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/mmu_notifier.h>
  33 #include <linux/swap.h>
  34 #include <linux/ksm.h>
  35
  36 #include <asm/tlbflush.h>
  37
  38 /*
  39  * A few notes about the KSM scanning process,
  40  * to make it easier to understand the data structures below:
  41  *
  42  * In order to reduce excessive scanning, KSM sorts the memory pages by their
  43  * contents into a data structure that holds pointers to the pages' locations.
  44  *
  45  * Since the contents of the pages may change at any moment, KSM cannot just
  46  * insert the pages into a normal sorted tree and expect it to find anything.
  47  * Therefore KSM uses two data structures - the stable and the unstable tree.
  48  *
  49  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
  50  * by their contents.  Because each such page is write-protected, searching on
  51  * this tree is fully assured to be working (except when pages are unmapped),
  52  * and therefore this tree is called the stable tree.
  53  *
  54  * In addition to the stable tree, KSM uses a second data structure called the
  55  * unstable tree: this tree holds pointers to pages which have been found to
  56  * be "unchanged for a period of time".  The unstable tree sorts these pages
  57  * by their contents, but since they are not write-protected, KSM cannot rely
  58  * upon the unstable tree to work correctly - the unstable tree is liable to
  59  * be corrupted as its contents are modified, and so it is called unstable.
  60  *
  61  * KSM solves this problem by several techniques:
  62  *
  63  * 1) The unstable tree is flushed every time KSM completes scanning all
  64  *    memory areas, and then the tree is rebuilt again from the beginning.
  65  * 2) KSM will only insert into the unstable tree, pages whose hash value
  66  *    has not changed since the previous scan of all memory areas.
  67  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
  68  *    colors of the nodes and not on their contents, assuring that even when
  69  *    the tree gets "corrupted" it won't get out of balance, so scanning time
  70  *    remains the same (also, searching and inserting nodes in an rbtree uses
  71  *    the same algorithm, so we have no overhead when we flush and rebuild).
  72  * 4) KSM never flushes the stable tree, which means that even if it were to
  73  *    take 10 attempts to find a page in the unstable tree, once it is found,
  74  *    it is secured in the stable tree.  (When we scan a new page, we first
  75  *    compare it against the stable tree, and then against the unstable tree.)
  76  */
  77
  78 /**
  79  * struct mm_slot - ksm information per mm that is being scanned
  80  * @link: link to the mm_slots hash list
  81  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
  82  * @rmap_list: head for this mm_slot's list of rmap_items
  83  * @mm: the mm that this information is valid for
  84  */
  85 struct mm_slot {
  86         struct hlist_node link;
  87         struct list_head mm_list;
  88         struct list_head rmap_list;
  89         struct mm_struct *mm;
  90 };
  91
  92 /**
  93  * struct ksm_scan - cursor for scanning
  94  * @mm_slot: the current mm_slot we are scanning
  95  * @address: the next address inside that to be scanned
  96  * @rmap_item: the current rmap that we are scanning inside the rmap_list
  97  * @seqnr: count of completed full scans (needed when removing unstable node)
  98  *
  99  * There is only the one ksm_scan instance of this cursor structure.
 100  */
 101 struct ksm_scan {
 102         struct mm_slot *mm_slot;
 103         unsigned long address;
 104         struct rmap_item *rmap_item;
 105         unsigned long seqnr;
 106 };
 107
 108 /**
 109  * struct rmap_item - reverse mapping item for virtual addresses
 110  * @link: link into mm_slot's rmap_list (rmap_list is per mm)
 111  * @mm: the memory structure this rmap_item is pointing into
 112  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
 113  * @oldchecksum: previous checksum of the page at that virtual address
 114  * @node: rb_node of this rmap_item in either unstable or stable tree
 115  * @next: next rmap_item hanging off the same node of the stable tree
 116  * @prev: previous rmap_item hanging off the same node of the stable tree
 117  */
 118 struct rmap_item {
 119         struct list_head link;
 120         struct mm_struct *mm;
 121         unsigned long address;          /* + low bits used for flags below */
 122         union {
 123                 unsigned int oldchecksum;               /* when unstable */
 124                 struct rmap_item *next;                 /* when stable */
 125         };
 126         union {
 127                 struct rb_node node;                    /* when tree node */
 128                 struct rmap_item *prev;                 /* in stable list */
 129         };
 130 };
 131
 132 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
 133 #define NODE_FLAG       0x100   /* is a node of unstable or stable tree */
 134 #define STABLE_FLAG     0x200   /* is a node or list item of stable tree */
 135
 136 /* The stable and unstable tree heads */
 137 static struct rb_root root_stable_tree = RB_ROOT;
 138 static struct rb_root root_unstable_tree = RB_ROOT;
 139
 140 #define MM_SLOTS_HASH_HEADS 1024
 141 static struct hlist_head *mm_slots_hash;
 142
 143 static struct mm_slot ksm_mm_head = {
 144         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
 145 };
 146 static struct ksm_scan ksm_scan = {
 147         .mm_slot = &ksm_mm_head,
 148 };
 149
 150 static struct kmem_cache *rmap_item_cache;
 151 static struct kmem_cache *mm_slot_cache;
 152
 153 /* The number of nodes in the stable tree */
 154 static unsigned long ksm_pages_shared;
 155
 156 /* The number of page slots additionally sharing those nodes */
 157 static unsigned long ksm_pages_sharing;
 158
 159 /* The number of nodes in the unstable tree */
 160 static unsigned long ksm_pages_unshared;
 161
 162 /* The number of rmap_items in use: to calculate pages_volatile */
 163 static unsigned long ksm_rmap_items;
 164
 165 /* Limit on the number of unswappable pages used */
 166 static unsigned long ksm_max_kernel_pages;
 167
 168 /* Number of pages ksmd should scan in one batch */
 169 static unsigned int ksm_thread_pages_to_scan = 100;
 170
 171 /* Milliseconds ksmd should sleep between batches */
 172 static unsigned int ksm_thread_sleep_millisecs = 20;
 173
 174 #define KSM_RUN_STOP    0
 175 #define KSM_RUN_MERGE   1
 176 #define KSM_RUN_UNMERGE 2
 177 static unsigned int ksm_run = KSM_RUN_STOP;
 178
 179 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
 180 static DEFINE_MUTEX(ksm_thread_mutex);
 181 static DEFINE_SPINLOCK(ksm_mmlist_lock);
 182
 183 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
 184                 sizeof(struct __struct), __alignof__(struct __struct),\
 185                 (__flags), NULL)
 186
 187 static int __init ksm_slab_init(void)
 188 {
 189         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
 190         if (!rmap_item_cache)
 191                 goto out;
 192
 193         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
 194         if (!mm_slot_cache)
 195                 goto out_free;
 196
 197         return 0;
 198
 199 out_free:
 200         kmem_cache_destroy(rmap_item_cache);
 201 out:
 202         return -ENOMEM;
 203 }
 204
 205 static void __init ksm_slab_free(void)
 206 {
 207         kmem_cache_destroy(mm_slot_cache);
 208         kmem_cache_destroy(rmap_item_cache);
 209         mm_slot_cache = NULL;
 210 }
 211
 212 static inline struct rmap_item *alloc_rmap_item(void)
 213 {
 214         struct rmap_item *rmap_item;
 215
 216         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
 217         if (rmap_item)
 218                 ksm_rmap_items++;
 219         return rmap_item;
 220 }
 221
 222 static inline void free_rmap_item(struct rmap_item *rmap_item)
 223 {
 224         ksm_rmap_items--;
 225         rmap_item->mm = NULL;   /* debug safety */
 226         kmem_cache_free(rmap_item_cache, rmap_item);
 227 }
 228
 229 static inline struct mm_slot *alloc_mm_slot(void)
 230 {
 231         if (!mm_slot_cache)     /* initialization failed */
 232                 return NULL;
 233         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 234 }
 235
 236 static inline void free_mm_slot(struct mm_slot *mm_slot)
 237 {
 238         kmem_cache_free(mm_slot_cache, mm_slot);
 239 }
 240
 241 static int __init mm_slots_hash_init(void)
 242 {
 243         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
 244                                 GFP_KERNEL);
 245         if (!mm_slots_hash)
 246                 return -ENOMEM;
 247         return 0;
 248 }
 249
 250 static void __init mm_slots_hash_free(void)
 251 {
 252         kfree(mm_slots_hash);
 253 }
 254
 255 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 256 {
 257         struct mm_slot *mm_slot;
 258         struct hlist_head *bucket;
 259         struct hlist_node *node;
 260
 261         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
 262                                 % MM_SLOTS_HASH_HEADS];
 263         hlist_for_each_entry(mm_slot, node, bucket, link) {
 264                 if (mm == mm_slot->mm)
 265                         return mm_slot;
 266         }
 267         return NULL;
 268 }
 269
 270 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 271                                     struct mm_slot *mm_slot)
 272 {
 273         struct hlist_head *bucket;
 274
 275         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
 276                                 % MM_SLOTS_HASH_HEADS];
 277         mm_slot->mm = mm;
 278         INIT_LIST_HEAD(&mm_slot->rmap_list);
 279         hlist_add_head(&mm_slot->link, bucket);
 280 }
 281
 282 static inline int in_stable_tree(struct rmap_item *rmap_item)
 283 {
 284         return rmap_item->address & STABLE_FLAG;
 285 }
 286
 287 /*
 288  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
 289  * page tables after it has passed through ksm_exit() - which, if necessary,
 290  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
 291  * a special flag: they can just back out as soon as mm_users goes to zero.
 292  * ksm_test_exit() is used throughout to make this test for exit: in some
 293  * places for correctness, in some places just to avoid unnecessary work.
 294  */
 295 static inline bool ksm_test_exit(struct mm_struct *mm)
 296 {
 297         return atomic_read(&mm->mm_users) == 0;
 298 }
 299
 300 /*
 301  * We use break_ksm to break COW on a ksm page: it's a stripped down
 302  *
 303  *      if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
 304  *              put_page(page);
 305  *
 306  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
 307  * in case the application has unmapped and remapped mm,addr meanwhile.
 308  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
 309  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
 310  */
 311 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
 312 {
 313         struct page *page;
 314         int ret = 0;
 315
 316         do {
 317                 cond_resched();
 318                 page = follow_page(vma, addr, FOLL_GET);
 319                 if (!page)
 320                         break;
 321                 if (PageKsm(page))
 322                         ret = handle_mm_fault(vma->vm_mm, vma, addr,
 323                                                         FAULT_FLAG_WRITE);
 324                 else
 325                         ret = VM_FAULT_WRITE;
 326                 put_page(page);
 327         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
 328         /*
 329          * We must loop because handle_mm_fault() may back out if there's
 330          * any difficulty e.g. if pte accessed bit gets updated concurrently.
 331          *
 332          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
 333          * COW has been broken, even if the vma does not permit VM_WRITE;
 334          * but note that a concurrent fault might break PageKsm for us.
 335          *
 336          * VM_FAULT_SIGBUS could occur if we race with truncation of the
 337          * backing file, which also invalidates anonymous pages: that's
 338          * okay, that truncation will have unmapped the PageKsm for us.
 339          *
 340          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
 341          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
 342          * current task has TIF_MEMDIE set, and will be OOM killed on return
 343          * to user; and ksmd, having no mm, would never be chosen for that.
 344          *
 345          * But if the mm is in a limited mem_cgroup, then the fault may fail
 346          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
 347          * even ksmd can fail in this way - though it's usually breaking ksm
 348          * just to undo a merge it made a moment before, so unlikely to oom.
 349          *
 350          * That's a pity: we might therefore have more kernel pages allocated
 351          * than we're counting as nodes in the stable tree; but ksm_do_scan
 352          * will retry to break_cow on each pass, so should recover the page
 353          * in due course.  The important thing is to not let VM_MERGEABLE
 354          * be cleared while any such pages might remain in the area.
 355          */
 356         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
 357 }
 358
 359 static void break_cow(struct mm_struct *mm, unsigned long addr)
 360 {
 361         struct vm_area_struct *vma;
 362
 363         down_read(&mm->mmap_sem);
 364         if (ksm_test_exit(mm))
 365                 goto out;
 366         vma = find_vma(mm, addr);
 367         if (!vma || vma->vm_start > addr)
 368                 goto out;
 369         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
 370                 goto out;
 371         break_ksm(vma, addr);
 372 out:
 373         up_read(&mm->mmap_sem);
 374 }
 375
 376 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
 377 {
 378         struct mm_struct *mm = rmap_item->mm;
 379         unsigned long addr = rmap_item->address;
 380         struct vm_area_struct *vma;
 381         struct page *page;
 382
 383         down_read(&mm->mmap_sem);
 384         if (ksm_test_exit(mm))
 385                 goto out;
 386         vma = find_vma(mm, addr);
 387         if (!vma || vma->vm_start > addr)
 388                 goto out;
 389         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
 390                 goto out;
 391
 392         page = follow_page(vma, addr, FOLL_GET);
 393         if (!page)
 394                 goto out;
 395         if (PageAnon(page)) {
 396                 flush_anon_page(vma, page, addr);
 397                 flush_dcache_page(page);
 398         } else {
 399                 put_page(page);
 400 out:            page = NULL;
 401         }
 402         up_read(&mm->mmap_sem);
 403         return page;
 404 }
 405
 406 /*
 407  * get_ksm_page: checks if the page at the virtual address in rmap_item
 408  * is still PageKsm, in which case we can trust the content of the page,
 409  * and it returns the gotten page; but NULL if the page has been zapped.
 410  */
 411 static struct page *get_ksm_page(struct rmap_item *rmap_item)
 412 {
 413         struct page *page;
 414
 415         page = get_mergeable_page(rmap_item);
 416         if (page && !PageKsm(page)) {
 417                 put_page(page);
 418                 page = NULL;
 419         }
 420         return page;
 421 }
 422
 423 /*
 424  * Removing rmap_item from stable or unstable tree.
 425  * This function will clean the information from the stable/unstable tree.
 426  */
 427 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
 428 {
 429         if (in_stable_tree(rmap_item)) {
 430                 struct rmap_item *next_item = rmap_item->next;
 431
 432                 if (rmap_item->address & NODE_FLAG) {
 433                         if (next_item) {
 434                                 rb_replace_node(&rmap_item->node,
 435                                                 &next_item->node,
 436                                                 &root_stable_tree);
 437                                 next_item->address |= NODE_FLAG;
 438                                 ksm_pages_sharing--;
 439                         } else {
 440                                 rb_erase(&rmap_item->node, &root_stable_tree);
 441                                 ksm_pages_shared--;
 442                         }
 443                 } else {
 444                         struct rmap_item *prev_item = rmap_item->prev;
 445
 446                         BUG_ON(prev_item->next != rmap_item);
 447                         prev_item->next = next_item;
 448                         if (next_item) {
 449                                 BUG_ON(next_item->prev != rmap_item);
 450                                 next_item->prev = rmap_item->prev;
 451                         }
 452                         ksm_pages_sharing--;
 453                 }
 454
 455                 rmap_item->next = NULL;
 456
 457         } else if (rmap_item->address & NODE_FLAG) {
 458                 unsigned char age;
 459                 /*
 460                  * Usually ksmd can and must skip the rb_erase, because
 461                  * root_unstable_tree was already reset to RB_ROOT.
 462                  * But be careful when an mm is exiting: do the rb_erase
 463                  * if this rmap_item was inserted by this scan, rather
 464                  * than left over from before.
 465                  */
 466                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
 467                 BUG_ON(age > 1);
 468                 if (!age)
 469                         rb_erase(&rmap_item->node, &root_unstable_tree);
 470                 ksm_pages_unshared--;
 471         }
 472
 473         rmap_item->address &= PAGE_MASK;
 474
 475         cond_resched();         /* we're called from many long loops */
 476 }
 477
 478 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
 479                                        struct list_head *cur)
 480 {
 481         struct rmap_item *rmap_item;
 482
 483         while (cur != &mm_slot->rmap_list) {
 484                 rmap_item = list_entry(cur, struct rmap_item, link);
 485                 cur = cur->next;
 486                 remove_rmap_item_from_tree(rmap_item);
 487                 list_del(&rmap_item->link);
 488                 free_rmap_item(rmap_item);
 489         }
 490 }
 491
 492 /*
 493  * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
 494  * than check every pte of a given vma, the locking doesn't quite work for
 495  * that - an rmap_item is assigned to the stable tree after inserting ksm
 496  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
 497  * rmap_items from parent to child at fork time (so as not to waste time
 498  * if exit comes before the next scan reaches it).
 499  *
 500  * Similarly, although we'd like to remove rmap_items (so updating counts
 501  * and freeing memory) when unmerging an area, it's easier to leave that
 502  * to the next pass of ksmd - consider, for example, how ksmd might be
 503  * in cmp_and_merge_page on one of the rmap_items we would be removing.
 504  */
 505 static int unmerge_ksm_pages(struct vm_area_struct *vma,
 506                              unsigned long start, unsigned long end)
 507 {
 508         unsigned long addr;
 509         int err = 0;
 510
 511         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
 512                 if (ksm_test_exit(vma->vm_mm))
 513                         break;
 514                 if (signal_pending(current))
 515                         err = -ERESTARTSYS;
 516                 else
 517                         err = break_ksm(vma, addr);
 518         }
 519         return err;
 520 }
 521
 522 #ifdef CONFIG_SYSFS
 523 /*
 524  * Only called through the sysfs control interface:
 525  */
 526 static int unmerge_and_remove_all_rmap_items(void)
 527 {
 528         struct mm_slot *mm_slot;
 529         struct mm_struct *mm;
 530         struct vm_area_struct *vma;
 531         int err = 0;
 532
 533         spin_lock(&ksm_mmlist_lock);
 534         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
 535                                                 struct mm_slot, mm_list);
 536         spin_unlock(&ksm_mmlist_lock);
 537
 538         for (mm_slot = ksm_scan.mm_slot;
 539                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
 540                 mm = mm_slot->mm;
 541                 down_read(&mm->mmap_sem);
 542                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
 543                         if (ksm_test_exit(mm))
 544                                 break;
 545                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
 546                                 continue;
 547                         err = unmerge_ksm_pages(vma,
 548                                                 vma->vm_start, vma->vm_end);
 549                         if (err)
 550                                 goto error;
 551                 }
 552
 553                 remove_trailing_rmap_items(mm_slot, mm_slot->rmap_list.next);
 554
 555                 spin_lock(&ksm_mmlist_lock);
 556                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
 557                                                 struct mm_slot, mm_list);
 558                 if (ksm_test_exit(mm)) {
 559                         hlist_del(&mm_slot->link);
 560                         list_del(&mm_slot->mm_list);
 561                         spin_unlock(&ksm_mmlist_lock);
 562
 563                         free_mm_slot(mm_slot);
 564                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
 565                         up_read(&mm->mmap_sem);
 566                         mmdrop(mm);
 567                 } else {
 568                         spin_unlock(&ksm_mmlist_lock);
 569                         up_read(&mm->mmap_sem);
 570                 }
 571         }
 572
 573         ksm_scan.seqnr = 0;
 574         return 0;
 575
 576 error:
 577         up_read(&mm->mmap_sem);
 578         spin_lock(&ksm_mmlist_lock);
 579         ksm_scan.mm_slot = &ksm_mm_head;
 580         spin_unlock(&ksm_mmlist_lock);
 581         return err;
 582 }
 583 #endif /* CONFIG_SYSFS */
 584
 585 static u32 calc_checksum(struct page *page)
 586 {
 587         u32 checksum;
 588         void *addr = kmap_atomic(page, KM_USER0);
 589         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
 590         kunmap_atomic(addr, KM_USER0);
 591         return checksum;
 592 }
 593
 594 static int memcmp_pages(struct page *page1, struct page *page2)
 595 {
 596         char *addr1, *addr2;
 597         int ret;
 598
 599         addr1 = kmap_atomic(page1, KM_USER0);
 600         addr2 = kmap_atomic(page2, KM_USER1);
 601         ret = memcmp(addr1, addr2, PAGE_SIZE);
 602         kunmap_atomic(addr2, KM_USER1);
 603         kunmap_atomic(addr1, KM_USER0);
 604         return ret;
 605 }
 606
 607 static inline int pages_identical(struct page *page1, struct page *page2)
 608 {
 609         return !memcmp_pages(page1, page2);
 610 }
 611
 612 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
 613                               pte_t *orig_pte)
 614 {
 615         struct mm_struct *mm = vma->vm_mm;
 616         unsigned long addr;
 617         pte_t *ptep;
 618         spinlock_t *ptl;
 619         int swapped;
 620         int err = -EFAULT;
 621
 622         addr = page_address_in_vma(page, vma);
 623         if (addr == -EFAULT)
 624                 goto out;
 625
 626         ptep = page_check_address(page, mm, addr, &ptl, 0);
 627         if (!ptep)
 628                 goto out;
 629
 630         if (pte_write(*ptep)) {
 631                 pte_t entry;
 632
 633                 swapped = PageSwapCache(page);
 634                 flush_cache_page(vma, addr, page_to_pfn(page));
 635                 /*
 636                  * Ok this is tricky, when get_user_pages_fast() run it doesnt
 637                  * take any lock, therefore the check that we are going to make
 638                  * with the pagecount against the mapcount is racey and
 639                  * O_DIRECT can happen right after the check.
 640                  * So we clear the pte and flush the tlb before the check
 641                  * this assure us that no O_DIRECT can happen after the check
 642                  * or in the middle of the check.
 643                  */
 644                 entry = ptep_clear_flush(vma, addr, ptep);
 645                 /*
 646                  * Check that no O_DIRECT or similar I/O is in progress on the
 647                  * page
 648                  */
 649                 if ((page_mapcount(page) + 2 + swapped) != page_count(page)) {
 650                         set_pte_at_notify(mm, addr, ptep, entry);
 651                         goto out_unlock;
 652                 }
 653                 entry = pte_wrprotect(entry);
 654                 set_pte_at_notify(mm, addr, ptep, entry);
 655         }
 656         *orig_pte = *ptep;
 657         err = 0;
 658
 659 out_unlock:
 660         pte_unmap_unlock(ptep, ptl);
 661 out:
 662         return err;
 663 }
 664
 665 /**
 666  * replace_page - replace page in vma by new ksm page
 667  * @vma:      vma that holds the pte pointing to oldpage
 668  * @oldpage:  the page we are replacing by newpage
 669  * @newpage:  the ksm page we replace oldpage by
 670  * @orig_pte: the original value of the pte
 671  *
 672  * Returns 0 on success, -EFAULT on failure.
 673  */
 674 static int replace_page(struct vm_area_struct *vma, struct page *oldpage,
 675                         struct page *newpage, pte_t orig_pte)
 676 {
 677         struct mm_struct *mm = vma->vm_mm;
 678         pgd_t *pgd;
 679         pud_t *pud;
 680         pmd_t *pmd;
 681         pte_t *ptep;
 682         spinlock_t *ptl;
 683         unsigned long addr;
 684         pgprot_t prot;
 685         int err = -EFAULT;
 686
 687         prot = vm_get_page_prot(vma->vm_flags & ~VM_WRITE);
 688
 689         addr = page_address_in_vma(oldpage, vma);
 690         if (addr == -EFAULT)
 691                 goto out;
 692
 693         pgd = pgd_offset(mm, addr);
 694         if (!pgd_present(*pgd))
 695                 goto out;
 696
 697         pud = pud_offset(pgd, addr);
 698         if (!pud_present(*pud))
 699                 goto out;
 700
 701         pmd = pmd_offset(pud, addr);
 702         if (!pmd_present(*pmd))
 703                 goto out;
 704
 705         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
 706         if (!pte_same(*ptep, orig_pte)) {
 707                 pte_unmap_unlock(ptep, ptl);
 708                 goto out;
 709         }
 710
 711         get_page(newpage);
 712         page_add_ksm_rmap(newpage);
 713
 714         flush_cache_page(vma, addr, pte_pfn(*ptep));
 715         ptep_clear_flush(vma, addr, ptep);
 716         set_pte_at_notify(mm, addr, ptep, mk_pte(newpage, prot));
 717
 718         page_remove_rmap(oldpage);
 719         put_page(oldpage);
 720
 721         pte_unmap_unlock(ptep, ptl);
 722         err = 0;
 723 out:
 724         return err;
 725 }
 726
 727 /*
 728  * try_to_merge_one_page - take two pages and merge them into one
 729  * @vma: the vma that hold the pte pointing into oldpage
 730  * @oldpage: the page that we want to replace with newpage
 731  * @newpage: the page that we want to map instead of oldpage
 732  *
 733  * Note:
 734  * oldpage should be a PageAnon page, while newpage should be a PageKsm page,
 735  * or a newly allocated kernel page which page_add_ksm_rmap will make PageKsm.
 736  *
 737  * This function returns 0 if the pages were merged, -EFAULT otherwise.
 738  */
 739 static int try_to_merge_one_page(struct vm_area_struct *vma,
 740                                  struct page *oldpage,
 741                                  struct page *newpage)
 742 {
 743         pte_t orig_pte = __pte(0);
 744         int err = -EFAULT;
 745
 746         if (!(vma->vm_flags & VM_MERGEABLE))
 747                 goto out;
 748
 749         if (!PageAnon(oldpage))
 750                 goto out;
 751
 752         get_page(newpage);
 753         get_page(oldpage);
 754
 755         /*
 756          * We need the page lock to read a stable PageSwapCache in
 757          * write_protect_page().  We use trylock_page() instead of
 758          * lock_page() because we don't want to wait here - we
 759          * prefer to continue scanning and merging different pages,
 760          * then come back to this page when it is unlocked.
 761          */
 762         if (!trylock_page(oldpage))
 763                 goto out_putpage;
 764         /*
 765          * If this anonymous page is mapped only here, its pte may need
 766          * to be write-protected.  If it's mapped elsewhere, all of its
 767          * ptes are necessarily already write-protected.  But in either
 768          * case, we need to lock and check page_count is not raised.
 769          */
 770         if (write_protect_page(vma, oldpage, &orig_pte)) {
 771                 unlock_page(oldpage);
 772                 goto out_putpage;
 773         }
 774         unlock_page(oldpage);
 775
 776         if (pages_identical(oldpage, newpage))
 777                 err = replace_page(vma, oldpage, newpage, orig_pte);
 778
 779 out_putpage:
 780         put_page(oldpage);
 781         put_page(newpage);
 782 out:
 783         return err;
 784 }
 785
 786 /*
 787  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
 788  * but no new kernel page is allocated: kpage must already be a ksm page.
 789  */
 790 static int try_to_merge_with_ksm_page(struct mm_struct *mm1,
 791                                       unsigned long addr1,
 792                                       struct page *page1,
 793                                       struct page *kpage)
 794 {
 795         struct vm_area_struct *vma;
 796         int err = -EFAULT;
 797
 798         down_read(&mm1->mmap_sem);
 799         if (ksm_test_exit(mm1))
 800                 goto out;
 801
 802         vma = find_vma(mm1, addr1);
 803         if (!vma || vma->vm_start > addr1)
 804                 goto out;
 805
 806         err = try_to_merge_one_page(vma, page1, kpage);
 807 out:
 808         up_read(&mm1->mmap_sem);
 809         return err;
 810 }
 811
 812 /*
 813  * try_to_merge_two_pages - take two identical pages and prepare them
 814  * to be merged into one page.
 815  *
 816  * This function returns 0 if we successfully mapped two identical pages
 817  * into one page, -EFAULT otherwise.
 818  *
 819  * Note that this function allocates a new kernel page: if one of the pages
 820  * is already a ksm page, try_to_merge_with_ksm_page should be used.
 821  */
 822 static int try_to_merge_two_pages(struct mm_struct *mm1, unsigned long addr1,
 823                                   struct page *page1, struct mm_struct *mm2,
 824                                   unsigned long addr2, struct page *page2)
 825 {
 826         struct vm_area_struct *vma;
 827         struct page *kpage;
 828         int err = -EFAULT;
 829
 830         /*
 831          * The number of nodes in the stable tree
 832          * is the number of kernel pages that we hold.
 833          */
 834         if (ksm_max_kernel_pages &&
 835             ksm_max_kernel_pages <= ksm_pages_shared)
 836                 return err;
 837
 838         kpage = alloc_page(GFP_HIGHUSER);
 839         if (!kpage)
 840                 return err;
 841
 842         down_read(&mm1->mmap_sem);
 843         if (ksm_test_exit(mm1)) {
 844                 up_read(&mm1->mmap_sem);
 845                 goto out;
 846         }
 847         vma = find_vma(mm1, addr1);
 848         if (!vma || vma->vm_start > addr1) {
 849                 up_read(&mm1->mmap_sem);
 850                 goto out;
 851         }
 852
 853         copy_user_highpage(kpage, page1, addr1, vma);
 854         err = try_to_merge_one_page(vma, page1, kpage);
 855         up_read(&mm1->mmap_sem);
 856
 857         if (!err) {
 858                 err = try_to_merge_with_ksm_page(mm2, addr2, page2, kpage);
 859                 /*
 860                  * If that fails, we have a ksm page with only one pte
 861                  * pointing to it: so break it.
 862                  */
 863                 if (err)
 864                         break_cow(mm1, addr1);
 865         }
 866 out:
 867         put_page(kpage);
 868         return err;
 869 }
 870
 871 /*
 872  * stable_tree_search - search page inside the stable tree
 873  * @page: the page that we are searching identical pages to.
 874  * @page2: pointer into identical page that we are holding inside the stable
 875  *         tree that we have found.
 876  * @rmap_item: the reverse mapping item
 877  *
 878  * This function checks if there is a page inside the stable tree
 879  * with identical content to the page that we are scanning right now.
 880  *
 881  * This function return rmap_item pointer to the identical item if found,
 882  * NULL otherwise.
 883  */
 884 static struct rmap_item *stable_tree_search(struct page *page,
 885                                             struct page **page2,
 886                                             struct rmap_item *rmap_item)
 887 {
 888         struct rb_node *node = root_stable_tree.rb_node;
 889
 890         while (node) {
 891                 struct rmap_item *tree_rmap_item, *next_rmap_item;
 892                 int ret;
 893
 894                 tree_rmap_item = rb_entry(node, struct rmap_item, node);
 895                 while (tree_rmap_item) {
 896                         BUG_ON(!in_stable_tree(tree_rmap_item));
 897                         cond_resched();
 898                         page2[0] = get_ksm_page(tree_rmap_item);
 899                         if (page2[0])
 900                                 break;
 901                         next_rmap_item = tree_rmap_item->next;
 902                         remove_rmap_item_from_tree(tree_rmap_item);
 903                         tree_rmap_item = next_rmap_item;
 904                 }
 905                 if (!tree_rmap_item)
 906                         return NULL;
 907
 908                 ret = memcmp_pages(page, page2[0]);
 909
 910                 if (ret < 0) {
 911                         put_page(page2[0]);
 912                         node = node->rb_left;
 913                 } else if (ret > 0) {
 914                         put_page(page2[0]);
 915                         node = node->rb_right;
 916                 } else {
 917                         return tree_rmap_item;
 918                 }
 919         }
 920
 921         return NULL;
 922 }
 923
 924 /*
 925  * stable_tree_insert - insert rmap_item pointing to new ksm page
 926  * into the stable tree.
 927  *
 928  * @page: the page that we are searching identical page to inside the stable
 929  *        tree.
 930  * @rmap_item: pointer to the reverse mapping item.
 931  *
 932  * This function returns rmap_item if success, NULL otherwise.
 933  */
 934 static struct rmap_item *stable_tree_insert(struct page *page,
 935                                             struct rmap_item *rmap_item)
 936 {
 937         struct rb_node **new = &root_stable_tree.rb_node;
 938         struct rb_node *parent = NULL;
 939
 940         while (*new) {
 941                 struct rmap_item *tree_rmap_item, *next_rmap_item;
 942                 struct page *tree_page;
 943                 int ret;
 944
 945                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
 946                 while (tree_rmap_item) {
 947                         BUG_ON(!in_stable_tree(tree_rmap_item));
 948                         cond_resched();
 949                         tree_page = get_ksm_page(tree_rmap_item);
 950                         if (tree_page)
 951                                 break;
 952                         next_rmap_item = tree_rmap_item->next;
 953                         remove_rmap_item_from_tree(tree_rmap_item);
 954                         tree_rmap_item = next_rmap_item;
 955                 }
 956                 if (!tree_rmap_item)
 957                         return NULL;
 958
 959                 ret = memcmp_pages(page, tree_page);
 960                 put_page(tree_page);
 961
 962                 parent = *new;
 963                 if (ret < 0)
 964                         new = &parent->rb_left;
 965                 else if (ret > 0)
 966                         new = &parent->rb_right;
 967                 else {
 968                         /*
 969                          * It is not a bug that stable_tree_search() didn't
 970                          * find this node: because at that time our page was
 971                          * not yet write-protected, so may have changed since.
 972                          */
 973                         return NULL;
 974                 }
 975         }
 976
 977         rmap_item->address |= NODE_FLAG | STABLE_FLAG;
 978         rmap_item->next = NULL;
 979         rb_link_node(&rmap_item->node, parent, new);
 980         rb_insert_color(&rmap_item->node, &root_stable_tree);
 981
 982         ksm_pages_shared++;
 983         return rmap_item;
 984 }
 985
 986 /*
 987  * unstable_tree_search_insert - search and insert items into the unstable tree.
 988  *
 989  * @page: the page that we are going to search for identical page or to insert
 990  *        into the unstable tree
 991  * @page2: pointer into identical page that was found inside the unstable tree
 992  * @rmap_item: the reverse mapping item of page
 993  *
 994  * This function searches for a page in the unstable tree identical to the
 995  * page currently being scanned; and if no identical page is found in the
 996  * tree, we insert rmap_item as a new object into the unstable tree.
 997  *
 998  * This function returns pointer to rmap_item found to be identical
 999  * to the currently scanned page, NULL otherwise.
1000  *
1001  * This function does both searching and inserting, because they share
1002  * the same walking algorithm in an rbtree.
1003  */
1004 static struct rmap_item *unstable_tree_search_insert(struct page *page,
1005                                                 struct page **page2,
1006                                                 struct rmap_item *rmap_item)
1007 {
1008         struct rb_node **new = &root_unstable_tree.rb_node;
1009         struct rb_node *parent = NULL;
1010
1011         while (*new) {
1012                 struct rmap_item *tree_rmap_item;
1013                 int ret;
1014
1015                 cond_resched();
1016                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1017                 page2[0] = get_mergeable_page(tree_rmap_item);
1018                 if (!page2[0])
1019                         return NULL;
1020
1021                 /*
1022                  * Don't substitute an unswappable ksm page
1023                  * just for one good swappable forked page.
1024                  */
1025                 if (page == page2[0]) {
1026                         put_page(page2[0]);
1027                         return NULL;
1028                 }
1029
1030                 ret = memcmp_pages(page, page2[0]);
1031
1032                 parent = *new;
1033                 if (ret < 0) {
1034                         put_page(page2[0]);
1035                         new = &parent->rb_left;
1036                 } else if (ret > 0) {
1037                         put_page(page2[0]);
1038                         new = &parent->rb_right;
1039                 } else {
1040                         return tree_rmap_item;
1041                 }
1042         }
1043
1044         rmap_item->address |= NODE_FLAG;
1045         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1046         rb_link_node(&rmap_item->node, parent, new);
1047         rb_insert_color(&rmap_item->node, &root_unstable_tree);
1048
1049         ksm_pages_unshared++;
1050         return NULL;
1051 }
1052
1053 /*
1054  * stable_tree_append - add another rmap_item to the linked list of
1055  * rmap_items hanging off a given node of the stable tree, all sharing
1056  * the same ksm page.
1057  */
1058 static void stable_tree_append(struct rmap_item *rmap_item,
1059                                struct rmap_item *tree_rmap_item)
1060 {
1061         rmap_item->next = tree_rmap_item->next;
1062         rmap_item->prev = tree_rmap_item;
1063
1064         if (tree_rmap_item->next)
1065                 tree_rmap_item->next->prev = rmap_item;
1066
1067         tree_rmap_item->next = rmap_item;
1068         rmap_item->address |= STABLE_FLAG;
1069
1070         ksm_pages_sharing++;
1071 }
1072
1073 /*
1074  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1075  * if not, compare checksum to previous and if it's the same, see if page can
1076  * be inserted into the unstable tree, or merged with a page already there and
1077  * both transferred to the stable tree.
1078  *
1079  * @page: the page that we are searching identical page to.
1080  * @rmap_item: the reverse mapping into the virtual address of this page
1081  */
1082 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1083 {
1084         struct page *page2[1];
1085         struct rmap_item *tree_rmap_item;
1086         unsigned int checksum;
1087         int err;
1088
1089         if (in_stable_tree(rmap_item))
1090                 remove_rmap_item_from_tree(rmap_item);
1091
1092         /* We first start with searching the page inside the stable tree */
1093         tree_rmap_item = stable_tree_search(page, page2, rmap_item);
1094         if (tree_rmap_item) {
1095                 if (page == page2[0])                   /* forked */
1096                         err = 0;
1097                 else
1098                         err = try_to_merge_with_ksm_page(rmap_item->mm,
1099                                                          rmap_item->address,
1100                                                          page, page2[0]);
1101                 put_page(page2[0]);
1102
1103                 if (!err) {
1104                         /*
1105                          * The page was successfully merged:
1106                          * add its rmap_item to the stable tree.
1107                          */
1108                         stable_tree_append(rmap_item, tree_rmap_item);
1109                 }
1110                 return;
1111         }
1112
1113         /*
1114          * A ksm page might have got here by fork, but its other
1115          * references have already been removed from the stable tree.
1116          * Or it might be left over from a break_ksm which failed
1117          * when the mem_cgroup had reached its limit: try again now.
1118          */
1119         if (PageKsm(page))
1120                 break_cow(rmap_item->mm, rmap_item->address);
1121
1122         /*
1123          * In case the hash value of the page was changed from the last time we
1124          * have calculated it, this page to be changed frequely, therefore we
1125          * don't want to insert it to the unstable tree, and we don't want to
1126          * waste our time to search if there is something identical to it there.
1127          */
1128         checksum = calc_checksum(page);
1129         if (rmap_item->oldchecksum != checksum) {
1130                 rmap_item->oldchecksum = checksum;
1131                 return;
1132         }
1133
1134         tree_rmap_item = unstable_tree_search_insert(page, page2, rmap_item);
1135         if (tree_rmap_item) {
1136                 err = try_to_merge_two_pages(rmap_item->mm,
1137                                              rmap_item->address, page,
1138                                              tree_rmap_item->mm,
1139                                              tree_rmap_item->address, page2[0]);
1140                 /*
1141                  * As soon as we merge this page, we want to remove the
1142                  * rmap_item of the page we have merged with from the unstable
1143                  * tree, and insert it instead as new node in the stable tree.
1144                  */
1145                 if (!err) {
1146                         rb_erase(&tree_rmap_item->node, &root_unstable_tree);
1147                         tree_rmap_item->address &= ~NODE_FLAG;
1148                         ksm_pages_unshared--;
1149
1150                         /*
1151                          * If we fail to insert the page into the stable tree,
1152                          * we will have 2 virtual addresses that are pointing
1153                          * to a ksm page left outside the stable tree,
1154                          * in which case we need to break_cow on both.
1155                          */
1156                         if (stable_tree_insert(page2[0], tree_rmap_item))
1157                                 stable_tree_append(rmap_item, tree_rmap_item);
1158                         else {
1159                                 break_cow(tree_rmap_item->mm,
1160                                                 tree_rmap_item->address);
1161                                 break_cow(rmap_item->mm, rmap_item->address);
1162                         }
1163                 }
1164
1165                 put_page(page2[0]);
1166         }
1167 }
1168
1169 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1170                                             struct list_head *cur,
1171                                             unsigned long addr)
1172 {
1173         struct rmap_item *rmap_item;
1174
1175         while (cur != &mm_slot->rmap_list) {
1176                 rmap_item = list_entry(cur, struct rmap_item, link);
1177                 if ((rmap_item->address & PAGE_MASK) == addr) {
1178                         if (!in_stable_tree(rmap_item))
1179                                 remove_rmap_item_from_tree(rmap_item);
1180                         return rmap_item;
1181                 }
1182                 if (rmap_item->address > addr)
1183                         break;
1184                 cur = cur->next;
1185                 remove_rmap_item_from_tree(rmap_item);
1186                 list_del(&rmap_item->link);
1187                 free_rmap_item(rmap_item);
1188         }
1189
1190         rmap_item = alloc_rmap_item();
1191         if (rmap_item) {
1192                 /* It has already been zeroed */
1193                 rmap_item->mm = mm_slot->mm;
1194                 rmap_item->address = addr;
1195                 list_add_tail(&rmap_item->link, cur);
1196         }
1197         return rmap_item;
1198 }
1199
1200 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1201 {
1202         struct mm_struct *mm;
1203         struct mm_slot *slot;
1204         struct vm_area_struct *vma;
1205         struct rmap_item *rmap_item;
1206
1207         if (list_empty(&ksm_mm_head.mm_list))
1208                 return NULL;
1209
1210         slot = ksm_scan.mm_slot;
1211         if (slot == &ksm_mm_head) {
1212                 root_unstable_tree = RB_ROOT;
1213
1214                 spin_lock(&ksm_mmlist_lock);
1215                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1216                 ksm_scan.mm_slot = slot;
1217                 spin_unlock(&ksm_mmlist_lock);
1218 next_mm:
1219                 ksm_scan.address = 0;
1220                 ksm_scan.rmap_item = list_entry(&slot->rmap_list,
1221                                                 struct rmap_item, link);
1222         }
1223
1224         mm = slot->mm;
1225         down_read(&mm->mmap_sem);
1226         if (ksm_test_exit(mm))
1227                 vma = NULL;
1228         else
1229                 vma = find_vma(mm, ksm_scan.address);
1230
1231         for (; vma; vma = vma->vm_next) {
1232                 if (!(vma->vm_flags & VM_MERGEABLE))
1233                         continue;
1234                 if (ksm_scan.address < vma->vm_start)
1235                         ksm_scan.address = vma->vm_start;
1236                 if (!vma->anon_vma)
1237                         ksm_scan.address = vma->vm_end;
1238
1239                 while (ksm_scan.address < vma->vm_end) {
1240                         if (ksm_test_exit(mm))
1241                                 break;
1242                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1243                         if (*page && PageAnon(*page)) {
1244                                 flush_anon_page(vma, *page, ksm_scan.address);
1245                                 flush_dcache_page(*page);
1246                                 rmap_item = get_next_rmap_item(slot,
1247                                         ksm_scan.rmap_item->link.next,
1248                                         ksm_scan.address);
1249                                 if (rmap_item) {
1250                                         ksm_scan.rmap_item = rmap_item;
1251                                         ksm_scan.address += PAGE_SIZE;
1252                                 } else
1253                                         put_page(*page);
1254                                 up_read(&mm->mmap_sem);
1255                                 return rmap_item;
1256                         }
1257                         if (*page)
1258                                 put_page(*page);
1259                         ksm_scan.address += PAGE_SIZE;
1260                         cond_resched();
1261                 }
1262         }
1263
1264         if (ksm_test_exit(mm)) {
1265                 ksm_scan.address = 0;
1266                 ksm_scan.rmap_item = list_entry(&slot->rmap_list,
1267                                                 struct rmap_item, link);
1268         }
1269         /*
1270          * Nuke all the rmap_items that are above this current rmap:
1271          * because there were no VM_MERGEABLE vmas with such addresses.
1272          */
1273         remove_trailing_rmap_items(slot, ksm_scan.rmap_item->link.next);
1274
1275         spin_lock(&ksm_mmlist_lock);
1276         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1277                                                 struct mm_slot, mm_list);
1278         if (ksm_scan.address == 0) {
1279                 /*
1280                  * We've completed a full scan of all vmas, holding mmap_sem
1281                  * throughout, and found no VM_MERGEABLE: so do the same as
1282                  * __ksm_exit does to remove this mm from all our lists now.
1283                  * This applies either when cleaning up after __ksm_exit
1284                  * (but beware: we can reach here even before __ksm_exit),
1285                  * or when all VM_MERGEABLE areas have been unmapped (and
1286                  * mmap_sem then protects against race with MADV_MERGEABLE).
1287                  */
1288                 hlist_del(&slot->link);
1289                 list_del(&slot->mm_list);
1290                 spin_unlock(&ksm_mmlist_lock);
1291
1292                 free_mm_slot(slot);
1293                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1294                 up_read(&mm->mmap_sem);
1295                 mmdrop(mm);
1296         } else {
1297                 spin_unlock(&ksm_mmlist_lock);
1298                 up_read(&mm->mmap_sem);
1299         }
1300
1301         /* Repeat until we've completed scanning the whole list */
1302         slot = ksm_scan.mm_slot;
1303         if (slot != &ksm_mm_head)
1304                 goto next_mm;
1305
1306         ksm_scan.seqnr++;
1307         return NULL;
1308 }
1309
1310 /**
1311  * ksm_do_scan  - the ksm scanner main worker function.
1312  * @scan_npages - number of pages we want to scan before we return.
1313  */
1314 static void ksm_do_scan(unsigned int scan_npages)
1315 {
1316         struct rmap_item *rmap_item;
1317         struct page *page;
1318
1319         while (scan_npages--) {
1320                 cond_resched();
1321                 rmap_item = scan_get_next_rmap_item(&page);
1322                 if (!rmap_item)
1323                         return;
1324                 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1325                         cmp_and_merge_page(page, rmap_item);
1326                 else if (page_mapcount(page) == 1) {
1327                         /*
1328                          * Replace now-unshared ksm page by ordinary page.
1329                          */
1330                         break_cow(rmap_item->mm, rmap_item->address);
1331                         remove_rmap_item_from_tree(rmap_item);
1332                         rmap_item->oldchecksum = calc_checksum(page);
1333                 }
1334                 put_page(page);
1335         }
1336 }
1337
1338 static int ksmd_should_run(void)
1339 {
1340         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1341 }
1342
1343 static int ksm_scan_thread(void *nothing)
1344 {
1345         set_user_nice(current, 5);
1346
1347         while (!kthread_should_stop()) {
1348                 mutex_lock(&ksm_thread_mutex);
1349                 if (ksmd_should_run())
1350                         ksm_do_scan(ksm_thread_pages_to_scan);
1351                 mutex_unlock(&ksm_thread_mutex);
1352
1353                 if (ksmd_should_run()) {
1354                         schedule_timeout_interruptible(
1355                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1356                 } else {
1357                         wait_event_interruptible(ksm_thread_wait,
1358                                 ksmd_should_run() || kthread_should_stop());
1359                 }
1360         }
1361         return 0;
1362 }
1363
1364 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1365                 unsigned long end, int advice, unsigned long *vm_flags)
1366 {
1367         struct mm_struct *mm = vma->vm_mm;
1368         int err;
1369
1370         switch (advice) {
1371         case MADV_MERGEABLE:
1372                 /*
1373                  * Be somewhat over-protective for now!
1374                  */
1375                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1376                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1377                                  VM_RESERVED  | VM_HUGETLB | VM_INSERTPAGE |
1378                                  VM_MIXEDMAP  | VM_SAO))
1379                         return 0;               /* just ignore the advice */
1380
1381                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1382                         err = __ksm_enter(mm);
1383                         if (err)
1384                                 return err;
1385                 }
1386
1387                 *vm_flags |= VM_MERGEABLE;
1388                 break;
1389
1390         case MADV_UNMERGEABLE:
1391                 if (!(*vm_flags & VM_MERGEABLE))
1392                         return 0;               /* just ignore the advice */
1393
1394                 if (vma->anon_vma) {
1395                         err = unmerge_ksm_pages(vma, start, end);
1396                         if (err)
1397                                 return err;
1398                 }
1399
1400                 *vm_flags &= ~VM_MERGEABLE;
1401                 break;
1402         }
1403
1404         return 0;
1405 }
1406
1407 int __ksm_enter(struct mm_struct *mm)
1408 {
1409         struct mm_slot *mm_slot;
1410         int needs_wakeup;
1411
1412         mm_slot = alloc_mm_slot();
1413         if (!mm_slot)
1414                 return -ENOMEM;
1415
1416         /* Check ksm_run too?  Would need tighter locking */
1417         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1418
1419         spin_lock(&ksm_mmlist_lock);
1420         insert_to_mm_slots_hash(mm, mm_slot);
1421         /*
1422          * Insert just behind the scanning cursor, to let the area settle
1423          * down a little; when fork is followed by immediate exec, we don't
1424          * want ksmd to waste time setting up and tearing down an rmap_list.
1425          */
1426         list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1427         spin_unlock(&ksm_mmlist_lock);
1428
1429         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1430         atomic_inc(&mm->mm_count);
1431
1432         if (needs_wakeup)
1433                 wake_up_interruptible(&ksm_thread_wait);
1434
1435         return 0;
1436 }
1437
1438 void __ksm_exit(struct mm_struct *mm)
1439 {
1440         struct mm_slot *mm_slot;
1441         int easy_to_free = 0;
1442
1443         /*
1444          * This process is exiting: if it's straightforward (as is the
1445          * case when ksmd was never running), free mm_slot immediately.
1446          * But if it's at the cursor or has rmap_items linked to it, use
1447          * mmap_sem to synchronize with any break_cows before pagetables
1448          * are freed, and leave the mm_slot on the list for ksmd to free.
1449          * Beware: ksm may already have noticed it exiting and freed the slot.
1450          */
1451
1452         spin_lock(&ksm_mmlist_lock);
1453         mm_slot = get_mm_slot(mm);
1454         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1455                 if (list_empty(&mm_slot->rmap_list)) {
1456                         hlist_del(&mm_slot->link);
1457                         list_del(&mm_slot->mm_list);
1458                         easy_to_free = 1;
1459                 } else {
1460                         list_move(&mm_slot->mm_list,
1461                                   &ksm_scan.mm_slot->mm_list);
1462                 }
1463         }
1464         spin_unlock(&ksm_mmlist_lock);
1465
1466         if (easy_to_free) {
1467                 free_mm_slot(mm_slot);
1468                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1469                 mmdrop(mm);
1470         } else if (mm_slot) {
1471                 down_write(&mm->mmap_sem);
1472                 up_write(&mm->mmap_sem);
1473         }
1474 }
1475
1476 #ifdef CONFIG_SYSFS
1477 /*
1478  * This all compiles without CONFIG_SYSFS, but is a waste of space.
1479  */
1480
1481 #define KSM_ATTR_RO(_name) \
1482         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1483 #define KSM_ATTR(_name) \
1484         static struct kobj_attribute _name##_attr = \
1485                 __ATTR(_name, 0644, _name##_show, _name##_store)
1486
1487 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1488                                     struct kobj_attribute *attr, char *buf)
1489 {
1490         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1491 }
1492
1493 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1494                                      struct kobj_attribute *attr,
1495                                      const char *buf, size_t count)
1496 {
1497         unsigned long msecs;
1498         int err;
1499
1500         err = strict_strtoul(buf, 10, &msecs);
1501         if (err || msecs > UINT_MAX)
1502                 return -EINVAL;
1503
1504         ksm_thread_sleep_millisecs = msecs;
1505
1506         return count;
1507 }
1508 KSM_ATTR(sleep_millisecs);
1509
1510 static ssize_t pages_to_scan_show(struct kobject *kobj,
1511                                   struct kobj_attribute *attr, char *buf)
1512 {
1513         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1514 }
1515
1516 static ssize_t pages_to_scan_store(struct kobject *kobj,
1517                                    struct kobj_attribute *attr,
1518                                    const char *buf, size_t count)
1519 {
1520         int err;
1521         unsigned long nr_pages;
1522
1523         err = strict_strtoul(buf, 10, &nr_pages);
1524         if (err || nr_pages > UINT_MAX)
1525                 return -EINVAL;
1526
1527         ksm_thread_pages_to_scan = nr_pages;
1528
1529         return count;
1530 }
1531 KSM_ATTR(pages_to_scan);
1532
1533 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1534                         char *buf)
1535 {
1536         return sprintf(buf, "%u\n", ksm_run);
1537 }
1538
1539 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1540                          const char *buf, size_t count)
1541 {
1542         int err;
1543         unsigned long flags;
1544
1545         err = strict_strtoul(buf, 10, &flags);
1546         if (err || flags > UINT_MAX)
1547                 return -EINVAL;
1548         if (flags > KSM_RUN_UNMERGE)
1549                 return -EINVAL;
1550
1551         /*
1552          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1553          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1554          * breaking COW to free the unswappable pages_shared (but leaves
1555          * mm_slots on the list for when ksmd may be set running again).
1556          */
1557
1558         mutex_lock(&ksm_thread_mutex);
1559         if (ksm_run != flags) {
1560                 ksm_run = flags;
1561                 if (flags & KSM_RUN_UNMERGE) {
1562                         current->flags |= PF_OOM_ORIGIN;
1563                         err = unmerge_and_remove_all_rmap_items();
1564                         current->flags &= ~PF_OOM_ORIGIN;
1565                         if (err) {
1566                                 ksm_run = KSM_RUN_STOP;
1567                                 count = err;
1568                         }
1569                 }
1570         }
1571         mutex_unlock(&ksm_thread_mutex);
1572
1573         if (flags & KSM_RUN_MERGE)
1574                 wake_up_interruptible(&ksm_thread_wait);
1575
1576         return count;
1577 }
1578 KSM_ATTR(run);
1579
1580 static ssize_t max_kernel_pages_store(struct kobject *kobj,
1581                                       struct kobj_attribute *attr,
1582                                       const char *buf, size_t count)
1583 {
1584         int err;
1585         unsigned long nr_pages;
1586
1587         err = strict_strtoul(buf, 10, &nr_pages);
1588         if (err)
1589                 return -EINVAL;
1590
1591         ksm_max_kernel_pages = nr_pages;
1592
1593         return count;
1594 }
1595
1596 static ssize_t max_kernel_pages_show(struct kobject *kobj,
1597                                      struct kobj_attribute *attr, char *buf)
1598 {
1599         return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1600 }
1601 KSM_ATTR(max_kernel_pages);
1602
1603 static ssize_t pages_shared_show(struct kobject *kobj,
1604                                  struct kobj_attribute *attr, char *buf)
1605 {
1606         return sprintf(buf, "%lu\n", ksm_pages_shared);
1607 }
1608 KSM_ATTR_RO(pages_shared);
1609
1610 static ssize_t pages_sharing_show(struct kobject *kobj,
1611                                   struct kobj_attribute *attr, char *buf)
1612 {
1613         return sprintf(buf, "%lu\n", ksm_pages_sharing);
1614 }
1615 KSM_ATTR_RO(pages_sharing);
1616
1617 static ssize_t pages_unshared_show(struct kobject *kobj,
1618                                    struct kobj_attribute *attr, char *buf)
1619 {
1620         return sprintf(buf, "%lu\n", ksm_pages_unshared);
1621 }
1622 KSM_ATTR_RO(pages_unshared);
1623
1624 static ssize_t pages_volatile_show(struct kobject *kobj,
1625                                    struct kobj_attribute *attr, char *buf)
1626 {
1627         long ksm_pages_volatile;
1628
1629         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1630                                 - ksm_pages_sharing - ksm_pages_unshared;
1631         /*
1632          * It was not worth any locking to calculate that statistic,
1633          * but it might therefore sometimes be negative: conceal that.
1634          */
1635         if (ksm_pages_volatile < 0)
1636                 ksm_pages_volatile = 0;
1637         return sprintf(buf, "%ld\n", ksm_pages_volatile);
1638 }
1639 KSM_ATTR_RO(pages_volatile);
1640
1641 static ssize_t full_scans_show(struct kobject *kobj,
1642                                struct kobj_attribute *attr, char *buf)
1643 {
1644         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1645 }
1646 KSM_ATTR_RO(full_scans);
1647
1648 static struct attribute *ksm_attrs[] = {
1649         &sleep_millisecs_attr.attr,
1650         &pages_to_scan_attr.attr,
1651         &run_attr.attr,
1652         &max_kernel_pages_attr.attr,
1653         &pages_shared_attr.attr,
1654         &pages_sharing_attr.attr,
1655         &pages_unshared_attr.attr,
1656         &pages_volatile_attr.attr,
1657         &full_scans_attr.attr,
1658         NULL,
1659 };
1660
1661 static struct attribute_group ksm_attr_group = {
1662         .attrs = ksm_attrs,
1663         .name = "ksm",
1664 };
1665 #endif /* CONFIG_SYSFS */
1666
1667 static int __init ksm_init(void)
1668 {
1669         struct task_struct *ksm_thread;
1670         int err;
1671
1672         ksm_max_kernel_pages = totalram_pages / 4;
1673
1674         err = ksm_slab_init();
1675         if (err)
1676                 goto out;
1677
1678         err = mm_slots_hash_init();
1679         if (err)
1680                 goto out_free1;
1681
1682         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1683         if (IS_ERR(ksm_thread)) {
1684                 printk(KERN_ERR "ksm: creating kthread failed\n");
1685                 err = PTR_ERR(ksm_thread);
1686                 goto out_free2;
1687         }
1688
1689 #ifdef CONFIG_SYSFS
1690         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1691         if (err) {
1692                 printk(KERN_ERR "ksm: register sysfs failed\n");
1693                 kthread_stop(ksm_thread);
1694                 goto out_free2;
1695         }
1696 #else
1697         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
1698
1699 #endif /* CONFIG_SYSFS */
1700
1701         return 0;
1702
1703 out_free2:
1704         mm_slots_hash_free();
1705 out_free1:
1706         ksm_slab_free();
1707 out:
1708         return err;
1709 }
1710 module_init(ksm_init)