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/xxhash.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 * DOC: Overview
56 *
57 * A few notes about the KSM scanning process,
58 * to make it easier to understand the data structures below:
59 *
60 * In order to reduce excessive scanning, KSM sorts the memory pages by their
61 * contents into a data structure that holds pointers to the pages' locations.
62 *
63 * Since the contents of the pages may change at any moment, KSM cannot just
64 * insert the pages into a normal sorted tree and expect it to find anything.
65 * Therefore KSM uses two data structures - the stable and the unstable tree.
66 *
67 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
68 * by their contents. Because each such page is write-protected, searching on
69 * this tree is fully assured to be working (except when pages are unmapped),
70 * and therefore this tree is called the stable tree.
71 *
72 * The stable tree node includes information required for reverse
73 * mapping from a KSM page to virtual addresses that map this page.
74 *
75 * In order to avoid large latencies of the rmap walks on KSM pages,
76 * KSM maintains two types of nodes in the stable tree:
77 *
78 * * the regular nodes that keep the reverse mapping structures in a
79 * linked list
80 * * the "chains" that link nodes ("dups") that represent the same
81 * write protected memory content, but each "dup" corresponds to a
82 * different KSM page copy of that content
83 *
84 * Internally, the regular nodes, "dups" and "chains" are represented
85 * using the same :c:type:`struct stable_node` structure.
86 *
87 * In addition to the stable tree, KSM uses a second data structure called the
88 * unstable tree: this tree holds pointers to pages which have been found to
89 * be "unchanged for a period of time". The unstable tree sorts these pages
90 * by their contents, but since they are not write-protected, KSM cannot rely
91 * upon the unstable tree to work correctly - the unstable tree is liable to
92 * be corrupted as its contents are modified, and so it is called unstable.
93 *
94 * KSM solves this problem by several techniques:
95 *
96 * 1) The unstable tree is flushed every time KSM completes scanning all
97 * memory areas, and then the tree is rebuilt again from the beginning.
98 * 2) KSM will only insert into the unstable tree, pages whose hash value
99 * has not changed since the previous scan of all memory areas.
100 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
101 * colors of the nodes and not on their contents, assuring that even when
102 * the tree gets "corrupted" it won't get out of balance, so scanning time
103 * remains the same (also, searching and inserting nodes in an rbtree uses
104 * the same algorithm, so we have no overhead when we flush and rebuild).
105 * 4) KSM never flushes the stable tree, which means that even if it were to
106 * take 10 attempts to find a page in the unstable tree, once it is found,
107 * it is secured in the stable tree. (When we scan a new page, we first
108 * compare it against the stable tree, and then against the unstable tree.)
109 *
110 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
111 * stable trees and multiple unstable trees: one of each for each NUMA node.
112 */
113
114/**
115 * struct mm_slot - ksm information per mm that is being scanned
116 * @link: link to the mm_slots hash list
117 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
118 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
119 * @mm: the mm that this information is valid for
120 */
121struct mm_slot {
122 struct hlist_node link;
123 struct list_head mm_list;
124 struct rmap_item *rmap_list;
125 struct mm_struct *mm;
126};
127
128/**
129 * struct ksm_scan - cursor for scanning
130 * @mm_slot: the current mm_slot we are scanning
131 * @address: the next address inside that to be scanned
132 * @rmap_list: link to the next rmap to be scanned in the rmap_list
133 * @seqnr: count of completed full scans (needed when removing unstable node)
134 *
135 * There is only the one ksm_scan instance of this cursor structure.
136 */
137struct ksm_scan {
138 struct mm_slot *mm_slot;
139 unsigned long address;
140 struct rmap_item **rmap_list;
141 unsigned long seqnr;
142};
143
144/**
145 * struct stable_node - node of the stable rbtree
146 * @node: rb node of this ksm page in the stable tree
147 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
148 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
149 * @list: linked into migrate_nodes, pending placement in the proper node tree
150 * @hlist: hlist head of rmap_items using this ksm page
151 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
152 * @chain_prune_time: time of the last full garbage collection
153 * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
154 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
155 */
156struct stable_node {
157 union {
158 struct rb_node node; /* when node of stable tree */
159 struct { /* when listed for migration */
160 struct list_head *head;
161 struct {
162 struct hlist_node hlist_dup;
163 struct list_head list;
164 };
165 };
166 };
167 struct hlist_head hlist;
168 union {
169 unsigned long kpfn;
170 unsigned long chain_prune_time;
171 };
172 /*
173 * STABLE_NODE_CHAIN can be any negative number in
174 * rmap_hlist_len negative range, but better not -1 to be able
175 * to reliably detect underflows.
176 */
177#define STABLE_NODE_CHAIN -1024
178 int rmap_hlist_len;
179#ifdef CONFIG_NUMA
180 int nid;
181#endif
182};
183
184/**
185 * struct rmap_item - reverse mapping item for virtual addresses
186 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
187 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
188 * @nid: NUMA node id of unstable tree in which linked (may not match page)
189 * @mm: the memory structure this rmap_item is pointing into
190 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
191 * @oldchecksum: previous checksum of the page at that virtual address
192 * @node: rb node of this rmap_item in the unstable tree
193 * @head: pointer to stable_node heading this list in the stable tree
194 * @hlist: link into hlist of rmap_items hanging off that stable_node
195 */
196struct rmap_item {
197 struct rmap_item *rmap_list;
198 union {
199 struct anon_vma *anon_vma; /* when stable */
200#ifdef CONFIG_NUMA
201 int nid; /* when node of unstable tree */
202#endif
203 };
204 struct mm_struct *mm;
205 unsigned long address; /* + low bits used for flags below */
206 unsigned int oldchecksum; /* when unstable */
207 union {
208 struct rb_node node; /* when node of unstable tree */
209 struct { /* when listed from stable tree */
210 struct stable_node *head;
211 struct hlist_node hlist;
212 };
213 };
214};
215
216#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
217#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
218#define STABLE_FLAG 0x200 /* is listed from the stable tree */
219#define KSM_FLAG_MASK (SEQNR_MASK|UNSTABLE_FLAG|STABLE_FLAG)
220 /* to mask all the flags */
221
222/* The stable and unstable tree heads */
223static struct rb_root one_stable_tree[1] = { RB_ROOT };
224static struct rb_root one_unstable_tree[1] = { RB_ROOT };
225static struct rb_root *root_stable_tree = one_stable_tree;
226static struct rb_root *root_unstable_tree = one_unstable_tree;
227
228/* Recently migrated nodes of stable tree, pending proper placement */
229static LIST_HEAD(migrate_nodes);
230#define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
231
232#define MM_SLOTS_HASH_BITS 10
233static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
234
235static struct mm_slot ksm_mm_head = {
236 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
237};
238static struct ksm_scan ksm_scan = {
239 .mm_slot = &ksm_mm_head,
240};
241
242static struct kmem_cache *rmap_item_cache;
243static struct kmem_cache *stable_node_cache;
244static struct kmem_cache *mm_slot_cache;
245
246/* The number of nodes in the stable tree */
247static unsigned long ksm_pages_shared;
248
249/* The number of page slots additionally sharing those nodes */
250static unsigned long ksm_pages_sharing;
251
252/* The number of nodes in the unstable tree */
253static unsigned long ksm_pages_unshared;
254
255/* The number of rmap_items in use: to calculate pages_volatile */
256static unsigned long ksm_rmap_items;
257
258/* The number of stable_node chains */
259static unsigned long ksm_stable_node_chains;
260
261/* The number of stable_node dups linked to the stable_node chains */
262static unsigned long ksm_stable_node_dups;
263
264/* Delay in pruning stale stable_node_dups in the stable_node_chains */
265static int ksm_stable_node_chains_prune_millisecs = 2000;
266
267/* Maximum number of page slots sharing a stable node */
268static int ksm_max_page_sharing = 256;
269
270/* Number of pages ksmd should scan in one batch */
271static unsigned int ksm_thread_pages_to_scan = 100;
272
273/* Milliseconds ksmd should sleep between batches */
274static unsigned int ksm_thread_sleep_millisecs = 20;
275
276/* Checksum of an empty (zeroed) page */
277static unsigned int zero_checksum __read_mostly;
278
279/* Whether to merge empty (zeroed) pages with actual zero pages */
280static bool ksm_use_zero_pages __read_mostly;
281
282#ifdef CONFIG_NUMA
283/* Zeroed when merging across nodes is not allowed */
284static unsigned int ksm_merge_across_nodes = 1;
285static int ksm_nr_node_ids = 1;
286#else
287#define ksm_merge_across_nodes 1U
288#define ksm_nr_node_ids 1
289#endif
290
291#define KSM_RUN_STOP 0
292#define KSM_RUN_MERGE 1
293#define KSM_RUN_UNMERGE 2
294#define KSM_RUN_OFFLINE 4
295static unsigned long ksm_run = KSM_RUN_STOP;
296static void wait_while_offlining(void);
297
298static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
299static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait);
300static DEFINE_MUTEX(ksm_thread_mutex);
301static DEFINE_SPINLOCK(ksm_mmlist_lock);
302
303#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
304 sizeof(struct __struct), __alignof__(struct __struct),\
305 (__flags), NULL)
306
307static int __init ksm_slab_init(void)
308{
309 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
310 if (!rmap_item_cache)
311 goto out;
312
313 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
314 if (!stable_node_cache)
315 goto out_free1;
316
317 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
318 if (!mm_slot_cache)
319 goto out_free2;
320
321 return 0;
322
323out_free2:
324 kmem_cache_destroy(stable_node_cache);
325out_free1:
326 kmem_cache_destroy(rmap_item_cache);
327out:
328 return -ENOMEM;
329}
330
331static void __init ksm_slab_free(void)
332{
333 kmem_cache_destroy(mm_slot_cache);
334 kmem_cache_destroy(stable_node_cache);
335 kmem_cache_destroy(rmap_item_cache);
336 mm_slot_cache = NULL;
337}
338
339static __always_inline bool is_stable_node_chain(struct stable_node *chain)
340{
341 return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
342}
343
344static __always_inline bool is_stable_node_dup(struct stable_node *dup)
345{
346 return dup->head == STABLE_NODE_DUP_HEAD;
347}
348
349static inline void stable_node_chain_add_dup(struct stable_node *dup,
350 struct stable_node *chain)
351{
352 VM_BUG_ON(is_stable_node_dup(dup));
353 dup->head = STABLE_NODE_DUP_HEAD;
354 VM_BUG_ON(!is_stable_node_chain(chain));
355 hlist_add_head(&dup->hlist_dup, &chain->hlist);
356 ksm_stable_node_dups++;
357}
358
359static inline void __stable_node_dup_del(struct stable_node *dup)
360{
361 VM_BUG_ON(!is_stable_node_dup(dup));
362 hlist_del(&dup->hlist_dup);
363 ksm_stable_node_dups--;
364}
365
366static inline void stable_node_dup_del(struct stable_node *dup)
367{
368 VM_BUG_ON(is_stable_node_chain(dup));
369 if (is_stable_node_dup(dup))
370 __stable_node_dup_del(dup);
371 else
372 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
373#ifdef CONFIG_DEBUG_VM
374 dup->head = NULL;
375#endif
376}
377
378static inline struct rmap_item *alloc_rmap_item(void)
379{
380 struct rmap_item *rmap_item;
381
382 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
383 __GFP_NORETRY | __GFP_NOWARN);
384 if (rmap_item)
385 ksm_rmap_items++;
386 return rmap_item;
387}
388
389static inline void free_rmap_item(struct rmap_item *rmap_item)
390{
391 ksm_rmap_items--;
392 rmap_item->mm = NULL; /* debug safety */
393 kmem_cache_free(rmap_item_cache, rmap_item);
394}
395
396static inline struct stable_node *alloc_stable_node(void)
397{
398 /*
399 * The allocation can take too long with GFP_KERNEL when memory is under
400 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
401 * grants access to memory reserves, helping to avoid this problem.
402 */
403 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
404}
405
406static inline void free_stable_node(struct stable_node *stable_node)
407{
408 VM_BUG_ON(stable_node->rmap_hlist_len &&
409 !is_stable_node_chain(stable_node));
410 kmem_cache_free(stable_node_cache, stable_node);
411}
412
413static inline struct mm_slot *alloc_mm_slot(void)
414{
415 if (!mm_slot_cache) /* initialization failed */
416 return NULL;
417 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
418}
419
420static inline void free_mm_slot(struct mm_slot *mm_slot)
421{
422 kmem_cache_free(mm_slot_cache, mm_slot);
423}
424
425static struct mm_slot *get_mm_slot(struct mm_struct *mm)
426{
427 struct mm_slot *slot;
428
429 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
430 if (slot->mm == mm)
431 return slot;
432
433 return NULL;
434}
435
436static void insert_to_mm_slots_hash(struct mm_struct *mm,
437 struct mm_slot *mm_slot)
438{
439 mm_slot->mm = mm;
440 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
441}
442
443/*
444 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
445 * page tables after it has passed through ksm_exit() - which, if necessary,
446 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
447 * a special flag: they can just back out as soon as mm_users goes to zero.
448 * ksm_test_exit() is used throughout to make this test for exit: in some
449 * places for correctness, in some places just to avoid unnecessary work.
450 */
451static inline bool ksm_test_exit(struct mm_struct *mm)
452{
453 return atomic_read(&mm->mm_users) == 0;
454}
455
456/*
457 * We use break_ksm to break COW on a ksm page: it's a stripped down
458 *
459 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
460 * put_page(page);
461 *
462 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
463 * in case the application has unmapped and remapped mm,addr meanwhile.
464 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
465 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
466 *
467 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
468 * of the process that owns 'vma'. We also do not want to enforce
469 * protection keys here anyway.
470 */
471static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
472{
473 struct page *page;
474 vm_fault_t ret = 0;
475
476 do {
477 cond_resched();
478 page = follow_page(vma, addr,
479 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
480 if (IS_ERR_OR_NULL(page))
481 break;
482 if (PageKsm(page))
483 ret = handle_mm_fault(vma, addr,
484 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
485 else
486 ret = VM_FAULT_WRITE;
487 put_page(page);
488 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
489 /*
490 * We must loop because handle_mm_fault() may back out if there's
491 * any difficulty e.g. if pte accessed bit gets updated concurrently.
492 *
493 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
494 * COW has been broken, even if the vma does not permit VM_WRITE;
495 * but note that a concurrent fault might break PageKsm for us.
496 *
497 * VM_FAULT_SIGBUS could occur if we race with truncation of the
498 * backing file, which also invalidates anonymous pages: that's
499 * okay, that truncation will have unmapped the PageKsm for us.
500 *
501 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
502 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
503 * current task has TIF_MEMDIE set, and will be OOM killed on return
504 * to user; and ksmd, having no mm, would never be chosen for that.
505 *
506 * But if the mm is in a limited mem_cgroup, then the fault may fail
507 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
508 * even ksmd can fail in this way - though it's usually breaking ksm
509 * just to undo a merge it made a moment before, so unlikely to oom.
510 *
511 * That's a pity: we might therefore have more kernel pages allocated
512 * than we're counting as nodes in the stable tree; but ksm_do_scan
513 * will retry to break_cow on each pass, so should recover the page
514 * in due course. The important thing is to not let VM_MERGEABLE
515 * be cleared while any such pages might remain in the area.
516 */
517 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
518}
519
520static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
521 unsigned long addr)
522{
523 struct vm_area_struct *vma;
524 if (ksm_test_exit(mm))
525 return NULL;
526 vma = find_vma(mm, addr);
527 if (!vma || vma->vm_start > addr)
528 return NULL;
529 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
530 return NULL;
531 return vma;
532}
533
534static void break_cow(struct rmap_item *rmap_item)
535{
536 struct mm_struct *mm = rmap_item->mm;
537 unsigned long addr = rmap_item->address;
538 struct vm_area_struct *vma;
539
540 /*
541 * It is not an accident that whenever we want to break COW
542 * to undo, we also need to drop a reference to the anon_vma.
543 */
544 put_anon_vma(rmap_item->anon_vma);
545
546 down_read(&mm->mmap_sem);
547 vma = find_mergeable_vma(mm, addr);
548 if (vma)
549 break_ksm(vma, addr);
550 up_read(&mm->mmap_sem);
551}
552
553static struct page *get_mergeable_page(struct rmap_item *rmap_item)
554{
555 struct mm_struct *mm = rmap_item->mm;
556 unsigned long addr = rmap_item->address;
557 struct vm_area_struct *vma;
558 struct page *page;
559
560 down_read(&mm->mmap_sem);
561 vma = find_mergeable_vma(mm, addr);
562 if (!vma)
563 goto out;
564
565 page = follow_page(vma, addr, FOLL_GET);
566 if (IS_ERR_OR_NULL(page))
567 goto out;
568 if (PageAnon(page)) {
569 flush_anon_page(vma, page, addr);
570 flush_dcache_page(page);
571 } else {
572 put_page(page);
573out:
574 page = NULL;
575 }
576 up_read(&mm->mmap_sem);
577 return page;
578}
579
580/*
581 * This helper is used for getting right index into array of tree roots.
582 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
583 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
584 * every node has its own stable and unstable tree.
585 */
586static inline int get_kpfn_nid(unsigned long kpfn)
587{
588 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
589}
590
591static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
592 struct rb_root *root)
593{
594 struct stable_node *chain = alloc_stable_node();
595 VM_BUG_ON(is_stable_node_chain(dup));
596 if (likely(chain)) {
597 INIT_HLIST_HEAD(&chain->hlist);
598 chain->chain_prune_time = jiffies;
599 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
600#if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
601 chain->nid = NUMA_NO_NODE; /* debug */
602#endif
603 ksm_stable_node_chains++;
604
605 /*
606 * Put the stable node chain in the first dimension of
607 * the stable tree and at the same time remove the old
608 * stable node.
609 */
610 rb_replace_node(&dup->node, &chain->node, root);
611
612 /*
613 * Move the old stable node to the second dimension
614 * queued in the hlist_dup. The invariant is that all
615 * dup stable_nodes in the chain->hlist point to pages
616 * that are wrprotected and have the exact same
617 * content.
618 */
619 stable_node_chain_add_dup(dup, chain);
620 }
621 return chain;
622}
623
624static inline void free_stable_node_chain(struct stable_node *chain,
625 struct rb_root *root)
626{
627 rb_erase(&chain->node, root);
628 free_stable_node(chain);
629 ksm_stable_node_chains--;
630}
631
632static void remove_node_from_stable_tree(struct stable_node *stable_node)
633{
634 struct rmap_item *rmap_item;
635
636 /* check it's not STABLE_NODE_CHAIN or negative */
637 BUG_ON(stable_node->rmap_hlist_len < 0);
638
639 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
640 if (rmap_item->hlist.next)
641 ksm_pages_sharing--;
642 else
643 ksm_pages_shared--;
644 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
645 stable_node->rmap_hlist_len--;
646 put_anon_vma(rmap_item->anon_vma);
647 rmap_item->address &= PAGE_MASK;
648 cond_resched();
649 }
650
651 /*
652 * We need the second aligned pointer of the migrate_nodes
653 * list_head to stay clear from the rb_parent_color union
654 * (aligned and different than any node) and also different
655 * from &migrate_nodes. This will verify that future list.h changes
656 * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it.
657 */
658#if defined(GCC_VERSION) && GCC_VERSION >= 40903
659 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
660 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
661#endif
662
663 if (stable_node->head == &migrate_nodes)
664 list_del(&stable_node->list);
665 else
666 stable_node_dup_del(stable_node);
667 free_stable_node(stable_node);
668}
669
670enum get_ksm_page_flags {
671 GET_KSM_PAGE_NOLOCK,
672 GET_KSM_PAGE_LOCK,
673 GET_KSM_PAGE_TRYLOCK
674};
675
676/*
677 * get_ksm_page: checks if the page indicated by the stable node
678 * is still its ksm page, despite having held no reference to it.
679 * In which case we can trust the content of the page, and it
680 * returns the gotten page; but if the page has now been zapped,
681 * remove the stale node from the stable tree and return NULL.
682 * But beware, the stable node's page might be being migrated.
683 *
684 * You would expect the stable_node to hold a reference to the ksm page.
685 * But if it increments the page's count, swapping out has to wait for
686 * ksmd to come around again before it can free the page, which may take
687 * seconds or even minutes: much too unresponsive. So instead we use a
688 * "keyhole reference": access to the ksm page from the stable node peeps
689 * out through its keyhole to see if that page still holds the right key,
690 * pointing back to this stable node. This relies on freeing a PageAnon
691 * page to reset its page->mapping to NULL, and relies on no other use of
692 * a page to put something that might look like our key in page->mapping.
693 * is on its way to being freed; but it is an anomaly to bear in mind.
694 */
695static struct page *get_ksm_page(struct stable_node *stable_node,
696 enum get_ksm_page_flags flags)
697{
698 struct page *page;
699 void *expected_mapping;
700 unsigned long kpfn;
701
702 expected_mapping = (void *)((unsigned long)stable_node |
703 PAGE_MAPPING_KSM);
704again:
705 kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
706 page = pfn_to_page(kpfn);
707 if (READ_ONCE(page->mapping) != expected_mapping)
708 goto stale;
709
710 /*
711 * We cannot do anything with the page while its refcount is 0.
712 * Usually 0 means free, or tail of a higher-order page: in which
713 * case this node is no longer referenced, and should be freed;
714 * however, it might mean that the page is under page_ref_freeze().
715 * The __remove_mapping() case is easy, again the node is now stale;
716 * the same is in reuse_ksm_page() case; but if page is swapcache
717 * in migrate_page_move_mapping(), it might still be our page,
718 * in which case it's essential to keep the node.
719 */
720 while (!get_page_unless_zero(page)) {
721 /*
722 * Another check for page->mapping != expected_mapping would
723 * work here too. We have chosen the !PageSwapCache test to
724 * optimize the common case, when the page is or is about to
725 * be freed: PageSwapCache is cleared (under spin_lock_irq)
726 * in the ref_freeze section of __remove_mapping(); but Anon
727 * page->mapping reset to NULL later, in free_pages_prepare().
728 */
729 if (!PageSwapCache(page))
730 goto stale;
731 cpu_relax();
732 }
733
734 if (READ_ONCE(page->mapping) != expected_mapping) {
735 put_page(page);
736 goto stale;
737 }
738
739 if (flags == GET_KSM_PAGE_TRYLOCK) {
740 if (!trylock_page(page)) {
741 put_page(page);
742 return ERR_PTR(-EBUSY);
743 }
744 } else if (flags == GET_KSM_PAGE_LOCK)
745 lock_page(page);
746
747 if (flags != GET_KSM_PAGE_NOLOCK) {
748 if (READ_ONCE(page->mapping) != expected_mapping) {
749 unlock_page(page);
750 put_page(page);
751 goto stale;
752 }
753 }
754 return page;
755
756stale:
757 /*
758 * We come here from above when page->mapping or !PageSwapCache
759 * suggests that the node is stale; but it might be under migration.
760 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
761 * before checking whether node->kpfn has been changed.
762 */
763 smp_rmb();
764 if (READ_ONCE(stable_node->kpfn) != kpfn)
765 goto again;
766 remove_node_from_stable_tree(stable_node);
767 return NULL;
768}
769
770/*
771 * Removing rmap_item from stable or unstable tree.
772 * This function will clean the information from the stable/unstable tree.
773 */
774static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
775{
776 if (rmap_item->address & STABLE_FLAG) {
777 struct stable_node *stable_node;
778 struct page *page;
779
780 stable_node = rmap_item->head;
781 page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
782 if (!page)
783 goto out;
784
785 hlist_del(&rmap_item->hlist);
786 unlock_page(page);
787 put_page(page);
788
789 if (!hlist_empty(&stable_node->hlist))
790 ksm_pages_sharing--;
791 else
792 ksm_pages_shared--;
793 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
794 stable_node->rmap_hlist_len--;
795
796 put_anon_vma(rmap_item->anon_vma);
797 rmap_item->address &= PAGE_MASK;
798
799 } else if (rmap_item->address & UNSTABLE_FLAG) {
800 unsigned char age;
801 /*
802 * Usually ksmd can and must skip the rb_erase, because
803 * root_unstable_tree was already reset to RB_ROOT.
804 * But be careful when an mm is exiting: do the rb_erase
805 * if this rmap_item was inserted by this scan, rather
806 * than left over from before.
807 */
808 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
809 BUG_ON(age > 1);
810 if (!age)
811 rb_erase(&rmap_item->node,
812 root_unstable_tree + NUMA(rmap_item->nid));
813 ksm_pages_unshared--;
814 rmap_item->address &= PAGE_MASK;
815 }
816out:
817 cond_resched(); /* we're called from many long loops */
818}
819
820static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
821 struct rmap_item **rmap_list)
822{
823 while (*rmap_list) {
824 struct rmap_item *rmap_item = *rmap_list;
825 *rmap_list = rmap_item->rmap_list;
826 remove_rmap_item_from_tree(rmap_item);
827 free_rmap_item(rmap_item);
828 }
829}
830
831/*
832 * Though it's very tempting to unmerge rmap_items from stable tree rather
833 * than check every pte of a given vma, the locking doesn't quite work for
834 * that - an rmap_item is assigned to the stable tree after inserting ksm
835 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
836 * rmap_items from parent to child at fork time (so as not to waste time
837 * if exit comes before the next scan reaches it).
838 *
839 * Similarly, although we'd like to remove rmap_items (so updating counts
840 * and freeing memory) when unmerging an area, it's easier to leave that
841 * to the next pass of ksmd - consider, for example, how ksmd might be
842 * in cmp_and_merge_page on one of the rmap_items we would be removing.
843 */
844static int unmerge_ksm_pages(struct vm_area_struct *vma,
845 unsigned long start, unsigned long end)
846{
847 unsigned long addr;
848 int err = 0;
849
850 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
851 if (ksm_test_exit(vma->vm_mm))
852 break;
853 if (signal_pending(current))
854 err = -ERESTARTSYS;
855 else
856 err = break_ksm(vma, addr);
857 }
858 return err;
859}
860
861static inline struct stable_node *page_stable_node(struct page *page)
862{
863 return PageKsm(page) ? page_rmapping(page) : NULL;
864}
865
866static inline void set_page_stable_node(struct page *page,
867 struct stable_node *stable_node)
868{
869 page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
870}
871
872#ifdef CONFIG_SYSFS
873/*
874 * Only called through the sysfs control interface:
875 */
876static int remove_stable_node(struct stable_node *stable_node)
877{
878 struct page *page;
879 int err;
880
881 page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
882 if (!page) {
883 /*
884 * get_ksm_page did remove_node_from_stable_tree itself.
885 */
886 return 0;
887 }
888
889 if (WARN_ON_ONCE(page_mapped(page))) {
890 /*
891 * This should not happen: but if it does, just refuse to let
892 * merge_across_nodes be switched - there is no need to panic.
893 */
894 err = -EBUSY;
895 } else {
896 /*
897 * The stable node did not yet appear stale to get_ksm_page(),
898 * since that allows for an unmapped ksm page to be recognized
899 * right up until it is freed; but the node is safe to remove.
900 * This page might be in a pagevec waiting to be freed,
901 * or it might be PageSwapCache (perhaps under writeback),
902 * or it might have been removed from swapcache a moment ago.
903 */
904 set_page_stable_node(page, NULL);
905 remove_node_from_stable_tree(stable_node);
906 err = 0;
907 }
908
909 unlock_page(page);
910 put_page(page);
911 return err;
912}
913
914static int remove_stable_node_chain(struct stable_node *stable_node,
915 struct rb_root *root)
916{
917 struct stable_node *dup;
918 struct hlist_node *hlist_safe;
919
920 if (!is_stable_node_chain(stable_node)) {
921 VM_BUG_ON(is_stable_node_dup(stable_node));
922 if (remove_stable_node(stable_node))
923 return true;
924 else
925 return false;
926 }
927
928 hlist_for_each_entry_safe(dup, hlist_safe,
929 &stable_node->hlist, hlist_dup) {
930 VM_BUG_ON(!is_stable_node_dup(dup));
931 if (remove_stable_node(dup))
932 return true;
933 }
934 BUG_ON(!hlist_empty(&stable_node->hlist));
935 free_stable_node_chain(stable_node, root);
936 return false;
937}
938
939static int remove_all_stable_nodes(void)
940{
941 struct stable_node *stable_node, *next;
942 int nid;
943 int err = 0;
944
945 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
946 while (root_stable_tree[nid].rb_node) {
947 stable_node = rb_entry(root_stable_tree[nid].rb_node,
948 struct stable_node, node);
949 if (remove_stable_node_chain(stable_node,
950 root_stable_tree + nid)) {
951 err = -EBUSY;
952 break; /* proceed to next nid */
953 }
954 cond_resched();
955 }
956 }
957 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
958 if (remove_stable_node(stable_node))
959 err = -EBUSY;
960 cond_resched();
961 }
962 return err;
963}
964
965static int unmerge_and_remove_all_rmap_items(void)
966{
967 struct mm_slot *mm_slot;
968 struct mm_struct *mm;
969 struct vm_area_struct *vma;
970 int err = 0;
971
972 spin_lock(&ksm_mmlist_lock);
973 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
974 struct mm_slot, mm_list);
975 spin_unlock(&ksm_mmlist_lock);
976
977 for (mm_slot = ksm_scan.mm_slot;
978 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
979 mm = mm_slot->mm;
980 down_read(&mm->mmap_sem);
981 for (vma = mm->mmap; vma; vma = vma->vm_next) {
982 if (ksm_test_exit(mm))
983 break;
984 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
985 continue;
986 err = unmerge_ksm_pages(vma,
987 vma->vm_start, vma->vm_end);
988 if (err)
989 goto error;
990 }
991
992 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
993 up_read(&mm->mmap_sem);
994
995 spin_lock(&ksm_mmlist_lock);
996 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
997 struct mm_slot, mm_list);
998 if (ksm_test_exit(mm)) {
999 hash_del(&mm_slot->link);
1000 list_del(&mm_slot->mm_list);
1001 spin_unlock(&ksm_mmlist_lock);
1002
1003 free_mm_slot(mm_slot);
1004 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1005 mmdrop(mm);
1006 } else
1007 spin_unlock(&ksm_mmlist_lock);
1008 }
1009
1010 /* Clean up stable nodes, but don't worry if some are still busy */
1011 remove_all_stable_nodes();
1012 ksm_scan.seqnr = 0;
1013 return 0;
1014
1015error:
1016 up_read(&mm->mmap_sem);
1017 spin_lock(&ksm_mmlist_lock);
1018 ksm_scan.mm_slot = &ksm_mm_head;
1019 spin_unlock(&ksm_mmlist_lock);
1020 return err;
1021}
1022#endif /* CONFIG_SYSFS */
1023
1024static u32 calc_checksum(struct page *page)
1025{
1026 u32 checksum;
1027 void *addr = kmap_atomic(page);
1028 checksum = xxhash(addr, PAGE_SIZE, 0);
1029 kunmap_atomic(addr);
1030 return checksum;
1031}
1032
1033static int memcmp_pages(struct page *page1, struct page *page2)
1034{
1035 char *addr1, *addr2;
1036 int ret;
1037
1038 addr1 = kmap_atomic(page1);
1039 addr2 = kmap_atomic(page2);
1040 ret = memcmp(addr1, addr2, PAGE_SIZE);
1041 kunmap_atomic(addr2);
1042 kunmap_atomic(addr1);
1043 return ret;
1044}
1045
1046static inline int pages_identical(struct page *page1, struct page *page2)
1047{
1048 return !memcmp_pages(page1, page2);
1049}
1050
1051static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1052 pte_t *orig_pte)
1053{
1054 struct mm_struct *mm = vma->vm_mm;
1055 struct page_vma_mapped_walk pvmw = {
1056 .page = page,
1057 .vma = vma,
1058 };
1059 int swapped;
1060 int err = -EFAULT;
1061 struct mmu_notifier_range range;
1062
1063 pvmw.address = page_address_in_vma(page, vma);
1064 if (pvmw.address == -EFAULT)
1065 goto out;
1066
1067 BUG_ON(PageTransCompound(page));
1068
1069 mmu_notifier_range_init(&range, mm, pvmw.address,
1070 pvmw.address + PAGE_SIZE);
1071 mmu_notifier_invalidate_range_start(&range);
1072
1073 if (!page_vma_mapped_walk(&pvmw))
1074 goto out_mn;
1075 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1076 goto out_unlock;
1077
1078 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1079 (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1080 mm_tlb_flush_pending(mm)) {
1081 pte_t entry;
1082
1083 swapped = PageSwapCache(page);
1084 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1085 /*
1086 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1087 * take any lock, therefore the check that we are going to make
1088 * with the pagecount against the mapcount is racey and
1089 * O_DIRECT can happen right after the check.
1090 * So we clear the pte and flush the tlb before the check
1091 * this assure us that no O_DIRECT can happen after the check
1092 * or in the middle of the check.
1093 *
1094 * No need to notify as we are downgrading page table to read
1095 * only not changing it to point to a new page.
1096 *
1097 * See Documentation/vm/mmu_notifier.rst
1098 */
1099 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1100 /*
1101 * Check that no O_DIRECT or similar I/O is in progress on the
1102 * page
1103 */
1104 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1105 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1106 goto out_unlock;
1107 }
1108 if (pte_dirty(entry))
1109 set_page_dirty(page);
1110
1111 if (pte_protnone(entry))
1112 entry = pte_mkclean(pte_clear_savedwrite(entry));
1113 else
1114 entry = pte_mkclean(pte_wrprotect(entry));
1115 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1116 }
1117 *orig_pte = *pvmw.pte;
1118 err = 0;
1119
1120out_unlock:
1121 page_vma_mapped_walk_done(&pvmw);
1122out_mn:
1123 mmu_notifier_invalidate_range_end(&range);
1124out:
1125 return err;
1126}
1127
1128/**
1129 * replace_page - replace page in vma by new ksm page
1130 * @vma: vma that holds the pte pointing to page
1131 * @page: the page we are replacing by kpage
1132 * @kpage: the ksm page we replace page by
1133 * @orig_pte: the original value of the pte
1134 *
1135 * Returns 0 on success, -EFAULT on failure.
1136 */
1137static int replace_page(struct vm_area_struct *vma, struct page *page,
1138 struct page *kpage, pte_t orig_pte)
1139{
1140 struct mm_struct *mm = vma->vm_mm;
1141 pmd_t *pmd;
1142 pte_t *ptep;
1143 pte_t newpte;
1144 spinlock_t *ptl;
1145 unsigned long addr;
1146 int err = -EFAULT;
1147 struct mmu_notifier_range range;
1148
1149 addr = page_address_in_vma(page, vma);
1150 if (addr == -EFAULT)
1151 goto out;
1152
1153 pmd = mm_find_pmd(mm, addr);
1154 if (!pmd)
1155 goto out;
1156
1157 mmu_notifier_range_init(&range, mm, addr, addr + PAGE_SIZE);
1158 mmu_notifier_invalidate_range_start(&range);
1159
1160 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1161 if (!pte_same(*ptep, orig_pte)) {
1162 pte_unmap_unlock(ptep, ptl);
1163 goto out_mn;
1164 }
1165
1166 /*
1167 * No need to check ksm_use_zero_pages here: we can only have a
1168 * zero_page here if ksm_use_zero_pages was enabled alreaady.
1169 */
1170 if (!is_zero_pfn(page_to_pfn(kpage))) {
1171 get_page(kpage);
1172 page_add_anon_rmap(kpage, vma, addr, false);
1173 newpte = mk_pte(kpage, vma->vm_page_prot);
1174 } else {
1175 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1176 vma->vm_page_prot));
1177 /*
1178 * We're replacing an anonymous page with a zero page, which is
1179 * not anonymous. We need to do proper accounting otherwise we
1180 * will get wrong values in /proc, and a BUG message in dmesg
1181 * when tearing down the mm.
1182 */
1183 dec_mm_counter(mm, MM_ANONPAGES);
1184 }
1185
1186 flush_cache_page(vma, addr, pte_pfn(*ptep));
1187 /*
1188 * No need to notify as we are replacing a read only page with another
1189 * read only page with the same content.
1190 *
1191 * See Documentation/vm/mmu_notifier.rst
1192 */
1193 ptep_clear_flush(vma, addr, ptep);
1194 set_pte_at_notify(mm, addr, ptep, newpte);
1195
1196 page_remove_rmap(page, false);
1197 if (!page_mapped(page))
1198 try_to_free_swap(page);
1199 put_page(page);
1200
1201 pte_unmap_unlock(ptep, ptl);
1202 err = 0;
1203out_mn:
1204 mmu_notifier_invalidate_range_end(&range);
1205out:
1206 return err;
1207}
1208
1209/*
1210 * try_to_merge_one_page - take two pages and merge them into one
1211 * @vma: the vma that holds the pte pointing to page
1212 * @page: the PageAnon page that we want to replace with kpage
1213 * @kpage: the PageKsm page that we want to map instead of page,
1214 * or NULL the first time when we want to use page as kpage.
1215 *
1216 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1217 */
1218static int try_to_merge_one_page(struct vm_area_struct *vma,
1219 struct page *page, struct page *kpage)
1220{
1221 pte_t orig_pte = __pte(0);
1222 int err = -EFAULT;
1223
1224 if (page == kpage) /* ksm page forked */
1225 return 0;
1226
1227 if (!PageAnon(page))
1228 goto out;
1229
1230 /*
1231 * We need the page lock to read a stable PageSwapCache in
1232 * write_protect_page(). We use trylock_page() instead of
1233 * lock_page() because we don't want to wait here - we
1234 * prefer to continue scanning and merging different pages,
1235 * then come back to this page when it is unlocked.
1236 */
1237 if (!trylock_page(page))
1238 goto out;
1239
1240 if (PageTransCompound(page)) {
1241 if (split_huge_page(page))
1242 goto out_unlock;
1243 }
1244
1245 /*
1246 * If this anonymous page is mapped only here, its pte may need
1247 * to be write-protected. If it's mapped elsewhere, all of its
1248 * ptes are necessarily already write-protected. But in either
1249 * case, we need to lock and check page_count is not raised.
1250 */
1251 if (write_protect_page(vma, page, &orig_pte) == 0) {
1252 if (!kpage) {
1253 /*
1254 * While we hold page lock, upgrade page from
1255 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1256 * stable_tree_insert() will update stable_node.
1257 */
1258 set_page_stable_node(page, NULL);
1259 mark_page_accessed(page);
1260 /*
1261 * Page reclaim just frees a clean page with no dirty
1262 * ptes: make sure that the ksm page would be swapped.
1263 */
1264 if (!PageDirty(page))
1265 SetPageDirty(page);
1266 err = 0;
1267 } else if (pages_identical(page, kpage))
1268 err = replace_page(vma, page, kpage, orig_pte);
1269 }
1270
1271 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1272 munlock_vma_page(page);
1273 if (!PageMlocked(kpage)) {
1274 unlock_page(page);
1275 lock_page(kpage);
1276 mlock_vma_page(kpage);
1277 page = kpage; /* for final unlock */
1278 }
1279 }
1280
1281out_unlock:
1282 unlock_page(page);
1283out:
1284 return err;
1285}
1286
1287/*
1288 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1289 * but no new kernel page is allocated: kpage must already be a ksm page.
1290 *
1291 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1292 */
1293static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1294 struct page *page, struct page *kpage)
1295{
1296 struct mm_struct *mm = rmap_item->mm;
1297 struct vm_area_struct *vma;
1298 int err = -EFAULT;
1299
1300 down_read(&mm->mmap_sem);
1301 vma = find_mergeable_vma(mm, rmap_item->address);
1302 if (!vma)
1303 goto out;
1304
1305 err = try_to_merge_one_page(vma, page, kpage);
1306 if (err)
1307 goto out;
1308
1309 /* Unstable nid is in union with stable anon_vma: remove first */
1310 remove_rmap_item_from_tree(rmap_item);
1311
1312 /* Must get reference to anon_vma while still holding mmap_sem */
1313 rmap_item->anon_vma = vma->anon_vma;
1314 get_anon_vma(vma->anon_vma);
1315out:
1316 up_read(&mm->mmap_sem);
1317 return err;
1318}
1319
1320/*
1321 * try_to_merge_two_pages - take two identical pages and prepare them
1322 * to be merged into one page.
1323 *
1324 * This function returns the kpage if we successfully merged two identical
1325 * pages into one ksm page, NULL otherwise.
1326 *
1327 * Note that this function upgrades page to ksm page: if one of the pages
1328 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1329 */
1330static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1331 struct page *page,
1332 struct rmap_item *tree_rmap_item,
1333 struct page *tree_page)
1334{
1335 int err;
1336
1337 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1338 if (!err) {
1339 err = try_to_merge_with_ksm_page(tree_rmap_item,
1340 tree_page, page);
1341 /*
1342 * If that fails, we have a ksm page with only one pte
1343 * pointing to it: so break it.
1344 */
1345 if (err)
1346 break_cow(rmap_item);
1347 }
1348 return err ? NULL : page;
1349}
1350
1351static __always_inline
1352bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1353{
1354 VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1355 /*
1356 * Check that at least one mapping still exists, otherwise
1357 * there's no much point to merge and share with this
1358 * stable_node, as the underlying tree_page of the other
1359 * sharer is going to be freed soon.
1360 */
1361 return stable_node->rmap_hlist_len &&
1362 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1363}
1364
1365static __always_inline
1366bool is_page_sharing_candidate(struct stable_node *stable_node)
1367{
1368 return __is_page_sharing_candidate(stable_node, 0);
1369}
1370
1371static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1372 struct stable_node **_stable_node,
1373 struct rb_root *root,
1374 bool prune_stale_stable_nodes)
1375{
1376 struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1377 struct hlist_node *hlist_safe;
1378 struct page *_tree_page, *tree_page = NULL;
1379 int nr = 0;
1380 int found_rmap_hlist_len;
1381
1382 if (!prune_stale_stable_nodes ||
1383 time_before(jiffies, stable_node->chain_prune_time +
1384 msecs_to_jiffies(
1385 ksm_stable_node_chains_prune_millisecs)))
1386 prune_stale_stable_nodes = false;
1387 else
1388 stable_node->chain_prune_time = jiffies;
1389
1390 hlist_for_each_entry_safe(dup, hlist_safe,
1391 &stable_node->hlist, hlist_dup) {
1392 cond_resched();
1393 /*
1394 * We must walk all stable_node_dup to prune the stale
1395 * stable nodes during lookup.
1396 *
1397 * get_ksm_page can drop the nodes from the
1398 * stable_node->hlist if they point to freed pages
1399 * (that's why we do a _safe walk). The "dup"
1400 * stable_node parameter itself will be freed from
1401 * under us if it returns NULL.
1402 */
1403 _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK);
1404 if (!_tree_page)
1405 continue;
1406 nr += 1;
1407 if (is_page_sharing_candidate(dup)) {
1408 if (!found ||
1409 dup->rmap_hlist_len > found_rmap_hlist_len) {
1410 if (found)
1411 put_page(tree_page);
1412 found = dup;
1413 found_rmap_hlist_len = found->rmap_hlist_len;
1414 tree_page = _tree_page;
1415
1416 /* skip put_page for found dup */
1417 if (!prune_stale_stable_nodes)
1418 break;
1419 continue;
1420 }
1421 }
1422 put_page(_tree_page);
1423 }
1424
1425 if (found) {
1426 /*
1427 * nr is counting all dups in the chain only if
1428 * prune_stale_stable_nodes is true, otherwise we may
1429 * break the loop at nr == 1 even if there are
1430 * multiple entries.
1431 */
1432 if (prune_stale_stable_nodes && nr == 1) {
1433 /*
1434 * If there's not just one entry it would
1435 * corrupt memory, better BUG_ON. In KSM
1436 * context with no lock held it's not even
1437 * fatal.
1438 */
1439 BUG_ON(stable_node->hlist.first->next);
1440
1441 /*
1442 * There's just one entry and it is below the
1443 * deduplication limit so drop the chain.
1444 */
1445 rb_replace_node(&stable_node->node, &found->node,
1446 root);
1447 free_stable_node(stable_node);
1448 ksm_stable_node_chains--;
1449 ksm_stable_node_dups--;
1450 /*
1451 * NOTE: the caller depends on the stable_node
1452 * to be equal to stable_node_dup if the chain
1453 * was collapsed.
1454 */
1455 *_stable_node = found;
1456 /*
1457 * Just for robustneess as stable_node is
1458 * otherwise left as a stable pointer, the
1459 * compiler shall optimize it away at build
1460 * time.
1461 */
1462 stable_node = NULL;
1463 } else if (stable_node->hlist.first != &found->hlist_dup &&
1464 __is_page_sharing_candidate(found, 1)) {
1465 /*
1466 * If the found stable_node dup can accept one
1467 * more future merge (in addition to the one
1468 * that is underway) and is not at the head of
1469 * the chain, put it there so next search will
1470 * be quicker in the !prune_stale_stable_nodes
1471 * case.
1472 *
1473 * NOTE: it would be inaccurate to use nr > 1
1474 * instead of checking the hlist.first pointer
1475 * directly, because in the
1476 * prune_stale_stable_nodes case "nr" isn't
1477 * the position of the found dup in the chain,
1478 * but the total number of dups in the chain.
1479 */
1480 hlist_del(&found->hlist_dup);
1481 hlist_add_head(&found->hlist_dup,
1482 &stable_node->hlist);
1483 }
1484 }
1485
1486 *_stable_node_dup = found;
1487 return tree_page;
1488}
1489
1490static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1491 struct rb_root *root)
1492{
1493 if (!is_stable_node_chain(stable_node))
1494 return stable_node;
1495 if (hlist_empty(&stable_node->hlist)) {
1496 free_stable_node_chain(stable_node, root);
1497 return NULL;
1498 }
1499 return hlist_entry(stable_node->hlist.first,
1500 typeof(*stable_node), hlist_dup);
1501}
1502
1503/*
1504 * Like for get_ksm_page, this function can free the *_stable_node and
1505 * *_stable_node_dup if the returned tree_page is NULL.
1506 *
1507 * It can also free and overwrite *_stable_node with the found
1508 * stable_node_dup if the chain is collapsed (in which case
1509 * *_stable_node will be equal to *_stable_node_dup like if the chain
1510 * never existed). It's up to the caller to verify tree_page is not
1511 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1512 *
1513 * *_stable_node_dup is really a second output parameter of this
1514 * function and will be overwritten in all cases, the caller doesn't
1515 * need to initialize it.
1516 */
1517static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1518 struct stable_node **_stable_node,
1519 struct rb_root *root,
1520 bool prune_stale_stable_nodes)
1521{
1522 struct stable_node *stable_node = *_stable_node;
1523 if (!is_stable_node_chain(stable_node)) {
1524 if (is_page_sharing_candidate(stable_node)) {
1525 *_stable_node_dup = stable_node;
1526 return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK);
1527 }
1528 /*
1529 * _stable_node_dup set to NULL means the stable_node
1530 * reached the ksm_max_page_sharing limit.
1531 */
1532 *_stable_node_dup = NULL;
1533 return NULL;
1534 }
1535 return stable_node_dup(_stable_node_dup, _stable_node, root,
1536 prune_stale_stable_nodes);
1537}
1538
1539static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1540 struct stable_node **s_n,
1541 struct rb_root *root)
1542{
1543 return __stable_node_chain(s_n_d, s_n, root, true);
1544}
1545
1546static __always_inline struct page *chain(struct stable_node **s_n_d,
1547 struct stable_node *s_n,
1548 struct rb_root *root)
1549{
1550 struct stable_node *old_stable_node = s_n;
1551 struct page *tree_page;
1552
1553 tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1554 /* not pruning dups so s_n cannot have changed */
1555 VM_BUG_ON(s_n != old_stable_node);
1556 return tree_page;
1557}
1558
1559/*
1560 * stable_tree_search - search for page inside the stable tree
1561 *
1562 * This function checks if there is a page inside the stable tree
1563 * with identical content to the page that we are scanning right now.
1564 *
1565 * This function returns the stable tree node of identical content if found,
1566 * NULL otherwise.
1567 */
1568static struct page *stable_tree_search(struct page *page)
1569{
1570 int nid;
1571 struct rb_root *root;
1572 struct rb_node **new;
1573 struct rb_node *parent;
1574 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1575 struct stable_node *page_node;
1576
1577 page_node = page_stable_node(page);
1578 if (page_node && page_node->head != &migrate_nodes) {
1579 /* ksm page forked */
1580 get_page(page);
1581 return page;
1582 }
1583
1584 nid = get_kpfn_nid(page_to_pfn(page));
1585 root = root_stable_tree + nid;
1586again:
1587 new = &root->rb_node;
1588 parent = NULL;
1589
1590 while (*new) {
1591 struct page *tree_page;
1592 int ret;
1593
1594 cond_resched();
1595 stable_node = rb_entry(*new, struct stable_node, node);
1596 stable_node_any = NULL;
1597 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1598 /*
1599 * NOTE: stable_node may have been freed by
1600 * chain_prune() if the returned stable_node_dup is
1601 * not NULL. stable_node_dup may have been inserted in
1602 * the rbtree instead as a regular stable_node (in
1603 * order to collapse the stable_node chain if a single
1604 * stable_node dup was found in it). In such case the
1605 * stable_node is overwritten by the calleee to point
1606 * to the stable_node_dup that was collapsed in the
1607 * stable rbtree and stable_node will be equal to
1608 * stable_node_dup like if the chain never existed.
1609 */
1610 if (!stable_node_dup) {
1611 /*
1612 * Either all stable_node dups were full in
1613 * this stable_node chain, or this chain was
1614 * empty and should be rb_erased.
1615 */
1616 stable_node_any = stable_node_dup_any(stable_node,
1617 root);
1618 if (!stable_node_any) {
1619 /* rb_erase just run */
1620 goto again;
1621 }
1622 /*
1623 * Take any of the stable_node dups page of
1624 * this stable_node chain to let the tree walk
1625 * continue. All KSM pages belonging to the
1626 * stable_node dups in a stable_node chain
1627 * have the same content and they're
1628 * wrprotected at all times. Any will work
1629 * fine to continue the walk.
1630 */
1631 tree_page = get_ksm_page(stable_node_any,
1632 GET_KSM_PAGE_NOLOCK);
1633 }
1634 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1635 if (!tree_page) {
1636 /*
1637 * If we walked over a stale stable_node,
1638 * get_ksm_page() will call rb_erase() and it
1639 * may rebalance the tree from under us. So
1640 * restart the search from scratch. Returning
1641 * NULL would be safe too, but we'd generate
1642 * false negative insertions just because some
1643 * stable_node was stale.
1644 */
1645 goto again;
1646 }
1647
1648 ret = memcmp_pages(page, tree_page);
1649 put_page(tree_page);
1650
1651 parent = *new;
1652 if (ret < 0)
1653 new = &parent->rb_left;
1654 else if (ret > 0)
1655 new = &parent->rb_right;
1656 else {
1657 if (page_node) {
1658 VM_BUG_ON(page_node->head != &migrate_nodes);
1659 /*
1660 * Test if the migrated page should be merged
1661 * into a stable node dup. If the mapcount is
1662 * 1 we can migrate it with another KSM page
1663 * without adding it to the chain.
1664 */
1665 if (page_mapcount(page) > 1)
1666 goto chain_append;
1667 }
1668
1669 if (!stable_node_dup) {
1670 /*
1671 * If the stable_node is a chain and
1672 * we got a payload match in memcmp
1673 * but we cannot merge the scanned
1674 * page in any of the existing
1675 * stable_node dups because they're
1676 * all full, we need to wait the
1677 * scanned page to find itself a match
1678 * in the unstable tree to create a
1679 * brand new KSM page to add later to
1680 * the dups of this stable_node.
1681 */
1682 return NULL;
1683 }
1684
1685 /*
1686 * Lock and unlock the stable_node's page (which
1687 * might already have been migrated) so that page
1688 * migration is sure to notice its raised count.
1689 * It would be more elegant to return stable_node
1690 * than kpage, but that involves more changes.
1691 */
1692 tree_page = get_ksm_page(stable_node_dup,
1693 GET_KSM_PAGE_TRYLOCK);
1694
1695 if (PTR_ERR(tree_page) == -EBUSY)
1696 return ERR_PTR(-EBUSY);
1697
1698 if (unlikely(!tree_page))
1699 /*
1700 * The tree may have been rebalanced,
1701 * so re-evaluate parent and new.
1702 */
1703 goto again;
1704 unlock_page(tree_page);
1705
1706 if (get_kpfn_nid(stable_node_dup->kpfn) !=
1707 NUMA(stable_node_dup->nid)) {
1708 put_page(tree_page);
1709 goto replace;
1710 }
1711 return tree_page;
1712 }
1713 }
1714
1715 if (!page_node)
1716 return NULL;
1717
1718 list_del(&page_node->list);
1719 DO_NUMA(page_node->nid = nid);
1720 rb_link_node(&page_node->node, parent, new);
1721 rb_insert_color(&page_node->node, root);
1722out:
1723 if (is_page_sharing_candidate(page_node)) {
1724 get_page(page);
1725 return page;
1726 } else
1727 return NULL;
1728
1729replace:
1730 /*
1731 * If stable_node was a chain and chain_prune collapsed it,
1732 * stable_node has been updated to be the new regular
1733 * stable_node. A collapse of the chain is indistinguishable
1734 * from the case there was no chain in the stable
1735 * rbtree. Otherwise stable_node is the chain and
1736 * stable_node_dup is the dup to replace.
1737 */
1738 if (stable_node_dup == stable_node) {
1739 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1740 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1741 /* there is no chain */
1742 if (page_node) {
1743 VM_BUG_ON(page_node->head != &migrate_nodes);
1744 list_del(&page_node->list);
1745 DO_NUMA(page_node->nid = nid);
1746 rb_replace_node(&stable_node_dup->node,
1747 &page_node->node,
1748 root);
1749 if (is_page_sharing_candidate(page_node))
1750 get_page(page);
1751 else
1752 page = NULL;
1753 } else {
1754 rb_erase(&stable_node_dup->node, root);
1755 page = NULL;
1756 }
1757 } else {
1758 VM_BUG_ON(!is_stable_node_chain(stable_node));
1759 __stable_node_dup_del(stable_node_dup);
1760 if (page_node) {
1761 VM_BUG_ON(page_node->head != &migrate_nodes);
1762 list_del(&page_node->list);
1763 DO_NUMA(page_node->nid = nid);
1764 stable_node_chain_add_dup(page_node, stable_node);
1765 if (is_page_sharing_candidate(page_node))
1766 get_page(page);
1767 else
1768 page = NULL;
1769 } else {
1770 page = NULL;
1771 }
1772 }
1773 stable_node_dup->head = &migrate_nodes;
1774 list_add(&stable_node_dup->list, stable_node_dup->head);
1775 return page;
1776
1777chain_append:
1778 /* stable_node_dup could be null if it reached the limit */
1779 if (!stable_node_dup)
1780 stable_node_dup = stable_node_any;
1781 /*
1782 * If stable_node was a chain and chain_prune collapsed it,
1783 * stable_node has been updated to be the new regular
1784 * stable_node. A collapse of the chain is indistinguishable
1785 * from the case there was no chain in the stable
1786 * rbtree. Otherwise stable_node is the chain and
1787 * stable_node_dup is the dup to replace.
1788 */
1789 if (stable_node_dup == stable_node) {
1790 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1791 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1792 /* chain is missing so create it */
1793 stable_node = alloc_stable_node_chain(stable_node_dup,
1794 root);
1795 if (!stable_node)
1796 return NULL;
1797 }
1798 /*
1799 * Add this stable_node dup that was
1800 * migrated to the stable_node chain
1801 * of the current nid for this page
1802 * content.
1803 */
1804 VM_BUG_ON(!is_stable_node_chain(stable_node));
1805 VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1806 VM_BUG_ON(page_node->head != &migrate_nodes);
1807 list_del(&page_node->list);
1808 DO_NUMA(page_node->nid = nid);
1809 stable_node_chain_add_dup(page_node, stable_node);
1810 goto out;
1811}
1812
1813/*
1814 * stable_tree_insert - insert stable tree node pointing to new ksm page
1815 * into the stable tree.
1816 *
1817 * This function returns the stable tree node just allocated on success,
1818 * NULL otherwise.
1819 */
1820static struct stable_node *stable_tree_insert(struct page *kpage)
1821{
1822 int nid;
1823 unsigned long kpfn;
1824 struct rb_root *root;
1825 struct rb_node **new;
1826 struct rb_node *parent;
1827 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1828 bool need_chain = false;
1829
1830 kpfn = page_to_pfn(kpage);
1831 nid = get_kpfn_nid(kpfn);
1832 root = root_stable_tree + nid;
1833again:
1834 parent = NULL;
1835 new = &root->rb_node;
1836
1837 while (*new) {
1838 struct page *tree_page;
1839 int ret;
1840
1841 cond_resched();
1842 stable_node = rb_entry(*new, struct stable_node, node);
1843 stable_node_any = NULL;
1844 tree_page = chain(&stable_node_dup, stable_node, root);
1845 if (!stable_node_dup) {
1846 /*
1847 * Either all stable_node dups were full in
1848 * this stable_node chain, or this chain was
1849 * empty and should be rb_erased.
1850 */
1851 stable_node_any = stable_node_dup_any(stable_node,
1852 root);
1853 if (!stable_node_any) {
1854 /* rb_erase just run */
1855 goto again;
1856 }
1857 /*
1858 * Take any of the stable_node dups page of
1859 * this stable_node chain to let the tree walk
1860 * continue. All KSM pages belonging to the
1861 * stable_node dups in a stable_node chain
1862 * have the same content and they're
1863 * wrprotected at all times. Any will work
1864 * fine to continue the walk.
1865 */
1866 tree_page = get_ksm_page(stable_node_any,
1867 GET_KSM_PAGE_NOLOCK);
1868 }
1869 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1870 if (!tree_page) {
1871 /*
1872 * If we walked over a stale stable_node,
1873 * get_ksm_page() will call rb_erase() and it
1874 * may rebalance the tree from under us. So
1875 * restart the search from scratch. Returning
1876 * NULL would be safe too, but we'd generate
1877 * false negative insertions just because some
1878 * stable_node was stale.
1879 */
1880 goto again;
1881 }
1882
1883 ret = memcmp_pages(kpage, tree_page);
1884 put_page(tree_page);
1885
1886 parent = *new;
1887 if (ret < 0)
1888 new = &parent->rb_left;
1889 else if (ret > 0)
1890 new = &parent->rb_right;
1891 else {
1892 need_chain = true;
1893 break;
1894 }
1895 }
1896
1897 stable_node_dup = alloc_stable_node();
1898 if (!stable_node_dup)
1899 return NULL;
1900
1901 INIT_HLIST_HEAD(&stable_node_dup->hlist);
1902 stable_node_dup->kpfn = kpfn;
1903 set_page_stable_node(kpage, stable_node_dup);
1904 stable_node_dup->rmap_hlist_len = 0;
1905 DO_NUMA(stable_node_dup->nid = nid);
1906 if (!need_chain) {
1907 rb_link_node(&stable_node_dup->node, parent, new);
1908 rb_insert_color(&stable_node_dup->node, root);
1909 } else {
1910 if (!is_stable_node_chain(stable_node)) {
1911 struct stable_node *orig = stable_node;
1912 /* chain is missing so create it */
1913 stable_node = alloc_stable_node_chain(orig, root);
1914 if (!stable_node) {
1915 free_stable_node(stable_node_dup);
1916 return NULL;
1917 }
1918 }
1919 stable_node_chain_add_dup(stable_node_dup, stable_node);
1920 }
1921
1922 return stable_node_dup;
1923}
1924
1925/*
1926 * unstable_tree_search_insert - search for identical page,
1927 * else insert rmap_item into the unstable tree.
1928 *
1929 * This function searches for a page in the unstable tree identical to the
1930 * page currently being scanned; and if no identical page is found in the
1931 * tree, we insert rmap_item as a new object into the unstable tree.
1932 *
1933 * This function returns pointer to rmap_item found to be identical
1934 * to the currently scanned page, NULL otherwise.
1935 *
1936 * This function does both searching and inserting, because they share
1937 * the same walking algorithm in an rbtree.
1938 */
1939static
1940struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1941 struct page *page,
1942 struct page **tree_pagep)
1943{
1944 struct rb_node **new;
1945 struct rb_root *root;
1946 struct rb_node *parent = NULL;
1947 int nid;
1948
1949 nid = get_kpfn_nid(page_to_pfn(page));
1950 root = root_unstable_tree + nid;
1951 new = &root->rb_node;
1952
1953 while (*new) {
1954 struct rmap_item *tree_rmap_item;
1955 struct page *tree_page;
1956 int ret;
1957
1958 cond_resched();
1959 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1960 tree_page = get_mergeable_page(tree_rmap_item);
1961 if (!tree_page)
1962 return NULL;
1963
1964 /*
1965 * Don't substitute a ksm page for a forked page.
1966 */
1967 if (page == tree_page) {
1968 put_page(tree_page);
1969 return NULL;
1970 }
1971
1972 ret = memcmp_pages(page, tree_page);
1973
1974 parent = *new;
1975 if (ret < 0) {
1976 put_page(tree_page);
1977 new = &parent->rb_left;
1978 } else if (ret > 0) {
1979 put_page(tree_page);
1980 new = &parent->rb_right;
1981 } else if (!ksm_merge_across_nodes &&
1982 page_to_nid(tree_page) != nid) {
1983 /*
1984 * If tree_page has been migrated to another NUMA node,
1985 * it will be flushed out and put in the right unstable
1986 * tree next time: only merge with it when across_nodes.
1987 */
1988 put_page(tree_page);
1989 return NULL;
1990 } else {
1991 *tree_pagep = tree_page;
1992 return tree_rmap_item;
1993 }
1994 }
1995
1996 rmap_item->address |= UNSTABLE_FLAG;
1997 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1998 DO_NUMA(rmap_item->nid = nid);
1999 rb_link_node(&rmap_item->node, parent, new);
2000 rb_insert_color(&rmap_item->node, root);
2001
2002 ksm_pages_unshared++;
2003 return NULL;
2004}
2005
2006/*
2007 * stable_tree_append - add another rmap_item to the linked list of
2008 * rmap_items hanging off a given node of the stable tree, all sharing
2009 * the same ksm page.
2010 */
2011static void stable_tree_append(struct rmap_item *rmap_item,
2012 struct stable_node *stable_node,
2013 bool max_page_sharing_bypass)
2014{
2015 /*
2016 * rmap won't find this mapping if we don't insert the
2017 * rmap_item in the right stable_node
2018 * duplicate. page_migration could break later if rmap breaks,
2019 * so we can as well crash here. We really need to check for
2020 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
2021 * for other negative values as an undeflow if detected here
2022 * for the first time (and not when decreasing rmap_hlist_len)
2023 * would be sign of memory corruption in the stable_node.
2024 */
2025 BUG_ON(stable_node->rmap_hlist_len < 0);
2026
2027 stable_node->rmap_hlist_len++;
2028 if (!max_page_sharing_bypass)
2029 /* possibly non fatal but unexpected overflow, only warn */
2030 WARN_ON_ONCE(stable_node->rmap_hlist_len >
2031 ksm_max_page_sharing);
2032
2033 rmap_item->head = stable_node;
2034 rmap_item->address |= STABLE_FLAG;
2035 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2036
2037 if (rmap_item->hlist.next)
2038 ksm_pages_sharing++;
2039 else
2040 ksm_pages_shared++;
2041}
2042
2043/*
2044 * cmp_and_merge_page - first see if page can be merged into the stable tree;
2045 * if not, compare checksum to previous and if it's the same, see if page can
2046 * be inserted into the unstable tree, or merged with a page already there and
2047 * both transferred to the stable tree.
2048 *
2049 * @page: the page that we are searching identical page to.
2050 * @rmap_item: the reverse mapping into the virtual address of this page
2051 */
2052static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2053{
2054 struct mm_struct *mm = rmap_item->mm;
2055 struct rmap_item *tree_rmap_item;
2056 struct page *tree_page = NULL;
2057 struct stable_node *stable_node;
2058 struct page *kpage;
2059 unsigned int checksum;
2060 int err;
2061 bool max_page_sharing_bypass = false;
2062
2063 stable_node = page_stable_node(page);
2064 if (stable_node) {
2065 if (stable_node->head != &migrate_nodes &&
2066 get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2067 NUMA(stable_node->nid)) {
2068 stable_node_dup_del(stable_node);
2069 stable_node->head = &migrate_nodes;
2070 list_add(&stable_node->list, stable_node->head);
2071 }
2072 if (stable_node->head != &migrate_nodes &&
2073 rmap_item->head == stable_node)
2074 return;
2075 /*
2076 * If it's a KSM fork, allow it to go over the sharing limit
2077 * without warnings.
2078 */
2079 if (!is_page_sharing_candidate(stable_node))
2080 max_page_sharing_bypass = true;
2081 }
2082
2083 /* We first start with searching the page inside the stable tree */
2084 kpage = stable_tree_search(page);
2085 if (kpage == page && rmap_item->head == stable_node) {
2086 put_page(kpage);
2087 return;
2088 }
2089
2090 remove_rmap_item_from_tree(rmap_item);
2091
2092 if (kpage) {
2093 if (PTR_ERR(kpage) == -EBUSY)
2094 return;
2095
2096 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2097 if (!err) {
2098 /*
2099 * The page was successfully merged:
2100 * add its rmap_item to the stable tree.
2101 */
2102 lock_page(kpage);
2103 stable_tree_append(rmap_item, page_stable_node(kpage),
2104 max_page_sharing_bypass);
2105 unlock_page(kpage);
2106 }
2107 put_page(kpage);
2108 return;
2109 }
2110
2111 /*
2112 * If the hash value of the page has changed from the last time
2113 * we calculated it, this page is changing frequently: therefore we
2114 * don't want to insert it in the unstable tree, and we don't want
2115 * to waste our time searching for something identical to it there.
2116 */
2117 checksum = calc_checksum(page);
2118 if (rmap_item->oldchecksum != checksum) {
2119 rmap_item->oldchecksum = checksum;
2120 return;
2121 }
2122
2123 /*
2124 * Same checksum as an empty page. We attempt to merge it with the
2125 * appropriate zero page if the user enabled this via sysfs.
2126 */
2127 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2128 struct vm_area_struct *vma;
2129
2130 down_read(&mm->mmap_sem);
2131 vma = find_mergeable_vma(mm, rmap_item->address);
2132 err = try_to_merge_one_page(vma, page,
2133 ZERO_PAGE(rmap_item->address));
2134 up_read(&mm->mmap_sem);
2135 /*
2136 * In case of failure, the page was not really empty, so we
2137 * need to continue. Otherwise we're done.
2138 */
2139 if (!err)
2140 return;
2141 }
2142 tree_rmap_item =
2143 unstable_tree_search_insert(rmap_item, page, &tree_page);
2144 if (tree_rmap_item) {
2145 bool split;
2146
2147 kpage = try_to_merge_two_pages(rmap_item, page,
2148 tree_rmap_item, tree_page);
2149 /*
2150 * If both pages we tried to merge belong to the same compound
2151 * page, then we actually ended up increasing the reference
2152 * count of the same compound page twice, and split_huge_page
2153 * failed.
2154 * Here we set a flag if that happened, and we use it later to
2155 * try split_huge_page again. Since we call put_page right
2156 * afterwards, the reference count will be correct and
2157 * split_huge_page should succeed.
2158 */
2159 split = PageTransCompound(page)
2160 && compound_head(page) == compound_head(tree_page);
2161 put_page(tree_page);
2162 if (kpage) {
2163 /*
2164 * The pages were successfully merged: insert new
2165 * node in the stable tree and add both rmap_items.
2166 */
2167 lock_page(kpage);
2168 stable_node = stable_tree_insert(kpage);
2169 if (stable_node) {
2170 stable_tree_append(tree_rmap_item, stable_node,
2171 false);
2172 stable_tree_append(rmap_item, stable_node,
2173 false);
2174 }
2175 unlock_page(kpage);
2176
2177 /*
2178 * If we fail to insert the page into the stable tree,
2179 * we will have 2 virtual addresses that are pointing
2180 * to a ksm page left outside the stable tree,
2181 * in which case we need to break_cow on both.
2182 */
2183 if (!stable_node) {
2184 break_cow(tree_rmap_item);
2185 break_cow(rmap_item);
2186 }
2187 } else if (split) {
2188 /*
2189 * We are here if we tried to merge two pages and
2190 * failed because they both belonged to the same
2191 * compound page. We will split the page now, but no
2192 * merging will take place.
2193 * We do not want to add the cost of a full lock; if
2194 * the page is locked, it is better to skip it and
2195 * perhaps try again later.
2196 */
2197 if (!trylock_page(page))
2198 return;
2199 split_huge_page(page);
2200 unlock_page(page);
2201 }
2202 }
2203}
2204
2205static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2206 struct rmap_item **rmap_list,
2207 unsigned long addr)
2208{
2209 struct rmap_item *rmap_item;
2210
2211 while (*rmap_list) {
2212 rmap_item = *rmap_list;
2213 if ((rmap_item->address & PAGE_MASK) == addr)
2214 return rmap_item;
2215 if (rmap_item->address > addr)
2216 break;
2217 *rmap_list = rmap_item->rmap_list;
2218 remove_rmap_item_from_tree(rmap_item);
2219 free_rmap_item(rmap_item);
2220 }
2221
2222 rmap_item = alloc_rmap_item();
2223 if (rmap_item) {
2224 /* It has already been zeroed */
2225 rmap_item->mm = mm_slot->mm;
2226 rmap_item->address = addr;
2227 rmap_item->rmap_list = *rmap_list;
2228 *rmap_list = rmap_item;
2229 }
2230 return rmap_item;
2231}
2232
2233static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2234{
2235 struct mm_struct *mm;
2236 struct mm_slot *slot;
2237 struct vm_area_struct *vma;
2238 struct rmap_item *rmap_item;
2239 int nid;
2240
2241 if (list_empty(&ksm_mm_head.mm_list))
2242 return NULL;
2243
2244 slot = ksm_scan.mm_slot;
2245 if (slot == &ksm_mm_head) {
2246 /*
2247 * A number of pages can hang around indefinitely on per-cpu
2248 * pagevecs, raised page count preventing write_protect_page
2249 * from merging them. Though it doesn't really matter much,
2250 * it is puzzling to see some stuck in pages_volatile until
2251 * other activity jostles them out, and they also prevented
2252 * LTP's KSM test from succeeding deterministically; so drain
2253 * them here (here rather than on entry to ksm_do_scan(),
2254 * so we don't IPI too often when pages_to_scan is set low).
2255 */
2256 lru_add_drain_all();
2257
2258 /*
2259 * Whereas stale stable_nodes on the stable_tree itself
2260 * get pruned in the regular course of stable_tree_search(),
2261 * those moved out to the migrate_nodes list can accumulate:
2262 * so prune them once before each full scan.
2263 */
2264 if (!ksm_merge_across_nodes) {
2265 struct stable_node *stable_node, *next;
2266 struct page *page;
2267
2268 list_for_each_entry_safe(stable_node, next,
2269 &migrate_nodes, list) {
2270 page = get_ksm_page(stable_node,
2271 GET_KSM_PAGE_NOLOCK);
2272 if (page)
2273 put_page(page);
2274 cond_resched();
2275 }
2276 }
2277
2278 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2279 root_unstable_tree[nid] = RB_ROOT;
2280
2281 spin_lock(&ksm_mmlist_lock);
2282 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2283 ksm_scan.mm_slot = slot;
2284 spin_unlock(&ksm_mmlist_lock);
2285 /*
2286 * Although we tested list_empty() above, a racing __ksm_exit
2287 * of the last mm on the list may have removed it since then.
2288 */
2289 if (slot == &ksm_mm_head)
2290 return NULL;
2291next_mm:
2292 ksm_scan.address = 0;
2293 ksm_scan.rmap_list = &slot->rmap_list;
2294 }
2295
2296 mm = slot->mm;
2297 down_read(&mm->mmap_sem);
2298 if (ksm_test_exit(mm))
2299 vma = NULL;
2300 else
2301 vma = find_vma(mm, ksm_scan.address);
2302
2303 for (; vma; vma = vma->vm_next) {
2304 if (!(vma->vm_flags & VM_MERGEABLE))
2305 continue;
2306 if (ksm_scan.address < vma->vm_start)
2307 ksm_scan.address = vma->vm_start;
2308 if (!vma->anon_vma)
2309 ksm_scan.address = vma->vm_end;
2310
2311 while (ksm_scan.address < vma->vm_end) {
2312 if (ksm_test_exit(mm))
2313 break;
2314 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2315 if (IS_ERR_OR_NULL(*page)) {
2316 ksm_scan.address += PAGE_SIZE;
2317 cond_resched();
2318 continue;
2319 }
2320 if (PageAnon(*page)) {
2321 flush_anon_page(vma, *page, ksm_scan.address);
2322 flush_dcache_page(*page);
2323 rmap_item = get_next_rmap_item(slot,
2324 ksm_scan.rmap_list, ksm_scan.address);
2325 if (rmap_item) {
2326 ksm_scan.rmap_list =
2327 &rmap_item->rmap_list;
2328 ksm_scan.address += PAGE_SIZE;
2329 } else
2330 put_page(*page);
2331 up_read(&mm->mmap_sem);
2332 return rmap_item;
2333 }
2334 put_page(*page);
2335 ksm_scan.address += PAGE_SIZE;
2336 cond_resched();
2337 }
2338 }
2339
2340 if (ksm_test_exit(mm)) {
2341 ksm_scan.address = 0;
2342 ksm_scan.rmap_list = &slot->rmap_list;
2343 }
2344 /*
2345 * Nuke all the rmap_items that are above this current rmap:
2346 * because there were no VM_MERGEABLE vmas with such addresses.
2347 */
2348 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
2349
2350 spin_lock(&ksm_mmlist_lock);
2351 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2352 struct mm_slot, mm_list);
2353 if (ksm_scan.address == 0) {
2354 /*
2355 * We've completed a full scan of all vmas, holding mmap_sem
2356 * throughout, and found no VM_MERGEABLE: so do the same as
2357 * __ksm_exit does to remove this mm from all our lists now.
2358 * This applies either when cleaning up after __ksm_exit
2359 * (but beware: we can reach here even before __ksm_exit),
2360 * or when all VM_MERGEABLE areas have been unmapped (and
2361 * mmap_sem then protects against race with MADV_MERGEABLE).
2362 */
2363 hash_del(&slot->link);
2364 list_del(&slot->mm_list);
2365 spin_unlock(&ksm_mmlist_lock);
2366
2367 free_mm_slot(slot);
2368 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2369 up_read(&mm->mmap_sem);
2370 mmdrop(mm);
2371 } else {
2372 up_read(&mm->mmap_sem);
2373 /*
2374 * up_read(&mm->mmap_sem) first because after
2375 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2376 * already have been freed under us by __ksm_exit()
2377 * because the "mm_slot" is still hashed and
2378 * ksm_scan.mm_slot doesn't point to it anymore.
2379 */
2380 spin_unlock(&ksm_mmlist_lock);
2381 }
2382
2383 /* Repeat until we've completed scanning the whole list */
2384 slot = ksm_scan.mm_slot;
2385 if (slot != &ksm_mm_head)
2386 goto next_mm;
2387
2388 ksm_scan.seqnr++;
2389 return NULL;
2390}
2391
2392/**
2393 * ksm_do_scan - the ksm scanner main worker function.
2394 * @scan_npages: number of pages we want to scan before we return.
2395 */
2396static void ksm_do_scan(unsigned int scan_npages)
2397{
2398 struct rmap_item *rmap_item;
2399 struct page *uninitialized_var(page);
2400
2401 while (scan_npages-- && likely(!freezing(current))) {
2402 cond_resched();
2403 rmap_item = scan_get_next_rmap_item(&page);
2404 if (!rmap_item)
2405 return;
2406 cmp_and_merge_page(page, rmap_item);
2407 put_page(page);
2408 }
2409}
2410
2411static int ksmd_should_run(void)
2412{
2413 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2414}
2415
2416static int ksm_scan_thread(void *nothing)
2417{
2418 unsigned int sleep_ms;
2419
2420 set_freezable();
2421 set_user_nice(current, 5);
2422
2423 while (!kthread_should_stop()) {
2424 mutex_lock(&ksm_thread_mutex);
2425 wait_while_offlining();
2426 if (ksmd_should_run())
2427 ksm_do_scan(ksm_thread_pages_to_scan);
2428 mutex_unlock(&ksm_thread_mutex);
2429
2430 try_to_freeze();
2431
2432 if (ksmd_should_run()) {
2433 sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs);
2434 wait_event_interruptible_timeout(ksm_iter_wait,
2435 sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs),
2436 msecs_to_jiffies(sleep_ms));
2437 } else {
2438 wait_event_freezable(ksm_thread_wait,
2439 ksmd_should_run() || kthread_should_stop());
2440 }
2441 }
2442 return 0;
2443}
2444
2445int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2446 unsigned long end, int advice, unsigned long *vm_flags)
2447{
2448 struct mm_struct *mm = vma->vm_mm;
2449 int err;
2450
2451 switch (advice) {
2452 case MADV_MERGEABLE:
2453 /*
2454 * Be somewhat over-protective for now!
2455 */
2456 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
2457 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
2458 VM_HUGETLB | VM_MIXEDMAP))
2459 return 0; /* just ignore the advice */
2460
2461 if (vma_is_dax(vma))
2462 return 0;
2463
2464#ifdef VM_SAO
2465 if (*vm_flags & VM_SAO)
2466 return 0;
2467#endif
2468#ifdef VM_SPARC_ADI
2469 if (*vm_flags & VM_SPARC_ADI)
2470 return 0;
2471#endif
2472
2473 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2474 err = __ksm_enter(mm);
2475 if (err)
2476 return err;
2477 }
2478
2479 *vm_flags |= VM_MERGEABLE;
2480 break;
2481
2482 case MADV_UNMERGEABLE:
2483 if (!(*vm_flags & VM_MERGEABLE))
2484 return 0; /* just ignore the advice */
2485
2486 if (vma->anon_vma) {
2487 err = unmerge_ksm_pages(vma, start, end);
2488 if (err)
2489 return err;
2490 }
2491
2492 *vm_flags &= ~VM_MERGEABLE;
2493 break;
2494 }
2495
2496 return 0;
2497}
2498
2499int __ksm_enter(struct mm_struct *mm)
2500{
2501 struct mm_slot *mm_slot;
2502 int needs_wakeup;
2503
2504 mm_slot = alloc_mm_slot();
2505 if (!mm_slot)
2506 return -ENOMEM;
2507
2508 /* Check ksm_run too? Would need tighter locking */
2509 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2510
2511 spin_lock(&ksm_mmlist_lock);
2512 insert_to_mm_slots_hash(mm, mm_slot);
2513 /*
2514 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2515 * insert just behind the scanning cursor, to let the area settle
2516 * down a little; when fork is followed by immediate exec, we don't
2517 * want ksmd to waste time setting up and tearing down an rmap_list.
2518 *
2519 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2520 * scanning cursor, otherwise KSM pages in newly forked mms will be
2521 * missed: then we might as well insert at the end of the list.
2522 */
2523 if (ksm_run & KSM_RUN_UNMERGE)
2524 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2525 else
2526 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2527 spin_unlock(&ksm_mmlist_lock);
2528
2529 set_bit(MMF_VM_MERGEABLE, &mm->flags);
2530 mmgrab(mm);
2531
2532 if (needs_wakeup)
2533 wake_up_interruptible(&ksm_thread_wait);
2534
2535 return 0;
2536}
2537
2538void __ksm_exit(struct mm_struct *mm)
2539{
2540 struct mm_slot *mm_slot;
2541 int easy_to_free = 0;
2542
2543 /*
2544 * This process is exiting: if it's straightforward (as is the
2545 * case when ksmd was never running), free mm_slot immediately.
2546 * But if it's at the cursor or has rmap_items linked to it, use
2547 * mmap_sem to synchronize with any break_cows before pagetables
2548 * are freed, and leave the mm_slot on the list for ksmd to free.
2549 * Beware: ksm may already have noticed it exiting and freed the slot.
2550 */
2551
2552 spin_lock(&ksm_mmlist_lock);
2553 mm_slot = get_mm_slot(mm);
2554 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2555 if (!mm_slot->rmap_list) {
2556 hash_del(&mm_slot->link);
2557 list_del(&mm_slot->mm_list);
2558 easy_to_free = 1;
2559 } else {
2560 list_move(&mm_slot->mm_list,
2561 &ksm_scan.mm_slot->mm_list);
2562 }
2563 }
2564 spin_unlock(&ksm_mmlist_lock);
2565
2566 if (easy_to_free) {
2567 free_mm_slot(mm_slot);
2568 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2569 mmdrop(mm);
2570 } else if (mm_slot) {
2571 down_write(&mm->mmap_sem);
2572 up_write(&mm->mmap_sem);
2573 }
2574}
2575
2576struct page *ksm_might_need_to_copy(struct page *page,
2577 struct vm_area_struct *vma, unsigned long address)
2578{
2579 struct anon_vma *anon_vma = page_anon_vma(page);
2580 struct page *new_page;
2581
2582 if (PageKsm(page)) {
2583 if (page_stable_node(page) &&
2584 !(ksm_run & KSM_RUN_UNMERGE))
2585 return page; /* no need to copy it */
2586 } else if (!anon_vma) {
2587 return page; /* no need to copy it */
2588 } else if (anon_vma->root == vma->anon_vma->root &&
2589 page->index == linear_page_index(vma, address)) {
2590 return page; /* still no need to copy it */
2591 }
2592 if (!PageUptodate(page))
2593 return page; /* let do_swap_page report the error */
2594
2595 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2596 if (new_page) {
2597 copy_user_highpage(new_page, page, address, vma);
2598
2599 SetPageDirty(new_page);
2600 __SetPageUptodate(new_page);
2601 __SetPageLocked(new_page);
2602 }
2603
2604 return new_page;
2605}
2606
2607void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
2608{
2609 struct stable_node *stable_node;
2610 struct rmap_item *rmap_item;
2611 int search_new_forks = 0;
2612
2613 VM_BUG_ON_PAGE(!PageKsm(page), page);
2614
2615 /*
2616 * Rely on the page lock to protect against concurrent modifications
2617 * to that page's node of the stable tree.
2618 */
2619 VM_BUG_ON_PAGE(!PageLocked(page), page);
2620
2621 stable_node = page_stable_node(page);
2622 if (!stable_node)
2623 return;
2624again:
2625 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2626 struct anon_vma *anon_vma = rmap_item->anon_vma;
2627 struct anon_vma_chain *vmac;
2628 struct vm_area_struct *vma;
2629
2630 cond_resched();
2631 anon_vma_lock_read(anon_vma);
2632 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2633 0, ULONG_MAX) {
2634 unsigned long addr;
2635
2636 cond_resched();
2637 vma = vmac->vma;
2638
2639 /* Ignore the stable/unstable/sqnr flags */
2640 addr = rmap_item->address & ~KSM_FLAG_MASK;
2641
2642 if (addr < vma->vm_start || addr >= vma->vm_end)
2643 continue;
2644 /*
2645 * Initially we examine only the vma which covers this
2646 * rmap_item; but later, if there is still work to do,
2647 * we examine covering vmas in other mms: in case they
2648 * were forked from the original since ksmd passed.
2649 */
2650 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2651 continue;
2652
2653 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2654 continue;
2655
2656 if (!rwc->rmap_one(page, vma, addr, rwc->arg)) {
2657 anon_vma_unlock_read(anon_vma);
2658 return;
2659 }
2660 if (rwc->done && rwc->done(page)) {
2661 anon_vma_unlock_read(anon_vma);
2662 return;
2663 }
2664 }
2665 anon_vma_unlock_read(anon_vma);
2666 }
2667 if (!search_new_forks++)
2668 goto again;
2669}
2670
2671bool reuse_ksm_page(struct page *page,
2672 struct vm_area_struct *vma,
2673 unsigned long address)
2674{
2675#ifdef CONFIG_DEBUG_VM
2676 if (WARN_ON(is_zero_pfn(page_to_pfn(page))) ||
2677 WARN_ON(!page_mapped(page)) ||
2678 WARN_ON(!PageLocked(page))) {
2679 dump_page(page, "reuse_ksm_page");
2680 return false;
2681 }
2682#endif
2683
2684 if (PageSwapCache(page) || !page_stable_node(page))
2685 return false;
2686 /* Prohibit parallel get_ksm_page() */
2687 if (!page_ref_freeze(page, 1))
2688 return false;
2689
2690 page_move_anon_rmap(page, vma);
2691 page->index = linear_page_index(vma, address);
2692 page_ref_unfreeze(page, 1);
2693
2694 return true;
2695}
2696#ifdef CONFIG_MIGRATION
2697void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2698{
2699 struct stable_node *stable_node;
2700
2701 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2702 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2703 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2704
2705 stable_node = page_stable_node(newpage);
2706 if (stable_node) {
2707 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2708 stable_node->kpfn = page_to_pfn(newpage);
2709 /*
2710 * newpage->mapping was set in advance; now we need smp_wmb()
2711 * to make sure that the new stable_node->kpfn is visible
2712 * to get_ksm_page() before it can see that oldpage->mapping
2713 * has gone stale (or that PageSwapCache has been cleared).
2714 */
2715 smp_wmb();
2716 set_page_stable_node(oldpage, NULL);
2717 }
2718}
2719#endif /* CONFIG_MIGRATION */
2720
2721#ifdef CONFIG_MEMORY_HOTREMOVE
2722static void wait_while_offlining(void)
2723{
2724 while (ksm_run & KSM_RUN_OFFLINE) {
2725 mutex_unlock(&ksm_thread_mutex);
2726 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2727 TASK_UNINTERRUPTIBLE);
2728 mutex_lock(&ksm_thread_mutex);
2729 }
2730}
2731
2732static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2733 unsigned long start_pfn,
2734 unsigned long end_pfn)
2735{
2736 if (stable_node->kpfn >= start_pfn &&
2737 stable_node->kpfn < end_pfn) {
2738 /*
2739 * Don't get_ksm_page, page has already gone:
2740 * which is why we keep kpfn instead of page*
2741 */
2742 remove_node_from_stable_tree(stable_node);
2743 return true;
2744 }
2745 return false;
2746}
2747
2748static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2749 unsigned long start_pfn,
2750 unsigned long end_pfn,
2751 struct rb_root *root)
2752{
2753 struct stable_node *dup;
2754 struct hlist_node *hlist_safe;
2755
2756 if (!is_stable_node_chain(stable_node)) {
2757 VM_BUG_ON(is_stable_node_dup(stable_node));
2758 return stable_node_dup_remove_range(stable_node, start_pfn,
2759 end_pfn);
2760 }
2761
2762 hlist_for_each_entry_safe(dup, hlist_safe,
2763 &stable_node->hlist, hlist_dup) {
2764 VM_BUG_ON(!is_stable_node_dup(dup));
2765 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2766 }
2767 if (hlist_empty(&stable_node->hlist)) {
2768 free_stable_node_chain(stable_node, root);
2769 return true; /* notify caller that tree was rebalanced */
2770 } else
2771 return false;
2772}
2773
2774static void ksm_check_stable_tree(unsigned long start_pfn,
2775 unsigned long end_pfn)
2776{
2777 struct stable_node *stable_node, *next;
2778 struct rb_node *node;
2779 int nid;
2780
2781 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2782 node = rb_first(root_stable_tree + nid);
2783 while (node) {
2784 stable_node = rb_entry(node, struct stable_node, node);
2785 if (stable_node_chain_remove_range(stable_node,
2786 start_pfn, end_pfn,
2787 root_stable_tree +
2788 nid))
2789 node = rb_first(root_stable_tree + nid);
2790 else
2791 node = rb_next(node);
2792 cond_resched();
2793 }
2794 }
2795 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2796 if (stable_node->kpfn >= start_pfn &&
2797 stable_node->kpfn < end_pfn)
2798 remove_node_from_stable_tree(stable_node);
2799 cond_resched();
2800 }
2801}
2802
2803static int ksm_memory_callback(struct notifier_block *self,
2804 unsigned long action, void *arg)
2805{
2806 struct memory_notify *mn = arg;
2807
2808 switch (action) {
2809 case MEM_GOING_OFFLINE:
2810 /*
2811 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2812 * and remove_all_stable_nodes() while memory is going offline:
2813 * it is unsafe for them to touch the stable tree at this time.
2814 * But unmerge_ksm_pages(), rmap lookups and other entry points
2815 * which do not need the ksm_thread_mutex are all safe.
2816 */
2817 mutex_lock(&ksm_thread_mutex);
2818 ksm_run |= KSM_RUN_OFFLINE;
2819 mutex_unlock(&ksm_thread_mutex);
2820 break;
2821
2822 case MEM_OFFLINE:
2823 /*
2824 * Most of the work is done by page migration; but there might
2825 * be a few stable_nodes left over, still pointing to struct
2826 * pages which have been offlined: prune those from the tree,
2827 * otherwise get_ksm_page() might later try to access a
2828 * non-existent struct page.
2829 */
2830 ksm_check_stable_tree(mn->start_pfn,
2831 mn->start_pfn + mn->nr_pages);
2832 /* fallthrough */
2833
2834 case MEM_CANCEL_OFFLINE:
2835 mutex_lock(&ksm_thread_mutex);
2836 ksm_run &= ~KSM_RUN_OFFLINE;
2837 mutex_unlock(&ksm_thread_mutex);
2838
2839 smp_mb(); /* wake_up_bit advises this */
2840 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2841 break;
2842 }
2843 return NOTIFY_OK;
2844}
2845#else
2846static void wait_while_offlining(void)
2847{
2848}
2849#endif /* CONFIG_MEMORY_HOTREMOVE */
2850
2851#ifdef CONFIG_SYSFS
2852/*
2853 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2854 */
2855
2856#define KSM_ATTR_RO(_name) \
2857 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2858#define KSM_ATTR(_name) \
2859 static struct kobj_attribute _name##_attr = \
2860 __ATTR(_name, 0644, _name##_show, _name##_store)
2861
2862static ssize_t sleep_millisecs_show(struct kobject *kobj,
2863 struct kobj_attribute *attr, char *buf)
2864{
2865 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2866}
2867
2868static ssize_t sleep_millisecs_store(struct kobject *kobj,
2869 struct kobj_attribute *attr,
2870 const char *buf, size_t count)
2871{
2872 unsigned long msecs;
2873 int err;
2874
2875 err = kstrtoul(buf, 10, &msecs);
2876 if (err || msecs > UINT_MAX)
2877 return -EINVAL;
2878
2879 ksm_thread_sleep_millisecs = msecs;
2880 wake_up_interruptible(&ksm_iter_wait);
2881
2882 return count;
2883}
2884KSM_ATTR(sleep_millisecs);
2885
2886static ssize_t pages_to_scan_show(struct kobject *kobj,
2887 struct kobj_attribute *attr, char *buf)
2888{
2889 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2890}
2891
2892static ssize_t pages_to_scan_store(struct kobject *kobj,
2893 struct kobj_attribute *attr,
2894 const char *buf, size_t count)
2895{
2896 int err;
2897 unsigned long nr_pages;
2898
2899 err = kstrtoul(buf, 10, &nr_pages);
2900 if (err || nr_pages > UINT_MAX)
2901 return -EINVAL;
2902
2903 ksm_thread_pages_to_scan = nr_pages;
2904
2905 return count;
2906}
2907KSM_ATTR(pages_to_scan);
2908
2909static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2910 char *buf)
2911{
2912 return sprintf(buf, "%lu\n", ksm_run);
2913}
2914
2915static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2916 const char *buf, size_t count)
2917{
2918 int err;
2919 unsigned long flags;
2920
2921 err = kstrtoul(buf, 10, &flags);
2922 if (err || flags > UINT_MAX)
2923 return -EINVAL;
2924 if (flags > KSM_RUN_UNMERGE)
2925 return -EINVAL;
2926
2927 /*
2928 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2929 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2930 * breaking COW to free the pages_shared (but leaves mm_slots
2931 * on the list for when ksmd may be set running again).
2932 */
2933
2934 mutex_lock(&ksm_thread_mutex);
2935 wait_while_offlining();
2936 if (ksm_run != flags) {
2937 ksm_run = flags;
2938 if (flags & KSM_RUN_UNMERGE) {
2939 set_current_oom_origin();
2940 err = unmerge_and_remove_all_rmap_items();
2941 clear_current_oom_origin();
2942 if (err) {
2943 ksm_run = KSM_RUN_STOP;
2944 count = err;
2945 }
2946 }
2947 }
2948 mutex_unlock(&ksm_thread_mutex);
2949
2950 if (flags & KSM_RUN_MERGE)
2951 wake_up_interruptible(&ksm_thread_wait);
2952
2953 return count;
2954}
2955KSM_ATTR(run);
2956
2957#ifdef CONFIG_NUMA
2958static ssize_t merge_across_nodes_show(struct kobject *kobj,
2959 struct kobj_attribute *attr, char *buf)
2960{
2961 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2962}
2963
2964static ssize_t merge_across_nodes_store(struct kobject *kobj,
2965 struct kobj_attribute *attr,
2966 const char *buf, size_t count)
2967{
2968 int err;
2969 unsigned long knob;
2970
2971 err = kstrtoul(buf, 10, &knob);
2972 if (err)
2973 return err;
2974 if (knob > 1)
2975 return -EINVAL;
2976
2977 mutex_lock(&ksm_thread_mutex);
2978 wait_while_offlining();
2979 if (ksm_merge_across_nodes != knob) {
2980 if (ksm_pages_shared || remove_all_stable_nodes())
2981 err = -EBUSY;
2982 else if (root_stable_tree == one_stable_tree) {
2983 struct rb_root *buf;
2984 /*
2985 * This is the first time that we switch away from the
2986 * default of merging across nodes: must now allocate
2987 * a buffer to hold as many roots as may be needed.
2988 * Allocate stable and unstable together:
2989 * MAXSMP NODES_SHIFT 10 will use 16kB.
2990 */
2991 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2992 GFP_KERNEL);
2993 /* Let us assume that RB_ROOT is NULL is zero */
2994 if (!buf)
2995 err = -ENOMEM;
2996 else {
2997 root_stable_tree = buf;
2998 root_unstable_tree = buf + nr_node_ids;
2999 /* Stable tree is empty but not the unstable */
3000 root_unstable_tree[0] = one_unstable_tree[0];
3001 }
3002 }
3003 if (!err) {
3004 ksm_merge_across_nodes = knob;
3005 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
3006 }
3007 }
3008 mutex_unlock(&ksm_thread_mutex);
3009
3010 return err ? err : count;
3011}
3012KSM_ATTR(merge_across_nodes);
3013#endif
3014
3015static ssize_t use_zero_pages_show(struct kobject *kobj,
3016 struct kobj_attribute *attr, char *buf)
3017{
3018 return sprintf(buf, "%u\n", ksm_use_zero_pages);
3019}
3020static ssize_t use_zero_pages_store(struct kobject *kobj,
3021 struct kobj_attribute *attr,
3022 const char *buf, size_t count)
3023{
3024 int err;
3025 bool value;
3026
3027 err = kstrtobool(buf, &value);
3028 if (err)
3029 return -EINVAL;
3030
3031 ksm_use_zero_pages = value;
3032
3033 return count;
3034}
3035KSM_ATTR(use_zero_pages);
3036
3037static ssize_t max_page_sharing_show(struct kobject *kobj,
3038 struct kobj_attribute *attr, char *buf)
3039{
3040 return sprintf(buf, "%u\n", ksm_max_page_sharing);
3041}
3042
3043static ssize_t max_page_sharing_store(struct kobject *kobj,
3044 struct kobj_attribute *attr,
3045 const char *buf, size_t count)
3046{
3047 int err;
3048 int knob;
3049
3050 err = kstrtoint(buf, 10, &knob);
3051 if (err)
3052 return err;
3053 /*
3054 * When a KSM page is created it is shared by 2 mappings. This
3055 * being a signed comparison, it implicitly verifies it's not
3056 * negative.
3057 */
3058 if (knob < 2)
3059 return -EINVAL;
3060
3061 if (READ_ONCE(ksm_max_page_sharing) == knob)
3062 return count;
3063
3064 mutex_lock(&ksm_thread_mutex);
3065 wait_while_offlining();
3066 if (ksm_max_page_sharing != knob) {
3067 if (ksm_pages_shared || remove_all_stable_nodes())
3068 err = -EBUSY;
3069 else
3070 ksm_max_page_sharing = knob;
3071 }
3072 mutex_unlock(&ksm_thread_mutex);
3073
3074 return err ? err : count;
3075}
3076KSM_ATTR(max_page_sharing);
3077
3078static ssize_t pages_shared_show(struct kobject *kobj,
3079 struct kobj_attribute *attr, char *buf)
3080{
3081 return sprintf(buf, "%lu\n", ksm_pages_shared);
3082}
3083KSM_ATTR_RO(pages_shared);
3084
3085static ssize_t pages_sharing_show(struct kobject *kobj,
3086 struct kobj_attribute *attr, char *buf)
3087{
3088 return sprintf(buf, "%lu\n", ksm_pages_sharing);
3089}
3090KSM_ATTR_RO(pages_sharing);
3091
3092static ssize_t pages_unshared_show(struct kobject *kobj,
3093 struct kobj_attribute *attr, char *buf)
3094{
3095 return sprintf(buf, "%lu\n", ksm_pages_unshared);
3096}
3097KSM_ATTR_RO(pages_unshared);
3098
3099static ssize_t pages_volatile_show(struct kobject *kobj,
3100 struct kobj_attribute *attr, char *buf)
3101{
3102 long ksm_pages_volatile;
3103
3104 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3105 - ksm_pages_sharing - ksm_pages_unshared;
3106 /*
3107 * It was not worth any locking to calculate that statistic,
3108 * but it might therefore sometimes be negative: conceal that.
3109 */
3110 if (ksm_pages_volatile < 0)
3111 ksm_pages_volatile = 0;
3112 return sprintf(buf, "%ld\n", ksm_pages_volatile);
3113}
3114KSM_ATTR_RO(pages_volatile);
3115
3116static ssize_t stable_node_dups_show(struct kobject *kobj,
3117 struct kobj_attribute *attr, char *buf)
3118{
3119 return sprintf(buf, "%lu\n", ksm_stable_node_dups);
3120}
3121KSM_ATTR_RO(stable_node_dups);
3122
3123static ssize_t stable_node_chains_show(struct kobject *kobj,
3124 struct kobj_attribute *attr, char *buf)
3125{
3126 return sprintf(buf, "%lu\n", ksm_stable_node_chains);
3127}
3128KSM_ATTR_RO(stable_node_chains);
3129
3130static ssize_t
3131stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3132 struct kobj_attribute *attr,
3133 char *buf)
3134{
3135 return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3136}
3137
3138static ssize_t
3139stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3140 struct kobj_attribute *attr,
3141 const char *buf, size_t count)
3142{
3143 unsigned long msecs;
3144 int err;
3145
3146 err = kstrtoul(buf, 10, &msecs);
3147 if (err || msecs > UINT_MAX)
3148 return -EINVAL;
3149
3150 ksm_stable_node_chains_prune_millisecs = msecs;
3151
3152 return count;
3153}
3154KSM_ATTR(stable_node_chains_prune_millisecs);
3155
3156static ssize_t full_scans_show(struct kobject *kobj,
3157 struct kobj_attribute *attr, char *buf)
3158{
3159 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3160}
3161KSM_ATTR_RO(full_scans);
3162
3163static struct attribute *ksm_attrs[] = {
3164 &sleep_millisecs_attr.attr,
3165 &pages_to_scan_attr.attr,
3166 &run_attr.attr,
3167 &pages_shared_attr.attr,
3168 &pages_sharing_attr.attr,
3169 &pages_unshared_attr.attr,
3170 &pages_volatile_attr.attr,
3171 &full_scans_attr.attr,
3172#ifdef CONFIG_NUMA
3173 &merge_across_nodes_attr.attr,
3174#endif
3175 &max_page_sharing_attr.attr,
3176 &stable_node_chains_attr.attr,
3177 &stable_node_dups_attr.attr,
3178 &stable_node_chains_prune_millisecs_attr.attr,
3179 &use_zero_pages_attr.attr,
3180 NULL,
3181};
3182
3183static const struct attribute_group ksm_attr_group = {
3184 .attrs = ksm_attrs,
3185 .name = "ksm",
3186};
3187#endif /* CONFIG_SYSFS */
3188
3189static int __init ksm_init(void)
3190{
3191 struct task_struct *ksm_thread;
3192 int err;
3193
3194 /* The correct value depends on page size and endianness */
3195 zero_checksum = calc_checksum(ZERO_PAGE(0));
3196 /* Default to false for backwards compatibility */
3197 ksm_use_zero_pages = false;
3198
3199 err = ksm_slab_init();
3200 if (err)
3201 goto out;
3202
3203 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3204 if (IS_ERR(ksm_thread)) {
3205 pr_err("ksm: creating kthread failed\n");
3206 err = PTR_ERR(ksm_thread);
3207 goto out_free;
3208 }
3209
3210#ifdef CONFIG_SYSFS
3211 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3212 if (err) {
3213 pr_err("ksm: register sysfs failed\n");
3214 kthread_stop(ksm_thread);
3215 goto out_free;
3216 }
3217#else
3218 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
3219
3220#endif /* CONFIG_SYSFS */
3221
3222#ifdef CONFIG_MEMORY_HOTREMOVE
3223 /* There is no significance to this priority 100 */
3224 hotplug_memory_notifier(ksm_memory_callback, 100);
3225#endif
3226 return 0;
3227
3228out_free:
3229 ksm_slab_free();
3230out:
3231 return err;
3232}
3233subsys_initcall(ksm_init);
3234