1/*
   2 * Memory merging support.
   3 *
   4 * This code enables dynamic sharing of identical pages found in different
   5 * memory areas, even if they are not shared by fork()
   6 *
   7 * Copyright (C) 2008-2009 Red Hat, Inc.
   8 * Authors:
   9 *	Izik Eidus
  10 *	Andrea Arcangeli
  11 *	Chris Wright
  12 *	Hugh Dickins
  13 *
  14 * This work is licensed under the terms of the GNU GPL, version 2.
  15 */
  16
  17#include <linux/errno.h>
  18#include <linux/mm.h>
  19#include <linux/fs.h>
  20#include <linux/mman.h>
  21#include <linux/sched.h>
  22#include <linux/rwsem.h>
  23#include <linux/pagemap.h>
  24#include <linux/rmap.h>
  25#include <linux/spinlock.h>
  26#include <linux/jhash.h>
  27#include <linux/delay.h>
  28#include <linux/kthread.h>
  29#include <linux/wait.h>
  30#include <linux/slab.h>
  31#include <linux/rbtree.h>
  32#include <linux/memory.h>
  33#include <linux/mmu_notifier.h>
  34#include <linux/swap.h>
  35#include <linux/ksm.h>
  36#include <linux/hashtable.h>
  37#include <linux/freezer.h>
  38#include <linux/oom.h>
  39#include <linux/numa.h>
  40
  41#include <asm/tlbflush.h>
  42#include "internal.h"
  43
  44#ifdef CONFIG_NUMA
  45#define NUMA(x)		(x)
  46#define DO_NUMA(x)	do { (x); } while (0)
  47#else
  48#define NUMA(x)		(0)
  49#define DO_NUMA(x)	do { } while (0)
  50#endif
  51
  52/*
  53 * A few notes about the KSM scanning process,
  54 * to make it easier to understand the data structures below:
  55 *
  56 * In order to reduce excessive scanning, KSM sorts the memory pages by their
  57 * contents into a data structure that holds pointers to the pages' locations.
  58 *
  59 * Since the contents of the pages may change at any moment, KSM cannot just
  60 * insert the pages into a normal sorted tree and expect it to find anything.
  61 * Therefore KSM uses two data structures - the stable and the unstable tree.
  62 *
  63 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
  64 * by their contents.  Because each such page is write-protected, searching on
  65 * this tree is fully assured to be working (except when pages are unmapped),
  66 * and therefore this tree is called the stable tree.
  67 *
  68 * In addition to the stable tree, KSM uses a second data structure called the
  69 * unstable tree: this tree holds pointers to pages which have been found to
  70 * be "unchanged for a period of time".  The unstable tree sorts these pages
  71 * by their contents, but since they are not write-protected, KSM cannot rely
  72 * upon the unstable tree to work correctly - the unstable tree is liable to
  73 * be corrupted as its contents are modified, and so it is called unstable.
  74 *
  75 * KSM solves this problem by several techniques:
  76 *
  77 * 1) The unstable tree is flushed every time KSM completes scanning all
  78 *    memory areas, and then the tree is rebuilt again from the beginning.
  79 * 2) KSM will only insert into the unstable tree, pages whose hash value
  80 *    has not changed since the previous scan of all memory areas.
  81 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
  82 *    colors of the nodes and not on their contents, assuring that even when
  83 *    the tree gets "corrupted" it won't get out of balance, so scanning time
  84 *    remains the same (also, searching and inserting nodes in an rbtree uses
  85 *    the same algorithm, so we have no overhead when we flush and rebuild).
  86 * 4) KSM never flushes the stable tree, which means that even if it were to
  87 *    take 10 attempts to find a page in the unstable tree, once it is found,
  88 *    it is secured in the stable tree.  (When we scan a new page, we first
  89 *    compare it against the stable tree, and then against the unstable tree.)
  90 *
  91 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
  92 * stable trees and multiple unstable trees: one of each for each NUMA node.
  93 */
  94
  95/**
  96 * struct mm_slot - ksm information per mm that is being scanned
  97 * @link: link to the mm_slots hash list
  98 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
  99 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
 100 * @mm: the mm that this information is valid for
 101 */
 102struct mm_slot {
 103	struct hlist_node link;
 104	struct list_head mm_list;
 105	struct rmap_item *rmap_list;
 106	struct mm_struct *mm;
 107};
 108
 109/**
 110 * struct ksm_scan - cursor for scanning
 111 * @mm_slot: the current mm_slot we are scanning
 112 * @address: the next address inside that to be scanned
 113 * @rmap_list: link to the next rmap to be scanned in the rmap_list
 114 * @seqnr: count of completed full scans (needed when removing unstable node)
 115 *
 116 * There is only the one ksm_scan instance of this cursor structure.
 117 */
 118struct ksm_scan {
 119	struct mm_slot *mm_slot;
 120	unsigned long address;
 121	struct rmap_item **rmap_list;
 122	unsigned long seqnr;
 123};
 124
 125/**
 126 * struct stable_node - node of the stable rbtree
 127 * @node: rb node of this ksm page in the stable tree
 128 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
 129 * @list: linked into migrate_nodes, pending placement in the proper node tree
 130 * @hlist: hlist head of rmap_items using this ksm page
 131 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
 132 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
 133 */
 134struct stable_node {
 135	union {
 136		struct rb_node node;	/* when node of stable tree */
 137		struct {		/* when listed for migration */
 138			struct list_head *head;
 139			struct list_head list;
 140		};
 141	};
 142	struct hlist_head hlist;
 143	unsigned long kpfn;
 144#ifdef CONFIG_NUMA
 145	int nid;
 146#endif
 147};
 148
 149/**
 150 * struct rmap_item - reverse mapping item for virtual addresses
 151 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
 152 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
 153 * @nid: NUMA node id of unstable tree in which linked (may not match page)
 154 * @mm: the memory structure this rmap_item is pointing into
 155 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
 156 * @oldchecksum: previous checksum of the page at that virtual address
 157 * @node: rb node of this rmap_item in the unstable tree
 158 * @head: pointer to stable_node heading this list in the stable tree
 159 * @hlist: link into hlist of rmap_items hanging off that stable_node
 160 */
 161struct rmap_item {
 162	struct rmap_item *rmap_list;
 163	union {
 164		struct anon_vma *anon_vma;	/* when stable */
 165#ifdef CONFIG_NUMA
 166		int nid;		/* when node of unstable tree */
 167#endif
 168	};
 169	struct mm_struct *mm;
 170	unsigned long address;		/* + low bits used for flags below */
 171	unsigned int oldchecksum;	/* when unstable */
 172	union {
 173		struct rb_node node;	/* when node of unstable tree */
 174		struct {		/* when listed from stable tree */
 175			struct stable_node *head;
 176			struct hlist_node hlist;
 177		};
 178	};
 179};
 180
 181#define SEQNR_MASK	0x0ff	/* low bits of unstable tree seqnr */
 182#define UNSTABLE_FLAG	0x100	/* is a node of the unstable tree */
 183#define STABLE_FLAG	0x200	/* is listed from the stable tree */
 184
 185/* The stable and unstable tree heads */
 186static struct rb_root one_stable_tree[1] = { RB_ROOT };
 187static struct rb_root one_unstable_tree[1] = { RB_ROOT };
 188static struct rb_root *root_stable_tree = one_stable_tree;
 189static struct rb_root *root_unstable_tree = one_unstable_tree;
 190
 191/* Recently migrated nodes of stable tree, pending proper placement */
 192static LIST_HEAD(migrate_nodes);
 193
 194#define MM_SLOTS_HASH_BITS 10
 195static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
 196
 197static struct mm_slot ksm_mm_head = {
 198	.mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
 199};
 200static struct ksm_scan ksm_scan = {
 201	.mm_slot = &ksm_mm_head,
 202};
 203
 204static struct kmem_cache *rmap_item_cache;
 205static struct kmem_cache *stable_node_cache;
 206static struct kmem_cache *mm_slot_cache;
 207
 208/* The number of nodes in the stable tree */
 209static unsigned long ksm_pages_shared;
 210
 211/* The number of page slots additionally sharing those nodes */
 212static unsigned long ksm_pages_sharing;
 213
 214/* The number of nodes in the unstable tree */
 215static unsigned long ksm_pages_unshared;
 216
 217/* The number of rmap_items in use: to calculate pages_volatile */
 218static unsigned long ksm_rmap_items;
 219
 220/* Number of pages ksmd should scan in one batch */
 221static unsigned int ksm_thread_pages_to_scan = 100;
 222
 223/* Milliseconds ksmd should sleep between batches */
 224static unsigned int ksm_thread_sleep_millisecs = 20;
 225
 226#ifdef CONFIG_NUMA
 227/* Zeroed when merging across nodes is not allowed */
 228static unsigned int ksm_merge_across_nodes = 1;
 229static int ksm_nr_node_ids = 1;
 230#else
 231#define ksm_merge_across_nodes	1U
 232#define ksm_nr_node_ids		1
 233#endif
 234
 235#define KSM_RUN_STOP	0
 236#define KSM_RUN_MERGE	1
 237#define KSM_RUN_UNMERGE	2
 238#define KSM_RUN_OFFLINE	4
 239static unsigned long ksm_run = KSM_RUN_STOP;
 240static void wait_while_offlining(void);
 241
 242static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
 243static DEFINE_MUTEX(ksm_thread_mutex);
 244static DEFINE_SPINLOCK(ksm_mmlist_lock);
 245
 246#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
 247		sizeof(struct __struct), __alignof__(struct __struct),\
 248		(__flags), NULL)
 249
 250static int __init ksm_slab_init(void)
 251{
 252	rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
 253	if (!rmap_item_cache)
 254		goto out;
 255
 256	stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
 257	if (!stable_node_cache)
 258		goto out_free1;
 259
 260	mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
 261	if (!mm_slot_cache)
 262		goto out_free2;
 263
 264	return 0;
 265
 266out_free2:
 267	kmem_cache_destroy(stable_node_cache);
 268out_free1:
 269	kmem_cache_destroy(rmap_item_cache);
 270out:
 271	return -ENOMEM;
 272}
 273
 274static void __init ksm_slab_free(void)
 275{
 276	kmem_cache_destroy(mm_slot_cache);
 277	kmem_cache_destroy(stable_node_cache);
 278	kmem_cache_destroy(rmap_item_cache);
 279	mm_slot_cache = NULL;
 280}
 281
 282static inline struct rmap_item *alloc_rmap_item(void)
 283{
 284	struct rmap_item *rmap_item;
 285
 286	rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
 287	if (rmap_item)
 288		ksm_rmap_items++;
 289	return rmap_item;
 290}
 291
 292static inline void free_rmap_item(struct rmap_item *rmap_item)
 293{
 294	ksm_rmap_items--;
 295	rmap_item->mm = NULL;	/* debug safety */
 296	kmem_cache_free(rmap_item_cache, rmap_item);
 297}
 298
 299static inline struct stable_node *alloc_stable_node(void)
 300{
 301	return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
 302}
 303
 304static inline void free_stable_node(struct stable_node *stable_node)
 305{
 306	kmem_cache_free(stable_node_cache, stable_node);
 307}
 308
 309static inline struct mm_slot *alloc_mm_slot(void)
 310{
 311	if (!mm_slot_cache)	/* initialization failed */
 312		return NULL;
 313	return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 314}
 315
 316static inline void free_mm_slot(struct mm_slot *mm_slot)
 317{
 318	kmem_cache_free(mm_slot_cache, mm_slot);
 319}
 320
 321static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 322{
 323	struct mm_slot *slot;
 324
 325	hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
 326		if (slot->mm == mm)
 327			return slot;
 328
 329	return NULL;
 330}
 331
 332static void insert_to_mm_slots_hash(struct mm_struct *mm,
 333				    struct mm_slot *mm_slot)
 334{
 335	mm_slot->mm = mm;
 336	hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
 337}
 338
 339/*
 340 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
 341 * page tables after it has passed through ksm_exit() - which, if necessary,
 342 * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
 343 * a special flag: they can just back out as soon as mm_users goes to zero.
 344 * ksm_test_exit() is used throughout to make this test for exit: in some
 345 * places for correctness, in some places just to avoid unnecessary work.
 346 */
 347static inline bool ksm_test_exit(struct mm_struct *mm)
 348{
 349	return atomic_read(&mm->mm_users) == 0;
 350}
 351
 352/*
 353 * We use break_ksm to break COW on a ksm page: it's a stripped down
 354 *
 355 *	if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
 356 *		put_page(page);
 357 *
 358 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
 359 * in case the application has unmapped and remapped mm,addr meanwhile.
 360 * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
 361 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
 362 *
 363 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
 364 * of the process that owns 'vma'.  We also do not want to enforce
 365 * protection keys here anyway.
 366 */
 367static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
 368{
 369	struct page *page;
 370	int ret = 0;
 371
 372	do {
 373		cond_resched();
 374		page = follow_page(vma, addr,
 375				FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
 376		if (IS_ERR_OR_NULL(page))
 377			break;
 378		if (PageKsm(page))
 379			ret = handle_mm_fault(vma->vm_mm, vma, addr,
 380							FAULT_FLAG_WRITE |
 381							FAULT_FLAG_REMOTE);
 382		else
 383			ret = VM_FAULT_WRITE;
 384		put_page(page);
 385	} while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
 386	/*
 387	 * We must loop because handle_mm_fault() may back out if there's
 388	 * any difficulty e.g. if pte accessed bit gets updated concurrently.
 389	 *
 390	 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
 391	 * COW has been broken, even if the vma does not permit VM_WRITE;
 392	 * but note that a concurrent fault might break PageKsm for us.
 393	 *
 394	 * VM_FAULT_SIGBUS could occur if we race with truncation of the
 395	 * backing file, which also invalidates anonymous pages: that's
 396	 * okay, that truncation will have unmapped the PageKsm for us.
 397	 *
 398	 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
 399	 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
 400	 * current task has TIF_MEMDIE set, and will be OOM killed on return
 401	 * to user; and ksmd, having no mm, would never be chosen for that.
 402	 *
 403	 * But if the mm is in a limited mem_cgroup, then the fault may fail
 404	 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
 405	 * even ksmd can fail in this way - though it's usually breaking ksm
 406	 * just to undo a merge it made a moment before, so unlikely to oom.
 407	 *
 408	 * That's a pity: we might therefore have more kernel pages allocated
 409	 * than we're counting as nodes in the stable tree; but ksm_do_scan
 410	 * will retry to break_cow on each pass, so should recover the page
 411	 * in due course.  The important thing is to not let VM_MERGEABLE
 412	 * be cleared while any such pages might remain in the area.
 413	 */
 414	return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
 415}
 416
 417static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
 418		unsigned long addr)
 419{
 420	struct vm_area_struct *vma;
 421	if (ksm_test_exit(mm))
 422		return NULL;
 423	vma = find_vma(mm, addr);
 424	if (!vma || vma->vm_start > addr)
 425		return NULL;
 426	if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
 427		return NULL;
 428	return vma;
 429}
 430
 431static void break_cow(struct rmap_item *rmap_item)
 432{
 433	struct mm_struct *mm = rmap_item->mm;
 434	unsigned long addr = rmap_item->address;
 435	struct vm_area_struct *vma;
 436
 437	/*
 438	 * It is not an accident that whenever we want to break COW
 439	 * to undo, we also need to drop a reference to the anon_vma.
 440	 */
 441	put_anon_vma(rmap_item->anon_vma);
 442
 443	down_read(&mm->mmap_sem);
 444	vma = find_mergeable_vma(mm, addr);
 445	if (vma)
 446		break_ksm(vma, addr);
 447	up_read(&mm->mmap_sem);
 448}
 449
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 450static struct page *get_mergeable_page(struct rmap_item *rmap_item)
 451{
 452	struct mm_struct *mm = rmap_item->mm;
 453	unsigned long addr = rmap_item->address;
 454	struct vm_area_struct *vma;
 455	struct page *page;
 456
 457	down_read(&mm->mmap_sem);
 458	vma = find_mergeable_vma(mm, addr);
 459	if (!vma)
 460		goto out;
 461
 462	page = follow_page(vma, addr, FOLL_GET);
 463	if (IS_ERR_OR_NULL(page))
 464		goto out;
 465	if (PageAnon(page)) {
 466		flush_anon_page(vma, page, addr);
 467		flush_dcache_page(page);
 468	} else {
 469		put_page(page);
 470out:
 471		page = NULL;
 472	}
 473	up_read(&mm->mmap_sem);
 474	return page;
 475}
 476
 477/*
 478 * This helper is used for getting right index into array of tree roots.
 479 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
 480 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
 481 * every node has its own stable and unstable tree.
 482 */
 483static inline int get_kpfn_nid(unsigned long kpfn)
 484{
 485	return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
 486}
 487
 488static void remove_node_from_stable_tree(struct stable_node *stable_node)
 489{
 490	struct rmap_item *rmap_item;
 491
 492	hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
 493		if (rmap_item->hlist.next)
 494			ksm_pages_sharing--;
 495		else
 496			ksm_pages_shared--;
 497		put_anon_vma(rmap_item->anon_vma);
 498		rmap_item->address &= PAGE_MASK;
 499		cond_resched();
 500	}
 501
 502	if (stable_node->head == &migrate_nodes)
 503		list_del(&stable_node->list);
 504	else
 505		rb_erase(&stable_node->node,
 506			 root_stable_tree + NUMA(stable_node->nid));
 507	free_stable_node(stable_node);
 508}
 509
 510/*
 511 * get_ksm_page: checks if the page indicated by the stable node
 512 * is still its ksm page, despite having held no reference to it.
 513 * In which case we can trust the content of the page, and it
 514 * returns the gotten page; but if the page has now been zapped,
 515 * remove the stale node from the stable tree and return NULL.
 516 * But beware, the stable node's page might be being migrated.
 517 *
 518 * You would expect the stable_node to hold a reference to the ksm page.
 519 * But if it increments the page's count, swapping out has to wait for
 520 * ksmd to come around again before it can free the page, which may take
 521 * seconds or even minutes: much too unresponsive.  So instead we use a
 522 * "keyhole reference": access to the ksm page from the stable node peeps
 523 * out through its keyhole to see if that page still holds the right key,
 524 * pointing back to this stable node.  This relies on freeing a PageAnon
 525 * page to reset its page->mapping to NULL, and relies on no other use of
 526 * a page to put something that might look like our key in page->mapping.
 527 * is on its way to being freed; but it is an anomaly to bear in mind.
 528 */
 529static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
 530{
 531	struct page *page;
 532	void *expected_mapping;
 533	unsigned long kpfn;
 534
 535	expected_mapping = (void *)stable_node +
 536				(PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
 537again:
 538	kpfn = READ_ONCE(stable_node->kpfn);
 539	page = pfn_to_page(kpfn);
 540
 541	/*
 542	 * page is computed from kpfn, so on most architectures reading
 543	 * page->mapping is naturally ordered after reading node->kpfn,
 544	 * but on Alpha we need to be more careful.
 545	 */
 546	smp_read_barrier_depends();
 547	if (READ_ONCE(page->mapping) != expected_mapping)
 548		goto stale;
 549
 550	/*
 551	 * We cannot do anything with the page while its refcount is 0.
 552	 * Usually 0 means free, or tail of a higher-order page: in which
 553	 * case this node is no longer referenced, and should be freed;
 554	 * however, it might mean that the page is under page_freeze_refs().
 555	 * The __remove_mapping() case is easy, again the node is now stale;
 556	 * but if page is swapcache in migrate_page_move_mapping(), it might
 557	 * still be our page, in which case it's essential to keep the node.
 558	 */
 559	while (!get_page_unless_zero(page)) {
 560		/*
 561		 * Another check for page->mapping != expected_mapping would
 562		 * work here too.  We have chosen the !PageSwapCache test to
 563		 * optimize the common case, when the page is or is about to
 564		 * be freed: PageSwapCache is cleared (under spin_lock_irq)
 565		 * in the freeze_refs section of __remove_mapping(); but Anon
 566		 * page->mapping reset to NULL later, in free_pages_prepare().
 567		 */
 568		if (!PageSwapCache(page))
 569			goto stale;
 570		cpu_relax();
 571	}
 572
 573	if (READ_ONCE(page->mapping) != expected_mapping) {
 574		put_page(page);
 575		goto stale;
 576	}
 577
 578	if (lock_it) {
 579		lock_page(page);
 580		if (READ_ONCE(page->mapping) != expected_mapping) {
 581			unlock_page(page);
 582			put_page(page);
 583			goto stale;
 584		}
 585	}
 586	return page;
 587
 588stale:
 589	/*
 590	 * We come here from above when page->mapping or !PageSwapCache
 591	 * suggests that the node is stale; but it might be under migration.
 592	 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
 593	 * before checking whether node->kpfn has been changed.
 594	 */
 595	smp_rmb();
 596	if (READ_ONCE(stable_node->kpfn) != kpfn)
 597		goto again;
 598	remove_node_from_stable_tree(stable_node);
 599	return NULL;
 600}
 601
 602/*
 603 * Removing rmap_item from stable or unstable tree.
 604 * This function will clean the information from the stable/unstable tree.
 605 */
 606static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
 607{
 608	if (rmap_item->address & STABLE_FLAG) {
 609		struct stable_node *stable_node;
 610		struct page *page;
 611
 612		stable_node = rmap_item->head;
 613		page = get_ksm_page(stable_node, true);
 614		if (!page)
 615			goto out;
 616
 617		hlist_del(&rmap_item->hlist);
 618		unlock_page(page);
 619		put_page(page);
 620
 621		if (!hlist_empty(&stable_node->hlist))
 622			ksm_pages_sharing--;
 623		else
 624			ksm_pages_shared--;
 625
 626		put_anon_vma(rmap_item->anon_vma);
 627		rmap_item->address &= PAGE_MASK;
 628
 629	} else if (rmap_item->address & UNSTABLE_FLAG) {
 630		unsigned char age;
 631		/*
 632		 * Usually ksmd can and must skip the rb_erase, because
 633		 * root_unstable_tree was already reset to RB_ROOT.
 634		 * But be careful when an mm is exiting: do the rb_erase
 635		 * if this rmap_item was inserted by this scan, rather
 636		 * than left over from before.
 637		 */
 638		age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
 639		BUG_ON(age > 1);
 640		if (!age)
 641			rb_erase(&rmap_item->node,
 642				 root_unstable_tree + NUMA(rmap_item->nid));
 643		ksm_pages_unshared--;
 644		rmap_item->address &= PAGE_MASK;
 645	}
 646out:
 647	cond_resched();		/* we're called from many long loops */
 648}
 649
 650static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
 651				       struct rmap_item **rmap_list)
 652{
 653	while (*rmap_list) {
 654		struct rmap_item *rmap_item = *rmap_list;
 655		*rmap_list = rmap_item->rmap_list;
 656		remove_rmap_item_from_tree(rmap_item);
 657		free_rmap_item(rmap_item);
 658	}
 659}
 660
 661/*
 662 * Though it's very tempting to unmerge rmap_items from stable tree rather
 663 * than check every pte of a given vma, the locking doesn't quite work for
 664 * that - an rmap_item is assigned to the stable tree after inserting ksm
 665 * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
 666 * rmap_items from parent to child at fork time (so as not to waste time
 667 * if exit comes before the next scan reaches it).
 668 *
 669 * Similarly, although we'd like to remove rmap_items (so updating counts
 670 * and freeing memory) when unmerging an area, it's easier to leave that
 671 * to the next pass of ksmd - consider, for example, how ksmd might be
 672 * in cmp_and_merge_page on one of the rmap_items we would be removing.
 673 */
 674static int unmerge_ksm_pages(struct vm_area_struct *vma,
 675			     unsigned long start, unsigned long end)
 676{
 677	unsigned long addr;
 678	int err = 0;
 679
 680	for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
 681		if (ksm_test_exit(vma->vm_mm))
 682			break;
 683		if (signal_pending(current))
 684			err = -ERESTARTSYS;
 685		else
 686			err = break_ksm(vma, addr);
 687	}
 688	return err;
 689}
 690
 691#ifdef CONFIG_SYSFS
 692/*
 693 * Only called through the sysfs control interface:
 694 */
 695static int remove_stable_node(struct stable_node *stable_node)
 696{
 697	struct page *page;
 698	int err;
 699
 700	page = get_ksm_page(stable_node, true);
 701	if (!page) {
 702		/*
 703		 * get_ksm_page did remove_node_from_stable_tree itself.
 704		 */
 705		return 0;
 706	}
 707
 708	if (WARN_ON_ONCE(page_mapped(page))) {
 709		/*
 710		 * This should not happen: but if it does, just refuse to let
 711		 * merge_across_nodes be switched - there is no need to panic.
 712		 */
 713		err = -EBUSY;
 714	} else {
 715		/*
 716		 * The stable node did not yet appear stale to get_ksm_page(),
 717		 * since that allows for an unmapped ksm page to be recognized
 718		 * right up until it is freed; but the node is safe to remove.
 719		 * This page might be in a pagevec waiting to be freed,
 720		 * or it might be PageSwapCache (perhaps under writeback),
 721		 * or it might have been removed from swapcache a moment ago.
 722		 */
 723		set_page_stable_node(page, NULL);
 724		remove_node_from_stable_tree(stable_node);
 725		err = 0;
 726	}
 727
 728	unlock_page(page);
 729	put_page(page);
 730	return err;
 731}
 732
 733static int remove_all_stable_nodes(void)
 734{
 735	struct stable_node *stable_node, *next;
 
 736	int nid;
 737	int err = 0;
 738
 739	for (nid = 0; nid < ksm_nr_node_ids; nid++) {
 740		while (root_stable_tree[nid].rb_node) {
 741			stable_node = rb_entry(root_stable_tree[nid].rb_node,
 742						struct stable_node, node);
 743			if (remove_stable_node(stable_node)) {
 744				err = -EBUSY;
 745				break;	/* proceed to next nid */
 746			}
 747			cond_resched();
 748		}
 749	}
 750	list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
 
 751		if (remove_stable_node(stable_node))
 752			err = -EBUSY;
 753		cond_resched();
 754	}
 755	return err;
 756}
 757
 758static int unmerge_and_remove_all_rmap_items(void)
 759{
 760	struct mm_slot *mm_slot;
 761	struct mm_struct *mm;
 762	struct vm_area_struct *vma;
 763	int err = 0;
 764
 765	spin_lock(&ksm_mmlist_lock);
 766	ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
 767						struct mm_slot, mm_list);
 768	spin_unlock(&ksm_mmlist_lock);
 769
 770	for (mm_slot = ksm_scan.mm_slot;
 771			mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
 772		mm = mm_slot->mm;
 773		down_read(&mm->mmap_sem);
 774		for (vma = mm->mmap; vma; vma = vma->vm_next) {
 775			if (ksm_test_exit(mm))
 776				break;
 777			if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
 778				continue;
 779			err = unmerge_ksm_pages(vma,
 780						vma->vm_start, vma->vm_end);
 781			if (err)
 782				goto error;
 783		}
 784
 785		remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
 786		up_read(&mm->mmap_sem);
 787
 788		spin_lock(&ksm_mmlist_lock);
 789		ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
 790						struct mm_slot, mm_list);
 791		if (ksm_test_exit(mm)) {
 792			hash_del(&mm_slot->link);
 793			list_del(&mm_slot->mm_list);
 794			spin_unlock(&ksm_mmlist_lock);
 795
 796			free_mm_slot(mm_slot);
 797			clear_bit(MMF_VM_MERGEABLE, &mm->flags);
 
 798			mmdrop(mm);
 799		} else
 800			spin_unlock(&ksm_mmlist_lock);
 
 
 801	}
 802
 803	/* Clean up stable nodes, but don't worry if some are still busy */
 804	remove_all_stable_nodes();
 805	ksm_scan.seqnr = 0;
 806	return 0;
 807
 808error:
 809	up_read(&mm->mmap_sem);
 810	spin_lock(&ksm_mmlist_lock);
 811	ksm_scan.mm_slot = &ksm_mm_head;
 812	spin_unlock(&ksm_mmlist_lock);
 813	return err;
 814}
 815#endif /* CONFIG_SYSFS */
 816
 817static u32 calc_checksum(struct page *page)
 818{
 819	u32 checksum;
 820	void *addr = kmap_atomic(page);
 821	checksum = jhash2(addr, PAGE_SIZE / 4, 17);
 822	kunmap_atomic(addr);
 823	return checksum;
 824}
 825
 826static int memcmp_pages(struct page *page1, struct page *page2)
 827{
 828	char *addr1, *addr2;
 829	int ret;
 830
 831	addr1 = kmap_atomic(page1);
 832	addr2 = kmap_atomic(page2);
 833	ret = memcmp(addr1, addr2, PAGE_SIZE);
 834	kunmap_atomic(addr2);
 835	kunmap_atomic(addr1);
 836	return ret;
 837}
 838
 839static inline int pages_identical(struct page *page1, struct page *page2)
 840{
 841	return !memcmp_pages(page1, page2);
 842}
 843
 844static int write_protect_page(struct vm_area_struct *vma, struct page *page,
 845			      pte_t *orig_pte)
 846{
 847	struct mm_struct *mm = vma->vm_mm;
 848	unsigned long addr;
 849	pte_t *ptep;
 850	spinlock_t *ptl;
 851	int swapped;
 852	int err = -EFAULT;
 853	unsigned long mmun_start;	/* For mmu_notifiers */
 854	unsigned long mmun_end;		/* For mmu_notifiers */
 855
 856	addr = page_address_in_vma(page, vma);
 857	if (addr == -EFAULT)
 858		goto out;
 859
 860	BUG_ON(PageTransCompound(page));
 861
 862	mmun_start = addr;
 863	mmun_end   = addr + PAGE_SIZE;
 864	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 865
 866	ptep = page_check_address(page, mm, addr, &ptl, 0);
 867	if (!ptep)
 868		goto out_mn;
 869
 870	if (pte_write(*ptep) || pte_dirty(*ptep)) {
 871		pte_t entry;
 872
 873		swapped = PageSwapCache(page);
 874		flush_cache_page(vma, addr, page_to_pfn(page));
 875		/*
 876		 * Ok this is tricky, when get_user_pages_fast() run it doesn't
 877		 * take any lock, therefore the check that we are going to make
 878		 * with the pagecount against the mapcount is racey and
 879		 * O_DIRECT can happen right after the check.
 880		 * So we clear the pte and flush the tlb before the check
 881		 * this assure us that no O_DIRECT can happen after the check
 882		 * or in the middle of the check.
 883		 */
 884		entry = ptep_clear_flush_notify(vma, addr, ptep);
 885		/*
 886		 * Check that no O_DIRECT or similar I/O is in progress on the
 887		 * page
 888		 */
 889		if (page_mapcount(page) + 1 + swapped != page_count(page)) {
 890			set_pte_at(mm, addr, ptep, entry);
 891			goto out_unlock;
 892		}
 893		if (pte_dirty(entry))
 894			set_page_dirty(page);
 895		entry = pte_mkclean(pte_wrprotect(entry));
 896		set_pte_at_notify(mm, addr, ptep, entry);
 897	}
 898	*orig_pte = *ptep;
 899	err = 0;
 900
 901out_unlock:
 902	pte_unmap_unlock(ptep, ptl);
 903out_mn:
 904	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 905out:
 906	return err;
 907}
 908
 909/**
 910 * replace_page - replace page in vma by new ksm page
 911 * @vma:      vma that holds the pte pointing to page
 912 * @page:     the page we are replacing by kpage
 913 * @kpage:    the ksm page we replace page by
 914 * @orig_pte: the original value of the pte
 915 *
 916 * Returns 0 on success, -EFAULT on failure.
 917 */
 918static int replace_page(struct vm_area_struct *vma, struct page *page,
 919			struct page *kpage, pte_t orig_pte)
 920{
 921	struct mm_struct *mm = vma->vm_mm;
 922	pmd_t *pmd;
 923	pte_t *ptep;
 924	spinlock_t *ptl;
 925	unsigned long addr;
 926	int err = -EFAULT;
 927	unsigned long mmun_start;	/* For mmu_notifiers */
 928	unsigned long mmun_end;		/* For mmu_notifiers */
 929
 930	addr = page_address_in_vma(page, vma);
 931	if (addr == -EFAULT)
 932		goto out;
 933
 934	pmd = mm_find_pmd(mm, addr);
 935	if (!pmd)
 936		goto out;
 
 937
 938	mmun_start = addr;
 939	mmun_end   = addr + PAGE_SIZE;
 940	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 941
 942	ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
 943	if (!pte_same(*ptep, orig_pte)) {
 944		pte_unmap_unlock(ptep, ptl);
 945		goto out_mn;
 946	}
 947
 948	get_page(kpage);
 949	page_add_anon_rmap(kpage, vma, addr, false);
 950
 951	flush_cache_page(vma, addr, pte_pfn(*ptep));
 952	ptep_clear_flush_notify(vma, addr, ptep);
 953	set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
 954
 955	page_remove_rmap(page, false);
 956	if (!page_mapped(page))
 957		try_to_free_swap(page);
 958	put_page(page);
 959
 960	pte_unmap_unlock(ptep, ptl);
 961	err = 0;
 962out_mn:
 963	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 964out:
 965	return err;
 966}
 967
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 968/*
 969 * try_to_merge_one_page - take two pages and merge them into one
 970 * @vma: the vma that holds the pte pointing to page
 971 * @page: the PageAnon page that we want to replace with kpage
 972 * @kpage: the PageKsm page that we want to map instead of page,
 973 *         or NULL the first time when we want to use page as kpage.
 974 *
 975 * This function returns 0 if the pages were merged, -EFAULT otherwise.
 976 */
 977static int try_to_merge_one_page(struct vm_area_struct *vma,
 978				 struct page *page, struct page *kpage)
 979{
 980	pte_t orig_pte = __pte(0);
 981	int err = -EFAULT;
 982
 983	if (page == kpage)			/* ksm page forked */
 984		return 0;
 985
 
 
 
 
 
 986	if (!PageAnon(page))
 987		goto out;
 988
 989	/*
 990	 * We need the page lock to read a stable PageSwapCache in
 991	 * write_protect_page().  We use trylock_page() instead of
 992	 * lock_page() because we don't want to wait here - we
 993	 * prefer to continue scanning and merging different pages,
 994	 * then come back to this page when it is unlocked.
 995	 */
 996	if (!trylock_page(page))
 997		goto out;
 998
 999	if (PageTransCompound(page)) {
1000		err = split_huge_page(page);
1001		if (err)
1002			goto out_unlock;
1003	}
1004
1005	/*
1006	 * If this anonymous page is mapped only here, its pte may need
1007	 * to be write-protected.  If it's mapped elsewhere, all of its
1008	 * ptes are necessarily already write-protected.  But in either
1009	 * case, we need to lock and check page_count is not raised.
1010	 */
1011	if (write_protect_page(vma, page, &orig_pte) == 0) {
1012		if (!kpage) {
1013			/*
1014			 * While we hold page lock, upgrade page from
1015			 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1016			 * stable_tree_insert() will update stable_node.
1017			 */
1018			set_page_stable_node(page, NULL);
1019			mark_page_accessed(page);
1020			/*
1021			 * Page reclaim just frees a clean page with no dirty
1022			 * ptes: make sure that the ksm page would be swapped.
1023			 */
1024			if (!PageDirty(page))
1025				SetPageDirty(page);
1026			err = 0;
1027		} else if (pages_identical(page, kpage))
1028			err = replace_page(vma, page, kpage, orig_pte);
1029	}
1030
1031	if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1032		munlock_vma_page(page);
1033		if (!PageMlocked(kpage)) {
1034			unlock_page(page);
1035			lock_page(kpage);
1036			mlock_vma_page(kpage);
1037			page = kpage;		/* for final unlock */
1038		}
1039	}
1040
1041out_unlock:
1042	unlock_page(page);
1043out:
1044	return err;
1045}
1046
1047/*
1048 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1049 * but no new kernel page is allocated: kpage must already be a ksm page.
1050 *
1051 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1052 */
1053static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1054				      struct page *page, struct page *kpage)
1055{
1056	struct mm_struct *mm = rmap_item->mm;
1057	struct vm_area_struct *vma;
1058	int err = -EFAULT;
1059
1060	down_read(&mm->mmap_sem);
1061	vma = find_mergeable_vma(mm, rmap_item->address);
1062	if (!vma)
 
 
1063		goto out;
1064
1065	err = try_to_merge_one_page(vma, page, kpage);
1066	if (err)
1067		goto out;
1068
1069	/* Unstable nid is in union with stable anon_vma: remove first */
1070	remove_rmap_item_from_tree(rmap_item);
1071
1072	/* Must get reference to anon_vma while still holding mmap_sem */
1073	rmap_item->anon_vma = vma->anon_vma;
1074	get_anon_vma(vma->anon_vma);
1075out:
1076	up_read(&mm->mmap_sem);
1077	return err;
1078}
1079
1080/*
1081 * try_to_merge_two_pages - take two identical pages and prepare them
1082 * to be merged into one page.
1083 *
1084 * This function returns the kpage if we successfully merged two identical
1085 * pages into one ksm page, NULL otherwise.
1086 *
1087 * Note that this function upgrades page to ksm page: if one of the pages
1088 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1089 */
1090static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1091					   struct page *page,
1092					   struct rmap_item *tree_rmap_item,
1093					   struct page *tree_page)
1094{
1095	int err;
1096
1097	err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1098	if (!err) {
1099		err = try_to_merge_with_ksm_page(tree_rmap_item,
1100							tree_page, page);
1101		/*
1102		 * If that fails, we have a ksm page with only one pte
1103		 * pointing to it: so break it.
1104		 */
1105		if (err)
1106			break_cow(rmap_item);
1107	}
1108	return err ? NULL : page;
1109}
1110
1111/*
1112 * stable_tree_search - search for page inside the stable tree
1113 *
1114 * This function checks if there is a page inside the stable tree
1115 * with identical content to the page that we are scanning right now.
1116 *
1117 * This function returns the stable tree node of identical content if found,
1118 * NULL otherwise.
1119 */
1120static struct page *stable_tree_search(struct page *page)
1121{
1122	int nid;
1123	struct rb_root *root;
1124	struct rb_node **new;
1125	struct rb_node *parent;
1126	struct stable_node *stable_node;
1127	struct stable_node *page_node;
1128
1129	page_node = page_stable_node(page);
1130	if (page_node && page_node->head != &migrate_nodes) {
1131		/* ksm page forked */
1132		get_page(page);
1133		return page;
1134	}
1135
1136	nid = get_kpfn_nid(page_to_pfn(page));
1137	root = root_stable_tree + nid;
1138again:
1139	new = &root->rb_node;
1140	parent = NULL;
1141
1142	while (*new) {
1143		struct page *tree_page;
1144		int ret;
1145
1146		cond_resched();
1147		stable_node = rb_entry(*new, struct stable_node, node);
1148		tree_page = get_ksm_page(stable_node, false);
1149		if (!tree_page) {
1150			/*
1151			 * If we walked over a stale stable_node,
1152			 * get_ksm_page() will call rb_erase() and it
1153			 * may rebalance the tree from under us. So
1154			 * restart the search from scratch. Returning
1155			 * NULL would be safe too, but we'd generate
1156			 * false negative insertions just because some
1157			 * stable_node was stale.
1158			 */
1159			goto again;
1160		}
1161
1162		ret = memcmp_pages(page, tree_page);
1163		put_page(tree_page);
1164
1165		parent = *new;
1166		if (ret < 0)
1167			new = &parent->rb_left;
1168		else if (ret > 0)
1169			new = &parent->rb_right;
1170		else {
1171			/*
1172			 * Lock and unlock the stable_node's page (which
1173			 * might already have been migrated) so that page
1174			 * migration is sure to notice its raised count.
1175			 * It would be more elegant to return stable_node
1176			 * than kpage, but that involves more changes.
1177			 */
1178			tree_page = get_ksm_page(stable_node, true);
1179			if (tree_page) {
1180				unlock_page(tree_page);
1181				if (get_kpfn_nid(stable_node->kpfn) !=
1182						NUMA(stable_node->nid)) {
1183					put_page(tree_page);
1184					goto replace;
1185				}
1186				return tree_page;
1187			}
1188			/*
1189			 * There is now a place for page_node, but the tree may
1190			 * have been rebalanced, so re-evaluate parent and new.
1191			 */
1192			if (page_node)
1193				goto again;
1194			return NULL;
1195		}
1196	}
1197
1198	if (!page_node)
1199		return NULL;
1200
1201	list_del(&page_node->list);
1202	DO_NUMA(page_node->nid = nid);
1203	rb_link_node(&page_node->node, parent, new);
1204	rb_insert_color(&page_node->node, root);
1205	get_page(page);
1206	return page;
1207
1208replace:
1209	if (page_node) {
1210		list_del(&page_node->list);
1211		DO_NUMA(page_node->nid = nid);
1212		rb_replace_node(&stable_node->node, &page_node->node, root);
1213		get_page(page);
1214	} else {
1215		rb_erase(&stable_node->node, root);
1216		page = NULL;
1217	}
1218	stable_node->head = &migrate_nodes;
1219	list_add(&stable_node->list, stable_node->head);
1220	return page;
1221}
1222
1223/*
1224 * stable_tree_insert - insert stable tree node pointing to new ksm page
1225 * into the stable tree.
1226 *
1227 * This function returns the stable tree node just allocated on success,
1228 * NULL otherwise.
1229 */
1230static struct stable_node *stable_tree_insert(struct page *kpage)
1231{
1232	int nid;
1233	unsigned long kpfn;
1234	struct rb_root *root;
1235	struct rb_node **new;
1236	struct rb_node *parent;
1237	struct stable_node *stable_node;
1238
1239	kpfn = page_to_pfn(kpage);
1240	nid = get_kpfn_nid(kpfn);
1241	root = root_stable_tree + nid;
1242again:
1243	parent = NULL;
1244	new = &root->rb_node;
1245
1246	while (*new) {
1247		struct page *tree_page;
1248		int ret;
1249
1250		cond_resched();
1251		stable_node = rb_entry(*new, struct stable_node, node);
1252		tree_page = get_ksm_page(stable_node, false);
1253		if (!tree_page) {
1254			/*
1255			 * If we walked over a stale stable_node,
1256			 * get_ksm_page() will call rb_erase() and it
1257			 * may rebalance the tree from under us. So
1258			 * restart the search from scratch. Returning
1259			 * NULL would be safe too, but we'd generate
1260			 * false negative insertions just because some
1261			 * stable_node was stale.
1262			 */
1263			goto again;
1264		}
1265
1266		ret = memcmp_pages(kpage, tree_page);
1267		put_page(tree_page);
1268
1269		parent = *new;
1270		if (ret < 0)
1271			new = &parent->rb_left;
1272		else if (ret > 0)
1273			new = &parent->rb_right;
1274		else {
1275			/*
1276			 * It is not a bug that stable_tree_search() didn't
1277			 * find this node: because at that time our page was
1278			 * not yet write-protected, so may have changed since.
1279			 */
1280			return NULL;
1281		}
1282	}
1283
1284	stable_node = alloc_stable_node();
1285	if (!stable_node)
1286		return NULL;
1287
1288	INIT_HLIST_HEAD(&stable_node->hlist);
1289	stable_node->kpfn = kpfn;
1290	set_page_stable_node(kpage, stable_node);
1291	DO_NUMA(stable_node->nid = nid);
1292	rb_link_node(&stable_node->node, parent, new);
1293	rb_insert_color(&stable_node->node, root);
1294
1295	return stable_node;
1296}
1297
1298/*
1299 * unstable_tree_search_insert - search for identical page,
1300 * else insert rmap_item into the unstable tree.
1301 *
1302 * This function searches for a page in the unstable tree identical to the
1303 * page currently being scanned; and if no identical page is found in the
1304 * tree, we insert rmap_item as a new object into the unstable tree.
1305 *
1306 * This function returns pointer to rmap_item found to be identical
1307 * to the currently scanned page, NULL otherwise.
1308 *
1309 * This function does both searching and inserting, because they share
1310 * the same walking algorithm in an rbtree.
1311 */
1312static
1313struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1314					      struct page *page,
1315					      struct page **tree_pagep)
1316{
1317	struct rb_node **new;
1318	struct rb_root *root;
1319	struct rb_node *parent = NULL;
1320	int nid;
1321
1322	nid = get_kpfn_nid(page_to_pfn(page));
1323	root = root_unstable_tree + nid;
1324	new = &root->rb_node;
1325
1326	while (*new) {
1327		struct rmap_item *tree_rmap_item;
1328		struct page *tree_page;
1329		int ret;
1330
1331		cond_resched();
1332		tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1333		tree_page = get_mergeable_page(tree_rmap_item);
1334		if (!tree_page)
1335			return NULL;
1336
1337		/*
1338		 * Don't substitute a ksm page for a forked page.
1339		 */
1340		if (page == tree_page) {
1341			put_page(tree_page);
1342			return NULL;
1343		}
1344
1345		ret = memcmp_pages(page, tree_page);
1346
1347		parent = *new;
1348		if (ret < 0) {
1349			put_page(tree_page);
1350			new = &parent->rb_left;
1351		} else if (ret > 0) {
1352			put_page(tree_page);
1353			new = &parent->rb_right;
1354		} else if (!ksm_merge_across_nodes &&
1355			   page_to_nid(tree_page) != nid) {
1356			/*
1357			 * If tree_page has been migrated to another NUMA node,
1358			 * it will be flushed out and put in the right unstable
1359			 * tree next time: only merge with it when across_nodes.
1360			 */
1361			put_page(tree_page);
1362			return NULL;
1363		} else {
1364			*tree_pagep = tree_page;
1365			return tree_rmap_item;
1366		}
1367	}
1368
1369	rmap_item->address |= UNSTABLE_FLAG;
1370	rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1371	DO_NUMA(rmap_item->nid = nid);
1372	rb_link_node(&rmap_item->node, parent, new);
1373	rb_insert_color(&rmap_item->node, root);
1374
1375	ksm_pages_unshared++;
1376	return NULL;
1377}
1378
1379/*
1380 * stable_tree_append - add another rmap_item to the linked list of
1381 * rmap_items hanging off a given node of the stable tree, all sharing
1382 * the same ksm page.
1383 */
1384static void stable_tree_append(struct rmap_item *rmap_item,
1385			       struct stable_node *stable_node)
1386{
1387	rmap_item->head = stable_node;
1388	rmap_item->address |= STABLE_FLAG;
1389	hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1390
1391	if (rmap_item->hlist.next)
1392		ksm_pages_sharing++;
1393	else
1394		ksm_pages_shared++;
1395}
1396
1397/*
1398 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1399 * if not, compare checksum to previous and if it's the same, see if page can
1400 * be inserted into the unstable tree, or merged with a page already there and
1401 * both transferred to the stable tree.
1402 *
1403 * @page: the page that we are searching identical page to.
1404 * @rmap_item: the reverse mapping into the virtual address of this page
1405 */
1406static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1407{
1408	struct rmap_item *tree_rmap_item;
1409	struct page *tree_page = NULL;
1410	struct stable_node *stable_node;
1411	struct page *kpage;
1412	unsigned int checksum;
1413	int err;
1414
1415	stable_node = page_stable_node(page);
1416	if (stable_node) {
1417		if (stable_node->head != &migrate_nodes &&
1418		    get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1419			rb_erase(&stable_node->node,
1420				 root_stable_tree + NUMA(stable_node->nid));
1421			stable_node->head = &migrate_nodes;
1422			list_add(&stable_node->list, stable_node->head);
1423		}
1424		if (stable_node->head != &migrate_nodes &&
1425		    rmap_item->head == stable_node)
1426			return;
1427	}
1428
1429	/* We first start with searching the page inside the stable tree */
1430	kpage = stable_tree_search(page);
1431	if (kpage == page && rmap_item->head == stable_node) {
1432		put_page(kpage);
1433		return;
1434	}
1435
1436	remove_rmap_item_from_tree(rmap_item);
1437
1438	if (kpage) {
1439		err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1440		if (!err) {
1441			/*
1442			 * The page was successfully merged:
1443			 * add its rmap_item to the stable tree.
1444			 */
1445			lock_page(kpage);
1446			stable_tree_append(rmap_item, page_stable_node(kpage));
1447			unlock_page(kpage);
1448		}
1449		put_page(kpage);
1450		return;
1451	}
1452
1453	/*
1454	 * If the hash value of the page has changed from the last time
1455	 * we calculated it, this page is changing frequently: therefore we
1456	 * don't want to insert it in the unstable tree, and we don't want
1457	 * to waste our time searching for something identical to it there.
1458	 */
1459	checksum = calc_checksum(page);
1460	if (rmap_item->oldchecksum != checksum) {
1461		rmap_item->oldchecksum = checksum;
1462		return;
1463	}
1464
1465	tree_rmap_item =
1466		unstable_tree_search_insert(rmap_item, page, &tree_page);
1467	if (tree_rmap_item) {
1468		kpage = try_to_merge_two_pages(rmap_item, page,
1469						tree_rmap_item, tree_page);
1470		put_page(tree_page);
1471		if (kpage) {
1472			/*
1473			 * The pages were successfully merged: insert new
1474			 * node in the stable tree and add both rmap_items.
1475			 */
1476			lock_page(kpage);
1477			stable_node = stable_tree_insert(kpage);
1478			if (stable_node) {
1479				stable_tree_append(tree_rmap_item, stable_node);
1480				stable_tree_append(rmap_item, stable_node);
1481			}
1482			unlock_page(kpage);
1483
1484			/*
1485			 * If we fail to insert the page into the stable tree,
1486			 * we will have 2 virtual addresses that are pointing
1487			 * to a ksm page left outside the stable tree,
1488			 * in which case we need to break_cow on both.
1489			 */
1490			if (!stable_node) {
1491				break_cow(tree_rmap_item);
1492				break_cow(rmap_item);
1493			}
1494		}
1495	}
1496}
1497
1498static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1499					    struct rmap_item **rmap_list,
1500					    unsigned long addr)
1501{
1502	struct rmap_item *rmap_item;
1503
1504	while (*rmap_list) {
1505		rmap_item = *rmap_list;
1506		if ((rmap_item->address & PAGE_MASK) == addr)
1507			return rmap_item;
1508		if (rmap_item->address > addr)
1509			break;
1510		*rmap_list = rmap_item->rmap_list;
1511		remove_rmap_item_from_tree(rmap_item);
1512		free_rmap_item(rmap_item);
1513	}
1514
1515	rmap_item = alloc_rmap_item();
1516	if (rmap_item) {
1517		/* It has already been zeroed */
1518		rmap_item->mm = mm_slot->mm;
1519		rmap_item->address = addr;
1520		rmap_item->rmap_list = *rmap_list;
1521		*rmap_list = rmap_item;
1522	}
1523	return rmap_item;
1524}
1525
1526static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1527{
1528	struct mm_struct *mm;
1529	struct mm_slot *slot;
1530	struct vm_area_struct *vma;
1531	struct rmap_item *rmap_item;
1532	int nid;
1533
1534	if (list_empty(&ksm_mm_head.mm_list))
1535		return NULL;
1536
1537	slot = ksm_scan.mm_slot;
1538	if (slot == &ksm_mm_head) {
1539		/*
1540		 * A number of pages can hang around indefinitely on per-cpu
1541		 * pagevecs, raised page count preventing write_protect_page
1542		 * from merging them.  Though it doesn't really matter much,
1543		 * it is puzzling to see some stuck in pages_volatile until
1544		 * other activity jostles them out, and they also prevented
1545		 * LTP's KSM test from succeeding deterministically; so drain
1546		 * them here (here rather than on entry to ksm_do_scan(),
1547		 * so we don't IPI too often when pages_to_scan is set low).
1548		 */
1549		lru_add_drain_all();
1550
1551		/*
1552		 * Whereas stale stable_nodes on the stable_tree itself
1553		 * get pruned in the regular course of stable_tree_search(),
1554		 * those moved out to the migrate_nodes list can accumulate:
1555		 * so prune them once before each full scan.
1556		 */
1557		if (!ksm_merge_across_nodes) {
1558			struct stable_node *stable_node, *next;
 
1559			struct page *page;
1560
1561			list_for_each_entry_safe(stable_node, next,
1562						 &migrate_nodes, list) {
 
1563				page = get_ksm_page(stable_node, false);
1564				if (page)
1565					put_page(page);
1566				cond_resched();
1567			}
1568		}
1569
1570		for (nid = 0; nid < ksm_nr_node_ids; nid++)
1571			root_unstable_tree[nid] = RB_ROOT;
1572
1573		spin_lock(&ksm_mmlist_lock);
1574		slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1575		ksm_scan.mm_slot = slot;
1576		spin_unlock(&ksm_mmlist_lock);
1577		/*
1578		 * Although we tested list_empty() above, a racing __ksm_exit
1579		 * of the last mm on the list may have removed it since then.
1580		 */
1581		if (slot == &ksm_mm_head)
1582			return NULL;
1583next_mm:
1584		ksm_scan.address = 0;
1585		ksm_scan.rmap_list = &slot->rmap_list;
1586	}
1587
1588	mm = slot->mm;
1589	down_read(&mm->mmap_sem);
1590	if (ksm_test_exit(mm))
1591		vma = NULL;
1592	else
1593		vma = find_vma(mm, ksm_scan.address);
1594
1595	for (; vma; vma = vma->vm_next) {
1596		if (!(vma->vm_flags & VM_MERGEABLE))
1597			continue;
1598		if (ksm_scan.address < vma->vm_start)
1599			ksm_scan.address = vma->vm_start;
1600		if (!vma->anon_vma)
1601			ksm_scan.address = vma->vm_end;
1602
1603		while (ksm_scan.address < vma->vm_end) {
1604			if (ksm_test_exit(mm))
1605				break;
1606			*page = follow_page(vma, ksm_scan.address, FOLL_GET);
1607			if (IS_ERR_OR_NULL(*page)) {
1608				ksm_scan.address += PAGE_SIZE;
1609				cond_resched();
1610				continue;
1611			}
1612			if (PageAnon(*page)) {
 
1613				flush_anon_page(vma, *page, ksm_scan.address);
1614				flush_dcache_page(*page);
1615				rmap_item = get_next_rmap_item(slot,
1616					ksm_scan.rmap_list, ksm_scan.address);
1617				if (rmap_item) {
1618					ksm_scan.rmap_list =
1619							&rmap_item->rmap_list;
1620					ksm_scan.address += PAGE_SIZE;
1621				} else
1622					put_page(*page);
1623				up_read(&mm->mmap_sem);
1624				return rmap_item;
1625			}
1626			put_page(*page);
1627			ksm_scan.address += PAGE_SIZE;
1628			cond_resched();
1629		}
1630	}
1631
1632	if (ksm_test_exit(mm)) {
1633		ksm_scan.address = 0;
1634		ksm_scan.rmap_list = &slot->rmap_list;
1635	}
1636	/*
1637	 * Nuke all the rmap_items that are above this current rmap:
1638	 * because there were no VM_MERGEABLE vmas with such addresses.
1639	 */
1640	remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1641
1642	spin_lock(&ksm_mmlist_lock);
1643	ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1644						struct mm_slot, mm_list);
1645	if (ksm_scan.address == 0) {
1646		/*
1647		 * We've completed a full scan of all vmas, holding mmap_sem
1648		 * throughout, and found no VM_MERGEABLE: so do the same as
1649		 * __ksm_exit does to remove this mm from all our lists now.
1650		 * This applies either when cleaning up after __ksm_exit
1651		 * (but beware: we can reach here even before __ksm_exit),
1652		 * or when all VM_MERGEABLE areas have been unmapped (and
1653		 * mmap_sem then protects against race with MADV_MERGEABLE).
1654		 */
1655		hash_del(&slot->link);
1656		list_del(&slot->mm_list);
1657		spin_unlock(&ksm_mmlist_lock);
1658
1659		free_mm_slot(slot);
1660		clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1661		up_read(&mm->mmap_sem);
1662		mmdrop(mm);
1663	} else {
1664		up_read(&mm->mmap_sem);
1665		/*
1666		 * up_read(&mm->mmap_sem) first because after
1667		 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
1668		 * already have been freed under us by __ksm_exit()
1669		 * because the "mm_slot" is still hashed and
1670		 * ksm_scan.mm_slot doesn't point to it anymore.
1671		 */
1672		spin_unlock(&ksm_mmlist_lock);
 
1673	}
1674
1675	/* Repeat until we've completed scanning the whole list */
1676	slot = ksm_scan.mm_slot;
1677	if (slot != &ksm_mm_head)
1678		goto next_mm;
1679
1680	ksm_scan.seqnr++;
1681	return NULL;
1682}
1683
1684/**
1685 * ksm_do_scan  - the ksm scanner main worker function.
1686 * @scan_npages - number of pages we want to scan before we return.
1687 */
1688static void ksm_do_scan(unsigned int scan_npages)
1689{
1690	struct rmap_item *rmap_item;
1691	struct page *uninitialized_var(page);
1692
1693	while (scan_npages-- && likely(!freezing(current))) {
1694		cond_resched();
1695		rmap_item = scan_get_next_rmap_item(&page);
1696		if (!rmap_item)
1697			return;
1698		cmp_and_merge_page(page, rmap_item);
1699		put_page(page);
1700	}
1701}
1702
1703static int ksmd_should_run(void)
1704{
1705	return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1706}
1707
1708static int ksm_scan_thread(void *nothing)
1709{
1710	set_freezable();
1711	set_user_nice(current, 5);
1712
1713	while (!kthread_should_stop()) {
1714		mutex_lock(&ksm_thread_mutex);
1715		wait_while_offlining();
1716		if (ksmd_should_run())
1717			ksm_do_scan(ksm_thread_pages_to_scan);
1718		mutex_unlock(&ksm_thread_mutex);
1719
1720		try_to_freeze();
1721
1722		if (ksmd_should_run()) {
1723			schedule_timeout_interruptible(
1724				msecs_to_jiffies(ksm_thread_sleep_millisecs));
1725		} else {
1726			wait_event_freezable(ksm_thread_wait,
1727				ksmd_should_run() || kthread_should_stop());
1728		}
1729	}
1730	return 0;
1731}
1732
1733int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1734		unsigned long end, int advice, unsigned long *vm_flags)
1735{
1736	struct mm_struct *mm = vma->vm_mm;
1737	int err;
1738
1739	switch (advice) {
1740	case MADV_MERGEABLE:
1741		/*
1742		 * Be somewhat over-protective for now!
1743		 */
1744		if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1745				 VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1746				 VM_HUGETLB | VM_MIXEDMAP))
1747			return 0;		/* just ignore the advice */
1748
1749#ifdef VM_SAO
1750		if (*vm_flags & VM_SAO)
1751			return 0;
1752#endif
1753
1754		if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1755			err = __ksm_enter(mm);
1756			if (err)
1757				return err;
1758		}
1759
1760		*vm_flags |= VM_MERGEABLE;
1761		break;
1762
1763	case MADV_UNMERGEABLE:
1764		if (!(*vm_flags & VM_MERGEABLE))
1765			return 0;		/* just ignore the advice */
1766
1767		if (vma->anon_vma) {
1768			err = unmerge_ksm_pages(vma, start, end);
1769			if (err)
1770				return err;
1771		}
1772
1773		*vm_flags &= ~VM_MERGEABLE;
1774		break;
1775	}
1776
1777	return 0;
1778}
1779
1780int __ksm_enter(struct mm_struct *mm)
1781{
1782	struct mm_slot *mm_slot;
1783	int needs_wakeup;
1784
1785	mm_slot = alloc_mm_slot();
1786	if (!mm_slot)
1787		return -ENOMEM;
1788
1789	/* Check ksm_run too?  Would need tighter locking */
1790	needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1791
1792	spin_lock(&ksm_mmlist_lock);
1793	insert_to_mm_slots_hash(mm, mm_slot);
1794	/*
1795	 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1796	 * insert just behind the scanning cursor, to let the area settle
1797	 * down a little; when fork is followed by immediate exec, we don't
1798	 * want ksmd to waste time setting up and tearing down an rmap_list.
1799	 *
1800	 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1801	 * scanning cursor, otherwise KSM pages in newly forked mms will be
1802	 * missed: then we might as well insert at the end of the list.
1803	 */
1804	if (ksm_run & KSM_RUN_UNMERGE)
1805		list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1806	else
1807		list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1808	spin_unlock(&ksm_mmlist_lock);
1809
1810	set_bit(MMF_VM_MERGEABLE, &mm->flags);
1811	atomic_inc(&mm->mm_count);
1812
1813	if (needs_wakeup)
1814		wake_up_interruptible(&ksm_thread_wait);
1815
1816	return 0;
1817}
1818
1819void __ksm_exit(struct mm_struct *mm)
1820{
1821	struct mm_slot *mm_slot;
1822	int easy_to_free = 0;
1823
1824	/*
1825	 * This process is exiting: if it's straightforward (as is the
1826	 * case when ksmd was never running), free mm_slot immediately.
1827	 * But if it's at the cursor or has rmap_items linked to it, use
1828	 * mmap_sem to synchronize with any break_cows before pagetables
1829	 * are freed, and leave the mm_slot on the list for ksmd to free.
1830	 * Beware: ksm may already have noticed it exiting and freed the slot.
1831	 */
1832
1833	spin_lock(&ksm_mmlist_lock);
1834	mm_slot = get_mm_slot(mm);
1835	if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1836		if (!mm_slot->rmap_list) {
1837			hash_del(&mm_slot->link);
1838			list_del(&mm_slot->mm_list);
1839			easy_to_free = 1;
1840		} else {
1841			list_move(&mm_slot->mm_list,
1842				  &ksm_scan.mm_slot->mm_list);
1843		}
1844	}
1845	spin_unlock(&ksm_mmlist_lock);
1846
1847	if (easy_to_free) {
1848		free_mm_slot(mm_slot);
1849		clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1850		mmdrop(mm);
1851	} else if (mm_slot) {
1852		down_write(&mm->mmap_sem);
1853		up_write(&mm->mmap_sem);
1854	}
1855}
1856
1857struct page *ksm_might_need_to_copy(struct page *page,
1858			struct vm_area_struct *vma, unsigned long address)
1859{
1860	struct anon_vma *anon_vma = page_anon_vma(page);
1861	struct page *new_page;
1862
1863	if (PageKsm(page)) {
1864		if (page_stable_node(page) &&
1865		    !(ksm_run & KSM_RUN_UNMERGE))
1866			return page;	/* no need to copy it */
1867	} else if (!anon_vma) {
1868		return page;		/* no need to copy it */
1869	} else if (anon_vma->root == vma->anon_vma->root &&
1870		 page->index == linear_page_index(vma, address)) {
1871		return page;		/* still no need to copy it */
1872	}
1873	if (!PageUptodate(page))
1874		return page;		/* let do_swap_page report the error */
1875
1876	new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1877	if (new_page) {
1878		copy_user_highpage(new_page, page, address, vma);
1879
1880		SetPageDirty(new_page);
1881		__SetPageUptodate(new_page);
1882		__SetPageLocked(new_page);
1883	}
1884
1885	return new_page;
1886}
1887
1888int rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
1889{
1890	struct stable_node *stable_node;
1891	struct rmap_item *rmap_item;
1892	int ret = SWAP_AGAIN;
1893	int search_new_forks = 0;
1894
1895	VM_BUG_ON_PAGE(!PageKsm(page), page);
1896
1897	/*
1898	 * Rely on the page lock to protect against concurrent modifications
1899	 * to that page's node of the stable tree.
1900	 */
1901	VM_BUG_ON_PAGE(!PageLocked(page), page);
1902
1903	stable_node = page_stable_node(page);
1904	if (!stable_node)
1905		return ret;
1906again:
1907	hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
1908		struct anon_vma *anon_vma = rmap_item->anon_vma;
1909		struct anon_vma_chain *vmac;
1910		struct vm_area_struct *vma;
1911
1912		cond_resched();
1913		anon_vma_lock_read(anon_vma);
1914		anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1915					       0, ULONG_MAX) {
1916			cond_resched();
1917			vma = vmac->vma;
1918			if (rmap_item->address < vma->vm_start ||
1919			    rmap_item->address >= vma->vm_end)
1920				continue;
1921			/*
1922			 * Initially we examine only the vma which covers this
1923			 * rmap_item; but later, if there is still work to do,
1924			 * we examine covering vmas in other mms: in case they
1925			 * were forked from the original since ksmd passed.
1926			 */
1927			if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1928				continue;
1929
1930			if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1931				continue;
1932
1933			ret = rwc->rmap_one(page, vma,
1934					rmap_item->address, rwc->arg);
1935			if (ret != SWAP_AGAIN) {
1936				anon_vma_unlock_read(anon_vma);
1937				goto out;
1938			}
1939			if (rwc->done && rwc->done(page)) {
1940				anon_vma_unlock_read(anon_vma);
1941				goto out;
1942			}
1943		}
1944		anon_vma_unlock_read(anon_vma);
1945	}
1946	if (!search_new_forks++)
1947		goto again;
1948out:
1949	return ret;
1950}
1951
1952#ifdef CONFIG_MIGRATION
1953void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1954{
1955	struct stable_node *stable_node;
1956
1957	VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
1958	VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1959	VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
1960
1961	stable_node = page_stable_node(newpage);
1962	if (stable_node) {
1963		VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
1964		stable_node->kpfn = page_to_pfn(newpage);
1965		/*
1966		 * newpage->mapping was set in advance; now we need smp_wmb()
1967		 * to make sure that the new stable_node->kpfn is visible
1968		 * to get_ksm_page() before it can see that oldpage->mapping
1969		 * has gone stale (or that PageSwapCache has been cleared).
1970		 */
1971		smp_wmb();
1972		set_page_stable_node(oldpage, NULL);
1973	}
1974}
1975#endif /* CONFIG_MIGRATION */
1976
1977#ifdef CONFIG_MEMORY_HOTREMOVE
 
 
 
 
 
 
1978static void wait_while_offlining(void)
1979{
1980	while (ksm_run & KSM_RUN_OFFLINE) {
1981		mutex_unlock(&ksm_thread_mutex);
1982		wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
1983			    TASK_UNINTERRUPTIBLE);
1984		mutex_lock(&ksm_thread_mutex);
1985	}
1986}
1987
1988static void ksm_check_stable_tree(unsigned long start_pfn,
1989				  unsigned long end_pfn)
1990{
1991	struct stable_node *stable_node, *next;
 
1992	struct rb_node *node;
1993	int nid;
1994
1995	for (nid = 0; nid < ksm_nr_node_ids; nid++) {
1996		node = rb_first(root_stable_tree + nid);
1997		while (node) {
1998			stable_node = rb_entry(node, struct stable_node, node);
1999			if (stable_node->kpfn >= start_pfn &&
2000			    stable_node->kpfn < end_pfn) {
2001				/*
2002				 * Don't get_ksm_page, page has already gone:
2003				 * which is why we keep kpfn instead of page*
2004				 */
2005				remove_node_from_stable_tree(stable_node);
2006				node = rb_first(root_stable_tree + nid);
2007			} else
2008				node = rb_next(node);
2009			cond_resched();
2010		}
2011	}
2012	list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
 
2013		if (stable_node->kpfn >= start_pfn &&
2014		    stable_node->kpfn < end_pfn)
2015			remove_node_from_stable_tree(stable_node);
2016		cond_resched();
2017	}
2018}
2019
2020static int ksm_memory_callback(struct notifier_block *self,
2021			       unsigned long action, void *arg)
2022{
2023	struct memory_notify *mn = arg;
2024
2025	switch (action) {
2026	case MEM_GOING_OFFLINE:
2027		/*
2028		 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2029		 * and remove_all_stable_nodes() while memory is going offline:
2030		 * it is unsafe for them to touch the stable tree at this time.
2031		 * But unmerge_ksm_pages(), rmap lookups and other entry points
2032		 * which do not need the ksm_thread_mutex are all safe.
2033		 */
2034		mutex_lock(&ksm_thread_mutex);
2035		ksm_run |= KSM_RUN_OFFLINE;
2036		mutex_unlock(&ksm_thread_mutex);
2037		break;
2038
2039	case MEM_OFFLINE:
2040		/*
2041		 * Most of the work is done by page migration; but there might
2042		 * be a few stable_nodes left over, still pointing to struct
2043		 * pages which have been offlined: prune those from the tree,
2044		 * otherwise get_ksm_page() might later try to access a
2045		 * non-existent struct page.
2046		 */
2047		ksm_check_stable_tree(mn->start_pfn,
2048				      mn->start_pfn + mn->nr_pages);
2049		/* fallthrough */
2050
2051	case MEM_CANCEL_OFFLINE:
2052		mutex_lock(&ksm_thread_mutex);
2053		ksm_run &= ~KSM_RUN_OFFLINE;
2054		mutex_unlock(&ksm_thread_mutex);
2055
2056		smp_mb();	/* wake_up_bit advises this */
2057		wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2058		break;
2059	}
2060	return NOTIFY_OK;
2061}
2062#else
2063static void wait_while_offlining(void)
2064{
2065}
2066#endif /* CONFIG_MEMORY_HOTREMOVE */
2067
2068#ifdef CONFIG_SYSFS
2069/*
2070 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2071 */
2072
2073#define KSM_ATTR_RO(_name) \
2074	static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2075#define KSM_ATTR(_name) \
2076	static struct kobj_attribute _name##_attr = \
2077		__ATTR(_name, 0644, _name##_show, _name##_store)
2078
2079static ssize_t sleep_millisecs_show(struct kobject *kobj,
2080				    struct kobj_attribute *attr, char *buf)
2081{
2082	return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2083}
2084
2085static ssize_t sleep_millisecs_store(struct kobject *kobj,
2086				     struct kobj_attribute *attr,
2087				     const char *buf, size_t count)
2088{
2089	unsigned long msecs;
2090	int err;
2091
2092	err = kstrtoul(buf, 10, &msecs);
2093	if (err || msecs > UINT_MAX)
2094		return -EINVAL;
2095
2096	ksm_thread_sleep_millisecs = msecs;
2097
2098	return count;
2099}
2100KSM_ATTR(sleep_millisecs);
2101
2102static ssize_t pages_to_scan_show(struct kobject *kobj,
2103				  struct kobj_attribute *attr, char *buf)
2104{
2105	return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2106}
2107
2108static ssize_t pages_to_scan_store(struct kobject *kobj,
2109				   struct kobj_attribute *attr,
2110				   const char *buf, size_t count)
2111{
2112	int err;
2113	unsigned long nr_pages;
2114
2115	err = kstrtoul(buf, 10, &nr_pages);
2116	if (err || nr_pages > UINT_MAX)
2117		return -EINVAL;
2118
2119	ksm_thread_pages_to_scan = nr_pages;
2120
2121	return count;
2122}
2123KSM_ATTR(pages_to_scan);
2124
2125static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2126			char *buf)
2127{
2128	return sprintf(buf, "%lu\n", ksm_run);
2129}
2130
2131static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2132			 const char *buf, size_t count)
2133{
2134	int err;
2135	unsigned long flags;
2136
2137	err = kstrtoul(buf, 10, &flags);
2138	if (err || flags > UINT_MAX)
2139		return -EINVAL;
2140	if (flags > KSM_RUN_UNMERGE)
2141		return -EINVAL;
2142
2143	/*
2144	 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2145	 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2146	 * breaking COW to free the pages_shared (but leaves mm_slots
2147	 * on the list for when ksmd may be set running again).
2148	 */
2149
2150	mutex_lock(&ksm_thread_mutex);
2151	wait_while_offlining();
2152	if (ksm_run != flags) {
2153		ksm_run = flags;
2154		if (flags & KSM_RUN_UNMERGE) {
2155			set_current_oom_origin();
2156			err = unmerge_and_remove_all_rmap_items();
2157			clear_current_oom_origin();
2158			if (err) {
2159				ksm_run = KSM_RUN_STOP;
2160				count = err;
2161			}
2162		}
2163	}
2164	mutex_unlock(&ksm_thread_mutex);
2165
2166	if (flags & KSM_RUN_MERGE)
2167		wake_up_interruptible(&ksm_thread_wait);
2168
2169	return count;
2170}
2171KSM_ATTR(run);
2172
2173#ifdef CONFIG_NUMA
2174static ssize_t merge_across_nodes_show(struct kobject *kobj,
2175				struct kobj_attribute *attr, char *buf)
2176{
2177	return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2178}
2179
2180static ssize_t merge_across_nodes_store(struct kobject *kobj,
2181				   struct kobj_attribute *attr,
2182				   const char *buf, size_t count)
2183{
2184	int err;
2185	unsigned long knob;
2186
2187	err = kstrtoul(buf, 10, &knob);
2188	if (err)
2189		return err;
2190	if (knob > 1)
2191		return -EINVAL;
2192
2193	mutex_lock(&ksm_thread_mutex);
2194	wait_while_offlining();
2195	if (ksm_merge_across_nodes != knob) {
2196		if (ksm_pages_shared || remove_all_stable_nodes())
2197			err = -EBUSY;
2198		else if (root_stable_tree == one_stable_tree) {
2199			struct rb_root *buf;
2200			/*
2201			 * This is the first time that we switch away from the
2202			 * default of merging across nodes: must now allocate
2203			 * a buffer to hold as many roots as may be needed.
2204			 * Allocate stable and unstable together:
2205			 * MAXSMP NODES_SHIFT 10 will use 16kB.
2206			 */
2207			buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2208				      GFP_KERNEL);
2209			/* Let us assume that RB_ROOT is NULL is zero */
2210			if (!buf)
2211				err = -ENOMEM;
2212			else {
2213				root_stable_tree = buf;
2214				root_unstable_tree = buf + nr_node_ids;
2215				/* Stable tree is empty but not the unstable */
2216				root_unstable_tree[0] = one_unstable_tree[0];
2217			}
2218		}
2219		if (!err) {
2220			ksm_merge_across_nodes = knob;
2221			ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2222		}
2223	}
2224	mutex_unlock(&ksm_thread_mutex);
2225
2226	return err ? err : count;
2227}
2228KSM_ATTR(merge_across_nodes);
2229#endif
2230
2231static ssize_t pages_shared_show(struct kobject *kobj,
2232				 struct kobj_attribute *attr, char *buf)
2233{
2234	return sprintf(buf, "%lu\n", ksm_pages_shared);
2235}
2236KSM_ATTR_RO(pages_shared);
2237
2238static ssize_t pages_sharing_show(struct kobject *kobj,
2239				  struct kobj_attribute *attr, char *buf)
2240{
2241	return sprintf(buf, "%lu\n", ksm_pages_sharing);
2242}
2243KSM_ATTR_RO(pages_sharing);
2244
2245static ssize_t pages_unshared_show(struct kobject *kobj,
2246				   struct kobj_attribute *attr, char *buf)
2247{
2248	return sprintf(buf, "%lu\n", ksm_pages_unshared);
2249}
2250KSM_ATTR_RO(pages_unshared);
2251
2252static ssize_t pages_volatile_show(struct kobject *kobj,
2253				   struct kobj_attribute *attr, char *buf)
2254{
2255	long ksm_pages_volatile;
2256
2257	ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2258				- ksm_pages_sharing - ksm_pages_unshared;
2259	/*
2260	 * It was not worth any locking to calculate that statistic,
2261	 * but it might therefore sometimes be negative: conceal that.
2262	 */
2263	if (ksm_pages_volatile < 0)
2264		ksm_pages_volatile = 0;
2265	return sprintf(buf, "%ld\n", ksm_pages_volatile);
2266}
2267KSM_ATTR_RO(pages_volatile);
2268
2269static ssize_t full_scans_show(struct kobject *kobj,
2270			       struct kobj_attribute *attr, char *buf)
2271{
2272	return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2273}
2274KSM_ATTR_RO(full_scans);
2275
2276static struct attribute *ksm_attrs[] = {
2277	&sleep_millisecs_attr.attr,
2278	&pages_to_scan_attr.attr,
2279	&run_attr.attr,
2280	&pages_shared_attr.attr,
2281	&pages_sharing_attr.attr,
2282	&pages_unshared_attr.attr,
2283	&pages_volatile_attr.attr,
2284	&full_scans_attr.attr,
2285#ifdef CONFIG_NUMA
2286	&merge_across_nodes_attr.attr,
2287#endif
2288	NULL,
2289};
2290
2291static struct attribute_group ksm_attr_group = {
2292	.attrs = ksm_attrs,
2293	.name = "ksm",
2294};
2295#endif /* CONFIG_SYSFS */
2296
2297static int __init ksm_init(void)
2298{
2299	struct task_struct *ksm_thread;
2300	int err;
2301
2302	err = ksm_slab_init();
2303	if (err)
2304		goto out;
2305
2306	ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2307	if (IS_ERR(ksm_thread)) {
2308		pr_err("ksm: creating kthread failed\n");
2309		err = PTR_ERR(ksm_thread);
2310		goto out_free;
2311	}
2312
2313#ifdef CONFIG_SYSFS
2314	err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2315	if (err) {
2316		pr_err("ksm: register sysfs failed\n");
2317		kthread_stop(ksm_thread);
2318		goto out_free;
2319	}
2320#else
2321	ksm_run = KSM_RUN_MERGE;	/* no way for user to start it */
2322
2323#endif /* CONFIG_SYSFS */
2324
2325#ifdef CONFIG_MEMORY_HOTREMOVE
2326	/* There is no significance to this priority 100 */
2327	hotplug_memory_notifier(ksm_memory_callback, 100);
2328#endif
2329	return 0;
2330
2331out_free:
2332	ksm_slab_free();
2333out:
2334	return err;
2335}
2336subsys_initcall(ksm_init);