		 * Don't get_ksm_page, page has already gone:
2712		 * which is why we keep kpfn instead of page*
2713		 */
2714		remove_node_from_stable_tree(stable_node);
2715		return true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2716	}
2717	return false;
 
 
 
2718}
2719
2720static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2721					   unsigned long start_pfn,
2722					   unsigned long end_pfn,
2723					   struct rb_root *root)
2724{
2725	struct stable_node *dup;
2726	struct hlist_node *hlist_safe;
2727
2728	if (!is_stable_node_chain(stable_node)) {
2729		VM_BUG_ON(is_stable_node_dup(stable_node));
2730		return stable_node_dup_remove_range(stable_node, start_pfn,
2731						    end_pfn);
2732	}
2733
2734	hlist_for_each_entry_safe(dup, hlist_safe,
2735				  &stable_node->hlist, hlist_dup) {
2736		VM_BUG_ON(!is_stable_node_dup(dup));
2737		stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2738	}
2739	if (hlist_empty(&stable_node->hlist)) {
2740		free_stable_node_chain(stable_node, root);
2741		return true; /* notify caller that tree was rebalanced */
2742	} else
2743		return false;
2744}
 
2745
2746static void ksm_check_stable_tree(unsigned long start_pfn,
2747				  unsigned long end_pfn)
 
2748{
2749	struct stable_node *stable_node, *next;
2750	struct rb_node *node;
2751	int nid;
2752
2753	for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2754		node = rb_first(root_stable_tree + nid);
2755		while (node) {
2756			stable_node = rb_entry(node, struct stable_node, node);
2757			if (stable_node_chain_remove_range(stable_node,
2758							   start_pfn, end_pfn,
2759							   root_stable_tree +
2760							   nid))
2761				node = rb_first(root_stable_tree + nid);
2762			else
2763				node = rb_next(node);
2764			cond_resched();
2765		}
2766	}
2767	list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2768		if (stable_node->kpfn >= start_pfn &&
2769		    stable_node->kpfn < end_pfn)
2770			remove_node_from_stable_tree(stable_node);
2771		cond_resched();
2772	}
 
2773}
2774
2775static int ksm_memory_callback(struct notifier_block *self,
2776			       unsigned long action, void *arg)
2777{
2778	struct memory_notify *mn = arg;
 
2779
2780	switch (action) {
2781	case MEM_GOING_OFFLINE:
2782		/*
2783		 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2784		 * and remove_all_stable_nodes() while memory is going offline:
2785		 * it is unsafe for them to touch the stable tree at this time.
2786		 * But unmerge_ksm_pages(), rmap lookups and other entry points
2787		 * which do not need the ksm_thread_mutex are all safe.
 
 
2788		 */
2789		mutex_lock(&ksm_thread_mutex);
2790		ksm_run |= KSM_RUN_OFFLINE;
2791		mutex_unlock(&ksm_thread_mutex);
2792		break;
2793
2794	case MEM_OFFLINE:
2795		/*
2796		 * Most of the work is done by page migration; but there might
2797		 * be a few stable_nodes left over, still pointing to struct
2798		 * pages which have been offlined: prune those from the tree,
2799		 * otherwise get_ksm_page() might later try to access a
2800		 * non-existent struct page.
2801		 */
2802		ksm_check_stable_tree(mn->start_pfn,
2803				      mn->start_pfn + mn->nr_pages);
2804		fallthrough;
 
 
2805	case MEM_CANCEL_OFFLINE:
2806		mutex_lock(&ksm_thread_mutex);
2807		ksm_run &= ~KSM_RUN_OFFLINE;
2808		mutex_unlock(&ksm_thread_mutex);
2809
2810		smp_mb();	/* wake_up_bit advises this */
2811		wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2812		break;
2813	}
2814	return NOTIFY_OK;
2815}
2816#else
2817static void wait_while_offlining(void)
2818{
2819}
2820#endif /* CONFIG_MEMORY_HOTREMOVE */
2821
2822#ifdef CONFIG_SYSFS
2823/*
2824 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2825 */
2826
2827#define KSM_ATTR_RO(_name) \
2828	static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2829#define KSM_ATTR(_name) \
2830	static struct kobj_attribute _name##_attr = \
2831		__ATTR(_name, 0644, _name##_show, _name##_store)
2832
2833static ssize_t sleep_millisecs_show(struct kobject *kobj,
2834				    struct kobj_attribute *attr, char *buf)
2835{
2836	return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2837}
2838
2839static ssize_t sleep_millisecs_store(struct kobject *kobj,
2840				     struct kobj_attribute *attr,
2841				     const char *buf, size_t count)
2842{
2843	unsigned long msecs;
2844	int err;
2845
2846	err = kstrtoul(buf, 10, &msecs);
2847	if (err || msecs > UINT_MAX)
2848		return -EINVAL;
2849
2850	ksm_thread_sleep_millisecs = msecs;
2851	wake_up_interruptible(&ksm_iter_wait);
2852
2853	return count;
2854}
2855KSM_ATTR(sleep_millisecs);
2856
2857static ssize_t pages_to_scan_show(struct kobject *kobj,
2858				  struct kobj_attribute *attr, char *buf)
2859{
2860	return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2861}
2862
2863static ssize_t pages_to_scan_store(struct kobject *kobj,
2864				   struct kobj_attribute *attr,
2865				   const char *buf, size_t count)
2866{
2867	int err;
2868	unsigned long nr_pages;
2869
2870	err = kstrtoul(buf, 10, &nr_pages);
2871	if (err || nr_pages > UINT_MAX)
2872		return -EINVAL;
2873
2874	ksm_thread_pages_to_scan = nr_pages;
2875
2876	return count;
2877}
2878KSM_ATTR(pages_to_scan);
2879
2880static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2881			char *buf)
2882{
2883	return sprintf(buf, "%lu\n", ksm_run);
2884}
2885
2886static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2887			 const char *buf, size_t count)
2888{
2889	int err;
2890	unsigned long flags;
2891
2892	err = kstrtoul(buf, 10, &flags);
2893	if (err || flags > UINT_MAX)
2894		return -EINVAL;
2895	if (flags > KSM_RUN_UNMERGE)
2896		return -EINVAL;
2897
2898	/*
2899	 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2900	 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2901	 * breaking COW to free the pages_shared (but leaves mm_slots
2902	 * on the list for when ksmd may be set running again).
2903	 */
2904
2905	mutex_lock(&ksm_thread_mutex);
2906	wait_while_offlining();
2907	if (ksm_run != flags) {
2908		ksm_run = flags;
2909		if (flags & KSM_RUN_UNMERGE) {
2910			set_current_oom_origin();
 
 
2911			err = unmerge_and_remove_all_rmap_items();
2912			clear_current_oom_origin();
 
2913			if (err) {
2914				ksm_run = KSM_RUN_STOP;
2915				count = err;
2916			}
2917		}
2918	}
2919	mutex_unlock(&ksm_thread_mutex);
2920
2921	if (flags & KSM_RUN_MERGE)
2922		wake_up_interruptible(&ksm_thread_wait);
2923
2924	return count;
2925}
2926KSM_ATTR(run);
2927
2928#ifdef CONFIG_NUMA
2929static ssize_t merge_across_nodes_show(struct kobject *kobj,
2930				struct kobj_attribute *attr, char *buf)
2931{
2932	return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2933}
2934
2935static ssize_t merge_across_nodes_store(struct kobject *kobj,
2936				   struct kobj_attribute *attr,
2937				   const char *buf, size_t count)
2938{
2939	int err;
2940	unsigned long knob;
2941
2942	err = kstrtoul(buf, 10, &knob);
2943	if (err)
2944		return err;
2945	if (knob > 1)
2946		return -EINVAL;
2947
2948	mutex_lock(&ksm_thread_mutex);
2949	wait_while_offlining();
2950	if (ksm_merge_across_nodes != knob) {
2951		if (ksm_pages_shared || remove_all_stable_nodes())
2952			err = -EBUSY;
2953		else if (root_stable_tree == one_stable_tree) {
2954			struct rb_root *buf;
2955			/*
2956			 * This is the first time that we switch away from the
2957			 * default of merging across nodes: must now allocate
2958			 * a buffer to hold as many roots as may be needed.
2959			 * Allocate stable and unstable together:
2960			 * MAXSMP NODES_SHIFT 10 will use 16kB.
2961			 */
2962			buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2963				      GFP_KERNEL);
2964			/* Let us assume that RB_ROOT is NULL is zero */
2965			if (!buf)
2966				err = -ENOMEM;
2967			else {
2968				root_stable_tree = buf;
2969				root_unstable_tree = buf + nr_node_ids;
2970				/* Stable tree is empty but not the unstable */
2971				root_unstable_tree[0] = one_unstable_tree[0];
2972			}
2973		}
2974		if (!err) {
2975			ksm_merge_across_nodes = knob;
2976			ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2977		}
2978	}
2979	mutex_unlock(&ksm_thread_mutex);
2980
2981	return err ? err : count;
2982}
2983KSM_ATTR(merge_across_nodes);
2984#endif
2985
2986static ssize_t use_zero_pages_show(struct kobject *kobj,
2987				struct kobj_attribute *attr, char *buf)
2988{
2989	return sprintf(buf, "%u\n", ksm_use_zero_pages);
2990}
2991static ssize_t use_zero_pages_store(struct kobject *kobj,
2992				   struct kobj_attribute *attr,
2993				   const char *buf, size_t count)
2994{
2995	int err;
2996	bool value;
2997
2998	err = kstrtobool(buf, &value);
2999	if (err)
3000		return -EINVAL;
3001
3002	ksm_use_zero_pages = value;
3003
3004	return count;
3005}
3006KSM_ATTR(use_zero_pages);
3007
3008static ssize_t max_page_sharing_show(struct kobject *kobj,
3009				     struct kobj_attribute *attr, char *buf)
3010{
3011	return sprintf(buf, "%u\n", ksm_max_page_sharing);
3012}
3013
3014static ssize_t max_page_sharing_store(struct kobject *kobj,
3015				      struct kobj_attribute *attr,
3016				      const char *buf, size_t count)
3017{
3018	int err;
3019	int knob;
3020
3021	err = kstrtoint(buf, 10, &knob);
3022	if (err)
3023		return err;
3024	/*
3025	 * When a KSM page is created it is shared by 2 mappings. This
3026	 * being a signed comparison, it implicitly verifies it's not
3027	 * negative.
3028	 */
3029	if (knob < 2)
3030		return -EINVAL;
3031
3032	if (READ_ONCE(ksm_max_page_sharing) == knob)
3033		return count;
3034
3035	mutex_lock(&ksm_thread_mutex);
3036	wait_while_offlining();
3037	if (ksm_max_page_sharing != knob) {
3038		if (ksm_pages_shared || remove_all_stable_nodes())
3039			err = -EBUSY;
3040		else
3041			ksm_max_page_sharing = knob;
3042	}
3043	mutex_unlock(&ksm_thread_mutex);
3044
3045	return err ? err : count;
3046}
3047KSM_ATTR(max_page_sharing);
3048
3049static ssize_t pages_shared_show(struct kobject *kobj,
3050				 struct kobj_attribute *attr, char *buf)
3051{
3052	return sprintf(buf, "%lu\n", ksm_pages_shared);
3053}
3054KSM_ATTR_RO(pages_shared);
3055
3056static ssize_t pages_sharing_show(struct kobject *kobj,
3057				  struct kobj_attribute *attr, char *buf)
3058{
3059	return sprintf(buf, "%lu\n", ksm_pages_sharing);
3060}
3061KSM_ATTR_RO(pages_sharing);
3062
3063static ssize_t pages_unshared_show(struct kobject *kobj,
3064				   struct kobj_attribute *attr, char *buf)
3065{
3066	return sprintf(buf, "%lu\n", ksm_pages_unshared);
3067}
3068KSM_ATTR_RO(pages_unshared);
3069
3070static ssize_t pages_volatile_show(struct kobject *kobj,
3071				   struct kobj_attribute *attr, char *buf)
3072{
3073	long ksm_pages_volatile;
3074
3075	ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3076				- ksm_pages_sharing - ksm_pages_unshared;
3077	/*
3078	 * It was not worth any locking to calculate that statistic,
3079	 * but it might therefore sometimes be negative: conceal that.
3080	 */
3081	if (ksm_pages_volatile < 0)
3082		ksm_pages_volatile = 0;
3083	return sprintf(buf, "%ld\n", ksm_pages_volatile);
3084}
3085KSM_ATTR_RO(pages_volatile);
3086
3087static ssize_t stable_node_dups_show(struct kobject *kobj,
3088				     struct kobj_attribute *attr, char *buf)
3089{
3090	return sprintf(buf, "%lu\n", ksm_stable_node_dups);
3091}
3092KSM_ATTR_RO(stable_node_dups);
3093
3094static ssize_t stable_node_chains_show(struct kobject *kobj,
3095				       struct kobj_attribute *attr, char *buf)
3096{
3097	return sprintf(buf, "%lu\n", ksm_stable_node_chains);
3098}
3099KSM_ATTR_RO(stable_node_chains);
3100
3101static ssize_t
3102stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3103					struct kobj_attribute *attr,
3104					char *buf)
3105{
3106	return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3107}
3108
3109static ssize_t
3110stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3111					 struct kobj_attribute *attr,
3112					 const char *buf, size_t count)
3113{
3114	unsigned long msecs;
3115	int err;
3116
3117	err = kstrtoul(buf, 10, &msecs);
3118	if (err || msecs > UINT_MAX)
3119		return -EINVAL;
3120
3121	ksm_stable_node_chains_prune_millisecs = msecs;
3122
3123	return count;
3124}
3125KSM_ATTR(stable_node_chains_prune_millisecs);
3126
3127static ssize_t full_scans_show(struct kobject *kobj,
3128			       struct kobj_attribute *attr, char *buf)
3129{
3130	return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3131}
3132KSM_ATTR_RO(full_scans);
3133
3134static struct attribute *ksm_attrs[] = {
3135	&sleep_millisecs_attr.attr,
3136	&pages_to_scan_attr.attr,
3137	&run_attr.attr,
3138	&pages_shared_attr.attr,
3139	&pages_sharing_attr.attr,
3140	&pages_unshared_attr.attr,
3141	&pages_volatile_attr.attr,
3142	&full_scans_attr.attr,
3143#ifdef CONFIG_NUMA
3144	&merge_across_nodes_attr.attr,
3145#endif
3146	&max_page_sharing_attr.attr,
3147	&stable_node_chains_attr.attr,
3148	&stable_node_dups_attr.attr,
3149	&stable_node_chains_prune_millisecs_attr.attr,
3150	&use_zero_pages_attr.attr,
3151	NULL,
3152};
3153
3154static const struct attribute_group ksm_attr_group = {
3155	.attrs = ksm_attrs,
3156	.name = "ksm",
3157};
3158#endif /* CONFIG_SYSFS */
3159
3160static int __init ksm_init(void)
3161{
3162	struct task_struct *ksm_thread;
3163	int err;
3164
3165	/* The correct value depends on page size and endianness */
3166	zero_checksum = calc_checksum(ZERO_PAGE(0));
3167	/* Default to false for backwards compatibility */
3168	ksm_use_zero_pages = false;
3169
3170	err = ksm_slab_init();
3171	if (err)
3172		goto out;
3173
3174	ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3175	if (IS_ERR(ksm_thread)) {
3176		pr_err("ksm: creating kthread failed\n");
3177		err = PTR_ERR(ksm_thread);
3178		goto out_free;
3179	}
3180
3181#ifdef CONFIG_SYSFS
3182	err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3183	if (err) {
3184		pr_err("ksm: register sysfs failed\n");
3185		kthread_stop(ksm_thread);
3186		goto out_free;
3187	}
3188#else
3189	ksm_run = KSM_RUN_MERGE;	/* no way for user to start it */
3190
3191#endif /* CONFIG_SYSFS */
3192
3193#ifdef CONFIG_MEMORY_HOTREMOVE
3194	/* There is no significance to this priority 100 */
 
 
 
3195	hotplug_memory_notifier(ksm_memory_callback, 100);
3196#endif
3197	return 0;
3198
3199out_free:
3200	ksm_slab_free();
3201out:
3202	return err;
3203}
3204subsys_initcall(ksm_init);