mm/ksm.c

   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/hash.h>
  37 #include <linux/freezer.h>
  38 #include <linux/oom.h>
  39
  40 #include <asm/tlbflush.h>
  41 #include "internal.h"
  42
  43 /*
  44  * A few notes about the KSM scanning process,
  45  * to make it easier to understand the data structures below:
  46  *
  47  * In order to reduce excessive scanning, KSM sorts the memory pages by their
  48  * contents into a data structure that holds pointers to the pages' locations.
  49  *
  50  * Since the contents of the pages may change at any moment, KSM cannot just
  51  * insert the pages into a normal sorted tree and expect it to find anything.
  52  * Therefore KSM uses two data structures - the stable and the unstable tree.
  53  *
  54  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
  55  * by their contents.  Because each such page is write-protected, searching on
  56  * this tree is fully assured to be working (except when pages are unmapped),
  57  * and therefore this tree is called the stable tree.
  58  *
  59  * In addition to the stable tree, KSM uses a second data structure called the
  60  * unstable tree: this tree holds pointers to pages which have been found to
  61  * be "unchanged for a period of time".  The unstable tree sorts these pages
  62  * by their contents, but since they are not write-protected, KSM cannot rely
  63  * upon the unstable tree to work correctly - the unstable tree is liable to
  64  * be corrupted as its contents are modified, and so it is called unstable.
  65  *
  66  * KSM solves this problem by several techniques:
  67  *
  68  * 1) The unstable tree is flushed every time KSM completes scanning all
  69  *    memory areas, and then the tree is rebuilt again from the beginning.
  70  * 2) KSM will only insert into the unstable tree, pages whose hash value
  71  *    has not changed since the previous scan of all memory areas.
  72  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
  73  *    colors of the nodes and not on their contents, assuring that even when
  74  *    the tree gets "corrupted" it won't get out of balance, so scanning time
  75  *    remains the same (also, searching and inserting nodes in an rbtree uses
  76  *    the same algorithm, so we have no overhead when we flush and rebuild).
  77  * 4) KSM never flushes the stable tree, which means that even if it were to
  78  *    take 10 attempts to find a page in the unstable tree, once it is found,
  79  *    it is secured in the stable tree.  (When we scan a new page, we first
  80  *    compare it against the stable tree, and then against the unstable tree.)
  81  */
  82
  83 /**
  84  * struct mm_slot - ksm information per mm that is being scanned
  85  * @link: link to the mm_slots hash list
  86  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
  87  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
  88  * @mm: the mm that this information is valid for
  89  */
  90 struct mm_slot {
  91         struct hlist_node link;
  92         struct list_head mm_list;
  93         struct rmap_item *rmap_list;
  94         struct mm_struct *mm;
  95 };
  96
  97 /**
  98  * struct ksm_scan - cursor for scanning
  99  * @mm_slot: the current mm_slot we are scanning
 100  * @address: the next address inside that to be scanned
 101  * @rmap_list: link to the next rmap to be scanned in the rmap_list
 102  * @seqnr: count of completed full scans (needed when removing unstable node)
 103  *
 104  * There is only the one ksm_scan instance of this cursor structure.
 105  */
 106 struct ksm_scan {
 107         struct mm_slot *mm_slot;
 108         unsigned long address;
 109         struct rmap_item **rmap_list;
 110         unsigned long seqnr;
 111 };
 112
 113 /**
 114  * struct stable_node - node of the stable rbtree
 115  * @node: rb node of this ksm page in the stable tree
 116  * @hlist: hlist head of rmap_items using this ksm page
 117  * @kpfn: page frame number of this ksm page
 118  */
 119 struct stable_node {
 120         struct rb_node node;
 121         struct hlist_head hlist;
 122         unsigned long kpfn;
 123 };
 124
 125 /**
 126  * struct rmap_item - reverse mapping item for virtual addresses
 127  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
 128  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
 129  * @mm: the memory structure this rmap_item is pointing into
 130  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
 131  * @oldchecksum: previous checksum of the page at that virtual address
 132  * @node: rb node of this rmap_item in the unstable tree
 133  * @head: pointer to stable_node heading this list in the stable tree
 134  * @hlist: link into hlist of rmap_items hanging off that stable_node
 135  */
 136 struct rmap_item {
 137         struct rmap_item *rmap_list;
 138         struct anon_vma *anon_vma;      /* when stable */
 139         struct mm_struct *mm;
 140         unsigned long address;          /* + low bits used for flags below */
 141         unsigned int oldchecksum;       /* when unstable */
 142         union {
 143                 struct rb_node node;    /* when node of unstable tree */
 144                 struct {                /* when listed from stable tree */
 145                         struct stable_node *head;
 146                         struct hlist_node hlist;
 147                 };
 148         };
 149 };
 150
 151 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
 152 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
 153 #define STABLE_FLAG     0x200   /* is listed from the stable tree */
 154
 155 /* The stable and unstable tree heads */
 156 static struct rb_root root_stable_tree = RB_ROOT;
 157 static struct rb_root root_unstable_tree = RB_ROOT;
 158
 159 #define MM_SLOTS_HASH_SHIFT 10
 160 #define MM_SLOTS_HASH_HEADS (1 << MM_SLOTS_HASH_SHIFT)
 161 static struct hlist_head mm_slots_hash[MM_SLOTS_HASH_HEADS];
 162
 163 static struct mm_slot ksm_mm_head = {
 164         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
 165 };
 166 static struct ksm_scan ksm_scan = {
 167         .mm_slot = &ksm_mm_head,
 168 };
 169
 170 static struct kmem_cache *rmap_item_cache;
 171 static struct kmem_cache *stable_node_cache;
 172 static struct kmem_cache *mm_slot_cache;
 173
 174 /* The number of nodes in the stable tree */
 175 static unsigned long ksm_pages_shared;
 176
 177 /* The number of page slots additionally sharing those nodes */
 178 static unsigned long ksm_pages_sharing;
 179
 180 /* The number of nodes in the unstable tree */
 181 static unsigned long ksm_pages_unshared;
 182
 183 /* The number of rmap_items in use: to calculate pages_volatile */
 184 static unsigned long ksm_rmap_items;
 185
 186 /* Number of pages ksmd should scan in one batch */
 187 static unsigned int ksm_thread_pages_to_scan = 100;
 188
 189 /* Milliseconds ksmd should sleep between batches */
 190 static unsigned int ksm_thread_sleep_millisecs = 20;
 191
 192 #define KSM_RUN_STOP    0
 193 #define KSM_RUN_MERGE   1
 194 #define KSM_RUN_UNMERGE 2
 195 static unsigned int ksm_run = KSM_RUN_STOP;
 196
 197 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
 198 static DEFINE_MUTEX(ksm_thread_mutex);
 199 static DEFINE_SPINLOCK(ksm_mmlist_lock);
 200
 201 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
 202                 sizeof(struct __struct), __alignof__(struct __struct),\
 203                 (__flags), NULL)
 204
 205 static int __init ksm_slab_init(void)
 206 {
 207         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
 208         if (!rmap_item_cache)
 209                 goto out;
 210
 211         stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
 212         if (!stable_node_cache)
 213                 goto out_free1;
 214
 215         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
 216         if (!mm_slot_cache)
 217                 goto out_free2;
 218
 219         return 0;
 220
 221 out_free2:
 222         kmem_cache_destroy(stable_node_cache);
 223 out_free1:
 224         kmem_cache_destroy(rmap_item_cache);
 225 out:
 226         return -ENOMEM;
 227 }
 228
 229 static void __init ksm_slab_free(void)
 230 {
 231         kmem_cache_destroy(mm_slot_cache);
 232         kmem_cache_destroy(stable_node_cache);
 233         kmem_cache_destroy(rmap_item_cache);
 234         mm_slot_cache = NULL;
 235 }
 236
 237 static inline struct rmap_item *alloc_rmap_item(void)
 238 {
 239         struct rmap_item *rmap_item;
 240
 241         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
 242                                                 __GFP_NORETRY | __GFP_NOWARN);
 243         if (rmap_item)
 244                 ksm_rmap_items++;
 245         return rmap_item;
 246 }
 247
 248 static inline void free_rmap_item(struct rmap_item *rmap_item)
 249 {
 250         ksm_rmap_items--;
 251         rmap_item->mm = NULL;   /* debug safety */
 252         kmem_cache_free(rmap_item_cache, rmap_item);
 253 }
 254
 255 static inline struct stable_node *alloc_stable_node(void)
 256 {
 257         return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
 258 }
 259
 260 static inline void free_stable_node(struct stable_node *stable_node)
 261 {
 262         kmem_cache_free(stable_node_cache, stable_node);
 263 }
 264
 265 static inline struct mm_slot *alloc_mm_slot(void)
 266 {
 267         if (!mm_slot_cache)     /* initialization failed */
 268                 return NULL;
 269         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 270 }
 271
 272 static inline void free_mm_slot(struct mm_slot *mm_slot)
 273 {
 274         kmem_cache_free(mm_slot_cache, mm_slot);
 275 }
 276
 277 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 278 {
 279         struct mm_slot *mm_slot;
 280         struct hlist_head *bucket;
 281         struct hlist_node *node;
 282
 283         bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
 284         hlist_for_each_entry(mm_slot, node, bucket, link) {
 285                 if (mm == mm_slot->mm)
 286                         return mm_slot;
 287         }
 288         return NULL;
 289 }
 290
 291 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 292                                     struct mm_slot *mm_slot)
 293 {
 294         struct hlist_head *bucket;
 295
 296         bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
 297         mm_slot->mm = mm;
 298         hlist_add_head(&mm_slot->link, bucket);
 299 }
 300
 301 static inline int in_stable_tree(struct rmap_item *rmap_item)
 302 {
 303         return rmap_item->address & STABLE_FLAG;
 304 }
 305
 306 /*
 307  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
 308  * page tables after it has passed through ksm_exit() - which, if necessary,
 309  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
 310  * a special flag: they can just back out as soon as mm_users goes to zero.
 311  * ksm_test_exit() is used throughout to make this test for exit: in some
 312  * places for correctness, in some places just to avoid unnecessary work.
 313  */
 314 static inline bool ksm_test_exit(struct mm_struct *mm)
 315 {
 316         return atomic_read(&mm->mm_users) == 0;
 317 }
 318
 319 /*
 320  * We use break_ksm to break COW on a ksm page: it's a stripped down
 321  *
 322  *      if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
 323  *              put_page(page);
 324  *
 325  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
 326  * in case the application has unmapped and remapped mm,addr meanwhile.
 327  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
 328  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
 329  */
 330 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
 331 {
 332         struct page *page;
 333         int ret = 0;
 334
 335         do {
 336                 cond_resched();
 337                 page = follow_page(vma, addr, FOLL_GET);
 338                 if (IS_ERR_OR_NULL(page))
 339                         break;
 340                 if (PageKsm(page))
 341                         ret = handle_mm_fault(vma->vm_mm, vma, addr,
 342                                                         FAULT_FLAG_WRITE);
 343                 else
 344                         ret = VM_FAULT_WRITE;
 345                 put_page(page);
 346         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
 347         /*
 348          * We must loop because handle_mm_fault() may back out if there's
 349          * any difficulty e.g. if pte accessed bit gets updated concurrently.
 350          *
 351          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
 352          * COW has been broken, even if the vma does not permit VM_WRITE;
 353          * but note that a concurrent fault might break PageKsm for us.
 354          *
 355          * VM_FAULT_SIGBUS could occur if we race with truncation of the
 356          * backing file, which also invalidates anonymous pages: that's
 357          * okay, that truncation will have unmapped the PageKsm for us.
 358          *
 359          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
 360          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
 361          * current task has TIF_MEMDIE set, and will be OOM killed on return
 362          * to user; and ksmd, having no mm, would never be chosen for that.
 363          *
 364          * But if the mm is in a limited mem_cgroup, then the fault may fail
 365          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
 366          * even ksmd can fail in this way - though it's usually breaking ksm
 367          * just to undo a merge it made a moment before, so unlikely to oom.
 368          *
 369          * That's a pity: we might therefore have more kernel pages allocated
 370          * than we're counting as nodes in the stable tree; but ksm_do_scan
 371          * will retry to break_cow on each pass, so should recover the page
 372          * in due course.  The important thing is to not let VM_MERGEABLE
 373          * be cleared while any such pages might remain in the area.
 374          */
 375         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
 376 }
 377
 378 static void break_cow(struct rmap_item *rmap_item)
 379 {
 380         struct mm_struct *mm = rmap_item->mm;
 381         unsigned long addr = rmap_item->address;
 382         struct vm_area_struct *vma;
 383
 384         /*
 385          * It is not an accident that whenever we want to break COW
 386          * to undo, we also need to drop a reference to the anon_vma.
 387          */
 388         put_anon_vma(rmap_item->anon_vma);
 389
 390         down_read(&mm->mmap_sem);
 391         if (ksm_test_exit(mm))
 392                 goto out;
 393         vma = find_vma(mm, addr);
 394         if (!vma || vma->vm_start > addr)
 395                 goto out;
 396         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
 397                 goto out;
 398         break_ksm(vma, addr);
 399 out:
 400         up_read(&mm->mmap_sem);
 401 }
 402
 403 static struct page *page_trans_compound_anon(struct page *page)
 404 {
 405         if (PageTransCompound(page)) {
 406                 struct page *head = compound_trans_head(page);
 407                 /*
 408                  * head may actually be splitted and freed from under
 409                  * us but it's ok here.
 410                  */
 411                 if (PageAnon(head))
 412                         return head;
 413         }
 414         return NULL;
 415 }
 416
 417 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
 418 {
 419         struct mm_struct *mm = rmap_item->mm;
 420         unsigned long addr = rmap_item->address;
 421         struct vm_area_struct *vma;
 422         struct page *page;
 423
 424         down_read(&mm->mmap_sem);
 425         if (ksm_test_exit(mm))
 426                 goto out;
 427         vma = find_vma(mm, addr);
 428         if (!vma || vma->vm_start > addr)
 429                 goto out;
 430         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
 431                 goto out;
 432
 433         page = follow_page(vma, addr, FOLL_GET);
 434         if (IS_ERR_OR_NULL(page))
 435                 goto out;
 436         if (PageAnon(page) || page_trans_compound_anon(page)) {
 437                 flush_anon_page(vma, page, addr);
 438                 flush_dcache_page(page);
 439         } else {
 440                 put_page(page);
 441 out:            page = NULL;
 442         }
 443         up_read(&mm->mmap_sem);
 444         return page;
 445 }
 446
 447 static void remove_node_from_stable_tree(struct stable_node *stable_node)
 448 {
 449         struct rmap_item *rmap_item;
 450         struct hlist_node *hlist;
 451
 452         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
 453                 if (rmap_item->hlist.next)
 454                         ksm_pages_sharing--;
 455                 else
 456                         ksm_pages_shared--;
 457                 put_anon_vma(rmap_item->anon_vma);
 458                 rmap_item->address &= PAGE_MASK;
 459                 cond_resched();
 460         }
 461
 462         rb_erase(&stable_node->node, &root_stable_tree);
 463         free_stable_node(stable_node);
 464 }
 465
 466 /*
 467  * get_ksm_page: checks if the page indicated by the stable node
 468  * is still its ksm page, despite having held no reference to it.
 469  * In which case we can trust the content of the page, and it
 470  * returns the gotten page; but if the page has now been zapped,
 471  * remove the stale node from the stable tree and return NULL.
 472  *
 473  * You would expect the stable_node to hold a reference to the ksm page.
 474  * But if it increments the page's count, swapping out has to wait for
 475  * ksmd to come around again before it can free the page, which may take
 476  * seconds or even minutes: much too unresponsive.  So instead we use a
 477  * "keyhole reference": access to the ksm page from the stable node peeps
 478  * out through its keyhole to see if that page still holds the right key,
 479  * pointing back to this stable node.  This relies on freeing a PageAnon
 480  * page to reset its page->mapping to NULL, and relies on no other use of
 481  * a page to put something that might look like our key in page->mapping.
 482  *
 483  * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
 484  * but this is different - made simpler by ksm_thread_mutex being held, but
 485  * interesting for assuming that no other use of the struct page could ever
 486  * put our expected_mapping into page->mapping (or a field of the union which
 487  * coincides with page->mapping).  The RCU calls are not for KSM at all, but
 488  * to keep the page_count protocol described with page_cache_get_speculative.
 489  *
 490  * Note: it is possible that get_ksm_page() will return NULL one moment,
 491  * then page the next, if the page is in between page_freeze_refs() and
 492  * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
 493  * is on its way to being freed; but it is an anomaly to bear in mind.
 494  */
 495 static struct page *get_ksm_page(struct stable_node *stable_node)
 496 {
 497         struct page *page;
 498         void *expected_mapping;
 499
 500         page = pfn_to_page(stable_node->kpfn);
 501         expected_mapping = (void *)stable_node +
 502                                 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
 503         rcu_read_lock();
 504         if (page->mapping != expected_mapping)
 505                 goto stale;
 506         if (!get_page_unless_zero(page))
 507                 goto stale;
 508         if (page->mapping != expected_mapping) {
 509                 put_page(page);
 510                 goto stale;
 511         }
 512         rcu_read_unlock();
 513         return page;
 514 stale:
 515         rcu_read_unlock();
 516         remove_node_from_stable_tree(stable_node);
 517         return NULL;
 518 }
 519
 520 /*
 521  * Removing rmap_item from stable or unstable tree.
 522  * This function will clean the information from the stable/unstable tree.
 523  */
 524 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
 525 {
 526         if (rmap_item->address & STABLE_FLAG) {
 527                 struct stable_node *stable_node;
 528                 struct page *page;
 529
 530                 stable_node = rmap_item->head;
 531                 page = get_ksm_page(stable_node);
 532                 if (!page)
 533                         goto out;
 534
 535                 lock_page(page);
 536                 hlist_del(&rmap_item->hlist);
 537                 unlock_page(page);
 538                 put_page(page);
 539
 540                 if (stable_node->hlist.first)
 541                         ksm_pages_sharing--;
 542                 else
 543                         ksm_pages_shared--;
 544
 545                 put_anon_vma(rmap_item->anon_vma);
 546                 rmap_item->address &= PAGE_MASK;
 547
 548         } else if (rmap_item->address & UNSTABLE_FLAG) {
 549                 unsigned char age;
 550                 /*
 551                  * Usually ksmd can and must skip the rb_erase, because
 552                  * root_unstable_tree was already reset to RB_ROOT.
 553                  * But be careful when an mm is exiting: do the rb_erase
 554                  * if this rmap_item was inserted by this scan, rather
 555                  * than left over from before.
 556                  */
 557                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
 558                 BUG_ON(age > 1);
 559                 if (!age)
 560                         rb_erase(&rmap_item->node, &root_unstable_tree);
 561
 562                 ksm_pages_unshared--;
 563                 rmap_item->address &= PAGE_MASK;
 564         }
 565 out:
 566         cond_resched();         /* we're called from many long loops */
 567 }
 568
 569 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
 570                                        struct rmap_item **rmap_list)
 571 {
 572         while (*rmap_list) {
 573                 struct rmap_item *rmap_item = *rmap_list;
 574                 *rmap_list = rmap_item->rmap_list;
 575                 remove_rmap_item_from_tree(rmap_item);
 576                 free_rmap_item(rmap_item);
 577         }
 578 }
 579
 580 /*
 581  * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
 582  * than check every pte of a given vma, the locking doesn't quite work for
 583  * that - an rmap_item is assigned to the stable tree after inserting ksm
 584  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
 585  * rmap_items from parent to child at fork time (so as not to waste time
 586  * if exit comes before the next scan reaches it).
 587  *
 588  * Similarly, although we'd like to remove rmap_items (so updating counts
 589  * and freeing memory) when unmerging an area, it's easier to leave that
 590  * to the next pass of ksmd - consider, for example, how ksmd might be
 591  * in cmp_and_merge_page on one of the rmap_items we would be removing.
 592  */
 593 static int unmerge_ksm_pages(struct vm_area_struct *vma,
 594                              unsigned long start, unsigned long end)
 595 {
 596         unsigned long addr;
 597         int err = 0;
 598
 599         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
 600                 if (ksm_test_exit(vma->vm_mm))
 601                         break;
 602                 if (signal_pending(current))
 603                         err = -ERESTARTSYS;
 604                 else
 605                         err = break_ksm(vma, addr);
 606         }
 607         return err;
 608 }
 609
 610 #ifdef CONFIG_SYSFS
 611 /*
 612  * Only called through the sysfs control interface:
 613  */
 614 static int unmerge_and_remove_all_rmap_items(void)
 615 {
 616         struct mm_slot *mm_slot;
 617         struct mm_struct *mm;
 618         struct vm_area_struct *vma;
 619         int err = 0;
 620
 621         spin_lock(&ksm_mmlist_lock);
 622         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
 623                                                 struct mm_slot, mm_list);
 624         spin_unlock(&ksm_mmlist_lock);
 625
 626         for (mm_slot = ksm_scan.mm_slot;
 627                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
 628                 mm = mm_slot->mm;
 629                 down_read(&mm->mmap_sem);
 630                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
 631                         if (ksm_test_exit(mm))
 632                                 break;
 633                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
 634                                 continue;
 635                         err = unmerge_ksm_pages(vma,
 636                                                 vma->vm_start, vma->vm_end);
 637                         if (err)
 638                                 goto error;
 639                 }
 640
 641                 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
 642
 643                 spin_lock(&ksm_mmlist_lock);
 644                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
 645                                                 struct mm_slot, mm_list);
 646                 if (ksm_test_exit(mm)) {
 647                         hlist_del(&mm_slot->link);
 648                         list_del(&mm_slot->mm_list);
 649                         spin_unlock(&ksm_mmlist_lock);
 650
 651                         free_mm_slot(mm_slot);
 652                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
 653                         up_read(&mm->mmap_sem);
 654                         mmdrop(mm);
 655                 } else {
 656                         spin_unlock(&ksm_mmlist_lock);
 657                         up_read(&mm->mmap_sem);
 658                 }
 659         }
 660
 661         ksm_scan.seqnr = 0;
 662         return 0;
 663
 664 error:
 665         up_read(&mm->mmap_sem);
 666         spin_lock(&ksm_mmlist_lock);
 667         ksm_scan.mm_slot = &ksm_mm_head;
 668         spin_unlock(&ksm_mmlist_lock);
 669         return err;
 670 }
 671 #endif /* CONFIG_SYSFS */
 672
 673 static u32 calc_checksum(struct page *page)
 674 {
 675         u32 checksum;
 676         void *addr = kmap_atomic(page, KM_USER0);
 677         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
 678         kunmap_atomic(addr, KM_USER0);
 679         return checksum;
 680 }
 681
 682 static int memcmp_pages(struct page *page1, struct page *page2)
 683 {
 684         char *addr1, *addr2;
 685         int ret;
 686
 687         addr1 = kmap_atomic(page1, KM_USER0);
 688         addr2 = kmap_atomic(page2, KM_USER1);
 689         ret = memcmp(addr1, addr2, PAGE_SIZE);
 690         kunmap_atomic(addr2, KM_USER1);
 691         kunmap_atomic(addr1, KM_USER0);
 692         return ret;
 693 }
 694
 695 static inline int pages_identical(struct page *page1, struct page *page2)
 696 {
 697         return !memcmp_pages(page1, page2);
 698 }
 699
 700 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
 701                               pte_t *orig_pte)
 702 {
 703         struct mm_struct *mm = vma->vm_mm;
 704         unsigned long addr;
 705         pte_t *ptep;
 706         spinlock_t *ptl;
 707         int swapped;
 708         int err = -EFAULT;
 709
 710         addr = page_address_in_vma(page, vma);
 711         if (addr == -EFAULT)
 712                 goto out;
 713
 714         BUG_ON(PageTransCompound(page));
 715         ptep = page_check_address(page, mm, addr, &ptl, 0);
 716         if (!ptep)
 717                 goto out;
 718
 719         if (pte_write(*ptep) || pte_dirty(*ptep)) {
 720                 pte_t entry;
 721
 722                 swapped = PageSwapCache(page);
 723                 flush_cache_page(vma, addr, page_to_pfn(page));
 724                 /*
 725                  * Ok this is tricky, when get_user_pages_fast() run it doesn't
 726                  * take any lock, therefore the check that we are going to make
 727                  * with the pagecount against the mapcount is racey and
 728                  * O_DIRECT can happen right after the check.
 729                  * So we clear the pte and flush the tlb before the check
 730                  * this assure us that no O_DIRECT can happen after the check
 731                  * or in the middle of the check.
 732                  */
 733                 entry = ptep_clear_flush(vma, addr, ptep);
 734                 /*
 735                  * Check that no O_DIRECT or similar I/O is in progress on the
 736                  * page
 737                  */
 738                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
 739                         set_pte_at(mm, addr, ptep, entry);
 740                         goto out_unlock;
 741                 }
 742                 if (pte_dirty(entry))
 743                         set_page_dirty(page);
 744                 entry = pte_mkclean(pte_wrprotect(entry));
 745                 set_pte_at_notify(mm, addr, ptep, entry);
 746         }
 747         *orig_pte = *ptep;
 748         err = 0;
 749
 750 out_unlock:
 751         pte_unmap_unlock(ptep, ptl);
 752 out:
 753         return err;
 754 }
 755
 756 /**
 757  * replace_page - replace page in vma by new ksm page
 758  * @vma:      vma that holds the pte pointing to page
 759  * @page:     the page we are replacing by kpage
 760  * @kpage:    the ksm page we replace page by
 761  * @orig_pte: the original value of the pte
 762  *
 763  * Returns 0 on success, -EFAULT on failure.
 764  */
 765 static int replace_page(struct vm_area_struct *vma, struct page *page,
 766                         struct page *kpage, pte_t orig_pte)
 767 {
 768         struct mm_struct *mm = vma->vm_mm;
 769         pgd_t *pgd;
 770         pud_t *pud;
 771         pmd_t *pmd;
 772         pte_t *ptep;
 773         spinlock_t *ptl;
 774         unsigned long addr;
 775         int err = -EFAULT;
 776
 777         addr = page_address_in_vma(page, vma);
 778         if (addr == -EFAULT)
 779                 goto out;
 780
 781         pgd = pgd_offset(mm, addr);
 782         if (!pgd_present(*pgd))
 783                 goto out;
 784
 785         pud = pud_offset(pgd, addr);
 786         if (!pud_present(*pud))
 787                 goto out;
 788
 789         pmd = pmd_offset(pud, addr);
 790         BUG_ON(pmd_trans_huge(*pmd));
 791         if (!pmd_present(*pmd))
 792                 goto out;
 793
 794         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
 795         if (!pte_same(*ptep, orig_pte)) {
 796                 pte_unmap_unlock(ptep, ptl);
 797                 goto out;
 798         }
 799
 800         get_page(kpage);
 801         page_add_anon_rmap(kpage, vma, addr);
 802
 803         flush_cache_page(vma, addr, pte_pfn(*ptep));
 804         ptep_clear_flush(vma, addr, ptep);
 805         set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
 806
 807         page_remove_rmap(page);
 808         if (!page_mapped(page))
 809                 try_to_free_swap(page);
 810         put_page(page);
 811
 812         pte_unmap_unlock(ptep, ptl);
 813         err = 0;
 814 out:
 815         return err;
 816 }
 817
 818 static int page_trans_compound_anon_split(struct page *page)
 819 {
 820         int ret = 0;
 821         struct page *transhuge_head = page_trans_compound_anon(page);
 822         if (transhuge_head) {
 823                 /* Get the reference on the head to split it. */
 824                 if (get_page_unless_zero(transhuge_head)) {
 825                         /*
 826                          * Recheck we got the reference while the head
 827                          * was still anonymous.
 828                          */
 829                         if (PageAnon(transhuge_head))
 830                                 ret = split_huge_page(transhuge_head);
 831                         else
 832                                 /*
 833                                  * Retry later if split_huge_page run
 834                                  * from under us.
 835                                  */
 836                                 ret = 1;
 837                         put_page(transhuge_head);
 838                 } else
 839                         /* Retry later if split_huge_page run from under us. */
 840                         ret = 1;
 841         }
 842         return ret;
 843 }
 844
 845 /*
 846  * try_to_merge_one_page - take two pages and merge them into one
 847  * @vma: the vma that holds the pte pointing to page
 848  * @page: the PageAnon page that we want to replace with kpage
 849  * @kpage: the PageKsm page that we want to map instead of page,
 850  *         or NULL the first time when we want to use page as kpage.
 851  *
 852  * This function returns 0 if the pages were merged, -EFAULT otherwise.
 853  */
 854 static int try_to_merge_one_page(struct vm_area_struct *vma,
 855                                  struct page *page, struct page *kpage)
 856 {
 857         pte_t orig_pte = __pte(0);
 858         int err = -EFAULT;
 859
 860         if (page == kpage)                      /* ksm page forked */
 861                 return 0;
 862
 863         if (!(vma->vm_flags & VM_MERGEABLE))
 864                 goto out;
 865         if (PageTransCompound(page) && page_trans_compound_anon_split(page))
 866                 goto out;
 867         BUG_ON(PageTransCompound(page));
 868         if (!PageAnon(page))
 869                 goto out;
 870
 871         /*
 872          * We need the page lock to read a stable PageSwapCache in
 873          * write_protect_page().  We use trylock_page() instead of
 874          * lock_page() because we don't want to wait here - we
 875          * prefer to continue scanning and merging different pages,
 876          * then come back to this page when it is unlocked.
 877          */
 878         if (!trylock_page(page))
 879                 goto out;
 880         /*
 881          * If this anonymous page is mapped only here, its pte may need
 882          * to be write-protected.  If it's mapped elsewhere, all of its
 883          * ptes are necessarily already write-protected.  But in either
 884          * case, we need to lock and check page_count is not raised.
 885          */
 886         if (write_protect_page(vma, page, &orig_pte) == 0) {
 887                 if (!kpage) {
 888                         /*
 889                          * While we hold page lock, upgrade page from
 890                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
 891                          * stable_tree_insert() will update stable_node.
 892                          */
 893                         set_page_stable_node(page, NULL);
 894                         mark_page_accessed(page);
 895                         err = 0;
 896                 } else if (pages_identical(page, kpage))
 897                         err = replace_page(vma, page, kpage, orig_pte);
 898         }
 899
 900         if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
 901                 munlock_vma_page(page);
 902                 if (!PageMlocked(kpage)) {
 903                         unlock_page(page);
 904                         lock_page(kpage);
 905                         mlock_vma_page(kpage);
 906                         page = kpage;           /* for final unlock */
 907                 }
 908         }
 909
 910         unlock_page(page);
 911 out:
 912         return err;
 913 }
 914
 915 /*
 916  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
 917  * but no new kernel page is allocated: kpage must already be a ksm page.
 918  *
 919  * This function returns 0 if the pages were merged, -EFAULT otherwise.
 920  */
 921 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
 922                                       struct page *page, struct page *kpage)
 923 {
 924         struct mm_struct *mm = rmap_item->mm;
 925         struct vm_area_struct *vma;
 926         int err = -EFAULT;
 927
 928         down_read(&mm->mmap_sem);
 929         if (ksm_test_exit(mm))
 930                 goto out;
 931         vma = find_vma(mm, rmap_item->address);
 932         if (!vma || vma->vm_start > rmap_item->address)
 933                 goto out;
 934
 935         err = try_to_merge_one_page(vma, page, kpage);
 936         if (err)
 937                 goto out;
 938
 939         /* Must get reference to anon_vma while still holding mmap_sem */
 940         rmap_item->anon_vma = vma->anon_vma;
 941         get_anon_vma(vma->anon_vma);
 942 out:
 943         up_read(&mm->mmap_sem);
 944         return err;
 945 }
 946
 947 /*
 948  * try_to_merge_two_pages - take two identical pages and prepare them
 949  * to be merged into one page.
 950  *
 951  * This function returns the kpage if we successfully merged two identical
 952  * pages into one ksm page, NULL otherwise.
 953  *
 954  * Note that this function upgrades page to ksm page: if one of the pages
 955  * is already a ksm page, try_to_merge_with_ksm_page should be used.
 956  */
 957 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
 958                                            struct page *page,
 959                                            struct rmap_item *tree_rmap_item,
 960                                            struct page *tree_page)
 961 {
 962         int err;
 963
 964         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
 965         if (!err) {
 966                 err = try_to_merge_with_ksm_page(tree_rmap_item,
 967                                                         tree_page, page);
 968                 /*
 969                  * If that fails, we have a ksm page with only one pte
 970                  * pointing to it: so break it.
 971                  */
 972                 if (err)
 973                         break_cow(rmap_item);
 974         }
 975         return err ? NULL : page;
 976 }
 977
 978 /*
 979  * stable_tree_search - search for page inside the stable tree
 980  *
 981  * This function checks if there is a page inside the stable tree
 982  * with identical content to the page that we are scanning right now.
 983  *
 984  * This function returns the stable tree node of identical content if found,
 985  * NULL otherwise.
 986  */
 987 static struct page *stable_tree_search(struct page *page)
 988 {
 989         struct rb_node *node = root_stable_tree.rb_node;
 990         struct stable_node *stable_node;
 991
 992         stable_node = page_stable_node(page);
 993         if (stable_node) {                      /* ksm page forked */
 994                 get_page(page);
 995                 return page;
 996         }
 997
 998         while (node) {
 999                 struct page *tree_page;
1000                 int ret;
1001
1002                 cond_resched();
1003                 stable_node = rb_entry(node, struct stable_node, node);
1004                 tree_page = get_ksm_page(stable_node);
1005                 if (!tree_page)
1006                         return NULL;
1007
1008                 ret = memcmp_pages(page, tree_page);
1009
1010                 if (ret < 0) {
1011                         put_page(tree_page);
1012                         node = node->rb_left;
1013                 } else if (ret > 0) {
1014                         put_page(tree_page);
1015                         node = node->rb_right;
1016                 } else
1017                         return tree_page;
1018         }
1019
1020         return NULL;
1021 }
1022
1023 /*
1024  * stable_tree_insert - insert rmap_item pointing to new ksm page
1025  * into the stable tree.
1026  *
1027  * This function returns the stable tree node just allocated on success,
1028  * NULL otherwise.
1029  */
1030 static struct stable_node *stable_tree_insert(struct page *kpage)
1031 {
1032         struct rb_node **new = &root_stable_tree.rb_node;
1033         struct rb_node *parent = NULL;
1034         struct stable_node *stable_node;
1035
1036         while (*new) {
1037                 struct page *tree_page;
1038                 int ret;
1039
1040                 cond_resched();
1041                 stable_node = rb_entry(*new, struct stable_node, node);
1042                 tree_page = get_ksm_page(stable_node);
1043                 if (!tree_page)
1044                         return NULL;
1045
1046                 ret = memcmp_pages(kpage, tree_page);
1047                 put_page(tree_page);
1048
1049                 parent = *new;
1050                 if (ret < 0)
1051                         new = &parent->rb_left;
1052                 else if (ret > 0)
1053                         new = &parent->rb_right;
1054                 else {
1055                         /*
1056                          * It is not a bug that stable_tree_search() didn't
1057                          * find this node: because at that time our page was
1058                          * not yet write-protected, so may have changed since.
1059                          */
1060                         return NULL;
1061                 }
1062         }
1063
1064         stable_node = alloc_stable_node();
1065         if (!stable_node)
1066                 return NULL;
1067
1068         rb_link_node(&stable_node->node, parent, new);
1069         rb_insert_color(&stable_node->node, &root_stable_tree);
1070
1071         INIT_HLIST_HEAD(&stable_node->hlist);
1072
1073         stable_node->kpfn = page_to_pfn(kpage);
1074         set_page_stable_node(kpage, stable_node);
1075
1076         return stable_node;
1077 }
1078
1079 /*
1080  * unstable_tree_search_insert - search for identical page,
1081  * else insert rmap_item into the unstable tree.
1082  *
1083  * This function searches for a page in the unstable tree identical to the
1084  * page currently being scanned; and if no identical page is found in the
1085  * tree, we insert rmap_item as a new object into the unstable tree.
1086  *
1087  * This function returns pointer to rmap_item found to be identical
1088  * to the currently scanned page, NULL otherwise.
1089  *
1090  * This function does both searching and inserting, because they share
1091  * the same walking algorithm in an rbtree.
1092  */
1093 static
1094 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1095                                               struct page *page,
1096                                               struct page **tree_pagep)
1097
1098 {
1099         struct rb_node **new = &root_unstable_tree.rb_node;
1100         struct rb_node *parent = NULL;
1101
1102         while (*new) {
1103                 struct rmap_item *tree_rmap_item;
1104                 struct page *tree_page;
1105                 int ret;
1106
1107                 cond_resched();
1108                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1109                 tree_page = get_mergeable_page(tree_rmap_item);
1110                 if (IS_ERR_OR_NULL(tree_page))
1111                         return NULL;
1112
1113                 /*
1114                  * Don't substitute a ksm page for a forked page.
1115                  */
1116                 if (page == tree_page) {
1117                         put_page(tree_page);
1118                         return NULL;
1119                 }
1120
1121                 ret = memcmp_pages(page, tree_page);
1122
1123                 parent = *new;
1124                 if (ret < 0) {
1125                         put_page(tree_page);
1126                         new = &parent->rb_left;
1127                 } else if (ret > 0) {
1128                         put_page(tree_page);
1129                         new = &parent->rb_right;
1130                 } else {
1131                         *tree_pagep = tree_page;
1132                         return tree_rmap_item;
1133                 }
1134         }
1135
1136         rmap_item->address |= UNSTABLE_FLAG;
1137         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1138         rb_link_node(&rmap_item->node, parent, new);
1139         rb_insert_color(&rmap_item->node, &root_unstable_tree);
1140
1141         ksm_pages_unshared++;
1142         return NULL;
1143 }
1144
1145 /*
1146  * stable_tree_append - add another rmap_item to the linked list of
1147  * rmap_items hanging off a given node of the stable tree, all sharing
1148  * the same ksm page.
1149  */
1150 static void stable_tree_append(struct rmap_item *rmap_item,
1151                                struct stable_node *stable_node)
1152 {
1153         rmap_item->head = stable_node;
1154         rmap_item->address |= STABLE_FLAG;
1155         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1156
1157         if (rmap_item->hlist.next)
1158                 ksm_pages_sharing++;
1159         else
1160                 ksm_pages_shared++;
1161 }
1162
1163 /*
1164  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1165  * if not, compare checksum to previous and if it's the same, see if page can
1166  * be inserted into the unstable tree, or merged with a page already there and
1167  * both transferred to the stable tree.
1168  *
1169  * @page: the page that we are searching identical page to.
1170  * @rmap_item: the reverse mapping into the virtual address of this page
1171  */
1172 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1173 {
1174         struct rmap_item *tree_rmap_item;
1175         struct page *tree_page = NULL;
1176         struct stable_node *stable_node;
1177         struct page *kpage;
1178         unsigned int checksum;
1179         int err;
1180
1181         remove_rmap_item_from_tree(rmap_item);
1182
1183         /* We first start with searching the page inside the stable tree */
1184         kpage = stable_tree_search(page);
1185         if (kpage) {
1186                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1187                 if (!err) {
1188                         /*
1189                          * The page was successfully merged:
1190                          * add its rmap_item to the stable tree.
1191                          */
1192                         lock_page(kpage);
1193                         stable_tree_append(rmap_item, page_stable_node(kpage));
1194                         unlock_page(kpage);
1195                 }
1196                 put_page(kpage);
1197                 return;
1198         }
1199
1200         /*
1201          * If the hash value of the page has changed from the last time
1202          * we calculated it, this page is changing frequently: therefore we
1203          * don't want to insert it in the unstable tree, and we don't want
1204          * to waste our time searching for something identical to it there.
1205          */
1206         checksum = calc_checksum(page);
1207         if (rmap_item->oldchecksum != checksum) {
1208                 rmap_item->oldchecksum = checksum;
1209                 return;
1210         }
1211
1212         tree_rmap_item =
1213                 unstable_tree_search_insert(rmap_item, page, &tree_page);
1214         if (tree_rmap_item) {
1215                 kpage = try_to_merge_two_pages(rmap_item, page,
1216                                                 tree_rmap_item, tree_page);
1217                 put_page(tree_page);
1218                 /*
1219                  * As soon as we merge this page, we want to remove the
1220                  * rmap_item of the page we have merged with from the unstable
1221                  * tree, and insert it instead as new node in the stable tree.
1222                  */
1223                 if (kpage) {
1224                         remove_rmap_item_from_tree(tree_rmap_item);
1225
1226                         lock_page(kpage);
1227                         stable_node = stable_tree_insert(kpage);
1228                         if (stable_node) {
1229                                 stable_tree_append(tree_rmap_item, stable_node);
1230                                 stable_tree_append(rmap_item, stable_node);
1231                         }
1232                         unlock_page(kpage);
1233
1234                         /*
1235                          * If we fail to insert the page into the stable tree,
1236                          * we will have 2 virtual addresses that are pointing
1237                          * to a ksm page left outside the stable tree,
1238                          * in which case we need to break_cow on both.
1239                          */
1240                         if (!stable_node) {
1241                                 break_cow(tree_rmap_item);
1242                                 break_cow(rmap_item);
1243                         }
1244                 }
1245         }
1246 }
1247
1248 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1249                                             struct rmap_item **rmap_list,
1250                                             unsigned long addr)
1251 {
1252         struct rmap_item *rmap_item;
1253
1254         while (*rmap_list) {
1255                 rmap_item = *rmap_list;
1256                 if ((rmap_item->address & PAGE_MASK) == addr)
1257                         return rmap_item;
1258                 if (rmap_item->address > addr)
1259                         break;
1260                 *rmap_list = rmap_item->rmap_list;
1261                 remove_rmap_item_from_tree(rmap_item);
1262                 free_rmap_item(rmap_item);
1263         }
1264
1265         rmap_item = alloc_rmap_item();
1266         if (rmap_item) {
1267                 /* It has already been zeroed */
1268                 rmap_item->mm = mm_slot->mm;
1269                 rmap_item->address = addr;
1270                 rmap_item->rmap_list = *rmap_list;
1271                 *rmap_list = rmap_item;
1272         }
1273         return rmap_item;
1274 }
1275
1276 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1277 {
1278         struct mm_struct *mm;
1279         struct mm_slot *slot;
1280         struct vm_area_struct *vma;
1281         struct rmap_item *rmap_item;
1282
1283         if (list_empty(&ksm_mm_head.mm_list))
1284                 return NULL;
1285
1286         slot = ksm_scan.mm_slot;
1287         if (slot == &ksm_mm_head) {
1288                 /*
1289                  * A number of pages can hang around indefinitely on per-cpu
1290                  * pagevecs, raised page count preventing write_protect_page
1291                  * from merging them.  Though it doesn't really matter much,
1292                  * it is puzzling to see some stuck in pages_volatile until
1293                  * other activity jostles them out, and they also prevented
1294                  * LTP's KSM test from succeeding deterministically; so drain
1295                  * them here (here rather than on entry to ksm_do_scan(),
1296                  * so we don't IPI too often when pages_to_scan is set low).
1297                  */
1298                 lru_add_drain_all();
1299
1300                 root_unstable_tree = RB_ROOT;
1301
1302                 spin_lock(&ksm_mmlist_lock);
1303                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1304                 ksm_scan.mm_slot = slot;
1305                 spin_unlock(&ksm_mmlist_lock);
1306                 /*
1307                  * Although we tested list_empty() above, a racing __ksm_exit
1308                  * of the last mm on the list may have removed it since then.
1309                  */
1310                 if (slot == &ksm_mm_head)
1311                         return NULL;
1312 next_mm:
1313                 ksm_scan.address = 0;
1314                 ksm_scan.rmap_list = &slot->rmap_list;
1315         }
1316
1317         mm = slot->mm;
1318         down_read(&mm->mmap_sem);
1319         if (ksm_test_exit(mm))
1320                 vma = NULL;
1321         else
1322                 vma = find_vma(mm, ksm_scan.address);
1323
1324         for (; vma; vma = vma->vm_next) {
1325                 if (!(vma->vm_flags & VM_MERGEABLE))
1326                         continue;
1327                 if (ksm_scan.address < vma->vm_start)
1328                         ksm_scan.address = vma->vm_start;
1329                 if (!vma->anon_vma)
1330                         ksm_scan.address = vma->vm_end;
1331
1332                 while (ksm_scan.address < vma->vm_end) {
1333                         if (ksm_test_exit(mm))
1334                                 break;
1335                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1336                         if (IS_ERR_OR_NULL(*page)) {
1337                                 ksm_scan.address += PAGE_SIZE;
1338                                 cond_resched();
1339                                 continue;
1340                         }
1341                         if (PageAnon(*page) ||
1342                             page_trans_compound_anon(*page)) {
1343                                 flush_anon_page(vma, *page, ksm_scan.address);
1344                                 flush_dcache_page(*page);
1345                                 rmap_item = get_next_rmap_item(slot,
1346                                         ksm_scan.rmap_list, ksm_scan.address);
1347                                 if (rmap_item) {
1348                                         ksm_scan.rmap_list =
1349                                                         &rmap_item->rmap_list;
1350                                         ksm_scan.address += PAGE_SIZE;
1351                                 } else
1352                                         put_page(*page);
1353                                 up_read(&mm->mmap_sem);
1354                                 return rmap_item;
1355                         }
1356                         put_page(*page);
1357                         ksm_scan.address += PAGE_SIZE;
1358                         cond_resched();
1359                 }
1360         }
1361
1362         if (ksm_test_exit(mm)) {
1363                 ksm_scan.address = 0;
1364                 ksm_scan.rmap_list = &slot->rmap_list;
1365         }
1366         /*
1367          * Nuke all the rmap_items that are above this current rmap:
1368          * because there were no VM_MERGEABLE vmas with such addresses.
1369          */
1370         remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1371
1372         spin_lock(&ksm_mmlist_lock);
1373         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1374                                                 struct mm_slot, mm_list);
1375         if (ksm_scan.address == 0) {
1376                 /*
1377                  * We've completed a full scan of all vmas, holding mmap_sem
1378                  * throughout, and found no VM_MERGEABLE: so do the same as
1379                  * __ksm_exit does to remove this mm from all our lists now.
1380                  * This applies either when cleaning up after __ksm_exit
1381                  * (but beware: we can reach here even before __ksm_exit),
1382                  * or when all VM_MERGEABLE areas have been unmapped (and
1383                  * mmap_sem then protects against race with MADV_MERGEABLE).
1384                  */
1385                 hlist_del(&slot->link);
1386                 list_del(&slot->mm_list);
1387                 spin_unlock(&ksm_mmlist_lock);
1388
1389                 free_mm_slot(slot);
1390                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1391                 up_read(&mm->mmap_sem);
1392                 mmdrop(mm);
1393         } else {
1394                 spin_unlock(&ksm_mmlist_lock);
1395                 up_read(&mm->mmap_sem);
1396         }
1397
1398         /* Repeat until we've completed scanning the whole list */
1399         slot = ksm_scan.mm_slot;
1400         if (slot != &ksm_mm_head)
1401                 goto next_mm;
1402
1403         ksm_scan.seqnr++;
1404         return NULL;
1405 }
1406
1407 /**
1408  * ksm_do_scan  - the ksm scanner main worker function.
1409  * @scan_npages - number of pages we want to scan before we return.
1410  */
1411 static void ksm_do_scan(unsigned int scan_npages)
1412 {
1413         struct rmap_item *rmap_item;
1414         struct page *uninitialized_var(page);
1415
1416         while (scan_npages-- && likely(!freezing(current))) {
1417                 cond_resched();
1418                 rmap_item = scan_get_next_rmap_item(&page);
1419                 if (!rmap_item)
1420                         return;
1421                 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1422                         cmp_and_merge_page(page, rmap_item);
1423                 put_page(page);
1424         }
1425 }
1426
1427 static int ksmd_should_run(void)
1428 {
1429         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1430 }
1431
1432 static int ksm_scan_thread(void *nothing)
1433 {
1434         set_freezable();
1435         set_user_nice(current, 5);
1436
1437         while (!kthread_should_stop()) {
1438                 mutex_lock(&ksm_thread_mutex);
1439                 if (ksmd_should_run())
1440                         ksm_do_scan(ksm_thread_pages_to_scan);
1441                 mutex_unlock(&ksm_thread_mutex);
1442
1443                 try_to_freeze();
1444
1445                 if (ksmd_should_run()) {
1446                         schedule_timeout_interruptible(
1447                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1448                 } else {
1449                         wait_event_freezable(ksm_thread_wait,
1450                                 ksmd_should_run() || kthread_should_stop());
1451                 }
1452         }
1453         return 0;
1454 }
1455
1456 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1457                 unsigned long end, int advice, unsigned long *vm_flags)
1458 {
1459         struct mm_struct *mm = vma->vm_mm;
1460         int err;
1461
1462         switch (advice) {
1463         case MADV_MERGEABLE:
1464                 /*
1465                  * Be somewhat over-protective for now!
1466                  */
1467                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1468                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1469                                  VM_RESERVED  | VM_HUGETLB | VM_INSERTPAGE |
1470                                  VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
1471                         return 0;               /* just ignore the advice */
1472
1473                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1474                         err = __ksm_enter(mm);
1475                         if (err)
1476                                 return err;
1477                 }
1478
1479                 *vm_flags |= VM_MERGEABLE;
1480                 break;
1481
1482         case MADV_UNMERGEABLE:
1483                 if (!(*vm_flags & VM_MERGEABLE))
1484                         return 0;               /* just ignore the advice */
1485
1486                 if (vma->anon_vma) {
1487                         err = unmerge_ksm_pages(vma, start, end);
1488                         if (err)
1489                                 return err;
1490                 }
1491
1492                 *vm_flags &= ~VM_MERGEABLE;
1493                 break;
1494         }
1495
1496         return 0;
1497 }
1498
1499 int __ksm_enter(struct mm_struct *mm)
1500 {
1501         struct mm_slot *mm_slot;
1502         int needs_wakeup;
1503
1504         mm_slot = alloc_mm_slot();
1505         if (!mm_slot)
1506                 return -ENOMEM;
1507
1508         /* Check ksm_run too?  Would need tighter locking */
1509         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1510
1511         spin_lock(&ksm_mmlist_lock);
1512         insert_to_mm_slots_hash(mm, mm_slot);
1513         /*
1514          * Insert just behind the scanning cursor, to let the area settle
1515          * down a little; when fork is followed by immediate exec, we don't
1516          * want ksmd to waste time setting up and tearing down an rmap_list.
1517          */
1518         list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1519         spin_unlock(&ksm_mmlist_lock);
1520
1521         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1522         atomic_inc(&mm->mm_count);
1523
1524         if (needs_wakeup)
1525                 wake_up_interruptible(&ksm_thread_wait);
1526
1527         return 0;
1528 }
1529
1530 void __ksm_exit(struct mm_struct *mm)
1531 {
1532         struct mm_slot *mm_slot;
1533         int easy_to_free = 0;
1534
1535         /*
1536          * This process is exiting: if it's straightforward (as is the
1537          * case when ksmd was never running), free mm_slot immediately.
1538          * But if it's at the cursor or has rmap_items linked to it, use
1539          * mmap_sem to synchronize with any break_cows before pagetables
1540          * are freed, and leave the mm_slot on the list for ksmd to free.
1541          * Beware: ksm may already have noticed it exiting and freed the slot.
1542          */
1543
1544         spin_lock(&ksm_mmlist_lock);
1545         mm_slot = get_mm_slot(mm);
1546         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1547                 if (!mm_slot->rmap_list) {
1548                         hlist_del(&mm_slot->link);
1549                         list_del(&mm_slot->mm_list);
1550                         easy_to_free = 1;
1551                 } else {
1552                         list_move(&mm_slot->mm_list,
1553                                   &ksm_scan.mm_slot->mm_list);
1554                 }
1555         }
1556         spin_unlock(&ksm_mmlist_lock);
1557
1558         if (easy_to_free) {
1559                 free_mm_slot(mm_slot);
1560                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1561                 mmdrop(mm);
1562         } else if (mm_slot) {
1563                 down_write(&mm->mmap_sem);
1564                 up_write(&mm->mmap_sem);
1565         }
1566 }
1567
1568 struct page *ksm_does_need_to_copy(struct page *page,
1569                         struct vm_area_struct *vma, unsigned long address)
1570 {
1571         struct page *new_page;
1572
1573         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1574         if (new_page) {
1575                 copy_user_highpage(new_page, page, address, vma);
1576
1577                 SetPageDirty(new_page);
1578                 __SetPageUptodate(new_page);
1579                 SetPageSwapBacked(new_page);
1580                 __set_page_locked(new_page);
1581
1582                 if (page_evictable(new_page, vma))
1583                         lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1584                 else
1585                         add_page_to_unevictable_list(new_page);
1586         }
1587
1588         return new_page;
1589 }
1590
1591 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1592                         unsigned long *vm_flags)
1593 {
1594         struct stable_node *stable_node;
1595         struct rmap_item *rmap_item;
1596         struct hlist_node *hlist;
1597         unsigned int mapcount = page_mapcount(page);
1598         int referenced = 0;
1599         int search_new_forks = 0;
1600
1601         VM_BUG_ON(!PageKsm(page));
1602         VM_BUG_ON(!PageLocked(page));
1603
1604         stable_node = page_stable_node(page);
1605         if (!stable_node)
1606                 return 0;
1607 again:
1608         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1609                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1610                 struct anon_vma_chain *vmac;
1611                 struct vm_area_struct *vma;
1612
1613                 anon_vma_lock(anon_vma);
1614                 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1615                         vma = vmac->vma;
1616                         if (rmap_item->address < vma->vm_start ||
1617                             rmap_item->address >= vma->vm_end)
1618                                 continue;
1619                         /*
1620                          * Initially we examine only the vma which covers this
1621                          * rmap_item; but later, if there is still work to do,
1622                          * we examine covering vmas in other mms: in case they
1623                          * were forked from the original since ksmd passed.
1624                          */
1625                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1626                                 continue;
1627
1628                         if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1629                                 continue;
1630
1631                         referenced += page_referenced_one(page, vma,
1632                                 rmap_item->address, &mapcount, vm_flags);
1633                         if (!search_new_forks || !mapcount)
1634                                 break;
1635                 }
1636                 anon_vma_unlock(anon_vma);
1637                 if (!mapcount)
1638                         goto out;
1639         }
1640         if (!search_new_forks++)
1641                 goto again;
1642 out:
1643         return referenced;
1644 }
1645
1646 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1647 {
1648         struct stable_node *stable_node;
1649         struct hlist_node *hlist;
1650         struct rmap_item *rmap_item;
1651         int ret = SWAP_AGAIN;
1652         int search_new_forks = 0;
1653
1654         VM_BUG_ON(!PageKsm(page));
1655         VM_BUG_ON(!PageLocked(page));
1656
1657         stable_node = page_stable_node(page);
1658         if (!stable_node)
1659                 return SWAP_FAIL;
1660 again:
1661         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1662                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1663                 struct anon_vma_chain *vmac;
1664                 struct vm_area_struct *vma;
1665
1666                 anon_vma_lock(anon_vma);
1667                 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1668                         vma = vmac->vma;
1669                         if (rmap_item->address < vma->vm_start ||
1670                             rmap_item->address >= vma->vm_end)
1671                                 continue;
1672                         /*
1673                          * Initially we examine only the vma which covers this
1674                          * rmap_item; but later, if there is still work to do,
1675                          * we examine covering vmas in other mms: in case they
1676                          * were forked from the original since ksmd passed.
1677                          */
1678                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1679                                 continue;
1680
1681                         ret = try_to_unmap_one(page, vma,
1682                                         rmap_item->address, flags);
1683                         if (ret != SWAP_AGAIN || !page_mapped(page)) {
1684                                 anon_vma_unlock(anon_vma);
1685                                 goto out;
1686                         }
1687                 }
1688                 anon_vma_unlock(anon_vma);
1689         }
1690         if (!search_new_forks++)
1691                 goto again;
1692 out:
1693         return ret;
1694 }
1695
1696 #ifdef CONFIG_MIGRATION
1697 int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1698                   struct vm_area_struct *, unsigned long, void *), void *arg)
1699 {
1700         struct stable_node *stable_node;
1701         struct hlist_node *hlist;
1702         struct rmap_item *rmap_item;
1703         int ret = SWAP_AGAIN;
1704         int search_new_forks = 0;
1705
1706         VM_BUG_ON(!PageKsm(page));
1707         VM_BUG_ON(!PageLocked(page));
1708
1709         stable_node = page_stable_node(page);
1710         if (!stable_node)
1711                 return ret;
1712 again:
1713         hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1714                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1715                 struct anon_vma_chain *vmac;
1716                 struct vm_area_struct *vma;
1717
1718                 anon_vma_lock(anon_vma);
1719                 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1720                         vma = vmac->vma;
1721                         if (rmap_item->address < vma->vm_start ||
1722                             rmap_item->address >= vma->vm_end)
1723                                 continue;
1724                         /*
1725                          * Initially we examine only the vma which covers this
1726                          * rmap_item; but later, if there is still work to do,
1727                          * we examine covering vmas in other mms: in case they
1728                          * were forked from the original since ksmd passed.
1729                          */
1730                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1731                                 continue;
1732
1733                         ret = rmap_one(page, vma, rmap_item->address, arg);
1734                         if (ret != SWAP_AGAIN) {
1735                                 anon_vma_unlock(anon_vma);
1736                                 goto out;
1737                         }
1738                 }
1739                 anon_vma_unlock(anon_vma);
1740         }
1741         if (!search_new_forks++)
1742                 goto again;
1743 out:
1744         return ret;
1745 }
1746
1747 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1748 {
1749         struct stable_node *stable_node;
1750
1751         VM_BUG_ON(!PageLocked(oldpage));
1752         VM_BUG_ON(!PageLocked(newpage));
1753         VM_BUG_ON(newpage->mapping != oldpage->mapping);
1754
1755         stable_node = page_stable_node(newpage);
1756         if (stable_node) {
1757                 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1758                 stable_node->kpfn = page_to_pfn(newpage);
1759         }
1760 }
1761 #endif /* CONFIG_MIGRATION */
1762
1763 #ifdef CONFIG_MEMORY_HOTREMOVE
1764 static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1765                                                  unsigned long end_pfn)
1766 {
1767         struct rb_node *node;
1768
1769         for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1770                 struct stable_node *stable_node;
1771
1772                 stable_node = rb_entry(node, struct stable_node, node);
1773                 if (stable_node->kpfn >= start_pfn &&
1774                     stable_node->kpfn < end_pfn)
1775                         return stable_node;
1776         }
1777         return NULL;
1778 }
1779
1780 static int ksm_memory_callback(struct notifier_block *self,
1781                                unsigned long action, void *arg)
1782 {
1783         struct memory_notify *mn = arg;
1784         struct stable_node *stable_node;
1785
1786         switch (action) {
1787         case MEM_GOING_OFFLINE:
1788                 /*
1789                  * Keep it very simple for now: just lock out ksmd and
1790                  * MADV_UNMERGEABLE while any memory is going offline.
1791                  * mutex_lock_nested() is necessary because lockdep was alarmed
1792                  * that here we take ksm_thread_mutex inside notifier chain
1793                  * mutex, and later take notifier chain mutex inside
1794                  * ksm_thread_mutex to unlock it.   But that's safe because both
1795                  * are inside mem_hotplug_mutex.
1796                  */
1797                 mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
1798                 break;
1799
1800         case MEM_OFFLINE:
1801                 /*
1802                  * Most of the work is done by page migration; but there might
1803                  * be a few stable_nodes left over, still pointing to struct
1804                  * pages which have been offlined: prune those from the tree.
1805                  */
1806                 while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1807                                         mn->start_pfn + mn->nr_pages)) != NULL)
1808                         remove_node_from_stable_tree(stable_node);
1809                 /* fallthrough */
1810
1811         case MEM_CANCEL_OFFLINE:
1812                 mutex_unlock(&ksm_thread_mutex);
1813                 break;
1814         }
1815         return NOTIFY_OK;
1816 }
1817 #endif /* CONFIG_MEMORY_HOTREMOVE */
1818
1819 #ifdef CONFIG_SYSFS
1820 /*
1821  * This all compiles without CONFIG_SYSFS, but is a waste of space.
1822  */
1823
1824 #define KSM_ATTR_RO(_name) \
1825         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1826 #define KSM_ATTR(_name) \
1827         static struct kobj_attribute _name##_attr = \
1828                 __ATTR(_name, 0644, _name##_show, _name##_store)
1829
1830 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1831                                     struct kobj_attribute *attr, char *buf)
1832 {
1833         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1834 }
1835
1836 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1837                                      struct kobj_attribute *attr,
1838                                      const char *buf, size_t count)
1839 {
1840         unsigned long msecs;
1841         int err;
1842
1843         err = strict_strtoul(buf, 10, &msecs);
1844         if (err || msecs > UINT_MAX)
1845                 return -EINVAL;
1846
1847         ksm_thread_sleep_millisecs = msecs;
1848
1849         return count;
1850 }
1851 KSM_ATTR(sleep_millisecs);
1852
1853 static ssize_t pages_to_scan_show(struct kobject *kobj,
1854                                   struct kobj_attribute *attr, char *buf)
1855 {
1856         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1857 }
1858
1859 static ssize_t pages_to_scan_store(struct kobject *kobj,
1860                                    struct kobj_attribute *attr,
1861                                    const char *buf, size_t count)
1862 {
1863         int err;
1864         unsigned long nr_pages;
1865
1866         err = strict_strtoul(buf, 10, &nr_pages);
1867         if (err || nr_pages > UINT_MAX)
1868                 return -EINVAL;
1869
1870         ksm_thread_pages_to_scan = nr_pages;
1871
1872         return count;
1873 }
1874 KSM_ATTR(pages_to_scan);
1875
1876 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1877                         char *buf)
1878 {
1879         return sprintf(buf, "%u\n", ksm_run);
1880 }
1881
1882 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1883                          const char *buf, size_t count)
1884 {
1885         int err;
1886         unsigned long flags;
1887
1888         err = strict_strtoul(buf, 10, &flags);
1889         if (err || flags > UINT_MAX)
1890                 return -EINVAL;
1891         if (flags > KSM_RUN_UNMERGE)
1892                 return -EINVAL;
1893
1894         /*
1895          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1896          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1897          * breaking COW to free the pages_shared (but leaves mm_slots
1898          * on the list for when ksmd may be set running again).
1899          */
1900
1901         mutex_lock(&ksm_thread_mutex);
1902         if (ksm_run != flags) {
1903                 ksm_run = flags;
1904                 if (flags & KSM_RUN_UNMERGE) {
1905                         int oom_score_adj;
1906
1907                         oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
1908                         err = unmerge_and_remove_all_rmap_items();
1909                         compare_swap_oom_score_adj(OOM_SCORE_ADJ_MAX,
1910                                                                 oom_score_adj);
1911                         if (err) {
1912                                 ksm_run = KSM_RUN_STOP;
1913                                 count = err;
1914                         }
1915                 }
1916         }
1917         mutex_unlock(&ksm_thread_mutex);
1918
1919         if (flags & KSM_RUN_MERGE)
1920                 wake_up_interruptible(&ksm_thread_wait);
1921
1922         return count;
1923 }
1924 KSM_ATTR(run);
1925
1926 static ssize_t pages_shared_show(struct kobject *kobj,
1927                                  struct kobj_attribute *attr, char *buf)
1928 {
1929         return sprintf(buf, "%lu\n", ksm_pages_shared);
1930 }
1931 KSM_ATTR_RO(pages_shared);
1932
1933 static ssize_t pages_sharing_show(struct kobject *kobj,
1934                                   struct kobj_attribute *attr, char *buf)
1935 {
1936         return sprintf(buf, "%lu\n", ksm_pages_sharing);
1937 }
1938 KSM_ATTR_RO(pages_sharing);
1939
1940 static ssize_t pages_unshared_show(struct kobject *kobj,
1941                                    struct kobj_attribute *attr, char *buf)
1942 {
1943         return sprintf(buf, "%lu\n", ksm_pages_unshared);
1944 }
1945 KSM_ATTR_RO(pages_unshared);
1946
1947 static ssize_t pages_volatile_show(struct kobject *kobj,
1948                                    struct kobj_attribute *attr, char *buf)
1949 {
1950         long ksm_pages_volatile;
1951
1952         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1953                                 - ksm_pages_sharing - ksm_pages_unshared;
1954         /*
1955          * It was not worth any locking to calculate that statistic,
1956          * but it might therefore sometimes be negative: conceal that.
1957          */
1958         if (ksm_pages_volatile < 0)
1959                 ksm_pages_volatile = 0;
1960         return sprintf(buf, "%ld\n", ksm_pages_volatile);
1961 }
1962 KSM_ATTR_RO(pages_volatile);
1963
1964 static ssize_t full_scans_show(struct kobject *kobj,
1965                                struct kobj_attribute *attr, char *buf)
1966 {
1967         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1968 }
1969 KSM_ATTR_RO(full_scans);
1970
1971 static struct attribute *ksm_attrs[] = {
1972         &sleep_millisecs_attr.attr,
1973         &pages_to_scan_attr.attr,
1974         &run_attr.attr,
1975         &pages_shared_attr.attr,
1976         &pages_sharing_attr.attr,
1977         &pages_unshared_attr.attr,
1978         &pages_volatile_attr.attr,
1979         &full_scans_attr.attr,
1980         NULL,
1981 };
1982
1983 static struct attribute_group ksm_attr_group = {
1984         .attrs = ksm_attrs,
1985         .name = "ksm",
1986 };
1987 #endif /* CONFIG_SYSFS */
1988
1989 static int __init ksm_init(void)
1990 {
1991         struct task_struct *ksm_thread;
1992         int err;
1993
1994         err = ksm_slab_init();
1995         if (err)
1996                 goto out;
1997
1998         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1999         if (IS_ERR(ksm_thread)) {
2000                 printk(KERN_ERR "ksm: creating kthread failed\n");
2001                 err = PTR_ERR(ksm_thread);
2002                 goto out_free;
2003         }
2004
2005 #ifdef CONFIG_SYSFS
2006         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2007         if (err) {
2008                 printk(KERN_ERR "ksm: register sysfs failed\n");
2009                 kthread_stop(ksm_thread);
2010                 goto out_free;
2011         }
2012 #else
2013         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
2014
2015 #endif /* CONFIG_SYSFS */
2016
2017 #ifdef CONFIG_MEMORY_HOTREMOVE
2018         /*
2019          * Choose a high priority since the callback takes ksm_thread_mutex:
2020          * later callbacks could only be taking locks which nest within that.
2021          */
2022         hotplug_memory_notifier(ksm_memory_callback, 100);
2023 #endif
2024         return 0;
2025
2026 out_free:
2027         ksm_slab_free();
2028 out:
2029         return err;
2030 }
2031 module_init(ksm_init)