]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blame - mm/ksm.c
staging: fsl-dpaa2/eth: Only store bpid in priv struct
[mirror_ubuntu-bionic-kernel.git] / mm / ksm.c
CommitLineData
f8af4da3 1/*
31dbd01f
IE
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 *
36b2528d 7 * Copyright (C) 2008-2009 Red Hat, Inc.
31dbd01f
IE
8 * Authors:
9 * Izik Eidus
10 * Andrea Arcangeli
11 * Chris Wright
36b2528d 12 * Hugh Dickins
31dbd01f
IE
13 *
14 * This work is licensed under the terms of the GNU GPL, version 2.
f8af4da3
HD
15 */
16
17#include <linux/errno.h>
31dbd01f
IE
18#include <linux/mm.h>
19#include <linux/fs.h>
f8af4da3 20#include <linux/mman.h>
31dbd01f 21#include <linux/sched.h>
6e84f315 22#include <linux/sched/mm.h>
f7ccbae4 23#include <linux/sched/coredump.h>
31dbd01f
IE
24#include <linux/rwsem.h>
25#include <linux/pagemap.h>
26#include <linux/rmap.h>
27#include <linux/spinlock.h>
28#include <linux/jhash.h>
29#include <linux/delay.h>
30#include <linux/kthread.h>
31#include <linux/wait.h>
32#include <linux/slab.h>
33#include <linux/rbtree.h>
62b61f61 34#include <linux/memory.h>
31dbd01f 35#include <linux/mmu_notifier.h>
2c6854fd 36#include <linux/swap.h>
f8af4da3 37#include <linux/ksm.h>
4ca3a69b 38#include <linux/hashtable.h>
878aee7d 39#include <linux/freezer.h>
72788c38 40#include <linux/oom.h>
90bd6fd3 41#include <linux/numa.h>
f8af4da3 42
31dbd01f 43#include <asm/tlbflush.h>
73848b46 44#include "internal.h"
31dbd01f 45
e850dcf5
HD
46#ifdef CONFIG_NUMA
47#define NUMA(x) (x)
48#define DO_NUMA(x) do { (x); } while (0)
49#else
50#define NUMA(x) (0)
51#define DO_NUMA(x) do { } while (0)
52#endif
53
31dbd01f
IE
54/*
55 * A few notes about the KSM scanning process,
56 * to make it easier to understand the data structures below:
57 *
58 * In order to reduce excessive scanning, KSM sorts the memory pages by their
59 * contents into a data structure that holds pointers to the pages' locations.
60 *
61 * Since the contents of the pages may change at any moment, KSM cannot just
62 * insert the pages into a normal sorted tree and expect it to find anything.
63 * Therefore KSM uses two data structures - the stable and the unstable tree.
64 *
65 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
66 * by their contents. Because each such page is write-protected, searching on
67 * this tree is fully assured to be working (except when pages are unmapped),
68 * and therefore this tree is called the stable tree.
69 *
70 * In addition to the stable tree, KSM uses a second data structure called the
71 * unstable tree: this tree holds pointers to pages which have been found to
72 * be "unchanged for a period of time". The unstable tree sorts these pages
73 * by their contents, but since they are not write-protected, KSM cannot rely
74 * upon the unstable tree to work correctly - the unstable tree is liable to
75 * be corrupted as its contents are modified, and so it is called unstable.
76 *
77 * KSM solves this problem by several techniques:
78 *
79 * 1) The unstable tree is flushed every time KSM completes scanning all
80 * memory areas, and then the tree is rebuilt again from the beginning.
81 * 2) KSM will only insert into the unstable tree, pages whose hash value
82 * has not changed since the previous scan of all memory areas.
83 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
84 * colors of the nodes and not on their contents, assuring that even when
85 * the tree gets "corrupted" it won't get out of balance, so scanning time
86 * remains the same (also, searching and inserting nodes in an rbtree uses
87 * the same algorithm, so we have no overhead when we flush and rebuild).
88 * 4) KSM never flushes the stable tree, which means that even if it were to
89 * take 10 attempts to find a page in the unstable tree, once it is found,
90 * it is secured in the stable tree. (When we scan a new page, we first
91 * compare it against the stable tree, and then against the unstable tree.)
8fdb3dbf
HD
92 *
93 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
94 * stable trees and multiple unstable trees: one of each for each NUMA node.
31dbd01f
IE
95 */
96
97/**
98 * struct mm_slot - ksm information per mm that is being scanned
99 * @link: link to the mm_slots hash list
100 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
6514d511 101 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
31dbd01f
IE
102 * @mm: the mm that this information is valid for
103 */
104struct mm_slot {
105 struct hlist_node link;
106 struct list_head mm_list;
6514d511 107 struct rmap_item *rmap_list;
31dbd01f
IE
108 struct mm_struct *mm;
109};
110
111/**
112 * struct ksm_scan - cursor for scanning
113 * @mm_slot: the current mm_slot we are scanning
114 * @address: the next address inside that to be scanned
6514d511 115 * @rmap_list: link to the next rmap to be scanned in the rmap_list
31dbd01f
IE
116 * @seqnr: count of completed full scans (needed when removing unstable node)
117 *
118 * There is only the one ksm_scan instance of this cursor structure.
119 */
120struct ksm_scan {
121 struct mm_slot *mm_slot;
122 unsigned long address;
6514d511 123 struct rmap_item **rmap_list;
31dbd01f
IE
124 unsigned long seqnr;
125};
126
7b6ba2c7
HD
127/**
128 * struct stable_node - node of the stable rbtree
129 * @node: rb node of this ksm page in the stable tree
4146d2d6
HD
130 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
131 * @list: linked into migrate_nodes, pending placement in the proper node tree
7b6ba2c7 132 * @hlist: hlist head of rmap_items using this ksm page
4146d2d6
HD
133 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
134 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
7b6ba2c7
HD
135 */
136struct stable_node {
4146d2d6
HD
137 union {
138 struct rb_node node; /* when node of stable tree */
139 struct { /* when listed for migration */
140 struct list_head *head;
141 struct list_head list;
142 };
143 };
7b6ba2c7 144 struct hlist_head hlist;
62b61f61 145 unsigned long kpfn;
4146d2d6
HD
146#ifdef CONFIG_NUMA
147 int nid;
148#endif
7b6ba2c7
HD
149};
150
31dbd01f
IE
151/**
152 * struct rmap_item - reverse mapping item for virtual addresses
6514d511 153 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
db114b83 154 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
bc56620b 155 * @nid: NUMA node id of unstable tree in which linked (may not match page)
31dbd01f
IE
156 * @mm: the memory structure this rmap_item is pointing into
157 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
158 * @oldchecksum: previous checksum of the page at that virtual address
7b6ba2c7
HD
159 * @node: rb node of this rmap_item in the unstable tree
160 * @head: pointer to stable_node heading this list in the stable tree
161 * @hlist: link into hlist of rmap_items hanging off that stable_node
31dbd01f
IE
162 */
163struct rmap_item {
6514d511 164 struct rmap_item *rmap_list;
bc56620b
HD
165 union {
166 struct anon_vma *anon_vma; /* when stable */
167#ifdef CONFIG_NUMA
168 int nid; /* when node of unstable tree */
169#endif
170 };
31dbd01f
IE
171 struct mm_struct *mm;
172 unsigned long address; /* + low bits used for flags below */
7b6ba2c7 173 unsigned int oldchecksum; /* when unstable */
31dbd01f 174 union {
7b6ba2c7
HD
175 struct rb_node node; /* when node of unstable tree */
176 struct { /* when listed from stable tree */
177 struct stable_node *head;
178 struct hlist_node hlist;
179 };
31dbd01f
IE
180 };
181};
182
183#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
7b6ba2c7
HD
184#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
185#define STABLE_FLAG 0x200 /* is listed from the stable tree */
31dbd01f
IE
186
187/* The stable and unstable tree heads */
ef53d16c
HD
188static struct rb_root one_stable_tree[1] = { RB_ROOT };
189static struct rb_root one_unstable_tree[1] = { RB_ROOT };
190static struct rb_root *root_stable_tree = one_stable_tree;
191static struct rb_root *root_unstable_tree = one_unstable_tree;
31dbd01f 192
4146d2d6
HD
193/* Recently migrated nodes of stable tree, pending proper placement */
194static LIST_HEAD(migrate_nodes);
195
4ca3a69b
SL
196#define MM_SLOTS_HASH_BITS 10
197static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
31dbd01f
IE
198
199static struct mm_slot ksm_mm_head = {
200 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
201};
202static struct ksm_scan ksm_scan = {
203 .mm_slot = &ksm_mm_head,
204};
205
206static struct kmem_cache *rmap_item_cache;
7b6ba2c7 207static struct kmem_cache *stable_node_cache;
31dbd01f
IE
208static struct kmem_cache *mm_slot_cache;
209
210/* The number of nodes in the stable tree */
b4028260 211static unsigned long ksm_pages_shared;
31dbd01f 212
e178dfde 213/* The number of page slots additionally sharing those nodes */
b4028260 214static unsigned long ksm_pages_sharing;
31dbd01f 215
473b0ce4
HD
216/* The number of nodes in the unstable tree */
217static unsigned long ksm_pages_unshared;
218
219/* The number of rmap_items in use: to calculate pages_volatile */
220static unsigned long ksm_rmap_items;
221
31dbd01f 222/* Number of pages ksmd should scan in one batch */
2c6854fd 223static unsigned int ksm_thread_pages_to_scan = 100;
31dbd01f
IE
224
225/* Milliseconds ksmd should sleep between batches */
2ffd8679 226static unsigned int ksm_thread_sleep_millisecs = 20;
31dbd01f 227
e86c59b1
CI
228/* Checksum of an empty (zeroed) page */
229static unsigned int zero_checksum __read_mostly;
230
231/* Whether to merge empty (zeroed) pages with actual zero pages */
232static bool ksm_use_zero_pages __read_mostly;
233
e850dcf5 234#ifdef CONFIG_NUMA
90bd6fd3
PH
235/* Zeroed when merging across nodes is not allowed */
236static unsigned int ksm_merge_across_nodes = 1;
ef53d16c 237static int ksm_nr_node_ids = 1;
e850dcf5
HD
238#else
239#define ksm_merge_across_nodes 1U
ef53d16c 240#define ksm_nr_node_ids 1
e850dcf5 241#endif
90bd6fd3 242
31dbd01f
IE
243#define KSM_RUN_STOP 0
244#define KSM_RUN_MERGE 1
245#define KSM_RUN_UNMERGE 2
ef4d43a8
HD
246#define KSM_RUN_OFFLINE 4
247static unsigned long ksm_run = KSM_RUN_STOP;
248static void wait_while_offlining(void);
31dbd01f
IE
249
250static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
251static DEFINE_MUTEX(ksm_thread_mutex);
252static DEFINE_SPINLOCK(ksm_mmlist_lock);
253
254#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
255 sizeof(struct __struct), __alignof__(struct __struct),\
256 (__flags), NULL)
257
258static int __init ksm_slab_init(void)
259{
260 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
261 if (!rmap_item_cache)
262 goto out;
263
7b6ba2c7
HD
264 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
265 if (!stable_node_cache)
266 goto out_free1;
267
31dbd01f
IE
268 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
269 if (!mm_slot_cache)
7b6ba2c7 270 goto out_free2;
31dbd01f
IE
271
272 return 0;
273
7b6ba2c7
HD
274out_free2:
275 kmem_cache_destroy(stable_node_cache);
276out_free1:
31dbd01f
IE
277 kmem_cache_destroy(rmap_item_cache);
278out:
279 return -ENOMEM;
280}
281
282static void __init ksm_slab_free(void)
283{
284 kmem_cache_destroy(mm_slot_cache);
7b6ba2c7 285 kmem_cache_destroy(stable_node_cache);
31dbd01f
IE
286 kmem_cache_destroy(rmap_item_cache);
287 mm_slot_cache = NULL;
288}
289
290static inline struct rmap_item *alloc_rmap_item(void)
291{
473b0ce4
HD
292 struct rmap_item *rmap_item;
293
5b398e41 294 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
295 __GFP_NORETRY | __GFP_NOWARN);
473b0ce4
HD
296 if (rmap_item)
297 ksm_rmap_items++;
298 return rmap_item;
31dbd01f
IE
299}
300
301static inline void free_rmap_item(struct rmap_item *rmap_item)
302{
473b0ce4 303 ksm_rmap_items--;
31dbd01f
IE
304 rmap_item->mm = NULL; /* debug safety */
305 kmem_cache_free(rmap_item_cache, rmap_item);
306}
307
7b6ba2c7
HD
308static inline struct stable_node *alloc_stable_node(void)
309{
6213055f 310 /*
311 * The allocation can take too long with GFP_KERNEL when memory is under
312 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
313 * grants access to memory reserves, helping to avoid this problem.
314 */
315 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
7b6ba2c7
HD
316}
317
318static inline void free_stable_node(struct stable_node *stable_node)
319{
320 kmem_cache_free(stable_node_cache, stable_node);
321}
322
31dbd01f
IE
323static inline struct mm_slot *alloc_mm_slot(void)
324{
325 if (!mm_slot_cache) /* initialization failed */
326 return NULL;
327 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
328}
329
330static inline void free_mm_slot(struct mm_slot *mm_slot)
331{
332 kmem_cache_free(mm_slot_cache, mm_slot);
333}
334
31dbd01f
IE
335static struct mm_slot *get_mm_slot(struct mm_struct *mm)
336{
4ca3a69b
SL
337 struct mm_slot *slot;
338
b67bfe0d 339 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
4ca3a69b
SL
340 if (slot->mm == mm)
341 return slot;
31dbd01f 342
31dbd01f
IE
343 return NULL;
344}
345
346static void insert_to_mm_slots_hash(struct mm_struct *mm,
347 struct mm_slot *mm_slot)
348{
31dbd01f 349 mm_slot->mm = mm;
4ca3a69b 350 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
31dbd01f
IE
351}
352
a913e182
HD
353/*
354 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
355 * page tables after it has passed through ksm_exit() - which, if necessary,
356 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
357 * a special flag: they can just back out as soon as mm_users goes to zero.
358 * ksm_test_exit() is used throughout to make this test for exit: in some
359 * places for correctness, in some places just to avoid unnecessary work.
360 */
361static inline bool ksm_test_exit(struct mm_struct *mm)
362{
363 return atomic_read(&mm->mm_users) == 0;
364}
365
31dbd01f
IE
366/*
367 * We use break_ksm to break COW on a ksm page: it's a stripped down
368 *
d4edcf0d 369 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
31dbd01f
IE
370 * put_page(page);
371 *
372 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
373 * in case the application has unmapped and remapped mm,addr meanwhile.
374 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
375 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
1b2ee126
DH
376 *
377 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
378 * of the process that owns 'vma'. We also do not want to enforce
379 * protection keys here anyway.
31dbd01f 380 */
d952b791 381static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
31dbd01f
IE
382{
383 struct page *page;
d952b791 384 int ret = 0;
31dbd01f
IE
385
386 do {
387 cond_resched();
1b2ee126
DH
388 page = follow_page(vma, addr,
389 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
22eccdd7 390 if (IS_ERR_OR_NULL(page))
31dbd01f
IE
391 break;
392 if (PageKsm(page))
dcddffd4
KS
393 ret = handle_mm_fault(vma, addr,
394 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
31dbd01f
IE
395 else
396 ret = VM_FAULT_WRITE;
397 put_page(page);
33692f27 398 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
d952b791
HD
399 /*
400 * We must loop because handle_mm_fault() may back out if there's
401 * any difficulty e.g. if pte accessed bit gets updated concurrently.
402 *
403 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
404 * COW has been broken, even if the vma does not permit VM_WRITE;
405 * but note that a concurrent fault might break PageKsm for us.
406 *
407 * VM_FAULT_SIGBUS could occur if we race with truncation of the
408 * backing file, which also invalidates anonymous pages: that's
409 * okay, that truncation will have unmapped the PageKsm for us.
410 *
411 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
412 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
413 * current task has TIF_MEMDIE set, and will be OOM killed on return
414 * to user; and ksmd, having no mm, would never be chosen for that.
415 *
416 * But if the mm is in a limited mem_cgroup, then the fault may fail
417 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
418 * even ksmd can fail in this way - though it's usually breaking ksm
419 * just to undo a merge it made a moment before, so unlikely to oom.
420 *
421 * That's a pity: we might therefore have more kernel pages allocated
422 * than we're counting as nodes in the stable tree; but ksm_do_scan
423 * will retry to break_cow on each pass, so should recover the page
424 * in due course. The important thing is to not let VM_MERGEABLE
425 * be cleared while any such pages might remain in the area.
426 */
427 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
31dbd01f
IE
428}
429
ef694222
BL
430static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
431 unsigned long addr)
432{
433 struct vm_area_struct *vma;
434 if (ksm_test_exit(mm))
435 return NULL;
436 vma = find_vma(mm, addr);
437 if (!vma || vma->vm_start > addr)
438 return NULL;
439 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
440 return NULL;
441 return vma;
442}
443
8dd3557a 444static void break_cow(struct rmap_item *rmap_item)
31dbd01f 445{
8dd3557a
HD
446 struct mm_struct *mm = rmap_item->mm;
447 unsigned long addr = rmap_item->address;
31dbd01f
IE
448 struct vm_area_struct *vma;
449
4035c07a
HD
450 /*
451 * It is not an accident that whenever we want to break COW
452 * to undo, we also need to drop a reference to the anon_vma.
453 */
9e60109f 454 put_anon_vma(rmap_item->anon_vma);
4035c07a 455
81464e30 456 down_read(&mm->mmap_sem);
ef694222
BL
457 vma = find_mergeable_vma(mm, addr);
458 if (vma)
459 break_ksm(vma, addr);
31dbd01f
IE
460 up_read(&mm->mmap_sem);
461}
462
463static struct page *get_mergeable_page(struct rmap_item *rmap_item)
464{
465 struct mm_struct *mm = rmap_item->mm;
466 unsigned long addr = rmap_item->address;
467 struct vm_area_struct *vma;
468 struct page *page;
469
470 down_read(&mm->mmap_sem);
ef694222
BL
471 vma = find_mergeable_vma(mm, addr);
472 if (!vma)
31dbd01f
IE
473 goto out;
474
475 page = follow_page(vma, addr, FOLL_GET);
22eccdd7 476 if (IS_ERR_OR_NULL(page))
31dbd01f 477 goto out;
f765f540 478 if (PageAnon(page)) {
31dbd01f
IE
479 flush_anon_page(vma, page, addr);
480 flush_dcache_page(page);
481 } else {
482 put_page(page);
c8f95ed1
AA
483out:
484 page = NULL;
31dbd01f
IE
485 }
486 up_read(&mm->mmap_sem);
487 return page;
488}
489
90bd6fd3
PH
490/*
491 * This helper is used for getting right index into array of tree roots.
492 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
493 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
494 * every node has its own stable and unstable tree.
495 */
496static inline int get_kpfn_nid(unsigned long kpfn)
497{
d8fc16a8 498 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
90bd6fd3
PH
499}
500
4035c07a
HD
501static void remove_node_from_stable_tree(struct stable_node *stable_node)
502{
503 struct rmap_item *rmap_item;
4035c07a 504
b67bfe0d 505 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
4035c07a
HD
506 if (rmap_item->hlist.next)
507 ksm_pages_sharing--;
508 else
509 ksm_pages_shared--;
9e60109f 510 put_anon_vma(rmap_item->anon_vma);
4035c07a
HD
511 rmap_item->address &= PAGE_MASK;
512 cond_resched();
513 }
514
4146d2d6
HD
515 if (stable_node->head == &migrate_nodes)
516 list_del(&stable_node->list);
517 else
518 rb_erase(&stable_node->node,
ef53d16c 519 root_stable_tree + NUMA(stable_node->nid));
4035c07a
HD
520 free_stable_node(stable_node);
521}
522
523/*
524 * get_ksm_page: checks if the page indicated by the stable node
525 * is still its ksm page, despite having held no reference to it.
526 * In which case we can trust the content of the page, and it
527 * returns the gotten page; but if the page has now been zapped,
528 * remove the stale node from the stable tree and return NULL.
c8d6553b 529 * But beware, the stable node's page might be being migrated.
4035c07a
HD
530 *
531 * You would expect the stable_node to hold a reference to the ksm page.
532 * But if it increments the page's count, swapping out has to wait for
533 * ksmd to come around again before it can free the page, which may take
534 * seconds or even minutes: much too unresponsive. So instead we use a
535 * "keyhole reference": access to the ksm page from the stable node peeps
536 * out through its keyhole to see if that page still holds the right key,
537 * pointing back to this stable node. This relies on freeing a PageAnon
538 * page to reset its page->mapping to NULL, and relies on no other use of
539 * a page to put something that might look like our key in page->mapping.
4035c07a
HD
540 * is on its way to being freed; but it is an anomaly to bear in mind.
541 */
8fdb3dbf 542static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
4035c07a
HD
543{
544 struct page *page;
545 void *expected_mapping;
c8d6553b 546 unsigned long kpfn;
4035c07a 547
bda807d4
MK
548 expected_mapping = (void *)((unsigned long)stable_node |
549 PAGE_MAPPING_KSM);
c8d6553b 550again:
4db0c3c2 551 kpfn = READ_ONCE(stable_node->kpfn);
c8d6553b
HD
552 page = pfn_to_page(kpfn);
553
554 /*
555 * page is computed from kpfn, so on most architectures reading
556 * page->mapping is naturally ordered after reading node->kpfn,
557 * but on Alpha we need to be more careful.
558 */
559 smp_read_barrier_depends();
4db0c3c2 560 if (READ_ONCE(page->mapping) != expected_mapping)
4035c07a 561 goto stale;
c8d6553b
HD
562
563 /*
564 * We cannot do anything with the page while its refcount is 0.
565 * Usually 0 means free, or tail of a higher-order page: in which
566 * case this node is no longer referenced, and should be freed;
567 * however, it might mean that the page is under page_freeze_refs().
568 * The __remove_mapping() case is easy, again the node is now stale;
569 * but if page is swapcache in migrate_page_move_mapping(), it might
570 * still be our page, in which case it's essential to keep the node.
571 */
572 while (!get_page_unless_zero(page)) {
573 /*
574 * Another check for page->mapping != expected_mapping would
575 * work here too. We have chosen the !PageSwapCache test to
576 * optimize the common case, when the page is or is about to
577 * be freed: PageSwapCache is cleared (under spin_lock_irq)
578 * in the freeze_refs section of __remove_mapping(); but Anon
579 * page->mapping reset to NULL later, in free_pages_prepare().
580 */
581 if (!PageSwapCache(page))
582 goto stale;
583 cpu_relax();
584 }
585
4db0c3c2 586 if (READ_ONCE(page->mapping) != expected_mapping) {
4035c07a
HD
587 put_page(page);
588 goto stale;
589 }
c8d6553b 590
8fdb3dbf 591 if (lock_it) {
8aafa6a4 592 lock_page(page);
4db0c3c2 593 if (READ_ONCE(page->mapping) != expected_mapping) {
8aafa6a4
HD
594 unlock_page(page);
595 put_page(page);
596 goto stale;
597 }
598 }
4035c07a 599 return page;
c8d6553b 600
4035c07a 601stale:
c8d6553b
HD
602 /*
603 * We come here from above when page->mapping or !PageSwapCache
604 * suggests that the node is stale; but it might be under migration.
605 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
606 * before checking whether node->kpfn has been changed.
607 */
608 smp_rmb();
4db0c3c2 609 if (READ_ONCE(stable_node->kpfn) != kpfn)
c8d6553b 610 goto again;
4035c07a
HD
611 remove_node_from_stable_tree(stable_node);
612 return NULL;
613}
614
31dbd01f
IE
615/*
616 * Removing rmap_item from stable or unstable tree.
617 * This function will clean the information from the stable/unstable tree.
618 */
619static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
620{
7b6ba2c7
HD
621 if (rmap_item->address & STABLE_FLAG) {
622 struct stable_node *stable_node;
5ad64688 623 struct page *page;
31dbd01f 624
7b6ba2c7 625 stable_node = rmap_item->head;
8aafa6a4 626 page = get_ksm_page(stable_node, true);
4035c07a
HD
627 if (!page)
628 goto out;
5ad64688 629
7b6ba2c7 630 hlist_del(&rmap_item->hlist);
4035c07a
HD
631 unlock_page(page);
632 put_page(page);
08beca44 633
98666f8a 634 if (!hlist_empty(&stable_node->hlist))
4035c07a
HD
635 ksm_pages_sharing--;
636 else
7b6ba2c7 637 ksm_pages_shared--;
31dbd01f 638
9e60109f 639 put_anon_vma(rmap_item->anon_vma);
93d17715 640 rmap_item->address &= PAGE_MASK;
31dbd01f 641
7b6ba2c7 642 } else if (rmap_item->address & UNSTABLE_FLAG) {
31dbd01f
IE
643 unsigned char age;
644 /*
9ba69294 645 * Usually ksmd can and must skip the rb_erase, because
31dbd01f 646 * root_unstable_tree was already reset to RB_ROOT.
9ba69294
HD
647 * But be careful when an mm is exiting: do the rb_erase
648 * if this rmap_item was inserted by this scan, rather
649 * than left over from before.
31dbd01f
IE
650 */
651 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
cd551f97 652 BUG_ON(age > 1);
31dbd01f 653 if (!age)
90bd6fd3 654 rb_erase(&rmap_item->node,
ef53d16c 655 root_unstable_tree + NUMA(rmap_item->nid));
473b0ce4 656 ksm_pages_unshared--;
93d17715 657 rmap_item->address &= PAGE_MASK;
31dbd01f 658 }
4035c07a 659out:
31dbd01f
IE
660 cond_resched(); /* we're called from many long loops */
661}
662
31dbd01f 663static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
6514d511 664 struct rmap_item **rmap_list)
31dbd01f 665{
6514d511
HD
666 while (*rmap_list) {
667 struct rmap_item *rmap_item = *rmap_list;
668 *rmap_list = rmap_item->rmap_list;
31dbd01f 669 remove_rmap_item_from_tree(rmap_item);
31dbd01f
IE
670 free_rmap_item(rmap_item);
671 }
672}
673
674/*
e850dcf5 675 * Though it's very tempting to unmerge rmap_items from stable tree rather
31dbd01f
IE
676 * than check every pte of a given vma, the locking doesn't quite work for
677 * that - an rmap_item is assigned to the stable tree after inserting ksm
678 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
679 * rmap_items from parent to child at fork time (so as not to waste time
680 * if exit comes before the next scan reaches it).
81464e30
HD
681 *
682 * Similarly, although we'd like to remove rmap_items (so updating counts
683 * and freeing memory) when unmerging an area, it's easier to leave that
684 * to the next pass of ksmd - consider, for example, how ksmd might be
685 * in cmp_and_merge_page on one of the rmap_items we would be removing.
31dbd01f 686 */
d952b791
HD
687static int unmerge_ksm_pages(struct vm_area_struct *vma,
688 unsigned long start, unsigned long end)
31dbd01f
IE
689{
690 unsigned long addr;
d952b791 691 int err = 0;
31dbd01f 692
d952b791 693 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
9ba69294
HD
694 if (ksm_test_exit(vma->vm_mm))
695 break;
d952b791
HD
696 if (signal_pending(current))
697 err = -ERESTARTSYS;
698 else
699 err = break_ksm(vma, addr);
700 }
701 return err;
31dbd01f
IE
702}
703
2ffd8679
HD
704#ifdef CONFIG_SYSFS
705/*
706 * Only called through the sysfs control interface:
707 */
cbf86cfe
HD
708static int remove_stable_node(struct stable_node *stable_node)
709{
710 struct page *page;
711 int err;
712
713 page = get_ksm_page(stable_node, true);
714 if (!page) {
715 /*
716 * get_ksm_page did remove_node_from_stable_tree itself.
717 */
718 return 0;
719 }
720
8fdb3dbf
HD
721 if (WARN_ON_ONCE(page_mapped(page))) {
722 /*
723 * This should not happen: but if it does, just refuse to let
724 * merge_across_nodes be switched - there is no need to panic.
725 */
cbf86cfe 726 err = -EBUSY;
8fdb3dbf 727 } else {
cbf86cfe 728 /*
8fdb3dbf
HD
729 * The stable node did not yet appear stale to get_ksm_page(),
730 * since that allows for an unmapped ksm page to be recognized
731 * right up until it is freed; but the node is safe to remove.
cbf86cfe
HD
732 * This page might be in a pagevec waiting to be freed,
733 * or it might be PageSwapCache (perhaps under writeback),
734 * or it might have been removed from swapcache a moment ago.
735 */
736 set_page_stable_node(page, NULL);
737 remove_node_from_stable_tree(stable_node);
738 err = 0;
739 }
740
741 unlock_page(page);
742 put_page(page);
743 return err;
744}
745
746static int remove_all_stable_nodes(void)
747{
03640418 748 struct stable_node *stable_node, *next;
cbf86cfe
HD
749 int nid;
750 int err = 0;
751
ef53d16c 752 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
cbf86cfe
HD
753 while (root_stable_tree[nid].rb_node) {
754 stable_node = rb_entry(root_stable_tree[nid].rb_node,
755 struct stable_node, node);
756 if (remove_stable_node(stable_node)) {
757 err = -EBUSY;
758 break; /* proceed to next nid */
759 }
760 cond_resched();
761 }
762 }
03640418 763 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
4146d2d6
HD
764 if (remove_stable_node(stable_node))
765 err = -EBUSY;
766 cond_resched();
767 }
cbf86cfe
HD
768 return err;
769}
770
d952b791 771static int unmerge_and_remove_all_rmap_items(void)
31dbd01f
IE
772{
773 struct mm_slot *mm_slot;
774 struct mm_struct *mm;
775 struct vm_area_struct *vma;
d952b791
HD
776 int err = 0;
777
778 spin_lock(&ksm_mmlist_lock);
9ba69294 779 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
d952b791
HD
780 struct mm_slot, mm_list);
781 spin_unlock(&ksm_mmlist_lock);
31dbd01f 782
9ba69294
HD
783 for (mm_slot = ksm_scan.mm_slot;
784 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
31dbd01f
IE
785 mm = mm_slot->mm;
786 down_read(&mm->mmap_sem);
787 for (vma = mm->mmap; vma; vma = vma->vm_next) {
9ba69294
HD
788 if (ksm_test_exit(mm))
789 break;
31dbd01f
IE
790 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
791 continue;
d952b791
HD
792 err = unmerge_ksm_pages(vma,
793 vma->vm_start, vma->vm_end);
9ba69294
HD
794 if (err)
795 goto error;
31dbd01f 796 }
9ba69294 797
6514d511 798 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
7496fea9 799 up_read(&mm->mmap_sem);
d952b791
HD
800
801 spin_lock(&ksm_mmlist_lock);
9ba69294 802 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
d952b791 803 struct mm_slot, mm_list);
9ba69294 804 if (ksm_test_exit(mm)) {
4ca3a69b 805 hash_del(&mm_slot->link);
9ba69294
HD
806 list_del(&mm_slot->mm_list);
807 spin_unlock(&ksm_mmlist_lock);
808
809 free_mm_slot(mm_slot);
810 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
9ba69294 811 mmdrop(mm);
7496fea9 812 } else
9ba69294 813 spin_unlock(&ksm_mmlist_lock);
31dbd01f
IE
814 }
815
cbf86cfe
HD
816 /* Clean up stable nodes, but don't worry if some are still busy */
817 remove_all_stable_nodes();
d952b791 818 ksm_scan.seqnr = 0;
9ba69294
HD
819 return 0;
820
821error:
822 up_read(&mm->mmap_sem);
31dbd01f 823 spin_lock(&ksm_mmlist_lock);
d952b791 824 ksm_scan.mm_slot = &ksm_mm_head;
31dbd01f 825 spin_unlock(&ksm_mmlist_lock);
d952b791 826 return err;
31dbd01f 827}
2ffd8679 828#endif /* CONFIG_SYSFS */
31dbd01f 829
31dbd01f
IE
830static u32 calc_checksum(struct page *page)
831{
832 u32 checksum;
9b04c5fe 833 void *addr = kmap_atomic(page);
31dbd01f 834 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
9b04c5fe 835 kunmap_atomic(addr);
31dbd01f
IE
836 return checksum;
837}
838
839static int memcmp_pages(struct page *page1, struct page *page2)
840{
841 char *addr1, *addr2;
842 int ret;
843
9b04c5fe
CW
844 addr1 = kmap_atomic(page1);
845 addr2 = kmap_atomic(page2);
31dbd01f 846 ret = memcmp(addr1, addr2, PAGE_SIZE);
9b04c5fe
CW
847 kunmap_atomic(addr2);
848 kunmap_atomic(addr1);
31dbd01f
IE
849 return ret;
850}
851
852static inline int pages_identical(struct page *page1, struct page *page2)
853{
854 return !memcmp_pages(page1, page2);
855}
856
857static int write_protect_page(struct vm_area_struct *vma, struct page *page,
858 pte_t *orig_pte)
859{
860 struct mm_struct *mm = vma->vm_mm;
36eaff33
KS
861 struct page_vma_mapped_walk pvmw = {
862 .page = page,
863 .vma = vma,
864 };
31dbd01f
IE
865 int swapped;
866 int err = -EFAULT;
6bdb913f
HE
867 unsigned long mmun_start; /* For mmu_notifiers */
868 unsigned long mmun_end; /* For mmu_notifiers */
31dbd01f 869
36eaff33
KS
870 pvmw.address = page_address_in_vma(page, vma);
871 if (pvmw.address == -EFAULT)
31dbd01f
IE
872 goto out;
873
29ad768c 874 BUG_ON(PageTransCompound(page));
6bdb913f 875
36eaff33
KS
876 mmun_start = pvmw.address;
877 mmun_end = pvmw.address + PAGE_SIZE;
6bdb913f
HE
878 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
879
36eaff33 880 if (!page_vma_mapped_walk(&pvmw))
6bdb913f 881 goto out_mn;
36eaff33
KS
882 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
883 goto out_unlock;
31dbd01f 884
595cd8f2
AK
885 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
886 (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte))) {
31dbd01f
IE
887 pte_t entry;
888
889 swapped = PageSwapCache(page);
36eaff33 890 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
31dbd01f 891 /*
25985edc 892 * Ok this is tricky, when get_user_pages_fast() run it doesn't
31dbd01f
IE
893 * take any lock, therefore the check that we are going to make
894 * with the pagecount against the mapcount is racey and
895 * O_DIRECT can happen right after the check.
896 * So we clear the pte and flush the tlb before the check
897 * this assure us that no O_DIRECT can happen after the check
898 * or in the middle of the check.
899 */
36eaff33 900 entry = ptep_clear_flush_notify(vma, pvmw.address, pvmw.pte);
31dbd01f
IE
901 /*
902 * Check that no O_DIRECT or similar I/O is in progress on the
903 * page
904 */
31e855ea 905 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
36eaff33 906 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
31dbd01f
IE
907 goto out_unlock;
908 }
4e31635c
HD
909 if (pte_dirty(entry))
910 set_page_dirty(page);
595cd8f2
AK
911
912 if (pte_protnone(entry))
913 entry = pte_mkclean(pte_clear_savedwrite(entry));
914 else
915 entry = pte_mkclean(pte_wrprotect(entry));
36eaff33 916 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
31dbd01f 917 }
36eaff33 918 *orig_pte = *pvmw.pte;
31dbd01f
IE
919 err = 0;
920
921out_unlock:
36eaff33 922 page_vma_mapped_walk_done(&pvmw);
6bdb913f
HE
923out_mn:
924 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
31dbd01f
IE
925out:
926 return err;
927}
928
929/**
930 * replace_page - replace page in vma by new ksm page
8dd3557a
HD
931 * @vma: vma that holds the pte pointing to page
932 * @page: the page we are replacing by kpage
933 * @kpage: the ksm page we replace page by
31dbd01f
IE
934 * @orig_pte: the original value of the pte
935 *
936 * Returns 0 on success, -EFAULT on failure.
937 */
8dd3557a
HD
938static int replace_page(struct vm_area_struct *vma, struct page *page,
939 struct page *kpage, pte_t orig_pte)
31dbd01f
IE
940{
941 struct mm_struct *mm = vma->vm_mm;
31dbd01f
IE
942 pmd_t *pmd;
943 pte_t *ptep;
e86c59b1 944 pte_t newpte;
31dbd01f
IE
945 spinlock_t *ptl;
946 unsigned long addr;
31dbd01f 947 int err = -EFAULT;
6bdb913f
HE
948 unsigned long mmun_start; /* For mmu_notifiers */
949 unsigned long mmun_end; /* For mmu_notifiers */
31dbd01f 950
8dd3557a 951 addr = page_address_in_vma(page, vma);
31dbd01f
IE
952 if (addr == -EFAULT)
953 goto out;
954
6219049a
BL
955 pmd = mm_find_pmd(mm, addr);
956 if (!pmd)
31dbd01f 957 goto out;
31dbd01f 958
6bdb913f
HE
959 mmun_start = addr;
960 mmun_end = addr + PAGE_SIZE;
961 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
962
31dbd01f
IE
963 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
964 if (!pte_same(*ptep, orig_pte)) {
965 pte_unmap_unlock(ptep, ptl);
6bdb913f 966 goto out_mn;
31dbd01f
IE
967 }
968
e86c59b1
CI
969 /*
970 * No need to check ksm_use_zero_pages here: we can only have a
971 * zero_page here if ksm_use_zero_pages was enabled alreaady.
972 */
973 if (!is_zero_pfn(page_to_pfn(kpage))) {
974 get_page(kpage);
975 page_add_anon_rmap(kpage, vma, addr, false);
976 newpte = mk_pte(kpage, vma->vm_page_prot);
977 } else {
978 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
979 vma->vm_page_prot));
980 }
31dbd01f
IE
981
982 flush_cache_page(vma, addr, pte_pfn(*ptep));
34ee645e 983 ptep_clear_flush_notify(vma, addr, ptep);
e86c59b1 984 set_pte_at_notify(mm, addr, ptep, newpte);
31dbd01f 985
d281ee61 986 page_remove_rmap(page, false);
ae52a2ad
HD
987 if (!page_mapped(page))
988 try_to_free_swap(page);
8dd3557a 989 put_page(page);
31dbd01f
IE
990
991 pte_unmap_unlock(ptep, ptl);
992 err = 0;
6bdb913f
HE
993out_mn:
994 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
31dbd01f
IE
995out:
996 return err;
997}
998
999/*
1000 * try_to_merge_one_page - take two pages and merge them into one
8dd3557a
HD
1001 * @vma: the vma that holds the pte pointing to page
1002 * @page: the PageAnon page that we want to replace with kpage
80e14822
HD
1003 * @kpage: the PageKsm page that we want to map instead of page,
1004 * or NULL the first time when we want to use page as kpage.
31dbd01f
IE
1005 *
1006 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1007 */
1008static int try_to_merge_one_page(struct vm_area_struct *vma,
8dd3557a 1009 struct page *page, struct page *kpage)
31dbd01f
IE
1010{
1011 pte_t orig_pte = __pte(0);
1012 int err = -EFAULT;
1013
db114b83
HD
1014 if (page == kpage) /* ksm page forked */
1015 return 0;
1016
8dd3557a 1017 if (!PageAnon(page))
31dbd01f
IE
1018 goto out;
1019
31dbd01f
IE
1020 /*
1021 * We need the page lock to read a stable PageSwapCache in
1022 * write_protect_page(). We use trylock_page() instead of
1023 * lock_page() because we don't want to wait here - we
1024 * prefer to continue scanning and merging different pages,
1025 * then come back to this page when it is unlocked.
1026 */
8dd3557a 1027 if (!trylock_page(page))
31e855ea 1028 goto out;
f765f540
KS
1029
1030 if (PageTransCompound(page)) {
a7306c34 1031 if (split_huge_page(page))
f765f540
KS
1032 goto out_unlock;
1033 }
1034
31dbd01f
IE
1035 /*
1036 * If this anonymous page is mapped only here, its pte may need
1037 * to be write-protected. If it's mapped elsewhere, all of its
1038 * ptes are necessarily already write-protected. But in either
1039 * case, we need to lock and check page_count is not raised.
1040 */
80e14822
HD
1041 if (write_protect_page(vma, page, &orig_pte) == 0) {
1042 if (!kpage) {
1043 /*
1044 * While we hold page lock, upgrade page from
1045 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1046 * stable_tree_insert() will update stable_node.
1047 */
1048 set_page_stable_node(page, NULL);
1049 mark_page_accessed(page);
337ed7eb
MK
1050 /*
1051 * Page reclaim just frees a clean page with no dirty
1052 * ptes: make sure that the ksm page would be swapped.
1053 */
1054 if (!PageDirty(page))
1055 SetPageDirty(page);
80e14822
HD
1056 err = 0;
1057 } else if (pages_identical(page, kpage))
1058 err = replace_page(vma, page, kpage, orig_pte);
1059 }
31dbd01f 1060
80e14822 1061 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
73848b46 1062 munlock_vma_page(page);
5ad64688
HD
1063 if (!PageMlocked(kpage)) {
1064 unlock_page(page);
5ad64688
HD
1065 lock_page(kpage);
1066 mlock_vma_page(kpage);
1067 page = kpage; /* for final unlock */
1068 }
1069 }
73848b46 1070
f765f540 1071out_unlock:
8dd3557a 1072 unlock_page(page);
31dbd01f
IE
1073out:
1074 return err;
1075}
1076
81464e30
HD
1077/*
1078 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1079 * but no new kernel page is allocated: kpage must already be a ksm page.
8dd3557a
HD
1080 *
1081 * This function returns 0 if the pages were merged, -EFAULT otherwise.
81464e30 1082 */
8dd3557a
HD
1083static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1084 struct page *page, struct page *kpage)
81464e30 1085{
8dd3557a 1086 struct mm_struct *mm = rmap_item->mm;
81464e30
HD
1087 struct vm_area_struct *vma;
1088 int err = -EFAULT;
1089
8dd3557a 1090 down_read(&mm->mmap_sem);
85c6e8dd
AA
1091 vma = find_mergeable_vma(mm, rmap_item->address);
1092 if (!vma)
81464e30
HD
1093 goto out;
1094
8dd3557a 1095 err = try_to_merge_one_page(vma, page, kpage);
db114b83
HD
1096 if (err)
1097 goto out;
1098
bc56620b
HD
1099 /* Unstable nid is in union with stable anon_vma: remove first */
1100 remove_rmap_item_from_tree(rmap_item);
1101
db114b83 1102 /* Must get reference to anon_vma while still holding mmap_sem */
9e60109f
PZ
1103 rmap_item->anon_vma = vma->anon_vma;
1104 get_anon_vma(vma->anon_vma);
81464e30 1105out:
8dd3557a 1106 up_read(&mm->mmap_sem);
81464e30
HD
1107 return err;
1108}
1109
31dbd01f
IE
1110/*
1111 * try_to_merge_two_pages - take two identical pages and prepare them
1112 * to be merged into one page.
1113 *
8dd3557a
HD
1114 * This function returns the kpage if we successfully merged two identical
1115 * pages into one ksm page, NULL otherwise.
31dbd01f 1116 *
80e14822 1117 * Note that this function upgrades page to ksm page: if one of the pages
31dbd01f
IE
1118 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1119 */
8dd3557a
HD
1120static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1121 struct page *page,
1122 struct rmap_item *tree_rmap_item,
1123 struct page *tree_page)
31dbd01f 1124{
80e14822 1125 int err;
31dbd01f 1126
80e14822 1127 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
31dbd01f 1128 if (!err) {
8dd3557a 1129 err = try_to_merge_with_ksm_page(tree_rmap_item,
80e14822 1130 tree_page, page);
31dbd01f 1131 /*
81464e30
HD
1132 * If that fails, we have a ksm page with only one pte
1133 * pointing to it: so break it.
31dbd01f 1134 */
4035c07a 1135 if (err)
8dd3557a 1136 break_cow(rmap_item);
31dbd01f 1137 }
80e14822 1138 return err ? NULL : page;
31dbd01f
IE
1139}
1140
31dbd01f 1141/*
8dd3557a 1142 * stable_tree_search - search for page inside the stable tree
31dbd01f
IE
1143 *
1144 * This function checks if there is a page inside the stable tree
1145 * with identical content to the page that we are scanning right now.
1146 *
7b6ba2c7 1147 * This function returns the stable tree node of identical content if found,
31dbd01f
IE
1148 * NULL otherwise.
1149 */
62b61f61 1150static struct page *stable_tree_search(struct page *page)
31dbd01f 1151{
90bd6fd3 1152 int nid;
ef53d16c 1153 struct rb_root *root;
4146d2d6
HD
1154 struct rb_node **new;
1155 struct rb_node *parent;
1156 struct stable_node *stable_node;
1157 struct stable_node *page_node;
31dbd01f 1158
4146d2d6
HD
1159 page_node = page_stable_node(page);
1160 if (page_node && page_node->head != &migrate_nodes) {
1161 /* ksm page forked */
08beca44 1162 get_page(page);
62b61f61 1163 return page;
08beca44
HD
1164 }
1165
90bd6fd3 1166 nid = get_kpfn_nid(page_to_pfn(page));
ef53d16c 1167 root = root_stable_tree + nid;
4146d2d6 1168again:
ef53d16c 1169 new = &root->rb_node;
4146d2d6 1170 parent = NULL;
90bd6fd3 1171
4146d2d6 1172 while (*new) {
4035c07a 1173 struct page *tree_page;
31dbd01f
IE
1174 int ret;
1175
08beca44 1176 cond_resched();
4146d2d6 1177 stable_node = rb_entry(*new, struct stable_node, node);
8aafa6a4 1178 tree_page = get_ksm_page(stable_node, false);
f2e5ff85
AA
1179 if (!tree_page) {
1180 /*
1181 * If we walked over a stale stable_node,
1182 * get_ksm_page() will call rb_erase() and it
1183 * may rebalance the tree from under us. So
1184 * restart the search from scratch. Returning
1185 * NULL would be safe too, but we'd generate
1186 * false negative insertions just because some
1187 * stable_node was stale.
1188 */
1189 goto again;
1190 }
31dbd01f 1191
4035c07a 1192 ret = memcmp_pages(page, tree_page);
c8d6553b 1193 put_page(tree_page);
31dbd01f 1194
4146d2d6 1195 parent = *new;
c8d6553b 1196 if (ret < 0)
4146d2d6 1197 new = &parent->rb_left;
c8d6553b 1198 else if (ret > 0)
4146d2d6 1199 new = &parent->rb_right;
c8d6553b
HD
1200 else {
1201 /*
1202 * Lock and unlock the stable_node's page (which
1203 * might already have been migrated) so that page
1204 * migration is sure to notice its raised count.
1205 * It would be more elegant to return stable_node
1206 * than kpage, but that involves more changes.
1207 */
1208 tree_page = get_ksm_page(stable_node, true);
4146d2d6 1209 if (tree_page) {
c8d6553b 1210 unlock_page(tree_page);
4146d2d6
HD
1211 if (get_kpfn_nid(stable_node->kpfn) !=
1212 NUMA(stable_node->nid)) {
1213 put_page(tree_page);
1214 goto replace;
1215 }
1216 return tree_page;
1217 }
1218 /*
1219 * There is now a place for page_node, but the tree may
1220 * have been rebalanced, so re-evaluate parent and new.
1221 */
1222 if (page_node)
1223 goto again;
1224 return NULL;
c8d6553b 1225 }
31dbd01f
IE
1226 }
1227
4146d2d6
HD
1228 if (!page_node)
1229 return NULL;
1230
1231 list_del(&page_node->list);
1232 DO_NUMA(page_node->nid = nid);
1233 rb_link_node(&page_node->node, parent, new);
ef53d16c 1234 rb_insert_color(&page_node->node, root);
4146d2d6
HD
1235 get_page(page);
1236 return page;
1237
1238replace:
1239 if (page_node) {
1240 list_del(&page_node->list);
1241 DO_NUMA(page_node->nid = nid);
ef53d16c 1242 rb_replace_node(&stable_node->node, &page_node->node, root);
4146d2d6
HD
1243 get_page(page);
1244 } else {
ef53d16c 1245 rb_erase(&stable_node->node, root);
4146d2d6
HD
1246 page = NULL;
1247 }
1248 stable_node->head = &migrate_nodes;
1249 list_add(&stable_node->list, stable_node->head);
1250 return page;
31dbd01f
IE
1251}
1252
1253/*
e850dcf5 1254 * stable_tree_insert - insert stable tree node pointing to new ksm page
31dbd01f
IE
1255 * into the stable tree.
1256 *
7b6ba2c7
HD
1257 * This function returns the stable tree node just allocated on success,
1258 * NULL otherwise.
31dbd01f 1259 */
7b6ba2c7 1260static struct stable_node *stable_tree_insert(struct page *kpage)
31dbd01f 1261{
90bd6fd3
PH
1262 int nid;
1263 unsigned long kpfn;
ef53d16c 1264 struct rb_root *root;
90bd6fd3 1265 struct rb_node **new;
f2e5ff85 1266 struct rb_node *parent;
7b6ba2c7 1267 struct stable_node *stable_node;
31dbd01f 1268
90bd6fd3
PH
1269 kpfn = page_to_pfn(kpage);
1270 nid = get_kpfn_nid(kpfn);
ef53d16c 1271 root = root_stable_tree + nid;
f2e5ff85
AA
1272again:
1273 parent = NULL;
ef53d16c 1274 new = &root->rb_node;
90bd6fd3 1275
31dbd01f 1276 while (*new) {
4035c07a 1277 struct page *tree_page;
31dbd01f
IE
1278 int ret;
1279
08beca44 1280 cond_resched();
7b6ba2c7 1281 stable_node = rb_entry(*new, struct stable_node, node);
8aafa6a4 1282 tree_page = get_ksm_page(stable_node, false);
f2e5ff85
AA
1283 if (!tree_page) {
1284 /*
1285 * If we walked over a stale stable_node,
1286 * get_ksm_page() will call rb_erase() and it
1287 * may rebalance the tree from under us. So
1288 * restart the search from scratch. Returning
1289 * NULL would be safe too, but we'd generate
1290 * false negative insertions just because some
1291 * stable_node was stale.
1292 */
1293 goto again;
1294 }
31dbd01f 1295
4035c07a
HD
1296 ret = memcmp_pages(kpage, tree_page);
1297 put_page(tree_page);
31dbd01f
IE
1298
1299 parent = *new;
1300 if (ret < 0)
1301 new = &parent->rb_left;
1302 else if (ret > 0)
1303 new = &parent->rb_right;
1304 else {
1305 /*
1306 * It is not a bug that stable_tree_search() didn't
1307 * find this node: because at that time our page was
1308 * not yet write-protected, so may have changed since.
1309 */
1310 return NULL;
1311 }
1312 }
1313
7b6ba2c7
HD
1314 stable_node = alloc_stable_node();
1315 if (!stable_node)
1316 return NULL;
31dbd01f 1317
7b6ba2c7 1318 INIT_HLIST_HEAD(&stable_node->hlist);
90bd6fd3 1319 stable_node->kpfn = kpfn;
08beca44 1320 set_page_stable_node(kpage, stable_node);
4146d2d6 1321 DO_NUMA(stable_node->nid = nid);
e850dcf5 1322 rb_link_node(&stable_node->node, parent, new);
ef53d16c 1323 rb_insert_color(&stable_node->node, root);
08beca44 1324
7b6ba2c7 1325 return stable_node;
31dbd01f
IE
1326}
1327
1328/*
8dd3557a
HD
1329 * unstable_tree_search_insert - search for identical page,
1330 * else insert rmap_item into the unstable tree.
31dbd01f
IE
1331 *
1332 * This function searches for a page in the unstable tree identical to the
1333 * page currently being scanned; and if no identical page is found in the
1334 * tree, we insert rmap_item as a new object into the unstable tree.
1335 *
1336 * This function returns pointer to rmap_item found to be identical
1337 * to the currently scanned page, NULL otherwise.
1338 *
1339 * This function does both searching and inserting, because they share
1340 * the same walking algorithm in an rbtree.
1341 */
8dd3557a
HD
1342static
1343struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1344 struct page *page,
1345 struct page **tree_pagep)
31dbd01f 1346{
90bd6fd3
PH
1347 struct rb_node **new;
1348 struct rb_root *root;
31dbd01f 1349 struct rb_node *parent = NULL;
90bd6fd3
PH
1350 int nid;
1351
1352 nid = get_kpfn_nid(page_to_pfn(page));
ef53d16c 1353 root = root_unstable_tree + nid;
90bd6fd3 1354 new = &root->rb_node;
31dbd01f
IE
1355
1356 while (*new) {
1357 struct rmap_item *tree_rmap_item;
8dd3557a 1358 struct page *tree_page;
31dbd01f
IE
1359 int ret;
1360
d178f27f 1361 cond_resched();
31dbd01f 1362 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
8dd3557a 1363 tree_page = get_mergeable_page(tree_rmap_item);
c8f95ed1 1364 if (!tree_page)
31dbd01f
IE
1365 return NULL;
1366
1367 /*
8dd3557a 1368 * Don't substitute a ksm page for a forked page.
31dbd01f 1369 */
8dd3557a
HD
1370 if (page == tree_page) {
1371 put_page(tree_page);
31dbd01f
IE
1372 return NULL;
1373 }
1374
8dd3557a 1375 ret = memcmp_pages(page, tree_page);
31dbd01f
IE
1376
1377 parent = *new;
1378 if (ret < 0) {
8dd3557a 1379 put_page(tree_page);
31dbd01f
IE
1380 new = &parent->rb_left;
1381 } else if (ret > 0) {
8dd3557a 1382 put_page(tree_page);
31dbd01f 1383 new = &parent->rb_right;
b599cbdf
HD
1384 } else if (!ksm_merge_across_nodes &&
1385 page_to_nid(tree_page) != nid) {
1386 /*
1387 * If tree_page has been migrated to another NUMA node,
1388 * it will be flushed out and put in the right unstable
1389 * tree next time: only merge with it when across_nodes.
1390 */
1391 put_page(tree_page);
1392 return NULL;
31dbd01f 1393 } else {
8dd3557a 1394 *tree_pagep = tree_page;
31dbd01f
IE
1395 return tree_rmap_item;
1396 }
1397 }
1398
7b6ba2c7 1399 rmap_item->address |= UNSTABLE_FLAG;
31dbd01f 1400 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
e850dcf5 1401 DO_NUMA(rmap_item->nid = nid);
31dbd01f 1402 rb_link_node(&rmap_item->node, parent, new);
90bd6fd3 1403 rb_insert_color(&rmap_item->node, root);
31dbd01f 1404
473b0ce4 1405 ksm_pages_unshared++;
31dbd01f
IE
1406 return NULL;
1407}
1408
1409/*
1410 * stable_tree_append - add another rmap_item to the linked list of
1411 * rmap_items hanging off a given node of the stable tree, all sharing
1412 * the same ksm page.
1413 */
1414static void stable_tree_append(struct rmap_item *rmap_item,
7b6ba2c7 1415 struct stable_node *stable_node)
31dbd01f 1416{
7b6ba2c7 1417 rmap_item->head = stable_node;
31dbd01f 1418 rmap_item->address |= STABLE_FLAG;
7b6ba2c7 1419 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
e178dfde 1420
7b6ba2c7
HD
1421 if (rmap_item->hlist.next)
1422 ksm_pages_sharing++;
1423 else
1424 ksm_pages_shared++;
31dbd01f
IE
1425}
1426
1427/*
81464e30
HD
1428 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1429 * if not, compare checksum to previous and if it's the same, see if page can
1430 * be inserted into the unstable tree, or merged with a page already there and
1431 * both transferred to the stable tree.
31dbd01f
IE
1432 *
1433 * @page: the page that we are searching identical page to.
1434 * @rmap_item: the reverse mapping into the virtual address of this page
1435 */
1436static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1437{
31dbd01f 1438 struct rmap_item *tree_rmap_item;
8dd3557a 1439 struct page *tree_page = NULL;
7b6ba2c7 1440 struct stable_node *stable_node;
8dd3557a 1441 struct page *kpage;
31dbd01f
IE
1442 unsigned int checksum;
1443 int err;
1444
4146d2d6
HD
1445 stable_node = page_stable_node(page);
1446 if (stable_node) {
1447 if (stable_node->head != &migrate_nodes &&
1448 get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1449 rb_erase(&stable_node->node,
ef53d16c 1450 root_stable_tree + NUMA(stable_node->nid));
4146d2d6
HD
1451 stable_node->head = &migrate_nodes;
1452 list_add(&stable_node->list, stable_node->head);
1453 }
1454 if (stable_node->head != &migrate_nodes &&
1455 rmap_item->head == stable_node)
1456 return;
1457 }
31dbd01f
IE
1458
1459 /* We first start with searching the page inside the stable tree */
62b61f61 1460 kpage = stable_tree_search(page);
4146d2d6
HD
1461 if (kpage == page && rmap_item->head == stable_node) {
1462 put_page(kpage);
1463 return;
1464 }
1465
1466 remove_rmap_item_from_tree(rmap_item);
1467
62b61f61 1468 if (kpage) {
08beca44 1469 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
31dbd01f
IE
1470 if (!err) {
1471 /*
1472 * The page was successfully merged:
1473 * add its rmap_item to the stable tree.
1474 */
5ad64688 1475 lock_page(kpage);
62b61f61 1476 stable_tree_append(rmap_item, page_stable_node(kpage));
5ad64688 1477 unlock_page(kpage);
31dbd01f 1478 }
8dd3557a 1479 put_page(kpage);
31dbd01f
IE
1480 return;
1481 }
1482
1483 /*
4035c07a
HD
1484 * If the hash value of the page has changed from the last time
1485 * we calculated it, this page is changing frequently: therefore we
1486 * don't want to insert it in the unstable tree, and we don't want
1487 * to waste our time searching for something identical to it there.
31dbd01f
IE
1488 */
1489 checksum = calc_checksum(page);
1490 if (rmap_item->oldchecksum != checksum) {
1491 rmap_item->oldchecksum = checksum;
1492 return;
1493 }
1494
e86c59b1
CI
1495 /*
1496 * Same checksum as an empty page. We attempt to merge it with the
1497 * appropriate zero page if the user enabled this via sysfs.
1498 */
1499 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
1500 struct vm_area_struct *vma;
1501
1502 vma = find_mergeable_vma(rmap_item->mm, rmap_item->address);
1503 err = try_to_merge_one_page(vma, page,
1504 ZERO_PAGE(rmap_item->address));
1505 /*
1506 * In case of failure, the page was not really empty, so we
1507 * need to continue. Otherwise we're done.
1508 */
1509 if (!err)
1510 return;
1511 }
8dd3557a
HD
1512 tree_rmap_item =
1513 unstable_tree_search_insert(rmap_item, page, &tree_page);
31dbd01f 1514 if (tree_rmap_item) {
8dd3557a
HD
1515 kpage = try_to_merge_two_pages(rmap_item, page,
1516 tree_rmap_item, tree_page);
1517 put_page(tree_page);
8dd3557a 1518 if (kpage) {
bc56620b
HD
1519 /*
1520 * The pages were successfully merged: insert new
1521 * node in the stable tree and add both rmap_items.
1522 */
5ad64688 1523 lock_page(kpage);
7b6ba2c7
HD
1524 stable_node = stable_tree_insert(kpage);
1525 if (stable_node) {
1526 stable_tree_append(tree_rmap_item, stable_node);
1527 stable_tree_append(rmap_item, stable_node);
1528 }
5ad64688 1529 unlock_page(kpage);
7b6ba2c7 1530
31dbd01f
IE
1531 /*
1532 * If we fail to insert the page into the stable tree,
1533 * we will have 2 virtual addresses that are pointing
1534 * to a ksm page left outside the stable tree,
1535 * in which case we need to break_cow on both.
1536 */
7b6ba2c7 1537 if (!stable_node) {
8dd3557a
HD
1538 break_cow(tree_rmap_item);
1539 break_cow(rmap_item);
31dbd01f
IE
1540 }
1541 }
31dbd01f
IE
1542 }
1543}
1544
1545static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
6514d511 1546 struct rmap_item **rmap_list,
31dbd01f
IE
1547 unsigned long addr)
1548{
1549 struct rmap_item *rmap_item;
1550
6514d511
HD
1551 while (*rmap_list) {
1552 rmap_item = *rmap_list;
93d17715 1553 if ((rmap_item->address & PAGE_MASK) == addr)
31dbd01f 1554 return rmap_item;
31dbd01f
IE
1555 if (rmap_item->address > addr)
1556 break;
6514d511 1557 *rmap_list = rmap_item->rmap_list;
31dbd01f 1558 remove_rmap_item_from_tree(rmap_item);
31dbd01f
IE
1559 free_rmap_item(rmap_item);
1560 }
1561
1562 rmap_item = alloc_rmap_item();
1563 if (rmap_item) {
1564 /* It has already been zeroed */
1565 rmap_item->mm = mm_slot->mm;
1566 rmap_item->address = addr;
6514d511
HD
1567 rmap_item->rmap_list = *rmap_list;
1568 *rmap_list = rmap_item;
31dbd01f
IE
1569 }
1570 return rmap_item;
1571}
1572
1573static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1574{
1575 struct mm_struct *mm;
1576 struct mm_slot *slot;
1577 struct vm_area_struct *vma;
1578 struct rmap_item *rmap_item;
90bd6fd3 1579 int nid;
31dbd01f
IE
1580
1581 if (list_empty(&ksm_mm_head.mm_list))
1582 return NULL;
1583
1584 slot = ksm_scan.mm_slot;
1585 if (slot == &ksm_mm_head) {
2919bfd0
HD
1586 /*
1587 * A number of pages can hang around indefinitely on per-cpu
1588 * pagevecs, raised page count preventing write_protect_page
1589 * from merging them. Though it doesn't really matter much,
1590 * it is puzzling to see some stuck in pages_volatile until
1591 * other activity jostles them out, and they also prevented
1592 * LTP's KSM test from succeeding deterministically; so drain
1593 * them here (here rather than on entry to ksm_do_scan(),
1594 * so we don't IPI too often when pages_to_scan is set low).
1595 */
1596 lru_add_drain_all();
1597
4146d2d6
HD
1598 /*
1599 * Whereas stale stable_nodes on the stable_tree itself
1600 * get pruned in the regular course of stable_tree_search(),
1601 * those moved out to the migrate_nodes list can accumulate:
1602 * so prune them once before each full scan.
1603 */
1604 if (!ksm_merge_across_nodes) {
03640418 1605 struct stable_node *stable_node, *next;
4146d2d6
HD
1606 struct page *page;
1607
03640418
GT
1608 list_for_each_entry_safe(stable_node, next,
1609 &migrate_nodes, list) {
4146d2d6
HD
1610 page = get_ksm_page(stable_node, false);
1611 if (page)
1612 put_page(page);
1613 cond_resched();
1614 }
1615 }
1616
ef53d16c 1617 for (nid = 0; nid < ksm_nr_node_ids; nid++)
90bd6fd3 1618 root_unstable_tree[nid] = RB_ROOT;
31dbd01f
IE
1619
1620 spin_lock(&ksm_mmlist_lock);
1621 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1622 ksm_scan.mm_slot = slot;
1623 spin_unlock(&ksm_mmlist_lock);
2b472611
HD
1624 /*
1625 * Although we tested list_empty() above, a racing __ksm_exit
1626 * of the last mm on the list may have removed it since then.
1627 */
1628 if (slot == &ksm_mm_head)
1629 return NULL;
31dbd01f
IE
1630next_mm:
1631 ksm_scan.address = 0;
6514d511 1632 ksm_scan.rmap_list = &slot->rmap_list;
31dbd01f
IE
1633 }
1634
1635 mm = slot->mm;
1636 down_read(&mm->mmap_sem);
9ba69294
HD
1637 if (ksm_test_exit(mm))
1638 vma = NULL;
1639 else
1640 vma = find_vma(mm, ksm_scan.address);
1641
1642 for (; vma; vma = vma->vm_next) {
31dbd01f
IE
1643 if (!(vma->vm_flags & VM_MERGEABLE))
1644 continue;
1645 if (ksm_scan.address < vma->vm_start)
1646 ksm_scan.address = vma->vm_start;
1647 if (!vma->anon_vma)
1648 ksm_scan.address = vma->vm_end;
1649
1650 while (ksm_scan.address < vma->vm_end) {
9ba69294
HD
1651 if (ksm_test_exit(mm))
1652 break;
31dbd01f 1653 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
21ae5b01
AA
1654 if (IS_ERR_OR_NULL(*page)) {
1655 ksm_scan.address += PAGE_SIZE;
1656 cond_resched();
1657 continue;
1658 }
f765f540 1659 if (PageAnon(*page)) {
31dbd01f
IE
1660 flush_anon_page(vma, *page, ksm_scan.address);
1661 flush_dcache_page(*page);
1662 rmap_item = get_next_rmap_item(slot,
6514d511 1663 ksm_scan.rmap_list, ksm_scan.address);
31dbd01f 1664 if (rmap_item) {
6514d511
HD
1665 ksm_scan.rmap_list =
1666 &rmap_item->rmap_list;
31dbd01f
IE
1667 ksm_scan.address += PAGE_SIZE;
1668 } else
1669 put_page(*page);
1670 up_read(&mm->mmap_sem);
1671 return rmap_item;
1672 }
21ae5b01 1673 put_page(*page);
31dbd01f
IE
1674 ksm_scan.address += PAGE_SIZE;
1675 cond_resched();
1676 }
1677 }
1678
9ba69294
HD
1679 if (ksm_test_exit(mm)) {
1680 ksm_scan.address = 0;
6514d511 1681 ksm_scan.rmap_list = &slot->rmap_list;
9ba69294 1682 }
31dbd01f
IE
1683 /*
1684 * Nuke all the rmap_items that are above this current rmap:
1685 * because there were no VM_MERGEABLE vmas with such addresses.
1686 */
6514d511 1687 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
31dbd01f
IE
1688
1689 spin_lock(&ksm_mmlist_lock);
cd551f97
HD
1690 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1691 struct mm_slot, mm_list);
1692 if (ksm_scan.address == 0) {
1693 /*
1694 * We've completed a full scan of all vmas, holding mmap_sem
1695 * throughout, and found no VM_MERGEABLE: so do the same as
1696 * __ksm_exit does to remove this mm from all our lists now.
9ba69294
HD
1697 * This applies either when cleaning up after __ksm_exit
1698 * (but beware: we can reach here even before __ksm_exit),
1699 * or when all VM_MERGEABLE areas have been unmapped (and
1700 * mmap_sem then protects against race with MADV_MERGEABLE).
cd551f97 1701 */
4ca3a69b 1702 hash_del(&slot->link);
cd551f97 1703 list_del(&slot->mm_list);
9ba69294
HD
1704 spin_unlock(&ksm_mmlist_lock);
1705
cd551f97
HD
1706 free_mm_slot(slot);
1707 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
9ba69294
HD
1708 up_read(&mm->mmap_sem);
1709 mmdrop(mm);
1710 } else {
9ba69294 1711 up_read(&mm->mmap_sem);
7496fea9
ZC
1712 /*
1713 * up_read(&mm->mmap_sem) first because after
1714 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
1715 * already have been freed under us by __ksm_exit()
1716 * because the "mm_slot" is still hashed and
1717 * ksm_scan.mm_slot doesn't point to it anymore.
1718 */
1719 spin_unlock(&ksm_mmlist_lock);
cd551f97 1720 }
31dbd01f
IE
1721
1722 /* Repeat until we've completed scanning the whole list */
cd551f97 1723 slot = ksm_scan.mm_slot;
31dbd01f
IE
1724 if (slot != &ksm_mm_head)
1725 goto next_mm;
1726
31dbd01f
IE
1727 ksm_scan.seqnr++;
1728 return NULL;
1729}
1730
1731/**
1732 * ksm_do_scan - the ksm scanner main worker function.
1733 * @scan_npages - number of pages we want to scan before we return.
1734 */
1735static void ksm_do_scan(unsigned int scan_npages)
1736{
1737 struct rmap_item *rmap_item;
22eccdd7 1738 struct page *uninitialized_var(page);
31dbd01f 1739
878aee7d 1740 while (scan_npages-- && likely(!freezing(current))) {
31dbd01f
IE
1741 cond_resched();
1742 rmap_item = scan_get_next_rmap_item(&page);
1743 if (!rmap_item)
1744 return;
4146d2d6 1745 cmp_and_merge_page(page, rmap_item);
31dbd01f
IE
1746 put_page(page);
1747 }
1748}
1749
6e158384
HD
1750static int ksmd_should_run(void)
1751{
1752 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1753}
1754
31dbd01f
IE
1755static int ksm_scan_thread(void *nothing)
1756{
878aee7d 1757 set_freezable();
339aa624 1758 set_user_nice(current, 5);
31dbd01f
IE
1759
1760 while (!kthread_should_stop()) {
6e158384 1761 mutex_lock(&ksm_thread_mutex);
ef4d43a8 1762 wait_while_offlining();
6e158384 1763 if (ksmd_should_run())
31dbd01f 1764 ksm_do_scan(ksm_thread_pages_to_scan);
6e158384
HD
1765 mutex_unlock(&ksm_thread_mutex);
1766
878aee7d
AA
1767 try_to_freeze();
1768
6e158384 1769 if (ksmd_should_run()) {
31dbd01f
IE
1770 schedule_timeout_interruptible(
1771 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1772 } else {
878aee7d 1773 wait_event_freezable(ksm_thread_wait,
6e158384 1774 ksmd_should_run() || kthread_should_stop());
31dbd01f
IE
1775 }
1776 }
1777 return 0;
1778}
1779
f8af4da3
HD
1780int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1781 unsigned long end, int advice, unsigned long *vm_flags)
1782{
1783 struct mm_struct *mm = vma->vm_mm;
d952b791 1784 int err;
f8af4da3
HD
1785
1786 switch (advice) {
1787 case MADV_MERGEABLE:
1788 /*
1789 * Be somewhat over-protective for now!
1790 */
1791 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1792 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
0661a336 1793 VM_HUGETLB | VM_MIXEDMAP))
f8af4da3
HD
1794 return 0; /* just ignore the advice */
1795
cc2383ec
KK
1796#ifdef VM_SAO
1797 if (*vm_flags & VM_SAO)
1798 return 0;
1799#endif
1800
d952b791
HD
1801 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1802 err = __ksm_enter(mm);
1803 if (err)
1804 return err;
1805 }
f8af4da3
HD
1806
1807 *vm_flags |= VM_MERGEABLE;
1808 break;
1809
1810 case MADV_UNMERGEABLE:
1811 if (!(*vm_flags & VM_MERGEABLE))
1812 return 0; /* just ignore the advice */
1813
d952b791
HD
1814 if (vma->anon_vma) {
1815 err = unmerge_ksm_pages(vma, start, end);
1816 if (err)
1817 return err;
1818 }
f8af4da3
HD
1819
1820 *vm_flags &= ~VM_MERGEABLE;
1821 break;
1822 }
1823
1824 return 0;
1825}
1826
1827int __ksm_enter(struct mm_struct *mm)
1828{
6e158384
HD
1829 struct mm_slot *mm_slot;
1830 int needs_wakeup;
1831
1832 mm_slot = alloc_mm_slot();
31dbd01f
IE
1833 if (!mm_slot)
1834 return -ENOMEM;
1835
6e158384
HD
1836 /* Check ksm_run too? Would need tighter locking */
1837 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1838
31dbd01f
IE
1839 spin_lock(&ksm_mmlist_lock);
1840 insert_to_mm_slots_hash(mm, mm_slot);
1841 /*
cbf86cfe
HD
1842 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1843 * insert just behind the scanning cursor, to let the area settle
31dbd01f
IE
1844 * down a little; when fork is followed by immediate exec, we don't
1845 * want ksmd to waste time setting up and tearing down an rmap_list.
cbf86cfe
HD
1846 *
1847 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1848 * scanning cursor, otherwise KSM pages in newly forked mms will be
1849 * missed: then we might as well insert at the end of the list.
31dbd01f 1850 */
cbf86cfe
HD
1851 if (ksm_run & KSM_RUN_UNMERGE)
1852 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1853 else
1854 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
31dbd01f
IE
1855 spin_unlock(&ksm_mmlist_lock);
1856
f8af4da3 1857 set_bit(MMF_VM_MERGEABLE, &mm->flags);
f1f10076 1858 mmgrab(mm);
6e158384
HD
1859
1860 if (needs_wakeup)
1861 wake_up_interruptible(&ksm_thread_wait);
1862
f8af4da3
HD
1863 return 0;
1864}
1865
1c2fb7a4 1866void __ksm_exit(struct mm_struct *mm)
f8af4da3 1867{
cd551f97 1868 struct mm_slot *mm_slot;
9ba69294 1869 int easy_to_free = 0;
cd551f97 1870
31dbd01f 1871 /*
9ba69294
HD
1872 * This process is exiting: if it's straightforward (as is the
1873 * case when ksmd was never running), free mm_slot immediately.
1874 * But if it's at the cursor or has rmap_items linked to it, use
1875 * mmap_sem to synchronize with any break_cows before pagetables
1876 * are freed, and leave the mm_slot on the list for ksmd to free.
1877 * Beware: ksm may already have noticed it exiting and freed the slot.
31dbd01f 1878 */
9ba69294 1879
cd551f97
HD
1880 spin_lock(&ksm_mmlist_lock);
1881 mm_slot = get_mm_slot(mm);
9ba69294 1882 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
6514d511 1883 if (!mm_slot->rmap_list) {
4ca3a69b 1884 hash_del(&mm_slot->link);
9ba69294
HD
1885 list_del(&mm_slot->mm_list);
1886 easy_to_free = 1;
1887 } else {
1888 list_move(&mm_slot->mm_list,
1889 &ksm_scan.mm_slot->mm_list);
1890 }
cd551f97 1891 }
cd551f97
HD
1892 spin_unlock(&ksm_mmlist_lock);
1893
9ba69294
HD
1894 if (easy_to_free) {
1895 free_mm_slot(mm_slot);
1896 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1897 mmdrop(mm);
1898 } else if (mm_slot) {
9ba69294
HD
1899 down_write(&mm->mmap_sem);
1900 up_write(&mm->mmap_sem);
9ba69294 1901 }
31dbd01f
IE
1902}
1903
cbf86cfe 1904struct page *ksm_might_need_to_copy(struct page *page,
5ad64688
HD
1905 struct vm_area_struct *vma, unsigned long address)
1906{
cbf86cfe 1907 struct anon_vma *anon_vma = page_anon_vma(page);
5ad64688
HD
1908 struct page *new_page;
1909
cbf86cfe
HD
1910 if (PageKsm(page)) {
1911 if (page_stable_node(page) &&
1912 !(ksm_run & KSM_RUN_UNMERGE))
1913 return page; /* no need to copy it */
1914 } else if (!anon_vma) {
1915 return page; /* no need to copy it */
1916 } else if (anon_vma->root == vma->anon_vma->root &&
1917 page->index == linear_page_index(vma, address)) {
1918 return page; /* still no need to copy it */
1919 }
1920 if (!PageUptodate(page))
1921 return page; /* let do_swap_page report the error */
1922
5ad64688
HD
1923 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1924 if (new_page) {
1925 copy_user_highpage(new_page, page, address, vma);
1926
1927 SetPageDirty(new_page);
1928 __SetPageUptodate(new_page);
48c935ad 1929 __SetPageLocked(new_page);
5ad64688
HD
1930 }
1931
5ad64688
HD
1932 return new_page;
1933}
1934
1df631ae 1935void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
e9995ef9
HD
1936{
1937 struct stable_node *stable_node;
e9995ef9 1938 struct rmap_item *rmap_item;
e9995ef9
HD
1939 int search_new_forks = 0;
1940
309381fe 1941 VM_BUG_ON_PAGE(!PageKsm(page), page);
9f32624b
JK
1942
1943 /*
1944 * Rely on the page lock to protect against concurrent modifications
1945 * to that page's node of the stable tree.
1946 */
309381fe 1947 VM_BUG_ON_PAGE(!PageLocked(page), page);
e9995ef9
HD
1948
1949 stable_node = page_stable_node(page);
1950 if (!stable_node)
1df631ae 1951 return;
e9995ef9 1952again:
b67bfe0d 1953 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
e9995ef9 1954 struct anon_vma *anon_vma = rmap_item->anon_vma;
5beb4930 1955 struct anon_vma_chain *vmac;
e9995ef9
HD
1956 struct vm_area_struct *vma;
1957
ad12695f 1958 cond_resched();
b6b19f25 1959 anon_vma_lock_read(anon_vma);
bf181b9f
ML
1960 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1961 0, ULONG_MAX) {
ad12695f 1962 cond_resched();
5beb4930 1963 vma = vmac->vma;
e9995ef9
HD
1964 if (rmap_item->address < vma->vm_start ||
1965 rmap_item->address >= vma->vm_end)
1966 continue;
1967 /*
1968 * Initially we examine only the vma which covers this
1969 * rmap_item; but later, if there is still work to do,
1970 * we examine covering vmas in other mms: in case they
1971 * were forked from the original since ksmd passed.
1972 */
1973 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1974 continue;
1975
0dd1c7bb
JK
1976 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1977 continue;
1978
e4b82222 1979 if (!rwc->rmap_one(page, vma,
1df631ae 1980 rmap_item->address, rwc->arg)) {
b6b19f25 1981 anon_vma_unlock_read(anon_vma);
1df631ae 1982 return;
e9995ef9 1983 }
0dd1c7bb
JK
1984 if (rwc->done && rwc->done(page)) {
1985 anon_vma_unlock_read(anon_vma);
1df631ae 1986 return;
0dd1c7bb 1987 }
e9995ef9 1988 }
b6b19f25 1989 anon_vma_unlock_read(anon_vma);
e9995ef9
HD
1990 }
1991 if (!search_new_forks++)
1992 goto again;
e9995ef9
HD
1993}
1994
52629506 1995#ifdef CONFIG_MIGRATION
e9995ef9
HD
1996void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1997{
1998 struct stable_node *stable_node;
1999
309381fe
SL
2000 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2001 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2002 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
e9995ef9
HD
2003
2004 stable_node = page_stable_node(newpage);
2005 if (stable_node) {
309381fe 2006 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
62b61f61 2007 stable_node->kpfn = page_to_pfn(newpage);
c8d6553b
HD
2008 /*
2009 * newpage->mapping was set in advance; now we need smp_wmb()
2010 * to make sure that the new stable_node->kpfn is visible
2011 * to get_ksm_page() before it can see that oldpage->mapping
2012 * has gone stale (or that PageSwapCache has been cleared).
2013 */
2014 smp_wmb();
2015 set_page_stable_node(oldpage, NULL);
e9995ef9
HD
2016 }
2017}
2018#endif /* CONFIG_MIGRATION */
2019
62b61f61 2020#ifdef CONFIG_MEMORY_HOTREMOVE
ef4d43a8
HD
2021static void wait_while_offlining(void)
2022{
2023 while (ksm_run & KSM_RUN_OFFLINE) {
2024 mutex_unlock(&ksm_thread_mutex);
2025 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
74316201 2026 TASK_UNINTERRUPTIBLE);
ef4d43a8
HD
2027 mutex_lock(&ksm_thread_mutex);
2028 }
2029}
2030
ee0ea59c
HD
2031static void ksm_check_stable_tree(unsigned long start_pfn,
2032 unsigned long end_pfn)
62b61f61 2033{
03640418 2034 struct stable_node *stable_node, *next;
62b61f61 2035 struct rb_node *node;
90bd6fd3 2036 int nid;
62b61f61 2037
ef53d16c
HD
2038 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2039 node = rb_first(root_stable_tree + nid);
ee0ea59c 2040 while (node) {
90bd6fd3
PH
2041 stable_node = rb_entry(node, struct stable_node, node);
2042 if (stable_node->kpfn >= start_pfn &&
ee0ea59c
HD
2043 stable_node->kpfn < end_pfn) {
2044 /*
2045 * Don't get_ksm_page, page has already gone:
2046 * which is why we keep kpfn instead of page*
2047 */
2048 remove_node_from_stable_tree(stable_node);
ef53d16c 2049 node = rb_first(root_stable_tree + nid);
ee0ea59c
HD
2050 } else
2051 node = rb_next(node);
2052 cond_resched();
90bd6fd3 2053 }
ee0ea59c 2054 }
03640418 2055 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
4146d2d6
HD
2056 if (stable_node->kpfn >= start_pfn &&
2057 stable_node->kpfn < end_pfn)
2058 remove_node_from_stable_tree(stable_node);
2059 cond_resched();
2060 }
62b61f61
HD
2061}
2062
2063static int ksm_memory_callback(struct notifier_block *self,
2064 unsigned long action, void *arg)
2065{
2066 struct memory_notify *mn = arg;
62b61f61
HD
2067
2068 switch (action) {
2069 case MEM_GOING_OFFLINE:
2070 /*
ef4d43a8
HD
2071 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2072 * and remove_all_stable_nodes() while memory is going offline:
2073 * it is unsafe for them to touch the stable tree at this time.
2074 * But unmerge_ksm_pages(), rmap lookups and other entry points
2075 * which do not need the ksm_thread_mutex are all safe.
62b61f61 2076 */
ef4d43a8
HD
2077 mutex_lock(&ksm_thread_mutex);
2078 ksm_run |= KSM_RUN_OFFLINE;
2079 mutex_unlock(&ksm_thread_mutex);
62b61f61
HD
2080 break;
2081
2082 case MEM_OFFLINE:
2083 /*
2084 * Most of the work is done by page migration; but there might
2085 * be a few stable_nodes left over, still pointing to struct
ee0ea59c
HD
2086 * pages which have been offlined: prune those from the tree,
2087 * otherwise get_ksm_page() might later try to access a
2088 * non-existent struct page.
62b61f61 2089 */
ee0ea59c
HD
2090 ksm_check_stable_tree(mn->start_pfn,
2091 mn->start_pfn + mn->nr_pages);
62b61f61
HD
2092 /* fallthrough */
2093
2094 case MEM_CANCEL_OFFLINE:
ef4d43a8
HD
2095 mutex_lock(&ksm_thread_mutex);
2096 ksm_run &= ~KSM_RUN_OFFLINE;
62b61f61 2097 mutex_unlock(&ksm_thread_mutex);
ef4d43a8
HD
2098
2099 smp_mb(); /* wake_up_bit advises this */
2100 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
62b61f61
HD
2101 break;
2102 }
2103 return NOTIFY_OK;
2104}
ef4d43a8
HD
2105#else
2106static void wait_while_offlining(void)
2107{
2108}
62b61f61
HD
2109#endif /* CONFIG_MEMORY_HOTREMOVE */
2110
2ffd8679
HD
2111#ifdef CONFIG_SYSFS
2112/*
2113 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2114 */
2115
31dbd01f
IE
2116#define KSM_ATTR_RO(_name) \
2117 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2118#define KSM_ATTR(_name) \
2119 static struct kobj_attribute _name##_attr = \
2120 __ATTR(_name, 0644, _name##_show, _name##_store)
2121
2122static ssize_t sleep_millisecs_show(struct kobject *kobj,
2123 struct kobj_attribute *attr, char *buf)
2124{
2125 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2126}
2127
2128static ssize_t sleep_millisecs_store(struct kobject *kobj,
2129 struct kobj_attribute *attr,
2130 const char *buf, size_t count)
2131{
2132 unsigned long msecs;
2133 int err;
2134
3dbb95f7 2135 err = kstrtoul(buf, 10, &msecs);
31dbd01f
IE
2136 if (err || msecs > UINT_MAX)
2137 return -EINVAL;
2138
2139 ksm_thread_sleep_millisecs = msecs;
2140
2141 return count;
2142}
2143KSM_ATTR(sleep_millisecs);
2144
2145static ssize_t pages_to_scan_show(struct kobject *kobj,
2146 struct kobj_attribute *attr, char *buf)
2147{
2148 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2149}
2150
2151static ssize_t pages_to_scan_store(struct kobject *kobj,
2152 struct kobj_attribute *attr,
2153 const char *buf, size_t count)
2154{
2155 int err;
2156 unsigned long nr_pages;
2157
3dbb95f7 2158 err = kstrtoul(buf, 10, &nr_pages);
31dbd01f
IE
2159 if (err || nr_pages > UINT_MAX)
2160 return -EINVAL;
2161
2162 ksm_thread_pages_to_scan = nr_pages;
2163
2164 return count;
2165}
2166KSM_ATTR(pages_to_scan);
2167
2168static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2169 char *buf)
2170{
ef4d43a8 2171 return sprintf(buf, "%lu\n", ksm_run);
31dbd01f
IE
2172}
2173
2174static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2175 const char *buf, size_t count)
2176{
2177 int err;
2178 unsigned long flags;
2179
3dbb95f7 2180 err = kstrtoul(buf, 10, &flags);
31dbd01f
IE
2181 if (err || flags > UINT_MAX)
2182 return -EINVAL;
2183 if (flags > KSM_RUN_UNMERGE)
2184 return -EINVAL;
2185
2186 /*
2187 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2188 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
d0f209f6
HD
2189 * breaking COW to free the pages_shared (but leaves mm_slots
2190 * on the list for when ksmd may be set running again).
31dbd01f
IE
2191 */
2192
2193 mutex_lock(&ksm_thread_mutex);
ef4d43a8 2194 wait_while_offlining();
31dbd01f
IE
2195 if (ksm_run != flags) {
2196 ksm_run = flags;
d952b791 2197 if (flags & KSM_RUN_UNMERGE) {
e1e12d2f 2198 set_current_oom_origin();
d952b791 2199 err = unmerge_and_remove_all_rmap_items();
e1e12d2f 2200 clear_current_oom_origin();
d952b791
HD
2201 if (err) {
2202 ksm_run = KSM_RUN_STOP;
2203 count = err;
2204 }
2205 }
31dbd01f
IE
2206 }
2207 mutex_unlock(&ksm_thread_mutex);
2208
2209 if (flags & KSM_RUN_MERGE)
2210 wake_up_interruptible(&ksm_thread_wait);
2211
2212 return count;
2213}
2214KSM_ATTR(run);
2215
90bd6fd3
PH
2216#ifdef CONFIG_NUMA
2217static ssize_t merge_across_nodes_show(struct kobject *kobj,
2218 struct kobj_attribute *attr, char *buf)
2219{
2220 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2221}
2222
2223static ssize_t merge_across_nodes_store(struct kobject *kobj,
2224 struct kobj_attribute *attr,
2225 const char *buf, size_t count)
2226{
2227 int err;
2228 unsigned long knob;
2229
2230 err = kstrtoul(buf, 10, &knob);
2231 if (err)
2232 return err;
2233 if (knob > 1)
2234 return -EINVAL;
2235
2236 mutex_lock(&ksm_thread_mutex);
ef4d43a8 2237 wait_while_offlining();
90bd6fd3 2238 if (ksm_merge_across_nodes != knob) {
cbf86cfe 2239 if (ksm_pages_shared || remove_all_stable_nodes())
90bd6fd3 2240 err = -EBUSY;
ef53d16c
HD
2241 else if (root_stable_tree == one_stable_tree) {
2242 struct rb_root *buf;
2243 /*
2244 * This is the first time that we switch away from the
2245 * default of merging across nodes: must now allocate
2246 * a buffer to hold as many roots as may be needed.
2247 * Allocate stable and unstable together:
2248 * MAXSMP NODES_SHIFT 10 will use 16kB.
2249 */
bafe1e14
JP
2250 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2251 GFP_KERNEL);
ef53d16c
HD
2252 /* Let us assume that RB_ROOT is NULL is zero */
2253 if (!buf)
2254 err = -ENOMEM;
2255 else {
2256 root_stable_tree = buf;
2257 root_unstable_tree = buf + nr_node_ids;
2258 /* Stable tree is empty but not the unstable */
2259 root_unstable_tree[0] = one_unstable_tree[0];
2260 }
2261 }
2262 if (!err) {
90bd6fd3 2263 ksm_merge_across_nodes = knob;
ef53d16c
HD
2264 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2265 }
90bd6fd3
PH
2266 }
2267 mutex_unlock(&ksm_thread_mutex);
2268
2269 return err ? err : count;
2270}
2271KSM_ATTR(merge_across_nodes);
2272#endif
2273
e86c59b1
CI
2274static ssize_t use_zero_pages_show(struct kobject *kobj,
2275 struct kobj_attribute *attr, char *buf)
2276{
2277 return sprintf(buf, "%u\n", ksm_use_zero_pages);
2278}
2279static ssize_t use_zero_pages_store(struct kobject *kobj,
2280 struct kobj_attribute *attr,
2281 const char *buf, size_t count)
2282{
2283 int err;
2284 bool value;
2285
2286 err = kstrtobool(buf, &value);
2287 if (err)
2288 return -EINVAL;
2289
2290 ksm_use_zero_pages = value;
2291
2292 return count;
2293}
2294KSM_ATTR(use_zero_pages);
2295
b4028260
HD
2296static ssize_t pages_shared_show(struct kobject *kobj,
2297 struct kobj_attribute *attr, char *buf)
2298{
2299 return sprintf(buf, "%lu\n", ksm_pages_shared);
2300}
2301KSM_ATTR_RO(pages_shared);
2302
2303static ssize_t pages_sharing_show(struct kobject *kobj,
2304 struct kobj_attribute *attr, char *buf)
2305{
e178dfde 2306 return sprintf(buf, "%lu\n", ksm_pages_sharing);
b4028260
HD
2307}
2308KSM_ATTR_RO(pages_sharing);
2309
473b0ce4
HD
2310static ssize_t pages_unshared_show(struct kobject *kobj,
2311 struct kobj_attribute *attr, char *buf)
2312{
2313 return sprintf(buf, "%lu\n", ksm_pages_unshared);
2314}
2315KSM_ATTR_RO(pages_unshared);
2316
2317static ssize_t pages_volatile_show(struct kobject *kobj,
2318 struct kobj_attribute *attr, char *buf)
2319{
2320 long ksm_pages_volatile;
2321
2322 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2323 - ksm_pages_sharing - ksm_pages_unshared;
2324 /*
2325 * It was not worth any locking to calculate that statistic,
2326 * but it might therefore sometimes be negative: conceal that.
2327 */
2328 if (ksm_pages_volatile < 0)
2329 ksm_pages_volatile = 0;
2330 return sprintf(buf, "%ld\n", ksm_pages_volatile);
2331}
2332KSM_ATTR_RO(pages_volatile);
2333
2334static ssize_t full_scans_show(struct kobject *kobj,
2335 struct kobj_attribute *attr, char *buf)
2336{
2337 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2338}
2339KSM_ATTR_RO(full_scans);
2340
31dbd01f
IE
2341static struct attribute *ksm_attrs[] = {
2342 &sleep_millisecs_attr.attr,
2343 &pages_to_scan_attr.attr,
2344 &run_attr.attr,
b4028260
HD
2345 &pages_shared_attr.attr,
2346 &pages_sharing_attr.attr,
473b0ce4
HD
2347 &pages_unshared_attr.attr,
2348 &pages_volatile_attr.attr,
2349 &full_scans_attr.attr,
90bd6fd3
PH
2350#ifdef CONFIG_NUMA
2351 &merge_across_nodes_attr.attr,
2352#endif
e86c59b1 2353 &use_zero_pages_attr.attr,
31dbd01f
IE
2354 NULL,
2355};
2356
2357static struct attribute_group ksm_attr_group = {
2358 .attrs = ksm_attrs,
2359 .name = "ksm",
2360};
2ffd8679 2361#endif /* CONFIG_SYSFS */
31dbd01f
IE
2362
2363static int __init ksm_init(void)
2364{
2365 struct task_struct *ksm_thread;
2366 int err;
2367
e86c59b1
CI
2368 /* The correct value depends on page size and endianness */
2369 zero_checksum = calc_checksum(ZERO_PAGE(0));
2370 /* Default to false for backwards compatibility */
2371 ksm_use_zero_pages = false;
2372
31dbd01f
IE
2373 err = ksm_slab_init();
2374 if (err)
2375 goto out;
2376
31dbd01f
IE
2377 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2378 if (IS_ERR(ksm_thread)) {
25acde31 2379 pr_err("ksm: creating kthread failed\n");
31dbd01f 2380 err = PTR_ERR(ksm_thread);
d9f8984c 2381 goto out_free;
31dbd01f
IE
2382 }
2383
2ffd8679 2384#ifdef CONFIG_SYSFS
31dbd01f
IE
2385 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2386 if (err) {
25acde31 2387 pr_err("ksm: register sysfs failed\n");
2ffd8679 2388 kthread_stop(ksm_thread);
d9f8984c 2389 goto out_free;
31dbd01f 2390 }
c73602ad
HD
2391#else
2392 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2393
2ffd8679 2394#endif /* CONFIG_SYSFS */
31dbd01f 2395
62b61f61 2396#ifdef CONFIG_MEMORY_HOTREMOVE
ef4d43a8 2397 /* There is no significance to this priority 100 */
62b61f61
HD
2398 hotplug_memory_notifier(ksm_memory_callback, 100);
2399#endif
31dbd01f
IE
2400 return 0;
2401
d9f8984c 2402out_free:
31dbd01f
IE
2403 ksm_slab_free();
2404out:
2405 return err;
f8af4da3 2406}
a64fb3cd 2407subsys_initcall(ksm_init);