Root/mm/ksm.c

1/*
2 * Memory merging support.
3 *
4 * This code enables dynamic sharing of identical pages found in different
5 * memory areas, even if they are not shared by fork()
6 *
7 * Copyright (C) 2008-2009 Red Hat, Inc.
8 * Authors:
9 * Izik Eidus
10 * Andrea Arcangeli
11 * Chris Wright
12 * Hugh Dickins
13 *
14 * This work is licensed under the terms of the GNU GPL, version 2.
15 */
16
17#include <linux/errno.h>
18#include <linux/mm.h>
19#include <linux/fs.h>
20#include <linux/mman.h>
21#include <linux/sched.h>
22#include <linux/rwsem.h>
23#include <linux/pagemap.h>
24#include <linux/rmap.h>
25#include <linux/spinlock.h>
26#include <linux/jhash.h>
27#include <linux/delay.h>
28#include <linux/kthread.h>
29#include <linux/wait.h>
30#include <linux/slab.h>
31#include <linux/rbtree.h>
32#include <linux/memory.h>
33#include <linux/mmu_notifier.h>
34#include <linux/swap.h>
35#include <linux/ksm.h>
36#include <linux/hash.h>
37#include <linux/freezer.h>
38#include <linux/oom.h>
39
40#include <asm/tlbflush.h>
41#include "internal.h"
42
43/*
44 * A few notes about the KSM scanning process,
45 * to make it easier to understand the data structures below:
46 *
47 * In order to reduce excessive scanning, KSM sorts the memory pages by their
48 * contents into a data structure that holds pointers to the pages' locations.
49 *
50 * Since the contents of the pages may change at any moment, KSM cannot just
51 * insert the pages into a normal sorted tree and expect it to find anything.
52 * Therefore KSM uses two data structures - the stable and the unstable tree.
53 *
54 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
55 * by their contents. Because each such page is write-protected, searching on
56 * this tree is fully assured to be working (except when pages are unmapped),
57 * and therefore this tree is called the stable tree.
58 *
59 * In addition to the stable tree, KSM uses a second data structure called the
60 * unstable tree: this tree holds pointers to pages which have been found to
61 * be "unchanged for a period of time". The unstable tree sorts these pages
62 * by their contents, but since they are not write-protected, KSM cannot rely
63 * upon the unstable tree to work correctly - the unstable tree is liable to
64 * be corrupted as its contents are modified, and so it is called unstable.
65 *
66 * KSM solves this problem by several techniques:
67 *
68 * 1) The unstable tree is flushed every time KSM completes scanning all
69 * memory areas, and then the tree is rebuilt again from the beginning.
70 * 2) KSM will only insert into the unstable tree, pages whose hash value
71 * has not changed since the previous scan of all memory areas.
72 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
73 * colors of the nodes and not on their contents, assuring that even when
74 * the tree gets "corrupted" it won't get out of balance, so scanning time
75 * remains the same (also, searching and inserting nodes in an rbtree uses
76 * the same algorithm, so we have no overhead when we flush and rebuild).
77 * 4) KSM never flushes the stable tree, which means that even if it were to
78 * take 10 attempts to find a page in the unstable tree, once it is found,
79 * it is secured in the stable tree. (When we scan a new page, we first
80 * compare it against the stable tree, and then against the unstable tree.)
81 */
82
83/**
84 * struct mm_slot - ksm information per mm that is being scanned
85 * @link: link to the mm_slots hash list
86 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
87 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
88 * @mm: the mm that this information is valid for
89 */
90struct mm_slot {
91    struct hlist_node link;
92    struct list_head mm_list;
93    struct rmap_item *rmap_list;
94    struct mm_struct *mm;
95};
96
97/**
98 * struct ksm_scan - cursor for scanning
99 * @mm_slot: the current mm_slot we are scanning
100 * @address: the next address inside that to be scanned
101 * @rmap_list: link to the next rmap to be scanned in the rmap_list
102 * @seqnr: count of completed full scans (needed when removing unstable node)
103 *
104 * There is only the one ksm_scan instance of this cursor structure.
105 */
106struct ksm_scan {
107    struct mm_slot *mm_slot;
108    unsigned long address;
109    struct rmap_item **rmap_list;
110    unsigned long seqnr;
111};
112
113/**
114 * struct stable_node - node of the stable rbtree
115 * @node: rb node of this ksm page in the stable tree
116 * @hlist: hlist head of rmap_items using this ksm page
117 * @kpfn: page frame number of this ksm page
118 */
119struct stable_node {
120    struct rb_node node;
121    struct hlist_head hlist;
122    unsigned long kpfn;
123};
124
125/**
126 * struct rmap_item - reverse mapping item for virtual addresses
127 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
128 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
129 * @mm: the memory structure this rmap_item is pointing into
130 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
131 * @oldchecksum: previous checksum of the page at that virtual address
132 * @node: rb node of this rmap_item in the unstable tree
133 * @head: pointer to stable_node heading this list in the stable tree
134 * @hlist: link into hlist of rmap_items hanging off that stable_node
135 */
136struct rmap_item {
137    struct rmap_item *rmap_list;
138    struct anon_vma *anon_vma; /* when stable */
139    struct mm_struct *mm;
140    unsigned long address; /* + low bits used for flags below */
141    unsigned int oldchecksum; /* when unstable */
142    union {
143        struct rb_node node; /* when node of unstable tree */
144        struct { /* when listed from stable tree */
145            struct stable_node *head;
146            struct hlist_node hlist;
147        };
148    };
149};
150
151#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
152#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
153#define STABLE_FLAG 0x200 /* is listed from the stable tree */
154
155/* The stable and unstable tree heads */
156static struct rb_root root_stable_tree = RB_ROOT;
157static struct rb_root root_unstable_tree = RB_ROOT;
158
159#define MM_SLOTS_HASH_SHIFT 10
160#define MM_SLOTS_HASH_HEADS (1 << MM_SLOTS_HASH_SHIFT)
161static struct hlist_head mm_slots_hash[MM_SLOTS_HASH_HEADS];
162
163static struct mm_slot ksm_mm_head = {
164    .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
165};
166static struct ksm_scan ksm_scan = {
167    .mm_slot = &ksm_mm_head,
168};
169
170static struct kmem_cache *rmap_item_cache;
171static struct kmem_cache *stable_node_cache;
172static struct kmem_cache *mm_slot_cache;
173
174/* The number of nodes in the stable tree */
175static unsigned long ksm_pages_shared;
176
177/* The number of page slots additionally sharing those nodes */
178static unsigned long ksm_pages_sharing;
179
180/* The number of nodes in the unstable tree */
181static unsigned long ksm_pages_unshared;
182
183/* The number of rmap_items in use: to calculate pages_volatile */
184static unsigned long ksm_rmap_items;
185
186/* Number of pages ksmd should scan in one batch */
187static unsigned int ksm_thread_pages_to_scan = 100;
188
189/* Milliseconds ksmd should sleep between batches */
190static unsigned int ksm_thread_sleep_millisecs = 20;
191
192#define KSM_RUN_STOP 0
193#define KSM_RUN_MERGE 1
194#define KSM_RUN_UNMERGE 2
195static unsigned int ksm_run = KSM_RUN_STOP;
196
197static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
198static DEFINE_MUTEX(ksm_thread_mutex);
199static DEFINE_SPINLOCK(ksm_mmlist_lock);
200
201#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
202        sizeof(struct __struct), __alignof__(struct __struct),\
203        (__flags), NULL)
204
205static int __init ksm_slab_init(void)
206{
207    rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
208    if (!rmap_item_cache)
209        goto out;
210
211    stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
212    if (!stable_node_cache)
213        goto out_free1;
214
215    mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
216    if (!mm_slot_cache)
217        goto out_free2;
218
219    return 0;
220
221out_free2:
222    kmem_cache_destroy(stable_node_cache);
223out_free1:
224    kmem_cache_destroy(rmap_item_cache);
225out:
226    return -ENOMEM;
227}
228
229static void __init ksm_slab_free(void)
230{
231    kmem_cache_destroy(mm_slot_cache);
232    kmem_cache_destroy(stable_node_cache);
233    kmem_cache_destroy(rmap_item_cache);
234    mm_slot_cache = NULL;
235}
236
237static inline struct rmap_item *alloc_rmap_item(void)
238{
239    struct rmap_item *rmap_item;
240
241    rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
242    if (rmap_item)
243        ksm_rmap_items++;
244    return rmap_item;
245}
246
247static inline void free_rmap_item(struct rmap_item *rmap_item)
248{
249    ksm_rmap_items--;
250    rmap_item->mm = NULL; /* debug safety */
251    kmem_cache_free(rmap_item_cache, rmap_item);
252}
253
254static inline struct stable_node *alloc_stable_node(void)
255{
256    return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
257}
258
259static inline void free_stable_node(struct stable_node *stable_node)
260{
261    kmem_cache_free(stable_node_cache, stable_node);
262}
263
264static inline struct mm_slot *alloc_mm_slot(void)
265{
266    if (!mm_slot_cache) /* initialization failed */
267        return NULL;
268    return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
269}
270
271static inline void free_mm_slot(struct mm_slot *mm_slot)
272{
273    kmem_cache_free(mm_slot_cache, mm_slot);
274}
275
276static struct mm_slot *get_mm_slot(struct mm_struct *mm)
277{
278    struct mm_slot *mm_slot;
279    struct hlist_head *bucket;
280    struct hlist_node *node;
281
282    bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
283    hlist_for_each_entry(mm_slot, node, bucket, link) {
284        if (mm == mm_slot->mm)
285            return mm_slot;
286    }
287    return NULL;
288}
289
290static void insert_to_mm_slots_hash(struct mm_struct *mm,
291                    struct mm_slot *mm_slot)
292{
293    struct hlist_head *bucket;
294
295    bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
296    mm_slot->mm = mm;
297    hlist_add_head(&mm_slot->link, bucket);
298}
299
300static inline int in_stable_tree(struct rmap_item *rmap_item)
301{
302    return rmap_item->address & STABLE_FLAG;
303}
304
305/*
306 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
307 * page tables after it has passed through ksm_exit() - which, if necessary,
308 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
309 * a special flag: they can just back out as soon as mm_users goes to zero.
310 * ksm_test_exit() is used throughout to make this test for exit: in some
311 * places for correctness, in some places just to avoid unnecessary work.
312 */
313static inline bool ksm_test_exit(struct mm_struct *mm)
314{
315    return atomic_read(&mm->mm_users) == 0;
316}
317
318/*
319 * We use break_ksm to break COW on a ksm page: it's a stripped down
320 *
321 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
322 * put_page(page);
323 *
324 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
325 * in case the application has unmapped and remapped mm,addr meanwhile.
326 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
327 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
328 */
329static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
330{
331    struct page *page;
332    int ret = 0;
333
334    do {
335        cond_resched();
336        page = follow_page(vma, addr, FOLL_GET);
337        if (IS_ERR_OR_NULL(page))
338            break;
339        if (PageKsm(page))
340            ret = handle_mm_fault(vma->vm_mm, vma, addr,
341                            FAULT_FLAG_WRITE);
342        else
343            ret = VM_FAULT_WRITE;
344        put_page(page);
345    } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
346    /*
347     * We must loop because handle_mm_fault() may back out if there's
348     * any difficulty e.g. if pte accessed bit gets updated concurrently.
349     *
350     * VM_FAULT_WRITE is what we have been hoping for: it indicates that
351     * COW has been broken, even if the vma does not permit VM_WRITE;
352     * but note that a concurrent fault might break PageKsm for us.
353     *
354     * VM_FAULT_SIGBUS could occur if we race with truncation of the
355     * backing file, which also invalidates anonymous pages: that's
356     * okay, that truncation will have unmapped the PageKsm for us.
357     *
358     * VM_FAULT_OOM: at the time of writing (late July 2009), setting
359     * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
360     * current task has TIF_MEMDIE set, and will be OOM killed on return
361     * to user; and ksmd, having no mm, would never be chosen for that.
362     *
363     * But if the mm is in a limited mem_cgroup, then the fault may fail
364     * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
365     * even ksmd can fail in this way - though it's usually breaking ksm
366     * just to undo a merge it made a moment before, so unlikely to oom.
367     *
368     * That's a pity: we might therefore have more kernel pages allocated
369     * than we're counting as nodes in the stable tree; but ksm_do_scan
370     * will retry to break_cow on each pass, so should recover the page
371     * in due course. The important thing is to not let VM_MERGEABLE
372     * be cleared while any such pages might remain in the area.
373     */
374    return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
375}
376
377static void break_cow(struct rmap_item *rmap_item)
378{
379    struct mm_struct *mm = rmap_item->mm;
380    unsigned long addr = rmap_item->address;
381    struct vm_area_struct *vma;
382
383    /*
384     * It is not an accident that whenever we want to break COW
385     * to undo, we also need to drop a reference to the anon_vma.
386     */
387    put_anon_vma(rmap_item->anon_vma);
388
389    down_read(&mm->mmap_sem);
390    if (ksm_test_exit(mm))
391        goto out;
392    vma = find_vma(mm, addr);
393    if (!vma || vma->vm_start > addr)
394        goto out;
395    if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
396        goto out;
397    break_ksm(vma, addr);
398out:
399    up_read(&mm->mmap_sem);
400}
401
402static struct page *page_trans_compound_anon(struct page *page)
403{
404    if (PageTransCompound(page)) {
405        struct page *head = compound_trans_head(page);
406        /*
407         * head may actually be splitted and freed from under
408         * us but it's ok here.
409         */
410        if (PageAnon(head))
411            return head;
412    }
413    return NULL;
414}
415
416static struct page *get_mergeable_page(struct rmap_item *rmap_item)
417{
418    struct mm_struct *mm = rmap_item->mm;
419    unsigned long addr = rmap_item->address;
420    struct vm_area_struct *vma;
421    struct page *page;
422
423    down_read(&mm->mmap_sem);
424    if (ksm_test_exit(mm))
425        goto out;
426    vma = find_vma(mm, addr);
427    if (!vma || vma->vm_start > addr)
428        goto out;
429    if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
430        goto out;
431
432    page = follow_page(vma, addr, FOLL_GET);
433    if (IS_ERR_OR_NULL(page))
434        goto out;
435    if (PageAnon(page) || page_trans_compound_anon(page)) {
436        flush_anon_page(vma, page, addr);
437        flush_dcache_page(page);
438    } else {
439        put_page(page);
440out: page = NULL;
441    }
442    up_read(&mm->mmap_sem);
443    return page;
444}
445
446static void remove_node_from_stable_tree(struct stable_node *stable_node)
447{
448    struct rmap_item *rmap_item;
449    struct hlist_node *hlist;
450
451    hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
452        if (rmap_item->hlist.next)
453            ksm_pages_sharing--;
454        else
455            ksm_pages_shared--;
456        put_anon_vma(rmap_item->anon_vma);
457        rmap_item->address &= PAGE_MASK;
458        cond_resched();
459    }
460
461    rb_erase(&stable_node->node, &root_stable_tree);
462    free_stable_node(stable_node);
463}
464
465/*
466 * get_ksm_page: checks if the page indicated by the stable node
467 * is still its ksm page, despite having held no reference to it.
468 * In which case we can trust the content of the page, and it
469 * returns the gotten page; but if the page has now been zapped,
470 * remove the stale node from the stable tree and return NULL.
471 *
472 * You would expect the stable_node to hold a reference to the ksm page.
473 * But if it increments the page's count, swapping out has to wait for
474 * ksmd to come around again before it can free the page, which may take
475 * seconds or even minutes: much too unresponsive. So instead we use a
476 * "keyhole reference": access to the ksm page from the stable node peeps
477 * out through its keyhole to see if that page still holds the right key,
478 * pointing back to this stable node. This relies on freeing a PageAnon
479 * page to reset its page->mapping to NULL, and relies on no other use of
480 * a page to put something that might look like our key in page->mapping.
481 *
482 * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
483 * but this is different - made simpler by ksm_thread_mutex being held, but
484 * interesting for assuming that no other use of the struct page could ever
485 * put our expected_mapping into page->mapping (or a field of the union which
486 * coincides with page->mapping). The RCU calls are not for KSM at all, but
487 * to keep the page_count protocol described with page_cache_get_speculative.
488 *
489 * Note: it is possible that get_ksm_page() will return NULL one moment,
490 * then page the next, if the page is in between page_freeze_refs() and
491 * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
492 * is on its way to being freed; but it is an anomaly to bear in mind.
493 */
494static struct page *get_ksm_page(struct stable_node *stable_node)
495{
496    struct page *page;
497    void *expected_mapping;
498
499    page = pfn_to_page(stable_node->kpfn);
500    expected_mapping = (void *)stable_node +
501                (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
502    rcu_read_lock();
503    if (page->mapping != expected_mapping)
504        goto stale;
505    if (!get_page_unless_zero(page))
506        goto stale;
507    if (page->mapping != expected_mapping) {
508        put_page(page);
509        goto stale;
510    }
511    rcu_read_unlock();
512    return page;
513stale:
514    rcu_read_unlock();
515    remove_node_from_stable_tree(stable_node);
516    return NULL;
517}
518
519/*
520 * Removing rmap_item from stable or unstable tree.
521 * This function will clean the information from the stable/unstable tree.
522 */
523static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
524{
525    if (rmap_item->address & STABLE_FLAG) {
526        struct stable_node *stable_node;
527        struct page *page;
528
529        stable_node = rmap_item->head;
530        page = get_ksm_page(stable_node);
531        if (!page)
532            goto out;
533
534        lock_page(page);
535        hlist_del(&rmap_item->hlist);
536        unlock_page(page);
537        put_page(page);
538
539        if (stable_node->hlist.first)
540            ksm_pages_sharing--;
541        else
542            ksm_pages_shared--;
543
544        put_anon_vma(rmap_item->anon_vma);
545        rmap_item->address &= PAGE_MASK;
546
547    } else if (rmap_item->address & UNSTABLE_FLAG) {
548        unsigned char age;
549        /*
550         * Usually ksmd can and must skip the rb_erase, because
551         * root_unstable_tree was already reset to RB_ROOT.
552         * But be careful when an mm is exiting: do the rb_erase
553         * if this rmap_item was inserted by this scan, rather
554         * than left over from before.
555         */
556        age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
557        BUG_ON(age > 1);
558        if (!age)
559            rb_erase(&rmap_item->node, &root_unstable_tree);
560
561        ksm_pages_unshared--;
562        rmap_item->address &= PAGE_MASK;
563    }
564out:
565    cond_resched(); /* we're called from many long loops */
566}
567
568static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
569                       struct rmap_item **rmap_list)
570{
571    while (*rmap_list) {
572        struct rmap_item *rmap_item = *rmap_list;
573        *rmap_list = rmap_item->rmap_list;
574        remove_rmap_item_from_tree(rmap_item);
575        free_rmap_item(rmap_item);
576    }
577}
578
579/*
580 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
581 * than check every pte of a given vma, the locking doesn't quite work for
582 * that - an rmap_item is assigned to the stable tree after inserting ksm
583 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
584 * rmap_items from parent to child at fork time (so as not to waste time
585 * if exit comes before the next scan reaches it).
586 *
587 * Similarly, although we'd like to remove rmap_items (so updating counts
588 * and freeing memory) when unmerging an area, it's easier to leave that
589 * to the next pass of ksmd - consider, for example, how ksmd might be
590 * in cmp_and_merge_page on one of the rmap_items we would be removing.
591 */
592static int unmerge_ksm_pages(struct vm_area_struct *vma,
593                 unsigned long start, unsigned long end)
594{
595    unsigned long addr;
596    int err = 0;
597
598    for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
599        if (ksm_test_exit(vma->vm_mm))
600            break;
601        if (signal_pending(current))
602            err = -ERESTARTSYS;
603        else
604            err = break_ksm(vma, addr);
605    }
606    return err;
607}
608
609#ifdef CONFIG_SYSFS
610/*
611 * Only called through the sysfs control interface:
612 */
613static int unmerge_and_remove_all_rmap_items(void)
614{
615    struct mm_slot *mm_slot;
616    struct mm_struct *mm;
617    struct vm_area_struct *vma;
618    int err = 0;
619
620    spin_lock(&ksm_mmlist_lock);
621    ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
622                        struct mm_slot, mm_list);
623    spin_unlock(&ksm_mmlist_lock);
624
625    for (mm_slot = ksm_scan.mm_slot;
626            mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
627        mm = mm_slot->mm;
628        down_read(&mm->mmap_sem);
629        for (vma = mm->mmap; vma; vma = vma->vm_next) {
630            if (ksm_test_exit(mm))
631                break;
632            if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
633                continue;
634            err = unmerge_ksm_pages(vma,
635                        vma->vm_start, vma->vm_end);
636            if (err)
637                goto error;
638        }
639
640        remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
641
642        spin_lock(&ksm_mmlist_lock);
643        ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
644                        struct mm_slot, mm_list);
645        if (ksm_test_exit(mm)) {
646            hlist_del(&mm_slot->link);
647            list_del(&mm_slot->mm_list);
648            spin_unlock(&ksm_mmlist_lock);
649
650            free_mm_slot(mm_slot);
651            clear_bit(MMF_VM_MERGEABLE, &mm->flags);
652            up_read(&mm->mmap_sem);
653            mmdrop(mm);
654        } else {
655            spin_unlock(&ksm_mmlist_lock);
656            up_read(&mm->mmap_sem);
657        }
658    }
659
660    ksm_scan.seqnr = 0;
661    return 0;
662
663error:
664    up_read(&mm->mmap_sem);
665    spin_lock(&ksm_mmlist_lock);
666    ksm_scan.mm_slot = &ksm_mm_head;
667    spin_unlock(&ksm_mmlist_lock);
668    return err;
669}
670#endif /* CONFIG_SYSFS */
671
672static u32 calc_checksum(struct page *page)
673{
674    u32 checksum;
675    void *addr = kmap_atomic(page, KM_USER0);
676    checksum = jhash2(addr, PAGE_SIZE / 4, 17);
677    kunmap_atomic(addr, KM_USER0);
678    return checksum;
679}
680
681static int memcmp_pages(struct page *page1, struct page *page2)
682{
683    char *addr1, *addr2;
684    int ret;
685
686    addr1 = kmap_atomic(page1, KM_USER0);
687    addr2 = kmap_atomic(page2, KM_USER1);
688    ret = memcmp(addr1, addr2, PAGE_SIZE);
689    kunmap_atomic(addr2, KM_USER1);
690    kunmap_atomic(addr1, KM_USER0);
691    return ret;
692}
693
694static inline int pages_identical(struct page *page1, struct page *page2)
695{
696    return !memcmp_pages(page1, page2);
697}
698
699static int write_protect_page(struct vm_area_struct *vma, struct page *page,
700                  pte_t *orig_pte)
701{
702    struct mm_struct *mm = vma->vm_mm;
703    unsigned long addr;
704    pte_t *ptep;
705    spinlock_t *ptl;
706    int swapped;
707    int err = -EFAULT;
708
709    addr = page_address_in_vma(page, vma);
710    if (addr == -EFAULT)
711        goto out;
712
713    BUG_ON(PageTransCompound(page));
714    ptep = page_check_address(page, mm, addr, &ptl, 0);
715    if (!ptep)
716        goto out;
717
718    if (pte_write(*ptep) || pte_dirty(*ptep)) {
719        pte_t entry;
720
721        swapped = PageSwapCache(page);
722        flush_cache_page(vma, addr, page_to_pfn(page));
723        /*
724         * Ok this is tricky, when get_user_pages_fast() run it doesn't
725         * take any lock, therefore the check that we are going to make
726         * with the pagecount against the mapcount is racey and
727         * O_DIRECT can happen right after the check.
728         * So we clear the pte and flush the tlb before the check
729         * this assure us that no O_DIRECT can happen after the check
730         * or in the middle of the check.
731         */
732        entry = ptep_clear_flush(vma, addr, ptep);
733        /*
734         * Check that no O_DIRECT or similar I/O is in progress on the
735         * page
736         */
737        if (page_mapcount(page) + 1 + swapped != page_count(page)) {
738            set_pte_at(mm, addr, ptep, entry);
739            goto out_unlock;
740        }
741        if (pte_dirty(entry))
742            set_page_dirty(page);
743        entry = pte_mkclean(pte_wrprotect(entry));
744        set_pte_at_notify(mm, addr, ptep, entry);
745    }
746    *orig_pte = *ptep;
747    err = 0;
748
749out_unlock:
750    pte_unmap_unlock(ptep, ptl);
751out:
752    return err;
753}
754
755/**
756 * replace_page - replace page in vma by new ksm page
757 * @vma: vma that holds the pte pointing to page
758 * @page: the page we are replacing by kpage
759 * @kpage: the ksm page we replace page by
760 * @orig_pte: the original value of the pte
761 *
762 * Returns 0 on success, -EFAULT on failure.
763 */
764static int replace_page(struct vm_area_struct *vma, struct page *page,
765            struct page *kpage, pte_t orig_pte)
766{
767    struct mm_struct *mm = vma->vm_mm;
768    pgd_t *pgd;
769    pud_t *pud;
770    pmd_t *pmd;
771    pte_t *ptep;
772    spinlock_t *ptl;
773    unsigned long addr;
774    int err = -EFAULT;
775
776    addr = page_address_in_vma(page, vma);
777    if (addr == -EFAULT)
778        goto out;
779
780    pgd = pgd_offset(mm, addr);
781    if (!pgd_present(*pgd))
782        goto out;
783
784    pud = pud_offset(pgd, addr);
785    if (!pud_present(*pud))
786        goto out;
787
788    pmd = pmd_offset(pud, addr);
789    BUG_ON(pmd_trans_huge(*pmd));
790    if (!pmd_present(*pmd))
791        goto out;
792
793    ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
794    if (!pte_same(*ptep, orig_pte)) {
795        pte_unmap_unlock(ptep, ptl);
796        goto out;
797    }
798
799    get_page(kpage);
800    page_add_anon_rmap(kpage, vma, addr);
801
802    flush_cache_page(vma, addr, pte_pfn(*ptep));
803    ptep_clear_flush(vma, addr, ptep);
804    set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
805
806    page_remove_rmap(page);
807    if (!page_mapped(page))
808        try_to_free_swap(page);
809    put_page(page);
810
811    pte_unmap_unlock(ptep, ptl);
812    err = 0;
813out:
814    return err;
815}
816
817static int page_trans_compound_anon_split(struct page *page)
818{
819    int ret = 0;
820    struct page *transhuge_head = page_trans_compound_anon(page);
821    if (transhuge_head) {
822        /* Get the reference on the head to split it. */
823        if (get_page_unless_zero(transhuge_head)) {
824            /*
825             * Recheck we got the reference while the head
826             * was still anonymous.
827             */
828            if (PageAnon(transhuge_head))
829                ret = split_huge_page(transhuge_head);
830            else
831                /*
832                 * Retry later if split_huge_page run
833                 * from under us.
834                 */
835                ret = 1;
836            put_page(transhuge_head);
837        } else
838            /* Retry later if split_huge_page run from under us. */
839            ret = 1;
840    }
841    return ret;
842}
843
844/*
845 * try_to_merge_one_page - take two pages and merge them into one
846 * @vma: the vma that holds the pte pointing to page
847 * @page: the PageAnon page that we want to replace with kpage
848 * @kpage: the PageKsm page that we want to map instead of page,
849 * or NULL the first time when we want to use page as kpage.
850 *
851 * This function returns 0 if the pages were merged, -EFAULT otherwise.
852 */
853static int try_to_merge_one_page(struct vm_area_struct *vma,
854                 struct page *page, struct page *kpage)
855{
856    pte_t orig_pte = __pte(0);
857    int err = -EFAULT;
858
859    if (page == kpage) /* ksm page forked */
860        return 0;
861
862    if (!(vma->vm_flags & VM_MERGEABLE))
863        goto out;
864    if (PageTransCompound(page) && page_trans_compound_anon_split(page))
865        goto out;
866    BUG_ON(PageTransCompound(page));
867    if (!PageAnon(page))
868        goto out;
869
870    /*
871     * We need the page lock to read a stable PageSwapCache in
872     * write_protect_page(). We use trylock_page() instead of
873     * lock_page() because we don't want to wait here - we
874     * prefer to continue scanning and merging different pages,
875     * then come back to this page when it is unlocked.
876     */
877    if (!trylock_page(page))
878        goto out;
879    /*
880     * If this anonymous page is mapped only here, its pte may need
881     * to be write-protected. If it's mapped elsewhere, all of its
882     * ptes are necessarily already write-protected. But in either
883     * case, we need to lock and check page_count is not raised.
884     */
885    if (write_protect_page(vma, page, &orig_pte) == 0) {
886        if (!kpage) {
887            /*
888             * While we hold page lock, upgrade page from
889             * PageAnon+anon_vma to PageKsm+NULL stable_node:
890             * stable_tree_insert() will update stable_node.
891             */
892            set_page_stable_node(page, NULL);
893            mark_page_accessed(page);
894            err = 0;
895        } else if (pages_identical(page, kpage))
896            err = replace_page(vma, page, kpage, orig_pte);
897    }
898
899    if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
900        munlock_vma_page(page);
901        if (!PageMlocked(kpage)) {
902            unlock_page(page);
903            lock_page(kpage);
904            mlock_vma_page(kpage);
905            page = kpage; /* for final unlock */
906        }
907    }
908
909    unlock_page(page);
910out:
911    return err;
912}
913
914/*
915 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
916 * but no new kernel page is allocated: kpage must already be a ksm page.
917 *
918 * This function returns 0 if the pages were merged, -EFAULT otherwise.
919 */
920static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
921                      struct page *page, struct page *kpage)
922{
923    struct mm_struct *mm = rmap_item->mm;
924    struct vm_area_struct *vma;
925    int err = -EFAULT;
926
927    down_read(&mm->mmap_sem);
928    if (ksm_test_exit(mm))
929        goto out;
930    vma = find_vma(mm, rmap_item->address);
931    if (!vma || vma->vm_start > rmap_item->address)
932        goto out;
933
934    err = try_to_merge_one_page(vma, page, kpage);
935    if (err)
936        goto out;
937
938    /* Must get reference to anon_vma while still holding mmap_sem */
939    rmap_item->anon_vma = vma->anon_vma;
940    get_anon_vma(vma->anon_vma);
941out:
942    up_read(&mm->mmap_sem);
943    return err;
944}
945
946/*
947 * try_to_merge_two_pages - take two identical pages and prepare them
948 * to be merged into one page.
949 *
950 * This function returns the kpage if we successfully merged two identical
951 * pages into one ksm page, NULL otherwise.
952 *
953 * Note that this function upgrades page to ksm page: if one of the pages
954 * is already a ksm page, try_to_merge_with_ksm_page should be used.
955 */
956static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
957                       struct page *page,
958                       struct rmap_item *tree_rmap_item,
959                       struct page *tree_page)
960{
961    int err;
962
963    err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
964    if (!err) {
965        err = try_to_merge_with_ksm_page(tree_rmap_item,
966                            tree_page, page);
967        /*
968         * If that fails, we have a ksm page with only one pte
969         * pointing to it: so break it.
970         */
971        if (err)
972            break_cow(rmap_item);
973    }
974    return err ? NULL : page;
975}
976
977/*
978 * stable_tree_search - search for page inside the stable tree
979 *
980 * This function checks if there is a page inside the stable tree
981 * with identical content to the page that we are scanning right now.
982 *
983 * This function returns the stable tree node of identical content if found,
984 * NULL otherwise.
985 */
986static struct page *stable_tree_search(struct page *page)
987{
988    struct rb_node *node = root_stable_tree.rb_node;
989    struct stable_node *stable_node;
990
991    stable_node = page_stable_node(page);
992    if (stable_node) { /* ksm page forked */
993        get_page(page);
994        return page;
995    }
996
997    while (node) {
998        struct page *tree_page;
999        int ret;
1000
1001        cond_resched();
1002        stable_node = rb_entry(node, struct stable_node, node);
1003        tree_page = get_ksm_page(stable_node);
1004        if (!tree_page)
1005            return NULL;
1006
1007        ret = memcmp_pages(page, tree_page);
1008
1009        if (ret < 0) {
1010            put_page(tree_page);
1011            node = node->rb_left;
1012        } else if (ret > 0) {
1013            put_page(tree_page);
1014            node = node->rb_right;
1015        } else
1016            return tree_page;
1017    }
1018
1019    return NULL;
1020}
1021
1022/*
1023 * stable_tree_insert - insert rmap_item pointing to new ksm page
1024 * into the stable tree.
1025 *
1026 * This function returns the stable tree node just allocated on success,
1027 * NULL otherwise.
1028 */
1029static struct stable_node *stable_tree_insert(struct page *kpage)
1030{
1031    struct rb_node **new = &root_stable_tree.rb_node;
1032    struct rb_node *parent = NULL;
1033    struct stable_node *stable_node;
1034
1035    while (*new) {
1036        struct page *tree_page;
1037        int ret;
1038
1039        cond_resched();
1040        stable_node = rb_entry(*new, struct stable_node, node);
1041        tree_page = get_ksm_page(stable_node);
1042        if (!tree_page)
1043            return NULL;
1044
1045        ret = memcmp_pages(kpage, tree_page);
1046        put_page(tree_page);
1047
1048        parent = *new;
1049        if (ret < 0)
1050            new = &parent->rb_left;
1051        else if (ret > 0)
1052            new = &parent->rb_right;
1053        else {
1054            /*
1055             * It is not a bug that stable_tree_search() didn't
1056             * find this node: because at that time our page was
1057             * not yet write-protected, so may have changed since.
1058             */
1059            return NULL;
1060        }
1061    }
1062
1063    stable_node = alloc_stable_node();
1064    if (!stable_node)
1065        return NULL;
1066
1067    rb_link_node(&stable_node->node, parent, new);
1068    rb_insert_color(&stable_node->node, &root_stable_tree);
1069
1070    INIT_HLIST_HEAD(&stable_node->hlist);
1071
1072    stable_node->kpfn = page_to_pfn(kpage);
1073    set_page_stable_node(kpage, stable_node);
1074
1075    return stable_node;
1076}
1077
1078/*
1079 * unstable_tree_search_insert - search for identical page,
1080 * else insert rmap_item into the unstable tree.
1081 *
1082 * This function searches for a page in the unstable tree identical to the
1083 * page currently being scanned; and if no identical page is found in the
1084 * tree, we insert rmap_item as a new object into the unstable tree.
1085 *
1086 * This function returns pointer to rmap_item found to be identical
1087 * to the currently scanned page, NULL otherwise.
1088 *
1089 * This function does both searching and inserting, because they share
1090 * the same walking algorithm in an rbtree.
1091 */
1092static
1093struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1094                          struct page *page,
1095                          struct page **tree_pagep)
1096
1097{
1098    struct rb_node **new = &root_unstable_tree.rb_node;
1099    struct rb_node *parent = NULL;
1100
1101    while (*new) {
1102        struct rmap_item *tree_rmap_item;
1103        struct page *tree_page;
1104        int ret;
1105
1106        cond_resched();
1107        tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1108        tree_page = get_mergeable_page(tree_rmap_item);
1109        if (IS_ERR_OR_NULL(tree_page))
1110            return NULL;
1111
1112        /*
1113         * Don't substitute a ksm page for a forked page.
1114         */
1115        if (page == tree_page) {
1116            put_page(tree_page);
1117            return NULL;
1118        }
1119
1120        ret = memcmp_pages(page, tree_page);
1121
1122        parent = *new;
1123        if (ret < 0) {
1124            put_page(tree_page);
1125            new = &parent->rb_left;
1126        } else if (ret > 0) {
1127            put_page(tree_page);
1128            new = &parent->rb_right;
1129        } else {
1130            *tree_pagep = tree_page;
1131            return tree_rmap_item;
1132        }
1133    }
1134
1135    rmap_item->address |= UNSTABLE_FLAG;
1136    rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1137    rb_link_node(&rmap_item->node, parent, new);
1138    rb_insert_color(&rmap_item->node, &root_unstable_tree);
1139
1140    ksm_pages_unshared++;
1141    return NULL;
1142}
1143
1144/*
1145 * stable_tree_append - add another rmap_item to the linked list of
1146 * rmap_items hanging off a given node of the stable tree, all sharing
1147 * the same ksm page.
1148 */
1149static void stable_tree_append(struct rmap_item *rmap_item,
1150                   struct stable_node *stable_node)
1151{
1152    rmap_item->head = stable_node;
1153    rmap_item->address |= STABLE_FLAG;
1154    hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1155
1156    if (rmap_item->hlist.next)
1157        ksm_pages_sharing++;
1158    else
1159        ksm_pages_shared++;
1160}
1161
1162/*
1163 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1164 * if not, compare checksum to previous and if it's the same, see if page can
1165 * be inserted into the unstable tree, or merged with a page already there and
1166 * both transferred to the stable tree.
1167 *
1168 * @page: the page that we are searching identical page to.
1169 * @rmap_item: the reverse mapping into the virtual address of this page
1170 */
1171static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1172{
1173    struct rmap_item *tree_rmap_item;
1174    struct page *tree_page = NULL;
1175    struct stable_node *stable_node;
1176    struct page *kpage;
1177    unsigned int checksum;
1178    int err;
1179
1180    remove_rmap_item_from_tree(rmap_item);
1181
1182    /* We first start with searching the page inside the stable tree */
1183    kpage = stable_tree_search(page);
1184    if (kpage) {
1185        err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1186        if (!err) {
1187            /*
1188             * The page was successfully merged:
1189             * add its rmap_item to the stable tree.
1190             */
1191            lock_page(kpage);
1192            stable_tree_append(rmap_item, page_stable_node(kpage));
1193            unlock_page(kpage);
1194        }
1195        put_page(kpage);
1196        return;
1197    }
1198
1199    /*
1200     * If the hash value of the page has changed from the last time
1201     * we calculated it, this page is changing frequently: therefore we
1202     * don't want to insert it in the unstable tree, and we don't want
1203     * to waste our time searching for something identical to it there.
1204     */
1205    checksum = calc_checksum(page);
1206    if (rmap_item->oldchecksum != checksum) {
1207        rmap_item->oldchecksum = checksum;
1208        return;
1209    }
1210
1211    tree_rmap_item =
1212        unstable_tree_search_insert(rmap_item, page, &tree_page);
1213    if (tree_rmap_item) {
1214        kpage = try_to_merge_two_pages(rmap_item, page,
1215                        tree_rmap_item, tree_page);
1216        put_page(tree_page);
1217        /*
1218         * As soon as we merge this page, we want to remove the
1219         * rmap_item of the page we have merged with from the unstable
1220         * tree, and insert it instead as new node in the stable tree.
1221         */
1222        if (kpage) {
1223            remove_rmap_item_from_tree(tree_rmap_item);
1224
1225            lock_page(kpage);
1226            stable_node = stable_tree_insert(kpage);
1227            if (stable_node) {
1228                stable_tree_append(tree_rmap_item, stable_node);
1229                stable_tree_append(rmap_item, stable_node);
1230            }
1231            unlock_page(kpage);
1232
1233            /*
1234             * If we fail to insert the page into the stable tree,
1235             * we will have 2 virtual addresses that are pointing
1236             * to a ksm page left outside the stable tree,
1237             * in which case we need to break_cow on both.
1238             */
1239            if (!stable_node) {
1240                break_cow(tree_rmap_item);
1241                break_cow(rmap_item);
1242            }
1243        }
1244    }
1245}
1246
1247static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1248                        struct rmap_item **rmap_list,
1249                        unsigned long addr)
1250{
1251    struct rmap_item *rmap_item;
1252
1253    while (*rmap_list) {
1254        rmap_item = *rmap_list;
1255        if ((rmap_item->address & PAGE_MASK) == addr)
1256            return rmap_item;
1257        if (rmap_item->address > addr)
1258            break;
1259        *rmap_list = rmap_item->rmap_list;
1260        remove_rmap_item_from_tree(rmap_item);
1261        free_rmap_item(rmap_item);
1262    }
1263
1264    rmap_item = alloc_rmap_item();
1265    if (rmap_item) {
1266        /* It has already been zeroed */
1267        rmap_item->mm = mm_slot->mm;
1268        rmap_item->address = addr;
1269        rmap_item->rmap_list = *rmap_list;
1270        *rmap_list = rmap_item;
1271    }
1272    return rmap_item;
1273}
1274
1275static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1276{
1277    struct mm_struct *mm;
1278    struct mm_slot *slot;
1279    struct vm_area_struct *vma;
1280    struct rmap_item *rmap_item;
1281
1282    if (list_empty(&ksm_mm_head.mm_list))
1283        return NULL;
1284
1285    slot = ksm_scan.mm_slot;
1286    if (slot == &ksm_mm_head) {
1287        /*
1288         * A number of pages can hang around indefinitely on per-cpu
1289         * pagevecs, raised page count preventing write_protect_page
1290         * from merging them. Though it doesn't really matter much,
1291         * it is puzzling to see some stuck in pages_volatile until
1292         * other activity jostles them out, and they also prevented
1293         * LTP's KSM test from succeeding deterministically; so drain
1294         * them here (here rather than on entry to ksm_do_scan(),
1295         * so we don't IPI too often when pages_to_scan is set low).
1296         */
1297        lru_add_drain_all();
1298
1299        root_unstable_tree = RB_ROOT;
1300
1301        spin_lock(&ksm_mmlist_lock);
1302        slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1303        ksm_scan.mm_slot = slot;
1304        spin_unlock(&ksm_mmlist_lock);
1305        /*
1306         * Although we tested list_empty() above, a racing __ksm_exit
1307         * of the last mm on the list may have removed it since then.
1308         */
1309        if (slot == &ksm_mm_head)
1310            return NULL;
1311next_mm:
1312        ksm_scan.address = 0;
1313        ksm_scan.rmap_list = &slot->rmap_list;
1314    }
1315
1316    mm = slot->mm;
1317    down_read(&mm->mmap_sem);
1318    if (ksm_test_exit(mm))
1319        vma = NULL;
1320    else
1321        vma = find_vma(mm, ksm_scan.address);
1322
1323    for (; vma; vma = vma->vm_next) {
1324        if (!(vma->vm_flags & VM_MERGEABLE))
1325            continue;
1326        if (ksm_scan.address < vma->vm_start)
1327            ksm_scan.address = vma->vm_start;
1328        if (!vma->anon_vma)
1329            ksm_scan.address = vma->vm_end;
1330
1331        while (ksm_scan.address < vma->vm_end) {
1332            if (ksm_test_exit(mm))
1333                break;
1334            *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1335            if (IS_ERR_OR_NULL(*page)) {
1336                ksm_scan.address += PAGE_SIZE;
1337                cond_resched();
1338                continue;
1339            }
1340            if (PageAnon(*page) ||
1341                page_trans_compound_anon(*page)) {
1342                flush_anon_page(vma, *page, ksm_scan.address);
1343                flush_dcache_page(*page);
1344                rmap_item = get_next_rmap_item(slot,
1345                    ksm_scan.rmap_list, ksm_scan.address);
1346                if (rmap_item) {
1347                    ksm_scan.rmap_list =
1348                            &rmap_item->rmap_list;
1349                    ksm_scan.address += PAGE_SIZE;
1350                } else
1351                    put_page(*page);
1352                up_read(&mm->mmap_sem);
1353                return rmap_item;
1354            }
1355            put_page(*page);
1356            ksm_scan.address += PAGE_SIZE;
1357            cond_resched();
1358        }
1359    }
1360
1361    if (ksm_test_exit(mm)) {
1362        ksm_scan.address = 0;
1363        ksm_scan.rmap_list = &slot->rmap_list;
1364    }
1365    /*
1366     * Nuke all the rmap_items that are above this current rmap:
1367     * because there were no VM_MERGEABLE vmas with such addresses.
1368     */
1369    remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1370
1371    spin_lock(&ksm_mmlist_lock);
1372    ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1373                        struct mm_slot, mm_list);
1374    if (ksm_scan.address == 0) {
1375        /*
1376         * We've completed a full scan of all vmas, holding mmap_sem
1377         * throughout, and found no VM_MERGEABLE: so do the same as
1378         * __ksm_exit does to remove this mm from all our lists now.
1379         * This applies either when cleaning up after __ksm_exit
1380         * (but beware: we can reach here even before __ksm_exit),
1381         * or when all VM_MERGEABLE areas have been unmapped (and
1382         * mmap_sem then protects against race with MADV_MERGEABLE).
1383         */
1384        hlist_del(&slot->link);
1385        list_del(&slot->mm_list);
1386        spin_unlock(&ksm_mmlist_lock);
1387
1388        free_mm_slot(slot);
1389        clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1390        up_read(&mm->mmap_sem);
1391        mmdrop(mm);
1392    } else {
1393        spin_unlock(&ksm_mmlist_lock);
1394        up_read(&mm->mmap_sem);
1395    }
1396
1397    /* Repeat until we've completed scanning the whole list */
1398    slot = ksm_scan.mm_slot;
1399    if (slot != &ksm_mm_head)
1400        goto next_mm;
1401
1402    ksm_scan.seqnr++;
1403    return NULL;
1404}
1405
1406/**
1407 * ksm_do_scan - the ksm scanner main worker function.
1408 * @scan_npages - number of pages we want to scan before we return.
1409 */
1410static void ksm_do_scan(unsigned int scan_npages)
1411{
1412    struct rmap_item *rmap_item;
1413    struct page *uninitialized_var(page);
1414
1415    while (scan_npages-- && likely(!freezing(current))) {
1416        cond_resched();
1417        rmap_item = scan_get_next_rmap_item(&page);
1418        if (!rmap_item)
1419            return;
1420        if (!PageKsm(page) || !in_stable_tree(rmap_item))
1421            cmp_and_merge_page(page, rmap_item);
1422        put_page(page);
1423    }
1424}
1425
1426static int ksmd_should_run(void)
1427{
1428    return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1429}
1430
1431static int ksm_scan_thread(void *nothing)
1432{
1433    set_freezable();
1434    set_user_nice(current, 5);
1435
1436    while (!kthread_should_stop()) {
1437        mutex_lock(&ksm_thread_mutex);
1438        if (ksmd_should_run())
1439            ksm_do_scan(ksm_thread_pages_to_scan);
1440        mutex_unlock(&ksm_thread_mutex);
1441
1442        try_to_freeze();
1443
1444        if (ksmd_should_run()) {
1445            schedule_timeout_interruptible(
1446                msecs_to_jiffies(ksm_thread_sleep_millisecs));
1447        } else {
1448            wait_event_freezable(ksm_thread_wait,
1449                ksmd_should_run() || kthread_should_stop());
1450        }
1451    }
1452    return 0;
1453}
1454
1455int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1456        unsigned long end, int advice, unsigned long *vm_flags)
1457{
1458    struct mm_struct *mm = vma->vm_mm;
1459    int err;
1460
1461    switch (advice) {
1462    case MADV_MERGEABLE:
1463        /*
1464         * Be somewhat over-protective for now!
1465         */
1466        if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1467                 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1468                 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1469                 VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
1470            return 0; /* just ignore the advice */
1471
1472        if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1473            err = __ksm_enter(mm);
1474            if (err)
1475                return err;
1476        }
1477
1478        *vm_flags |= VM_MERGEABLE;
1479        break;
1480
1481    case MADV_UNMERGEABLE:
1482        if (!(*vm_flags & VM_MERGEABLE))
1483            return 0; /* just ignore the advice */
1484
1485        if (vma->anon_vma) {
1486            err = unmerge_ksm_pages(vma, start, end);
1487            if (err)
1488                return err;
1489        }
1490
1491        *vm_flags &= ~VM_MERGEABLE;
1492        break;
1493    }
1494
1495    return 0;
1496}
1497
1498int __ksm_enter(struct mm_struct *mm)
1499{
1500    struct mm_slot *mm_slot;
1501    int needs_wakeup;
1502
1503    mm_slot = alloc_mm_slot();
1504    if (!mm_slot)
1505        return -ENOMEM;
1506
1507    /* Check ksm_run too? Would need tighter locking */
1508    needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1509
1510    spin_lock(&ksm_mmlist_lock);
1511    insert_to_mm_slots_hash(mm, mm_slot);
1512    /*
1513     * Insert just behind the scanning cursor, to let the area settle
1514     * down a little; when fork is followed by immediate exec, we don't
1515     * want ksmd to waste time setting up and tearing down an rmap_list.
1516     */
1517    list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1518    spin_unlock(&ksm_mmlist_lock);
1519
1520    set_bit(MMF_VM_MERGEABLE, &mm->flags);
1521    atomic_inc(&mm->mm_count);
1522
1523    if (needs_wakeup)
1524        wake_up_interruptible(&ksm_thread_wait);
1525
1526    return 0;
1527}
1528
1529void __ksm_exit(struct mm_struct *mm)
1530{
1531    struct mm_slot *mm_slot;
1532    int easy_to_free = 0;
1533
1534    /*
1535     * This process is exiting: if it's straightforward (as is the
1536     * case when ksmd was never running), free mm_slot immediately.
1537     * But if it's at the cursor or has rmap_items linked to it, use
1538     * mmap_sem to synchronize with any break_cows before pagetables
1539     * are freed, and leave the mm_slot on the list for ksmd to free.
1540     * Beware: ksm may already have noticed it exiting and freed the slot.
1541     */
1542
1543    spin_lock(&ksm_mmlist_lock);
1544    mm_slot = get_mm_slot(mm);
1545    if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1546        if (!mm_slot->rmap_list) {
1547            hlist_del(&mm_slot->link);
1548            list_del(&mm_slot->mm_list);
1549            easy_to_free = 1;
1550        } else {
1551            list_move(&mm_slot->mm_list,
1552                  &ksm_scan.mm_slot->mm_list);
1553        }
1554    }
1555    spin_unlock(&ksm_mmlist_lock);
1556
1557    if (easy_to_free) {
1558        free_mm_slot(mm_slot);
1559        clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1560        mmdrop(mm);
1561    } else if (mm_slot) {
1562        down_write(&mm->mmap_sem);
1563        up_write(&mm->mmap_sem);
1564    }
1565}
1566
1567struct page *ksm_does_need_to_copy(struct page *page,
1568            struct vm_area_struct *vma, unsigned long address)
1569{
1570    struct page *new_page;
1571
1572    new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1573    if (new_page) {
1574        copy_user_highpage(new_page, page, address, vma);
1575
1576        SetPageDirty(new_page);
1577        __SetPageUptodate(new_page);
1578        SetPageSwapBacked(new_page);
1579        __set_page_locked(new_page);
1580
1581        if (page_evictable(new_page, vma))
1582            lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1583        else
1584            add_page_to_unevictable_list(new_page);
1585    }
1586
1587    return new_page;
1588}
1589
1590int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1591            unsigned long *vm_flags)
1592{
1593    struct stable_node *stable_node;
1594    struct rmap_item *rmap_item;
1595    struct hlist_node *hlist;
1596    unsigned int mapcount = page_mapcount(page);
1597    int referenced = 0;
1598    int search_new_forks = 0;
1599
1600    VM_BUG_ON(!PageKsm(page));
1601    VM_BUG_ON(!PageLocked(page));
1602
1603    stable_node = page_stable_node(page);
1604    if (!stable_node)
1605        return 0;
1606again:
1607    hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1608        struct anon_vma *anon_vma = rmap_item->anon_vma;
1609        struct anon_vma_chain *vmac;
1610        struct vm_area_struct *vma;
1611
1612        anon_vma_lock(anon_vma);
1613        list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1614            vma = vmac->vma;
1615            if (rmap_item->address < vma->vm_start ||
1616                rmap_item->address >= vma->vm_end)
1617                continue;
1618            /*
1619             * Initially we examine only the vma which covers this
1620             * rmap_item; but later, if there is still work to do,
1621             * we examine covering vmas in other mms: in case they
1622             * were forked from the original since ksmd passed.
1623             */
1624            if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1625                continue;
1626
1627            if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1628                continue;
1629
1630            referenced += page_referenced_one(page, vma,
1631                rmap_item->address, &mapcount, vm_flags);
1632            if (!search_new_forks || !mapcount)
1633                break;
1634        }
1635        anon_vma_unlock(anon_vma);
1636        if (!mapcount)
1637            goto out;
1638    }
1639    if (!search_new_forks++)
1640        goto again;
1641out:
1642    return referenced;
1643}
1644
1645int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1646{
1647    struct stable_node *stable_node;
1648    struct hlist_node *hlist;
1649    struct rmap_item *rmap_item;
1650    int ret = SWAP_AGAIN;
1651    int search_new_forks = 0;
1652
1653    VM_BUG_ON(!PageKsm(page));
1654    VM_BUG_ON(!PageLocked(page));
1655
1656    stable_node = page_stable_node(page);
1657    if (!stable_node)
1658        return SWAP_FAIL;
1659again:
1660    hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1661        struct anon_vma *anon_vma = rmap_item->anon_vma;
1662        struct anon_vma_chain *vmac;
1663        struct vm_area_struct *vma;
1664
1665        anon_vma_lock(anon_vma);
1666        list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1667            vma = vmac->vma;
1668            if (rmap_item->address < vma->vm_start ||
1669                rmap_item->address >= vma->vm_end)
1670                continue;
1671            /*
1672             * Initially we examine only the vma which covers this
1673             * rmap_item; but later, if there is still work to do,
1674             * we examine covering vmas in other mms: in case they
1675             * were forked from the original since ksmd passed.
1676             */
1677            if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1678                continue;
1679
1680            ret = try_to_unmap_one(page, vma,
1681                    rmap_item->address, flags);
1682            if (ret != SWAP_AGAIN || !page_mapped(page)) {
1683                anon_vma_unlock(anon_vma);
1684                goto out;
1685            }
1686        }
1687        anon_vma_unlock(anon_vma);
1688    }
1689    if (!search_new_forks++)
1690        goto again;
1691out:
1692    return ret;
1693}
1694
1695#ifdef CONFIG_MIGRATION
1696int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1697          struct vm_area_struct *, unsigned long, void *), void *arg)
1698{
1699    struct stable_node *stable_node;
1700    struct hlist_node *hlist;
1701    struct rmap_item *rmap_item;
1702    int ret = SWAP_AGAIN;
1703    int search_new_forks = 0;
1704
1705    VM_BUG_ON(!PageKsm(page));
1706    VM_BUG_ON(!PageLocked(page));
1707
1708    stable_node = page_stable_node(page);
1709    if (!stable_node)
1710        return ret;
1711again:
1712    hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1713        struct anon_vma *anon_vma = rmap_item->anon_vma;
1714        struct anon_vma_chain *vmac;
1715        struct vm_area_struct *vma;
1716
1717        anon_vma_lock(anon_vma);
1718        list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1719            vma = vmac->vma;
1720            if (rmap_item->address < vma->vm_start ||
1721                rmap_item->address >= vma->vm_end)
1722                continue;
1723            /*
1724             * Initially we examine only the vma which covers this
1725             * rmap_item; but later, if there is still work to do,
1726             * we examine covering vmas in other mms: in case they
1727             * were forked from the original since ksmd passed.
1728             */
1729            if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1730                continue;
1731
1732            ret = rmap_one(page, vma, rmap_item->address, arg);
1733            if (ret != SWAP_AGAIN) {
1734                anon_vma_unlock(anon_vma);
1735                goto out;
1736            }
1737        }
1738        anon_vma_unlock(anon_vma);
1739    }
1740    if (!search_new_forks++)
1741        goto again;
1742out:
1743    return ret;
1744}
1745
1746void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1747{
1748    struct stable_node *stable_node;
1749
1750    VM_BUG_ON(!PageLocked(oldpage));
1751    VM_BUG_ON(!PageLocked(newpage));
1752    VM_BUG_ON(newpage->mapping != oldpage->mapping);
1753
1754    stable_node = page_stable_node(newpage);
1755    if (stable_node) {
1756        VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1757        stable_node->kpfn = page_to_pfn(newpage);
1758    }
1759}
1760#endif /* CONFIG_MIGRATION */
1761
1762#ifdef CONFIG_MEMORY_HOTREMOVE
1763static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1764                         unsigned long end_pfn)
1765{
1766    struct rb_node *node;
1767
1768    for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1769        struct stable_node *stable_node;
1770
1771        stable_node = rb_entry(node, struct stable_node, node);
1772        if (stable_node->kpfn >= start_pfn &&
1773            stable_node->kpfn < end_pfn)
1774            return stable_node;
1775    }
1776    return NULL;
1777}
1778
1779static int ksm_memory_callback(struct notifier_block *self,
1780                   unsigned long action, void *arg)
1781{
1782    struct memory_notify *mn = arg;
1783    struct stable_node *stable_node;
1784
1785    switch (action) {
1786    case MEM_GOING_OFFLINE:
1787        /*
1788         * Keep it very simple for now: just lock out ksmd and
1789         * MADV_UNMERGEABLE while any memory is going offline.
1790         * mutex_lock_nested() is necessary because lockdep was alarmed
1791         * that here we take ksm_thread_mutex inside notifier chain
1792         * mutex, and later take notifier chain mutex inside
1793         * ksm_thread_mutex to unlock it. But that's safe because both
1794         * are inside mem_hotplug_mutex.
1795         */
1796        mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
1797        break;
1798
1799    case MEM_OFFLINE:
1800        /*
1801         * Most of the work is done by page migration; but there might
1802         * be a few stable_nodes left over, still pointing to struct
1803         * pages which have been offlined: prune those from the tree.
1804         */
1805        while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1806                    mn->start_pfn + mn->nr_pages)) != NULL)
1807            remove_node_from_stable_tree(stable_node);
1808        /* fallthrough */
1809
1810    case MEM_CANCEL_OFFLINE:
1811        mutex_unlock(&ksm_thread_mutex);
1812        break;
1813    }
1814    return NOTIFY_OK;
1815}
1816#endif /* CONFIG_MEMORY_HOTREMOVE */
1817
1818#ifdef CONFIG_SYSFS
1819/*
1820 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1821 */
1822
1823#define KSM_ATTR_RO(_name) \
1824    static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1825#define KSM_ATTR(_name) \
1826    static struct kobj_attribute _name##_attr = \
1827        __ATTR(_name, 0644, _name##_show, _name##_store)
1828
1829static ssize_t sleep_millisecs_show(struct kobject *kobj,
1830                    struct kobj_attribute *attr, char *buf)
1831{
1832    return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1833}
1834
1835static ssize_t sleep_millisecs_store(struct kobject *kobj,
1836                     struct kobj_attribute *attr,
1837                     const char *buf, size_t count)
1838{
1839    unsigned long msecs;
1840    int err;
1841
1842    err = strict_strtoul(buf, 10, &msecs);
1843    if (err || msecs > UINT_MAX)
1844        return -EINVAL;
1845
1846    ksm_thread_sleep_millisecs = msecs;
1847
1848    return count;
1849}
1850KSM_ATTR(sleep_millisecs);
1851
1852static ssize_t pages_to_scan_show(struct kobject *kobj,
1853                  struct kobj_attribute *attr, char *buf)
1854{
1855    return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1856}
1857
1858static ssize_t pages_to_scan_store(struct kobject *kobj,
1859                   struct kobj_attribute *attr,
1860                   const char *buf, size_t count)
1861{
1862    int err;
1863    unsigned long nr_pages;
1864
1865    err = strict_strtoul(buf, 10, &nr_pages);
1866    if (err || nr_pages > UINT_MAX)
1867        return -EINVAL;
1868
1869    ksm_thread_pages_to_scan = nr_pages;
1870
1871    return count;
1872}
1873KSM_ATTR(pages_to_scan);
1874
1875static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1876            char *buf)
1877{
1878    return sprintf(buf, "%u\n", ksm_run);
1879}
1880
1881static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1882             const char *buf, size_t count)
1883{
1884    int err;
1885    unsigned long flags;
1886
1887    err = strict_strtoul(buf, 10, &flags);
1888    if (err || flags > UINT_MAX)
1889        return -EINVAL;
1890    if (flags > KSM_RUN_UNMERGE)
1891        return -EINVAL;
1892
1893    /*
1894     * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1895     * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1896     * breaking COW to free the pages_shared (but leaves mm_slots
1897     * on the list for when ksmd may be set running again).
1898     */
1899
1900    mutex_lock(&ksm_thread_mutex);
1901    if (ksm_run != flags) {
1902        ksm_run = flags;
1903        if (flags & KSM_RUN_UNMERGE) {
1904            int oom_score_adj;
1905
1906            oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
1907            err = unmerge_and_remove_all_rmap_items();
1908            test_set_oom_score_adj(oom_score_adj);
1909            if (err) {
1910                ksm_run = KSM_RUN_STOP;
1911                count = err;
1912            }
1913        }
1914    }
1915    mutex_unlock(&ksm_thread_mutex);
1916
1917    if (flags & KSM_RUN_MERGE)
1918        wake_up_interruptible(&ksm_thread_wait);
1919
1920    return count;
1921}
1922KSM_ATTR(run);
1923
1924static ssize_t pages_shared_show(struct kobject *kobj,
1925                 struct kobj_attribute *attr, char *buf)
1926{
1927    return sprintf(buf, "%lu\n", ksm_pages_shared);
1928}
1929KSM_ATTR_RO(pages_shared);
1930
1931static ssize_t pages_sharing_show(struct kobject *kobj,
1932                  struct kobj_attribute *attr, char *buf)
1933{
1934    return sprintf(buf, "%lu\n", ksm_pages_sharing);
1935}
1936KSM_ATTR_RO(pages_sharing);
1937
1938static ssize_t pages_unshared_show(struct kobject *kobj,
1939                   struct kobj_attribute *attr, char *buf)
1940{
1941    return sprintf(buf, "%lu\n", ksm_pages_unshared);
1942}
1943KSM_ATTR_RO(pages_unshared);
1944
1945static ssize_t pages_volatile_show(struct kobject *kobj,
1946                   struct kobj_attribute *attr, char *buf)
1947{
1948    long ksm_pages_volatile;
1949
1950    ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1951                - ksm_pages_sharing - ksm_pages_unshared;
1952    /*
1953     * It was not worth any locking to calculate that statistic,
1954     * but it might therefore sometimes be negative: conceal that.
1955     */
1956    if (ksm_pages_volatile < 0)
1957        ksm_pages_volatile = 0;
1958    return sprintf(buf, "%ld\n", ksm_pages_volatile);
1959}
1960KSM_ATTR_RO(pages_volatile);
1961
1962static ssize_t full_scans_show(struct kobject *kobj,
1963                   struct kobj_attribute *attr, char *buf)
1964{
1965    return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1966}
1967KSM_ATTR_RO(full_scans);
1968
1969static struct attribute *ksm_attrs[] = {
1970    &sleep_millisecs_attr.attr,
1971    &pages_to_scan_attr.attr,
1972    &run_attr.attr,
1973    &pages_shared_attr.attr,
1974    &pages_sharing_attr.attr,
1975    &pages_unshared_attr.attr,
1976    &pages_volatile_attr.attr,
1977    &full_scans_attr.attr,
1978    NULL,
1979};
1980
1981static struct attribute_group ksm_attr_group = {
1982    .attrs = ksm_attrs,
1983    .name = "ksm",
1984};
1985#endif /* CONFIG_SYSFS */
1986
1987static int __init ksm_init(void)
1988{
1989    struct task_struct *ksm_thread;
1990    int err;
1991
1992    err = ksm_slab_init();
1993    if (err)
1994        goto out;
1995
1996    ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1997    if (IS_ERR(ksm_thread)) {
1998        printk(KERN_ERR "ksm: creating kthread failed\n");
1999        err = PTR_ERR(ksm_thread);
2000        goto out_free;
2001    }
2002
2003#ifdef CONFIG_SYSFS
2004    err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2005    if (err) {
2006        printk(KERN_ERR "ksm: register sysfs failed\n");
2007        kthread_stop(ksm_thread);
2008        goto out_free;
2009    }
2010#else
2011    ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2012
2013#endif /* CONFIG_SYSFS */
2014
2015#ifdef CONFIG_MEMORY_HOTREMOVE
2016    /*
2017     * Choose a high priority since the callback takes ksm_thread_mutex:
2018     * later callbacks could only be taking locks which nest within that.
2019     */
2020    hotplug_memory_notifier(ksm_memory_callback, 100);
2021#endif
2022    return 0;
2023
2024out_free:
2025    ksm_slab_free();
2026out:
2027    return err;
2028}
2029module_init(ksm_init)
2030

Archive Download this file



interactive