Loading...
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/sched/mm.h>
23#include <linux/sched/coredump.h>
24#include <linux/rwsem.h>
25#include <linux/pagemap.h>
26#include <linux/rmap.h>
27#include <linux/spinlock.h>
28#include <linux/jhash.h>
29#include <linux/delay.h>
30#include <linux/kthread.h>
31#include <linux/wait.h>
32#include <linux/slab.h>
33#include <linux/rbtree.h>
34#include <linux/memory.h>
35#include <linux/mmu_notifier.h>
36#include <linux/swap.h>
37#include <linux/ksm.h>
38#include <linux/hashtable.h>
39#include <linux/freezer.h>
40#include <linux/oom.h>
41#include <linux/numa.h>
42
43#include <asm/tlbflush.h>
44#include "internal.h"
45
46#ifdef CONFIG_NUMA
47#define NUMA(x) (x)
48#define DO_NUMA(x) do { (x); } while (0)
49#else
50#define NUMA(x) (0)
51#define DO_NUMA(x) do { } while (0)
52#endif
53
54/*
55 * A few notes about the KSM scanning process,
56 * to make it easier to understand the data structures below:
57 *
58 * In order to reduce excessive scanning, KSM sorts the memory pages by their
59 * contents into a data structure that holds pointers to the pages' locations.
60 *
61 * Since the contents of the pages may change at any moment, KSM cannot just
62 * insert the pages into a normal sorted tree and expect it to find anything.
63 * Therefore KSM uses two data structures - the stable and the unstable tree.
64 *
65 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
66 * by their contents. Because each such page is write-protected, searching on
67 * this tree is fully assured to be working (except when pages are unmapped),
68 * and therefore this tree is called the stable tree.
69 *
70 * In addition to the stable tree, KSM uses a second data structure called the
71 * unstable tree: this tree holds pointers to pages which have been found to
72 * be "unchanged for a period of time". The unstable tree sorts these pages
73 * by their contents, but since they are not write-protected, KSM cannot rely
74 * upon the unstable tree to work correctly - the unstable tree is liable to
75 * be corrupted as its contents are modified, and so it is called unstable.
76 *
77 * KSM solves this problem by several techniques:
78 *
79 * 1) The unstable tree is flushed every time KSM completes scanning all
80 * memory areas, and then the tree is rebuilt again from the beginning.
81 * 2) KSM will only insert into the unstable tree, pages whose hash value
82 * has not changed since the previous scan of all memory areas.
83 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
84 * colors of the nodes and not on their contents, assuring that even when
85 * the tree gets "corrupted" it won't get out of balance, so scanning time
86 * remains the same (also, searching and inserting nodes in an rbtree uses
87 * the same algorithm, so we have no overhead when we flush and rebuild).
88 * 4) KSM never flushes the stable tree, which means that even if it were to
89 * take 10 attempts to find a page in the unstable tree, once it is found,
90 * it is secured in the stable tree. (When we scan a new page, we first
91 * compare it against the stable tree, and then against the unstable tree.)
92 *
93 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
94 * stable trees and multiple unstable trees: one of each for each NUMA node.
95 */
96
97/**
98 * struct mm_slot - ksm information per mm that is being scanned
99 * @link: link to the mm_slots hash list
100 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
101 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
102 * @mm: the mm that this information is valid for
103 */
104struct mm_slot {
105 struct hlist_node link;
106 struct list_head mm_list;
107 struct rmap_item *rmap_list;
108 struct mm_struct *mm;
109};
110
111/**
112 * struct ksm_scan - cursor for scanning
113 * @mm_slot: the current mm_slot we are scanning
114 * @address: the next address inside that to be scanned
115 * @rmap_list: link to the next rmap to be scanned in the rmap_list
116 * @seqnr: count of completed full scans (needed when removing unstable node)
117 *
118 * There is only the one ksm_scan instance of this cursor structure.
119 */
120struct ksm_scan {
121 struct mm_slot *mm_slot;
122 unsigned long address;
123 struct rmap_item **rmap_list;
124 unsigned long seqnr;
125};
126
127/**
128 * struct stable_node - node of the stable rbtree
129 * @node: rb node of this ksm page in the stable tree
130 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
131 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
132 * @list: linked into migrate_nodes, pending placement in the proper node tree
133 * @hlist: hlist head of rmap_items using this ksm page
134 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
135 * @chain_prune_time: time of the last full garbage collection
136 * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
137 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
138 */
139struct stable_node {
140 union {
141 struct rb_node node; /* when node of stable tree */
142 struct { /* when listed for migration */
143 struct list_head *head;
144 struct {
145 struct hlist_node hlist_dup;
146 struct list_head list;
147 };
148 };
149 };
150 struct hlist_head hlist;
151 union {
152 unsigned long kpfn;
153 unsigned long chain_prune_time;
154 };
155 /*
156 * STABLE_NODE_CHAIN can be any negative number in
157 * rmap_hlist_len negative range, but better not -1 to be able
158 * to reliably detect underflows.
159 */
160#define STABLE_NODE_CHAIN -1024
161 int rmap_hlist_len;
162#ifdef CONFIG_NUMA
163 int nid;
164#endif
165};
166
167/**
168 * struct rmap_item - reverse mapping item for virtual addresses
169 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
170 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
171 * @nid: NUMA node id of unstable tree in which linked (may not match page)
172 * @mm: the memory structure this rmap_item is pointing into
173 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
174 * @oldchecksum: previous checksum of the page at that virtual address
175 * @node: rb node of this rmap_item in the unstable tree
176 * @head: pointer to stable_node heading this list in the stable tree
177 * @hlist: link into hlist of rmap_items hanging off that stable_node
178 */
179struct rmap_item {
180 struct rmap_item *rmap_list;
181 union {
182 struct anon_vma *anon_vma; /* when stable */
183#ifdef CONFIG_NUMA
184 int nid; /* when node of unstable tree */
185#endif
186 };
187 struct mm_struct *mm;
188 unsigned long address; /* + low bits used for flags below */
189 unsigned int oldchecksum; /* when unstable */
190 union {
191 struct rb_node node; /* when node of unstable tree */
192 struct { /* when listed from stable tree */
193 struct stable_node *head;
194 struct hlist_node hlist;
195 };
196 };
197};
198
199#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
200#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
201#define STABLE_FLAG 0x200 /* is listed from the stable tree */
202
203/* The stable and unstable tree heads */
204static struct rb_root one_stable_tree[1] = { RB_ROOT };
205static struct rb_root one_unstable_tree[1] = { RB_ROOT };
206static struct rb_root *root_stable_tree = one_stable_tree;
207static struct rb_root *root_unstable_tree = one_unstable_tree;
208
209/* Recently migrated nodes of stable tree, pending proper placement */
210static LIST_HEAD(migrate_nodes);
211#define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
212
213#define MM_SLOTS_HASH_BITS 10
214static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
215
216static struct mm_slot ksm_mm_head = {
217 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
218};
219static struct ksm_scan ksm_scan = {
220 .mm_slot = &ksm_mm_head,
221};
222
223static struct kmem_cache *rmap_item_cache;
224static struct kmem_cache *stable_node_cache;
225static struct kmem_cache *mm_slot_cache;
226
227/* The number of nodes in the stable tree */
228static unsigned long ksm_pages_shared;
229
230/* The number of page slots additionally sharing those nodes */
231static unsigned long ksm_pages_sharing;
232
233/* The number of nodes in the unstable tree */
234static unsigned long ksm_pages_unshared;
235
236/* The number of rmap_items in use: to calculate pages_volatile */
237static unsigned long ksm_rmap_items;
238
239/* The number of stable_node chains */
240static unsigned long ksm_stable_node_chains;
241
242/* The number of stable_node dups linked to the stable_node chains */
243static unsigned long ksm_stable_node_dups;
244
245/* Delay in pruning stale stable_node_dups in the stable_node_chains */
246static int ksm_stable_node_chains_prune_millisecs = 2000;
247
248/* Maximum number of page slots sharing a stable node */
249static int ksm_max_page_sharing = 256;
250
251/* Number of pages ksmd should scan in one batch */
252static unsigned int ksm_thread_pages_to_scan = 100;
253
254/* Milliseconds ksmd should sleep between batches */
255static unsigned int ksm_thread_sleep_millisecs = 20;
256
257/* Checksum of an empty (zeroed) page */
258static unsigned int zero_checksum __read_mostly;
259
260/* Whether to merge empty (zeroed) pages with actual zero pages */
261static bool ksm_use_zero_pages __read_mostly;
262
263#ifdef CONFIG_NUMA
264/* Zeroed when merging across nodes is not allowed */
265static unsigned int ksm_merge_across_nodes = 1;
266static int ksm_nr_node_ids = 1;
267#else
268#define ksm_merge_across_nodes 1U
269#define ksm_nr_node_ids 1
270#endif
271
272#define KSM_RUN_STOP 0
273#define KSM_RUN_MERGE 1
274#define KSM_RUN_UNMERGE 2
275#define KSM_RUN_OFFLINE 4
276static unsigned long ksm_run = KSM_RUN_STOP;
277static void wait_while_offlining(void);
278
279static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
280static DEFINE_MUTEX(ksm_thread_mutex);
281static DEFINE_SPINLOCK(ksm_mmlist_lock);
282
283#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
284 sizeof(struct __struct), __alignof__(struct __struct),\
285 (__flags), NULL)
286
287static int __init ksm_slab_init(void)
288{
289 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
290 if (!rmap_item_cache)
291 goto out;
292
293 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
294 if (!stable_node_cache)
295 goto out_free1;
296
297 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
298 if (!mm_slot_cache)
299 goto out_free2;
300
301 return 0;
302
303out_free2:
304 kmem_cache_destroy(stable_node_cache);
305out_free1:
306 kmem_cache_destroy(rmap_item_cache);
307out:
308 return -ENOMEM;
309}
310
311static void __init ksm_slab_free(void)
312{
313 kmem_cache_destroy(mm_slot_cache);
314 kmem_cache_destroy(stable_node_cache);
315 kmem_cache_destroy(rmap_item_cache);
316 mm_slot_cache = NULL;
317}
318
319static __always_inline bool is_stable_node_chain(struct stable_node *chain)
320{
321 return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
322}
323
324static __always_inline bool is_stable_node_dup(struct stable_node *dup)
325{
326 return dup->head == STABLE_NODE_DUP_HEAD;
327}
328
329static inline void stable_node_chain_add_dup(struct stable_node *dup,
330 struct stable_node *chain)
331{
332 VM_BUG_ON(is_stable_node_dup(dup));
333 dup->head = STABLE_NODE_DUP_HEAD;
334 VM_BUG_ON(!is_stable_node_chain(chain));
335 hlist_add_head(&dup->hlist_dup, &chain->hlist);
336 ksm_stable_node_dups++;
337}
338
339static inline void __stable_node_dup_del(struct stable_node *dup)
340{
341 VM_BUG_ON(!is_stable_node_dup(dup));
342 hlist_del(&dup->hlist_dup);
343 ksm_stable_node_dups--;
344}
345
346static inline void stable_node_dup_del(struct stable_node *dup)
347{
348 VM_BUG_ON(is_stable_node_chain(dup));
349 if (is_stable_node_dup(dup))
350 __stable_node_dup_del(dup);
351 else
352 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
353#ifdef CONFIG_DEBUG_VM
354 dup->head = NULL;
355#endif
356}
357
358static inline struct rmap_item *alloc_rmap_item(void)
359{
360 struct rmap_item *rmap_item;
361
362 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
363 __GFP_NORETRY | __GFP_NOWARN);
364 if (rmap_item)
365 ksm_rmap_items++;
366 return rmap_item;
367}
368
369static inline void free_rmap_item(struct rmap_item *rmap_item)
370{
371 ksm_rmap_items--;
372 rmap_item->mm = NULL; /* debug safety */
373 kmem_cache_free(rmap_item_cache, rmap_item);
374}
375
376static inline struct stable_node *alloc_stable_node(void)
377{
378 /*
379 * The allocation can take too long with GFP_KERNEL when memory is under
380 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
381 * grants access to memory reserves, helping to avoid this problem.
382 */
383 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
384}
385
386static inline void free_stable_node(struct stable_node *stable_node)
387{
388 VM_BUG_ON(stable_node->rmap_hlist_len &&
389 !is_stable_node_chain(stable_node));
390 kmem_cache_free(stable_node_cache, stable_node);
391}
392
393static inline struct mm_slot *alloc_mm_slot(void)
394{
395 if (!mm_slot_cache) /* initialization failed */
396 return NULL;
397 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
398}
399
400static inline void free_mm_slot(struct mm_slot *mm_slot)
401{
402 kmem_cache_free(mm_slot_cache, mm_slot);
403}
404
405static struct mm_slot *get_mm_slot(struct mm_struct *mm)
406{
407 struct mm_slot *slot;
408
409 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
410 if (slot->mm == mm)
411 return slot;
412
413 return NULL;
414}
415
416static void insert_to_mm_slots_hash(struct mm_struct *mm,
417 struct mm_slot *mm_slot)
418{
419 mm_slot->mm = mm;
420 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
421}
422
423/*
424 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
425 * page tables after it has passed through ksm_exit() - which, if necessary,
426 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
427 * a special flag: they can just back out as soon as mm_users goes to zero.
428 * ksm_test_exit() is used throughout to make this test for exit: in some
429 * places for correctness, in some places just to avoid unnecessary work.
430 */
431static inline bool ksm_test_exit(struct mm_struct *mm)
432{
433 return atomic_read(&mm->mm_users) == 0;
434}
435
436/*
437 * We use break_ksm to break COW on a ksm page: it's a stripped down
438 *
439 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
440 * put_page(page);
441 *
442 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
443 * in case the application has unmapped and remapped mm,addr meanwhile.
444 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
445 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
446 *
447 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
448 * of the process that owns 'vma'. We also do not want to enforce
449 * protection keys here anyway.
450 */
451static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
452{
453 struct page *page;
454 int ret = 0;
455
456 do {
457 cond_resched();
458 page = follow_page(vma, addr,
459 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
460 if (IS_ERR_OR_NULL(page))
461 break;
462 if (PageKsm(page))
463 ret = handle_mm_fault(vma, addr,
464 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
465 else
466 ret = VM_FAULT_WRITE;
467 put_page(page);
468 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
469 /*
470 * We must loop because handle_mm_fault() may back out if there's
471 * any difficulty e.g. if pte accessed bit gets updated concurrently.
472 *
473 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
474 * COW has been broken, even if the vma does not permit VM_WRITE;
475 * but note that a concurrent fault might break PageKsm for us.
476 *
477 * VM_FAULT_SIGBUS could occur if we race with truncation of the
478 * backing file, which also invalidates anonymous pages: that's
479 * okay, that truncation will have unmapped the PageKsm for us.
480 *
481 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
482 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
483 * current task has TIF_MEMDIE set, and will be OOM killed on return
484 * to user; and ksmd, having no mm, would never be chosen for that.
485 *
486 * But if the mm is in a limited mem_cgroup, then the fault may fail
487 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
488 * even ksmd can fail in this way - though it's usually breaking ksm
489 * just to undo a merge it made a moment before, so unlikely to oom.
490 *
491 * That's a pity: we might therefore have more kernel pages allocated
492 * than we're counting as nodes in the stable tree; but ksm_do_scan
493 * will retry to break_cow on each pass, so should recover the page
494 * in due course. The important thing is to not let VM_MERGEABLE
495 * be cleared while any such pages might remain in the area.
496 */
497 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
498}
499
500static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
501 unsigned long addr)
502{
503 struct vm_area_struct *vma;
504 if (ksm_test_exit(mm))
505 return NULL;
506 vma = find_vma(mm, addr);
507 if (!vma || vma->vm_start > addr)
508 return NULL;
509 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
510 return NULL;
511 return vma;
512}
513
514static void break_cow(struct rmap_item *rmap_item)
515{
516 struct mm_struct *mm = rmap_item->mm;
517 unsigned long addr = rmap_item->address;
518 struct vm_area_struct *vma;
519
520 /*
521 * It is not an accident that whenever we want to break COW
522 * to undo, we also need to drop a reference to the anon_vma.
523 */
524 put_anon_vma(rmap_item->anon_vma);
525
526 down_read(&mm->mmap_sem);
527 vma = find_mergeable_vma(mm, addr);
528 if (vma)
529 break_ksm(vma, addr);
530 up_read(&mm->mmap_sem);
531}
532
533static struct page *get_mergeable_page(struct rmap_item *rmap_item)
534{
535 struct mm_struct *mm = rmap_item->mm;
536 unsigned long addr = rmap_item->address;
537 struct vm_area_struct *vma;
538 struct page *page;
539
540 down_read(&mm->mmap_sem);
541 vma = find_mergeable_vma(mm, addr);
542 if (!vma)
543 goto out;
544
545 page = follow_page(vma, addr, FOLL_GET);
546 if (IS_ERR_OR_NULL(page))
547 goto out;
548 if (PageAnon(page)) {
549 flush_anon_page(vma, page, addr);
550 flush_dcache_page(page);
551 } else {
552 put_page(page);
553out:
554 page = NULL;
555 }
556 up_read(&mm->mmap_sem);
557 return page;
558}
559
560/*
561 * This helper is used for getting right index into array of tree roots.
562 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
563 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
564 * every node has its own stable and unstable tree.
565 */
566static inline int get_kpfn_nid(unsigned long kpfn)
567{
568 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
569}
570
571static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
572 struct rb_root *root)
573{
574 struct stable_node *chain = alloc_stable_node();
575 VM_BUG_ON(is_stable_node_chain(dup));
576 if (likely(chain)) {
577 INIT_HLIST_HEAD(&chain->hlist);
578 chain->chain_prune_time = jiffies;
579 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
580#if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
581 chain->nid = -1; /* debug */
582#endif
583 ksm_stable_node_chains++;
584
585 /*
586 * Put the stable node chain in the first dimension of
587 * the stable tree and at the same time remove the old
588 * stable node.
589 */
590 rb_replace_node(&dup->node, &chain->node, root);
591
592 /*
593 * Move the old stable node to the second dimension
594 * queued in the hlist_dup. The invariant is that all
595 * dup stable_nodes in the chain->hlist point to pages
596 * that are wrprotected and have the exact same
597 * content.
598 */
599 stable_node_chain_add_dup(dup, chain);
600 }
601 return chain;
602}
603
604static inline void free_stable_node_chain(struct stable_node *chain,
605 struct rb_root *root)
606{
607 rb_erase(&chain->node, root);
608 free_stable_node(chain);
609 ksm_stable_node_chains--;
610}
611
612static void remove_node_from_stable_tree(struct stable_node *stable_node)
613{
614 struct rmap_item *rmap_item;
615
616 /* check it's not STABLE_NODE_CHAIN or negative */
617 BUG_ON(stable_node->rmap_hlist_len < 0);
618
619 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
620 if (rmap_item->hlist.next)
621 ksm_pages_sharing--;
622 else
623 ksm_pages_shared--;
624 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
625 stable_node->rmap_hlist_len--;
626 put_anon_vma(rmap_item->anon_vma);
627 rmap_item->address &= PAGE_MASK;
628 cond_resched();
629 }
630
631 /*
632 * We need the second aligned pointer of the migrate_nodes
633 * list_head to stay clear from the rb_parent_color union
634 * (aligned and different than any node) and also different
635 * from &migrate_nodes. This will verify that future list.h changes
636 * don't break STABLE_NODE_DUP_HEAD.
637 */
638#if GCC_VERSION >= 40903 /* only recent gcc can handle it */
639 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
640 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
641#endif
642
643 if (stable_node->head == &migrate_nodes)
644 list_del(&stable_node->list);
645 else
646 stable_node_dup_del(stable_node);
647 free_stable_node(stable_node);
648}
649
650/*
651 * get_ksm_page: checks if the page indicated by the stable node
652 * is still its ksm page, despite having held no reference to it.
653 * In which case we can trust the content of the page, and it
654 * returns the gotten page; but if the page has now been zapped,
655 * remove the stale node from the stable tree and return NULL.
656 * But beware, the stable node's page might be being migrated.
657 *
658 * You would expect the stable_node to hold a reference to the ksm page.
659 * But if it increments the page's count, swapping out has to wait for
660 * ksmd to come around again before it can free the page, which may take
661 * seconds or even minutes: much too unresponsive. So instead we use a
662 * "keyhole reference": access to the ksm page from the stable node peeps
663 * out through its keyhole to see if that page still holds the right key,
664 * pointing back to this stable node. This relies on freeing a PageAnon
665 * page to reset its page->mapping to NULL, and relies on no other use of
666 * a page to put something that might look like our key in page->mapping.
667 * is on its way to being freed; but it is an anomaly to bear in mind.
668 */
669static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
670{
671 struct page *page;
672 void *expected_mapping;
673 unsigned long kpfn;
674
675 expected_mapping = (void *)((unsigned long)stable_node |
676 PAGE_MAPPING_KSM);
677again:
678 kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
679 page = pfn_to_page(kpfn);
680 if (READ_ONCE(page->mapping) != expected_mapping)
681 goto stale;
682
683 /*
684 * We cannot do anything with the page while its refcount is 0.
685 * Usually 0 means free, or tail of a higher-order page: in which
686 * case this node is no longer referenced, and should be freed;
687 * however, it might mean that the page is under page_freeze_refs().
688 * The __remove_mapping() case is easy, again the node is now stale;
689 * but if page is swapcache in migrate_page_move_mapping(), it might
690 * still be our page, in which case it's essential to keep the node.
691 */
692 while (!get_page_unless_zero(page)) {
693 /*
694 * Another check for page->mapping != expected_mapping would
695 * work here too. We have chosen the !PageSwapCache test to
696 * optimize the common case, when the page is or is about to
697 * be freed: PageSwapCache is cleared (under spin_lock_irq)
698 * in the freeze_refs section of __remove_mapping(); but Anon
699 * page->mapping reset to NULL later, in free_pages_prepare().
700 */
701 if (!PageSwapCache(page))
702 goto stale;
703 cpu_relax();
704 }
705
706 if (READ_ONCE(page->mapping) != expected_mapping) {
707 put_page(page);
708 goto stale;
709 }
710
711 if (lock_it) {
712 lock_page(page);
713 if (READ_ONCE(page->mapping) != expected_mapping) {
714 unlock_page(page);
715 put_page(page);
716 goto stale;
717 }
718 }
719 return page;
720
721stale:
722 /*
723 * We come here from above when page->mapping or !PageSwapCache
724 * suggests that the node is stale; but it might be under migration.
725 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
726 * before checking whether node->kpfn has been changed.
727 */
728 smp_rmb();
729 if (READ_ONCE(stable_node->kpfn) != kpfn)
730 goto again;
731 remove_node_from_stable_tree(stable_node);
732 return NULL;
733}
734
735/*
736 * Removing rmap_item from stable or unstable tree.
737 * This function will clean the information from the stable/unstable tree.
738 */
739static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
740{
741 if (rmap_item->address & STABLE_FLAG) {
742 struct stable_node *stable_node;
743 struct page *page;
744
745 stable_node = rmap_item->head;
746 page = get_ksm_page(stable_node, true);
747 if (!page)
748 goto out;
749
750 hlist_del(&rmap_item->hlist);
751 unlock_page(page);
752 put_page(page);
753
754 if (!hlist_empty(&stable_node->hlist))
755 ksm_pages_sharing--;
756 else
757 ksm_pages_shared--;
758 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
759 stable_node->rmap_hlist_len--;
760
761 put_anon_vma(rmap_item->anon_vma);
762 rmap_item->address &= PAGE_MASK;
763
764 } else if (rmap_item->address & UNSTABLE_FLAG) {
765 unsigned char age;
766 /*
767 * Usually ksmd can and must skip the rb_erase, because
768 * root_unstable_tree was already reset to RB_ROOT.
769 * But be careful when an mm is exiting: do the rb_erase
770 * if this rmap_item was inserted by this scan, rather
771 * than left over from before.
772 */
773 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
774 BUG_ON(age > 1);
775 if (!age)
776 rb_erase(&rmap_item->node,
777 root_unstable_tree + NUMA(rmap_item->nid));
778 ksm_pages_unshared--;
779 rmap_item->address &= PAGE_MASK;
780 }
781out:
782 cond_resched(); /* we're called from many long loops */
783}
784
785static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
786 struct rmap_item **rmap_list)
787{
788 while (*rmap_list) {
789 struct rmap_item *rmap_item = *rmap_list;
790 *rmap_list = rmap_item->rmap_list;
791 remove_rmap_item_from_tree(rmap_item);
792 free_rmap_item(rmap_item);
793 }
794}
795
796/*
797 * Though it's very tempting to unmerge rmap_items from stable tree rather
798 * than check every pte of a given vma, the locking doesn't quite work for
799 * that - an rmap_item is assigned to the stable tree after inserting ksm
800 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
801 * rmap_items from parent to child at fork time (so as not to waste time
802 * if exit comes before the next scan reaches it).
803 *
804 * Similarly, although we'd like to remove rmap_items (so updating counts
805 * and freeing memory) when unmerging an area, it's easier to leave that
806 * to the next pass of ksmd - consider, for example, how ksmd might be
807 * in cmp_and_merge_page on one of the rmap_items we would be removing.
808 */
809static int unmerge_ksm_pages(struct vm_area_struct *vma,
810 unsigned long start, unsigned long end)
811{
812 unsigned long addr;
813 int err = 0;
814
815 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
816 if (ksm_test_exit(vma->vm_mm))
817 break;
818 if (signal_pending(current))
819 err = -ERESTARTSYS;
820 else
821 err = break_ksm(vma, addr);
822 }
823 return err;
824}
825
826#ifdef CONFIG_SYSFS
827/*
828 * Only called through the sysfs control interface:
829 */
830static int remove_stable_node(struct stable_node *stable_node)
831{
832 struct page *page;
833 int err;
834
835 page = get_ksm_page(stable_node, true);
836 if (!page) {
837 /*
838 * get_ksm_page did remove_node_from_stable_tree itself.
839 */
840 return 0;
841 }
842
843 if (WARN_ON_ONCE(page_mapped(page))) {
844 /*
845 * This should not happen: but if it does, just refuse to let
846 * merge_across_nodes be switched - there is no need to panic.
847 */
848 err = -EBUSY;
849 } else {
850 /*
851 * The stable node did not yet appear stale to get_ksm_page(),
852 * since that allows for an unmapped ksm page to be recognized
853 * right up until it is freed; but the node is safe to remove.
854 * This page might be in a pagevec waiting to be freed,
855 * or it might be PageSwapCache (perhaps under writeback),
856 * or it might have been removed from swapcache a moment ago.
857 */
858 set_page_stable_node(page, NULL);
859 remove_node_from_stable_tree(stable_node);
860 err = 0;
861 }
862
863 unlock_page(page);
864 put_page(page);
865 return err;
866}
867
868static int remove_stable_node_chain(struct stable_node *stable_node,
869 struct rb_root *root)
870{
871 struct stable_node *dup;
872 struct hlist_node *hlist_safe;
873
874 if (!is_stable_node_chain(stable_node)) {
875 VM_BUG_ON(is_stable_node_dup(stable_node));
876 if (remove_stable_node(stable_node))
877 return true;
878 else
879 return false;
880 }
881
882 hlist_for_each_entry_safe(dup, hlist_safe,
883 &stable_node->hlist, hlist_dup) {
884 VM_BUG_ON(!is_stable_node_dup(dup));
885 if (remove_stable_node(dup))
886 return true;
887 }
888 BUG_ON(!hlist_empty(&stable_node->hlist));
889 free_stable_node_chain(stable_node, root);
890 return false;
891}
892
893static int remove_all_stable_nodes(void)
894{
895 struct stable_node *stable_node, *next;
896 int nid;
897 int err = 0;
898
899 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
900 while (root_stable_tree[nid].rb_node) {
901 stable_node = rb_entry(root_stable_tree[nid].rb_node,
902 struct stable_node, node);
903 if (remove_stable_node_chain(stable_node,
904 root_stable_tree + nid)) {
905 err = -EBUSY;
906 break; /* proceed to next nid */
907 }
908 cond_resched();
909 }
910 }
911 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
912 if (remove_stable_node(stable_node))
913 err = -EBUSY;
914 cond_resched();
915 }
916 return err;
917}
918
919static int unmerge_and_remove_all_rmap_items(void)
920{
921 struct mm_slot *mm_slot;
922 struct mm_struct *mm;
923 struct vm_area_struct *vma;
924 int err = 0;
925
926 spin_lock(&ksm_mmlist_lock);
927 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
928 struct mm_slot, mm_list);
929 spin_unlock(&ksm_mmlist_lock);
930
931 for (mm_slot = ksm_scan.mm_slot;
932 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
933 mm = mm_slot->mm;
934 down_read(&mm->mmap_sem);
935 for (vma = mm->mmap; vma; vma = vma->vm_next) {
936 if (ksm_test_exit(mm))
937 break;
938 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
939 continue;
940 err = unmerge_ksm_pages(vma,
941 vma->vm_start, vma->vm_end);
942 if (err)
943 goto error;
944 }
945
946 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
947 up_read(&mm->mmap_sem);
948
949 spin_lock(&ksm_mmlist_lock);
950 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
951 struct mm_slot, mm_list);
952 if (ksm_test_exit(mm)) {
953 hash_del(&mm_slot->link);
954 list_del(&mm_slot->mm_list);
955 spin_unlock(&ksm_mmlist_lock);
956
957 free_mm_slot(mm_slot);
958 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
959 mmdrop(mm);
960 } else
961 spin_unlock(&ksm_mmlist_lock);
962 }
963
964 /* Clean up stable nodes, but don't worry if some are still busy */
965 remove_all_stable_nodes();
966 ksm_scan.seqnr = 0;
967 return 0;
968
969error:
970 up_read(&mm->mmap_sem);
971 spin_lock(&ksm_mmlist_lock);
972 ksm_scan.mm_slot = &ksm_mm_head;
973 spin_unlock(&ksm_mmlist_lock);
974 return err;
975}
976#endif /* CONFIG_SYSFS */
977
978static u32 calc_checksum(struct page *page)
979{
980 u32 checksum;
981 void *addr = kmap_atomic(page);
982 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
983 kunmap_atomic(addr);
984 return checksum;
985}
986
987static int memcmp_pages(struct page *page1, struct page *page2)
988{
989 char *addr1, *addr2;
990 int ret;
991
992 addr1 = kmap_atomic(page1);
993 addr2 = kmap_atomic(page2);
994 ret = memcmp(addr1, addr2, PAGE_SIZE);
995 kunmap_atomic(addr2);
996 kunmap_atomic(addr1);
997 return ret;
998}
999
1000static inline int pages_identical(struct page *page1, struct page *page2)
1001{
1002 return !memcmp_pages(page1, page2);
1003}
1004
1005static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1006 pte_t *orig_pte)
1007{
1008 struct mm_struct *mm = vma->vm_mm;
1009 struct page_vma_mapped_walk pvmw = {
1010 .page = page,
1011 .vma = vma,
1012 };
1013 int swapped;
1014 int err = -EFAULT;
1015 unsigned long mmun_start; /* For mmu_notifiers */
1016 unsigned long mmun_end; /* For mmu_notifiers */
1017
1018 pvmw.address = page_address_in_vma(page, vma);
1019 if (pvmw.address == -EFAULT)
1020 goto out;
1021
1022 BUG_ON(PageTransCompound(page));
1023
1024 mmun_start = pvmw.address;
1025 mmun_end = pvmw.address + PAGE_SIZE;
1026 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1027
1028 if (!page_vma_mapped_walk(&pvmw))
1029 goto out_mn;
1030 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1031 goto out_unlock;
1032
1033 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1034 (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1035 mm_tlb_flush_pending(mm)) {
1036 pte_t entry;
1037
1038 swapped = PageSwapCache(page);
1039 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1040 /*
1041 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1042 * take any lock, therefore the check that we are going to make
1043 * with the pagecount against the mapcount is racey and
1044 * O_DIRECT can happen right after the check.
1045 * So we clear the pte and flush the tlb before the check
1046 * this assure us that no O_DIRECT can happen after the check
1047 * or in the middle of the check.
1048 *
1049 * No need to notify as we are downgrading page table to read
1050 * only not changing it to point to a new page.
1051 *
1052 * See Documentation/vm/mmu_notifier.txt
1053 */
1054 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1055 /*
1056 * Check that no O_DIRECT or similar I/O is in progress on the
1057 * page
1058 */
1059 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1060 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1061 goto out_unlock;
1062 }
1063 if (pte_dirty(entry))
1064 set_page_dirty(page);
1065
1066 if (pte_protnone(entry))
1067 entry = pte_mkclean(pte_clear_savedwrite(entry));
1068 else
1069 entry = pte_mkclean(pte_wrprotect(entry));
1070 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1071 }
1072 *orig_pte = *pvmw.pte;
1073 err = 0;
1074
1075out_unlock:
1076 page_vma_mapped_walk_done(&pvmw);
1077out_mn:
1078 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1079out:
1080 return err;
1081}
1082
1083/**
1084 * replace_page - replace page in vma by new ksm page
1085 * @vma: vma that holds the pte pointing to page
1086 * @page: the page we are replacing by kpage
1087 * @kpage: the ksm page we replace page by
1088 * @orig_pte: the original value of the pte
1089 *
1090 * Returns 0 on success, -EFAULT on failure.
1091 */
1092static int replace_page(struct vm_area_struct *vma, struct page *page,
1093 struct page *kpage, pte_t orig_pte)
1094{
1095 struct mm_struct *mm = vma->vm_mm;
1096 pmd_t *pmd;
1097 pte_t *ptep;
1098 pte_t newpte;
1099 spinlock_t *ptl;
1100 unsigned long addr;
1101 int err = -EFAULT;
1102 unsigned long mmun_start; /* For mmu_notifiers */
1103 unsigned long mmun_end; /* For mmu_notifiers */
1104
1105 addr = page_address_in_vma(page, vma);
1106 if (addr == -EFAULT)
1107 goto out;
1108
1109 pmd = mm_find_pmd(mm, addr);
1110 if (!pmd)
1111 goto out;
1112
1113 mmun_start = addr;
1114 mmun_end = addr + PAGE_SIZE;
1115 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1116
1117 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1118 if (!pte_same(*ptep, orig_pte)) {
1119 pte_unmap_unlock(ptep, ptl);
1120 goto out_mn;
1121 }
1122
1123 /*
1124 * No need to check ksm_use_zero_pages here: we can only have a
1125 * zero_page here if ksm_use_zero_pages was enabled alreaady.
1126 */
1127 if (!is_zero_pfn(page_to_pfn(kpage))) {
1128 get_page(kpage);
1129 page_add_anon_rmap(kpage, vma, addr, false);
1130 newpte = mk_pte(kpage, vma->vm_page_prot);
1131 } else {
1132 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1133 vma->vm_page_prot));
1134 /*
1135 * We're replacing an anonymous page with a zero page, which is
1136 * not anonymous. We need to do proper accounting otherwise we
1137 * will get wrong values in /proc, and a BUG message in dmesg
1138 * when tearing down the mm.
1139 */
1140 dec_mm_counter(mm, MM_ANONPAGES);
1141 }
1142
1143 flush_cache_page(vma, addr, pte_pfn(*ptep));
1144 /*
1145 * No need to notify as we are replacing a read only page with another
1146 * read only page with the same content.
1147 *
1148 * See Documentation/vm/mmu_notifier.txt
1149 */
1150 ptep_clear_flush(vma, addr, ptep);
1151 set_pte_at_notify(mm, addr, ptep, newpte);
1152
1153 page_remove_rmap(page, false);
1154 if (!page_mapped(page))
1155 try_to_free_swap(page);
1156 put_page(page);
1157
1158 pte_unmap_unlock(ptep, ptl);
1159 err = 0;
1160out_mn:
1161 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1162out:
1163 return err;
1164}
1165
1166/*
1167 * try_to_merge_one_page - take two pages and merge them into one
1168 * @vma: the vma that holds the pte pointing to page
1169 * @page: the PageAnon page that we want to replace with kpage
1170 * @kpage: the PageKsm page that we want to map instead of page,
1171 * or NULL the first time when we want to use page as kpage.
1172 *
1173 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1174 */
1175static int try_to_merge_one_page(struct vm_area_struct *vma,
1176 struct page *page, struct page *kpage)
1177{
1178 pte_t orig_pte = __pte(0);
1179 int err = -EFAULT;
1180
1181 if (page == kpage) /* ksm page forked */
1182 return 0;
1183
1184 if (!PageAnon(page))
1185 goto out;
1186
1187 /*
1188 * We need the page lock to read a stable PageSwapCache in
1189 * write_protect_page(). We use trylock_page() instead of
1190 * lock_page() because we don't want to wait here - we
1191 * prefer to continue scanning and merging different pages,
1192 * then come back to this page when it is unlocked.
1193 */
1194 if (!trylock_page(page))
1195 goto out;
1196
1197 if (PageTransCompound(page)) {
1198 if (split_huge_page(page))
1199 goto out_unlock;
1200 }
1201
1202 /*
1203 * If this anonymous page is mapped only here, its pte may need
1204 * to be write-protected. If it's mapped elsewhere, all of its
1205 * ptes are necessarily already write-protected. But in either
1206 * case, we need to lock and check page_count is not raised.
1207 */
1208 if (write_protect_page(vma, page, &orig_pte) == 0) {
1209 if (!kpage) {
1210 /*
1211 * While we hold page lock, upgrade page from
1212 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1213 * stable_tree_insert() will update stable_node.
1214 */
1215 set_page_stable_node(page, NULL);
1216 mark_page_accessed(page);
1217 /*
1218 * Page reclaim just frees a clean page with no dirty
1219 * ptes: make sure that the ksm page would be swapped.
1220 */
1221 if (!PageDirty(page))
1222 SetPageDirty(page);
1223 err = 0;
1224 } else if (pages_identical(page, kpage))
1225 err = replace_page(vma, page, kpage, orig_pte);
1226 }
1227
1228 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1229 munlock_vma_page(page);
1230 if (!PageMlocked(kpage)) {
1231 unlock_page(page);
1232 lock_page(kpage);
1233 mlock_vma_page(kpage);
1234 page = kpage; /* for final unlock */
1235 }
1236 }
1237
1238out_unlock:
1239 unlock_page(page);
1240out:
1241 return err;
1242}
1243
1244/*
1245 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1246 * but no new kernel page is allocated: kpage must already be a ksm page.
1247 *
1248 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1249 */
1250static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1251 struct page *page, struct page *kpage)
1252{
1253 struct mm_struct *mm = rmap_item->mm;
1254 struct vm_area_struct *vma;
1255 int err = -EFAULT;
1256
1257 down_read(&mm->mmap_sem);
1258 vma = find_mergeable_vma(mm, rmap_item->address);
1259 if (!vma)
1260 goto out;
1261
1262 err = try_to_merge_one_page(vma, page, kpage);
1263 if (err)
1264 goto out;
1265
1266 /* Unstable nid is in union with stable anon_vma: remove first */
1267 remove_rmap_item_from_tree(rmap_item);
1268
1269 /* Must get reference to anon_vma while still holding mmap_sem */
1270 rmap_item->anon_vma = vma->anon_vma;
1271 get_anon_vma(vma->anon_vma);
1272out:
1273 up_read(&mm->mmap_sem);
1274 return err;
1275}
1276
1277/*
1278 * try_to_merge_two_pages - take two identical pages and prepare them
1279 * to be merged into one page.
1280 *
1281 * This function returns the kpage if we successfully merged two identical
1282 * pages into one ksm page, NULL otherwise.
1283 *
1284 * Note that this function upgrades page to ksm page: if one of the pages
1285 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1286 */
1287static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1288 struct page *page,
1289 struct rmap_item *tree_rmap_item,
1290 struct page *tree_page)
1291{
1292 int err;
1293
1294 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1295 if (!err) {
1296 err = try_to_merge_with_ksm_page(tree_rmap_item,
1297 tree_page, page);
1298 /*
1299 * If that fails, we have a ksm page with only one pte
1300 * pointing to it: so break it.
1301 */
1302 if (err)
1303 break_cow(rmap_item);
1304 }
1305 return err ? NULL : page;
1306}
1307
1308static __always_inline
1309bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1310{
1311 VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1312 /*
1313 * Check that at least one mapping still exists, otherwise
1314 * there's no much point to merge and share with this
1315 * stable_node, as the underlying tree_page of the other
1316 * sharer is going to be freed soon.
1317 */
1318 return stable_node->rmap_hlist_len &&
1319 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1320}
1321
1322static __always_inline
1323bool is_page_sharing_candidate(struct stable_node *stable_node)
1324{
1325 return __is_page_sharing_candidate(stable_node, 0);
1326}
1327
1328static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1329 struct stable_node **_stable_node,
1330 struct rb_root *root,
1331 bool prune_stale_stable_nodes)
1332{
1333 struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1334 struct hlist_node *hlist_safe;
1335 struct page *_tree_page, *tree_page = NULL;
1336 int nr = 0;
1337 int found_rmap_hlist_len;
1338
1339 if (!prune_stale_stable_nodes ||
1340 time_before(jiffies, stable_node->chain_prune_time +
1341 msecs_to_jiffies(
1342 ksm_stable_node_chains_prune_millisecs)))
1343 prune_stale_stable_nodes = false;
1344 else
1345 stable_node->chain_prune_time = jiffies;
1346
1347 hlist_for_each_entry_safe(dup, hlist_safe,
1348 &stable_node->hlist, hlist_dup) {
1349 cond_resched();
1350 /*
1351 * We must walk all stable_node_dup to prune the stale
1352 * stable nodes during lookup.
1353 *
1354 * get_ksm_page can drop the nodes from the
1355 * stable_node->hlist if they point to freed pages
1356 * (that's why we do a _safe walk). The "dup"
1357 * stable_node parameter itself will be freed from
1358 * under us if it returns NULL.
1359 */
1360 _tree_page = get_ksm_page(dup, false);
1361 if (!_tree_page)
1362 continue;
1363 nr += 1;
1364 if (is_page_sharing_candidate(dup)) {
1365 if (!found ||
1366 dup->rmap_hlist_len > found_rmap_hlist_len) {
1367 if (found)
1368 put_page(tree_page);
1369 found = dup;
1370 found_rmap_hlist_len = found->rmap_hlist_len;
1371 tree_page = _tree_page;
1372
1373 /* skip put_page for found dup */
1374 if (!prune_stale_stable_nodes)
1375 break;
1376 continue;
1377 }
1378 }
1379 put_page(_tree_page);
1380 }
1381
1382 if (found) {
1383 /*
1384 * nr is counting all dups in the chain only if
1385 * prune_stale_stable_nodes is true, otherwise we may
1386 * break the loop at nr == 1 even if there are
1387 * multiple entries.
1388 */
1389 if (prune_stale_stable_nodes && nr == 1) {
1390 /*
1391 * If there's not just one entry it would
1392 * corrupt memory, better BUG_ON. In KSM
1393 * context with no lock held it's not even
1394 * fatal.
1395 */
1396 BUG_ON(stable_node->hlist.first->next);
1397
1398 /*
1399 * There's just one entry and it is below the
1400 * deduplication limit so drop the chain.
1401 */
1402 rb_replace_node(&stable_node->node, &found->node,
1403 root);
1404 free_stable_node(stable_node);
1405 ksm_stable_node_chains--;
1406 ksm_stable_node_dups--;
1407 /*
1408 * NOTE: the caller depends on the stable_node
1409 * to be equal to stable_node_dup if the chain
1410 * was collapsed.
1411 */
1412 *_stable_node = found;
1413 /*
1414 * Just for robustneess as stable_node is
1415 * otherwise left as a stable pointer, the
1416 * compiler shall optimize it away at build
1417 * time.
1418 */
1419 stable_node = NULL;
1420 } else if (stable_node->hlist.first != &found->hlist_dup &&
1421 __is_page_sharing_candidate(found, 1)) {
1422 /*
1423 * If the found stable_node dup can accept one
1424 * more future merge (in addition to the one
1425 * that is underway) and is not at the head of
1426 * the chain, put it there so next search will
1427 * be quicker in the !prune_stale_stable_nodes
1428 * case.
1429 *
1430 * NOTE: it would be inaccurate to use nr > 1
1431 * instead of checking the hlist.first pointer
1432 * directly, because in the
1433 * prune_stale_stable_nodes case "nr" isn't
1434 * the position of the found dup in the chain,
1435 * but the total number of dups in the chain.
1436 */
1437 hlist_del(&found->hlist_dup);
1438 hlist_add_head(&found->hlist_dup,
1439 &stable_node->hlist);
1440 }
1441 }
1442
1443 *_stable_node_dup = found;
1444 return tree_page;
1445}
1446
1447static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1448 struct rb_root *root)
1449{
1450 if (!is_stable_node_chain(stable_node))
1451 return stable_node;
1452 if (hlist_empty(&stable_node->hlist)) {
1453 free_stable_node_chain(stable_node, root);
1454 return NULL;
1455 }
1456 return hlist_entry(stable_node->hlist.first,
1457 typeof(*stable_node), hlist_dup);
1458}
1459
1460/*
1461 * Like for get_ksm_page, this function can free the *_stable_node and
1462 * *_stable_node_dup if the returned tree_page is NULL.
1463 *
1464 * It can also free and overwrite *_stable_node with the found
1465 * stable_node_dup if the chain is collapsed (in which case
1466 * *_stable_node will be equal to *_stable_node_dup like if the chain
1467 * never existed). It's up to the caller to verify tree_page is not
1468 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1469 *
1470 * *_stable_node_dup is really a second output parameter of this
1471 * function and will be overwritten in all cases, the caller doesn't
1472 * need to initialize it.
1473 */
1474static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1475 struct stable_node **_stable_node,
1476 struct rb_root *root,
1477 bool prune_stale_stable_nodes)
1478{
1479 struct stable_node *stable_node = *_stable_node;
1480 if (!is_stable_node_chain(stable_node)) {
1481 if (is_page_sharing_candidate(stable_node)) {
1482 *_stable_node_dup = stable_node;
1483 return get_ksm_page(stable_node, false);
1484 }
1485 /*
1486 * _stable_node_dup set to NULL means the stable_node
1487 * reached the ksm_max_page_sharing limit.
1488 */
1489 *_stable_node_dup = NULL;
1490 return NULL;
1491 }
1492 return stable_node_dup(_stable_node_dup, _stable_node, root,
1493 prune_stale_stable_nodes);
1494}
1495
1496static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1497 struct stable_node **s_n,
1498 struct rb_root *root)
1499{
1500 return __stable_node_chain(s_n_d, s_n, root, true);
1501}
1502
1503static __always_inline struct page *chain(struct stable_node **s_n_d,
1504 struct stable_node *s_n,
1505 struct rb_root *root)
1506{
1507 struct stable_node *old_stable_node = s_n;
1508 struct page *tree_page;
1509
1510 tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1511 /* not pruning dups so s_n cannot have changed */
1512 VM_BUG_ON(s_n != old_stable_node);
1513 return tree_page;
1514}
1515
1516/*
1517 * stable_tree_search - search for page inside the stable tree
1518 *
1519 * This function checks if there is a page inside the stable tree
1520 * with identical content to the page that we are scanning right now.
1521 *
1522 * This function returns the stable tree node of identical content if found,
1523 * NULL otherwise.
1524 */
1525static struct page *stable_tree_search(struct page *page)
1526{
1527 int nid;
1528 struct rb_root *root;
1529 struct rb_node **new;
1530 struct rb_node *parent;
1531 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1532 struct stable_node *page_node;
1533
1534 page_node = page_stable_node(page);
1535 if (page_node && page_node->head != &migrate_nodes) {
1536 /* ksm page forked */
1537 get_page(page);
1538 return page;
1539 }
1540
1541 nid = get_kpfn_nid(page_to_pfn(page));
1542 root = root_stable_tree + nid;
1543again:
1544 new = &root->rb_node;
1545 parent = NULL;
1546
1547 while (*new) {
1548 struct page *tree_page;
1549 int ret;
1550
1551 cond_resched();
1552 stable_node = rb_entry(*new, struct stable_node, node);
1553 stable_node_any = NULL;
1554 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1555 /*
1556 * NOTE: stable_node may have been freed by
1557 * chain_prune() if the returned stable_node_dup is
1558 * not NULL. stable_node_dup may have been inserted in
1559 * the rbtree instead as a regular stable_node (in
1560 * order to collapse the stable_node chain if a single
1561 * stable_node dup was found in it). In such case the
1562 * stable_node is overwritten by the calleee to point
1563 * to the stable_node_dup that was collapsed in the
1564 * stable rbtree and stable_node will be equal to
1565 * stable_node_dup like if the chain never existed.
1566 */
1567 if (!stable_node_dup) {
1568 /*
1569 * Either all stable_node dups were full in
1570 * this stable_node chain, or this chain was
1571 * empty and should be rb_erased.
1572 */
1573 stable_node_any = stable_node_dup_any(stable_node,
1574 root);
1575 if (!stable_node_any) {
1576 /* rb_erase just run */
1577 goto again;
1578 }
1579 /*
1580 * Take any of the stable_node dups page of
1581 * this stable_node chain to let the tree walk
1582 * continue. All KSM pages belonging to the
1583 * stable_node dups in a stable_node chain
1584 * have the same content and they're
1585 * wrprotected at all times. Any will work
1586 * fine to continue the walk.
1587 */
1588 tree_page = get_ksm_page(stable_node_any, false);
1589 }
1590 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1591 if (!tree_page) {
1592 /*
1593 * If we walked over a stale stable_node,
1594 * get_ksm_page() will call rb_erase() and it
1595 * may rebalance the tree from under us. So
1596 * restart the search from scratch. Returning
1597 * NULL would be safe too, but we'd generate
1598 * false negative insertions just because some
1599 * stable_node was stale.
1600 */
1601 goto again;
1602 }
1603
1604 ret = memcmp_pages(page, tree_page);
1605 put_page(tree_page);
1606
1607 parent = *new;
1608 if (ret < 0)
1609 new = &parent->rb_left;
1610 else if (ret > 0)
1611 new = &parent->rb_right;
1612 else {
1613 if (page_node) {
1614 VM_BUG_ON(page_node->head != &migrate_nodes);
1615 /*
1616 * Test if the migrated page should be merged
1617 * into a stable node dup. If the mapcount is
1618 * 1 we can migrate it with another KSM page
1619 * without adding it to the chain.
1620 */
1621 if (page_mapcount(page) > 1)
1622 goto chain_append;
1623 }
1624
1625 if (!stable_node_dup) {
1626 /*
1627 * If the stable_node is a chain and
1628 * we got a payload match in memcmp
1629 * but we cannot merge the scanned
1630 * page in any of the existing
1631 * stable_node dups because they're
1632 * all full, we need to wait the
1633 * scanned page to find itself a match
1634 * in the unstable tree to create a
1635 * brand new KSM page to add later to
1636 * the dups of this stable_node.
1637 */
1638 return NULL;
1639 }
1640
1641 /*
1642 * Lock and unlock the stable_node's page (which
1643 * might already have been migrated) so that page
1644 * migration is sure to notice its raised count.
1645 * It would be more elegant to return stable_node
1646 * than kpage, but that involves more changes.
1647 */
1648 tree_page = get_ksm_page(stable_node_dup, true);
1649 if (unlikely(!tree_page))
1650 /*
1651 * The tree may have been rebalanced,
1652 * so re-evaluate parent and new.
1653 */
1654 goto again;
1655 unlock_page(tree_page);
1656
1657 if (get_kpfn_nid(stable_node_dup->kpfn) !=
1658 NUMA(stable_node_dup->nid)) {
1659 put_page(tree_page);
1660 goto replace;
1661 }
1662 return tree_page;
1663 }
1664 }
1665
1666 if (!page_node)
1667 return NULL;
1668
1669 list_del(&page_node->list);
1670 DO_NUMA(page_node->nid = nid);
1671 rb_link_node(&page_node->node, parent, new);
1672 rb_insert_color(&page_node->node, root);
1673out:
1674 if (is_page_sharing_candidate(page_node)) {
1675 get_page(page);
1676 return page;
1677 } else
1678 return NULL;
1679
1680replace:
1681 /*
1682 * If stable_node was a chain and chain_prune collapsed it,
1683 * stable_node has been updated to be the new regular
1684 * stable_node. A collapse of the chain is indistinguishable
1685 * from the case there was no chain in the stable
1686 * rbtree. Otherwise stable_node is the chain and
1687 * stable_node_dup is the dup to replace.
1688 */
1689 if (stable_node_dup == stable_node) {
1690 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1691 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1692 /* there is no chain */
1693 if (page_node) {
1694 VM_BUG_ON(page_node->head != &migrate_nodes);
1695 list_del(&page_node->list);
1696 DO_NUMA(page_node->nid = nid);
1697 rb_replace_node(&stable_node_dup->node,
1698 &page_node->node,
1699 root);
1700 if (is_page_sharing_candidate(page_node))
1701 get_page(page);
1702 else
1703 page = NULL;
1704 } else {
1705 rb_erase(&stable_node_dup->node, root);
1706 page = NULL;
1707 }
1708 } else {
1709 VM_BUG_ON(!is_stable_node_chain(stable_node));
1710 __stable_node_dup_del(stable_node_dup);
1711 if (page_node) {
1712 VM_BUG_ON(page_node->head != &migrate_nodes);
1713 list_del(&page_node->list);
1714 DO_NUMA(page_node->nid = nid);
1715 stable_node_chain_add_dup(page_node, stable_node);
1716 if (is_page_sharing_candidate(page_node))
1717 get_page(page);
1718 else
1719 page = NULL;
1720 } else {
1721 page = NULL;
1722 }
1723 }
1724 stable_node_dup->head = &migrate_nodes;
1725 list_add(&stable_node_dup->list, stable_node_dup->head);
1726 return page;
1727
1728chain_append:
1729 /* stable_node_dup could be null if it reached the limit */
1730 if (!stable_node_dup)
1731 stable_node_dup = stable_node_any;
1732 /*
1733 * If stable_node was a chain and chain_prune collapsed it,
1734 * stable_node has been updated to be the new regular
1735 * stable_node. A collapse of the chain is indistinguishable
1736 * from the case there was no chain in the stable
1737 * rbtree. Otherwise stable_node is the chain and
1738 * stable_node_dup is the dup to replace.
1739 */
1740 if (stable_node_dup == stable_node) {
1741 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1742 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1743 /* chain is missing so create it */
1744 stable_node = alloc_stable_node_chain(stable_node_dup,
1745 root);
1746 if (!stable_node)
1747 return NULL;
1748 }
1749 /*
1750 * Add this stable_node dup that was
1751 * migrated to the stable_node chain
1752 * of the current nid for this page
1753 * content.
1754 */
1755 VM_BUG_ON(!is_stable_node_chain(stable_node));
1756 VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1757 VM_BUG_ON(page_node->head != &migrate_nodes);
1758 list_del(&page_node->list);
1759 DO_NUMA(page_node->nid = nid);
1760 stable_node_chain_add_dup(page_node, stable_node);
1761 goto out;
1762}
1763
1764/*
1765 * stable_tree_insert - insert stable tree node pointing to new ksm page
1766 * into the stable tree.
1767 *
1768 * This function returns the stable tree node just allocated on success,
1769 * NULL otherwise.
1770 */
1771static struct stable_node *stable_tree_insert(struct page *kpage)
1772{
1773 int nid;
1774 unsigned long kpfn;
1775 struct rb_root *root;
1776 struct rb_node **new;
1777 struct rb_node *parent;
1778 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1779 bool need_chain = false;
1780
1781 kpfn = page_to_pfn(kpage);
1782 nid = get_kpfn_nid(kpfn);
1783 root = root_stable_tree + nid;
1784again:
1785 parent = NULL;
1786 new = &root->rb_node;
1787
1788 while (*new) {
1789 struct page *tree_page;
1790 int ret;
1791
1792 cond_resched();
1793 stable_node = rb_entry(*new, struct stable_node, node);
1794 stable_node_any = NULL;
1795 tree_page = chain(&stable_node_dup, stable_node, root);
1796 if (!stable_node_dup) {
1797 /*
1798 * Either all stable_node dups were full in
1799 * this stable_node chain, or this chain was
1800 * empty and should be rb_erased.
1801 */
1802 stable_node_any = stable_node_dup_any(stable_node,
1803 root);
1804 if (!stable_node_any) {
1805 /* rb_erase just run */
1806 goto again;
1807 }
1808 /*
1809 * Take any of the stable_node dups page of
1810 * this stable_node chain to let the tree walk
1811 * continue. All KSM pages belonging to the
1812 * stable_node dups in a stable_node chain
1813 * have the same content and they're
1814 * wrprotected at all times. Any will work
1815 * fine to continue the walk.
1816 */
1817 tree_page = get_ksm_page(stable_node_any, false);
1818 }
1819 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1820 if (!tree_page) {
1821 /*
1822 * If we walked over a stale stable_node,
1823 * get_ksm_page() will call rb_erase() and it
1824 * may rebalance the tree from under us. So
1825 * restart the search from scratch. Returning
1826 * NULL would be safe too, but we'd generate
1827 * false negative insertions just because some
1828 * stable_node was stale.
1829 */
1830 goto again;
1831 }
1832
1833 ret = memcmp_pages(kpage, tree_page);
1834 put_page(tree_page);
1835
1836 parent = *new;
1837 if (ret < 0)
1838 new = &parent->rb_left;
1839 else if (ret > 0)
1840 new = &parent->rb_right;
1841 else {
1842 need_chain = true;
1843 break;
1844 }
1845 }
1846
1847 stable_node_dup = alloc_stable_node();
1848 if (!stable_node_dup)
1849 return NULL;
1850
1851 INIT_HLIST_HEAD(&stable_node_dup->hlist);
1852 stable_node_dup->kpfn = kpfn;
1853 set_page_stable_node(kpage, stable_node_dup);
1854 stable_node_dup->rmap_hlist_len = 0;
1855 DO_NUMA(stable_node_dup->nid = nid);
1856 if (!need_chain) {
1857 rb_link_node(&stable_node_dup->node, parent, new);
1858 rb_insert_color(&stable_node_dup->node, root);
1859 } else {
1860 if (!is_stable_node_chain(stable_node)) {
1861 struct stable_node *orig = stable_node;
1862 /* chain is missing so create it */
1863 stable_node = alloc_stable_node_chain(orig, root);
1864 if (!stable_node) {
1865 free_stable_node(stable_node_dup);
1866 return NULL;
1867 }
1868 }
1869 stable_node_chain_add_dup(stable_node_dup, stable_node);
1870 }
1871
1872 return stable_node_dup;
1873}
1874
1875/*
1876 * unstable_tree_search_insert - search for identical page,
1877 * else insert rmap_item into the unstable tree.
1878 *
1879 * This function searches for a page in the unstable tree identical to the
1880 * page currently being scanned; and if no identical page is found in the
1881 * tree, we insert rmap_item as a new object into the unstable tree.
1882 *
1883 * This function returns pointer to rmap_item found to be identical
1884 * to the currently scanned page, NULL otherwise.
1885 *
1886 * This function does both searching and inserting, because they share
1887 * the same walking algorithm in an rbtree.
1888 */
1889static
1890struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1891 struct page *page,
1892 struct page **tree_pagep)
1893{
1894 struct rb_node **new;
1895 struct rb_root *root;
1896 struct rb_node *parent = NULL;
1897 int nid;
1898
1899 nid = get_kpfn_nid(page_to_pfn(page));
1900 root = root_unstable_tree + nid;
1901 new = &root->rb_node;
1902
1903 while (*new) {
1904 struct rmap_item *tree_rmap_item;
1905 struct page *tree_page;
1906 int ret;
1907
1908 cond_resched();
1909 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1910 tree_page = get_mergeable_page(tree_rmap_item);
1911 if (!tree_page)
1912 return NULL;
1913
1914 /*
1915 * Don't substitute a ksm page for a forked page.
1916 */
1917 if (page == tree_page) {
1918 put_page(tree_page);
1919 return NULL;
1920 }
1921
1922 ret = memcmp_pages(page, tree_page);
1923
1924 parent = *new;
1925 if (ret < 0) {
1926 put_page(tree_page);
1927 new = &parent->rb_left;
1928 } else if (ret > 0) {
1929 put_page(tree_page);
1930 new = &parent->rb_right;
1931 } else if (!ksm_merge_across_nodes &&
1932 page_to_nid(tree_page) != nid) {
1933 /*
1934 * If tree_page has been migrated to another NUMA node,
1935 * it will be flushed out and put in the right unstable
1936 * tree next time: only merge with it when across_nodes.
1937 */
1938 put_page(tree_page);
1939 return NULL;
1940 } else {
1941 *tree_pagep = tree_page;
1942 return tree_rmap_item;
1943 }
1944 }
1945
1946 rmap_item->address |= UNSTABLE_FLAG;
1947 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1948 DO_NUMA(rmap_item->nid = nid);
1949 rb_link_node(&rmap_item->node, parent, new);
1950 rb_insert_color(&rmap_item->node, root);
1951
1952 ksm_pages_unshared++;
1953 return NULL;
1954}
1955
1956/*
1957 * stable_tree_append - add another rmap_item to the linked list of
1958 * rmap_items hanging off a given node of the stable tree, all sharing
1959 * the same ksm page.
1960 */
1961static void stable_tree_append(struct rmap_item *rmap_item,
1962 struct stable_node *stable_node,
1963 bool max_page_sharing_bypass)
1964{
1965 /*
1966 * rmap won't find this mapping if we don't insert the
1967 * rmap_item in the right stable_node
1968 * duplicate. page_migration could break later if rmap breaks,
1969 * so we can as well crash here. We really need to check for
1970 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
1971 * for other negative values as an undeflow if detected here
1972 * for the first time (and not when decreasing rmap_hlist_len)
1973 * would be sign of memory corruption in the stable_node.
1974 */
1975 BUG_ON(stable_node->rmap_hlist_len < 0);
1976
1977 stable_node->rmap_hlist_len++;
1978 if (!max_page_sharing_bypass)
1979 /* possibly non fatal but unexpected overflow, only warn */
1980 WARN_ON_ONCE(stable_node->rmap_hlist_len >
1981 ksm_max_page_sharing);
1982
1983 rmap_item->head = stable_node;
1984 rmap_item->address |= STABLE_FLAG;
1985 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1986
1987 if (rmap_item->hlist.next)
1988 ksm_pages_sharing++;
1989 else
1990 ksm_pages_shared++;
1991}
1992
1993/*
1994 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1995 * if not, compare checksum to previous and if it's the same, see if page can
1996 * be inserted into the unstable tree, or merged with a page already there and
1997 * both transferred to the stable tree.
1998 *
1999 * @page: the page that we are searching identical page to.
2000 * @rmap_item: the reverse mapping into the virtual address of this page
2001 */
2002static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2003{
2004 struct mm_struct *mm = rmap_item->mm;
2005 struct rmap_item *tree_rmap_item;
2006 struct page *tree_page = NULL;
2007 struct stable_node *stable_node;
2008 struct page *kpage;
2009 unsigned int checksum;
2010 int err;
2011 bool max_page_sharing_bypass = false;
2012
2013 stable_node = page_stable_node(page);
2014 if (stable_node) {
2015 if (stable_node->head != &migrate_nodes &&
2016 get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2017 NUMA(stable_node->nid)) {
2018 stable_node_dup_del(stable_node);
2019 stable_node->head = &migrate_nodes;
2020 list_add(&stable_node->list, stable_node->head);
2021 }
2022 if (stable_node->head != &migrate_nodes &&
2023 rmap_item->head == stable_node)
2024 return;
2025 /*
2026 * If it's a KSM fork, allow it to go over the sharing limit
2027 * without warnings.
2028 */
2029 if (!is_page_sharing_candidate(stable_node))
2030 max_page_sharing_bypass = true;
2031 }
2032
2033 /* We first start with searching the page inside the stable tree */
2034 kpage = stable_tree_search(page);
2035 if (kpage == page && rmap_item->head == stable_node) {
2036 put_page(kpage);
2037 return;
2038 }
2039
2040 remove_rmap_item_from_tree(rmap_item);
2041
2042 if (kpage) {
2043 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2044 if (!err) {
2045 /*
2046 * The page was successfully merged:
2047 * add its rmap_item to the stable tree.
2048 */
2049 lock_page(kpage);
2050 stable_tree_append(rmap_item, page_stable_node(kpage),
2051 max_page_sharing_bypass);
2052 unlock_page(kpage);
2053 }
2054 put_page(kpage);
2055 return;
2056 }
2057
2058 /*
2059 * If the hash value of the page has changed from the last time
2060 * we calculated it, this page is changing frequently: therefore we
2061 * don't want to insert it in the unstable tree, and we don't want
2062 * to waste our time searching for something identical to it there.
2063 */
2064 checksum = calc_checksum(page);
2065 if (rmap_item->oldchecksum != checksum) {
2066 rmap_item->oldchecksum = checksum;
2067 return;
2068 }
2069
2070 /*
2071 * Same checksum as an empty page. We attempt to merge it with the
2072 * appropriate zero page if the user enabled this via sysfs.
2073 */
2074 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2075 struct vm_area_struct *vma;
2076
2077 down_read(&mm->mmap_sem);
2078 vma = find_mergeable_vma(mm, rmap_item->address);
2079 err = try_to_merge_one_page(vma, page,
2080 ZERO_PAGE(rmap_item->address));
2081 up_read(&mm->mmap_sem);
2082 /*
2083 * In case of failure, the page was not really empty, so we
2084 * need to continue. Otherwise we're done.
2085 */
2086 if (!err)
2087 return;
2088 }
2089 tree_rmap_item =
2090 unstable_tree_search_insert(rmap_item, page, &tree_page);
2091 if (tree_rmap_item) {
2092 bool split;
2093
2094 kpage = try_to_merge_two_pages(rmap_item, page,
2095 tree_rmap_item, tree_page);
2096 /*
2097 * If both pages we tried to merge belong to the same compound
2098 * page, then we actually ended up increasing the reference
2099 * count of the same compound page twice, and split_huge_page
2100 * failed.
2101 * Here we set a flag if that happened, and we use it later to
2102 * try split_huge_page again. Since we call put_page right
2103 * afterwards, the reference count will be correct and
2104 * split_huge_page should succeed.
2105 */
2106 split = PageTransCompound(page)
2107 && compound_head(page) == compound_head(tree_page);
2108 put_page(tree_page);
2109 if (kpage) {
2110 /*
2111 * The pages were successfully merged: insert new
2112 * node in the stable tree and add both rmap_items.
2113 */
2114 lock_page(kpage);
2115 stable_node = stable_tree_insert(kpage);
2116 if (stable_node) {
2117 stable_tree_append(tree_rmap_item, stable_node,
2118 false);
2119 stable_tree_append(rmap_item, stable_node,
2120 false);
2121 }
2122 unlock_page(kpage);
2123
2124 /*
2125 * If we fail to insert the page into the stable tree,
2126 * we will have 2 virtual addresses that are pointing
2127 * to a ksm page left outside the stable tree,
2128 * in which case we need to break_cow on both.
2129 */
2130 if (!stable_node) {
2131 break_cow(tree_rmap_item);
2132 break_cow(rmap_item);
2133 }
2134 } else if (split) {
2135 /*
2136 * We are here if we tried to merge two pages and
2137 * failed because they both belonged to the same
2138 * compound page. We will split the page now, but no
2139 * merging will take place.
2140 * We do not want to add the cost of a full lock; if
2141 * the page is locked, it is better to skip it and
2142 * perhaps try again later.
2143 */
2144 if (!trylock_page(page))
2145 return;
2146 split_huge_page(page);
2147 unlock_page(page);
2148 }
2149 }
2150}
2151
2152static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2153 struct rmap_item **rmap_list,
2154 unsigned long addr)
2155{
2156 struct rmap_item *rmap_item;
2157
2158 while (*rmap_list) {
2159 rmap_item = *rmap_list;
2160 if ((rmap_item->address & PAGE_MASK) == addr)
2161 return rmap_item;
2162 if (rmap_item->address > addr)
2163 break;
2164 *rmap_list = rmap_item->rmap_list;
2165 remove_rmap_item_from_tree(rmap_item);
2166 free_rmap_item(rmap_item);
2167 }
2168
2169 rmap_item = alloc_rmap_item();
2170 if (rmap_item) {
2171 /* It has already been zeroed */
2172 rmap_item->mm = mm_slot->mm;
2173 rmap_item->address = addr;
2174 rmap_item->rmap_list = *rmap_list;
2175 *rmap_list = rmap_item;
2176 }
2177 return rmap_item;
2178}
2179
2180static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2181{
2182 struct mm_struct *mm;
2183 struct mm_slot *slot;
2184 struct vm_area_struct *vma;
2185 struct rmap_item *rmap_item;
2186 int nid;
2187
2188 if (list_empty(&ksm_mm_head.mm_list))
2189 return NULL;
2190
2191 slot = ksm_scan.mm_slot;
2192 if (slot == &ksm_mm_head) {
2193 /*
2194 * A number of pages can hang around indefinitely on per-cpu
2195 * pagevecs, raised page count preventing write_protect_page
2196 * from merging them. Though it doesn't really matter much,
2197 * it is puzzling to see some stuck in pages_volatile until
2198 * other activity jostles them out, and they also prevented
2199 * LTP's KSM test from succeeding deterministically; so drain
2200 * them here (here rather than on entry to ksm_do_scan(),
2201 * so we don't IPI too often when pages_to_scan is set low).
2202 */
2203 lru_add_drain_all();
2204
2205 /*
2206 * Whereas stale stable_nodes on the stable_tree itself
2207 * get pruned in the regular course of stable_tree_search(),
2208 * those moved out to the migrate_nodes list can accumulate:
2209 * so prune them once before each full scan.
2210 */
2211 if (!ksm_merge_across_nodes) {
2212 struct stable_node *stable_node, *next;
2213 struct page *page;
2214
2215 list_for_each_entry_safe(stable_node, next,
2216 &migrate_nodes, list) {
2217 page = get_ksm_page(stable_node, false);
2218 if (page)
2219 put_page(page);
2220 cond_resched();
2221 }
2222 }
2223
2224 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2225 root_unstable_tree[nid] = RB_ROOT;
2226
2227 spin_lock(&ksm_mmlist_lock);
2228 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2229 ksm_scan.mm_slot = slot;
2230 spin_unlock(&ksm_mmlist_lock);
2231 /*
2232 * Although we tested list_empty() above, a racing __ksm_exit
2233 * of the last mm on the list may have removed it since then.
2234 */
2235 if (slot == &ksm_mm_head)
2236 return NULL;
2237next_mm:
2238 ksm_scan.address = 0;
2239 ksm_scan.rmap_list = &slot->rmap_list;
2240 }
2241
2242 mm = slot->mm;
2243 down_read(&mm->mmap_sem);
2244 if (ksm_test_exit(mm))
2245 vma = NULL;
2246 else
2247 vma = find_vma(mm, ksm_scan.address);
2248
2249 for (; vma; vma = vma->vm_next) {
2250 if (!(vma->vm_flags & VM_MERGEABLE))
2251 continue;
2252 if (ksm_scan.address < vma->vm_start)
2253 ksm_scan.address = vma->vm_start;
2254 if (!vma->anon_vma)
2255 ksm_scan.address = vma->vm_end;
2256
2257 while (ksm_scan.address < vma->vm_end) {
2258 if (ksm_test_exit(mm))
2259 break;
2260 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2261 if (IS_ERR_OR_NULL(*page)) {
2262 ksm_scan.address += PAGE_SIZE;
2263 cond_resched();
2264 continue;
2265 }
2266 if (PageAnon(*page)) {
2267 flush_anon_page(vma, *page, ksm_scan.address);
2268 flush_dcache_page(*page);
2269 rmap_item = get_next_rmap_item(slot,
2270 ksm_scan.rmap_list, ksm_scan.address);
2271 if (rmap_item) {
2272 ksm_scan.rmap_list =
2273 &rmap_item->rmap_list;
2274 ksm_scan.address += PAGE_SIZE;
2275 } else
2276 put_page(*page);
2277 up_read(&mm->mmap_sem);
2278 return rmap_item;
2279 }
2280 put_page(*page);
2281 ksm_scan.address += PAGE_SIZE;
2282 cond_resched();
2283 }
2284 }
2285
2286 if (ksm_test_exit(mm)) {
2287 ksm_scan.address = 0;
2288 ksm_scan.rmap_list = &slot->rmap_list;
2289 }
2290 /*
2291 * Nuke all the rmap_items that are above this current rmap:
2292 * because there were no VM_MERGEABLE vmas with such addresses.
2293 */
2294 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
2295
2296 spin_lock(&ksm_mmlist_lock);
2297 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2298 struct mm_slot, mm_list);
2299 if (ksm_scan.address == 0) {
2300 /*
2301 * We've completed a full scan of all vmas, holding mmap_sem
2302 * throughout, and found no VM_MERGEABLE: so do the same as
2303 * __ksm_exit does to remove this mm from all our lists now.
2304 * This applies either when cleaning up after __ksm_exit
2305 * (but beware: we can reach here even before __ksm_exit),
2306 * or when all VM_MERGEABLE areas have been unmapped (and
2307 * mmap_sem then protects against race with MADV_MERGEABLE).
2308 */
2309 hash_del(&slot->link);
2310 list_del(&slot->mm_list);
2311 spin_unlock(&ksm_mmlist_lock);
2312
2313 free_mm_slot(slot);
2314 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2315 up_read(&mm->mmap_sem);
2316 mmdrop(mm);
2317 } else {
2318 up_read(&mm->mmap_sem);
2319 /*
2320 * up_read(&mm->mmap_sem) first because after
2321 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2322 * already have been freed under us by __ksm_exit()
2323 * because the "mm_slot" is still hashed and
2324 * ksm_scan.mm_slot doesn't point to it anymore.
2325 */
2326 spin_unlock(&ksm_mmlist_lock);
2327 }
2328
2329 /* Repeat until we've completed scanning the whole list */
2330 slot = ksm_scan.mm_slot;
2331 if (slot != &ksm_mm_head)
2332 goto next_mm;
2333
2334 ksm_scan.seqnr++;
2335 return NULL;
2336}
2337
2338/**
2339 * ksm_do_scan - the ksm scanner main worker function.
2340 * @scan_npages: number of pages we want to scan before we return.
2341 */
2342static void ksm_do_scan(unsigned int scan_npages)
2343{
2344 struct rmap_item *rmap_item;
2345 struct page *uninitialized_var(page);
2346
2347 while (scan_npages-- && likely(!freezing(current))) {
2348 cond_resched();
2349 rmap_item = scan_get_next_rmap_item(&page);
2350 if (!rmap_item)
2351 return;
2352 cmp_and_merge_page(page, rmap_item);
2353 put_page(page);
2354 }
2355}
2356
2357static int ksmd_should_run(void)
2358{
2359 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2360}
2361
2362static int ksm_scan_thread(void *nothing)
2363{
2364 set_freezable();
2365 set_user_nice(current, 5);
2366
2367 while (!kthread_should_stop()) {
2368 mutex_lock(&ksm_thread_mutex);
2369 wait_while_offlining();
2370 if (ksmd_should_run())
2371 ksm_do_scan(ksm_thread_pages_to_scan);
2372 mutex_unlock(&ksm_thread_mutex);
2373
2374 try_to_freeze();
2375
2376 if (ksmd_should_run()) {
2377 schedule_timeout_interruptible(
2378 msecs_to_jiffies(ksm_thread_sleep_millisecs));
2379 } else {
2380 wait_event_freezable(ksm_thread_wait,
2381 ksmd_should_run() || kthread_should_stop());
2382 }
2383 }
2384 return 0;
2385}
2386
2387int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2388 unsigned long end, int advice, unsigned long *vm_flags)
2389{
2390 struct mm_struct *mm = vma->vm_mm;
2391 int err;
2392
2393 switch (advice) {
2394 case MADV_MERGEABLE:
2395 /*
2396 * Be somewhat over-protective for now!
2397 */
2398 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
2399 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
2400 VM_HUGETLB | VM_MIXEDMAP))
2401 return 0; /* just ignore the advice */
2402
2403#ifdef VM_SAO
2404 if (*vm_flags & VM_SAO)
2405 return 0;
2406#endif
2407#ifdef VM_SPARC_ADI
2408 if (*vm_flags & VM_SPARC_ADI)
2409 return 0;
2410#endif
2411
2412 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2413 err = __ksm_enter(mm);
2414 if (err)
2415 return err;
2416 }
2417
2418 *vm_flags |= VM_MERGEABLE;
2419 break;
2420
2421 case MADV_UNMERGEABLE:
2422 if (!(*vm_flags & VM_MERGEABLE))
2423 return 0; /* just ignore the advice */
2424
2425 if (vma->anon_vma) {
2426 err = unmerge_ksm_pages(vma, start, end);
2427 if (err)
2428 return err;
2429 }
2430
2431 *vm_flags &= ~VM_MERGEABLE;
2432 break;
2433 }
2434
2435 return 0;
2436}
2437
2438int __ksm_enter(struct mm_struct *mm)
2439{
2440 struct mm_slot *mm_slot;
2441 int needs_wakeup;
2442
2443 mm_slot = alloc_mm_slot();
2444 if (!mm_slot)
2445 return -ENOMEM;
2446
2447 /* Check ksm_run too? Would need tighter locking */
2448 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2449
2450 spin_lock(&ksm_mmlist_lock);
2451 insert_to_mm_slots_hash(mm, mm_slot);
2452 /*
2453 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2454 * insert just behind the scanning cursor, to let the area settle
2455 * down a little; when fork is followed by immediate exec, we don't
2456 * want ksmd to waste time setting up and tearing down an rmap_list.
2457 *
2458 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2459 * scanning cursor, otherwise KSM pages in newly forked mms will be
2460 * missed: then we might as well insert at the end of the list.
2461 */
2462 if (ksm_run & KSM_RUN_UNMERGE)
2463 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2464 else
2465 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2466 spin_unlock(&ksm_mmlist_lock);
2467
2468 set_bit(MMF_VM_MERGEABLE, &mm->flags);
2469 mmgrab(mm);
2470
2471 if (needs_wakeup)
2472 wake_up_interruptible(&ksm_thread_wait);
2473
2474 return 0;
2475}
2476
2477void __ksm_exit(struct mm_struct *mm)
2478{
2479 struct mm_slot *mm_slot;
2480 int easy_to_free = 0;
2481
2482 /*
2483 * This process is exiting: if it's straightforward (as is the
2484 * case when ksmd was never running), free mm_slot immediately.
2485 * But if it's at the cursor or has rmap_items linked to it, use
2486 * mmap_sem to synchronize with any break_cows before pagetables
2487 * are freed, and leave the mm_slot on the list for ksmd to free.
2488 * Beware: ksm may already have noticed it exiting and freed the slot.
2489 */
2490
2491 spin_lock(&ksm_mmlist_lock);
2492 mm_slot = get_mm_slot(mm);
2493 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2494 if (!mm_slot->rmap_list) {
2495 hash_del(&mm_slot->link);
2496 list_del(&mm_slot->mm_list);
2497 easy_to_free = 1;
2498 } else {
2499 list_move(&mm_slot->mm_list,
2500 &ksm_scan.mm_slot->mm_list);
2501 }
2502 }
2503 spin_unlock(&ksm_mmlist_lock);
2504
2505 if (easy_to_free) {
2506 free_mm_slot(mm_slot);
2507 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2508 mmdrop(mm);
2509 } else if (mm_slot) {
2510 down_write(&mm->mmap_sem);
2511 up_write(&mm->mmap_sem);
2512 }
2513}
2514
2515struct page *ksm_might_need_to_copy(struct page *page,
2516 struct vm_area_struct *vma, unsigned long address)
2517{
2518 struct anon_vma *anon_vma = page_anon_vma(page);
2519 struct page *new_page;
2520
2521 if (PageKsm(page)) {
2522 if (page_stable_node(page) &&
2523 !(ksm_run & KSM_RUN_UNMERGE))
2524 return page; /* no need to copy it */
2525 } else if (!anon_vma) {
2526 return page; /* no need to copy it */
2527 } else if (anon_vma->root == vma->anon_vma->root &&
2528 page->index == linear_page_index(vma, address)) {
2529 return page; /* still no need to copy it */
2530 }
2531 if (!PageUptodate(page))
2532 return page; /* let do_swap_page report the error */
2533
2534 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2535 if (new_page) {
2536 copy_user_highpage(new_page, page, address, vma);
2537
2538 SetPageDirty(new_page);
2539 __SetPageUptodate(new_page);
2540 __SetPageLocked(new_page);
2541 }
2542
2543 return new_page;
2544}
2545
2546void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
2547{
2548 struct stable_node *stable_node;
2549 struct rmap_item *rmap_item;
2550 int search_new_forks = 0;
2551
2552 VM_BUG_ON_PAGE(!PageKsm(page), page);
2553
2554 /*
2555 * Rely on the page lock to protect against concurrent modifications
2556 * to that page's node of the stable tree.
2557 */
2558 VM_BUG_ON_PAGE(!PageLocked(page), page);
2559
2560 stable_node = page_stable_node(page);
2561 if (!stable_node)
2562 return;
2563again:
2564 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2565 struct anon_vma *anon_vma = rmap_item->anon_vma;
2566 struct anon_vma_chain *vmac;
2567 struct vm_area_struct *vma;
2568
2569 cond_resched();
2570 anon_vma_lock_read(anon_vma);
2571 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2572 0, ULONG_MAX) {
2573 cond_resched();
2574 vma = vmac->vma;
2575 if (rmap_item->address < vma->vm_start ||
2576 rmap_item->address >= vma->vm_end)
2577 continue;
2578 /*
2579 * Initially we examine only the vma which covers this
2580 * rmap_item; but later, if there is still work to do,
2581 * we examine covering vmas in other mms: in case they
2582 * were forked from the original since ksmd passed.
2583 */
2584 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2585 continue;
2586
2587 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2588 continue;
2589
2590 if (!rwc->rmap_one(page, vma,
2591 rmap_item->address, rwc->arg)) {
2592 anon_vma_unlock_read(anon_vma);
2593 return;
2594 }
2595 if (rwc->done && rwc->done(page)) {
2596 anon_vma_unlock_read(anon_vma);
2597 return;
2598 }
2599 }
2600 anon_vma_unlock_read(anon_vma);
2601 }
2602 if (!search_new_forks++)
2603 goto again;
2604}
2605
2606#ifdef CONFIG_MIGRATION
2607void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2608{
2609 struct stable_node *stable_node;
2610
2611 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2612 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2613 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2614
2615 stable_node = page_stable_node(newpage);
2616 if (stable_node) {
2617 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2618 stable_node->kpfn = page_to_pfn(newpage);
2619 /*
2620 * newpage->mapping was set in advance; now we need smp_wmb()
2621 * to make sure that the new stable_node->kpfn is visible
2622 * to get_ksm_page() before it can see that oldpage->mapping
2623 * has gone stale (or that PageSwapCache has been cleared).
2624 */
2625 smp_wmb();
2626 set_page_stable_node(oldpage, NULL);
2627 }
2628}
2629#endif /* CONFIG_MIGRATION */
2630
2631#ifdef CONFIG_MEMORY_HOTREMOVE
2632static void wait_while_offlining(void)
2633{
2634 while (ksm_run & KSM_RUN_OFFLINE) {
2635 mutex_unlock(&ksm_thread_mutex);
2636 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2637 TASK_UNINTERRUPTIBLE);
2638 mutex_lock(&ksm_thread_mutex);
2639 }
2640}
2641
2642static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2643 unsigned long start_pfn,
2644 unsigned long end_pfn)
2645{
2646 if (stable_node->kpfn >= start_pfn &&
2647 stable_node->kpfn < end_pfn) {
2648 /*
2649 * Don't get_ksm_page, page has already gone:
2650 * which is why we keep kpfn instead of page*
2651 */
2652 remove_node_from_stable_tree(stable_node);
2653 return true;
2654 }
2655 return false;
2656}
2657
2658static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2659 unsigned long start_pfn,
2660 unsigned long end_pfn,
2661 struct rb_root *root)
2662{
2663 struct stable_node *dup;
2664 struct hlist_node *hlist_safe;
2665
2666 if (!is_stable_node_chain(stable_node)) {
2667 VM_BUG_ON(is_stable_node_dup(stable_node));
2668 return stable_node_dup_remove_range(stable_node, start_pfn,
2669 end_pfn);
2670 }
2671
2672 hlist_for_each_entry_safe(dup, hlist_safe,
2673 &stable_node->hlist, hlist_dup) {
2674 VM_BUG_ON(!is_stable_node_dup(dup));
2675 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2676 }
2677 if (hlist_empty(&stable_node->hlist)) {
2678 free_stable_node_chain(stable_node, root);
2679 return true; /* notify caller that tree was rebalanced */
2680 } else
2681 return false;
2682}
2683
2684static void ksm_check_stable_tree(unsigned long start_pfn,
2685 unsigned long end_pfn)
2686{
2687 struct stable_node *stable_node, *next;
2688 struct rb_node *node;
2689 int nid;
2690
2691 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2692 node = rb_first(root_stable_tree + nid);
2693 while (node) {
2694 stable_node = rb_entry(node, struct stable_node, node);
2695 if (stable_node_chain_remove_range(stable_node,
2696 start_pfn, end_pfn,
2697 root_stable_tree +
2698 nid))
2699 node = rb_first(root_stable_tree + nid);
2700 else
2701 node = rb_next(node);
2702 cond_resched();
2703 }
2704 }
2705 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2706 if (stable_node->kpfn >= start_pfn &&
2707 stable_node->kpfn < end_pfn)
2708 remove_node_from_stable_tree(stable_node);
2709 cond_resched();
2710 }
2711}
2712
2713static int ksm_memory_callback(struct notifier_block *self,
2714 unsigned long action, void *arg)
2715{
2716 struct memory_notify *mn = arg;
2717
2718 switch (action) {
2719 case MEM_GOING_OFFLINE:
2720 /*
2721 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2722 * and remove_all_stable_nodes() while memory is going offline:
2723 * it is unsafe for them to touch the stable tree at this time.
2724 * But unmerge_ksm_pages(), rmap lookups and other entry points
2725 * which do not need the ksm_thread_mutex are all safe.
2726 */
2727 mutex_lock(&ksm_thread_mutex);
2728 ksm_run |= KSM_RUN_OFFLINE;
2729 mutex_unlock(&ksm_thread_mutex);
2730 break;
2731
2732 case MEM_OFFLINE:
2733 /*
2734 * Most of the work is done by page migration; but there might
2735 * be a few stable_nodes left over, still pointing to struct
2736 * pages which have been offlined: prune those from the tree,
2737 * otherwise get_ksm_page() might later try to access a
2738 * non-existent struct page.
2739 */
2740 ksm_check_stable_tree(mn->start_pfn,
2741 mn->start_pfn + mn->nr_pages);
2742 /* fallthrough */
2743
2744 case MEM_CANCEL_OFFLINE:
2745 mutex_lock(&ksm_thread_mutex);
2746 ksm_run &= ~KSM_RUN_OFFLINE;
2747 mutex_unlock(&ksm_thread_mutex);
2748
2749 smp_mb(); /* wake_up_bit advises this */
2750 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2751 break;
2752 }
2753 return NOTIFY_OK;
2754}
2755#else
2756static void wait_while_offlining(void)
2757{
2758}
2759#endif /* CONFIG_MEMORY_HOTREMOVE */
2760
2761#ifdef CONFIG_SYSFS
2762/*
2763 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2764 */
2765
2766#define KSM_ATTR_RO(_name) \
2767 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2768#define KSM_ATTR(_name) \
2769 static struct kobj_attribute _name##_attr = \
2770 __ATTR(_name, 0644, _name##_show, _name##_store)
2771
2772static ssize_t sleep_millisecs_show(struct kobject *kobj,
2773 struct kobj_attribute *attr, char *buf)
2774{
2775 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2776}
2777
2778static ssize_t sleep_millisecs_store(struct kobject *kobj,
2779 struct kobj_attribute *attr,
2780 const char *buf, size_t count)
2781{
2782 unsigned long msecs;
2783 int err;
2784
2785 err = kstrtoul(buf, 10, &msecs);
2786 if (err || msecs > UINT_MAX)
2787 return -EINVAL;
2788
2789 ksm_thread_sleep_millisecs = msecs;
2790
2791 return count;
2792}
2793KSM_ATTR(sleep_millisecs);
2794
2795static ssize_t pages_to_scan_show(struct kobject *kobj,
2796 struct kobj_attribute *attr, char *buf)
2797{
2798 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2799}
2800
2801static ssize_t pages_to_scan_store(struct kobject *kobj,
2802 struct kobj_attribute *attr,
2803 const char *buf, size_t count)
2804{
2805 int err;
2806 unsigned long nr_pages;
2807
2808 err = kstrtoul(buf, 10, &nr_pages);
2809 if (err || nr_pages > UINT_MAX)
2810 return -EINVAL;
2811
2812 ksm_thread_pages_to_scan = nr_pages;
2813
2814 return count;
2815}
2816KSM_ATTR(pages_to_scan);
2817
2818static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2819 char *buf)
2820{
2821 return sprintf(buf, "%lu\n", ksm_run);
2822}
2823
2824static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2825 const char *buf, size_t count)
2826{
2827 int err;
2828 unsigned long flags;
2829
2830 err = kstrtoul(buf, 10, &flags);
2831 if (err || flags > UINT_MAX)
2832 return -EINVAL;
2833 if (flags > KSM_RUN_UNMERGE)
2834 return -EINVAL;
2835
2836 /*
2837 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2838 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2839 * breaking COW to free the pages_shared (but leaves mm_slots
2840 * on the list for when ksmd may be set running again).
2841 */
2842
2843 mutex_lock(&ksm_thread_mutex);
2844 wait_while_offlining();
2845 if (ksm_run != flags) {
2846 ksm_run = flags;
2847 if (flags & KSM_RUN_UNMERGE) {
2848 set_current_oom_origin();
2849 err = unmerge_and_remove_all_rmap_items();
2850 clear_current_oom_origin();
2851 if (err) {
2852 ksm_run = KSM_RUN_STOP;
2853 count = err;
2854 }
2855 }
2856 }
2857 mutex_unlock(&ksm_thread_mutex);
2858
2859 if (flags & KSM_RUN_MERGE)
2860 wake_up_interruptible(&ksm_thread_wait);
2861
2862 return count;
2863}
2864KSM_ATTR(run);
2865
2866#ifdef CONFIG_NUMA
2867static ssize_t merge_across_nodes_show(struct kobject *kobj,
2868 struct kobj_attribute *attr, char *buf)
2869{
2870 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2871}
2872
2873static ssize_t merge_across_nodes_store(struct kobject *kobj,
2874 struct kobj_attribute *attr,
2875 const char *buf, size_t count)
2876{
2877 int err;
2878 unsigned long knob;
2879
2880 err = kstrtoul(buf, 10, &knob);
2881 if (err)
2882 return err;
2883 if (knob > 1)
2884 return -EINVAL;
2885
2886 mutex_lock(&ksm_thread_mutex);
2887 wait_while_offlining();
2888 if (ksm_merge_across_nodes != knob) {
2889 if (ksm_pages_shared || remove_all_stable_nodes())
2890 err = -EBUSY;
2891 else if (root_stable_tree == one_stable_tree) {
2892 struct rb_root *buf;
2893 /*
2894 * This is the first time that we switch away from the
2895 * default of merging across nodes: must now allocate
2896 * a buffer to hold as many roots as may be needed.
2897 * Allocate stable and unstable together:
2898 * MAXSMP NODES_SHIFT 10 will use 16kB.
2899 */
2900 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2901 GFP_KERNEL);
2902 /* Let us assume that RB_ROOT is NULL is zero */
2903 if (!buf)
2904 err = -ENOMEM;
2905 else {
2906 root_stable_tree = buf;
2907 root_unstable_tree = buf + nr_node_ids;
2908 /* Stable tree is empty but not the unstable */
2909 root_unstable_tree[0] = one_unstable_tree[0];
2910 }
2911 }
2912 if (!err) {
2913 ksm_merge_across_nodes = knob;
2914 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2915 }
2916 }
2917 mutex_unlock(&ksm_thread_mutex);
2918
2919 return err ? err : count;
2920}
2921KSM_ATTR(merge_across_nodes);
2922#endif
2923
2924static ssize_t use_zero_pages_show(struct kobject *kobj,
2925 struct kobj_attribute *attr, char *buf)
2926{
2927 return sprintf(buf, "%u\n", ksm_use_zero_pages);
2928}
2929static ssize_t use_zero_pages_store(struct kobject *kobj,
2930 struct kobj_attribute *attr,
2931 const char *buf, size_t count)
2932{
2933 int err;
2934 bool value;
2935
2936 err = kstrtobool(buf, &value);
2937 if (err)
2938 return -EINVAL;
2939
2940 ksm_use_zero_pages = value;
2941
2942 return count;
2943}
2944KSM_ATTR(use_zero_pages);
2945
2946static ssize_t max_page_sharing_show(struct kobject *kobj,
2947 struct kobj_attribute *attr, char *buf)
2948{
2949 return sprintf(buf, "%u\n", ksm_max_page_sharing);
2950}
2951
2952static ssize_t max_page_sharing_store(struct kobject *kobj,
2953 struct kobj_attribute *attr,
2954 const char *buf, size_t count)
2955{
2956 int err;
2957 int knob;
2958
2959 err = kstrtoint(buf, 10, &knob);
2960 if (err)
2961 return err;
2962 /*
2963 * When a KSM page is created it is shared by 2 mappings. This
2964 * being a signed comparison, it implicitly verifies it's not
2965 * negative.
2966 */
2967 if (knob < 2)
2968 return -EINVAL;
2969
2970 if (READ_ONCE(ksm_max_page_sharing) == knob)
2971 return count;
2972
2973 mutex_lock(&ksm_thread_mutex);
2974 wait_while_offlining();
2975 if (ksm_max_page_sharing != knob) {
2976 if (ksm_pages_shared || remove_all_stable_nodes())
2977 err = -EBUSY;
2978 else
2979 ksm_max_page_sharing = knob;
2980 }
2981 mutex_unlock(&ksm_thread_mutex);
2982
2983 return err ? err : count;
2984}
2985KSM_ATTR(max_page_sharing);
2986
2987static ssize_t pages_shared_show(struct kobject *kobj,
2988 struct kobj_attribute *attr, char *buf)
2989{
2990 return sprintf(buf, "%lu\n", ksm_pages_shared);
2991}
2992KSM_ATTR_RO(pages_shared);
2993
2994static ssize_t pages_sharing_show(struct kobject *kobj,
2995 struct kobj_attribute *attr, char *buf)
2996{
2997 return sprintf(buf, "%lu\n", ksm_pages_sharing);
2998}
2999KSM_ATTR_RO(pages_sharing);
3000
3001static ssize_t pages_unshared_show(struct kobject *kobj,
3002 struct kobj_attribute *attr, char *buf)
3003{
3004 return sprintf(buf, "%lu\n", ksm_pages_unshared);
3005}
3006KSM_ATTR_RO(pages_unshared);
3007
3008static ssize_t pages_volatile_show(struct kobject *kobj,
3009 struct kobj_attribute *attr, char *buf)
3010{
3011 long ksm_pages_volatile;
3012
3013 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3014 - ksm_pages_sharing - ksm_pages_unshared;
3015 /*
3016 * It was not worth any locking to calculate that statistic,
3017 * but it might therefore sometimes be negative: conceal that.
3018 */
3019 if (ksm_pages_volatile < 0)
3020 ksm_pages_volatile = 0;
3021 return sprintf(buf, "%ld\n", ksm_pages_volatile);
3022}
3023KSM_ATTR_RO(pages_volatile);
3024
3025static ssize_t stable_node_dups_show(struct kobject *kobj,
3026 struct kobj_attribute *attr, char *buf)
3027{
3028 return sprintf(buf, "%lu\n", ksm_stable_node_dups);
3029}
3030KSM_ATTR_RO(stable_node_dups);
3031
3032static ssize_t stable_node_chains_show(struct kobject *kobj,
3033 struct kobj_attribute *attr, char *buf)
3034{
3035 return sprintf(buf, "%lu\n", ksm_stable_node_chains);
3036}
3037KSM_ATTR_RO(stable_node_chains);
3038
3039static ssize_t
3040stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3041 struct kobj_attribute *attr,
3042 char *buf)
3043{
3044 return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3045}
3046
3047static ssize_t
3048stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3049 struct kobj_attribute *attr,
3050 const char *buf, size_t count)
3051{
3052 unsigned long msecs;
3053 int err;
3054
3055 err = kstrtoul(buf, 10, &msecs);
3056 if (err || msecs > UINT_MAX)
3057 return -EINVAL;
3058
3059 ksm_stable_node_chains_prune_millisecs = msecs;
3060
3061 return count;
3062}
3063KSM_ATTR(stable_node_chains_prune_millisecs);
3064
3065static ssize_t full_scans_show(struct kobject *kobj,
3066 struct kobj_attribute *attr, char *buf)
3067{
3068 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3069}
3070KSM_ATTR_RO(full_scans);
3071
3072static struct attribute *ksm_attrs[] = {
3073 &sleep_millisecs_attr.attr,
3074 &pages_to_scan_attr.attr,
3075 &run_attr.attr,
3076 &pages_shared_attr.attr,
3077 &pages_sharing_attr.attr,
3078 &pages_unshared_attr.attr,
3079 &pages_volatile_attr.attr,
3080 &full_scans_attr.attr,
3081#ifdef CONFIG_NUMA
3082 &merge_across_nodes_attr.attr,
3083#endif
3084 &max_page_sharing_attr.attr,
3085 &stable_node_chains_attr.attr,
3086 &stable_node_dups_attr.attr,
3087 &stable_node_chains_prune_millisecs_attr.attr,
3088 &use_zero_pages_attr.attr,
3089 NULL,
3090};
3091
3092static const struct attribute_group ksm_attr_group = {
3093 .attrs = ksm_attrs,
3094 .name = "ksm",
3095};
3096#endif /* CONFIG_SYSFS */
3097
3098static int __init ksm_init(void)
3099{
3100 struct task_struct *ksm_thread;
3101 int err;
3102
3103 /* The correct value depends on page size and endianness */
3104 zero_checksum = calc_checksum(ZERO_PAGE(0));
3105 /* Default to false for backwards compatibility */
3106 ksm_use_zero_pages = false;
3107
3108 err = ksm_slab_init();
3109 if (err)
3110 goto out;
3111
3112 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3113 if (IS_ERR(ksm_thread)) {
3114 pr_err("ksm: creating kthread failed\n");
3115 err = PTR_ERR(ksm_thread);
3116 goto out_free;
3117 }
3118
3119#ifdef CONFIG_SYSFS
3120 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3121 if (err) {
3122 pr_err("ksm: register sysfs failed\n");
3123 kthread_stop(ksm_thread);
3124 goto out_free;
3125 }
3126#else
3127 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
3128
3129#endif /* CONFIG_SYSFS */
3130
3131#ifdef CONFIG_MEMORY_HOTREMOVE
3132 /* There is no significance to this priority 100 */
3133 hotplug_memory_notifier(ksm_memory_callback, 100);
3134#endif
3135 return 0;
3136
3137out_free:
3138 ksm_slab_free();
3139out:
3140 return err;
3141}
3142subsys_initcall(ksm_init);
1/*
2 * Memory merging support.
3 *
4 * This code enables dynamic sharing of identical pages found in different
5 * memory areas, even if they are not shared by fork()
6 *
7 * Copyright (C) 2008-2009 Red Hat, Inc.
8 * Authors:
9 * Izik Eidus
10 * Andrea Arcangeli
11 * Chris Wright
12 * Hugh Dickins
13 *
14 * This work is licensed under the terms of the GNU GPL, version 2.
15 */
16
17#include <linux/errno.h>
18#include <linux/mm.h>
19#include <linux/fs.h>
20#include <linux/mman.h>
21#include <linux/sched.h>
22#include <linux/rwsem.h>
23#include <linux/pagemap.h>
24#include <linux/rmap.h>
25#include <linux/spinlock.h>
26#include <linux/jhash.h>
27#include <linux/delay.h>
28#include <linux/kthread.h>
29#include <linux/wait.h>
30#include <linux/slab.h>
31#include <linux/rbtree.h>
32#include <linux/memory.h>
33#include <linux/mmu_notifier.h>
34#include <linux/swap.h>
35#include <linux/ksm.h>
36#include <linux/hashtable.h>
37#include <linux/freezer.h>
38#include <linux/oom.h>
39#include <linux/numa.h>
40
41#include <asm/tlbflush.h>
42#include "internal.h"
43
44#ifdef CONFIG_NUMA
45#define NUMA(x) (x)
46#define DO_NUMA(x) do { (x); } while (0)
47#else
48#define NUMA(x) (0)
49#define DO_NUMA(x) do { } while (0)
50#endif
51
52/*
53 * A few notes about the KSM scanning process,
54 * to make it easier to understand the data structures below:
55 *
56 * In order to reduce excessive scanning, KSM sorts the memory pages by their
57 * contents into a data structure that holds pointers to the pages' locations.
58 *
59 * Since the contents of the pages may change at any moment, KSM cannot just
60 * insert the pages into a normal sorted tree and expect it to find anything.
61 * Therefore KSM uses two data structures - the stable and the unstable tree.
62 *
63 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
64 * by their contents. Because each such page is write-protected, searching on
65 * this tree is fully assured to be working (except when pages are unmapped),
66 * and therefore this tree is called the stable tree.
67 *
68 * In addition to the stable tree, KSM uses a second data structure called the
69 * unstable tree: this tree holds pointers to pages which have been found to
70 * be "unchanged for a period of time". The unstable tree sorts these pages
71 * by their contents, but since they are not write-protected, KSM cannot rely
72 * upon the unstable tree to work correctly - the unstable tree is liable to
73 * be corrupted as its contents are modified, and so it is called unstable.
74 *
75 * KSM solves this problem by several techniques:
76 *
77 * 1) The unstable tree is flushed every time KSM completes scanning all
78 * memory areas, and then the tree is rebuilt again from the beginning.
79 * 2) KSM will only insert into the unstable tree, pages whose hash value
80 * has not changed since the previous scan of all memory areas.
81 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
82 * colors of the nodes and not on their contents, assuring that even when
83 * the tree gets "corrupted" it won't get out of balance, so scanning time
84 * remains the same (also, searching and inserting nodes in an rbtree uses
85 * the same algorithm, so we have no overhead when we flush and rebuild).
86 * 4) KSM never flushes the stable tree, which means that even if it were to
87 * take 10 attempts to find a page in the unstable tree, once it is found,
88 * it is secured in the stable tree. (When we scan a new page, we first
89 * compare it against the stable tree, and then against the unstable tree.)
90 *
91 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
92 * stable trees and multiple unstable trees: one of each for each NUMA node.
93 */
94
95/**
96 * struct mm_slot - ksm information per mm that is being scanned
97 * @link: link to the mm_slots hash list
98 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
99 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
100 * @mm: the mm that this information is valid for
101 */
102struct mm_slot {
103 struct hlist_node link;
104 struct list_head mm_list;
105 struct rmap_item *rmap_list;
106 struct mm_struct *mm;
107};
108
109/**
110 * struct ksm_scan - cursor for scanning
111 * @mm_slot: the current mm_slot we are scanning
112 * @address: the next address inside that to be scanned
113 * @rmap_list: link to the next rmap to be scanned in the rmap_list
114 * @seqnr: count of completed full scans (needed when removing unstable node)
115 *
116 * There is only the one ksm_scan instance of this cursor structure.
117 */
118struct ksm_scan {
119 struct mm_slot *mm_slot;
120 unsigned long address;
121 struct rmap_item **rmap_list;
122 unsigned long seqnr;
123};
124
125/**
126 * struct stable_node - node of the stable rbtree
127 * @node: rb node of this ksm page in the stable tree
128 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
129 * @list: linked into migrate_nodes, pending placement in the proper node tree
130 * @hlist: hlist head of rmap_items using this ksm page
131 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
132 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
133 */
134struct stable_node {
135 union {
136 struct rb_node node; /* when node of stable tree */
137 struct { /* when listed for migration */
138 struct list_head *head;
139 struct list_head list;
140 };
141 };
142 struct hlist_head hlist;
143 unsigned long kpfn;
144#ifdef CONFIG_NUMA
145 int nid;
146#endif
147};
148
149/**
150 * struct rmap_item - reverse mapping item for virtual addresses
151 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
152 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
153 * @nid: NUMA node id of unstable tree in which linked (may not match page)
154 * @mm: the memory structure this rmap_item is pointing into
155 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
156 * @oldchecksum: previous checksum of the page at that virtual address
157 * @node: rb node of this rmap_item in the unstable tree
158 * @head: pointer to stable_node heading this list in the stable tree
159 * @hlist: link into hlist of rmap_items hanging off that stable_node
160 */
161struct rmap_item {
162 struct rmap_item *rmap_list;
163 union {
164 struct anon_vma *anon_vma; /* when stable */
165#ifdef CONFIG_NUMA
166 int nid; /* when node of unstable tree */
167#endif
168 };
169 struct mm_struct *mm;
170 unsigned long address; /* + low bits used for flags below */
171 unsigned int oldchecksum; /* when unstable */
172 union {
173 struct rb_node node; /* when node of unstable tree */
174 struct { /* when listed from stable tree */
175 struct stable_node *head;
176 struct hlist_node hlist;
177 };
178 };
179};
180
181#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
182#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
183#define STABLE_FLAG 0x200 /* is listed from the stable tree */
184
185/* The stable and unstable tree heads */
186static struct rb_root one_stable_tree[1] = { RB_ROOT };
187static struct rb_root one_unstable_tree[1] = { RB_ROOT };
188static struct rb_root *root_stable_tree = one_stable_tree;
189static struct rb_root *root_unstable_tree = one_unstable_tree;
190
191/* Recently migrated nodes of stable tree, pending proper placement */
192static LIST_HEAD(migrate_nodes);
193
194#define MM_SLOTS_HASH_BITS 10
195static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
196
197static struct mm_slot ksm_mm_head = {
198 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
199};
200static struct ksm_scan ksm_scan = {
201 .mm_slot = &ksm_mm_head,
202};
203
204static struct kmem_cache *rmap_item_cache;
205static struct kmem_cache *stable_node_cache;
206static struct kmem_cache *mm_slot_cache;
207
208/* The number of nodes in the stable tree */
209static unsigned long ksm_pages_shared;
210
211/* The number of page slots additionally sharing those nodes */
212static unsigned long ksm_pages_sharing;
213
214/* The number of nodes in the unstable tree */
215static unsigned long ksm_pages_unshared;
216
217/* The number of rmap_items in use: to calculate pages_volatile */
218static unsigned long ksm_rmap_items;
219
220/* Number of pages ksmd should scan in one batch */
221static unsigned int ksm_thread_pages_to_scan = 100;
222
223/* Milliseconds ksmd should sleep between batches */
224static unsigned int ksm_thread_sleep_millisecs = 20;
225
226#ifdef CONFIG_NUMA
227/* Zeroed when merging across nodes is not allowed */
228static unsigned int ksm_merge_across_nodes = 1;
229static int ksm_nr_node_ids = 1;
230#else
231#define ksm_merge_across_nodes 1U
232#define ksm_nr_node_ids 1
233#endif
234
235#define KSM_RUN_STOP 0
236#define KSM_RUN_MERGE 1
237#define KSM_RUN_UNMERGE 2
238#define KSM_RUN_OFFLINE 4
239static unsigned long ksm_run = KSM_RUN_STOP;
240static void wait_while_offlining(void);
241
242static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
243static DEFINE_MUTEX(ksm_thread_mutex);
244static DEFINE_SPINLOCK(ksm_mmlist_lock);
245
246#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
247 sizeof(struct __struct), __alignof__(struct __struct),\
248 (__flags), NULL)
249
250static int __init ksm_slab_init(void)
251{
252 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
253 if (!rmap_item_cache)
254 goto out;
255
256 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
257 if (!stable_node_cache)
258 goto out_free1;
259
260 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
261 if (!mm_slot_cache)
262 goto out_free2;
263
264 return 0;
265
266out_free2:
267 kmem_cache_destroy(stable_node_cache);
268out_free1:
269 kmem_cache_destroy(rmap_item_cache);
270out:
271 return -ENOMEM;
272}
273
274static void __init ksm_slab_free(void)
275{
276 kmem_cache_destroy(mm_slot_cache);
277 kmem_cache_destroy(stable_node_cache);
278 kmem_cache_destroy(rmap_item_cache);
279 mm_slot_cache = NULL;
280}
281
282static inline struct rmap_item *alloc_rmap_item(void)
283{
284 struct rmap_item *rmap_item;
285
286 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
287 __GFP_NORETRY | __GFP_NOWARN);
288 if (rmap_item)
289 ksm_rmap_items++;
290 return rmap_item;
291}
292
293static inline void free_rmap_item(struct rmap_item *rmap_item)
294{
295 ksm_rmap_items--;
296 rmap_item->mm = NULL; /* debug safety */
297 kmem_cache_free(rmap_item_cache, rmap_item);
298}
299
300static inline struct stable_node *alloc_stable_node(void)
301{
302 /*
303 * The allocation can take too long with GFP_KERNEL when memory is under
304 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
305 * grants access to memory reserves, helping to avoid this problem.
306 */
307 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
308}
309
310static inline void free_stable_node(struct stable_node *stable_node)
311{
312 kmem_cache_free(stable_node_cache, stable_node);
313}
314
315static inline struct mm_slot *alloc_mm_slot(void)
316{
317 if (!mm_slot_cache) /* initialization failed */
318 return NULL;
319 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
320}
321
322static inline void free_mm_slot(struct mm_slot *mm_slot)
323{
324 kmem_cache_free(mm_slot_cache, mm_slot);
325}
326
327static struct mm_slot *get_mm_slot(struct mm_struct *mm)
328{
329 struct mm_slot *slot;
330
331 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
332 if (slot->mm == mm)
333 return slot;
334
335 return NULL;
336}
337
338static void insert_to_mm_slots_hash(struct mm_struct *mm,
339 struct mm_slot *mm_slot)
340{
341 mm_slot->mm = mm;
342 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
343}
344
345/*
346 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
347 * page tables after it has passed through ksm_exit() - which, if necessary,
348 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
349 * a special flag: they can just back out as soon as mm_users goes to zero.
350 * ksm_test_exit() is used throughout to make this test for exit: in some
351 * places for correctness, in some places just to avoid unnecessary work.
352 */
353static inline bool ksm_test_exit(struct mm_struct *mm)
354{
355 return atomic_read(&mm->mm_users) == 0;
356}
357
358/*
359 * We use break_ksm to break COW on a ksm page: it's a stripped down
360 *
361 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
362 * put_page(page);
363 *
364 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
365 * in case the application has unmapped and remapped mm,addr meanwhile.
366 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
367 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
368 *
369 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
370 * of the process that owns 'vma'. We also do not want to enforce
371 * protection keys here anyway.
372 */
373static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
374{
375 struct page *page;
376 int ret = 0;
377
378 do {
379 cond_resched();
380 page = follow_page(vma, addr,
381 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
382 if (IS_ERR_OR_NULL(page))
383 break;
384 if (PageKsm(page))
385 ret = handle_mm_fault(vma, addr,
386 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
387 else
388 ret = VM_FAULT_WRITE;
389 put_page(page);
390 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
391 /*
392 * We must loop because handle_mm_fault() may back out if there's
393 * any difficulty e.g. if pte accessed bit gets updated concurrently.
394 *
395 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
396 * COW has been broken, even if the vma does not permit VM_WRITE;
397 * but note that a concurrent fault might break PageKsm for us.
398 *
399 * VM_FAULT_SIGBUS could occur if we race with truncation of the
400 * backing file, which also invalidates anonymous pages: that's
401 * okay, that truncation will have unmapped the PageKsm for us.
402 *
403 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
404 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
405 * current task has TIF_MEMDIE set, and will be OOM killed on return
406 * to user; and ksmd, having no mm, would never be chosen for that.
407 *
408 * But if the mm is in a limited mem_cgroup, then the fault may fail
409 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
410 * even ksmd can fail in this way - though it's usually breaking ksm
411 * just to undo a merge it made a moment before, so unlikely to oom.
412 *
413 * That's a pity: we might therefore have more kernel pages allocated
414 * than we're counting as nodes in the stable tree; but ksm_do_scan
415 * will retry to break_cow on each pass, so should recover the page
416 * in due course. The important thing is to not let VM_MERGEABLE
417 * be cleared while any such pages might remain in the area.
418 */
419 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
420}
421
422static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
423 unsigned long addr)
424{
425 struct vm_area_struct *vma;
426 if (ksm_test_exit(mm))
427 return NULL;
428 vma = find_vma(mm, addr);
429 if (!vma || vma->vm_start > addr)
430 return NULL;
431 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
432 return NULL;
433 return vma;
434}
435
436static void break_cow(struct rmap_item *rmap_item)
437{
438 struct mm_struct *mm = rmap_item->mm;
439 unsigned long addr = rmap_item->address;
440 struct vm_area_struct *vma;
441
442 /*
443 * It is not an accident that whenever we want to break COW
444 * to undo, we also need to drop a reference to the anon_vma.
445 */
446 put_anon_vma(rmap_item->anon_vma);
447
448 down_read(&mm->mmap_sem);
449 vma = find_mergeable_vma(mm, addr);
450 if (vma)
451 break_ksm(vma, addr);
452 up_read(&mm->mmap_sem);
453}
454
455static struct page *get_mergeable_page(struct rmap_item *rmap_item)
456{
457 struct mm_struct *mm = rmap_item->mm;
458 unsigned long addr = rmap_item->address;
459 struct vm_area_struct *vma;
460 struct page *page;
461
462 down_read(&mm->mmap_sem);
463 vma = find_mergeable_vma(mm, addr);
464 if (!vma)
465 goto out;
466
467 page = follow_page(vma, addr, FOLL_GET);
468 if (IS_ERR_OR_NULL(page))
469 goto out;
470 if (PageAnon(page)) {
471 flush_anon_page(vma, page, addr);
472 flush_dcache_page(page);
473 } else {
474 put_page(page);
475out:
476 page = NULL;
477 }
478 up_read(&mm->mmap_sem);
479 return page;
480}
481
482/*
483 * This helper is used for getting right index into array of tree roots.
484 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
485 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
486 * every node has its own stable and unstable tree.
487 */
488static inline int get_kpfn_nid(unsigned long kpfn)
489{
490 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
491}
492
493static void remove_node_from_stable_tree(struct stable_node *stable_node)
494{
495 struct rmap_item *rmap_item;
496
497 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
498 if (rmap_item->hlist.next)
499 ksm_pages_sharing--;
500 else
501 ksm_pages_shared--;
502 put_anon_vma(rmap_item->anon_vma);
503 rmap_item->address &= PAGE_MASK;
504 cond_resched();
505 }
506
507 if (stable_node->head == &migrate_nodes)
508 list_del(&stable_node->list);
509 else
510 rb_erase(&stable_node->node,
511 root_stable_tree + NUMA(stable_node->nid));
512 free_stable_node(stable_node);
513}
514
515/*
516 * get_ksm_page: checks if the page indicated by the stable node
517 * is still its ksm page, despite having held no reference to it.
518 * In which case we can trust the content of the page, and it
519 * returns the gotten page; but if the page has now been zapped,
520 * remove the stale node from the stable tree and return NULL.
521 * But beware, the stable node's page might be being migrated.
522 *
523 * You would expect the stable_node to hold a reference to the ksm page.
524 * But if it increments the page's count, swapping out has to wait for
525 * ksmd to come around again before it can free the page, which may take
526 * seconds or even minutes: much too unresponsive. So instead we use a
527 * "keyhole reference": access to the ksm page from the stable node peeps
528 * out through its keyhole to see if that page still holds the right key,
529 * pointing back to this stable node. This relies on freeing a PageAnon
530 * page to reset its page->mapping to NULL, and relies on no other use of
531 * a page to put something that might look like our key in page->mapping.
532 * is on its way to being freed; but it is an anomaly to bear in mind.
533 */
534static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
535{
536 struct page *page;
537 void *expected_mapping;
538 unsigned long kpfn;
539
540 expected_mapping = (void *)((unsigned long)stable_node |
541 PAGE_MAPPING_KSM);
542again:
543 kpfn = READ_ONCE(stable_node->kpfn);
544 page = pfn_to_page(kpfn);
545
546 /*
547 * page is computed from kpfn, so on most architectures reading
548 * page->mapping is naturally ordered after reading node->kpfn,
549 * but on Alpha we need to be more careful.
550 */
551 smp_read_barrier_depends();
552 if (READ_ONCE(page->mapping) != expected_mapping)
553 goto stale;
554
555 /*
556 * We cannot do anything with the page while its refcount is 0.
557 * Usually 0 means free, or tail of a higher-order page: in which
558 * case this node is no longer referenced, and should be freed;
559 * however, it might mean that the page is under page_freeze_refs().
560 * The __remove_mapping() case is easy, again the node is now stale;
561 * but if page is swapcache in migrate_page_move_mapping(), it might
562 * still be our page, in which case it's essential to keep the node.
563 */
564 while (!get_page_unless_zero(page)) {
565 /*
566 * Another check for page->mapping != expected_mapping would
567 * work here too. We have chosen the !PageSwapCache test to
568 * optimize the common case, when the page is or is about to
569 * be freed: PageSwapCache is cleared (under spin_lock_irq)
570 * in the freeze_refs section of __remove_mapping(); but Anon
571 * page->mapping reset to NULL later, in free_pages_prepare().
572 */
573 if (!PageSwapCache(page))
574 goto stale;
575 cpu_relax();
576 }
577
578 if (READ_ONCE(page->mapping) != expected_mapping) {
579 put_page(page);
580 goto stale;
581 }
582
583 if (lock_it) {
584 lock_page(page);
585 if (READ_ONCE(page->mapping) != expected_mapping) {
586 unlock_page(page);
587 put_page(page);
588 goto stale;
589 }
590 }
591 return page;
592
593stale:
594 /*
595 * We come here from above when page->mapping or !PageSwapCache
596 * suggests that the node is stale; but it might be under migration.
597 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
598 * before checking whether node->kpfn has been changed.
599 */
600 smp_rmb();
601 if (READ_ONCE(stable_node->kpfn) != kpfn)
602 goto again;
603 remove_node_from_stable_tree(stable_node);
604 return NULL;
605}
606
607/*
608 * Removing rmap_item from stable or unstable tree.
609 * This function will clean the information from the stable/unstable tree.
610 */
611static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
612{
613 if (rmap_item->address & STABLE_FLAG) {
614 struct stable_node *stable_node;
615 struct page *page;
616
617 stable_node = rmap_item->head;
618 page = get_ksm_page(stable_node, true);
619 if (!page)
620 goto out;
621
622 hlist_del(&rmap_item->hlist);
623 unlock_page(page);
624 put_page(page);
625
626 if (!hlist_empty(&stable_node->hlist))
627 ksm_pages_sharing--;
628 else
629 ksm_pages_shared--;
630
631 put_anon_vma(rmap_item->anon_vma);
632 rmap_item->address &= PAGE_MASK;
633
634 } else if (rmap_item->address & UNSTABLE_FLAG) {
635 unsigned char age;
636 /*
637 * Usually ksmd can and must skip the rb_erase, because
638 * root_unstable_tree was already reset to RB_ROOT.
639 * But be careful when an mm is exiting: do the rb_erase
640 * if this rmap_item was inserted by this scan, rather
641 * than left over from before.
642 */
643 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
644 BUG_ON(age > 1);
645 if (!age)
646 rb_erase(&rmap_item->node,
647 root_unstable_tree + NUMA(rmap_item->nid));
648 ksm_pages_unshared--;
649 rmap_item->address &= PAGE_MASK;
650 }
651out:
652 cond_resched(); /* we're called from many long loops */
653}
654
655static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
656 struct rmap_item **rmap_list)
657{
658 while (*rmap_list) {
659 struct rmap_item *rmap_item = *rmap_list;
660 *rmap_list = rmap_item->rmap_list;
661 remove_rmap_item_from_tree(rmap_item);
662 free_rmap_item(rmap_item);
663 }
664}
665
666/*
667 * Though it's very tempting to unmerge rmap_items from stable tree rather
668 * than check every pte of a given vma, the locking doesn't quite work for
669 * that - an rmap_item is assigned to the stable tree after inserting ksm
670 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
671 * rmap_items from parent to child at fork time (so as not to waste time
672 * if exit comes before the next scan reaches it).
673 *
674 * Similarly, although we'd like to remove rmap_items (so updating counts
675 * and freeing memory) when unmerging an area, it's easier to leave that
676 * to the next pass of ksmd - consider, for example, how ksmd might be
677 * in cmp_and_merge_page on one of the rmap_items we would be removing.
678 */
679static int unmerge_ksm_pages(struct vm_area_struct *vma,
680 unsigned long start, unsigned long end)
681{
682 unsigned long addr;
683 int err = 0;
684
685 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
686 if (ksm_test_exit(vma->vm_mm))
687 break;
688 if (signal_pending(current))
689 err = -ERESTARTSYS;
690 else
691 err = break_ksm(vma, addr);
692 }
693 return err;
694}
695
696#ifdef CONFIG_SYSFS
697/*
698 * Only called through the sysfs control interface:
699 */
700static int remove_stable_node(struct stable_node *stable_node)
701{
702 struct page *page;
703 int err;
704
705 page = get_ksm_page(stable_node, true);
706 if (!page) {
707 /*
708 * get_ksm_page did remove_node_from_stable_tree itself.
709 */
710 return 0;
711 }
712
713 if (WARN_ON_ONCE(page_mapped(page))) {
714 /*
715 * This should not happen: but if it does, just refuse to let
716 * merge_across_nodes be switched - there is no need to panic.
717 */
718 err = -EBUSY;
719 } else {
720 /*
721 * The stable node did not yet appear stale to get_ksm_page(),
722 * since that allows for an unmapped ksm page to be recognized
723 * right up until it is freed; but the node is safe to remove.
724 * This page might be in a pagevec waiting to be freed,
725 * or it might be PageSwapCache (perhaps under writeback),
726 * or it might have been removed from swapcache a moment ago.
727 */
728 set_page_stable_node(page, NULL);
729 remove_node_from_stable_tree(stable_node);
730 err = 0;
731 }
732
733 unlock_page(page);
734 put_page(page);
735 return err;
736}
737
738static int remove_all_stable_nodes(void)
739{
740 struct stable_node *stable_node, *next;
741 int nid;
742 int err = 0;
743
744 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
745 while (root_stable_tree[nid].rb_node) {
746 stable_node = rb_entry(root_stable_tree[nid].rb_node,
747 struct stable_node, node);
748 if (remove_stable_node(stable_node)) {
749 err = -EBUSY;
750 break; /* proceed to next nid */
751 }
752 cond_resched();
753 }
754 }
755 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
756 if (remove_stable_node(stable_node))
757 err = -EBUSY;
758 cond_resched();
759 }
760 return err;
761}
762
763static int unmerge_and_remove_all_rmap_items(void)
764{
765 struct mm_slot *mm_slot;
766 struct mm_struct *mm;
767 struct vm_area_struct *vma;
768 int err = 0;
769
770 spin_lock(&ksm_mmlist_lock);
771 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
772 struct mm_slot, mm_list);
773 spin_unlock(&ksm_mmlist_lock);
774
775 for (mm_slot = ksm_scan.mm_slot;
776 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
777 mm = mm_slot->mm;
778 down_read(&mm->mmap_sem);
779 for (vma = mm->mmap; vma; vma = vma->vm_next) {
780 if (ksm_test_exit(mm))
781 break;
782 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
783 continue;
784 err = unmerge_ksm_pages(vma,
785 vma->vm_start, vma->vm_end);
786 if (err)
787 goto error;
788 }
789
790 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
791 up_read(&mm->mmap_sem);
792
793 spin_lock(&ksm_mmlist_lock);
794 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
795 struct mm_slot, mm_list);
796 if (ksm_test_exit(mm)) {
797 hash_del(&mm_slot->link);
798 list_del(&mm_slot->mm_list);
799 spin_unlock(&ksm_mmlist_lock);
800
801 free_mm_slot(mm_slot);
802 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
803 mmdrop(mm);
804 } else
805 spin_unlock(&ksm_mmlist_lock);
806 }
807
808 /* Clean up stable nodes, but don't worry if some are still busy */
809 remove_all_stable_nodes();
810 ksm_scan.seqnr = 0;
811 return 0;
812
813error:
814 up_read(&mm->mmap_sem);
815 spin_lock(&ksm_mmlist_lock);
816 ksm_scan.mm_slot = &ksm_mm_head;
817 spin_unlock(&ksm_mmlist_lock);
818 return err;
819}
820#endif /* CONFIG_SYSFS */
821
822static u32 calc_checksum(struct page *page)
823{
824 u32 checksum;
825 void *addr = kmap_atomic(page);
826 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
827 kunmap_atomic(addr);
828 return checksum;
829}
830
831static int memcmp_pages(struct page *page1, struct page *page2)
832{
833 char *addr1, *addr2;
834 int ret;
835
836 addr1 = kmap_atomic(page1);
837 addr2 = kmap_atomic(page2);
838 ret = memcmp(addr1, addr2, PAGE_SIZE);
839 kunmap_atomic(addr2);
840 kunmap_atomic(addr1);
841 return ret;
842}
843
844static inline int pages_identical(struct page *page1, struct page *page2)
845{
846 return !memcmp_pages(page1, page2);
847}
848
849static int write_protect_page(struct vm_area_struct *vma, struct page *page,
850 pte_t *orig_pte)
851{
852 struct mm_struct *mm = vma->vm_mm;
853 unsigned long addr;
854 pte_t *ptep;
855 spinlock_t *ptl;
856 int swapped;
857 int err = -EFAULT;
858 unsigned long mmun_start; /* For mmu_notifiers */
859 unsigned long mmun_end; /* For mmu_notifiers */
860
861 addr = page_address_in_vma(page, vma);
862 if (addr == -EFAULT)
863 goto out;
864
865 BUG_ON(PageTransCompound(page));
866
867 mmun_start = addr;
868 mmun_end = addr + PAGE_SIZE;
869 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
870
871 ptep = page_check_address(page, mm, addr, &ptl, 0);
872 if (!ptep)
873 goto out_mn;
874
875 if (pte_write(*ptep) || pte_dirty(*ptep)) {
876 pte_t entry;
877
878 swapped = PageSwapCache(page);
879 flush_cache_page(vma, addr, page_to_pfn(page));
880 /*
881 * Ok this is tricky, when get_user_pages_fast() run it doesn't
882 * take any lock, therefore the check that we are going to make
883 * with the pagecount against the mapcount is racey and
884 * O_DIRECT can happen right after the check.
885 * So we clear the pte and flush the tlb before the check
886 * this assure us that no O_DIRECT can happen after the check
887 * or in the middle of the check.
888 */
889 entry = ptep_clear_flush_notify(vma, addr, ptep);
890 /*
891 * Check that no O_DIRECT or similar I/O is in progress on the
892 * page
893 */
894 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
895 set_pte_at(mm, addr, ptep, entry);
896 goto out_unlock;
897 }
898 if (pte_dirty(entry))
899 set_page_dirty(page);
900 entry = pte_mkclean(pte_wrprotect(entry));
901 set_pte_at_notify(mm, addr, ptep, entry);
902 }
903 *orig_pte = *ptep;
904 err = 0;
905
906out_unlock:
907 pte_unmap_unlock(ptep, ptl);
908out_mn:
909 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
910out:
911 return err;
912}
913
914/**
915 * replace_page - replace page in vma by new ksm page
916 * @vma: vma that holds the pte pointing to page
917 * @page: the page we are replacing by kpage
918 * @kpage: the ksm page we replace page by
919 * @orig_pte: the original value of the pte
920 *
921 * Returns 0 on success, -EFAULT on failure.
922 */
923static int replace_page(struct vm_area_struct *vma, struct page *page,
924 struct page *kpage, pte_t orig_pte)
925{
926 struct mm_struct *mm = vma->vm_mm;
927 pmd_t *pmd;
928 pte_t *ptep;
929 spinlock_t *ptl;
930 unsigned long addr;
931 int err = -EFAULT;
932 unsigned long mmun_start; /* For mmu_notifiers */
933 unsigned long mmun_end; /* For mmu_notifiers */
934
935 addr = page_address_in_vma(page, vma);
936 if (addr == -EFAULT)
937 goto out;
938
939 pmd = mm_find_pmd(mm, addr);
940 if (!pmd)
941 goto out;
942
943 mmun_start = addr;
944 mmun_end = addr + PAGE_SIZE;
945 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
946
947 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
948 if (!pte_same(*ptep, orig_pte)) {
949 pte_unmap_unlock(ptep, ptl);
950 goto out_mn;
951 }
952
953 get_page(kpage);
954 page_add_anon_rmap(kpage, vma, addr, false);
955
956 flush_cache_page(vma, addr, pte_pfn(*ptep));
957 ptep_clear_flush_notify(vma, addr, ptep);
958 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
959
960 page_remove_rmap(page, false);
961 if (!page_mapped(page))
962 try_to_free_swap(page);
963 put_page(page);
964
965 pte_unmap_unlock(ptep, ptl);
966 err = 0;
967out_mn:
968 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
969out:
970 return err;
971}
972
973/*
974 * try_to_merge_one_page - take two pages and merge them into one
975 * @vma: the vma that holds the pte pointing to page
976 * @page: the PageAnon page that we want to replace with kpage
977 * @kpage: the PageKsm page that we want to map instead of page,
978 * or NULL the first time when we want to use page as kpage.
979 *
980 * This function returns 0 if the pages were merged, -EFAULT otherwise.
981 */
982static int try_to_merge_one_page(struct vm_area_struct *vma,
983 struct page *page, struct page *kpage)
984{
985 pte_t orig_pte = __pte(0);
986 int err = -EFAULT;
987
988 if (page == kpage) /* ksm page forked */
989 return 0;
990
991 if (!PageAnon(page))
992 goto out;
993
994 /*
995 * We need the page lock to read a stable PageSwapCache in
996 * write_protect_page(). We use trylock_page() instead of
997 * lock_page() because we don't want to wait here - we
998 * prefer to continue scanning and merging different pages,
999 * 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
1046out_unlock:
1047 unlock_page(page);
1048out:
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 */
1058static 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);
1080out:
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 */
1095static 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 */
1125static 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;
1143again:
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
1213replace:
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 */
1235static 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;
1247again:
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 */
1317static
1318struct 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 */
1389static 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 */
1411static 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
1503static 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
1531static 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;
1588next_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 */
1693static 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
1708static int ksmd_should_run(void)
1709{
1710 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1711}
1712
1713static 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
1738int 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
1785int __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
1824void __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
1862struct 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
1893int 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;
1911again:
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;
1953out:
1954 return ret;
1955}
1956
1957#ifdef CONFIG_MIGRATION
1958void 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
1983static 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
1993static 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
2025static 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
2068static 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
2084static 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
2090static 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}
2105KSM_ATTR(sleep_millisecs);
2106
2107static 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
2113static 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}
2128KSM_ATTR(pages_to_scan);
2129
2130static 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
2136static 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}
2176KSM_ATTR(run);
2177
2178#ifdef CONFIG_NUMA
2179static 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
2185static 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}
2233KSM_ATTR(merge_across_nodes);
2234#endif
2235
2236static 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}
2241KSM_ATTR_RO(pages_shared);
2242
2243static 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}
2248KSM_ATTR_RO(pages_sharing);
2249
2250static 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}
2255KSM_ATTR_RO(pages_unshared);
2256
2257static 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}
2272KSM_ATTR_RO(pages_volatile);
2273
2274static 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}
2279KSM_ATTR_RO(full_scans);
2280
2281static 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
2296static struct attribute_group ksm_attr_group = {
2297 .attrs = ksm_attrs,
2298 .name = "ksm",
2299};
2300#endif /* CONFIG_SYSFS */
2301
2302static 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
2336out_free:
2337 ksm_slab_free();
2338out:
2339 return err;
2340}
2341subsys_initcall(ksm_init);