Back to home page

LXR

 
 

    


0001 /*
0002  * Memory merging support.
0003  *
0004  * This code enables dynamic sharing of identical pages found in different
0005  * memory areas, even if they are not shared by fork()
0006  *
0007  * Copyright (C) 2008-2009 Red Hat, Inc.
0008  * Authors:
0009  *  Izik Eidus
0010  *  Andrea Arcangeli
0011  *  Chris Wright
0012  *  Hugh Dickins
0013  *
0014  * This work is licensed under the terms of the GNU GPL, version 2.
0015  */
0016 
0017 #include <linux/errno.h>
0018 #include <linux/mm.h>
0019 #include <linux/fs.h>
0020 #include <linux/mman.h>
0021 #include <linux/sched.h>
0022 #include <linux/rwsem.h>
0023 #include <linux/pagemap.h>
0024 #include <linux/rmap.h>
0025 #include <linux/spinlock.h>
0026 #include <linux/jhash.h>
0027 #include <linux/delay.h>
0028 #include <linux/kthread.h>
0029 #include <linux/wait.h>
0030 #include <linux/slab.h>
0031 #include <linux/rbtree.h>
0032 #include <linux/memory.h>
0033 #include <linux/mmu_notifier.h>
0034 #include <linux/swap.h>
0035 #include <linux/ksm.h>
0036 #include <linux/hashtable.h>
0037 #include <linux/freezer.h>
0038 #include <linux/oom.h>
0039 #include <linux/numa.h>
0040 
0041 #include <asm/tlbflush.h>
0042 #include "internal.h"
0043 
0044 #ifdef CONFIG_NUMA
0045 #define NUMA(x)     (x)
0046 #define DO_NUMA(x)  do { (x); } while (0)
0047 #else
0048 #define NUMA(x)     (0)
0049 #define DO_NUMA(x)  do { } while (0)
0050 #endif
0051 
0052 /*
0053  * A few notes about the KSM scanning process,
0054  * to make it easier to understand the data structures below:
0055  *
0056  * In order to reduce excessive scanning, KSM sorts the memory pages by their
0057  * contents into a data structure that holds pointers to the pages' locations.
0058  *
0059  * Since the contents of the pages may change at any moment, KSM cannot just
0060  * insert the pages into a normal sorted tree and expect it to find anything.
0061  * Therefore KSM uses two data structures - the stable and the unstable tree.
0062  *
0063  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
0064  * by their contents.  Because each such page is write-protected, searching on
0065  * this tree is fully assured to be working (except when pages are unmapped),
0066  * and therefore this tree is called the stable tree.
0067  *
0068  * In addition to the stable tree, KSM uses a second data structure called the
0069  * unstable tree: this tree holds pointers to pages which have been found to
0070  * be "unchanged for a period of time".  The unstable tree sorts these pages
0071  * by their contents, but since they are not write-protected, KSM cannot rely
0072  * upon the unstable tree to work correctly - the unstable tree is liable to
0073  * be corrupted as its contents are modified, and so it is called unstable.
0074  *
0075  * KSM solves this problem by several techniques:
0076  *
0077  * 1) The unstable tree is flushed every time KSM completes scanning all
0078  *    memory areas, and then the tree is rebuilt again from the beginning.
0079  * 2) KSM will only insert into the unstable tree, pages whose hash value
0080  *    has not changed since the previous scan of all memory areas.
0081  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
0082  *    colors of the nodes and not on their contents, assuring that even when
0083  *    the tree gets "corrupted" it won't get out of balance, so scanning time
0084  *    remains the same (also, searching and inserting nodes in an rbtree uses
0085  *    the same algorithm, so we have no overhead when we flush and rebuild).
0086  * 4) KSM never flushes the stable tree, which means that even if it were to
0087  *    take 10 attempts to find a page in the unstable tree, once it is found,
0088  *    it is secured in the stable tree.  (When we scan a new page, we first
0089  *    compare it against the stable tree, and then against the unstable tree.)
0090  *
0091  * If the merge_across_nodes tunable is unset, then KSM maintains multiple
0092  * stable trees and multiple unstable trees: one of each for each NUMA node.
0093  */
0094 
0095 /**
0096  * struct mm_slot - ksm information per mm that is being scanned
0097  * @link: link to the mm_slots hash list
0098  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
0099  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
0100  * @mm: the mm that this information is valid for
0101  */
0102 struct mm_slot {
0103     struct hlist_node link;
0104     struct list_head mm_list;
0105     struct rmap_item *rmap_list;
0106     struct mm_struct *mm;
0107 };
0108 
0109 /**
0110  * struct ksm_scan - cursor for scanning
0111  * @mm_slot: the current mm_slot we are scanning
0112  * @address: the next address inside that to be scanned
0113  * @rmap_list: link to the next rmap to be scanned in the rmap_list
0114  * @seqnr: count of completed full scans (needed when removing unstable node)
0115  *
0116  * There is only the one ksm_scan instance of this cursor structure.
0117  */
0118 struct ksm_scan {
0119     struct mm_slot *mm_slot;
0120     unsigned long address;
0121     struct rmap_item **rmap_list;
0122     unsigned long seqnr;
0123 };
0124 
0125 /**
0126  * struct stable_node - node of the stable rbtree
0127  * @node: rb node of this ksm page in the stable tree
0128  * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
0129  * @list: linked into migrate_nodes, pending placement in the proper node tree
0130  * @hlist: hlist head of rmap_items using this ksm page
0131  * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
0132  * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
0133  */
0134 struct stable_node {
0135     union {
0136         struct rb_node node;    /* when node of stable tree */
0137         struct {        /* when listed for migration */
0138             struct list_head *head;
0139             struct list_head list;
0140         };
0141     };
0142     struct hlist_head hlist;
0143     unsigned long kpfn;
0144 #ifdef CONFIG_NUMA
0145     int nid;
0146 #endif
0147 };
0148 
0149 /**
0150  * struct rmap_item - reverse mapping item for virtual addresses
0151  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
0152  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
0153  * @nid: NUMA node id of unstable tree in which linked (may not match page)
0154  * @mm: the memory structure this rmap_item is pointing into
0155  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
0156  * @oldchecksum: previous checksum of the page at that virtual address
0157  * @node: rb node of this rmap_item in the unstable tree
0158  * @head: pointer to stable_node heading this list in the stable tree
0159  * @hlist: link into hlist of rmap_items hanging off that stable_node
0160  */
0161 struct rmap_item {
0162     struct rmap_item *rmap_list;
0163     union {
0164         struct anon_vma *anon_vma;  /* when stable */
0165 #ifdef CONFIG_NUMA
0166         int nid;        /* when node of unstable tree */
0167 #endif
0168     };
0169     struct mm_struct *mm;
0170     unsigned long address;      /* + low bits used for flags below */
0171     unsigned int oldchecksum;   /* when unstable */
0172     union {
0173         struct rb_node node;    /* when node of unstable tree */
0174         struct {        /* when listed from stable tree */
0175             struct stable_node *head;
0176             struct hlist_node hlist;
0177         };
0178     };
0179 };
0180 
0181 #define SEQNR_MASK  0x0ff   /* low bits of unstable tree seqnr */
0182 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
0183 #define STABLE_FLAG 0x200   /* is listed from the stable tree */
0184 
0185 /* The stable and unstable tree heads */
0186 static struct rb_root one_stable_tree[1] = { RB_ROOT };
0187 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
0188 static struct rb_root *root_stable_tree = one_stable_tree;
0189 static struct rb_root *root_unstable_tree = one_unstable_tree;
0190 
0191 /* Recently migrated nodes of stable tree, pending proper placement */
0192 static LIST_HEAD(migrate_nodes);
0193 
0194 #define MM_SLOTS_HASH_BITS 10
0195 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
0196 
0197 static struct mm_slot ksm_mm_head = {
0198     .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
0199 };
0200 static struct ksm_scan ksm_scan = {
0201     .mm_slot = &ksm_mm_head,
0202 };
0203 
0204 static struct kmem_cache *rmap_item_cache;
0205 static struct kmem_cache *stable_node_cache;
0206 static struct kmem_cache *mm_slot_cache;
0207 
0208 /* The number of nodes in the stable tree */
0209 static unsigned long ksm_pages_shared;
0210 
0211 /* The number of page slots additionally sharing those nodes */
0212 static unsigned long ksm_pages_sharing;
0213 
0214 /* The number of nodes in the unstable tree */
0215 static unsigned long ksm_pages_unshared;
0216 
0217 /* The number of rmap_items in use: to calculate pages_volatile */
0218 static unsigned long ksm_rmap_items;
0219 
0220 /* Number of pages ksmd should scan in one batch */
0221 static unsigned int ksm_thread_pages_to_scan = 100;
0222 
0223 /* Milliseconds ksmd should sleep between batches */
0224 static unsigned int ksm_thread_sleep_millisecs = 20;
0225 
0226 #ifdef CONFIG_NUMA
0227 /* Zeroed when merging across nodes is not allowed */
0228 static unsigned int ksm_merge_across_nodes = 1;
0229 static int ksm_nr_node_ids = 1;
0230 #else
0231 #define ksm_merge_across_nodes  1U
0232 #define ksm_nr_node_ids     1
0233 #endif
0234 
0235 #define KSM_RUN_STOP    0
0236 #define KSM_RUN_MERGE   1
0237 #define KSM_RUN_UNMERGE 2
0238 #define KSM_RUN_OFFLINE 4
0239 static unsigned long ksm_run = KSM_RUN_STOP;
0240 static void wait_while_offlining(void);
0241 
0242 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
0243 static DEFINE_MUTEX(ksm_thread_mutex);
0244 static DEFINE_SPINLOCK(ksm_mmlist_lock);
0245 
0246 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
0247         sizeof(struct __struct), __alignof__(struct __struct),\
0248         (__flags), NULL)
0249 
0250 static int __init ksm_slab_init(void)
0251 {
0252     rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
0253     if (!rmap_item_cache)
0254         goto out;
0255 
0256     stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
0257     if (!stable_node_cache)
0258         goto out_free1;
0259 
0260     mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
0261     if (!mm_slot_cache)
0262         goto out_free2;
0263 
0264     return 0;
0265 
0266 out_free2:
0267     kmem_cache_destroy(stable_node_cache);
0268 out_free1:
0269     kmem_cache_destroy(rmap_item_cache);
0270 out:
0271     return -ENOMEM;
0272 }
0273 
0274 static void __init ksm_slab_free(void)
0275 {
0276     kmem_cache_destroy(mm_slot_cache);
0277     kmem_cache_destroy(stable_node_cache);
0278     kmem_cache_destroy(rmap_item_cache);
0279     mm_slot_cache = NULL;
0280 }
0281 
0282 static inline struct rmap_item *alloc_rmap_item(void)
0283 {
0284     struct rmap_item *rmap_item;
0285 
0286     rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
0287                         __GFP_NORETRY | __GFP_NOWARN);
0288     if (rmap_item)
0289         ksm_rmap_items++;
0290     return rmap_item;
0291 }
0292 
0293 static inline void free_rmap_item(struct rmap_item *rmap_item)
0294 {
0295     ksm_rmap_items--;
0296     rmap_item->mm = NULL;   /* debug safety */
0297     kmem_cache_free(rmap_item_cache, rmap_item);
0298 }
0299 
0300 static inline struct stable_node *alloc_stable_node(void)
0301 {
0302     /*
0303      * The allocation can take too long with GFP_KERNEL when memory is under
0304      * pressure, which may lead to hung task warnings.  Adding __GFP_HIGH
0305      * grants access to memory reserves, helping to avoid this problem.
0306      */
0307     return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
0308 }
0309 
0310 static inline void free_stable_node(struct stable_node *stable_node)
0311 {
0312     kmem_cache_free(stable_node_cache, stable_node);
0313 }
0314 
0315 static inline struct mm_slot *alloc_mm_slot(void)
0316 {
0317     if (!mm_slot_cache) /* initialization failed */
0318         return NULL;
0319     return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
0320 }
0321 
0322 static inline void free_mm_slot(struct mm_slot *mm_slot)
0323 {
0324     kmem_cache_free(mm_slot_cache, mm_slot);
0325 }
0326 
0327 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
0328 {
0329     struct mm_slot *slot;
0330 
0331     hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
0332         if (slot->mm == mm)
0333             return slot;
0334 
0335     return NULL;
0336 }
0337 
0338 static void insert_to_mm_slots_hash(struct mm_struct *mm,
0339                     struct mm_slot *mm_slot)
0340 {
0341     mm_slot->mm = mm;
0342     hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
0343 }
0344 
0345 /*
0346  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
0347  * page tables after it has passed through ksm_exit() - which, if necessary,
0348  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
0349  * a special flag: they can just back out as soon as mm_users goes to zero.
0350  * ksm_test_exit() is used throughout to make this test for exit: in some
0351  * places for correctness, in some places just to avoid unnecessary work.
0352  */
0353 static inline bool ksm_test_exit(struct mm_struct *mm)
0354 {
0355     return atomic_read(&mm->mm_users) == 0;
0356 }
0357 
0358 /*
0359  * We use break_ksm to break COW on a ksm page: it's a stripped down
0360  *
0361  *  if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
0362  *      put_page(page);
0363  *
0364  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
0365  * in case the application has unmapped and remapped mm,addr meanwhile.
0366  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
0367  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
0368  *
0369  * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
0370  * of the process that owns 'vma'.  We also do not want to enforce
0371  * protection keys here anyway.
0372  */
0373 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
0374 {
0375     struct page *page;
0376     int ret = 0;
0377 
0378     do {
0379         cond_resched();
0380         page = follow_page(vma, addr,
0381                 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
0382         if (IS_ERR_OR_NULL(page))
0383             break;
0384         if (PageKsm(page))
0385             ret = handle_mm_fault(vma, addr,
0386                     FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
0387         else
0388             ret = VM_FAULT_WRITE;
0389         put_page(page);
0390     } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
0391     /*
0392      * We must loop because handle_mm_fault() may back out if there's
0393      * any difficulty e.g. if pte accessed bit gets updated concurrently.
0394      *
0395      * VM_FAULT_WRITE is what we have been hoping for: it indicates that
0396      * COW has been broken, even if the vma does not permit VM_WRITE;
0397      * but note that a concurrent fault might break PageKsm for us.
0398      *
0399      * VM_FAULT_SIGBUS could occur if we race with truncation of the
0400      * backing file, which also invalidates anonymous pages: that's
0401      * okay, that truncation will have unmapped the PageKsm for us.
0402      *
0403      * VM_FAULT_OOM: at the time of writing (late July 2009), setting
0404      * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
0405      * current task has TIF_MEMDIE set, and will be OOM killed on return
0406      * to user; and ksmd, having no mm, would never be chosen for that.
0407      *
0408      * But if the mm is in a limited mem_cgroup, then the fault may fail
0409      * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
0410      * even ksmd can fail in this way - though it's usually breaking ksm
0411      * just to undo a merge it made a moment before, so unlikely to oom.
0412      *
0413      * That's a pity: we might therefore have more kernel pages allocated
0414      * than we're counting as nodes in the stable tree; but ksm_do_scan
0415      * will retry to break_cow on each pass, so should recover the page
0416      * in due course.  The important thing is to not let VM_MERGEABLE
0417      * be cleared while any such pages might remain in the area.
0418      */
0419     return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
0420 }
0421 
0422 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
0423         unsigned long addr)
0424 {
0425     struct vm_area_struct *vma;
0426     if (ksm_test_exit(mm))
0427         return NULL;
0428     vma = find_vma(mm, addr);
0429     if (!vma || vma->vm_start > addr)
0430         return NULL;
0431     if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
0432         return NULL;
0433     return vma;
0434 }
0435 
0436 static void break_cow(struct rmap_item *rmap_item)
0437 {
0438     struct mm_struct *mm = rmap_item->mm;
0439     unsigned long addr = rmap_item->address;
0440     struct vm_area_struct *vma;
0441 
0442     /*
0443      * It is not an accident that whenever we want to break COW
0444      * to undo, we also need to drop a reference to the anon_vma.
0445      */
0446     put_anon_vma(rmap_item->anon_vma);
0447 
0448     down_read(&mm->mmap_sem);
0449     vma = find_mergeable_vma(mm, addr);
0450     if (vma)
0451         break_ksm(vma, addr);
0452     up_read(&mm->mmap_sem);
0453 }
0454 
0455 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
0456 {
0457     struct mm_struct *mm = rmap_item->mm;
0458     unsigned long addr = rmap_item->address;
0459     struct vm_area_struct *vma;
0460     struct page *page;
0461 
0462     down_read(&mm->mmap_sem);
0463     vma = find_mergeable_vma(mm, addr);
0464     if (!vma)
0465         goto out;
0466 
0467     page = follow_page(vma, addr, FOLL_GET);
0468     if (IS_ERR_OR_NULL(page))
0469         goto out;
0470     if (PageAnon(page)) {
0471         flush_anon_page(vma, page, addr);
0472         flush_dcache_page(page);
0473     } else {
0474         put_page(page);
0475 out:
0476         page = NULL;
0477     }
0478     up_read(&mm->mmap_sem);
0479     return page;
0480 }
0481 
0482 /*
0483  * This helper is used for getting right index into array of tree roots.
0484  * When merge_across_nodes knob is set to 1, there are only two rb-trees for
0485  * stable and unstable pages from all nodes with roots in index 0. Otherwise,
0486  * every node has its own stable and unstable tree.
0487  */
0488 static inline int get_kpfn_nid(unsigned long kpfn)
0489 {
0490     return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
0491 }
0492 
0493 static void remove_node_from_stable_tree(struct stable_node *stable_node)
0494 {
0495     struct rmap_item *rmap_item;
0496 
0497     hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
0498         if (rmap_item->hlist.next)
0499             ksm_pages_sharing--;
0500         else
0501             ksm_pages_shared--;
0502         put_anon_vma(rmap_item->anon_vma);
0503         rmap_item->address &= PAGE_MASK;
0504         cond_resched();
0505     }
0506 
0507     if (stable_node->head == &migrate_nodes)
0508         list_del(&stable_node->list);
0509     else
0510         rb_erase(&stable_node->node,
0511              root_stable_tree + NUMA(stable_node->nid));
0512     free_stable_node(stable_node);
0513 }
0514 
0515 /*
0516  * get_ksm_page: checks if the page indicated by the stable node
0517  * is still its ksm page, despite having held no reference to it.
0518  * In which case we can trust the content of the page, and it
0519  * returns the gotten page; but if the page has now been zapped,
0520  * remove the stale node from the stable tree and return NULL.
0521  * But beware, the stable node's page might be being migrated.
0522  *
0523  * You would expect the stable_node to hold a reference to the ksm page.
0524  * But if it increments the page's count, swapping out has to wait for
0525  * ksmd to come around again before it can free the page, which may take
0526  * seconds or even minutes: much too unresponsive.  So instead we use a
0527  * "keyhole reference": access to the ksm page from the stable node peeps
0528  * out through its keyhole to see if that page still holds the right key,
0529  * pointing back to this stable node.  This relies on freeing a PageAnon
0530  * page to reset its page->mapping to NULL, and relies on no other use of
0531  * a page to put something that might look like our key in page->mapping.
0532  * is on its way to being freed; but it is an anomaly to bear in mind.
0533  */
0534 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
0535 {
0536     struct page *page;
0537     void *expected_mapping;
0538     unsigned long kpfn;
0539 
0540     expected_mapping = (void *)((unsigned long)stable_node |
0541                     PAGE_MAPPING_KSM);
0542 again:
0543     kpfn = READ_ONCE(stable_node->kpfn);
0544     page = pfn_to_page(kpfn);
0545 
0546     /*
0547      * page is computed from kpfn, so on most architectures reading
0548      * page->mapping is naturally ordered after reading node->kpfn,
0549      * but on Alpha we need to be more careful.
0550      */
0551     smp_read_barrier_depends();
0552     if (READ_ONCE(page->mapping) != expected_mapping)
0553         goto stale;
0554 
0555     /*
0556      * We cannot do anything with the page while its refcount is 0.
0557      * Usually 0 means free, or tail of a higher-order page: in which
0558      * case this node is no longer referenced, and should be freed;
0559      * however, it might mean that the page is under page_freeze_refs().
0560      * The __remove_mapping() case is easy, again the node is now stale;
0561      * but if page is swapcache in migrate_page_move_mapping(), it might
0562      * still be our page, in which case it's essential to keep the node.
0563      */
0564     while (!get_page_unless_zero(page)) {
0565         /*
0566          * Another check for page->mapping != expected_mapping would
0567          * work here too.  We have chosen the !PageSwapCache test to
0568          * optimize the common case, when the page is or is about to
0569          * be freed: PageSwapCache is cleared (under spin_lock_irq)
0570          * in the freeze_refs section of __remove_mapping(); but Anon
0571          * page->mapping reset to NULL later, in free_pages_prepare().
0572          */
0573         if (!PageSwapCache(page))
0574             goto stale;
0575         cpu_relax();
0576     }
0577 
0578     if (READ_ONCE(page->mapping) != expected_mapping) {
0579         put_page(page);
0580         goto stale;
0581     }
0582 
0583     if (lock_it) {
0584         lock_page(page);
0585         if (READ_ONCE(page->mapping) != expected_mapping) {
0586             unlock_page(page);
0587             put_page(page);
0588             goto stale;
0589         }
0590     }
0591     return page;
0592 
0593 stale:
0594     /*
0595      * We come here from above when page->mapping or !PageSwapCache
0596      * suggests that the node is stale; but it might be under migration.
0597      * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
0598      * before checking whether node->kpfn has been changed.
0599      */
0600     smp_rmb();
0601     if (READ_ONCE(stable_node->kpfn) != kpfn)
0602         goto again;
0603     remove_node_from_stable_tree(stable_node);
0604     return NULL;
0605 }
0606 
0607 /*
0608  * Removing rmap_item from stable or unstable tree.
0609  * This function will clean the information from the stable/unstable tree.
0610  */
0611 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
0612 {
0613     if (rmap_item->address & STABLE_FLAG) {
0614         struct stable_node *stable_node;
0615         struct page *page;
0616 
0617         stable_node = rmap_item->head;
0618         page = get_ksm_page(stable_node, true);
0619         if (!page)
0620             goto out;
0621 
0622         hlist_del(&rmap_item->hlist);
0623         unlock_page(page);
0624         put_page(page);
0625 
0626         if (!hlist_empty(&stable_node->hlist))
0627             ksm_pages_sharing--;
0628         else
0629             ksm_pages_shared--;
0630 
0631         put_anon_vma(rmap_item->anon_vma);
0632         rmap_item->address &= PAGE_MASK;
0633 
0634     } else if (rmap_item->address & UNSTABLE_FLAG) {
0635         unsigned char age;
0636         /*
0637          * Usually ksmd can and must skip the rb_erase, because
0638          * root_unstable_tree was already reset to RB_ROOT.
0639          * But be careful when an mm is exiting: do the rb_erase
0640          * if this rmap_item was inserted by this scan, rather
0641          * than left over from before.
0642          */
0643         age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
0644         BUG_ON(age > 1);
0645         if (!age)
0646             rb_erase(&rmap_item->node,
0647                  root_unstable_tree + NUMA(rmap_item->nid));
0648         ksm_pages_unshared--;
0649         rmap_item->address &= PAGE_MASK;
0650     }
0651 out:
0652     cond_resched();     /* we're called from many long loops */
0653 }
0654 
0655 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
0656                        struct rmap_item **rmap_list)
0657 {
0658     while (*rmap_list) {
0659         struct rmap_item *rmap_item = *rmap_list;
0660         *rmap_list = rmap_item->rmap_list;
0661         remove_rmap_item_from_tree(rmap_item);
0662         free_rmap_item(rmap_item);
0663     }
0664 }
0665 
0666 /*
0667  * Though it's very tempting to unmerge rmap_items from stable tree rather
0668  * than check every pte of a given vma, the locking doesn't quite work for
0669  * that - an rmap_item is assigned to the stable tree after inserting ksm
0670  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
0671  * rmap_items from parent to child at fork time (so as not to waste time
0672  * if exit comes before the next scan reaches it).
0673  *
0674  * Similarly, although we'd like to remove rmap_items (so updating counts
0675  * and freeing memory) when unmerging an area, it's easier to leave that
0676  * to the next pass of ksmd - consider, for example, how ksmd might be
0677  * in cmp_and_merge_page on one of the rmap_items we would be removing.
0678  */
0679 static int unmerge_ksm_pages(struct vm_area_struct *vma,
0680                  unsigned long start, unsigned long end)
0681 {
0682     unsigned long addr;
0683     int err = 0;
0684 
0685     for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
0686         if (ksm_test_exit(vma->vm_mm))
0687             break;
0688         if (signal_pending(current))
0689             err = -ERESTARTSYS;
0690         else
0691             err = break_ksm(vma, addr);
0692     }
0693     return err;
0694 }
0695 
0696 #ifdef CONFIG_SYSFS
0697 /*
0698  * Only called through the sysfs control interface:
0699  */
0700 static int remove_stable_node(struct stable_node *stable_node)
0701 {
0702     struct page *page;
0703     int err;
0704 
0705     page = get_ksm_page(stable_node, true);
0706     if (!page) {
0707         /*
0708          * get_ksm_page did remove_node_from_stable_tree itself.
0709          */
0710         return 0;
0711     }
0712 
0713     if (WARN_ON_ONCE(page_mapped(page))) {
0714         /*
0715          * This should not happen: but if it does, just refuse to let
0716          * merge_across_nodes be switched - there is no need to panic.
0717          */
0718         err = -EBUSY;
0719     } else {
0720         /*
0721          * The stable node did not yet appear stale to get_ksm_page(),
0722          * since that allows for an unmapped ksm page to be recognized
0723          * right up until it is freed; but the node is safe to remove.
0724          * This page might be in a pagevec waiting to be freed,
0725          * or it might be PageSwapCache (perhaps under writeback),
0726          * or it might have been removed from swapcache a moment ago.
0727          */
0728         set_page_stable_node(page, NULL);
0729         remove_node_from_stable_tree(stable_node);
0730         err = 0;
0731     }
0732 
0733     unlock_page(page);
0734     put_page(page);
0735     return err;
0736 }
0737 
0738 static int remove_all_stable_nodes(void)
0739 {
0740     struct stable_node *stable_node, *next;
0741     int nid;
0742     int err = 0;
0743 
0744     for (nid = 0; nid < ksm_nr_node_ids; nid++) {
0745         while (root_stable_tree[nid].rb_node) {
0746             stable_node = rb_entry(root_stable_tree[nid].rb_node,
0747                         struct stable_node, node);
0748             if (remove_stable_node(stable_node)) {
0749                 err = -EBUSY;
0750                 break;  /* proceed to next nid */
0751             }
0752             cond_resched();
0753         }
0754     }
0755     list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
0756         if (remove_stable_node(stable_node))
0757             err = -EBUSY;
0758         cond_resched();
0759     }
0760     return err;
0761 }
0762 
0763 static int unmerge_and_remove_all_rmap_items(void)
0764 {
0765     struct mm_slot *mm_slot;
0766     struct mm_struct *mm;
0767     struct vm_area_struct *vma;
0768     int err = 0;
0769 
0770     spin_lock(&ksm_mmlist_lock);
0771     ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
0772                         struct mm_slot, mm_list);
0773     spin_unlock(&ksm_mmlist_lock);
0774 
0775     for (mm_slot = ksm_scan.mm_slot;
0776             mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
0777         mm = mm_slot->mm;
0778         down_read(&mm->mmap_sem);
0779         for (vma = mm->mmap; vma; vma = vma->vm_next) {
0780             if (ksm_test_exit(mm))
0781                 break;
0782             if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
0783                 continue;
0784             err = unmerge_ksm_pages(vma,
0785                         vma->vm_start, vma->vm_end);
0786             if (err)
0787                 goto error;
0788         }
0789 
0790         remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
0791         up_read(&mm->mmap_sem);
0792 
0793         spin_lock(&ksm_mmlist_lock);
0794         ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
0795                         struct mm_slot, mm_list);
0796         if (ksm_test_exit(mm)) {
0797             hash_del(&mm_slot->link);
0798             list_del(&mm_slot->mm_list);
0799             spin_unlock(&ksm_mmlist_lock);
0800 
0801             free_mm_slot(mm_slot);
0802             clear_bit(MMF_VM_MERGEABLE, &mm->flags);
0803             mmdrop(mm);
0804         } else
0805             spin_unlock(&ksm_mmlist_lock);
0806     }
0807 
0808     /* Clean up stable nodes, but don't worry if some are still busy */
0809     remove_all_stable_nodes();
0810     ksm_scan.seqnr = 0;
0811     return 0;
0812 
0813 error:
0814     up_read(&mm->mmap_sem);
0815     spin_lock(&ksm_mmlist_lock);
0816     ksm_scan.mm_slot = &ksm_mm_head;
0817     spin_unlock(&ksm_mmlist_lock);
0818     return err;
0819 }
0820 #endif /* CONFIG_SYSFS */
0821 
0822 static u32 calc_checksum(struct page *page)
0823 {
0824     u32 checksum;
0825     void *addr = kmap_atomic(page);
0826     checksum = jhash2(addr, PAGE_SIZE / 4, 17);
0827     kunmap_atomic(addr);
0828     return checksum;
0829 }
0830 
0831 static int memcmp_pages(struct page *page1, struct page *page2)
0832 {
0833     char *addr1, *addr2;
0834     int ret;
0835 
0836     addr1 = kmap_atomic(page1);
0837     addr2 = kmap_atomic(page2);
0838     ret = memcmp(addr1, addr2, PAGE_SIZE);
0839     kunmap_atomic(addr2);
0840     kunmap_atomic(addr1);
0841     return ret;
0842 }
0843 
0844 static inline int pages_identical(struct page *page1, struct page *page2)
0845 {
0846     return !memcmp_pages(page1, page2);
0847 }
0848 
0849 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
0850                   pte_t *orig_pte)
0851 {
0852     struct mm_struct *mm = vma->vm_mm;
0853     unsigned long addr;
0854     pte_t *ptep;
0855     spinlock_t *ptl;
0856     int swapped;
0857     int err = -EFAULT;
0858     unsigned long mmun_start;   /* For mmu_notifiers */
0859     unsigned long mmun_end;     /* For mmu_notifiers */
0860 
0861     addr = page_address_in_vma(page, vma);
0862     if (addr == -EFAULT)
0863         goto out;
0864 
0865     BUG_ON(PageTransCompound(page));
0866 
0867     mmun_start = addr;
0868     mmun_end   = addr + PAGE_SIZE;
0869     mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
0870 
0871     ptep = page_check_address(page, mm, addr, &ptl, 0);
0872     if (!ptep)
0873         goto out_mn;
0874 
0875     if (pte_write(*ptep) || pte_dirty(*ptep)) {
0876         pte_t entry;
0877 
0878         swapped = PageSwapCache(page);
0879         flush_cache_page(vma, addr, page_to_pfn(page));
0880         /*
0881          * Ok this is tricky, when get_user_pages_fast() run it doesn't
0882          * take any lock, therefore the check that we are going to make
0883          * with the pagecount against the mapcount is racey and
0884          * O_DIRECT can happen right after the check.
0885          * So we clear the pte and flush the tlb before the check
0886          * this assure us that no O_DIRECT can happen after the check
0887          * or in the middle of the check.
0888          */
0889         entry = ptep_clear_flush_notify(vma, addr, ptep);
0890         /*
0891          * Check that no O_DIRECT or similar I/O is in progress on the
0892          * page
0893          */
0894         if (page_mapcount(page) + 1 + swapped != page_count(page)) {
0895             set_pte_at(mm, addr, ptep, entry);
0896             goto out_unlock;
0897         }
0898         if (pte_dirty(entry))
0899             set_page_dirty(page);
0900         entry = pte_mkclean(pte_wrprotect(entry));
0901         set_pte_at_notify(mm, addr, ptep, entry);
0902     }
0903     *orig_pte = *ptep;
0904     err = 0;
0905 
0906 out_unlock:
0907     pte_unmap_unlock(ptep, ptl);
0908 out_mn:
0909     mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
0910 out:
0911     return err;
0912 }
0913 
0914 /**
0915  * replace_page - replace page in vma by new ksm page
0916  * @vma:      vma that holds the pte pointing to page
0917  * @page:     the page we are replacing by kpage
0918  * @kpage:    the ksm page we replace page by
0919  * @orig_pte: the original value of the pte
0920  *
0921  * Returns 0 on success, -EFAULT on failure.
0922  */
0923 static int replace_page(struct vm_area_struct *vma, struct page *page,
0924             struct page *kpage, pte_t orig_pte)
0925 {
0926     struct mm_struct *mm = vma->vm_mm;
0927     pmd_t *pmd;
0928     pte_t *ptep;
0929     spinlock_t *ptl;
0930     unsigned long addr;
0931     int err = -EFAULT;
0932     unsigned long mmun_start;   /* For mmu_notifiers */
0933     unsigned long mmun_end;     /* For mmu_notifiers */
0934 
0935     addr = page_address_in_vma(page, vma);
0936     if (addr == -EFAULT)
0937         goto out;
0938 
0939     pmd = mm_find_pmd(mm, addr);
0940     if (!pmd)
0941         goto out;
0942 
0943     mmun_start = addr;
0944     mmun_end   = addr + PAGE_SIZE;
0945     mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
0946 
0947     ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
0948     if (!pte_same(*ptep, orig_pte)) {
0949         pte_unmap_unlock(ptep, ptl);
0950         goto out_mn;
0951     }
0952 
0953     get_page(kpage);
0954     page_add_anon_rmap(kpage, vma, addr, false);
0955 
0956     flush_cache_page(vma, addr, pte_pfn(*ptep));
0957     ptep_clear_flush_notify(vma, addr, ptep);
0958     set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
0959 
0960     page_remove_rmap(page, false);
0961     if (!page_mapped(page))
0962         try_to_free_swap(page);
0963     put_page(page);
0964 
0965     pte_unmap_unlock(ptep, ptl);
0966     err = 0;
0967 out_mn:
0968     mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
0969 out:
0970     return err;
0971 }
0972 
0973 /*
0974  * try_to_merge_one_page - take two pages and merge them into one
0975  * @vma: the vma that holds the pte pointing to page
0976  * @page: the PageAnon page that we want to replace with kpage
0977  * @kpage: the PageKsm page that we want to map instead of page,
0978  *         or NULL the first time when we want to use page as kpage.
0979  *
0980  * This function returns 0 if the pages were merged, -EFAULT otherwise.
0981  */
0982 static int try_to_merge_one_page(struct vm_area_struct *vma,
0983                  struct page *page, struct page *kpage)
0984 {
0985     pte_t orig_pte = __pte(0);
0986     int err = -EFAULT;
0987 
0988     if (page == kpage)          /* ksm page forked */
0989         return 0;
0990 
0991     if (!PageAnon(page))
0992         goto out;
0993 
0994     /*
0995      * We need the page lock to read a stable PageSwapCache in
0996      * write_protect_page().  We use trylock_page() instead of
0997      * lock_page() because we don't want to wait here - we
0998      * prefer to continue scanning and merging different pages,
0999      * then come back to this page when it is unlocked.
1000      */
1001     if (!trylock_page(page))
1002         goto out;
1003 
1004     if (PageTransCompound(page)) {
1005         err = split_huge_page(page);
1006         if (err)
1007             goto out_unlock;
1008     }
1009 
1010     /*
1011      * If this anonymous page is mapped only here, its pte may need
1012      * to be write-protected.  If it's mapped elsewhere, all of its
1013      * ptes are necessarily already write-protected.  But in either
1014      * case, we need to lock and check page_count is not raised.
1015      */
1016     if (write_protect_page(vma, page, &orig_pte) == 0) {
1017         if (!kpage) {
1018             /*
1019              * While we hold page lock, upgrade page from
1020              * PageAnon+anon_vma to PageKsm+NULL stable_node:
1021              * stable_tree_insert() will update stable_node.
1022              */
1023             set_page_stable_node(page, NULL);
1024             mark_page_accessed(page);
1025             /*
1026              * Page reclaim just frees a clean page with no dirty
1027              * ptes: make sure that the ksm page would be swapped.
1028              */
1029             if (!PageDirty(page))
1030                 SetPageDirty(page);
1031             err = 0;
1032         } else if (pages_identical(page, kpage))
1033             err = replace_page(vma, page, kpage, orig_pte);
1034     }
1035 
1036     if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1037         munlock_vma_page(page);
1038         if (!PageMlocked(kpage)) {
1039             unlock_page(page);
1040             lock_page(kpage);
1041             mlock_vma_page(kpage);
1042             page = kpage;       /* for final unlock */
1043         }
1044     }
1045 
1046 out_unlock:
1047     unlock_page(page);
1048 out:
1049     return err;
1050 }
1051 
1052 /*
1053  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1054  * but no new kernel page is allocated: kpage must already be a ksm page.
1055  *
1056  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1057  */
1058 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1059                       struct page *page, struct page *kpage)
1060 {
1061     struct mm_struct *mm = rmap_item->mm;
1062     struct vm_area_struct *vma;
1063     int err = -EFAULT;
1064 
1065     down_read(&mm->mmap_sem);
1066     vma = find_mergeable_vma(mm, rmap_item->address);
1067     if (!vma)
1068         goto out;
1069 
1070     err = try_to_merge_one_page(vma, page, kpage);
1071     if (err)
1072         goto out;
1073 
1074     /* Unstable nid is in union with stable anon_vma: remove first */
1075     remove_rmap_item_from_tree(rmap_item);
1076 
1077     /* Must get reference to anon_vma while still holding mmap_sem */
1078     rmap_item->anon_vma = vma->anon_vma;
1079     get_anon_vma(vma->anon_vma);
1080 out:
1081     up_read(&mm->mmap_sem);
1082     return err;
1083 }
1084 
1085 /*
1086  * try_to_merge_two_pages - take two identical pages and prepare them
1087  * to be merged into one page.
1088  *
1089  * This function returns the kpage if we successfully merged two identical
1090  * pages into one ksm page, NULL otherwise.
1091  *
1092  * Note that this function upgrades page to ksm page: if one of the pages
1093  * is already a ksm page, try_to_merge_with_ksm_page should be used.
1094  */
1095 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1096                        struct page *page,
1097                        struct rmap_item *tree_rmap_item,
1098                        struct page *tree_page)
1099 {
1100     int err;
1101 
1102     err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1103     if (!err) {
1104         err = try_to_merge_with_ksm_page(tree_rmap_item,
1105                             tree_page, page);
1106         /*
1107          * If that fails, we have a ksm page with only one pte
1108          * pointing to it: so break it.
1109          */
1110         if (err)
1111             break_cow(rmap_item);
1112     }
1113     return err ? NULL : page;
1114 }
1115 
1116 /*
1117  * stable_tree_search - search for page inside the stable tree
1118  *
1119  * This function checks if there is a page inside the stable tree
1120  * with identical content to the page that we are scanning right now.
1121  *
1122  * This function returns the stable tree node of identical content if found,
1123  * NULL otherwise.
1124  */
1125 static struct page *stable_tree_search(struct page *page)
1126 {
1127     int nid;
1128     struct rb_root *root;
1129     struct rb_node **new;
1130     struct rb_node *parent;
1131     struct stable_node *stable_node;
1132     struct stable_node *page_node;
1133 
1134     page_node = page_stable_node(page);
1135     if (page_node && page_node->head != &migrate_nodes) {
1136         /* ksm page forked */
1137         get_page(page);
1138         return page;
1139     }
1140 
1141     nid = get_kpfn_nid(page_to_pfn(page));
1142     root = root_stable_tree + nid;
1143 again:
1144     new = &root->rb_node;
1145     parent = NULL;
1146 
1147     while (*new) {
1148         struct page *tree_page;
1149         int ret;
1150 
1151         cond_resched();
1152         stable_node = rb_entry(*new, struct stable_node, node);
1153         tree_page = get_ksm_page(stable_node, false);
1154         if (!tree_page) {
1155             /*
1156              * If we walked over a stale stable_node,
1157              * get_ksm_page() will call rb_erase() and it
1158              * may rebalance the tree from under us. So
1159              * restart the search from scratch. Returning
1160              * NULL would be safe too, but we'd generate
1161              * false negative insertions just because some
1162              * stable_node was stale.
1163              */
1164             goto again;
1165         }
1166 
1167         ret = memcmp_pages(page, tree_page);
1168         put_page(tree_page);
1169 
1170         parent = *new;
1171         if (ret < 0)
1172             new = &parent->rb_left;
1173         else if (ret > 0)
1174             new = &parent->rb_right;
1175         else {
1176             /*
1177              * Lock and unlock the stable_node's page (which
1178              * might already have been migrated) so that page
1179              * migration is sure to notice its raised count.
1180              * It would be more elegant to return stable_node
1181              * than kpage, but that involves more changes.
1182              */
1183             tree_page = get_ksm_page(stable_node, true);
1184             if (tree_page) {
1185                 unlock_page(tree_page);
1186                 if (get_kpfn_nid(stable_node->kpfn) !=
1187                         NUMA(stable_node->nid)) {
1188                     put_page(tree_page);
1189                     goto replace;
1190                 }
1191                 return tree_page;
1192             }
1193             /*
1194              * There is now a place for page_node, but the tree may
1195              * have been rebalanced, so re-evaluate parent and new.
1196              */
1197             if (page_node)
1198                 goto again;
1199             return NULL;
1200         }
1201     }
1202 
1203     if (!page_node)
1204         return NULL;
1205 
1206     list_del(&page_node->list);
1207     DO_NUMA(page_node->nid = nid);
1208     rb_link_node(&page_node->node, parent, new);
1209     rb_insert_color(&page_node->node, root);
1210     get_page(page);
1211     return page;
1212 
1213 replace:
1214     if (page_node) {
1215         list_del(&page_node->list);
1216         DO_NUMA(page_node->nid = nid);
1217         rb_replace_node(&stable_node->node, &page_node->node, root);
1218         get_page(page);
1219     } else {
1220         rb_erase(&stable_node->node, root);
1221         page = NULL;
1222     }
1223     stable_node->head = &migrate_nodes;
1224     list_add(&stable_node->list, stable_node->head);
1225     return page;
1226 }
1227 
1228 /*
1229  * stable_tree_insert - insert stable tree node pointing to new ksm page
1230  * into the stable tree.
1231  *
1232  * This function returns the stable tree node just allocated on success,
1233  * NULL otherwise.
1234  */
1235 static struct stable_node *stable_tree_insert(struct page *kpage)
1236 {
1237     int nid;
1238     unsigned long kpfn;
1239     struct rb_root *root;
1240     struct rb_node **new;
1241     struct rb_node *parent;
1242     struct stable_node *stable_node;
1243 
1244     kpfn = page_to_pfn(kpage);
1245     nid = get_kpfn_nid(kpfn);
1246     root = root_stable_tree + nid;
1247 again:
1248     parent = NULL;
1249     new = &root->rb_node;
1250 
1251     while (*new) {
1252         struct page *tree_page;
1253         int ret;
1254 
1255         cond_resched();
1256         stable_node = rb_entry(*new, struct stable_node, node);
1257         tree_page = get_ksm_page(stable_node, false);
1258         if (!tree_page) {
1259             /*
1260              * If we walked over a stale stable_node,
1261              * get_ksm_page() will call rb_erase() and it
1262              * may rebalance the tree from under us. So
1263              * restart the search from scratch. Returning
1264              * NULL would be safe too, but we'd generate
1265              * false negative insertions just because some
1266              * stable_node was stale.
1267              */
1268             goto again;
1269         }
1270 
1271         ret = memcmp_pages(kpage, tree_page);
1272         put_page(tree_page);
1273 
1274         parent = *new;
1275         if (ret < 0)
1276             new = &parent->rb_left;
1277         else if (ret > 0)
1278             new = &parent->rb_right;
1279         else {
1280             /*
1281              * It is not a bug that stable_tree_search() didn't
1282              * find this node: because at that time our page was
1283              * not yet write-protected, so may have changed since.
1284              */
1285             return NULL;
1286         }
1287     }
1288 
1289     stable_node = alloc_stable_node();
1290     if (!stable_node)
1291         return NULL;
1292 
1293     INIT_HLIST_HEAD(&stable_node->hlist);
1294     stable_node->kpfn = kpfn;
1295     set_page_stable_node(kpage, stable_node);
1296     DO_NUMA(stable_node->nid = nid);
1297     rb_link_node(&stable_node->node, parent, new);
1298     rb_insert_color(&stable_node->node, root);
1299 
1300     return stable_node;
1301 }
1302 
1303 /*
1304  * unstable_tree_search_insert - search for identical page,
1305  * else insert rmap_item into the unstable tree.
1306  *
1307  * This function searches for a page in the unstable tree identical to the
1308  * page currently being scanned; and if no identical page is found in the
1309  * tree, we insert rmap_item as a new object into the unstable tree.
1310  *
1311  * This function returns pointer to rmap_item found to be identical
1312  * to the currently scanned page, NULL otherwise.
1313  *
1314  * This function does both searching and inserting, because they share
1315  * the same walking algorithm in an rbtree.
1316  */
1317 static
1318 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1319                           struct page *page,
1320                           struct page **tree_pagep)
1321 {
1322     struct rb_node **new;
1323     struct rb_root *root;
1324     struct rb_node *parent = NULL;
1325     int nid;
1326 
1327     nid = get_kpfn_nid(page_to_pfn(page));
1328     root = root_unstable_tree + nid;
1329     new = &root->rb_node;
1330 
1331     while (*new) {
1332         struct rmap_item *tree_rmap_item;
1333         struct page *tree_page;
1334         int ret;
1335 
1336         cond_resched();
1337         tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1338         tree_page = get_mergeable_page(tree_rmap_item);
1339         if (!tree_page)
1340             return NULL;
1341 
1342         /*
1343          * Don't substitute a ksm page for a forked page.
1344          */
1345         if (page == tree_page) {
1346             put_page(tree_page);
1347             return NULL;
1348         }
1349 
1350         ret = memcmp_pages(page, tree_page);
1351 
1352         parent = *new;
1353         if (ret < 0) {
1354             put_page(tree_page);
1355             new = &parent->rb_left;
1356         } else if (ret > 0) {
1357             put_page(tree_page);
1358             new = &parent->rb_right;
1359         } else if (!ksm_merge_across_nodes &&
1360                page_to_nid(tree_page) != nid) {
1361             /*
1362              * If tree_page has been migrated to another NUMA node,
1363              * it will be flushed out and put in the right unstable
1364              * tree next time: only merge with it when across_nodes.
1365              */
1366             put_page(tree_page);
1367             return NULL;
1368         } else {
1369             *tree_pagep = tree_page;
1370             return tree_rmap_item;
1371         }
1372     }
1373 
1374     rmap_item->address |= UNSTABLE_FLAG;
1375     rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1376     DO_NUMA(rmap_item->nid = nid);
1377     rb_link_node(&rmap_item->node, parent, new);
1378     rb_insert_color(&rmap_item->node, root);
1379 
1380     ksm_pages_unshared++;
1381     return NULL;
1382 }
1383 
1384 /*
1385  * stable_tree_append - add another rmap_item to the linked list of
1386  * rmap_items hanging off a given node of the stable tree, all sharing
1387  * the same ksm page.
1388  */
1389 static void stable_tree_append(struct rmap_item *rmap_item,
1390                    struct stable_node *stable_node)
1391 {
1392     rmap_item->head = stable_node;
1393     rmap_item->address |= STABLE_FLAG;
1394     hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1395 
1396     if (rmap_item->hlist.next)
1397         ksm_pages_sharing++;
1398     else
1399         ksm_pages_shared++;
1400 }
1401 
1402 /*
1403  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1404  * if not, compare checksum to previous and if it's the same, see if page can
1405  * be inserted into the unstable tree, or merged with a page already there and
1406  * both transferred to the stable tree.
1407  *
1408  * @page: the page that we are searching identical page to.
1409  * @rmap_item: the reverse mapping into the virtual address of this page
1410  */
1411 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1412 {
1413     struct rmap_item *tree_rmap_item;
1414     struct page *tree_page = NULL;
1415     struct stable_node *stable_node;
1416     struct page *kpage;
1417     unsigned int checksum;
1418     int err;
1419 
1420     stable_node = page_stable_node(page);
1421     if (stable_node) {
1422         if (stable_node->head != &migrate_nodes &&
1423             get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1424             rb_erase(&stable_node->node,
1425                  root_stable_tree + NUMA(stable_node->nid));
1426             stable_node->head = &migrate_nodes;
1427             list_add(&stable_node->list, stable_node->head);
1428         }
1429         if (stable_node->head != &migrate_nodes &&
1430             rmap_item->head == stable_node)
1431             return;
1432     }
1433 
1434     /* We first start with searching the page inside the stable tree */
1435     kpage = stable_tree_search(page);
1436     if (kpage == page && rmap_item->head == stable_node) {
1437         put_page(kpage);
1438         return;
1439     }
1440 
1441     remove_rmap_item_from_tree(rmap_item);
1442 
1443     if (kpage) {
1444         err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1445         if (!err) {
1446             /*
1447              * The page was successfully merged:
1448              * add its rmap_item to the stable tree.
1449              */
1450             lock_page(kpage);
1451             stable_tree_append(rmap_item, page_stable_node(kpage));
1452             unlock_page(kpage);
1453         }
1454         put_page(kpage);
1455         return;
1456     }
1457 
1458     /*
1459      * If the hash value of the page has changed from the last time
1460      * we calculated it, this page is changing frequently: therefore we
1461      * don't want to insert it in the unstable tree, and we don't want
1462      * to waste our time searching for something identical to it there.
1463      */
1464     checksum = calc_checksum(page);
1465     if (rmap_item->oldchecksum != checksum) {
1466         rmap_item->oldchecksum = checksum;
1467         return;
1468     }
1469 
1470     tree_rmap_item =
1471         unstable_tree_search_insert(rmap_item, page, &tree_page);
1472     if (tree_rmap_item) {
1473         kpage = try_to_merge_two_pages(rmap_item, page,
1474                         tree_rmap_item, tree_page);
1475         put_page(tree_page);
1476         if (kpage) {
1477             /*
1478              * The pages were successfully merged: insert new
1479              * node in the stable tree and add both rmap_items.
1480              */
1481             lock_page(kpage);
1482             stable_node = stable_tree_insert(kpage);
1483             if (stable_node) {
1484                 stable_tree_append(tree_rmap_item, stable_node);
1485                 stable_tree_append(rmap_item, stable_node);
1486             }
1487             unlock_page(kpage);
1488 
1489             /*
1490              * If we fail to insert the page into the stable tree,
1491              * we will have 2 virtual addresses that are pointing
1492              * to a ksm page left outside the stable tree,
1493              * in which case we need to break_cow on both.
1494              */
1495             if (!stable_node) {
1496                 break_cow(tree_rmap_item);
1497                 break_cow(rmap_item);
1498             }
1499         }
1500     }
1501 }
1502 
1503 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1504                         struct rmap_item **rmap_list,
1505                         unsigned long addr)
1506 {
1507     struct rmap_item *rmap_item;
1508 
1509     while (*rmap_list) {
1510         rmap_item = *rmap_list;
1511         if ((rmap_item->address & PAGE_MASK) == addr)
1512             return rmap_item;
1513         if (rmap_item->address > addr)
1514             break;
1515         *rmap_list = rmap_item->rmap_list;
1516         remove_rmap_item_from_tree(rmap_item);
1517         free_rmap_item(rmap_item);
1518     }
1519 
1520     rmap_item = alloc_rmap_item();
1521     if (rmap_item) {
1522         /* It has already been zeroed */
1523         rmap_item->mm = mm_slot->mm;
1524         rmap_item->address = addr;
1525         rmap_item->rmap_list = *rmap_list;
1526         *rmap_list = rmap_item;
1527     }
1528     return rmap_item;
1529 }
1530 
1531 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1532 {
1533     struct mm_struct *mm;
1534     struct mm_slot *slot;
1535     struct vm_area_struct *vma;
1536     struct rmap_item *rmap_item;
1537     int nid;
1538 
1539     if (list_empty(&ksm_mm_head.mm_list))
1540         return NULL;
1541 
1542     slot = ksm_scan.mm_slot;
1543     if (slot == &ksm_mm_head) {
1544         /*
1545          * A number of pages can hang around indefinitely on per-cpu
1546          * pagevecs, raised page count preventing write_protect_page
1547          * from merging them.  Though it doesn't really matter much,
1548          * it is puzzling to see some stuck in pages_volatile until
1549          * other activity jostles them out, and they also prevented
1550          * LTP's KSM test from succeeding deterministically; so drain
1551          * them here (here rather than on entry to ksm_do_scan(),
1552          * so we don't IPI too often when pages_to_scan is set low).
1553          */
1554         lru_add_drain_all();
1555 
1556         /*
1557          * Whereas stale stable_nodes on the stable_tree itself
1558          * get pruned in the regular course of stable_tree_search(),
1559          * those moved out to the migrate_nodes list can accumulate:
1560          * so prune them once before each full scan.
1561          */
1562         if (!ksm_merge_across_nodes) {
1563             struct stable_node *stable_node, *next;
1564             struct page *page;
1565 
1566             list_for_each_entry_safe(stable_node, next,
1567                          &migrate_nodes, list) {
1568                 page = get_ksm_page(stable_node, false);
1569                 if (page)
1570                     put_page(page);
1571                 cond_resched();
1572             }
1573         }
1574 
1575         for (nid = 0; nid < ksm_nr_node_ids; nid++)
1576             root_unstable_tree[nid] = RB_ROOT;
1577 
1578         spin_lock(&ksm_mmlist_lock);
1579         slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1580         ksm_scan.mm_slot = slot;
1581         spin_unlock(&ksm_mmlist_lock);
1582         /*
1583          * Although we tested list_empty() above, a racing __ksm_exit
1584          * of the last mm on the list may have removed it since then.
1585          */
1586         if (slot == &ksm_mm_head)
1587             return NULL;
1588 next_mm:
1589         ksm_scan.address = 0;
1590         ksm_scan.rmap_list = &slot->rmap_list;
1591     }
1592 
1593     mm = slot->mm;
1594     down_read(&mm->mmap_sem);
1595     if (ksm_test_exit(mm))
1596         vma = NULL;
1597     else
1598         vma = find_vma(mm, ksm_scan.address);
1599 
1600     for (; vma; vma = vma->vm_next) {
1601         if (!(vma->vm_flags & VM_MERGEABLE))
1602             continue;
1603         if (ksm_scan.address < vma->vm_start)
1604             ksm_scan.address = vma->vm_start;
1605         if (!vma->anon_vma)
1606             ksm_scan.address = vma->vm_end;
1607 
1608         while (ksm_scan.address < vma->vm_end) {
1609             if (ksm_test_exit(mm))
1610                 break;
1611             *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1612             if (IS_ERR_OR_NULL(*page)) {
1613                 ksm_scan.address += PAGE_SIZE;
1614                 cond_resched();
1615                 continue;
1616             }
1617             if (PageAnon(*page)) {
1618                 flush_anon_page(vma, *page, ksm_scan.address);
1619                 flush_dcache_page(*page);
1620                 rmap_item = get_next_rmap_item(slot,
1621                     ksm_scan.rmap_list, ksm_scan.address);
1622                 if (rmap_item) {
1623                     ksm_scan.rmap_list =
1624                             &rmap_item->rmap_list;
1625                     ksm_scan.address += PAGE_SIZE;
1626                 } else
1627                     put_page(*page);
1628                 up_read(&mm->mmap_sem);
1629                 return rmap_item;
1630             }
1631             put_page(*page);
1632             ksm_scan.address += PAGE_SIZE;
1633             cond_resched();
1634         }
1635     }
1636 
1637     if (ksm_test_exit(mm)) {
1638         ksm_scan.address = 0;
1639         ksm_scan.rmap_list = &slot->rmap_list;
1640     }
1641     /*
1642      * Nuke all the rmap_items that are above this current rmap:
1643      * because there were no VM_MERGEABLE vmas with such addresses.
1644      */
1645     remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1646 
1647     spin_lock(&ksm_mmlist_lock);
1648     ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1649                         struct mm_slot, mm_list);
1650     if (ksm_scan.address == 0) {
1651         /*
1652          * We've completed a full scan of all vmas, holding mmap_sem
1653          * throughout, and found no VM_MERGEABLE: so do the same as
1654          * __ksm_exit does to remove this mm from all our lists now.
1655          * This applies either when cleaning up after __ksm_exit
1656          * (but beware: we can reach here even before __ksm_exit),
1657          * or when all VM_MERGEABLE areas have been unmapped (and
1658          * mmap_sem then protects against race with MADV_MERGEABLE).
1659          */
1660         hash_del(&slot->link);
1661         list_del(&slot->mm_list);
1662         spin_unlock(&ksm_mmlist_lock);
1663 
1664         free_mm_slot(slot);
1665         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1666         up_read(&mm->mmap_sem);
1667         mmdrop(mm);
1668     } else {
1669         up_read(&mm->mmap_sem);
1670         /*
1671          * up_read(&mm->mmap_sem) first because after
1672          * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
1673          * already have been freed under us by __ksm_exit()
1674          * because the "mm_slot" is still hashed and
1675          * ksm_scan.mm_slot doesn't point to it anymore.
1676          */
1677         spin_unlock(&ksm_mmlist_lock);
1678     }
1679 
1680     /* Repeat until we've completed scanning the whole list */
1681     slot = ksm_scan.mm_slot;
1682     if (slot != &ksm_mm_head)
1683         goto next_mm;
1684 
1685     ksm_scan.seqnr++;
1686     return NULL;
1687 }
1688 
1689 /**
1690  * ksm_do_scan  - the ksm scanner main worker function.
1691  * @scan_npages - number of pages we want to scan before we return.
1692  */
1693 static void ksm_do_scan(unsigned int scan_npages)
1694 {
1695     struct rmap_item *rmap_item;
1696     struct page *uninitialized_var(page);
1697 
1698     while (scan_npages-- && likely(!freezing(current))) {
1699         cond_resched();
1700         rmap_item = scan_get_next_rmap_item(&page);
1701         if (!rmap_item)
1702             return;
1703         cmp_and_merge_page(page, rmap_item);
1704         put_page(page);
1705     }
1706 }
1707 
1708 static int ksmd_should_run(void)
1709 {
1710     return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1711 }
1712 
1713 static int ksm_scan_thread(void *nothing)
1714 {
1715     set_freezable();
1716     set_user_nice(current, 5);
1717 
1718     while (!kthread_should_stop()) {
1719         mutex_lock(&ksm_thread_mutex);
1720         wait_while_offlining();
1721         if (ksmd_should_run())
1722             ksm_do_scan(ksm_thread_pages_to_scan);
1723         mutex_unlock(&ksm_thread_mutex);
1724 
1725         try_to_freeze();
1726 
1727         if (ksmd_should_run()) {
1728             schedule_timeout_interruptible(
1729                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1730         } else {
1731             wait_event_freezable(ksm_thread_wait,
1732                 ksmd_should_run() || kthread_should_stop());
1733         }
1734     }
1735     return 0;
1736 }
1737 
1738 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1739         unsigned long end, int advice, unsigned long *vm_flags)
1740 {
1741     struct mm_struct *mm = vma->vm_mm;
1742     int err;
1743 
1744     switch (advice) {
1745     case MADV_MERGEABLE:
1746         /*
1747          * Be somewhat over-protective for now!
1748          */
1749         if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1750                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1751                  VM_HUGETLB | VM_MIXEDMAP))
1752             return 0;       /* just ignore the advice */
1753 
1754 #ifdef VM_SAO
1755         if (*vm_flags & VM_SAO)
1756             return 0;
1757 #endif
1758 
1759         if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1760             err = __ksm_enter(mm);
1761             if (err)
1762                 return err;
1763         }
1764 
1765         *vm_flags |= VM_MERGEABLE;
1766         break;
1767 
1768     case MADV_UNMERGEABLE:
1769         if (!(*vm_flags & VM_MERGEABLE))
1770             return 0;       /* just ignore the advice */
1771 
1772         if (vma->anon_vma) {
1773             err = unmerge_ksm_pages(vma, start, end);
1774             if (err)
1775                 return err;
1776         }
1777 
1778         *vm_flags &= ~VM_MERGEABLE;
1779         break;
1780     }
1781 
1782     return 0;
1783 }
1784 
1785 int __ksm_enter(struct mm_struct *mm)
1786 {
1787     struct mm_slot *mm_slot;
1788     int needs_wakeup;
1789 
1790     mm_slot = alloc_mm_slot();
1791     if (!mm_slot)
1792         return -ENOMEM;
1793 
1794     /* Check ksm_run too?  Would need tighter locking */
1795     needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1796 
1797     spin_lock(&ksm_mmlist_lock);
1798     insert_to_mm_slots_hash(mm, mm_slot);
1799     /*
1800      * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1801      * insert just behind the scanning cursor, to let the area settle
1802      * down a little; when fork is followed by immediate exec, we don't
1803      * want ksmd to waste time setting up and tearing down an rmap_list.
1804      *
1805      * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1806      * scanning cursor, otherwise KSM pages in newly forked mms will be
1807      * missed: then we might as well insert at the end of the list.
1808      */
1809     if (ksm_run & KSM_RUN_UNMERGE)
1810         list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1811     else
1812         list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1813     spin_unlock(&ksm_mmlist_lock);
1814 
1815     set_bit(MMF_VM_MERGEABLE, &mm->flags);
1816     atomic_inc(&mm->mm_count);
1817 
1818     if (needs_wakeup)
1819         wake_up_interruptible(&ksm_thread_wait);
1820 
1821     return 0;
1822 }
1823 
1824 void __ksm_exit(struct mm_struct *mm)
1825 {
1826     struct mm_slot *mm_slot;
1827     int easy_to_free = 0;
1828 
1829     /*
1830      * This process is exiting: if it's straightforward (as is the
1831      * case when ksmd was never running), free mm_slot immediately.
1832      * But if it's at the cursor or has rmap_items linked to it, use
1833      * mmap_sem to synchronize with any break_cows before pagetables
1834      * are freed, and leave the mm_slot on the list for ksmd to free.
1835      * Beware: ksm may already have noticed it exiting and freed the slot.
1836      */
1837 
1838     spin_lock(&ksm_mmlist_lock);
1839     mm_slot = get_mm_slot(mm);
1840     if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1841         if (!mm_slot->rmap_list) {
1842             hash_del(&mm_slot->link);
1843             list_del(&mm_slot->mm_list);
1844             easy_to_free = 1;
1845         } else {
1846             list_move(&mm_slot->mm_list,
1847                   &ksm_scan.mm_slot->mm_list);
1848         }
1849     }
1850     spin_unlock(&ksm_mmlist_lock);
1851 
1852     if (easy_to_free) {
1853         free_mm_slot(mm_slot);
1854         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1855         mmdrop(mm);
1856     } else if (mm_slot) {
1857         down_write(&mm->mmap_sem);
1858         up_write(&mm->mmap_sem);
1859     }
1860 }
1861 
1862 struct page *ksm_might_need_to_copy(struct page *page,
1863             struct vm_area_struct *vma, unsigned long address)
1864 {
1865     struct anon_vma *anon_vma = page_anon_vma(page);
1866     struct page *new_page;
1867 
1868     if (PageKsm(page)) {
1869         if (page_stable_node(page) &&
1870             !(ksm_run & KSM_RUN_UNMERGE))
1871             return page;    /* no need to copy it */
1872     } else if (!anon_vma) {
1873         return page;        /* no need to copy it */
1874     } else if (anon_vma->root == vma->anon_vma->root &&
1875          page->index == linear_page_index(vma, address)) {
1876         return page;        /* still no need to copy it */
1877     }
1878     if (!PageUptodate(page))
1879         return page;        /* let do_swap_page report the error */
1880 
1881     new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1882     if (new_page) {
1883         copy_user_highpage(new_page, page, address, vma);
1884 
1885         SetPageDirty(new_page);
1886         __SetPageUptodate(new_page);
1887         __SetPageLocked(new_page);
1888     }
1889 
1890     return new_page;
1891 }
1892 
1893 int rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
1894 {
1895     struct stable_node *stable_node;
1896     struct rmap_item *rmap_item;
1897     int ret = SWAP_AGAIN;
1898     int search_new_forks = 0;
1899 
1900     VM_BUG_ON_PAGE(!PageKsm(page), page);
1901 
1902     /*
1903      * Rely on the page lock to protect against concurrent modifications
1904      * to that page's node of the stable tree.
1905      */
1906     VM_BUG_ON_PAGE(!PageLocked(page), page);
1907 
1908     stable_node = page_stable_node(page);
1909     if (!stable_node)
1910         return ret;
1911 again:
1912     hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
1913         struct anon_vma *anon_vma = rmap_item->anon_vma;
1914         struct anon_vma_chain *vmac;
1915         struct vm_area_struct *vma;
1916 
1917         cond_resched();
1918         anon_vma_lock_read(anon_vma);
1919         anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1920                            0, ULONG_MAX) {
1921             cond_resched();
1922             vma = vmac->vma;
1923             if (rmap_item->address < vma->vm_start ||
1924                 rmap_item->address >= vma->vm_end)
1925                 continue;
1926             /*
1927              * Initially we examine only the vma which covers this
1928              * rmap_item; but later, if there is still work to do,
1929              * we examine covering vmas in other mms: in case they
1930              * were forked from the original since ksmd passed.
1931              */
1932             if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1933                 continue;
1934 
1935             if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1936                 continue;
1937 
1938             ret = rwc->rmap_one(page, vma,
1939                     rmap_item->address, rwc->arg);
1940             if (ret != SWAP_AGAIN) {
1941                 anon_vma_unlock_read(anon_vma);
1942                 goto out;
1943             }
1944             if (rwc->done && rwc->done(page)) {
1945                 anon_vma_unlock_read(anon_vma);
1946                 goto out;
1947             }
1948         }
1949         anon_vma_unlock_read(anon_vma);
1950     }
1951     if (!search_new_forks++)
1952         goto again;
1953 out:
1954     return ret;
1955 }
1956 
1957 #ifdef CONFIG_MIGRATION
1958 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1959 {
1960     struct stable_node *stable_node;
1961 
1962     VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
1963     VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1964     VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
1965 
1966     stable_node = page_stable_node(newpage);
1967     if (stable_node) {
1968         VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
1969         stable_node->kpfn = page_to_pfn(newpage);
1970         /*
1971          * newpage->mapping was set in advance; now we need smp_wmb()
1972          * to make sure that the new stable_node->kpfn is visible
1973          * to get_ksm_page() before it can see that oldpage->mapping
1974          * has gone stale (or that PageSwapCache has been cleared).
1975          */
1976         smp_wmb();
1977         set_page_stable_node(oldpage, NULL);
1978     }
1979 }
1980 #endif /* CONFIG_MIGRATION */
1981 
1982 #ifdef CONFIG_MEMORY_HOTREMOVE
1983 static void wait_while_offlining(void)
1984 {
1985     while (ksm_run & KSM_RUN_OFFLINE) {
1986         mutex_unlock(&ksm_thread_mutex);
1987         wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
1988                 TASK_UNINTERRUPTIBLE);
1989         mutex_lock(&ksm_thread_mutex);
1990     }
1991 }
1992 
1993 static void ksm_check_stable_tree(unsigned long start_pfn,
1994                   unsigned long end_pfn)
1995 {
1996     struct stable_node *stable_node, *next;
1997     struct rb_node *node;
1998     int nid;
1999 
2000     for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2001         node = rb_first(root_stable_tree + nid);
2002         while (node) {
2003             stable_node = rb_entry(node, struct stable_node, node);
2004             if (stable_node->kpfn >= start_pfn &&
2005                 stable_node->kpfn < end_pfn) {
2006                 /*
2007                  * Don't get_ksm_page, page has already gone:
2008                  * which is why we keep kpfn instead of page*
2009                  */
2010                 remove_node_from_stable_tree(stable_node);
2011                 node = rb_first(root_stable_tree + nid);
2012             } else
2013                 node = rb_next(node);
2014             cond_resched();
2015         }
2016     }
2017     list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2018         if (stable_node->kpfn >= start_pfn &&
2019             stable_node->kpfn < end_pfn)
2020             remove_node_from_stable_tree(stable_node);
2021         cond_resched();
2022     }
2023 }
2024 
2025 static int ksm_memory_callback(struct notifier_block *self,
2026                    unsigned long action, void *arg)
2027 {
2028     struct memory_notify *mn = arg;
2029 
2030     switch (action) {
2031     case MEM_GOING_OFFLINE:
2032         /*
2033          * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2034          * and remove_all_stable_nodes() while memory is going offline:
2035          * it is unsafe for them to touch the stable tree at this time.
2036          * But unmerge_ksm_pages(), rmap lookups and other entry points
2037          * which do not need the ksm_thread_mutex are all safe.
2038          */
2039         mutex_lock(&ksm_thread_mutex);
2040         ksm_run |= KSM_RUN_OFFLINE;
2041         mutex_unlock(&ksm_thread_mutex);
2042         break;
2043 
2044     case MEM_OFFLINE:
2045         /*
2046          * Most of the work is done by page migration; but there might
2047          * be a few stable_nodes left over, still pointing to struct
2048          * pages which have been offlined: prune those from the tree,
2049          * otherwise get_ksm_page() might later try to access a
2050          * non-existent struct page.
2051          */
2052         ksm_check_stable_tree(mn->start_pfn,
2053                       mn->start_pfn + mn->nr_pages);
2054         /* fallthrough */
2055 
2056     case MEM_CANCEL_OFFLINE:
2057         mutex_lock(&ksm_thread_mutex);
2058         ksm_run &= ~KSM_RUN_OFFLINE;
2059         mutex_unlock(&ksm_thread_mutex);
2060 
2061         smp_mb();   /* wake_up_bit advises this */
2062         wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2063         break;
2064     }
2065     return NOTIFY_OK;
2066 }
2067 #else
2068 static void wait_while_offlining(void)
2069 {
2070 }
2071 #endif /* CONFIG_MEMORY_HOTREMOVE */
2072 
2073 #ifdef CONFIG_SYSFS
2074 /*
2075  * This all compiles without CONFIG_SYSFS, but is a waste of space.
2076  */
2077 
2078 #define KSM_ATTR_RO(_name) \
2079     static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2080 #define KSM_ATTR(_name) \
2081     static struct kobj_attribute _name##_attr = \
2082         __ATTR(_name, 0644, _name##_show, _name##_store)
2083 
2084 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2085                     struct kobj_attribute *attr, char *buf)
2086 {
2087     return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2088 }
2089 
2090 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2091                      struct kobj_attribute *attr,
2092                      const char *buf, size_t count)
2093 {
2094     unsigned long msecs;
2095     int err;
2096 
2097     err = kstrtoul(buf, 10, &msecs);
2098     if (err || msecs > UINT_MAX)
2099         return -EINVAL;
2100 
2101     ksm_thread_sleep_millisecs = msecs;
2102 
2103     return count;
2104 }
2105 KSM_ATTR(sleep_millisecs);
2106 
2107 static ssize_t pages_to_scan_show(struct kobject *kobj,
2108                   struct kobj_attribute *attr, char *buf)
2109 {
2110     return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2111 }
2112 
2113 static ssize_t pages_to_scan_store(struct kobject *kobj,
2114                    struct kobj_attribute *attr,
2115                    const char *buf, size_t count)
2116 {
2117     int err;
2118     unsigned long nr_pages;
2119 
2120     err = kstrtoul(buf, 10, &nr_pages);
2121     if (err || nr_pages > UINT_MAX)
2122         return -EINVAL;
2123 
2124     ksm_thread_pages_to_scan = nr_pages;
2125 
2126     return count;
2127 }
2128 KSM_ATTR(pages_to_scan);
2129 
2130 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2131             char *buf)
2132 {
2133     return sprintf(buf, "%lu\n", ksm_run);
2134 }
2135 
2136 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2137              const char *buf, size_t count)
2138 {
2139     int err;
2140     unsigned long flags;
2141 
2142     err = kstrtoul(buf, 10, &flags);
2143     if (err || flags > UINT_MAX)
2144         return -EINVAL;
2145     if (flags > KSM_RUN_UNMERGE)
2146         return -EINVAL;
2147 
2148     /*
2149      * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2150      * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2151      * breaking COW to free the pages_shared (but leaves mm_slots
2152      * on the list for when ksmd may be set running again).
2153      */
2154 
2155     mutex_lock(&ksm_thread_mutex);
2156     wait_while_offlining();
2157     if (ksm_run != flags) {
2158         ksm_run = flags;
2159         if (flags & KSM_RUN_UNMERGE) {
2160             set_current_oom_origin();
2161             err = unmerge_and_remove_all_rmap_items();
2162             clear_current_oom_origin();
2163             if (err) {
2164                 ksm_run = KSM_RUN_STOP;
2165                 count = err;
2166             }
2167         }
2168     }
2169     mutex_unlock(&ksm_thread_mutex);
2170 
2171     if (flags & KSM_RUN_MERGE)
2172         wake_up_interruptible(&ksm_thread_wait);
2173 
2174     return count;
2175 }
2176 KSM_ATTR(run);
2177 
2178 #ifdef CONFIG_NUMA
2179 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2180                 struct kobj_attribute *attr, char *buf)
2181 {
2182     return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2183 }
2184 
2185 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2186                    struct kobj_attribute *attr,
2187                    const char *buf, size_t count)
2188 {
2189     int err;
2190     unsigned long knob;
2191 
2192     err = kstrtoul(buf, 10, &knob);
2193     if (err)
2194         return err;
2195     if (knob > 1)
2196         return -EINVAL;
2197 
2198     mutex_lock(&ksm_thread_mutex);
2199     wait_while_offlining();
2200     if (ksm_merge_across_nodes != knob) {
2201         if (ksm_pages_shared || remove_all_stable_nodes())
2202             err = -EBUSY;
2203         else if (root_stable_tree == one_stable_tree) {
2204             struct rb_root *buf;
2205             /*
2206              * This is the first time that we switch away from the
2207              * default of merging across nodes: must now allocate
2208              * a buffer to hold as many roots as may be needed.
2209              * Allocate stable and unstable together:
2210              * MAXSMP NODES_SHIFT 10 will use 16kB.
2211              */
2212             buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2213                       GFP_KERNEL);
2214             /* Let us assume that RB_ROOT is NULL is zero */
2215             if (!buf)
2216                 err = -ENOMEM;
2217             else {
2218                 root_stable_tree = buf;
2219                 root_unstable_tree = buf + nr_node_ids;
2220                 /* Stable tree is empty but not the unstable */
2221                 root_unstable_tree[0] = one_unstable_tree[0];
2222             }
2223         }
2224         if (!err) {
2225             ksm_merge_across_nodes = knob;
2226             ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2227         }
2228     }
2229     mutex_unlock(&ksm_thread_mutex);
2230 
2231     return err ? err : count;
2232 }
2233 KSM_ATTR(merge_across_nodes);
2234 #endif
2235 
2236 static ssize_t pages_shared_show(struct kobject *kobj,
2237                  struct kobj_attribute *attr, char *buf)
2238 {
2239     return sprintf(buf, "%lu\n", ksm_pages_shared);
2240 }
2241 KSM_ATTR_RO(pages_shared);
2242 
2243 static ssize_t pages_sharing_show(struct kobject *kobj,
2244                   struct kobj_attribute *attr, char *buf)
2245 {
2246     return sprintf(buf, "%lu\n", ksm_pages_sharing);
2247 }
2248 KSM_ATTR_RO(pages_sharing);
2249 
2250 static ssize_t pages_unshared_show(struct kobject *kobj,
2251                    struct kobj_attribute *attr, char *buf)
2252 {
2253     return sprintf(buf, "%lu\n", ksm_pages_unshared);
2254 }
2255 KSM_ATTR_RO(pages_unshared);
2256 
2257 static ssize_t pages_volatile_show(struct kobject *kobj,
2258                    struct kobj_attribute *attr, char *buf)
2259 {
2260     long ksm_pages_volatile;
2261 
2262     ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2263                 - ksm_pages_sharing - ksm_pages_unshared;
2264     /*
2265      * It was not worth any locking to calculate that statistic,
2266      * but it might therefore sometimes be negative: conceal that.
2267      */
2268     if (ksm_pages_volatile < 0)
2269         ksm_pages_volatile = 0;
2270     return sprintf(buf, "%ld\n", ksm_pages_volatile);
2271 }
2272 KSM_ATTR_RO(pages_volatile);
2273 
2274 static ssize_t full_scans_show(struct kobject *kobj,
2275                    struct kobj_attribute *attr, char *buf)
2276 {
2277     return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2278 }
2279 KSM_ATTR_RO(full_scans);
2280 
2281 static struct attribute *ksm_attrs[] = {
2282     &sleep_millisecs_attr.attr,
2283     &pages_to_scan_attr.attr,
2284     &run_attr.attr,
2285     &pages_shared_attr.attr,
2286     &pages_sharing_attr.attr,
2287     &pages_unshared_attr.attr,
2288     &pages_volatile_attr.attr,
2289     &full_scans_attr.attr,
2290 #ifdef CONFIG_NUMA
2291     &merge_across_nodes_attr.attr,
2292 #endif
2293     NULL,
2294 };
2295 
2296 static struct attribute_group ksm_attr_group = {
2297     .attrs = ksm_attrs,
2298     .name = "ksm",
2299 };
2300 #endif /* CONFIG_SYSFS */
2301 
2302 static int __init ksm_init(void)
2303 {
2304     struct task_struct *ksm_thread;
2305     int err;
2306 
2307     err = ksm_slab_init();
2308     if (err)
2309         goto out;
2310 
2311     ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2312     if (IS_ERR(ksm_thread)) {
2313         pr_err("ksm: creating kthread failed\n");
2314         err = PTR_ERR(ksm_thread);
2315         goto out_free;
2316     }
2317 
2318 #ifdef CONFIG_SYSFS
2319     err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2320     if (err) {
2321         pr_err("ksm: register sysfs failed\n");
2322         kthread_stop(ksm_thread);
2323         goto out_free;
2324     }
2325 #else
2326     ksm_run = KSM_RUN_MERGE;    /* no way for user to start it */
2327 
2328 #endif /* CONFIG_SYSFS */
2329 
2330 #ifdef CONFIG_MEMORY_HOTREMOVE
2331     /* There is no significance to this priority 100 */
2332     hotplug_memory_notifier(ksm_memory_callback, 100);
2333 #endif
2334     return 0;
2335 
2336 out_free:
2337     ksm_slab_free();
2338 out:
2339     return err;
2340 }
2341 subsys_initcall(ksm_init);