Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Memory merging support.
4 *
5 * This code enables dynamic sharing of identical pages found in different
6 * memory areas, even if they are not shared by fork()
7 *
8 * Copyright (C) 2008-2009 Red Hat, Inc.
9 * Authors:
10 * Izik Eidus
11 * Andrea Arcangeli
12 * Chris Wright
13 * Hugh Dickins
14 */
15
16#include <linux/errno.h>
17#include <linux/mm.h>
18#include <linux/mm_inline.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#include <linux/pagewalk.h>
43
44#include <asm/tlbflush.h>
45#include "internal.h"
46#include "mm_slot.h"
47
48#ifdef CONFIG_NUMA
49#define NUMA(x) (x)
50#define DO_NUMA(x) do { (x); } while (0)
51#else
52#define NUMA(x) (0)
53#define DO_NUMA(x) do { } while (0)
54#endif
55
56/**
57 * DOC: Overview
58 *
59 * A few notes about the KSM scanning process,
60 * to make it easier to understand the data structures below:
61 *
62 * In order to reduce excessive scanning, KSM sorts the memory pages by their
63 * contents into a data structure that holds pointers to the pages' locations.
64 *
65 * Since the contents of the pages may change at any moment, KSM cannot just
66 * insert the pages into a normal sorted tree and expect it to find anything.
67 * Therefore KSM uses two data structures - the stable and the unstable tree.
68 *
69 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
70 * by their contents. Because each such page is write-protected, searching on
71 * this tree is fully assured to be working (except when pages are unmapped),
72 * and therefore this tree is called the stable tree.
73 *
74 * The stable tree node includes information required for reverse
75 * mapping from a KSM page to virtual addresses that map this page.
76 *
77 * In order to avoid large latencies of the rmap walks on KSM pages,
78 * KSM maintains two types of nodes in the stable tree:
79 *
80 * * the regular nodes that keep the reverse mapping structures in a
81 * linked list
82 * * the "chains" that link nodes ("dups") that represent the same
83 * write protected memory content, but each "dup" corresponds to a
84 * different KSM page copy of that content
85 *
86 * Internally, the regular nodes, "dups" and "chains" are represented
87 * using the same struct ksm_stable_node structure.
88 *
89 * In addition to the stable tree, KSM uses a second data structure called the
90 * unstable tree: this tree holds pointers to pages which have been found to
91 * be "unchanged for a period of time". The unstable tree sorts these pages
92 * by their contents, but since they are not write-protected, KSM cannot rely
93 * upon the unstable tree to work correctly - the unstable tree is liable to
94 * be corrupted as its contents are modified, and so it is called unstable.
95 *
96 * KSM solves this problem by several techniques:
97 *
98 * 1) The unstable tree is flushed every time KSM completes scanning all
99 * memory areas, and then the tree is rebuilt again from the beginning.
100 * 2) KSM will only insert into the unstable tree, pages whose hash value
101 * has not changed since the previous scan of all memory areas.
102 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
103 * colors of the nodes and not on their contents, assuring that even when
104 * the tree gets "corrupted" it won't get out of balance, so scanning time
105 * remains the same (also, searching and inserting nodes in an rbtree uses
106 * the same algorithm, so we have no overhead when we flush and rebuild).
107 * 4) KSM never flushes the stable tree, which means that even if it were to
108 * take 10 attempts to find a page in the unstable tree, once it is found,
109 * it is secured in the stable tree. (When we scan a new page, we first
110 * compare it against the stable tree, and then against the unstable tree.)
111 *
112 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
113 * stable trees and multiple unstable trees: one of each for each NUMA node.
114 */
115
116/**
117 * struct ksm_mm_slot - ksm information per mm that is being scanned
118 * @slot: hash lookup from mm to mm_slot
119 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
120 */
121struct ksm_mm_slot {
122 struct mm_slot slot;
123 struct ksm_rmap_item *rmap_list;
124};
125
126/**
127 * struct ksm_scan - cursor for scanning
128 * @mm_slot: the current mm_slot we are scanning
129 * @address: the next address inside that to be scanned
130 * @rmap_list: link to the next rmap to be scanned in the rmap_list
131 * @seqnr: count of completed full scans (needed when removing unstable node)
132 *
133 * There is only the one ksm_scan instance of this cursor structure.
134 */
135struct ksm_scan {
136 struct ksm_mm_slot *mm_slot;
137 unsigned long address;
138 struct ksm_rmap_item **rmap_list;
139 unsigned long seqnr;
140};
141
142/**
143 * struct ksm_stable_node - node of the stable rbtree
144 * @node: rb node of this ksm page in the stable tree
145 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
146 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
147 * @list: linked into migrate_nodes, pending placement in the proper node tree
148 * @hlist: hlist head of rmap_items using this ksm page
149 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
150 * @chain_prune_time: time of the last full garbage collection
151 * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
152 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
153 */
154struct ksm_stable_node {
155 union {
156 struct rb_node node; /* when node of stable tree */
157 struct { /* when listed for migration */
158 struct list_head *head;
159 struct {
160 struct hlist_node hlist_dup;
161 struct list_head list;
162 };
163 };
164 };
165 struct hlist_head hlist;
166 union {
167 unsigned long kpfn;
168 unsigned long chain_prune_time;
169 };
170 /*
171 * STABLE_NODE_CHAIN can be any negative number in
172 * rmap_hlist_len negative range, but better not -1 to be able
173 * to reliably detect underflows.
174 */
175#define STABLE_NODE_CHAIN -1024
176 int rmap_hlist_len;
177#ifdef CONFIG_NUMA
178 int nid;
179#endif
180};
181
182/**
183 * struct ksm_rmap_item - reverse mapping item for virtual addresses
184 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
185 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
186 * @nid: NUMA node id of unstable tree in which linked (may not match page)
187 * @mm: the memory structure this rmap_item is pointing into
188 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
189 * @oldchecksum: previous checksum of the page at that virtual address
190 * @node: rb node of this rmap_item in the unstable tree
191 * @head: pointer to stable_node heading this list in the stable tree
192 * @hlist: link into hlist of rmap_items hanging off that stable_node
193 */
194struct ksm_rmap_item {
195 struct ksm_rmap_item *rmap_list;
196 union {
197 struct anon_vma *anon_vma; /* when stable */
198#ifdef CONFIG_NUMA
199 int nid; /* when node of unstable tree */
200#endif
201 };
202 struct mm_struct *mm;
203 unsigned long address; /* + low bits used for flags below */
204 unsigned int oldchecksum; /* when unstable */
205 union {
206 struct rb_node node; /* when node of unstable tree */
207 struct { /* when listed from stable tree */
208 struct ksm_stable_node *head;
209 struct hlist_node hlist;
210 };
211 };
212};
213
214#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
215#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
216#define STABLE_FLAG 0x200 /* is listed from the stable tree */
217
218/* The stable and unstable tree heads */
219static struct rb_root one_stable_tree[1] = { RB_ROOT };
220static struct rb_root one_unstable_tree[1] = { RB_ROOT };
221static struct rb_root *root_stable_tree = one_stable_tree;
222static struct rb_root *root_unstable_tree = one_unstable_tree;
223
224/* Recently migrated nodes of stable tree, pending proper placement */
225static LIST_HEAD(migrate_nodes);
226#define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
227
228#define MM_SLOTS_HASH_BITS 10
229static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
230
231static struct ksm_mm_slot ksm_mm_head = {
232 .slot.mm_node = LIST_HEAD_INIT(ksm_mm_head.slot.mm_node),
233};
234static struct ksm_scan ksm_scan = {
235 .mm_slot = &ksm_mm_head,
236};
237
238static struct kmem_cache *rmap_item_cache;
239static struct kmem_cache *stable_node_cache;
240static struct kmem_cache *mm_slot_cache;
241
242/* The number of nodes in the stable tree */
243static unsigned long ksm_pages_shared;
244
245/* The number of page slots additionally sharing those nodes */
246static unsigned long ksm_pages_sharing;
247
248/* The number of nodes in the unstable tree */
249static unsigned long ksm_pages_unshared;
250
251/* The number of rmap_items in use: to calculate pages_volatile */
252static unsigned long ksm_rmap_items;
253
254/* The number of stable_node chains */
255static unsigned long ksm_stable_node_chains;
256
257/* The number of stable_node dups linked to the stable_node chains */
258static unsigned long ksm_stable_node_dups;
259
260/* Delay in pruning stale stable_node_dups in the stable_node_chains */
261static unsigned int ksm_stable_node_chains_prune_millisecs = 2000;
262
263/* Maximum number of page slots sharing a stable node */
264static int ksm_max_page_sharing = 256;
265
266/* Number of pages ksmd should scan in one batch */
267static unsigned int ksm_thread_pages_to_scan = 100;
268
269/* Milliseconds ksmd should sleep between batches */
270static unsigned int ksm_thread_sleep_millisecs = 20;
271
272/* Checksum of an empty (zeroed) page */
273static unsigned int zero_checksum __read_mostly;
274
275/* Whether to merge empty (zeroed) pages with actual zero pages */
276static bool ksm_use_zero_pages __read_mostly;
277
278#ifdef CONFIG_NUMA
279/* Zeroed when merging across nodes is not allowed */
280static unsigned int ksm_merge_across_nodes = 1;
281static int ksm_nr_node_ids = 1;
282#else
283#define ksm_merge_across_nodes 1U
284#define ksm_nr_node_ids 1
285#endif
286
287#define KSM_RUN_STOP 0
288#define KSM_RUN_MERGE 1
289#define KSM_RUN_UNMERGE 2
290#define KSM_RUN_OFFLINE 4
291static unsigned long ksm_run = KSM_RUN_STOP;
292static void wait_while_offlining(void);
293
294static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
295static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait);
296static DEFINE_MUTEX(ksm_thread_mutex);
297static DEFINE_SPINLOCK(ksm_mmlist_lock);
298
299#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create(#__struct,\
300 sizeof(struct __struct), __alignof__(struct __struct),\
301 (__flags), NULL)
302
303static int __init ksm_slab_init(void)
304{
305 rmap_item_cache = KSM_KMEM_CACHE(ksm_rmap_item, 0);
306 if (!rmap_item_cache)
307 goto out;
308
309 stable_node_cache = KSM_KMEM_CACHE(ksm_stable_node, 0);
310 if (!stable_node_cache)
311 goto out_free1;
312
313 mm_slot_cache = KSM_KMEM_CACHE(ksm_mm_slot, 0);
314 if (!mm_slot_cache)
315 goto out_free2;
316
317 return 0;
318
319out_free2:
320 kmem_cache_destroy(stable_node_cache);
321out_free1:
322 kmem_cache_destroy(rmap_item_cache);
323out:
324 return -ENOMEM;
325}
326
327static void __init ksm_slab_free(void)
328{
329 kmem_cache_destroy(mm_slot_cache);
330 kmem_cache_destroy(stable_node_cache);
331 kmem_cache_destroy(rmap_item_cache);
332 mm_slot_cache = NULL;
333}
334
335static __always_inline bool is_stable_node_chain(struct ksm_stable_node *chain)
336{
337 return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
338}
339
340static __always_inline bool is_stable_node_dup(struct ksm_stable_node *dup)
341{
342 return dup->head == STABLE_NODE_DUP_HEAD;
343}
344
345static inline void stable_node_chain_add_dup(struct ksm_stable_node *dup,
346 struct ksm_stable_node *chain)
347{
348 VM_BUG_ON(is_stable_node_dup(dup));
349 dup->head = STABLE_NODE_DUP_HEAD;
350 VM_BUG_ON(!is_stable_node_chain(chain));
351 hlist_add_head(&dup->hlist_dup, &chain->hlist);
352 ksm_stable_node_dups++;
353}
354
355static inline void __stable_node_dup_del(struct ksm_stable_node *dup)
356{
357 VM_BUG_ON(!is_stable_node_dup(dup));
358 hlist_del(&dup->hlist_dup);
359 ksm_stable_node_dups--;
360}
361
362static inline void stable_node_dup_del(struct ksm_stable_node *dup)
363{
364 VM_BUG_ON(is_stable_node_chain(dup));
365 if (is_stable_node_dup(dup))
366 __stable_node_dup_del(dup);
367 else
368 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
369#ifdef CONFIG_DEBUG_VM
370 dup->head = NULL;
371#endif
372}
373
374static inline struct ksm_rmap_item *alloc_rmap_item(void)
375{
376 struct ksm_rmap_item *rmap_item;
377
378 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
379 __GFP_NORETRY | __GFP_NOWARN);
380 if (rmap_item)
381 ksm_rmap_items++;
382 return rmap_item;
383}
384
385static inline void free_rmap_item(struct ksm_rmap_item *rmap_item)
386{
387 ksm_rmap_items--;
388 rmap_item->mm->ksm_rmap_items--;
389 rmap_item->mm = NULL; /* debug safety */
390 kmem_cache_free(rmap_item_cache, rmap_item);
391}
392
393static inline struct ksm_stable_node *alloc_stable_node(void)
394{
395 /*
396 * The allocation can take too long with GFP_KERNEL when memory is under
397 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
398 * grants access to memory reserves, helping to avoid this problem.
399 */
400 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
401}
402
403static inline void free_stable_node(struct ksm_stable_node *stable_node)
404{
405 VM_BUG_ON(stable_node->rmap_hlist_len &&
406 !is_stable_node_chain(stable_node));
407 kmem_cache_free(stable_node_cache, stable_node);
408}
409
410/*
411 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
412 * page tables after it has passed through ksm_exit() - which, if necessary,
413 * takes mmap_lock briefly to serialize against them. ksm_exit() does not set
414 * a special flag: they can just back out as soon as mm_users goes to zero.
415 * ksm_test_exit() is used throughout to make this test for exit: in some
416 * places for correctness, in some places just to avoid unnecessary work.
417 */
418static inline bool ksm_test_exit(struct mm_struct *mm)
419{
420 return atomic_read(&mm->mm_users) == 0;
421}
422
423static int break_ksm_pmd_entry(pmd_t *pmd, unsigned long addr, unsigned long next,
424 struct mm_walk *walk)
425{
426 struct page *page = NULL;
427 spinlock_t *ptl;
428 pte_t *pte;
429 int ret;
430
431 if (pmd_leaf(*pmd) || !pmd_present(*pmd))
432 return 0;
433
434 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
435 if (pte_present(*pte)) {
436 page = vm_normal_page(walk->vma, addr, *pte);
437 } else if (!pte_none(*pte)) {
438 swp_entry_t entry = pte_to_swp_entry(*pte);
439
440 /*
441 * As KSM pages remain KSM pages until freed, no need to wait
442 * here for migration to end.
443 */
444 if (is_migration_entry(entry))
445 page = pfn_swap_entry_to_page(entry);
446 }
447 ret = page && PageKsm(page);
448 pte_unmap_unlock(pte, ptl);
449 return ret;
450}
451
452static const struct mm_walk_ops break_ksm_ops = {
453 .pmd_entry = break_ksm_pmd_entry,
454};
455
456/*
457 * We use break_ksm to break COW on a ksm page by triggering unsharing,
458 * such that the ksm page will get replaced by an exclusive anonymous page.
459 *
460 * We take great care only to touch a ksm page, in a VM_MERGEABLE vma,
461 * in case the application has unmapped and remapped mm,addr meanwhile.
462 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
463 * mmap of /dev/mem, where we would not want to touch it.
464 *
465 * FAULT_FLAG_REMOTE/FOLL_REMOTE are because we do this outside the context
466 * of the process that owns 'vma'. We also do not want to enforce
467 * protection keys here anyway.
468 */
469static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
470{
471 vm_fault_t ret = 0;
472
473 do {
474 int ksm_page;
475
476 cond_resched();
477 ksm_page = walk_page_range_vma(vma, addr, addr + 1,
478 &break_ksm_ops, NULL);
479 if (WARN_ON_ONCE(ksm_page < 0))
480 return ksm_page;
481 if (!ksm_page)
482 return 0;
483 ret = handle_mm_fault(vma, addr,
484 FAULT_FLAG_UNSHARE | FAULT_FLAG_REMOTE,
485 NULL);
486 } while (!(ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
487 /*
488 * We must loop until we no longer find a KSM page because
489 * handle_mm_fault() may back out if there's any difficulty e.g. if
490 * pte accessed bit gets updated concurrently.
491 *
492 * VM_FAULT_SIGBUS could occur if we race with truncation of the
493 * backing file, which also invalidates anonymous pages: that's
494 * okay, that truncation will have unmapped the PageKsm for us.
495 *
496 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
497 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
498 * current task has TIF_MEMDIE set, and will be OOM killed on return
499 * to user; and ksmd, having no mm, would never be chosen for that.
500 *
501 * But if the mm is in a limited mem_cgroup, then the fault may fail
502 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
503 * even ksmd can fail in this way - though it's usually breaking ksm
504 * just to undo a merge it made a moment before, so unlikely to oom.
505 *
506 * That's a pity: we might therefore have more kernel pages allocated
507 * than we're counting as nodes in the stable tree; but ksm_do_scan
508 * will retry to break_cow on each pass, so should recover the page
509 * in due course. The important thing is to not let VM_MERGEABLE
510 * be cleared while any such pages might remain in the area.
511 */
512 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
513}
514
515static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
516 unsigned long addr)
517{
518 struct vm_area_struct *vma;
519 if (ksm_test_exit(mm))
520 return NULL;
521 vma = vma_lookup(mm, addr);
522 if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
523 return NULL;
524 return vma;
525}
526
527static void break_cow(struct ksm_rmap_item *rmap_item)
528{
529 struct mm_struct *mm = rmap_item->mm;
530 unsigned long addr = rmap_item->address;
531 struct vm_area_struct *vma;
532
533 /*
534 * It is not an accident that whenever we want to break COW
535 * to undo, we also need to drop a reference to the anon_vma.
536 */
537 put_anon_vma(rmap_item->anon_vma);
538
539 mmap_read_lock(mm);
540 vma = find_mergeable_vma(mm, addr);
541 if (vma)
542 break_ksm(vma, addr);
543 mmap_read_unlock(mm);
544}
545
546static struct page *get_mergeable_page(struct ksm_rmap_item *rmap_item)
547{
548 struct mm_struct *mm = rmap_item->mm;
549 unsigned long addr = rmap_item->address;
550 struct vm_area_struct *vma;
551 struct page *page;
552
553 mmap_read_lock(mm);
554 vma = find_mergeable_vma(mm, addr);
555 if (!vma)
556 goto out;
557
558 page = follow_page(vma, addr, FOLL_GET);
559 if (IS_ERR_OR_NULL(page))
560 goto out;
561 if (is_zone_device_page(page))
562 goto out_putpage;
563 if (PageAnon(page)) {
564 flush_anon_page(vma, page, addr);
565 flush_dcache_page(page);
566 } else {
567out_putpage:
568 put_page(page);
569out:
570 page = NULL;
571 }
572 mmap_read_unlock(mm);
573 return page;
574}
575
576/*
577 * This helper is used for getting right index into array of tree roots.
578 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
579 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
580 * every node has its own stable and unstable tree.
581 */
582static inline int get_kpfn_nid(unsigned long kpfn)
583{
584 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
585}
586
587static struct ksm_stable_node *alloc_stable_node_chain(struct ksm_stable_node *dup,
588 struct rb_root *root)
589{
590 struct ksm_stable_node *chain = alloc_stable_node();
591 VM_BUG_ON(is_stable_node_chain(dup));
592 if (likely(chain)) {
593 INIT_HLIST_HEAD(&chain->hlist);
594 chain->chain_prune_time = jiffies;
595 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
596#if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
597 chain->nid = NUMA_NO_NODE; /* debug */
598#endif
599 ksm_stable_node_chains++;
600
601 /*
602 * Put the stable node chain in the first dimension of
603 * the stable tree and at the same time remove the old
604 * stable node.
605 */
606 rb_replace_node(&dup->node, &chain->node, root);
607
608 /*
609 * Move the old stable node to the second dimension
610 * queued in the hlist_dup. The invariant is that all
611 * dup stable_nodes in the chain->hlist point to pages
612 * that are write protected and have the exact same
613 * content.
614 */
615 stable_node_chain_add_dup(dup, chain);
616 }
617 return chain;
618}
619
620static inline void free_stable_node_chain(struct ksm_stable_node *chain,
621 struct rb_root *root)
622{
623 rb_erase(&chain->node, root);
624 free_stable_node(chain);
625 ksm_stable_node_chains--;
626}
627
628static void remove_node_from_stable_tree(struct ksm_stable_node *stable_node)
629{
630 struct ksm_rmap_item *rmap_item;
631
632 /* check it's not STABLE_NODE_CHAIN or negative */
633 BUG_ON(stable_node->rmap_hlist_len < 0);
634
635 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
636 if (rmap_item->hlist.next)
637 ksm_pages_sharing--;
638 else
639 ksm_pages_shared--;
640
641 rmap_item->mm->ksm_merging_pages--;
642
643 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
644 stable_node->rmap_hlist_len--;
645 put_anon_vma(rmap_item->anon_vma);
646 rmap_item->address &= PAGE_MASK;
647 cond_resched();
648 }
649
650 /*
651 * We need the second aligned pointer of the migrate_nodes
652 * list_head to stay clear from the rb_parent_color union
653 * (aligned and different than any node) and also different
654 * from &migrate_nodes. This will verify that future list.h changes
655 * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it.
656 */
657 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
658 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
659
660 if (stable_node->head == &migrate_nodes)
661 list_del(&stable_node->list);
662 else
663 stable_node_dup_del(stable_node);
664 free_stable_node(stable_node);
665}
666
667enum get_ksm_page_flags {
668 GET_KSM_PAGE_NOLOCK,
669 GET_KSM_PAGE_LOCK,
670 GET_KSM_PAGE_TRYLOCK
671};
672
673/*
674 * get_ksm_page: checks if the page indicated by the stable node
675 * is still its ksm page, despite having held no reference to it.
676 * In which case we can trust the content of the page, and it
677 * returns the gotten page; but if the page has now been zapped,
678 * remove the stale node from the stable tree and return NULL.
679 * But beware, the stable node's page might be being migrated.
680 *
681 * You would expect the stable_node to hold a reference to the ksm page.
682 * But if it increments the page's count, swapping out has to wait for
683 * ksmd to come around again before it can free the page, which may take
684 * seconds or even minutes: much too unresponsive. So instead we use a
685 * "keyhole reference": access to the ksm page from the stable node peeps
686 * out through its keyhole to see if that page still holds the right key,
687 * pointing back to this stable node. This relies on freeing a PageAnon
688 * page to reset its page->mapping to NULL, and relies on no other use of
689 * a page to put something that might look like our key in page->mapping.
690 * is on its way to being freed; but it is an anomaly to bear in mind.
691 */
692static struct page *get_ksm_page(struct ksm_stable_node *stable_node,
693 enum get_ksm_page_flags flags)
694{
695 struct page *page;
696 void *expected_mapping;
697 unsigned long kpfn;
698
699 expected_mapping = (void *)((unsigned long)stable_node |
700 PAGE_MAPPING_KSM);
701again:
702 kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
703 page = pfn_to_page(kpfn);
704 if (READ_ONCE(page->mapping) != expected_mapping)
705 goto stale;
706
707 /*
708 * We cannot do anything with the page while its refcount is 0.
709 * Usually 0 means free, or tail of a higher-order page: in which
710 * case this node is no longer referenced, and should be freed;
711 * however, it might mean that the page is under page_ref_freeze().
712 * The __remove_mapping() case is easy, again the node is now stale;
713 * the same is in reuse_ksm_page() case; but if page is swapcache
714 * in folio_migrate_mapping(), it might still be our page,
715 * in which case it's essential to keep the node.
716 */
717 while (!get_page_unless_zero(page)) {
718 /*
719 * Another check for page->mapping != expected_mapping would
720 * work here too. We have chosen the !PageSwapCache test to
721 * optimize the common case, when the page is or is about to
722 * be freed: PageSwapCache is cleared (under spin_lock_irq)
723 * in the ref_freeze section of __remove_mapping(); but Anon
724 * page->mapping reset to NULL later, in free_pages_prepare().
725 */
726 if (!PageSwapCache(page))
727 goto stale;
728 cpu_relax();
729 }
730
731 if (READ_ONCE(page->mapping) != expected_mapping) {
732 put_page(page);
733 goto stale;
734 }
735
736 if (flags == GET_KSM_PAGE_TRYLOCK) {
737 if (!trylock_page(page)) {
738 put_page(page);
739 return ERR_PTR(-EBUSY);
740 }
741 } else if (flags == GET_KSM_PAGE_LOCK)
742 lock_page(page);
743
744 if (flags != GET_KSM_PAGE_NOLOCK) {
745 if (READ_ONCE(page->mapping) != expected_mapping) {
746 unlock_page(page);
747 put_page(page);
748 goto stale;
749 }
750 }
751 return page;
752
753stale:
754 /*
755 * We come here from above when page->mapping or !PageSwapCache
756 * suggests that the node is stale; but it might be under migration.
757 * We need smp_rmb(), matching the smp_wmb() in folio_migrate_ksm(),
758 * before checking whether node->kpfn has been changed.
759 */
760 smp_rmb();
761 if (READ_ONCE(stable_node->kpfn) != kpfn)
762 goto again;
763 remove_node_from_stable_tree(stable_node);
764 return NULL;
765}
766
767/*
768 * Removing rmap_item from stable or unstable tree.
769 * This function will clean the information from the stable/unstable tree.
770 */
771static void remove_rmap_item_from_tree(struct ksm_rmap_item *rmap_item)
772{
773 if (rmap_item->address & STABLE_FLAG) {
774 struct ksm_stable_node *stable_node;
775 struct page *page;
776
777 stable_node = rmap_item->head;
778 page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
779 if (!page)
780 goto out;
781
782 hlist_del(&rmap_item->hlist);
783 unlock_page(page);
784 put_page(page);
785
786 if (!hlist_empty(&stable_node->hlist))
787 ksm_pages_sharing--;
788 else
789 ksm_pages_shared--;
790
791 rmap_item->mm->ksm_merging_pages--;
792
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->head = NULL;
798 rmap_item->address &= PAGE_MASK;
799
800 } else if (rmap_item->address & UNSTABLE_FLAG) {
801 unsigned char age;
802 /*
803 * Usually ksmd can and must skip the rb_erase, because
804 * root_unstable_tree was already reset to RB_ROOT.
805 * But be careful when an mm is exiting: do the rb_erase
806 * if this rmap_item was inserted by this scan, rather
807 * than left over from before.
808 */
809 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
810 BUG_ON(age > 1);
811 if (!age)
812 rb_erase(&rmap_item->node,
813 root_unstable_tree + NUMA(rmap_item->nid));
814 ksm_pages_unshared--;
815 rmap_item->address &= PAGE_MASK;
816 }
817out:
818 cond_resched(); /* we're called from many long loops */
819}
820
821static void remove_trailing_rmap_items(struct ksm_rmap_item **rmap_list)
822{
823 while (*rmap_list) {
824 struct ksm_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_lock. 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 ksm_stable_node *folio_stable_node(struct folio *folio)
862{
863 return folio_test_ksm(folio) ? folio_raw_mapping(folio) : NULL;
864}
865
866static inline struct ksm_stable_node *page_stable_node(struct page *page)
867{
868 return folio_stable_node(page_folio(page));
869}
870
871static inline void set_page_stable_node(struct page *page,
872 struct ksm_stable_node *stable_node)
873{
874 VM_BUG_ON_PAGE(PageAnon(page) && PageAnonExclusive(page), page);
875 page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
876}
877
878#ifdef CONFIG_SYSFS
879/*
880 * Only called through the sysfs control interface:
881 */
882static int remove_stable_node(struct ksm_stable_node *stable_node)
883{
884 struct page *page;
885 int err;
886
887 page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
888 if (!page) {
889 /*
890 * get_ksm_page did remove_node_from_stable_tree itself.
891 */
892 return 0;
893 }
894
895 /*
896 * Page could be still mapped if this races with __mmput() running in
897 * between ksm_exit() and exit_mmap(). Just refuse to let
898 * merge_across_nodes/max_page_sharing be switched.
899 */
900 err = -EBUSY;
901 if (!page_mapped(page)) {
902 /*
903 * The stable node did not yet appear stale to get_ksm_page(),
904 * since that allows for an unmapped ksm page to be recognized
905 * right up until it is freed; but the node is safe to remove.
906 * This page might be in a pagevec waiting to be freed,
907 * or it might be PageSwapCache (perhaps under writeback),
908 * or it might have been removed from swapcache a moment ago.
909 */
910 set_page_stable_node(page, NULL);
911 remove_node_from_stable_tree(stable_node);
912 err = 0;
913 }
914
915 unlock_page(page);
916 put_page(page);
917 return err;
918}
919
920static int remove_stable_node_chain(struct ksm_stable_node *stable_node,
921 struct rb_root *root)
922{
923 struct ksm_stable_node *dup;
924 struct hlist_node *hlist_safe;
925
926 if (!is_stable_node_chain(stable_node)) {
927 VM_BUG_ON(is_stable_node_dup(stable_node));
928 if (remove_stable_node(stable_node))
929 return true;
930 else
931 return false;
932 }
933
934 hlist_for_each_entry_safe(dup, hlist_safe,
935 &stable_node->hlist, hlist_dup) {
936 VM_BUG_ON(!is_stable_node_dup(dup));
937 if (remove_stable_node(dup))
938 return true;
939 }
940 BUG_ON(!hlist_empty(&stable_node->hlist));
941 free_stable_node_chain(stable_node, root);
942 return false;
943}
944
945static int remove_all_stable_nodes(void)
946{
947 struct ksm_stable_node *stable_node, *next;
948 int nid;
949 int err = 0;
950
951 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
952 while (root_stable_tree[nid].rb_node) {
953 stable_node = rb_entry(root_stable_tree[nid].rb_node,
954 struct ksm_stable_node, node);
955 if (remove_stable_node_chain(stable_node,
956 root_stable_tree + nid)) {
957 err = -EBUSY;
958 break; /* proceed to next nid */
959 }
960 cond_resched();
961 }
962 }
963 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
964 if (remove_stable_node(stable_node))
965 err = -EBUSY;
966 cond_resched();
967 }
968 return err;
969}
970
971static int unmerge_and_remove_all_rmap_items(void)
972{
973 struct ksm_mm_slot *mm_slot;
974 struct mm_slot *slot;
975 struct mm_struct *mm;
976 struct vm_area_struct *vma;
977 int err = 0;
978
979 spin_lock(&ksm_mmlist_lock);
980 slot = list_entry(ksm_mm_head.slot.mm_node.next,
981 struct mm_slot, mm_node);
982 ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
983 spin_unlock(&ksm_mmlist_lock);
984
985 for (mm_slot = ksm_scan.mm_slot; mm_slot != &ksm_mm_head;
986 mm_slot = ksm_scan.mm_slot) {
987 VMA_ITERATOR(vmi, mm_slot->slot.mm, 0);
988
989 mm = mm_slot->slot.mm;
990 mmap_read_lock(mm);
991 for_each_vma(vmi, vma) {
992 if (ksm_test_exit(mm))
993 break;
994 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
995 continue;
996 err = unmerge_ksm_pages(vma,
997 vma->vm_start, vma->vm_end);
998 if (err)
999 goto error;
1000 }
1001
1002 remove_trailing_rmap_items(&mm_slot->rmap_list);
1003 mmap_read_unlock(mm);
1004
1005 spin_lock(&ksm_mmlist_lock);
1006 slot = list_entry(mm_slot->slot.mm_node.next,
1007 struct mm_slot, mm_node);
1008 ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
1009 if (ksm_test_exit(mm)) {
1010 hash_del(&mm_slot->slot.hash);
1011 list_del(&mm_slot->slot.mm_node);
1012 spin_unlock(&ksm_mmlist_lock);
1013
1014 mm_slot_free(mm_slot_cache, mm_slot);
1015 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1016 mmdrop(mm);
1017 } else
1018 spin_unlock(&ksm_mmlist_lock);
1019 }
1020
1021 /* Clean up stable nodes, but don't worry if some are still busy */
1022 remove_all_stable_nodes();
1023 ksm_scan.seqnr = 0;
1024 return 0;
1025
1026error:
1027 mmap_read_unlock(mm);
1028 spin_lock(&ksm_mmlist_lock);
1029 ksm_scan.mm_slot = &ksm_mm_head;
1030 spin_unlock(&ksm_mmlist_lock);
1031 return err;
1032}
1033#endif /* CONFIG_SYSFS */
1034
1035static u32 calc_checksum(struct page *page)
1036{
1037 u32 checksum;
1038 void *addr = kmap_atomic(page);
1039 checksum = xxhash(addr, PAGE_SIZE, 0);
1040 kunmap_atomic(addr);
1041 return checksum;
1042}
1043
1044static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1045 pte_t *orig_pte)
1046{
1047 struct mm_struct *mm = vma->vm_mm;
1048 DEFINE_PAGE_VMA_WALK(pvmw, page, vma, 0, 0);
1049 int swapped;
1050 int err = -EFAULT;
1051 struct mmu_notifier_range range;
1052 bool anon_exclusive;
1053
1054 pvmw.address = page_address_in_vma(page, vma);
1055 if (pvmw.address == -EFAULT)
1056 goto out;
1057
1058 BUG_ON(PageTransCompound(page));
1059
1060 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
1061 pvmw.address,
1062 pvmw.address + PAGE_SIZE);
1063 mmu_notifier_invalidate_range_start(&range);
1064
1065 if (!page_vma_mapped_walk(&pvmw))
1066 goto out_mn;
1067 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1068 goto out_unlock;
1069
1070 anon_exclusive = PageAnonExclusive(page);
1071 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1072 anon_exclusive || mm_tlb_flush_pending(mm)) {
1073 pte_t entry;
1074
1075 swapped = PageSwapCache(page);
1076 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1077 /*
1078 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1079 * take any lock, therefore the check that we are going to make
1080 * with the pagecount against the mapcount is racy and
1081 * O_DIRECT can happen right after the check.
1082 * So we clear the pte and flush the tlb before the check
1083 * this assure us that no O_DIRECT can happen after the check
1084 * or in the middle of the check.
1085 *
1086 * No need to notify as we are downgrading page table to read
1087 * only not changing it to point to a new page.
1088 *
1089 * See Documentation/mm/mmu_notifier.rst
1090 */
1091 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1092 /*
1093 * Check that no O_DIRECT or similar I/O is in progress on the
1094 * page
1095 */
1096 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1097 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1098 goto out_unlock;
1099 }
1100
1101 /* See page_try_share_anon_rmap(): clear PTE first. */
1102 if (anon_exclusive && page_try_share_anon_rmap(page)) {
1103 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1104 goto out_unlock;
1105 }
1106
1107 if (pte_dirty(entry))
1108 set_page_dirty(page);
1109 entry = pte_mkclean(entry);
1110
1111 if (pte_write(entry))
1112 entry = pte_wrprotect(entry);
1113
1114 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1115 }
1116 *orig_pte = *pvmw.pte;
1117 err = 0;
1118
1119out_unlock:
1120 page_vma_mapped_walk_done(&pvmw);
1121out_mn:
1122 mmu_notifier_invalidate_range_end(&range);
1123out:
1124 return err;
1125}
1126
1127/**
1128 * replace_page - replace page in vma by new ksm page
1129 * @vma: vma that holds the pte pointing to page
1130 * @page: the page we are replacing by kpage
1131 * @kpage: the ksm page we replace page by
1132 * @orig_pte: the original value of the pte
1133 *
1134 * Returns 0 on success, -EFAULT on failure.
1135 */
1136static int replace_page(struct vm_area_struct *vma, struct page *page,
1137 struct page *kpage, pte_t orig_pte)
1138{
1139 struct mm_struct *mm = vma->vm_mm;
1140 struct folio *folio;
1141 pmd_t *pmd;
1142 pmd_t pmde;
1143 pte_t *ptep;
1144 pte_t newpte;
1145 spinlock_t *ptl;
1146 unsigned long addr;
1147 int err = -EFAULT;
1148 struct mmu_notifier_range range;
1149
1150 addr = page_address_in_vma(page, vma);
1151 if (addr == -EFAULT)
1152 goto out;
1153
1154 pmd = mm_find_pmd(mm, addr);
1155 if (!pmd)
1156 goto out;
1157 /*
1158 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
1159 * without holding anon_vma lock for write. So when looking for a
1160 * genuine pmde (in which to find pte), test present and !THP together.
1161 */
1162 pmde = *pmd;
1163 barrier();
1164 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
1165 goto out;
1166
1167 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr,
1168 addr + PAGE_SIZE);
1169 mmu_notifier_invalidate_range_start(&range);
1170
1171 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1172 if (!pte_same(*ptep, orig_pte)) {
1173 pte_unmap_unlock(ptep, ptl);
1174 goto out_mn;
1175 }
1176 VM_BUG_ON_PAGE(PageAnonExclusive(page), page);
1177 VM_BUG_ON_PAGE(PageAnon(kpage) && PageAnonExclusive(kpage), kpage);
1178
1179 /*
1180 * No need to check ksm_use_zero_pages here: we can only have a
1181 * zero_page here if ksm_use_zero_pages was enabled already.
1182 */
1183 if (!is_zero_pfn(page_to_pfn(kpage))) {
1184 get_page(kpage);
1185 page_add_anon_rmap(kpage, vma, addr, RMAP_NONE);
1186 newpte = mk_pte(kpage, vma->vm_page_prot);
1187 } else {
1188 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1189 vma->vm_page_prot));
1190 /*
1191 * We're replacing an anonymous page with a zero page, which is
1192 * not anonymous. We need to do proper accounting otherwise we
1193 * will get wrong values in /proc, and a BUG message in dmesg
1194 * when tearing down the mm.
1195 */
1196 dec_mm_counter(mm, MM_ANONPAGES);
1197 }
1198
1199 flush_cache_page(vma, addr, pte_pfn(*ptep));
1200 /*
1201 * No need to notify as we are replacing a read only page with another
1202 * read only page with the same content.
1203 *
1204 * See Documentation/mm/mmu_notifier.rst
1205 */
1206 ptep_clear_flush(vma, addr, ptep);
1207 set_pte_at_notify(mm, addr, ptep, newpte);
1208
1209 folio = page_folio(page);
1210 page_remove_rmap(page, vma, false);
1211 if (!folio_mapped(folio))
1212 folio_free_swap(folio);
1213 folio_put(folio);
1214
1215 pte_unmap_unlock(ptep, ptl);
1216 err = 0;
1217out_mn:
1218 mmu_notifier_invalidate_range_end(&range);
1219out:
1220 return err;
1221}
1222
1223/*
1224 * try_to_merge_one_page - take two pages and merge them into one
1225 * @vma: the vma that holds the pte pointing to page
1226 * @page: the PageAnon page that we want to replace with kpage
1227 * @kpage: the PageKsm page that we want to map instead of page,
1228 * or NULL the first time when we want to use page as kpage.
1229 *
1230 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1231 */
1232static int try_to_merge_one_page(struct vm_area_struct *vma,
1233 struct page *page, struct page *kpage)
1234{
1235 pte_t orig_pte = __pte(0);
1236 int err = -EFAULT;
1237
1238 if (page == kpage) /* ksm page forked */
1239 return 0;
1240
1241 if (!PageAnon(page))
1242 goto out;
1243
1244 /*
1245 * We need the page lock to read a stable PageSwapCache in
1246 * write_protect_page(). We use trylock_page() instead of
1247 * lock_page() because we don't want to wait here - we
1248 * prefer to continue scanning and merging different pages,
1249 * then come back to this page when it is unlocked.
1250 */
1251 if (!trylock_page(page))
1252 goto out;
1253
1254 if (PageTransCompound(page)) {
1255 if (split_huge_page(page))
1256 goto out_unlock;
1257 }
1258
1259 /*
1260 * If this anonymous page is mapped only here, its pte may need
1261 * to be write-protected. If it's mapped elsewhere, all of its
1262 * ptes are necessarily already write-protected. But in either
1263 * case, we need to lock and check page_count is not raised.
1264 */
1265 if (write_protect_page(vma, page, &orig_pte) == 0) {
1266 if (!kpage) {
1267 /*
1268 * While we hold page lock, upgrade page from
1269 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1270 * stable_tree_insert() will update stable_node.
1271 */
1272 set_page_stable_node(page, NULL);
1273 mark_page_accessed(page);
1274 /*
1275 * Page reclaim just frees a clean page with no dirty
1276 * ptes: make sure that the ksm page would be swapped.
1277 */
1278 if (!PageDirty(page))
1279 SetPageDirty(page);
1280 err = 0;
1281 } else if (pages_identical(page, kpage))
1282 err = replace_page(vma, page, kpage, orig_pte);
1283 }
1284
1285out_unlock:
1286 unlock_page(page);
1287out:
1288 return err;
1289}
1290
1291/*
1292 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1293 * but no new kernel page is allocated: kpage must already be a ksm page.
1294 *
1295 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1296 */
1297static int try_to_merge_with_ksm_page(struct ksm_rmap_item *rmap_item,
1298 struct page *page, struct page *kpage)
1299{
1300 struct mm_struct *mm = rmap_item->mm;
1301 struct vm_area_struct *vma;
1302 int err = -EFAULT;
1303
1304 mmap_read_lock(mm);
1305 vma = find_mergeable_vma(mm, rmap_item->address);
1306 if (!vma)
1307 goto out;
1308
1309 err = try_to_merge_one_page(vma, page, kpage);
1310 if (err)
1311 goto out;
1312
1313 /* Unstable nid is in union with stable anon_vma: remove first */
1314 remove_rmap_item_from_tree(rmap_item);
1315
1316 /* Must get reference to anon_vma while still holding mmap_lock */
1317 rmap_item->anon_vma = vma->anon_vma;
1318 get_anon_vma(vma->anon_vma);
1319out:
1320 mmap_read_unlock(mm);
1321 return err;
1322}
1323
1324/*
1325 * try_to_merge_two_pages - take two identical pages and prepare them
1326 * to be merged into one page.
1327 *
1328 * This function returns the kpage if we successfully merged two identical
1329 * pages into one ksm page, NULL otherwise.
1330 *
1331 * Note that this function upgrades page to ksm page: if one of the pages
1332 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1333 */
1334static struct page *try_to_merge_two_pages(struct ksm_rmap_item *rmap_item,
1335 struct page *page,
1336 struct ksm_rmap_item *tree_rmap_item,
1337 struct page *tree_page)
1338{
1339 int err;
1340
1341 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1342 if (!err) {
1343 err = try_to_merge_with_ksm_page(tree_rmap_item,
1344 tree_page, page);
1345 /*
1346 * If that fails, we have a ksm page with only one pte
1347 * pointing to it: so break it.
1348 */
1349 if (err)
1350 break_cow(rmap_item);
1351 }
1352 return err ? NULL : page;
1353}
1354
1355static __always_inline
1356bool __is_page_sharing_candidate(struct ksm_stable_node *stable_node, int offset)
1357{
1358 VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1359 /*
1360 * Check that at least one mapping still exists, otherwise
1361 * there's no much point to merge and share with this
1362 * stable_node, as the underlying tree_page of the other
1363 * sharer is going to be freed soon.
1364 */
1365 return stable_node->rmap_hlist_len &&
1366 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1367}
1368
1369static __always_inline
1370bool is_page_sharing_candidate(struct ksm_stable_node *stable_node)
1371{
1372 return __is_page_sharing_candidate(stable_node, 0);
1373}
1374
1375static struct page *stable_node_dup(struct ksm_stable_node **_stable_node_dup,
1376 struct ksm_stable_node **_stable_node,
1377 struct rb_root *root,
1378 bool prune_stale_stable_nodes)
1379{
1380 struct ksm_stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1381 struct hlist_node *hlist_safe;
1382 struct page *_tree_page, *tree_page = NULL;
1383 int nr = 0;
1384 int found_rmap_hlist_len;
1385
1386 if (!prune_stale_stable_nodes ||
1387 time_before(jiffies, stable_node->chain_prune_time +
1388 msecs_to_jiffies(
1389 ksm_stable_node_chains_prune_millisecs)))
1390 prune_stale_stable_nodes = false;
1391 else
1392 stable_node->chain_prune_time = jiffies;
1393
1394 hlist_for_each_entry_safe(dup, hlist_safe,
1395 &stable_node->hlist, hlist_dup) {
1396 cond_resched();
1397 /*
1398 * We must walk all stable_node_dup to prune the stale
1399 * stable nodes during lookup.
1400 *
1401 * get_ksm_page can drop the nodes from the
1402 * stable_node->hlist if they point to freed pages
1403 * (that's why we do a _safe walk). The "dup"
1404 * stable_node parameter itself will be freed from
1405 * under us if it returns NULL.
1406 */
1407 _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK);
1408 if (!_tree_page)
1409 continue;
1410 nr += 1;
1411 if (is_page_sharing_candidate(dup)) {
1412 if (!found ||
1413 dup->rmap_hlist_len > found_rmap_hlist_len) {
1414 if (found)
1415 put_page(tree_page);
1416 found = dup;
1417 found_rmap_hlist_len = found->rmap_hlist_len;
1418 tree_page = _tree_page;
1419
1420 /* skip put_page for found dup */
1421 if (!prune_stale_stable_nodes)
1422 break;
1423 continue;
1424 }
1425 }
1426 put_page(_tree_page);
1427 }
1428
1429 if (found) {
1430 /*
1431 * nr is counting all dups in the chain only if
1432 * prune_stale_stable_nodes is true, otherwise we may
1433 * break the loop at nr == 1 even if there are
1434 * multiple entries.
1435 */
1436 if (prune_stale_stable_nodes && nr == 1) {
1437 /*
1438 * If there's not just one entry it would
1439 * corrupt memory, better BUG_ON. In KSM
1440 * context with no lock held it's not even
1441 * fatal.
1442 */
1443 BUG_ON(stable_node->hlist.first->next);
1444
1445 /*
1446 * There's just one entry and it is below the
1447 * deduplication limit so drop the chain.
1448 */
1449 rb_replace_node(&stable_node->node, &found->node,
1450 root);
1451 free_stable_node(stable_node);
1452 ksm_stable_node_chains--;
1453 ksm_stable_node_dups--;
1454 /*
1455 * NOTE: the caller depends on the stable_node
1456 * to be equal to stable_node_dup if the chain
1457 * was collapsed.
1458 */
1459 *_stable_node = found;
1460 /*
1461 * Just for robustness, as stable_node is
1462 * otherwise left as a stable pointer, the
1463 * compiler shall optimize it away at build
1464 * time.
1465 */
1466 stable_node = NULL;
1467 } else if (stable_node->hlist.first != &found->hlist_dup &&
1468 __is_page_sharing_candidate(found, 1)) {
1469 /*
1470 * If the found stable_node dup can accept one
1471 * more future merge (in addition to the one
1472 * that is underway) and is not at the head of
1473 * the chain, put it there so next search will
1474 * be quicker in the !prune_stale_stable_nodes
1475 * case.
1476 *
1477 * NOTE: it would be inaccurate to use nr > 1
1478 * instead of checking the hlist.first pointer
1479 * directly, because in the
1480 * prune_stale_stable_nodes case "nr" isn't
1481 * the position of the found dup in the chain,
1482 * but the total number of dups in the chain.
1483 */
1484 hlist_del(&found->hlist_dup);
1485 hlist_add_head(&found->hlist_dup,
1486 &stable_node->hlist);
1487 }
1488 }
1489
1490 *_stable_node_dup = found;
1491 return tree_page;
1492}
1493
1494static struct ksm_stable_node *stable_node_dup_any(struct ksm_stable_node *stable_node,
1495 struct rb_root *root)
1496{
1497 if (!is_stable_node_chain(stable_node))
1498 return stable_node;
1499 if (hlist_empty(&stable_node->hlist)) {
1500 free_stable_node_chain(stable_node, root);
1501 return NULL;
1502 }
1503 return hlist_entry(stable_node->hlist.first,
1504 typeof(*stable_node), hlist_dup);
1505}
1506
1507/*
1508 * Like for get_ksm_page, this function can free the *_stable_node and
1509 * *_stable_node_dup if the returned tree_page is NULL.
1510 *
1511 * It can also free and overwrite *_stable_node with the found
1512 * stable_node_dup if the chain is collapsed (in which case
1513 * *_stable_node will be equal to *_stable_node_dup like if the chain
1514 * never existed). It's up to the caller to verify tree_page is not
1515 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1516 *
1517 * *_stable_node_dup is really a second output parameter of this
1518 * function and will be overwritten in all cases, the caller doesn't
1519 * need to initialize it.
1520 */
1521static struct page *__stable_node_chain(struct ksm_stable_node **_stable_node_dup,
1522 struct ksm_stable_node **_stable_node,
1523 struct rb_root *root,
1524 bool prune_stale_stable_nodes)
1525{
1526 struct ksm_stable_node *stable_node = *_stable_node;
1527 if (!is_stable_node_chain(stable_node)) {
1528 if (is_page_sharing_candidate(stable_node)) {
1529 *_stable_node_dup = stable_node;
1530 return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK);
1531 }
1532 /*
1533 * _stable_node_dup set to NULL means the stable_node
1534 * reached the ksm_max_page_sharing limit.
1535 */
1536 *_stable_node_dup = NULL;
1537 return NULL;
1538 }
1539 return stable_node_dup(_stable_node_dup, _stable_node, root,
1540 prune_stale_stable_nodes);
1541}
1542
1543static __always_inline struct page *chain_prune(struct ksm_stable_node **s_n_d,
1544 struct ksm_stable_node **s_n,
1545 struct rb_root *root)
1546{
1547 return __stable_node_chain(s_n_d, s_n, root, true);
1548}
1549
1550static __always_inline struct page *chain(struct ksm_stable_node **s_n_d,
1551 struct ksm_stable_node *s_n,
1552 struct rb_root *root)
1553{
1554 struct ksm_stable_node *old_stable_node = s_n;
1555 struct page *tree_page;
1556
1557 tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1558 /* not pruning dups so s_n cannot have changed */
1559 VM_BUG_ON(s_n != old_stable_node);
1560 return tree_page;
1561}
1562
1563/*
1564 * stable_tree_search - search for page inside the stable tree
1565 *
1566 * This function checks if there is a page inside the stable tree
1567 * with identical content to the page that we are scanning right now.
1568 *
1569 * This function returns the stable tree node of identical content if found,
1570 * NULL otherwise.
1571 */
1572static struct page *stable_tree_search(struct page *page)
1573{
1574 int nid;
1575 struct rb_root *root;
1576 struct rb_node **new;
1577 struct rb_node *parent;
1578 struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any;
1579 struct ksm_stable_node *page_node;
1580
1581 page_node = page_stable_node(page);
1582 if (page_node && page_node->head != &migrate_nodes) {
1583 /* ksm page forked */
1584 get_page(page);
1585 return page;
1586 }
1587
1588 nid = get_kpfn_nid(page_to_pfn(page));
1589 root = root_stable_tree + nid;
1590again:
1591 new = &root->rb_node;
1592 parent = NULL;
1593
1594 while (*new) {
1595 struct page *tree_page;
1596 int ret;
1597
1598 cond_resched();
1599 stable_node = rb_entry(*new, struct ksm_stable_node, node);
1600 stable_node_any = NULL;
1601 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1602 /*
1603 * NOTE: stable_node may have been freed by
1604 * chain_prune() if the returned stable_node_dup is
1605 * not NULL. stable_node_dup may have been inserted in
1606 * the rbtree instead as a regular stable_node (in
1607 * order to collapse the stable_node chain if a single
1608 * stable_node dup was found in it). In such case the
1609 * stable_node is overwritten by the callee to point
1610 * to the stable_node_dup that was collapsed in the
1611 * stable rbtree and stable_node will be equal to
1612 * stable_node_dup like if the chain never existed.
1613 */
1614 if (!stable_node_dup) {
1615 /*
1616 * Either all stable_node dups were full in
1617 * this stable_node chain, or this chain was
1618 * empty and should be rb_erased.
1619 */
1620 stable_node_any = stable_node_dup_any(stable_node,
1621 root);
1622 if (!stable_node_any) {
1623 /* rb_erase just run */
1624 goto again;
1625 }
1626 /*
1627 * Take any of the stable_node dups page of
1628 * this stable_node chain to let the tree walk
1629 * continue. All KSM pages belonging to the
1630 * stable_node dups in a stable_node chain
1631 * have the same content and they're
1632 * write protected at all times. Any will work
1633 * fine to continue the walk.
1634 */
1635 tree_page = get_ksm_page(stable_node_any,
1636 GET_KSM_PAGE_NOLOCK);
1637 }
1638 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1639 if (!tree_page) {
1640 /*
1641 * If we walked over a stale stable_node,
1642 * get_ksm_page() will call rb_erase() and it
1643 * may rebalance the tree from under us. So
1644 * restart the search from scratch. Returning
1645 * NULL would be safe too, but we'd generate
1646 * false negative insertions just because some
1647 * stable_node was stale.
1648 */
1649 goto again;
1650 }
1651
1652 ret = memcmp_pages(page, tree_page);
1653 put_page(tree_page);
1654
1655 parent = *new;
1656 if (ret < 0)
1657 new = &parent->rb_left;
1658 else if (ret > 0)
1659 new = &parent->rb_right;
1660 else {
1661 if (page_node) {
1662 VM_BUG_ON(page_node->head != &migrate_nodes);
1663 /*
1664 * Test if the migrated page should be merged
1665 * into a stable node dup. If the mapcount is
1666 * 1 we can migrate it with another KSM page
1667 * without adding it to the chain.
1668 */
1669 if (page_mapcount(page) > 1)
1670 goto chain_append;
1671 }
1672
1673 if (!stable_node_dup) {
1674 /*
1675 * If the stable_node is a chain and
1676 * we got a payload match in memcmp
1677 * but we cannot merge the scanned
1678 * page in any of the existing
1679 * stable_node dups because they're
1680 * all full, we need to wait the
1681 * scanned page to find itself a match
1682 * in the unstable tree to create a
1683 * brand new KSM page to add later to
1684 * the dups of this stable_node.
1685 */
1686 return NULL;
1687 }
1688
1689 /*
1690 * Lock and unlock the stable_node's page (which
1691 * might already have been migrated) so that page
1692 * migration is sure to notice its raised count.
1693 * It would be more elegant to return stable_node
1694 * than kpage, but that involves more changes.
1695 */
1696 tree_page = get_ksm_page(stable_node_dup,
1697 GET_KSM_PAGE_TRYLOCK);
1698
1699 if (PTR_ERR(tree_page) == -EBUSY)
1700 return ERR_PTR(-EBUSY);
1701
1702 if (unlikely(!tree_page))
1703 /*
1704 * The tree may have been rebalanced,
1705 * so re-evaluate parent and new.
1706 */
1707 goto again;
1708 unlock_page(tree_page);
1709
1710 if (get_kpfn_nid(stable_node_dup->kpfn) !=
1711 NUMA(stable_node_dup->nid)) {
1712 put_page(tree_page);
1713 goto replace;
1714 }
1715 return tree_page;
1716 }
1717 }
1718
1719 if (!page_node)
1720 return NULL;
1721
1722 list_del(&page_node->list);
1723 DO_NUMA(page_node->nid = nid);
1724 rb_link_node(&page_node->node, parent, new);
1725 rb_insert_color(&page_node->node, root);
1726out:
1727 if (is_page_sharing_candidate(page_node)) {
1728 get_page(page);
1729 return page;
1730 } else
1731 return NULL;
1732
1733replace:
1734 /*
1735 * If stable_node was a chain and chain_prune collapsed it,
1736 * stable_node has been updated to be the new regular
1737 * stable_node. A collapse of the chain is indistinguishable
1738 * from the case there was no chain in the stable
1739 * rbtree. Otherwise stable_node is the chain and
1740 * stable_node_dup is the dup to replace.
1741 */
1742 if (stable_node_dup == stable_node) {
1743 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1744 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1745 /* there is no chain */
1746 if (page_node) {
1747 VM_BUG_ON(page_node->head != &migrate_nodes);
1748 list_del(&page_node->list);
1749 DO_NUMA(page_node->nid = nid);
1750 rb_replace_node(&stable_node_dup->node,
1751 &page_node->node,
1752 root);
1753 if (is_page_sharing_candidate(page_node))
1754 get_page(page);
1755 else
1756 page = NULL;
1757 } else {
1758 rb_erase(&stable_node_dup->node, root);
1759 page = NULL;
1760 }
1761 } else {
1762 VM_BUG_ON(!is_stable_node_chain(stable_node));
1763 __stable_node_dup_del(stable_node_dup);
1764 if (page_node) {
1765 VM_BUG_ON(page_node->head != &migrate_nodes);
1766 list_del(&page_node->list);
1767 DO_NUMA(page_node->nid = nid);
1768 stable_node_chain_add_dup(page_node, stable_node);
1769 if (is_page_sharing_candidate(page_node))
1770 get_page(page);
1771 else
1772 page = NULL;
1773 } else {
1774 page = NULL;
1775 }
1776 }
1777 stable_node_dup->head = &migrate_nodes;
1778 list_add(&stable_node_dup->list, stable_node_dup->head);
1779 return page;
1780
1781chain_append:
1782 /* stable_node_dup could be null if it reached the limit */
1783 if (!stable_node_dup)
1784 stable_node_dup = stable_node_any;
1785 /*
1786 * If stable_node was a chain and chain_prune collapsed it,
1787 * stable_node has been updated to be the new regular
1788 * stable_node. A collapse of the chain is indistinguishable
1789 * from the case there was no chain in the stable
1790 * rbtree. Otherwise stable_node is the chain and
1791 * stable_node_dup is the dup to replace.
1792 */
1793 if (stable_node_dup == stable_node) {
1794 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1795 /* chain is missing so create it */
1796 stable_node = alloc_stable_node_chain(stable_node_dup,
1797 root);
1798 if (!stable_node)
1799 return NULL;
1800 }
1801 /*
1802 * Add this stable_node dup that was
1803 * migrated to the stable_node chain
1804 * of the current nid for this page
1805 * content.
1806 */
1807 VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1808 VM_BUG_ON(page_node->head != &migrate_nodes);
1809 list_del(&page_node->list);
1810 DO_NUMA(page_node->nid = nid);
1811 stable_node_chain_add_dup(page_node, stable_node);
1812 goto out;
1813}
1814
1815/*
1816 * stable_tree_insert - insert stable tree node pointing to new ksm page
1817 * into the stable tree.
1818 *
1819 * This function returns the stable tree node just allocated on success,
1820 * NULL otherwise.
1821 */
1822static struct ksm_stable_node *stable_tree_insert(struct page *kpage)
1823{
1824 int nid;
1825 unsigned long kpfn;
1826 struct rb_root *root;
1827 struct rb_node **new;
1828 struct rb_node *parent;
1829 struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any;
1830 bool need_chain = false;
1831
1832 kpfn = page_to_pfn(kpage);
1833 nid = get_kpfn_nid(kpfn);
1834 root = root_stable_tree + nid;
1835again:
1836 parent = NULL;
1837 new = &root->rb_node;
1838
1839 while (*new) {
1840 struct page *tree_page;
1841 int ret;
1842
1843 cond_resched();
1844 stable_node = rb_entry(*new, struct ksm_stable_node, node);
1845 stable_node_any = NULL;
1846 tree_page = chain(&stable_node_dup, stable_node, root);
1847 if (!stable_node_dup) {
1848 /*
1849 * Either all stable_node dups were full in
1850 * this stable_node chain, or this chain was
1851 * empty and should be rb_erased.
1852 */
1853 stable_node_any = stable_node_dup_any(stable_node,
1854 root);
1855 if (!stable_node_any) {
1856 /* rb_erase just run */
1857 goto again;
1858 }
1859 /*
1860 * Take any of the stable_node dups page of
1861 * this stable_node chain to let the tree walk
1862 * continue. All KSM pages belonging to the
1863 * stable_node dups in a stable_node chain
1864 * have the same content and they're
1865 * write protected at all times. Any will work
1866 * fine to continue the walk.
1867 */
1868 tree_page = get_ksm_page(stable_node_any,
1869 GET_KSM_PAGE_NOLOCK);
1870 }
1871 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1872 if (!tree_page) {
1873 /*
1874 * If we walked over a stale stable_node,
1875 * get_ksm_page() will call rb_erase() and it
1876 * may rebalance the tree from under us. So
1877 * restart the search from scratch. Returning
1878 * NULL would be safe too, but we'd generate
1879 * false negative insertions just because some
1880 * stable_node was stale.
1881 */
1882 goto again;
1883 }
1884
1885 ret = memcmp_pages(kpage, tree_page);
1886 put_page(tree_page);
1887
1888 parent = *new;
1889 if (ret < 0)
1890 new = &parent->rb_left;
1891 else if (ret > 0)
1892 new = &parent->rb_right;
1893 else {
1894 need_chain = true;
1895 break;
1896 }
1897 }
1898
1899 stable_node_dup = alloc_stable_node();
1900 if (!stable_node_dup)
1901 return NULL;
1902
1903 INIT_HLIST_HEAD(&stable_node_dup->hlist);
1904 stable_node_dup->kpfn = kpfn;
1905 set_page_stable_node(kpage, stable_node_dup);
1906 stable_node_dup->rmap_hlist_len = 0;
1907 DO_NUMA(stable_node_dup->nid = nid);
1908 if (!need_chain) {
1909 rb_link_node(&stable_node_dup->node, parent, new);
1910 rb_insert_color(&stable_node_dup->node, root);
1911 } else {
1912 if (!is_stable_node_chain(stable_node)) {
1913 struct ksm_stable_node *orig = stable_node;
1914 /* chain is missing so create it */
1915 stable_node = alloc_stable_node_chain(orig, root);
1916 if (!stable_node) {
1917 free_stable_node(stable_node_dup);
1918 return NULL;
1919 }
1920 }
1921 stable_node_chain_add_dup(stable_node_dup, stable_node);
1922 }
1923
1924 return stable_node_dup;
1925}
1926
1927/*
1928 * unstable_tree_search_insert - search for identical page,
1929 * else insert rmap_item into the unstable tree.
1930 *
1931 * This function searches for a page in the unstable tree identical to the
1932 * page currently being scanned; and if no identical page is found in the
1933 * tree, we insert rmap_item as a new object into the unstable tree.
1934 *
1935 * This function returns pointer to rmap_item found to be identical
1936 * to the currently scanned page, NULL otherwise.
1937 *
1938 * This function does both searching and inserting, because they share
1939 * the same walking algorithm in an rbtree.
1940 */
1941static
1942struct ksm_rmap_item *unstable_tree_search_insert(struct ksm_rmap_item *rmap_item,
1943 struct page *page,
1944 struct page **tree_pagep)
1945{
1946 struct rb_node **new;
1947 struct rb_root *root;
1948 struct rb_node *parent = NULL;
1949 int nid;
1950
1951 nid = get_kpfn_nid(page_to_pfn(page));
1952 root = root_unstable_tree + nid;
1953 new = &root->rb_node;
1954
1955 while (*new) {
1956 struct ksm_rmap_item *tree_rmap_item;
1957 struct page *tree_page;
1958 int ret;
1959
1960 cond_resched();
1961 tree_rmap_item = rb_entry(*new, struct ksm_rmap_item, node);
1962 tree_page = get_mergeable_page(tree_rmap_item);
1963 if (!tree_page)
1964 return NULL;
1965
1966 /*
1967 * Don't substitute a ksm page for a forked page.
1968 */
1969 if (page == tree_page) {
1970 put_page(tree_page);
1971 return NULL;
1972 }
1973
1974 ret = memcmp_pages(page, tree_page);
1975
1976 parent = *new;
1977 if (ret < 0) {
1978 put_page(tree_page);
1979 new = &parent->rb_left;
1980 } else if (ret > 0) {
1981 put_page(tree_page);
1982 new = &parent->rb_right;
1983 } else if (!ksm_merge_across_nodes &&
1984 page_to_nid(tree_page) != nid) {
1985 /*
1986 * If tree_page has been migrated to another NUMA node,
1987 * it will be flushed out and put in the right unstable
1988 * tree next time: only merge with it when across_nodes.
1989 */
1990 put_page(tree_page);
1991 return NULL;
1992 } else {
1993 *tree_pagep = tree_page;
1994 return tree_rmap_item;
1995 }
1996 }
1997
1998 rmap_item->address |= UNSTABLE_FLAG;
1999 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
2000 DO_NUMA(rmap_item->nid = nid);
2001 rb_link_node(&rmap_item->node, parent, new);
2002 rb_insert_color(&rmap_item->node, root);
2003
2004 ksm_pages_unshared++;
2005 return NULL;
2006}
2007
2008/*
2009 * stable_tree_append - add another rmap_item to the linked list of
2010 * rmap_items hanging off a given node of the stable tree, all sharing
2011 * the same ksm page.
2012 */
2013static void stable_tree_append(struct ksm_rmap_item *rmap_item,
2014 struct ksm_stable_node *stable_node,
2015 bool max_page_sharing_bypass)
2016{
2017 /*
2018 * rmap won't find this mapping if we don't insert the
2019 * rmap_item in the right stable_node
2020 * duplicate. page_migration could break later if rmap breaks,
2021 * so we can as well crash here. We really need to check for
2022 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
2023 * for other negative values as an underflow if detected here
2024 * for the first time (and not when decreasing rmap_hlist_len)
2025 * would be sign of memory corruption in the stable_node.
2026 */
2027 BUG_ON(stable_node->rmap_hlist_len < 0);
2028
2029 stable_node->rmap_hlist_len++;
2030 if (!max_page_sharing_bypass)
2031 /* possibly non fatal but unexpected overflow, only warn */
2032 WARN_ON_ONCE(stable_node->rmap_hlist_len >
2033 ksm_max_page_sharing);
2034
2035 rmap_item->head = stable_node;
2036 rmap_item->address |= STABLE_FLAG;
2037 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2038
2039 if (rmap_item->hlist.next)
2040 ksm_pages_sharing++;
2041 else
2042 ksm_pages_shared++;
2043
2044 rmap_item->mm->ksm_merging_pages++;
2045}
2046
2047/*
2048 * cmp_and_merge_page - first see if page can be merged into the stable tree;
2049 * if not, compare checksum to previous and if it's the same, see if page can
2050 * be inserted into the unstable tree, or merged with a page already there and
2051 * both transferred to the stable tree.
2052 *
2053 * @page: the page that we are searching identical page to.
2054 * @rmap_item: the reverse mapping into the virtual address of this page
2055 */
2056static void cmp_and_merge_page(struct page *page, struct ksm_rmap_item *rmap_item)
2057{
2058 struct mm_struct *mm = rmap_item->mm;
2059 struct ksm_rmap_item *tree_rmap_item;
2060 struct page *tree_page = NULL;
2061 struct ksm_stable_node *stable_node;
2062 struct page *kpage;
2063 unsigned int checksum;
2064 int err;
2065 bool max_page_sharing_bypass = false;
2066
2067 stable_node = page_stable_node(page);
2068 if (stable_node) {
2069 if (stable_node->head != &migrate_nodes &&
2070 get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2071 NUMA(stable_node->nid)) {
2072 stable_node_dup_del(stable_node);
2073 stable_node->head = &migrate_nodes;
2074 list_add(&stable_node->list, stable_node->head);
2075 }
2076 if (stable_node->head != &migrate_nodes &&
2077 rmap_item->head == stable_node)
2078 return;
2079 /*
2080 * If it's a KSM fork, allow it to go over the sharing limit
2081 * without warnings.
2082 */
2083 if (!is_page_sharing_candidate(stable_node))
2084 max_page_sharing_bypass = true;
2085 }
2086
2087 /* We first start with searching the page inside the stable tree */
2088 kpage = stable_tree_search(page);
2089 if (kpage == page && rmap_item->head == stable_node) {
2090 put_page(kpage);
2091 return;
2092 }
2093
2094 remove_rmap_item_from_tree(rmap_item);
2095
2096 if (kpage) {
2097 if (PTR_ERR(kpage) == -EBUSY)
2098 return;
2099
2100 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2101 if (!err) {
2102 /*
2103 * The page was successfully merged:
2104 * add its rmap_item to the stable tree.
2105 */
2106 lock_page(kpage);
2107 stable_tree_append(rmap_item, page_stable_node(kpage),
2108 max_page_sharing_bypass);
2109 unlock_page(kpage);
2110 }
2111 put_page(kpage);
2112 return;
2113 }
2114
2115 /*
2116 * If the hash value of the page has changed from the last time
2117 * we calculated it, this page is changing frequently: therefore we
2118 * don't want to insert it in the unstable tree, and we don't want
2119 * to waste our time searching for something identical to it there.
2120 */
2121 checksum = calc_checksum(page);
2122 if (rmap_item->oldchecksum != checksum) {
2123 rmap_item->oldchecksum = checksum;
2124 return;
2125 }
2126
2127 /*
2128 * Same checksum as an empty page. We attempt to merge it with the
2129 * appropriate zero page if the user enabled this via sysfs.
2130 */
2131 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2132 struct vm_area_struct *vma;
2133
2134 mmap_read_lock(mm);
2135 vma = find_mergeable_vma(mm, rmap_item->address);
2136 if (vma) {
2137 err = try_to_merge_one_page(vma, page,
2138 ZERO_PAGE(rmap_item->address));
2139 } else {
2140 /*
2141 * If the vma is out of date, we do not need to
2142 * continue.
2143 */
2144 err = 0;
2145 }
2146 mmap_read_unlock(mm);
2147 /*
2148 * In case of failure, the page was not really empty, so we
2149 * need to continue. Otherwise we're done.
2150 */
2151 if (!err)
2152 return;
2153 }
2154 tree_rmap_item =
2155 unstable_tree_search_insert(rmap_item, page, &tree_page);
2156 if (tree_rmap_item) {
2157 bool split;
2158
2159 kpage = try_to_merge_two_pages(rmap_item, page,
2160 tree_rmap_item, tree_page);
2161 /*
2162 * If both pages we tried to merge belong to the same compound
2163 * page, then we actually ended up increasing the reference
2164 * count of the same compound page twice, and split_huge_page
2165 * failed.
2166 * Here we set a flag if that happened, and we use it later to
2167 * try split_huge_page again. Since we call put_page right
2168 * afterwards, the reference count will be correct and
2169 * split_huge_page should succeed.
2170 */
2171 split = PageTransCompound(page)
2172 && compound_head(page) == compound_head(tree_page);
2173 put_page(tree_page);
2174 if (kpage) {
2175 /*
2176 * The pages were successfully merged: insert new
2177 * node in the stable tree and add both rmap_items.
2178 */
2179 lock_page(kpage);
2180 stable_node = stable_tree_insert(kpage);
2181 if (stable_node) {
2182 stable_tree_append(tree_rmap_item, stable_node,
2183 false);
2184 stable_tree_append(rmap_item, stable_node,
2185 false);
2186 }
2187 unlock_page(kpage);
2188
2189 /*
2190 * If we fail to insert the page into the stable tree,
2191 * we will have 2 virtual addresses that are pointing
2192 * to a ksm page left outside the stable tree,
2193 * in which case we need to break_cow on both.
2194 */
2195 if (!stable_node) {
2196 break_cow(tree_rmap_item);
2197 break_cow(rmap_item);
2198 }
2199 } else if (split) {
2200 /*
2201 * We are here if we tried to merge two pages and
2202 * failed because they both belonged to the same
2203 * compound page. We will split the page now, but no
2204 * merging will take place.
2205 * We do not want to add the cost of a full lock; if
2206 * the page is locked, it is better to skip it and
2207 * perhaps try again later.
2208 */
2209 if (!trylock_page(page))
2210 return;
2211 split_huge_page(page);
2212 unlock_page(page);
2213 }
2214 }
2215}
2216
2217static struct ksm_rmap_item *get_next_rmap_item(struct ksm_mm_slot *mm_slot,
2218 struct ksm_rmap_item **rmap_list,
2219 unsigned long addr)
2220{
2221 struct ksm_rmap_item *rmap_item;
2222
2223 while (*rmap_list) {
2224 rmap_item = *rmap_list;
2225 if ((rmap_item->address & PAGE_MASK) == addr)
2226 return rmap_item;
2227 if (rmap_item->address > addr)
2228 break;
2229 *rmap_list = rmap_item->rmap_list;
2230 remove_rmap_item_from_tree(rmap_item);
2231 free_rmap_item(rmap_item);
2232 }
2233
2234 rmap_item = alloc_rmap_item();
2235 if (rmap_item) {
2236 /* It has already been zeroed */
2237 rmap_item->mm = mm_slot->slot.mm;
2238 rmap_item->mm->ksm_rmap_items++;
2239 rmap_item->address = addr;
2240 rmap_item->rmap_list = *rmap_list;
2241 *rmap_list = rmap_item;
2242 }
2243 return rmap_item;
2244}
2245
2246static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page)
2247{
2248 struct mm_struct *mm;
2249 struct ksm_mm_slot *mm_slot;
2250 struct mm_slot *slot;
2251 struct vm_area_struct *vma;
2252 struct ksm_rmap_item *rmap_item;
2253 struct vma_iterator vmi;
2254 int nid;
2255
2256 if (list_empty(&ksm_mm_head.slot.mm_node))
2257 return NULL;
2258
2259 mm_slot = ksm_scan.mm_slot;
2260 if (mm_slot == &ksm_mm_head) {
2261 /*
2262 * A number of pages can hang around indefinitely on per-cpu
2263 * pagevecs, raised page count preventing write_protect_page
2264 * from merging them. Though it doesn't really matter much,
2265 * it is puzzling to see some stuck in pages_volatile until
2266 * other activity jostles them out, and they also prevented
2267 * LTP's KSM test from succeeding deterministically; so drain
2268 * them here (here rather than on entry to ksm_do_scan(),
2269 * so we don't IPI too often when pages_to_scan is set low).
2270 */
2271 lru_add_drain_all();
2272
2273 /*
2274 * Whereas stale stable_nodes on the stable_tree itself
2275 * get pruned in the regular course of stable_tree_search(),
2276 * those moved out to the migrate_nodes list can accumulate:
2277 * so prune them once before each full scan.
2278 */
2279 if (!ksm_merge_across_nodes) {
2280 struct ksm_stable_node *stable_node, *next;
2281 struct page *page;
2282
2283 list_for_each_entry_safe(stable_node, next,
2284 &migrate_nodes, list) {
2285 page = get_ksm_page(stable_node,
2286 GET_KSM_PAGE_NOLOCK);
2287 if (page)
2288 put_page(page);
2289 cond_resched();
2290 }
2291 }
2292
2293 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2294 root_unstable_tree[nid] = RB_ROOT;
2295
2296 spin_lock(&ksm_mmlist_lock);
2297 slot = list_entry(mm_slot->slot.mm_node.next,
2298 struct mm_slot, mm_node);
2299 mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
2300 ksm_scan.mm_slot = mm_slot;
2301 spin_unlock(&ksm_mmlist_lock);
2302 /*
2303 * Although we tested list_empty() above, a racing __ksm_exit
2304 * of the last mm on the list may have removed it since then.
2305 */
2306 if (mm_slot == &ksm_mm_head)
2307 return NULL;
2308next_mm:
2309 ksm_scan.address = 0;
2310 ksm_scan.rmap_list = &mm_slot->rmap_list;
2311 }
2312
2313 slot = &mm_slot->slot;
2314 mm = slot->mm;
2315 vma_iter_init(&vmi, mm, ksm_scan.address);
2316
2317 mmap_read_lock(mm);
2318 if (ksm_test_exit(mm))
2319 goto no_vmas;
2320
2321 for_each_vma(vmi, vma) {
2322 if (!(vma->vm_flags & VM_MERGEABLE))
2323 continue;
2324 if (ksm_scan.address < vma->vm_start)
2325 ksm_scan.address = vma->vm_start;
2326 if (!vma->anon_vma)
2327 ksm_scan.address = vma->vm_end;
2328
2329 while (ksm_scan.address < vma->vm_end) {
2330 if (ksm_test_exit(mm))
2331 break;
2332 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2333 if (IS_ERR_OR_NULL(*page)) {
2334 ksm_scan.address += PAGE_SIZE;
2335 cond_resched();
2336 continue;
2337 }
2338 if (is_zone_device_page(*page))
2339 goto next_page;
2340 if (PageAnon(*page)) {
2341 flush_anon_page(vma, *page, ksm_scan.address);
2342 flush_dcache_page(*page);
2343 rmap_item = get_next_rmap_item(mm_slot,
2344 ksm_scan.rmap_list, ksm_scan.address);
2345 if (rmap_item) {
2346 ksm_scan.rmap_list =
2347 &rmap_item->rmap_list;
2348 ksm_scan.address += PAGE_SIZE;
2349 } else
2350 put_page(*page);
2351 mmap_read_unlock(mm);
2352 return rmap_item;
2353 }
2354next_page:
2355 put_page(*page);
2356 ksm_scan.address += PAGE_SIZE;
2357 cond_resched();
2358 }
2359 }
2360
2361 if (ksm_test_exit(mm)) {
2362no_vmas:
2363 ksm_scan.address = 0;
2364 ksm_scan.rmap_list = &mm_slot->rmap_list;
2365 }
2366 /*
2367 * Nuke all the rmap_items that are above this current rmap:
2368 * because there were no VM_MERGEABLE vmas with such addresses.
2369 */
2370 remove_trailing_rmap_items(ksm_scan.rmap_list);
2371
2372 spin_lock(&ksm_mmlist_lock);
2373 slot = list_entry(mm_slot->slot.mm_node.next,
2374 struct mm_slot, mm_node);
2375 ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
2376 if (ksm_scan.address == 0) {
2377 /*
2378 * We've completed a full scan of all vmas, holding mmap_lock
2379 * throughout, and found no VM_MERGEABLE: so do the same as
2380 * __ksm_exit does to remove this mm from all our lists now.
2381 * This applies either when cleaning up after __ksm_exit
2382 * (but beware: we can reach here even before __ksm_exit),
2383 * or when all VM_MERGEABLE areas have been unmapped (and
2384 * mmap_lock then protects against race with MADV_MERGEABLE).
2385 */
2386 hash_del(&mm_slot->slot.hash);
2387 list_del(&mm_slot->slot.mm_node);
2388 spin_unlock(&ksm_mmlist_lock);
2389
2390 mm_slot_free(mm_slot_cache, mm_slot);
2391 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2392 mmap_read_unlock(mm);
2393 mmdrop(mm);
2394 } else {
2395 mmap_read_unlock(mm);
2396 /*
2397 * mmap_read_unlock(mm) first because after
2398 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2399 * already have been freed under us by __ksm_exit()
2400 * because the "mm_slot" is still hashed and
2401 * ksm_scan.mm_slot doesn't point to it anymore.
2402 */
2403 spin_unlock(&ksm_mmlist_lock);
2404 }
2405
2406 /* Repeat until we've completed scanning the whole list */
2407 mm_slot = ksm_scan.mm_slot;
2408 if (mm_slot != &ksm_mm_head)
2409 goto next_mm;
2410
2411 ksm_scan.seqnr++;
2412 return NULL;
2413}
2414
2415/**
2416 * ksm_do_scan - the ksm scanner main worker function.
2417 * @scan_npages: number of pages we want to scan before we return.
2418 */
2419static void ksm_do_scan(unsigned int scan_npages)
2420{
2421 struct ksm_rmap_item *rmap_item;
2422 struct page *page;
2423
2424 while (scan_npages-- && likely(!freezing(current))) {
2425 cond_resched();
2426 rmap_item = scan_get_next_rmap_item(&page);
2427 if (!rmap_item)
2428 return;
2429 cmp_and_merge_page(page, rmap_item);
2430 put_page(page);
2431 }
2432}
2433
2434static int ksmd_should_run(void)
2435{
2436 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.slot.mm_node);
2437}
2438
2439static int ksm_scan_thread(void *nothing)
2440{
2441 unsigned int sleep_ms;
2442
2443 set_freezable();
2444 set_user_nice(current, 5);
2445
2446 while (!kthread_should_stop()) {
2447 mutex_lock(&ksm_thread_mutex);
2448 wait_while_offlining();
2449 if (ksmd_should_run())
2450 ksm_do_scan(ksm_thread_pages_to_scan);
2451 mutex_unlock(&ksm_thread_mutex);
2452
2453 try_to_freeze();
2454
2455 if (ksmd_should_run()) {
2456 sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs);
2457 wait_event_interruptible_timeout(ksm_iter_wait,
2458 sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs),
2459 msecs_to_jiffies(sleep_ms));
2460 } else {
2461 wait_event_freezable(ksm_thread_wait,
2462 ksmd_should_run() || kthread_should_stop());
2463 }
2464 }
2465 return 0;
2466}
2467
2468int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2469 unsigned long end, int advice, unsigned long *vm_flags)
2470{
2471 struct mm_struct *mm = vma->vm_mm;
2472 int err;
2473
2474 switch (advice) {
2475 case MADV_MERGEABLE:
2476 /*
2477 * Be somewhat over-protective for now!
2478 */
2479 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
2480 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
2481 VM_HUGETLB | VM_MIXEDMAP))
2482 return 0; /* just ignore the advice */
2483
2484 if (vma_is_dax(vma))
2485 return 0;
2486
2487#ifdef VM_SAO
2488 if (*vm_flags & VM_SAO)
2489 return 0;
2490#endif
2491#ifdef VM_SPARC_ADI
2492 if (*vm_flags & VM_SPARC_ADI)
2493 return 0;
2494#endif
2495
2496 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2497 err = __ksm_enter(mm);
2498 if (err)
2499 return err;
2500 }
2501
2502 *vm_flags |= VM_MERGEABLE;
2503 break;
2504
2505 case MADV_UNMERGEABLE:
2506 if (!(*vm_flags & VM_MERGEABLE))
2507 return 0; /* just ignore the advice */
2508
2509 if (vma->anon_vma) {
2510 err = unmerge_ksm_pages(vma, start, end);
2511 if (err)
2512 return err;
2513 }
2514
2515 *vm_flags &= ~VM_MERGEABLE;
2516 break;
2517 }
2518
2519 return 0;
2520}
2521EXPORT_SYMBOL_GPL(ksm_madvise);
2522
2523int __ksm_enter(struct mm_struct *mm)
2524{
2525 struct ksm_mm_slot *mm_slot;
2526 struct mm_slot *slot;
2527 int needs_wakeup;
2528
2529 mm_slot = mm_slot_alloc(mm_slot_cache);
2530 if (!mm_slot)
2531 return -ENOMEM;
2532
2533 slot = &mm_slot->slot;
2534
2535 /* Check ksm_run too? Would need tighter locking */
2536 needs_wakeup = list_empty(&ksm_mm_head.slot.mm_node);
2537
2538 spin_lock(&ksm_mmlist_lock);
2539 mm_slot_insert(mm_slots_hash, mm, slot);
2540 /*
2541 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2542 * insert just behind the scanning cursor, to let the area settle
2543 * down a little; when fork is followed by immediate exec, we don't
2544 * want ksmd to waste time setting up and tearing down an rmap_list.
2545 *
2546 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2547 * scanning cursor, otherwise KSM pages in newly forked mms will be
2548 * missed: then we might as well insert at the end of the list.
2549 */
2550 if (ksm_run & KSM_RUN_UNMERGE)
2551 list_add_tail(&slot->mm_node, &ksm_mm_head.slot.mm_node);
2552 else
2553 list_add_tail(&slot->mm_node, &ksm_scan.mm_slot->slot.mm_node);
2554 spin_unlock(&ksm_mmlist_lock);
2555
2556 set_bit(MMF_VM_MERGEABLE, &mm->flags);
2557 mmgrab(mm);
2558
2559 if (needs_wakeup)
2560 wake_up_interruptible(&ksm_thread_wait);
2561
2562 return 0;
2563}
2564
2565void __ksm_exit(struct mm_struct *mm)
2566{
2567 struct ksm_mm_slot *mm_slot;
2568 struct mm_slot *slot;
2569 int easy_to_free = 0;
2570
2571 /*
2572 * This process is exiting: if it's straightforward (as is the
2573 * case when ksmd was never running), free mm_slot immediately.
2574 * But if it's at the cursor or has rmap_items linked to it, use
2575 * mmap_lock to synchronize with any break_cows before pagetables
2576 * are freed, and leave the mm_slot on the list for ksmd to free.
2577 * Beware: ksm may already have noticed it exiting and freed the slot.
2578 */
2579
2580 spin_lock(&ksm_mmlist_lock);
2581 slot = mm_slot_lookup(mm_slots_hash, mm);
2582 mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
2583 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2584 if (!mm_slot->rmap_list) {
2585 hash_del(&slot->hash);
2586 list_del(&slot->mm_node);
2587 easy_to_free = 1;
2588 } else {
2589 list_move(&slot->mm_node,
2590 &ksm_scan.mm_slot->slot.mm_node);
2591 }
2592 }
2593 spin_unlock(&ksm_mmlist_lock);
2594
2595 if (easy_to_free) {
2596 mm_slot_free(mm_slot_cache, mm_slot);
2597 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2598 mmdrop(mm);
2599 } else if (mm_slot) {
2600 mmap_write_lock(mm);
2601 mmap_write_unlock(mm);
2602 }
2603}
2604
2605struct page *ksm_might_need_to_copy(struct page *page,
2606 struct vm_area_struct *vma, unsigned long address)
2607{
2608 struct folio *folio = page_folio(page);
2609 struct anon_vma *anon_vma = folio_anon_vma(folio);
2610 struct page *new_page;
2611
2612 if (PageKsm(page)) {
2613 if (page_stable_node(page) &&
2614 !(ksm_run & KSM_RUN_UNMERGE))
2615 return page; /* no need to copy it */
2616 } else if (!anon_vma) {
2617 return page; /* no need to copy it */
2618 } else if (page->index == linear_page_index(vma, address) &&
2619 anon_vma->root == vma->anon_vma->root) {
2620 return page; /* still no need to copy it */
2621 }
2622 if (!PageUptodate(page))
2623 return page; /* let do_swap_page report the error */
2624
2625 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2626 if (new_page &&
2627 mem_cgroup_charge(page_folio(new_page), vma->vm_mm, GFP_KERNEL)) {
2628 put_page(new_page);
2629 new_page = NULL;
2630 }
2631 if (new_page) {
2632 if (copy_mc_user_highpage(new_page, page, address, vma)) {
2633 put_page(new_page);
2634 memory_failure_queue(page_to_pfn(page), 0);
2635 return ERR_PTR(-EHWPOISON);
2636 }
2637 SetPageDirty(new_page);
2638 __SetPageUptodate(new_page);
2639 __SetPageLocked(new_page);
2640#ifdef CONFIG_SWAP
2641 count_vm_event(KSM_SWPIN_COPY);
2642#endif
2643 }
2644
2645 return new_page;
2646}
2647
2648void rmap_walk_ksm(struct folio *folio, struct rmap_walk_control *rwc)
2649{
2650 struct ksm_stable_node *stable_node;
2651 struct ksm_rmap_item *rmap_item;
2652 int search_new_forks = 0;
2653
2654 VM_BUG_ON_FOLIO(!folio_test_ksm(folio), folio);
2655
2656 /*
2657 * Rely on the page lock to protect against concurrent modifications
2658 * to that page's node of the stable tree.
2659 */
2660 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2661
2662 stable_node = folio_stable_node(folio);
2663 if (!stable_node)
2664 return;
2665again:
2666 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2667 struct anon_vma *anon_vma = rmap_item->anon_vma;
2668 struct anon_vma_chain *vmac;
2669 struct vm_area_struct *vma;
2670
2671 cond_resched();
2672 if (!anon_vma_trylock_read(anon_vma)) {
2673 if (rwc->try_lock) {
2674 rwc->contended = true;
2675 return;
2676 }
2677 anon_vma_lock_read(anon_vma);
2678 }
2679 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2680 0, ULONG_MAX) {
2681 unsigned long addr;
2682
2683 cond_resched();
2684 vma = vmac->vma;
2685
2686 /* Ignore the stable/unstable/sqnr flags */
2687 addr = rmap_item->address & PAGE_MASK;
2688
2689 if (addr < vma->vm_start || addr >= vma->vm_end)
2690 continue;
2691 /*
2692 * Initially we examine only the vma which covers this
2693 * rmap_item; but later, if there is still work to do,
2694 * we examine covering vmas in other mms: in case they
2695 * were forked from the original since ksmd passed.
2696 */
2697 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2698 continue;
2699
2700 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2701 continue;
2702
2703 if (!rwc->rmap_one(folio, vma, addr, rwc->arg)) {
2704 anon_vma_unlock_read(anon_vma);
2705 return;
2706 }
2707 if (rwc->done && rwc->done(folio)) {
2708 anon_vma_unlock_read(anon_vma);
2709 return;
2710 }
2711 }
2712 anon_vma_unlock_read(anon_vma);
2713 }
2714 if (!search_new_forks++)
2715 goto again;
2716}
2717
2718#ifdef CONFIG_MIGRATION
2719void folio_migrate_ksm(struct folio *newfolio, struct folio *folio)
2720{
2721 struct ksm_stable_node *stable_node;
2722
2723 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2724 VM_BUG_ON_FOLIO(!folio_test_locked(newfolio), newfolio);
2725 VM_BUG_ON_FOLIO(newfolio->mapping != folio->mapping, newfolio);
2726
2727 stable_node = folio_stable_node(folio);
2728 if (stable_node) {
2729 VM_BUG_ON_FOLIO(stable_node->kpfn != folio_pfn(folio), folio);
2730 stable_node->kpfn = folio_pfn(newfolio);
2731 /*
2732 * newfolio->mapping was set in advance; now we need smp_wmb()
2733 * to make sure that the new stable_node->kpfn is visible
2734 * to get_ksm_page() before it can see that folio->mapping
2735 * has gone stale (or that folio_test_swapcache has been cleared).
2736 */
2737 smp_wmb();
2738 set_page_stable_node(&folio->page, NULL);
2739 }
2740}
2741#endif /* CONFIG_MIGRATION */
2742
2743#ifdef CONFIG_MEMORY_HOTREMOVE
2744static void wait_while_offlining(void)
2745{
2746 while (ksm_run & KSM_RUN_OFFLINE) {
2747 mutex_unlock(&ksm_thread_mutex);
2748 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2749 TASK_UNINTERRUPTIBLE);
2750 mutex_lock(&ksm_thread_mutex);
2751 }
2752}
2753
2754static bool stable_node_dup_remove_range(struct ksm_stable_node *stable_node,
2755 unsigned long start_pfn,
2756 unsigned long end_pfn)
2757{
2758 if (stable_node->kpfn >= start_pfn &&
2759 stable_node->kpfn < end_pfn) {
2760 /*
2761 * Don't get_ksm_page, page has already gone:
2762 * which is why we keep kpfn instead of page*
2763 */
2764 remove_node_from_stable_tree(stable_node);
2765 return true;
2766 }
2767 return false;
2768}
2769
2770static bool stable_node_chain_remove_range(struct ksm_stable_node *stable_node,
2771 unsigned long start_pfn,
2772 unsigned long end_pfn,
2773 struct rb_root *root)
2774{
2775 struct ksm_stable_node *dup;
2776 struct hlist_node *hlist_safe;
2777
2778 if (!is_stable_node_chain(stable_node)) {
2779 VM_BUG_ON(is_stable_node_dup(stable_node));
2780 return stable_node_dup_remove_range(stable_node, start_pfn,
2781 end_pfn);
2782 }
2783
2784 hlist_for_each_entry_safe(dup, hlist_safe,
2785 &stable_node->hlist, hlist_dup) {
2786 VM_BUG_ON(!is_stable_node_dup(dup));
2787 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2788 }
2789 if (hlist_empty(&stable_node->hlist)) {
2790 free_stable_node_chain(stable_node, root);
2791 return true; /* notify caller that tree was rebalanced */
2792 } else
2793 return false;
2794}
2795
2796static void ksm_check_stable_tree(unsigned long start_pfn,
2797 unsigned long end_pfn)
2798{
2799 struct ksm_stable_node *stable_node, *next;
2800 struct rb_node *node;
2801 int nid;
2802
2803 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2804 node = rb_first(root_stable_tree + nid);
2805 while (node) {
2806 stable_node = rb_entry(node, struct ksm_stable_node, node);
2807 if (stable_node_chain_remove_range(stable_node,
2808 start_pfn, end_pfn,
2809 root_stable_tree +
2810 nid))
2811 node = rb_first(root_stable_tree + nid);
2812 else
2813 node = rb_next(node);
2814 cond_resched();
2815 }
2816 }
2817 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2818 if (stable_node->kpfn >= start_pfn &&
2819 stable_node->kpfn < end_pfn)
2820 remove_node_from_stable_tree(stable_node);
2821 cond_resched();
2822 }
2823}
2824
2825static int ksm_memory_callback(struct notifier_block *self,
2826 unsigned long action, void *arg)
2827{
2828 struct memory_notify *mn = arg;
2829
2830 switch (action) {
2831 case MEM_GOING_OFFLINE:
2832 /*
2833 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2834 * and remove_all_stable_nodes() while memory is going offline:
2835 * it is unsafe for them to touch the stable tree at this time.
2836 * But unmerge_ksm_pages(), rmap lookups and other entry points
2837 * which do not need the ksm_thread_mutex are all safe.
2838 */
2839 mutex_lock(&ksm_thread_mutex);
2840 ksm_run |= KSM_RUN_OFFLINE;
2841 mutex_unlock(&ksm_thread_mutex);
2842 break;
2843
2844 case MEM_OFFLINE:
2845 /*
2846 * Most of the work is done by page migration; but there might
2847 * be a few stable_nodes left over, still pointing to struct
2848 * pages which have been offlined: prune those from the tree,
2849 * otherwise get_ksm_page() might later try to access a
2850 * non-existent struct page.
2851 */
2852 ksm_check_stable_tree(mn->start_pfn,
2853 mn->start_pfn + mn->nr_pages);
2854 fallthrough;
2855 case MEM_CANCEL_OFFLINE:
2856 mutex_lock(&ksm_thread_mutex);
2857 ksm_run &= ~KSM_RUN_OFFLINE;
2858 mutex_unlock(&ksm_thread_mutex);
2859
2860 smp_mb(); /* wake_up_bit advises this */
2861 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2862 break;
2863 }
2864 return NOTIFY_OK;
2865}
2866#else
2867static void wait_while_offlining(void)
2868{
2869}
2870#endif /* CONFIG_MEMORY_HOTREMOVE */
2871
2872#ifdef CONFIG_SYSFS
2873/*
2874 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2875 */
2876
2877#define KSM_ATTR_RO(_name) \
2878 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2879#define KSM_ATTR(_name) \
2880 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
2881
2882static ssize_t sleep_millisecs_show(struct kobject *kobj,
2883 struct kobj_attribute *attr, char *buf)
2884{
2885 return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs);
2886}
2887
2888static ssize_t sleep_millisecs_store(struct kobject *kobj,
2889 struct kobj_attribute *attr,
2890 const char *buf, size_t count)
2891{
2892 unsigned int msecs;
2893 int err;
2894
2895 err = kstrtouint(buf, 10, &msecs);
2896 if (err)
2897 return -EINVAL;
2898
2899 ksm_thread_sleep_millisecs = msecs;
2900 wake_up_interruptible(&ksm_iter_wait);
2901
2902 return count;
2903}
2904KSM_ATTR(sleep_millisecs);
2905
2906static ssize_t pages_to_scan_show(struct kobject *kobj,
2907 struct kobj_attribute *attr, char *buf)
2908{
2909 return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan);
2910}
2911
2912static ssize_t pages_to_scan_store(struct kobject *kobj,
2913 struct kobj_attribute *attr,
2914 const char *buf, size_t count)
2915{
2916 unsigned int nr_pages;
2917 int err;
2918
2919 err = kstrtouint(buf, 10, &nr_pages);
2920 if (err)
2921 return -EINVAL;
2922
2923 ksm_thread_pages_to_scan = nr_pages;
2924
2925 return count;
2926}
2927KSM_ATTR(pages_to_scan);
2928
2929static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2930 char *buf)
2931{
2932 return sysfs_emit(buf, "%lu\n", ksm_run);
2933}
2934
2935static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2936 const char *buf, size_t count)
2937{
2938 unsigned int flags;
2939 int err;
2940
2941 err = kstrtouint(buf, 10, &flags);
2942 if (err)
2943 return -EINVAL;
2944 if (flags > KSM_RUN_UNMERGE)
2945 return -EINVAL;
2946
2947 /*
2948 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2949 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2950 * breaking COW to free the pages_shared (but leaves mm_slots
2951 * on the list for when ksmd may be set running again).
2952 */
2953
2954 mutex_lock(&ksm_thread_mutex);
2955 wait_while_offlining();
2956 if (ksm_run != flags) {
2957 ksm_run = flags;
2958 if (flags & KSM_RUN_UNMERGE) {
2959 set_current_oom_origin();
2960 err = unmerge_and_remove_all_rmap_items();
2961 clear_current_oom_origin();
2962 if (err) {
2963 ksm_run = KSM_RUN_STOP;
2964 count = err;
2965 }
2966 }
2967 }
2968 mutex_unlock(&ksm_thread_mutex);
2969
2970 if (flags & KSM_RUN_MERGE)
2971 wake_up_interruptible(&ksm_thread_wait);
2972
2973 return count;
2974}
2975KSM_ATTR(run);
2976
2977#ifdef CONFIG_NUMA
2978static ssize_t merge_across_nodes_show(struct kobject *kobj,
2979 struct kobj_attribute *attr, char *buf)
2980{
2981 return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes);
2982}
2983
2984static ssize_t merge_across_nodes_store(struct kobject *kobj,
2985 struct kobj_attribute *attr,
2986 const char *buf, size_t count)
2987{
2988 int err;
2989 unsigned long knob;
2990
2991 err = kstrtoul(buf, 10, &knob);
2992 if (err)
2993 return err;
2994 if (knob > 1)
2995 return -EINVAL;
2996
2997 mutex_lock(&ksm_thread_mutex);
2998 wait_while_offlining();
2999 if (ksm_merge_across_nodes != knob) {
3000 if (ksm_pages_shared || remove_all_stable_nodes())
3001 err = -EBUSY;
3002 else if (root_stable_tree == one_stable_tree) {
3003 struct rb_root *buf;
3004 /*
3005 * This is the first time that we switch away from the
3006 * default of merging across nodes: must now allocate
3007 * a buffer to hold as many roots as may be needed.
3008 * Allocate stable and unstable together:
3009 * MAXSMP NODES_SHIFT 10 will use 16kB.
3010 */
3011 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
3012 GFP_KERNEL);
3013 /* Let us assume that RB_ROOT is NULL is zero */
3014 if (!buf)
3015 err = -ENOMEM;
3016 else {
3017 root_stable_tree = buf;
3018 root_unstable_tree = buf + nr_node_ids;
3019 /* Stable tree is empty but not the unstable */
3020 root_unstable_tree[0] = one_unstable_tree[0];
3021 }
3022 }
3023 if (!err) {
3024 ksm_merge_across_nodes = knob;
3025 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
3026 }
3027 }
3028 mutex_unlock(&ksm_thread_mutex);
3029
3030 return err ? err : count;
3031}
3032KSM_ATTR(merge_across_nodes);
3033#endif
3034
3035static ssize_t use_zero_pages_show(struct kobject *kobj,
3036 struct kobj_attribute *attr, char *buf)
3037{
3038 return sysfs_emit(buf, "%u\n", ksm_use_zero_pages);
3039}
3040static ssize_t use_zero_pages_store(struct kobject *kobj,
3041 struct kobj_attribute *attr,
3042 const char *buf, size_t count)
3043{
3044 int err;
3045 bool value;
3046
3047 err = kstrtobool(buf, &value);
3048 if (err)
3049 return -EINVAL;
3050
3051 ksm_use_zero_pages = value;
3052
3053 return count;
3054}
3055KSM_ATTR(use_zero_pages);
3056
3057static ssize_t max_page_sharing_show(struct kobject *kobj,
3058 struct kobj_attribute *attr, char *buf)
3059{
3060 return sysfs_emit(buf, "%u\n", ksm_max_page_sharing);
3061}
3062
3063static ssize_t max_page_sharing_store(struct kobject *kobj,
3064 struct kobj_attribute *attr,
3065 const char *buf, size_t count)
3066{
3067 int err;
3068 int knob;
3069
3070 err = kstrtoint(buf, 10, &knob);
3071 if (err)
3072 return err;
3073 /*
3074 * When a KSM page is created it is shared by 2 mappings. This
3075 * being a signed comparison, it implicitly verifies it's not
3076 * negative.
3077 */
3078 if (knob < 2)
3079 return -EINVAL;
3080
3081 if (READ_ONCE(ksm_max_page_sharing) == knob)
3082 return count;
3083
3084 mutex_lock(&ksm_thread_mutex);
3085 wait_while_offlining();
3086 if (ksm_max_page_sharing != knob) {
3087 if (ksm_pages_shared || remove_all_stable_nodes())
3088 err = -EBUSY;
3089 else
3090 ksm_max_page_sharing = knob;
3091 }
3092 mutex_unlock(&ksm_thread_mutex);
3093
3094 return err ? err : count;
3095}
3096KSM_ATTR(max_page_sharing);
3097
3098static ssize_t pages_shared_show(struct kobject *kobj,
3099 struct kobj_attribute *attr, char *buf)
3100{
3101 return sysfs_emit(buf, "%lu\n", ksm_pages_shared);
3102}
3103KSM_ATTR_RO(pages_shared);
3104
3105static ssize_t pages_sharing_show(struct kobject *kobj,
3106 struct kobj_attribute *attr, char *buf)
3107{
3108 return sysfs_emit(buf, "%lu\n", ksm_pages_sharing);
3109}
3110KSM_ATTR_RO(pages_sharing);
3111
3112static ssize_t pages_unshared_show(struct kobject *kobj,
3113 struct kobj_attribute *attr, char *buf)
3114{
3115 return sysfs_emit(buf, "%lu\n", ksm_pages_unshared);
3116}
3117KSM_ATTR_RO(pages_unshared);
3118
3119static ssize_t pages_volatile_show(struct kobject *kobj,
3120 struct kobj_attribute *attr, char *buf)
3121{
3122 long ksm_pages_volatile;
3123
3124 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3125 - ksm_pages_sharing - ksm_pages_unshared;
3126 /*
3127 * It was not worth any locking to calculate that statistic,
3128 * but it might therefore sometimes be negative: conceal that.
3129 */
3130 if (ksm_pages_volatile < 0)
3131 ksm_pages_volatile = 0;
3132 return sysfs_emit(buf, "%ld\n", ksm_pages_volatile);
3133}
3134KSM_ATTR_RO(pages_volatile);
3135
3136static ssize_t stable_node_dups_show(struct kobject *kobj,
3137 struct kobj_attribute *attr, char *buf)
3138{
3139 return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups);
3140}
3141KSM_ATTR_RO(stable_node_dups);
3142
3143static ssize_t stable_node_chains_show(struct kobject *kobj,
3144 struct kobj_attribute *attr, char *buf)
3145{
3146 return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains);
3147}
3148KSM_ATTR_RO(stable_node_chains);
3149
3150static ssize_t
3151stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3152 struct kobj_attribute *attr,
3153 char *buf)
3154{
3155 return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3156}
3157
3158static ssize_t
3159stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3160 struct kobj_attribute *attr,
3161 const char *buf, size_t count)
3162{
3163 unsigned int msecs;
3164 int err;
3165
3166 err = kstrtouint(buf, 10, &msecs);
3167 if (err)
3168 return -EINVAL;
3169
3170 ksm_stable_node_chains_prune_millisecs = msecs;
3171
3172 return count;
3173}
3174KSM_ATTR(stable_node_chains_prune_millisecs);
3175
3176static ssize_t full_scans_show(struct kobject *kobj,
3177 struct kobj_attribute *attr, char *buf)
3178{
3179 return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr);
3180}
3181KSM_ATTR_RO(full_scans);
3182
3183static struct attribute *ksm_attrs[] = {
3184 &sleep_millisecs_attr.attr,
3185 &pages_to_scan_attr.attr,
3186 &run_attr.attr,
3187 &pages_shared_attr.attr,
3188 &pages_sharing_attr.attr,
3189 &pages_unshared_attr.attr,
3190 &pages_volatile_attr.attr,
3191 &full_scans_attr.attr,
3192#ifdef CONFIG_NUMA
3193 &merge_across_nodes_attr.attr,
3194#endif
3195 &max_page_sharing_attr.attr,
3196 &stable_node_chains_attr.attr,
3197 &stable_node_dups_attr.attr,
3198 &stable_node_chains_prune_millisecs_attr.attr,
3199 &use_zero_pages_attr.attr,
3200 NULL,
3201};
3202
3203static const struct attribute_group ksm_attr_group = {
3204 .attrs = ksm_attrs,
3205 .name = "ksm",
3206};
3207#endif /* CONFIG_SYSFS */
3208
3209static int __init ksm_init(void)
3210{
3211 struct task_struct *ksm_thread;
3212 int err;
3213
3214 /* The correct value depends on page size and endianness */
3215 zero_checksum = calc_checksum(ZERO_PAGE(0));
3216 /* Default to false for backwards compatibility */
3217 ksm_use_zero_pages = false;
3218
3219 err = ksm_slab_init();
3220 if (err)
3221 goto out;
3222
3223 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3224 if (IS_ERR(ksm_thread)) {
3225 pr_err("ksm: creating kthread failed\n");
3226 err = PTR_ERR(ksm_thread);
3227 goto out_free;
3228 }
3229
3230#ifdef CONFIG_SYSFS
3231 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3232 if (err) {
3233 pr_err("ksm: register sysfs failed\n");
3234 kthread_stop(ksm_thread);
3235 goto out_free;
3236 }
3237#else
3238 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
3239
3240#endif /* CONFIG_SYSFS */
3241
3242#ifdef CONFIG_MEMORY_HOTREMOVE
3243 /* There is no significance to this priority 100 */
3244 hotplug_memory_notifier(ksm_memory_callback, KSM_CALLBACK_PRI);
3245#endif
3246 return 0;
3247
3248out_free:
3249 ksm_slab_free();
3250out:
3251 return err;
3252}
3253subsys_initcall(ksm_init);
1/*
2 * Memory merging support.
3 *
4 * This code enables dynamic sharing of identical pages found in different
5 * memory areas, even if they are not shared by fork()
6 *
7 * Copyright (C) 2008-2009 Red Hat, Inc.
8 * Authors:
9 * Izik Eidus
10 * Andrea Arcangeli
11 * Chris Wright
12 * Hugh Dickins
13 *
14 * This work is licensed under the terms of the GNU GPL, version 2.
15 */
16
17#include <linux/errno.h>
18#include <linux/mm.h>
19#include <linux/fs.h>
20#include <linux/mman.h>
21#include <linux/sched.h>
22#include <linux/rwsem.h>
23#include <linux/pagemap.h>
24#include <linux/rmap.h>
25#include <linux/spinlock.h>
26#include <linux/jhash.h>
27#include <linux/delay.h>
28#include <linux/kthread.h>
29#include <linux/wait.h>
30#include <linux/slab.h>
31#include <linux/rbtree.h>
32#include <linux/memory.h>
33#include <linux/mmu_notifier.h>
34#include <linux/swap.h>
35#include <linux/ksm.h>
36#include <linux/hash.h>
37#include <linux/freezer.h>
38#include <linux/oom.h>
39
40#include <asm/tlbflush.h>
41#include "internal.h"
42
43/*
44 * A few notes about the KSM scanning process,
45 * to make it easier to understand the data structures below:
46 *
47 * In order to reduce excessive scanning, KSM sorts the memory pages by their
48 * contents into a data structure that holds pointers to the pages' locations.
49 *
50 * Since the contents of the pages may change at any moment, KSM cannot just
51 * insert the pages into a normal sorted tree and expect it to find anything.
52 * Therefore KSM uses two data structures - the stable and the unstable tree.
53 *
54 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
55 * by their contents. Because each such page is write-protected, searching on
56 * this tree is fully assured to be working (except when pages are unmapped),
57 * and therefore this tree is called the stable tree.
58 *
59 * In addition to the stable tree, KSM uses a second data structure called the
60 * unstable tree: this tree holds pointers to pages which have been found to
61 * be "unchanged for a period of time". The unstable tree sorts these pages
62 * by their contents, but since they are not write-protected, KSM cannot rely
63 * upon the unstable tree to work correctly - the unstable tree is liable to
64 * be corrupted as its contents are modified, and so it is called unstable.
65 *
66 * KSM solves this problem by several techniques:
67 *
68 * 1) The unstable tree is flushed every time KSM completes scanning all
69 * memory areas, and then the tree is rebuilt again from the beginning.
70 * 2) KSM will only insert into the unstable tree, pages whose hash value
71 * has not changed since the previous scan of all memory areas.
72 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
73 * colors of the nodes and not on their contents, assuring that even when
74 * the tree gets "corrupted" it won't get out of balance, so scanning time
75 * remains the same (also, searching and inserting nodes in an rbtree uses
76 * the same algorithm, so we have no overhead when we flush and rebuild).
77 * 4) KSM never flushes the stable tree, which means that even if it were to
78 * take 10 attempts to find a page in the unstable tree, once it is found,
79 * it is secured in the stable tree. (When we scan a new page, we first
80 * compare it against the stable tree, and then against the unstable tree.)
81 */
82
83/**
84 * struct mm_slot - ksm information per mm that is being scanned
85 * @link: link to the mm_slots hash list
86 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
87 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
88 * @mm: the mm that this information is valid for
89 */
90struct mm_slot {
91 struct hlist_node link;
92 struct list_head mm_list;
93 struct rmap_item *rmap_list;
94 struct mm_struct *mm;
95};
96
97/**
98 * struct ksm_scan - cursor for scanning
99 * @mm_slot: the current mm_slot we are scanning
100 * @address: the next address inside that to be scanned
101 * @rmap_list: link to the next rmap to be scanned in the rmap_list
102 * @seqnr: count of completed full scans (needed when removing unstable node)
103 *
104 * There is only the one ksm_scan instance of this cursor structure.
105 */
106struct ksm_scan {
107 struct mm_slot *mm_slot;
108 unsigned long address;
109 struct rmap_item **rmap_list;
110 unsigned long seqnr;
111};
112
113/**
114 * struct stable_node - node of the stable rbtree
115 * @node: rb node of this ksm page in the stable tree
116 * @hlist: hlist head of rmap_items using this ksm page
117 * @kpfn: page frame number of this ksm page
118 */
119struct stable_node {
120 struct rb_node node;
121 struct hlist_head hlist;
122 unsigned long kpfn;
123};
124
125/**
126 * struct rmap_item - reverse mapping item for virtual addresses
127 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
128 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
129 * @mm: the memory structure this rmap_item is pointing into
130 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
131 * @oldchecksum: previous checksum of the page at that virtual address
132 * @node: rb node of this rmap_item in the unstable tree
133 * @head: pointer to stable_node heading this list in the stable tree
134 * @hlist: link into hlist of rmap_items hanging off that stable_node
135 */
136struct rmap_item {
137 struct rmap_item *rmap_list;
138 struct anon_vma *anon_vma; /* when stable */
139 struct mm_struct *mm;
140 unsigned long address; /* + low bits used for flags below */
141 unsigned int oldchecksum; /* when unstable */
142 union {
143 struct rb_node node; /* when node of unstable tree */
144 struct { /* when listed from stable tree */
145 struct stable_node *head;
146 struct hlist_node hlist;
147 };
148 };
149};
150
151#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
152#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
153#define STABLE_FLAG 0x200 /* is listed from the stable tree */
154
155/* The stable and unstable tree heads */
156static struct rb_root root_stable_tree = RB_ROOT;
157static struct rb_root root_unstable_tree = RB_ROOT;
158
159#define MM_SLOTS_HASH_SHIFT 10
160#define MM_SLOTS_HASH_HEADS (1 << MM_SLOTS_HASH_SHIFT)
161static struct hlist_head mm_slots_hash[MM_SLOTS_HASH_HEADS];
162
163static struct mm_slot ksm_mm_head = {
164 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
165};
166static struct ksm_scan ksm_scan = {
167 .mm_slot = &ksm_mm_head,
168};
169
170static struct kmem_cache *rmap_item_cache;
171static struct kmem_cache *stable_node_cache;
172static struct kmem_cache *mm_slot_cache;
173
174/* The number of nodes in the stable tree */
175static unsigned long ksm_pages_shared;
176
177/* The number of page slots additionally sharing those nodes */
178static unsigned long ksm_pages_sharing;
179
180/* The number of nodes in the unstable tree */
181static unsigned long ksm_pages_unshared;
182
183/* The number of rmap_items in use: to calculate pages_volatile */
184static unsigned long ksm_rmap_items;
185
186/* Number of pages ksmd should scan in one batch */
187static unsigned int ksm_thread_pages_to_scan = 100;
188
189/* Milliseconds ksmd should sleep between batches */
190static unsigned int ksm_thread_sleep_millisecs = 20;
191
192#define KSM_RUN_STOP 0
193#define KSM_RUN_MERGE 1
194#define KSM_RUN_UNMERGE 2
195static unsigned int ksm_run = KSM_RUN_STOP;
196
197static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
198static DEFINE_MUTEX(ksm_thread_mutex);
199static DEFINE_SPINLOCK(ksm_mmlist_lock);
200
201#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
202 sizeof(struct __struct), __alignof__(struct __struct),\
203 (__flags), NULL)
204
205static int __init ksm_slab_init(void)
206{
207 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
208 if (!rmap_item_cache)
209 goto out;
210
211 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
212 if (!stable_node_cache)
213 goto out_free1;
214
215 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
216 if (!mm_slot_cache)
217 goto out_free2;
218
219 return 0;
220
221out_free2:
222 kmem_cache_destroy(stable_node_cache);
223out_free1:
224 kmem_cache_destroy(rmap_item_cache);
225out:
226 return -ENOMEM;
227}
228
229static void __init ksm_slab_free(void)
230{
231 kmem_cache_destroy(mm_slot_cache);
232 kmem_cache_destroy(stable_node_cache);
233 kmem_cache_destroy(rmap_item_cache);
234 mm_slot_cache = NULL;
235}
236
237static inline struct rmap_item *alloc_rmap_item(void)
238{
239 struct rmap_item *rmap_item;
240
241 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
242 if (rmap_item)
243 ksm_rmap_items++;
244 return rmap_item;
245}
246
247static inline void free_rmap_item(struct rmap_item *rmap_item)
248{
249 ksm_rmap_items--;
250 rmap_item->mm = NULL; /* debug safety */
251 kmem_cache_free(rmap_item_cache, rmap_item);
252}
253
254static inline struct stable_node *alloc_stable_node(void)
255{
256 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
257}
258
259static inline void free_stable_node(struct stable_node *stable_node)
260{
261 kmem_cache_free(stable_node_cache, stable_node);
262}
263
264static inline struct mm_slot *alloc_mm_slot(void)
265{
266 if (!mm_slot_cache) /* initialization failed */
267 return NULL;
268 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
269}
270
271static inline void free_mm_slot(struct mm_slot *mm_slot)
272{
273 kmem_cache_free(mm_slot_cache, mm_slot);
274}
275
276static struct mm_slot *get_mm_slot(struct mm_struct *mm)
277{
278 struct mm_slot *mm_slot;
279 struct hlist_head *bucket;
280 struct hlist_node *node;
281
282 bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
283 hlist_for_each_entry(mm_slot, node, bucket, link) {
284 if (mm == mm_slot->mm)
285 return mm_slot;
286 }
287 return NULL;
288}
289
290static void insert_to_mm_slots_hash(struct mm_struct *mm,
291 struct mm_slot *mm_slot)
292{
293 struct hlist_head *bucket;
294
295 bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
296 mm_slot->mm = mm;
297 hlist_add_head(&mm_slot->link, bucket);
298}
299
300static inline int in_stable_tree(struct rmap_item *rmap_item)
301{
302 return rmap_item->address & STABLE_FLAG;
303}
304
305/*
306 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
307 * page tables after it has passed through ksm_exit() - which, if necessary,
308 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
309 * a special flag: they can just back out as soon as mm_users goes to zero.
310 * ksm_test_exit() is used throughout to make this test for exit: in some
311 * places for correctness, in some places just to avoid unnecessary work.
312 */
313static inline bool ksm_test_exit(struct mm_struct *mm)
314{
315 return atomic_read(&mm->mm_users) == 0;
316}
317
318/*
319 * We use break_ksm to break COW on a ksm page: it's a stripped down
320 *
321 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
322 * put_page(page);
323 *
324 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
325 * in case the application has unmapped and remapped mm,addr meanwhile.
326 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
327 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
328 */
329static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
330{
331 struct page *page;
332 int ret = 0;
333
334 do {
335 cond_resched();
336 page = follow_page(vma, addr, FOLL_GET);
337 if (IS_ERR_OR_NULL(page))
338 break;
339 if (PageKsm(page))
340 ret = handle_mm_fault(vma->vm_mm, vma, addr,
341 FAULT_FLAG_WRITE);
342 else
343 ret = VM_FAULT_WRITE;
344 put_page(page);
345 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
346 /*
347 * We must loop because handle_mm_fault() may back out if there's
348 * any difficulty e.g. if pte accessed bit gets updated concurrently.
349 *
350 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
351 * COW has been broken, even if the vma does not permit VM_WRITE;
352 * but note that a concurrent fault might break PageKsm for us.
353 *
354 * VM_FAULT_SIGBUS could occur if we race with truncation of the
355 * backing file, which also invalidates anonymous pages: that's
356 * okay, that truncation will have unmapped the PageKsm for us.
357 *
358 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
359 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
360 * current task has TIF_MEMDIE set, and will be OOM killed on return
361 * to user; and ksmd, having no mm, would never be chosen for that.
362 *
363 * But if the mm is in a limited mem_cgroup, then the fault may fail
364 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
365 * even ksmd can fail in this way - though it's usually breaking ksm
366 * just to undo a merge it made a moment before, so unlikely to oom.
367 *
368 * That's a pity: we might therefore have more kernel pages allocated
369 * than we're counting as nodes in the stable tree; but ksm_do_scan
370 * will retry to break_cow on each pass, so should recover the page
371 * in due course. The important thing is to not let VM_MERGEABLE
372 * be cleared while any such pages might remain in the area.
373 */
374 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
375}
376
377static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
378 unsigned long addr)
379{
380 struct vm_area_struct *vma;
381 if (ksm_test_exit(mm))
382 return NULL;
383 vma = find_vma(mm, addr);
384 if (!vma || vma->vm_start > addr)
385 return NULL;
386 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
387 return NULL;
388 return vma;
389}
390
391static void break_cow(struct rmap_item *rmap_item)
392{
393 struct mm_struct *mm = rmap_item->mm;
394 unsigned long addr = rmap_item->address;
395 struct vm_area_struct *vma;
396
397 /*
398 * It is not an accident that whenever we want to break COW
399 * to undo, we also need to drop a reference to the anon_vma.
400 */
401 put_anon_vma(rmap_item->anon_vma);
402
403 down_read(&mm->mmap_sem);
404 vma = find_mergeable_vma(mm, addr);
405 if (vma)
406 break_ksm(vma, addr);
407 up_read(&mm->mmap_sem);
408}
409
410static struct page *page_trans_compound_anon(struct page *page)
411{
412 if (PageTransCompound(page)) {
413 struct page *head = compound_trans_head(page);
414 /*
415 * head may actually be splitted and freed from under
416 * us but it's ok here.
417 */
418 if (PageAnon(head))
419 return head;
420 }
421 return NULL;
422}
423
424static struct page *get_mergeable_page(struct rmap_item *rmap_item)
425{
426 struct mm_struct *mm = rmap_item->mm;
427 unsigned long addr = rmap_item->address;
428 struct vm_area_struct *vma;
429 struct page *page;
430
431 down_read(&mm->mmap_sem);
432 vma = find_mergeable_vma(mm, addr);
433 if (!vma)
434 goto out;
435
436 page = follow_page(vma, addr, FOLL_GET);
437 if (IS_ERR_OR_NULL(page))
438 goto out;
439 if (PageAnon(page) || page_trans_compound_anon(page)) {
440 flush_anon_page(vma, page, addr);
441 flush_dcache_page(page);
442 } else {
443 put_page(page);
444out: page = NULL;
445 }
446 up_read(&mm->mmap_sem);
447 return page;
448}
449
450static void remove_node_from_stable_tree(struct stable_node *stable_node)
451{
452 struct rmap_item *rmap_item;
453 struct hlist_node *hlist;
454
455 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
456 if (rmap_item->hlist.next)
457 ksm_pages_sharing--;
458 else
459 ksm_pages_shared--;
460 put_anon_vma(rmap_item->anon_vma);
461 rmap_item->address &= PAGE_MASK;
462 cond_resched();
463 }
464
465 rb_erase(&stable_node->node, &root_stable_tree);
466 free_stable_node(stable_node);
467}
468
469/*
470 * get_ksm_page: checks if the page indicated by the stable node
471 * is still its ksm page, despite having held no reference to it.
472 * In which case we can trust the content of the page, and it
473 * returns the gotten page; but if the page has now been zapped,
474 * remove the stale node from the stable tree and return NULL.
475 *
476 * You would expect the stable_node to hold a reference to the ksm page.
477 * But if it increments the page's count, swapping out has to wait for
478 * ksmd to come around again before it can free the page, which may take
479 * seconds or even minutes: much too unresponsive. So instead we use a
480 * "keyhole reference": access to the ksm page from the stable node peeps
481 * out through its keyhole to see if that page still holds the right key,
482 * pointing back to this stable node. This relies on freeing a PageAnon
483 * page to reset its page->mapping to NULL, and relies on no other use of
484 * a page to put something that might look like our key in page->mapping.
485 *
486 * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
487 * but this is different - made simpler by ksm_thread_mutex being held, but
488 * interesting for assuming that no other use of the struct page could ever
489 * put our expected_mapping into page->mapping (or a field of the union which
490 * coincides with page->mapping). The RCU calls are not for KSM at all, but
491 * to keep the page_count protocol described with page_cache_get_speculative.
492 *
493 * Note: it is possible that get_ksm_page() will return NULL one moment,
494 * then page the next, if the page is in between page_freeze_refs() and
495 * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
496 * is on its way to being freed; but it is an anomaly to bear in mind.
497 */
498static struct page *get_ksm_page(struct stable_node *stable_node)
499{
500 struct page *page;
501 void *expected_mapping;
502
503 page = pfn_to_page(stable_node->kpfn);
504 expected_mapping = (void *)stable_node +
505 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
506 rcu_read_lock();
507 if (page->mapping != expected_mapping)
508 goto stale;
509 if (!get_page_unless_zero(page))
510 goto stale;
511 if (page->mapping != expected_mapping) {
512 put_page(page);
513 goto stale;
514 }
515 rcu_read_unlock();
516 return page;
517stale:
518 rcu_read_unlock();
519 remove_node_from_stable_tree(stable_node);
520 return NULL;
521}
522
523/*
524 * Removing rmap_item from stable or unstable tree.
525 * This function will clean the information from the stable/unstable tree.
526 */
527static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
528{
529 if (rmap_item->address & STABLE_FLAG) {
530 struct stable_node *stable_node;
531 struct page *page;
532
533 stable_node = rmap_item->head;
534 page = get_ksm_page(stable_node);
535 if (!page)
536 goto out;
537
538 lock_page(page);
539 hlist_del(&rmap_item->hlist);
540 unlock_page(page);
541 put_page(page);
542
543 if (stable_node->hlist.first)
544 ksm_pages_sharing--;
545 else
546 ksm_pages_shared--;
547
548 put_anon_vma(rmap_item->anon_vma);
549 rmap_item->address &= PAGE_MASK;
550
551 } else if (rmap_item->address & UNSTABLE_FLAG) {
552 unsigned char age;
553 /*
554 * Usually ksmd can and must skip the rb_erase, because
555 * root_unstable_tree was already reset to RB_ROOT.
556 * But be careful when an mm is exiting: do the rb_erase
557 * if this rmap_item was inserted by this scan, rather
558 * than left over from before.
559 */
560 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
561 BUG_ON(age > 1);
562 if (!age)
563 rb_erase(&rmap_item->node, &root_unstable_tree);
564
565 ksm_pages_unshared--;
566 rmap_item->address &= PAGE_MASK;
567 }
568out:
569 cond_resched(); /* we're called from many long loops */
570}
571
572static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
573 struct rmap_item **rmap_list)
574{
575 while (*rmap_list) {
576 struct rmap_item *rmap_item = *rmap_list;
577 *rmap_list = rmap_item->rmap_list;
578 remove_rmap_item_from_tree(rmap_item);
579 free_rmap_item(rmap_item);
580 }
581}
582
583/*
584 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
585 * than check every pte of a given vma, the locking doesn't quite work for
586 * that - an rmap_item is assigned to the stable tree after inserting ksm
587 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
588 * rmap_items from parent to child at fork time (so as not to waste time
589 * if exit comes before the next scan reaches it).
590 *
591 * Similarly, although we'd like to remove rmap_items (so updating counts
592 * and freeing memory) when unmerging an area, it's easier to leave that
593 * to the next pass of ksmd - consider, for example, how ksmd might be
594 * in cmp_and_merge_page on one of the rmap_items we would be removing.
595 */
596static int unmerge_ksm_pages(struct vm_area_struct *vma,
597 unsigned long start, unsigned long end)
598{
599 unsigned long addr;
600 int err = 0;
601
602 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
603 if (ksm_test_exit(vma->vm_mm))
604 break;
605 if (signal_pending(current))
606 err = -ERESTARTSYS;
607 else
608 err = break_ksm(vma, addr);
609 }
610 return err;
611}
612
613#ifdef CONFIG_SYSFS
614/*
615 * Only called through the sysfs control interface:
616 */
617static int unmerge_and_remove_all_rmap_items(void)
618{
619 struct mm_slot *mm_slot;
620 struct mm_struct *mm;
621 struct vm_area_struct *vma;
622 int err = 0;
623
624 spin_lock(&ksm_mmlist_lock);
625 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
626 struct mm_slot, mm_list);
627 spin_unlock(&ksm_mmlist_lock);
628
629 for (mm_slot = ksm_scan.mm_slot;
630 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
631 mm = mm_slot->mm;
632 down_read(&mm->mmap_sem);
633 for (vma = mm->mmap; vma; vma = vma->vm_next) {
634 if (ksm_test_exit(mm))
635 break;
636 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
637 continue;
638 err = unmerge_ksm_pages(vma,
639 vma->vm_start, vma->vm_end);
640 if (err)
641 goto error;
642 }
643
644 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
645
646 spin_lock(&ksm_mmlist_lock);
647 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
648 struct mm_slot, mm_list);
649 if (ksm_test_exit(mm)) {
650 hlist_del(&mm_slot->link);
651 list_del(&mm_slot->mm_list);
652 spin_unlock(&ksm_mmlist_lock);
653
654 free_mm_slot(mm_slot);
655 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
656 up_read(&mm->mmap_sem);
657 mmdrop(mm);
658 } else {
659 spin_unlock(&ksm_mmlist_lock);
660 up_read(&mm->mmap_sem);
661 }
662 }
663
664 ksm_scan.seqnr = 0;
665 return 0;
666
667error:
668 up_read(&mm->mmap_sem);
669 spin_lock(&ksm_mmlist_lock);
670 ksm_scan.mm_slot = &ksm_mm_head;
671 spin_unlock(&ksm_mmlist_lock);
672 return err;
673}
674#endif /* CONFIG_SYSFS */
675
676static u32 calc_checksum(struct page *page)
677{
678 u32 checksum;
679 void *addr = kmap_atomic(page);
680 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
681 kunmap_atomic(addr);
682 return checksum;
683}
684
685static int memcmp_pages(struct page *page1, struct page *page2)
686{
687 char *addr1, *addr2;
688 int ret;
689
690 addr1 = kmap_atomic(page1);
691 addr2 = kmap_atomic(page2);
692 ret = memcmp(addr1, addr2, PAGE_SIZE);
693 kunmap_atomic(addr2);
694 kunmap_atomic(addr1);
695 return ret;
696}
697
698static inline int pages_identical(struct page *page1, struct page *page2)
699{
700 return !memcmp_pages(page1, page2);
701}
702
703static int write_protect_page(struct vm_area_struct *vma, struct page *page,
704 pte_t *orig_pte)
705{
706 struct mm_struct *mm = vma->vm_mm;
707 unsigned long addr;
708 pte_t *ptep;
709 spinlock_t *ptl;
710 int swapped;
711 int err = -EFAULT;
712
713 addr = page_address_in_vma(page, vma);
714 if (addr == -EFAULT)
715 goto out;
716
717 BUG_ON(PageTransCompound(page));
718 ptep = page_check_address(page, mm, addr, &ptl, 0);
719 if (!ptep)
720 goto out;
721
722 if (pte_write(*ptep) || pte_dirty(*ptep)) {
723 pte_t entry;
724
725 swapped = PageSwapCache(page);
726 flush_cache_page(vma, addr, page_to_pfn(page));
727 /*
728 * Ok this is tricky, when get_user_pages_fast() run it doesn't
729 * take any lock, therefore the check that we are going to make
730 * with the pagecount against the mapcount is racey and
731 * O_DIRECT can happen right after the check.
732 * So we clear the pte and flush the tlb before the check
733 * this assure us that no O_DIRECT can happen after the check
734 * or in the middle of the check.
735 */
736 entry = ptep_clear_flush(vma, addr, ptep);
737 /*
738 * Check that no O_DIRECT or similar I/O is in progress on the
739 * page
740 */
741 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
742 set_pte_at(mm, addr, ptep, entry);
743 goto out_unlock;
744 }
745 if (pte_dirty(entry))
746 set_page_dirty(page);
747 entry = pte_mkclean(pte_wrprotect(entry));
748 set_pte_at_notify(mm, addr, ptep, entry);
749 }
750 *orig_pte = *ptep;
751 err = 0;
752
753out_unlock:
754 pte_unmap_unlock(ptep, ptl);
755out:
756 return err;
757}
758
759/**
760 * replace_page - replace page in vma by new ksm page
761 * @vma: vma that holds the pte pointing to page
762 * @page: the page we are replacing by kpage
763 * @kpage: the ksm page we replace page by
764 * @orig_pte: the original value of the pte
765 *
766 * Returns 0 on success, -EFAULT on failure.
767 */
768static int replace_page(struct vm_area_struct *vma, struct page *page,
769 struct page *kpage, pte_t orig_pte)
770{
771 struct mm_struct *mm = vma->vm_mm;
772 pgd_t *pgd;
773 pud_t *pud;
774 pmd_t *pmd;
775 pte_t *ptep;
776 spinlock_t *ptl;
777 unsigned long addr;
778 int err = -EFAULT;
779
780 addr = page_address_in_vma(page, vma);
781 if (addr == -EFAULT)
782 goto out;
783
784 pgd = pgd_offset(mm, addr);
785 if (!pgd_present(*pgd))
786 goto out;
787
788 pud = pud_offset(pgd, addr);
789 if (!pud_present(*pud))
790 goto out;
791
792 pmd = pmd_offset(pud, addr);
793 BUG_ON(pmd_trans_huge(*pmd));
794 if (!pmd_present(*pmd))
795 goto out;
796
797 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
798 if (!pte_same(*ptep, orig_pte)) {
799 pte_unmap_unlock(ptep, ptl);
800 goto out;
801 }
802
803 get_page(kpage);
804 page_add_anon_rmap(kpage, vma, addr);
805
806 flush_cache_page(vma, addr, pte_pfn(*ptep));
807 ptep_clear_flush(vma, addr, ptep);
808 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
809
810 page_remove_rmap(page);
811 if (!page_mapped(page))
812 try_to_free_swap(page);
813 put_page(page);
814
815 pte_unmap_unlock(ptep, ptl);
816 err = 0;
817out:
818 return err;
819}
820
821static int page_trans_compound_anon_split(struct page *page)
822{
823 int ret = 0;
824 struct page *transhuge_head = page_trans_compound_anon(page);
825 if (transhuge_head) {
826 /* Get the reference on the head to split it. */
827 if (get_page_unless_zero(transhuge_head)) {
828 /*
829 * Recheck we got the reference while the head
830 * was still anonymous.
831 */
832 if (PageAnon(transhuge_head))
833 ret = split_huge_page(transhuge_head);
834 else
835 /*
836 * Retry later if split_huge_page run
837 * from under us.
838 */
839 ret = 1;
840 put_page(transhuge_head);
841 } else
842 /* Retry later if split_huge_page run from under us. */
843 ret = 1;
844 }
845 return ret;
846}
847
848/*
849 * try_to_merge_one_page - take two pages and merge them into one
850 * @vma: the vma that holds the pte pointing to page
851 * @page: the PageAnon page that we want to replace with kpage
852 * @kpage: the PageKsm page that we want to map instead of page,
853 * or NULL the first time when we want to use page as kpage.
854 *
855 * This function returns 0 if the pages were merged, -EFAULT otherwise.
856 */
857static int try_to_merge_one_page(struct vm_area_struct *vma,
858 struct page *page, struct page *kpage)
859{
860 pte_t orig_pte = __pte(0);
861 int err = -EFAULT;
862
863 if (page == kpage) /* ksm page forked */
864 return 0;
865
866 if (!(vma->vm_flags & VM_MERGEABLE))
867 goto out;
868 if (PageTransCompound(page) && page_trans_compound_anon_split(page))
869 goto out;
870 BUG_ON(PageTransCompound(page));
871 if (!PageAnon(page))
872 goto out;
873
874 /*
875 * We need the page lock to read a stable PageSwapCache in
876 * write_protect_page(). We use trylock_page() instead of
877 * lock_page() because we don't want to wait here - we
878 * prefer to continue scanning and merging different pages,
879 * then come back to this page when it is unlocked.
880 */
881 if (!trylock_page(page))
882 goto out;
883 /*
884 * If this anonymous page is mapped only here, its pte may need
885 * to be write-protected. If it's mapped elsewhere, all of its
886 * ptes are necessarily already write-protected. But in either
887 * case, we need to lock and check page_count is not raised.
888 */
889 if (write_protect_page(vma, page, &orig_pte) == 0) {
890 if (!kpage) {
891 /*
892 * While we hold page lock, upgrade page from
893 * PageAnon+anon_vma to PageKsm+NULL stable_node:
894 * stable_tree_insert() will update stable_node.
895 */
896 set_page_stable_node(page, NULL);
897 mark_page_accessed(page);
898 err = 0;
899 } else if (pages_identical(page, kpage))
900 err = replace_page(vma, page, kpage, orig_pte);
901 }
902
903 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
904 munlock_vma_page(page);
905 if (!PageMlocked(kpage)) {
906 unlock_page(page);
907 lock_page(kpage);
908 mlock_vma_page(kpage);
909 page = kpage; /* for final unlock */
910 }
911 }
912
913 unlock_page(page);
914out:
915 return err;
916}
917
918/*
919 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
920 * but no new kernel page is allocated: kpage must already be a ksm page.
921 *
922 * This function returns 0 if the pages were merged, -EFAULT otherwise.
923 */
924static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
925 struct page *page, struct page *kpage)
926{
927 struct mm_struct *mm = rmap_item->mm;
928 struct vm_area_struct *vma;
929 int err = -EFAULT;
930
931 down_read(&mm->mmap_sem);
932 if (ksm_test_exit(mm))
933 goto out;
934 vma = find_vma(mm, rmap_item->address);
935 if (!vma || vma->vm_start > rmap_item->address)
936 goto out;
937
938 err = try_to_merge_one_page(vma, page, kpage);
939 if (err)
940 goto out;
941
942 /* Must get reference to anon_vma while still holding mmap_sem */
943 rmap_item->anon_vma = vma->anon_vma;
944 get_anon_vma(vma->anon_vma);
945out:
946 up_read(&mm->mmap_sem);
947 return err;
948}
949
950/*
951 * try_to_merge_two_pages - take two identical pages and prepare them
952 * to be merged into one page.
953 *
954 * This function returns the kpage if we successfully merged two identical
955 * pages into one ksm page, NULL otherwise.
956 *
957 * Note that this function upgrades page to ksm page: if one of the pages
958 * is already a ksm page, try_to_merge_with_ksm_page should be used.
959 */
960static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
961 struct page *page,
962 struct rmap_item *tree_rmap_item,
963 struct page *tree_page)
964{
965 int err;
966
967 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
968 if (!err) {
969 err = try_to_merge_with_ksm_page(tree_rmap_item,
970 tree_page, page);
971 /*
972 * If that fails, we have a ksm page with only one pte
973 * pointing to it: so break it.
974 */
975 if (err)
976 break_cow(rmap_item);
977 }
978 return err ? NULL : page;
979}
980
981/*
982 * stable_tree_search - search for page inside the stable tree
983 *
984 * This function checks if there is a page inside the stable tree
985 * with identical content to the page that we are scanning right now.
986 *
987 * This function returns the stable tree node of identical content if found,
988 * NULL otherwise.
989 */
990static struct page *stable_tree_search(struct page *page)
991{
992 struct rb_node *node = root_stable_tree.rb_node;
993 struct stable_node *stable_node;
994
995 stable_node = page_stable_node(page);
996 if (stable_node) { /* ksm page forked */
997 get_page(page);
998 return page;
999 }
1000
1001 while (node) {
1002 struct page *tree_page;
1003 int ret;
1004
1005 cond_resched();
1006 stable_node = rb_entry(node, struct stable_node, node);
1007 tree_page = get_ksm_page(stable_node);
1008 if (!tree_page)
1009 return NULL;
1010
1011 ret = memcmp_pages(page, tree_page);
1012
1013 if (ret < 0) {
1014 put_page(tree_page);
1015 node = node->rb_left;
1016 } else if (ret > 0) {
1017 put_page(tree_page);
1018 node = node->rb_right;
1019 } else
1020 return tree_page;
1021 }
1022
1023 return NULL;
1024}
1025
1026/*
1027 * stable_tree_insert - insert rmap_item pointing to new ksm page
1028 * into the stable tree.
1029 *
1030 * This function returns the stable tree node just allocated on success,
1031 * NULL otherwise.
1032 */
1033static struct stable_node *stable_tree_insert(struct page *kpage)
1034{
1035 struct rb_node **new = &root_stable_tree.rb_node;
1036 struct rb_node *parent = NULL;
1037 struct stable_node *stable_node;
1038
1039 while (*new) {
1040 struct page *tree_page;
1041 int ret;
1042
1043 cond_resched();
1044 stable_node = rb_entry(*new, struct stable_node, node);
1045 tree_page = get_ksm_page(stable_node);
1046 if (!tree_page)
1047 return NULL;
1048
1049 ret = memcmp_pages(kpage, tree_page);
1050 put_page(tree_page);
1051
1052 parent = *new;
1053 if (ret < 0)
1054 new = &parent->rb_left;
1055 else if (ret > 0)
1056 new = &parent->rb_right;
1057 else {
1058 /*
1059 * It is not a bug that stable_tree_search() didn't
1060 * find this node: because at that time our page was
1061 * not yet write-protected, so may have changed since.
1062 */
1063 return NULL;
1064 }
1065 }
1066
1067 stable_node = alloc_stable_node();
1068 if (!stable_node)
1069 return NULL;
1070
1071 rb_link_node(&stable_node->node, parent, new);
1072 rb_insert_color(&stable_node->node, &root_stable_tree);
1073
1074 INIT_HLIST_HEAD(&stable_node->hlist);
1075
1076 stable_node->kpfn = page_to_pfn(kpage);
1077 set_page_stable_node(kpage, stable_node);
1078
1079 return stable_node;
1080}
1081
1082/*
1083 * unstable_tree_search_insert - search for identical page,
1084 * else insert rmap_item into the unstable tree.
1085 *
1086 * This function searches for a page in the unstable tree identical to the
1087 * page currently being scanned; and if no identical page is found in the
1088 * tree, we insert rmap_item as a new object into the unstable tree.
1089 *
1090 * This function returns pointer to rmap_item found to be identical
1091 * to the currently scanned page, NULL otherwise.
1092 *
1093 * This function does both searching and inserting, because they share
1094 * the same walking algorithm in an rbtree.
1095 */
1096static
1097struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1098 struct page *page,
1099 struct page **tree_pagep)
1100
1101{
1102 struct rb_node **new = &root_unstable_tree.rb_node;
1103 struct rb_node *parent = NULL;
1104
1105 while (*new) {
1106 struct rmap_item *tree_rmap_item;
1107 struct page *tree_page;
1108 int ret;
1109
1110 cond_resched();
1111 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1112 tree_page = get_mergeable_page(tree_rmap_item);
1113 if (IS_ERR_OR_NULL(tree_page))
1114 return NULL;
1115
1116 /*
1117 * Don't substitute a ksm page for a forked page.
1118 */
1119 if (page == tree_page) {
1120 put_page(tree_page);
1121 return NULL;
1122 }
1123
1124 ret = memcmp_pages(page, tree_page);
1125
1126 parent = *new;
1127 if (ret < 0) {
1128 put_page(tree_page);
1129 new = &parent->rb_left;
1130 } else if (ret > 0) {
1131 put_page(tree_page);
1132 new = &parent->rb_right;
1133 } else {
1134 *tree_pagep = tree_page;
1135 return tree_rmap_item;
1136 }
1137 }
1138
1139 rmap_item->address |= UNSTABLE_FLAG;
1140 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1141 rb_link_node(&rmap_item->node, parent, new);
1142 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1143
1144 ksm_pages_unshared++;
1145 return NULL;
1146}
1147
1148/*
1149 * stable_tree_append - add another rmap_item to the linked list of
1150 * rmap_items hanging off a given node of the stable tree, all sharing
1151 * the same ksm page.
1152 */
1153static void stable_tree_append(struct rmap_item *rmap_item,
1154 struct stable_node *stable_node)
1155{
1156 rmap_item->head = stable_node;
1157 rmap_item->address |= STABLE_FLAG;
1158 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1159
1160 if (rmap_item->hlist.next)
1161 ksm_pages_sharing++;
1162 else
1163 ksm_pages_shared++;
1164}
1165
1166/*
1167 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1168 * if not, compare checksum to previous and if it's the same, see if page can
1169 * be inserted into the unstable tree, or merged with a page already there and
1170 * both transferred to the stable tree.
1171 *
1172 * @page: the page that we are searching identical page to.
1173 * @rmap_item: the reverse mapping into the virtual address of this page
1174 */
1175static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1176{
1177 struct rmap_item *tree_rmap_item;
1178 struct page *tree_page = NULL;
1179 struct stable_node *stable_node;
1180 struct page *kpage;
1181 unsigned int checksum;
1182 int err;
1183
1184 remove_rmap_item_from_tree(rmap_item);
1185
1186 /* We first start with searching the page inside the stable tree */
1187 kpage = stable_tree_search(page);
1188 if (kpage) {
1189 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1190 if (!err) {
1191 /*
1192 * The page was successfully merged:
1193 * add its rmap_item to the stable tree.
1194 */
1195 lock_page(kpage);
1196 stable_tree_append(rmap_item, page_stable_node(kpage));
1197 unlock_page(kpage);
1198 }
1199 put_page(kpage);
1200 return;
1201 }
1202
1203 /*
1204 * If the hash value of the page has changed from the last time
1205 * we calculated it, this page is changing frequently: therefore we
1206 * don't want to insert it in the unstable tree, and we don't want
1207 * to waste our time searching for something identical to it there.
1208 */
1209 checksum = calc_checksum(page);
1210 if (rmap_item->oldchecksum != checksum) {
1211 rmap_item->oldchecksum = checksum;
1212 return;
1213 }
1214
1215 tree_rmap_item =
1216 unstable_tree_search_insert(rmap_item, page, &tree_page);
1217 if (tree_rmap_item) {
1218 kpage = try_to_merge_two_pages(rmap_item, page,
1219 tree_rmap_item, tree_page);
1220 put_page(tree_page);
1221 /*
1222 * As soon as we merge this page, we want to remove the
1223 * rmap_item of the page we have merged with from the unstable
1224 * tree, and insert it instead as new node in the stable tree.
1225 */
1226 if (kpage) {
1227 remove_rmap_item_from_tree(tree_rmap_item);
1228
1229 lock_page(kpage);
1230 stable_node = stable_tree_insert(kpage);
1231 if (stable_node) {
1232 stable_tree_append(tree_rmap_item, stable_node);
1233 stable_tree_append(rmap_item, stable_node);
1234 }
1235 unlock_page(kpage);
1236
1237 /*
1238 * If we fail to insert the page into the stable tree,
1239 * we will have 2 virtual addresses that are pointing
1240 * to a ksm page left outside the stable tree,
1241 * in which case we need to break_cow on both.
1242 */
1243 if (!stable_node) {
1244 break_cow(tree_rmap_item);
1245 break_cow(rmap_item);
1246 }
1247 }
1248 }
1249}
1250
1251static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1252 struct rmap_item **rmap_list,
1253 unsigned long addr)
1254{
1255 struct rmap_item *rmap_item;
1256
1257 while (*rmap_list) {
1258 rmap_item = *rmap_list;
1259 if ((rmap_item->address & PAGE_MASK) == addr)
1260 return rmap_item;
1261 if (rmap_item->address > addr)
1262 break;
1263 *rmap_list = rmap_item->rmap_list;
1264 remove_rmap_item_from_tree(rmap_item);
1265 free_rmap_item(rmap_item);
1266 }
1267
1268 rmap_item = alloc_rmap_item();
1269 if (rmap_item) {
1270 /* It has already been zeroed */
1271 rmap_item->mm = mm_slot->mm;
1272 rmap_item->address = addr;
1273 rmap_item->rmap_list = *rmap_list;
1274 *rmap_list = rmap_item;
1275 }
1276 return rmap_item;
1277}
1278
1279static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1280{
1281 struct mm_struct *mm;
1282 struct mm_slot *slot;
1283 struct vm_area_struct *vma;
1284 struct rmap_item *rmap_item;
1285
1286 if (list_empty(&ksm_mm_head.mm_list))
1287 return NULL;
1288
1289 slot = ksm_scan.mm_slot;
1290 if (slot == &ksm_mm_head) {
1291 /*
1292 * A number of pages can hang around indefinitely on per-cpu
1293 * pagevecs, raised page count preventing write_protect_page
1294 * from merging them. Though it doesn't really matter much,
1295 * it is puzzling to see some stuck in pages_volatile until
1296 * other activity jostles them out, and they also prevented
1297 * LTP's KSM test from succeeding deterministically; so drain
1298 * them here (here rather than on entry to ksm_do_scan(),
1299 * so we don't IPI too often when pages_to_scan is set low).
1300 */
1301 lru_add_drain_all();
1302
1303 root_unstable_tree = RB_ROOT;
1304
1305 spin_lock(&ksm_mmlist_lock);
1306 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1307 ksm_scan.mm_slot = slot;
1308 spin_unlock(&ksm_mmlist_lock);
1309 /*
1310 * Although we tested list_empty() above, a racing __ksm_exit
1311 * of the last mm on the list may have removed it since then.
1312 */
1313 if (slot == &ksm_mm_head)
1314 return NULL;
1315next_mm:
1316 ksm_scan.address = 0;
1317 ksm_scan.rmap_list = &slot->rmap_list;
1318 }
1319
1320 mm = slot->mm;
1321 down_read(&mm->mmap_sem);
1322 if (ksm_test_exit(mm))
1323 vma = NULL;
1324 else
1325 vma = find_vma(mm, ksm_scan.address);
1326
1327 for (; vma; vma = vma->vm_next) {
1328 if (!(vma->vm_flags & VM_MERGEABLE))
1329 continue;
1330 if (ksm_scan.address < vma->vm_start)
1331 ksm_scan.address = vma->vm_start;
1332 if (!vma->anon_vma)
1333 ksm_scan.address = vma->vm_end;
1334
1335 while (ksm_scan.address < vma->vm_end) {
1336 if (ksm_test_exit(mm))
1337 break;
1338 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1339 if (IS_ERR_OR_NULL(*page)) {
1340 ksm_scan.address += PAGE_SIZE;
1341 cond_resched();
1342 continue;
1343 }
1344 if (PageAnon(*page) ||
1345 page_trans_compound_anon(*page)) {
1346 flush_anon_page(vma, *page, ksm_scan.address);
1347 flush_dcache_page(*page);
1348 rmap_item = get_next_rmap_item(slot,
1349 ksm_scan.rmap_list, ksm_scan.address);
1350 if (rmap_item) {
1351 ksm_scan.rmap_list =
1352 &rmap_item->rmap_list;
1353 ksm_scan.address += PAGE_SIZE;
1354 } else
1355 put_page(*page);
1356 up_read(&mm->mmap_sem);
1357 return rmap_item;
1358 }
1359 put_page(*page);
1360 ksm_scan.address += PAGE_SIZE;
1361 cond_resched();
1362 }
1363 }
1364
1365 if (ksm_test_exit(mm)) {
1366 ksm_scan.address = 0;
1367 ksm_scan.rmap_list = &slot->rmap_list;
1368 }
1369 /*
1370 * Nuke all the rmap_items that are above this current rmap:
1371 * because there were no VM_MERGEABLE vmas with such addresses.
1372 */
1373 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1374
1375 spin_lock(&ksm_mmlist_lock);
1376 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1377 struct mm_slot, mm_list);
1378 if (ksm_scan.address == 0) {
1379 /*
1380 * We've completed a full scan of all vmas, holding mmap_sem
1381 * throughout, and found no VM_MERGEABLE: so do the same as
1382 * __ksm_exit does to remove this mm from all our lists now.
1383 * This applies either when cleaning up after __ksm_exit
1384 * (but beware: we can reach here even before __ksm_exit),
1385 * or when all VM_MERGEABLE areas have been unmapped (and
1386 * mmap_sem then protects against race with MADV_MERGEABLE).
1387 */
1388 hlist_del(&slot->link);
1389 list_del(&slot->mm_list);
1390 spin_unlock(&ksm_mmlist_lock);
1391
1392 free_mm_slot(slot);
1393 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1394 up_read(&mm->mmap_sem);
1395 mmdrop(mm);
1396 } else {
1397 spin_unlock(&ksm_mmlist_lock);
1398 up_read(&mm->mmap_sem);
1399 }
1400
1401 /* Repeat until we've completed scanning the whole list */
1402 slot = ksm_scan.mm_slot;
1403 if (slot != &ksm_mm_head)
1404 goto next_mm;
1405
1406 ksm_scan.seqnr++;
1407 return NULL;
1408}
1409
1410/**
1411 * ksm_do_scan - the ksm scanner main worker function.
1412 * @scan_npages - number of pages we want to scan before we return.
1413 */
1414static void ksm_do_scan(unsigned int scan_npages)
1415{
1416 struct rmap_item *rmap_item;
1417 struct page *uninitialized_var(page);
1418
1419 while (scan_npages-- && likely(!freezing(current))) {
1420 cond_resched();
1421 rmap_item = scan_get_next_rmap_item(&page);
1422 if (!rmap_item)
1423 return;
1424 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1425 cmp_and_merge_page(page, rmap_item);
1426 put_page(page);
1427 }
1428}
1429
1430static int ksmd_should_run(void)
1431{
1432 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1433}
1434
1435static int ksm_scan_thread(void *nothing)
1436{
1437 set_freezable();
1438 set_user_nice(current, 5);
1439
1440 while (!kthread_should_stop()) {
1441 mutex_lock(&ksm_thread_mutex);
1442 if (ksmd_should_run())
1443 ksm_do_scan(ksm_thread_pages_to_scan);
1444 mutex_unlock(&ksm_thread_mutex);
1445
1446 try_to_freeze();
1447
1448 if (ksmd_should_run()) {
1449 schedule_timeout_interruptible(
1450 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1451 } else {
1452 wait_event_freezable(ksm_thread_wait,
1453 ksmd_should_run() || kthread_should_stop());
1454 }
1455 }
1456 return 0;
1457}
1458
1459int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1460 unsigned long end, int advice, unsigned long *vm_flags)
1461{
1462 struct mm_struct *mm = vma->vm_mm;
1463 int err;
1464
1465 switch (advice) {
1466 case MADV_MERGEABLE:
1467 /*
1468 * Be somewhat over-protective for now!
1469 */
1470 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1471 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1472 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1473 VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
1474 return 0; /* just ignore the advice */
1475
1476 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1477 err = __ksm_enter(mm);
1478 if (err)
1479 return err;
1480 }
1481
1482 *vm_flags |= VM_MERGEABLE;
1483 break;
1484
1485 case MADV_UNMERGEABLE:
1486 if (!(*vm_flags & VM_MERGEABLE))
1487 return 0; /* just ignore the advice */
1488
1489 if (vma->anon_vma) {
1490 err = unmerge_ksm_pages(vma, start, end);
1491 if (err)
1492 return err;
1493 }
1494
1495 *vm_flags &= ~VM_MERGEABLE;
1496 break;
1497 }
1498
1499 return 0;
1500}
1501
1502int __ksm_enter(struct mm_struct *mm)
1503{
1504 struct mm_slot *mm_slot;
1505 int needs_wakeup;
1506
1507 mm_slot = alloc_mm_slot();
1508 if (!mm_slot)
1509 return -ENOMEM;
1510
1511 /* Check ksm_run too? Would need tighter locking */
1512 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1513
1514 spin_lock(&ksm_mmlist_lock);
1515 insert_to_mm_slots_hash(mm, mm_slot);
1516 /*
1517 * Insert just behind the scanning cursor, to let the area settle
1518 * down a little; when fork is followed by immediate exec, we don't
1519 * want ksmd to waste time setting up and tearing down an rmap_list.
1520 */
1521 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1522 spin_unlock(&ksm_mmlist_lock);
1523
1524 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1525 atomic_inc(&mm->mm_count);
1526
1527 if (needs_wakeup)
1528 wake_up_interruptible(&ksm_thread_wait);
1529
1530 return 0;
1531}
1532
1533void __ksm_exit(struct mm_struct *mm)
1534{
1535 struct mm_slot *mm_slot;
1536 int easy_to_free = 0;
1537
1538 /*
1539 * This process is exiting: if it's straightforward (as is the
1540 * case when ksmd was never running), free mm_slot immediately.
1541 * But if it's at the cursor or has rmap_items linked to it, use
1542 * mmap_sem to synchronize with any break_cows before pagetables
1543 * are freed, and leave the mm_slot on the list for ksmd to free.
1544 * Beware: ksm may already have noticed it exiting and freed the slot.
1545 */
1546
1547 spin_lock(&ksm_mmlist_lock);
1548 mm_slot = get_mm_slot(mm);
1549 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1550 if (!mm_slot->rmap_list) {
1551 hlist_del(&mm_slot->link);
1552 list_del(&mm_slot->mm_list);
1553 easy_to_free = 1;
1554 } else {
1555 list_move(&mm_slot->mm_list,
1556 &ksm_scan.mm_slot->mm_list);
1557 }
1558 }
1559 spin_unlock(&ksm_mmlist_lock);
1560
1561 if (easy_to_free) {
1562 free_mm_slot(mm_slot);
1563 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1564 mmdrop(mm);
1565 } else if (mm_slot) {
1566 down_write(&mm->mmap_sem);
1567 up_write(&mm->mmap_sem);
1568 }
1569}
1570
1571struct page *ksm_does_need_to_copy(struct page *page,
1572 struct vm_area_struct *vma, unsigned long address)
1573{
1574 struct page *new_page;
1575
1576 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1577 if (new_page) {
1578 copy_user_highpage(new_page, page, address, vma);
1579
1580 SetPageDirty(new_page);
1581 __SetPageUptodate(new_page);
1582 SetPageSwapBacked(new_page);
1583 __set_page_locked(new_page);
1584
1585 if (page_evictable(new_page, vma))
1586 lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1587 else
1588 add_page_to_unevictable_list(new_page);
1589 }
1590
1591 return new_page;
1592}
1593
1594int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1595 unsigned long *vm_flags)
1596{
1597 struct stable_node *stable_node;
1598 struct rmap_item *rmap_item;
1599 struct hlist_node *hlist;
1600 unsigned int mapcount = page_mapcount(page);
1601 int referenced = 0;
1602 int search_new_forks = 0;
1603
1604 VM_BUG_ON(!PageKsm(page));
1605 VM_BUG_ON(!PageLocked(page));
1606
1607 stable_node = page_stable_node(page);
1608 if (!stable_node)
1609 return 0;
1610again:
1611 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1612 struct anon_vma *anon_vma = rmap_item->anon_vma;
1613 struct anon_vma_chain *vmac;
1614 struct vm_area_struct *vma;
1615
1616 anon_vma_lock(anon_vma);
1617 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1618 vma = vmac->vma;
1619 if (rmap_item->address < vma->vm_start ||
1620 rmap_item->address >= vma->vm_end)
1621 continue;
1622 /*
1623 * Initially we examine only the vma which covers this
1624 * rmap_item; but later, if there is still work to do,
1625 * we examine covering vmas in other mms: in case they
1626 * were forked from the original since ksmd passed.
1627 */
1628 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1629 continue;
1630
1631 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1632 continue;
1633
1634 referenced += page_referenced_one(page, vma,
1635 rmap_item->address, &mapcount, vm_flags);
1636 if (!search_new_forks || !mapcount)
1637 break;
1638 }
1639 anon_vma_unlock(anon_vma);
1640 if (!mapcount)
1641 goto out;
1642 }
1643 if (!search_new_forks++)
1644 goto again;
1645out:
1646 return referenced;
1647}
1648
1649int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1650{
1651 struct stable_node *stable_node;
1652 struct hlist_node *hlist;
1653 struct rmap_item *rmap_item;
1654 int ret = SWAP_AGAIN;
1655 int search_new_forks = 0;
1656
1657 VM_BUG_ON(!PageKsm(page));
1658 VM_BUG_ON(!PageLocked(page));
1659
1660 stable_node = page_stable_node(page);
1661 if (!stable_node)
1662 return SWAP_FAIL;
1663again:
1664 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1665 struct anon_vma *anon_vma = rmap_item->anon_vma;
1666 struct anon_vma_chain *vmac;
1667 struct vm_area_struct *vma;
1668
1669 anon_vma_lock(anon_vma);
1670 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1671 vma = vmac->vma;
1672 if (rmap_item->address < vma->vm_start ||
1673 rmap_item->address >= vma->vm_end)
1674 continue;
1675 /*
1676 * Initially we examine only the vma which covers this
1677 * rmap_item; but later, if there is still work to do,
1678 * we examine covering vmas in other mms: in case they
1679 * were forked from the original since ksmd passed.
1680 */
1681 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1682 continue;
1683
1684 ret = try_to_unmap_one(page, vma,
1685 rmap_item->address, flags);
1686 if (ret != SWAP_AGAIN || !page_mapped(page)) {
1687 anon_vma_unlock(anon_vma);
1688 goto out;
1689 }
1690 }
1691 anon_vma_unlock(anon_vma);
1692 }
1693 if (!search_new_forks++)
1694 goto again;
1695out:
1696 return ret;
1697}
1698
1699#ifdef CONFIG_MIGRATION
1700int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1701 struct vm_area_struct *, unsigned long, void *), void *arg)
1702{
1703 struct stable_node *stable_node;
1704 struct hlist_node *hlist;
1705 struct rmap_item *rmap_item;
1706 int ret = SWAP_AGAIN;
1707 int search_new_forks = 0;
1708
1709 VM_BUG_ON(!PageKsm(page));
1710 VM_BUG_ON(!PageLocked(page));
1711
1712 stable_node = page_stable_node(page);
1713 if (!stable_node)
1714 return ret;
1715again:
1716 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1717 struct anon_vma *anon_vma = rmap_item->anon_vma;
1718 struct anon_vma_chain *vmac;
1719 struct vm_area_struct *vma;
1720
1721 anon_vma_lock(anon_vma);
1722 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1723 vma = vmac->vma;
1724 if (rmap_item->address < vma->vm_start ||
1725 rmap_item->address >= vma->vm_end)
1726 continue;
1727 /*
1728 * Initially we examine only the vma which covers this
1729 * rmap_item; but later, if there is still work to do,
1730 * we examine covering vmas in other mms: in case they
1731 * were forked from the original since ksmd passed.
1732 */
1733 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1734 continue;
1735
1736 ret = rmap_one(page, vma, rmap_item->address, arg);
1737 if (ret != SWAP_AGAIN) {
1738 anon_vma_unlock(anon_vma);
1739 goto out;
1740 }
1741 }
1742 anon_vma_unlock(anon_vma);
1743 }
1744 if (!search_new_forks++)
1745 goto again;
1746out:
1747 return ret;
1748}
1749
1750void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1751{
1752 struct stable_node *stable_node;
1753
1754 VM_BUG_ON(!PageLocked(oldpage));
1755 VM_BUG_ON(!PageLocked(newpage));
1756 VM_BUG_ON(newpage->mapping != oldpage->mapping);
1757
1758 stable_node = page_stable_node(newpage);
1759 if (stable_node) {
1760 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1761 stable_node->kpfn = page_to_pfn(newpage);
1762 }
1763}
1764#endif /* CONFIG_MIGRATION */
1765
1766#ifdef CONFIG_MEMORY_HOTREMOVE
1767static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1768 unsigned long end_pfn)
1769{
1770 struct rb_node *node;
1771
1772 for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1773 struct stable_node *stable_node;
1774
1775 stable_node = rb_entry(node, struct stable_node, node);
1776 if (stable_node->kpfn >= start_pfn &&
1777 stable_node->kpfn < end_pfn)
1778 return stable_node;
1779 }
1780 return NULL;
1781}
1782
1783static int ksm_memory_callback(struct notifier_block *self,
1784 unsigned long action, void *arg)
1785{
1786 struct memory_notify *mn = arg;
1787 struct stable_node *stable_node;
1788
1789 switch (action) {
1790 case MEM_GOING_OFFLINE:
1791 /*
1792 * Keep it very simple for now: just lock out ksmd and
1793 * MADV_UNMERGEABLE while any memory is going offline.
1794 * mutex_lock_nested() is necessary because lockdep was alarmed
1795 * that here we take ksm_thread_mutex inside notifier chain
1796 * mutex, and later take notifier chain mutex inside
1797 * ksm_thread_mutex to unlock it. But that's safe because both
1798 * are inside mem_hotplug_mutex.
1799 */
1800 mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
1801 break;
1802
1803 case MEM_OFFLINE:
1804 /*
1805 * Most of the work is done by page migration; but there might
1806 * be a few stable_nodes left over, still pointing to struct
1807 * pages which have been offlined: prune those from the tree.
1808 */
1809 while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1810 mn->start_pfn + mn->nr_pages)) != NULL)
1811 remove_node_from_stable_tree(stable_node);
1812 /* fallthrough */
1813
1814 case MEM_CANCEL_OFFLINE:
1815 mutex_unlock(&ksm_thread_mutex);
1816 break;
1817 }
1818 return NOTIFY_OK;
1819}
1820#endif /* CONFIG_MEMORY_HOTREMOVE */
1821
1822#ifdef CONFIG_SYSFS
1823/*
1824 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1825 */
1826
1827#define KSM_ATTR_RO(_name) \
1828 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1829#define KSM_ATTR(_name) \
1830 static struct kobj_attribute _name##_attr = \
1831 __ATTR(_name, 0644, _name##_show, _name##_store)
1832
1833static ssize_t sleep_millisecs_show(struct kobject *kobj,
1834 struct kobj_attribute *attr, char *buf)
1835{
1836 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1837}
1838
1839static ssize_t sleep_millisecs_store(struct kobject *kobj,
1840 struct kobj_attribute *attr,
1841 const char *buf, size_t count)
1842{
1843 unsigned long msecs;
1844 int err;
1845
1846 err = strict_strtoul(buf, 10, &msecs);
1847 if (err || msecs > UINT_MAX)
1848 return -EINVAL;
1849
1850 ksm_thread_sleep_millisecs = msecs;
1851
1852 return count;
1853}
1854KSM_ATTR(sleep_millisecs);
1855
1856static ssize_t pages_to_scan_show(struct kobject *kobj,
1857 struct kobj_attribute *attr, char *buf)
1858{
1859 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1860}
1861
1862static ssize_t pages_to_scan_store(struct kobject *kobj,
1863 struct kobj_attribute *attr,
1864 const char *buf, size_t count)
1865{
1866 int err;
1867 unsigned long nr_pages;
1868
1869 err = strict_strtoul(buf, 10, &nr_pages);
1870 if (err || nr_pages > UINT_MAX)
1871 return -EINVAL;
1872
1873 ksm_thread_pages_to_scan = nr_pages;
1874
1875 return count;
1876}
1877KSM_ATTR(pages_to_scan);
1878
1879static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1880 char *buf)
1881{
1882 return sprintf(buf, "%u\n", ksm_run);
1883}
1884
1885static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1886 const char *buf, size_t count)
1887{
1888 int err;
1889 unsigned long flags;
1890
1891 err = strict_strtoul(buf, 10, &flags);
1892 if (err || flags > UINT_MAX)
1893 return -EINVAL;
1894 if (flags > KSM_RUN_UNMERGE)
1895 return -EINVAL;
1896
1897 /*
1898 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1899 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1900 * breaking COW to free the pages_shared (but leaves mm_slots
1901 * on the list for when ksmd may be set running again).
1902 */
1903
1904 mutex_lock(&ksm_thread_mutex);
1905 if (ksm_run != flags) {
1906 ksm_run = flags;
1907 if (flags & KSM_RUN_UNMERGE) {
1908 int oom_score_adj;
1909
1910 oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
1911 err = unmerge_and_remove_all_rmap_items();
1912 compare_swap_oom_score_adj(OOM_SCORE_ADJ_MAX,
1913 oom_score_adj);
1914 if (err) {
1915 ksm_run = KSM_RUN_STOP;
1916 count = err;
1917 }
1918 }
1919 }
1920 mutex_unlock(&ksm_thread_mutex);
1921
1922 if (flags & KSM_RUN_MERGE)
1923 wake_up_interruptible(&ksm_thread_wait);
1924
1925 return count;
1926}
1927KSM_ATTR(run);
1928
1929static ssize_t pages_shared_show(struct kobject *kobj,
1930 struct kobj_attribute *attr, char *buf)
1931{
1932 return sprintf(buf, "%lu\n", ksm_pages_shared);
1933}
1934KSM_ATTR_RO(pages_shared);
1935
1936static ssize_t pages_sharing_show(struct kobject *kobj,
1937 struct kobj_attribute *attr, char *buf)
1938{
1939 return sprintf(buf, "%lu\n", ksm_pages_sharing);
1940}
1941KSM_ATTR_RO(pages_sharing);
1942
1943static ssize_t pages_unshared_show(struct kobject *kobj,
1944 struct kobj_attribute *attr, char *buf)
1945{
1946 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1947}
1948KSM_ATTR_RO(pages_unshared);
1949
1950static ssize_t pages_volatile_show(struct kobject *kobj,
1951 struct kobj_attribute *attr, char *buf)
1952{
1953 long ksm_pages_volatile;
1954
1955 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1956 - ksm_pages_sharing - ksm_pages_unshared;
1957 /*
1958 * It was not worth any locking to calculate that statistic,
1959 * but it might therefore sometimes be negative: conceal that.
1960 */
1961 if (ksm_pages_volatile < 0)
1962 ksm_pages_volatile = 0;
1963 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1964}
1965KSM_ATTR_RO(pages_volatile);
1966
1967static ssize_t full_scans_show(struct kobject *kobj,
1968 struct kobj_attribute *attr, char *buf)
1969{
1970 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1971}
1972KSM_ATTR_RO(full_scans);
1973
1974static struct attribute *ksm_attrs[] = {
1975 &sleep_millisecs_attr.attr,
1976 &pages_to_scan_attr.attr,
1977 &run_attr.attr,
1978 &pages_shared_attr.attr,
1979 &pages_sharing_attr.attr,
1980 &pages_unshared_attr.attr,
1981 &pages_volatile_attr.attr,
1982 &full_scans_attr.attr,
1983 NULL,
1984};
1985
1986static struct attribute_group ksm_attr_group = {
1987 .attrs = ksm_attrs,
1988 .name = "ksm",
1989};
1990#endif /* CONFIG_SYSFS */
1991
1992static int __init ksm_init(void)
1993{
1994 struct task_struct *ksm_thread;
1995 int err;
1996
1997 err = ksm_slab_init();
1998 if (err)
1999 goto out;
2000
2001 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2002 if (IS_ERR(ksm_thread)) {
2003 printk(KERN_ERR "ksm: creating kthread failed\n");
2004 err = PTR_ERR(ksm_thread);
2005 goto out_free;
2006 }
2007
2008#ifdef CONFIG_SYSFS
2009 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2010 if (err) {
2011 printk(KERN_ERR "ksm: register sysfs failed\n");
2012 kthread_stop(ksm_thread);
2013 goto out_free;
2014 }
2015#else
2016 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2017
2018#endif /* CONFIG_SYSFS */
2019
2020#ifdef CONFIG_MEMORY_HOTREMOVE
2021 /*
2022 * Choose a high priority since the callback takes ksm_thread_mutex:
2023 * later callbacks could only be taking locks which nest within that.
2024 */
2025 hotplug_memory_notifier(ksm_memory_callback, 100);
2026#endif
2027 return 0;
2028
2029out_free:
2030 ksm_slab_free();
2031out:
2032 return err;
2033}
2034module_init(ksm_init)