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/sched/cputime.h>
25 #include <linux/rwsem.h>
26 #include <linux/pagemap.h>
27 #include <linux/rmap.h>
28 #include <linux/spinlock.h>
29 #include <linux/xxhash.h>
30 #include <linux/delay.h>
31 #include <linux/kthread.h>
32 #include <linux/wait.h>
33 #include <linux/slab.h>
34 #include <linux/rbtree.h>
35 #include <linux/memory.h>
36 #include <linux/mmu_notifier.h>
37 #include <linux/swap.h>
38 #include <linux/ksm.h>
39 #include <linux/hashtable.h>
40 #include <linux/freezer.h>
41 #include <linux/oom.h>
42 #include <linux/numa.h>
43 #include <linux/pagewalk.h>
44
45 #include <asm/tlbflush.h>
46 #include "internal.h"
47 #include "mm_slot.h"
48
49 #define CREATE_TRACE_POINTS
50 #include <trace/events/ksm.h>
51
52 #ifdef CONFIG_NUMA
53 #define NUMA(x) (x)
54 #define DO_NUMA(x) do { (x); } while (0)
55 #else
56 #define NUMA(x) (0)
57 #define DO_NUMA(x) do { } while (0)
58 #endif
59
60 typedef u8 rmap_age_t;
61
62 /**
63 * DOC: Overview
64 *
65 * A few notes about the KSM scanning process,
66 * to make it easier to understand the data structures below:
67 *
68 * In order to reduce excessive scanning, KSM sorts the memory pages by their
69 * contents into a data structure that holds pointers to the pages' locations.
70 *
71 * Since the contents of the pages may change at any moment, KSM cannot just
72 * insert the pages into a normal sorted tree and expect it to find anything.
73 * Therefore KSM uses two data structures - the stable and the unstable tree.
74 *
75 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
76 * by their contents. Because each such page is write-protected, searching on
77 * this tree is fully assured to be working (except when pages are unmapped),
78 * and therefore this tree is called the stable tree.
79 *
80 * The stable tree node includes information required for reverse
81 * mapping from a KSM page to virtual addresses that map this page.
82 *
83 * In order to avoid large latencies of the rmap walks on KSM pages,
84 * KSM maintains two types of nodes in the stable tree:
85 *
86 * * the regular nodes that keep the reverse mapping structures in a
87 * linked list
88 * * the "chains" that link nodes ("dups") that represent the same
89 * write protected memory content, but each "dup" corresponds to a
90 * different KSM page copy of that content
91 *
92 * Internally, the regular nodes, "dups" and "chains" are represented
93 * using the same struct ksm_stable_node structure.
94 *
95 * In addition to the stable tree, KSM uses a second data structure called the
96 * unstable tree: this tree holds pointers to pages which have been found to
97 * be "unchanged for a period of time". The unstable tree sorts these pages
98 * by their contents, but since they are not write-protected, KSM cannot rely
99 * upon the unstable tree to work correctly - the unstable tree is liable to
100 * be corrupted as its contents are modified, and so it is called unstable.
101 *
102 * KSM solves this problem by several techniques:
103 *
104 * 1) The unstable tree is flushed every time KSM completes scanning all
105 * memory areas, and then the tree is rebuilt again from the beginning.
106 * 2) KSM will only insert into the unstable tree, pages whose hash value
107 * has not changed since the previous scan of all memory areas.
108 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
109 * colors of the nodes and not on their contents, assuring that even when
110 * the tree gets "corrupted" it won't get out of balance, so scanning time
111 * remains the same (also, searching and inserting nodes in an rbtree uses
112 * the same algorithm, so we have no overhead when we flush and rebuild).
113 * 4) KSM never flushes the stable tree, which means that even if it were to
114 * take 10 attempts to find a page in the unstable tree, once it is found,
115 * it is secured in the stable tree. (When we scan a new page, we first
116 * compare it against the stable tree, and then against the unstable tree.)
117 *
118 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
119 * stable trees and multiple unstable trees: one of each for each NUMA node.
120 */
121
122 /**
123 * struct ksm_mm_slot - ksm information per mm that is being scanned
124 * @slot: hash lookup from mm to mm_slot
125 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
126 */
127 struct ksm_mm_slot {
128 struct mm_slot slot;
129 struct ksm_rmap_item *rmap_list;
130 };
131
132 /**
133 * struct ksm_scan - cursor for scanning
134 * @mm_slot: the current mm_slot we are scanning
135 * @address: the next address inside that to be scanned
136 * @rmap_list: link to the next rmap to be scanned in the rmap_list
137 * @seqnr: count of completed full scans (needed when removing unstable node)
138 *
139 * There is only the one ksm_scan instance of this cursor structure.
140 */
141 struct ksm_scan {
142 struct ksm_mm_slot *mm_slot;
143 unsigned long address;
144 struct ksm_rmap_item **rmap_list;
145 unsigned long seqnr;
146 };
147
148 /**
149 * struct ksm_stable_node - node of the stable rbtree
150 * @node: rb node of this ksm page in the stable tree
151 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
152 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
153 * @list: linked into migrate_nodes, pending placement in the proper node tree
154 * @hlist: hlist head of rmap_items using this ksm page
155 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
156 * @chain_prune_time: time of the last full garbage collection
157 * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
158 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
159 */
160 struct ksm_stable_node {
161 union {
162 struct rb_node node; /* when node of stable tree */
163 struct { /* when listed for migration */
164 struct list_head *head;
165 struct {
166 struct hlist_node hlist_dup;
167 struct list_head list;
168 };
169 };
170 };
171 struct hlist_head hlist;
172 union {
173 unsigned long kpfn;
174 unsigned long chain_prune_time;
175 };
176 /*
177 * STABLE_NODE_CHAIN can be any negative number in
178 * rmap_hlist_len negative range, but better not -1 to be able
179 * to reliably detect underflows.
180 */
181 #define STABLE_NODE_CHAIN -1024
182 int rmap_hlist_len;
183 #ifdef CONFIG_NUMA
184 int nid;
185 #endif
186 };
187
188 /**
189 * struct ksm_rmap_item - reverse mapping item for virtual addresses
190 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
191 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
192 * @nid: NUMA node id of unstable tree in which linked (may not match page)
193 * @mm: the memory structure this rmap_item is pointing into
194 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
195 * @oldchecksum: previous checksum of the page at that virtual address
196 * @node: rb node of this rmap_item in the unstable tree
197 * @head: pointer to stable_node heading this list in the stable tree
198 * @hlist: link into hlist of rmap_items hanging off that stable_node
199 * @age: number of scan iterations since creation
200 * @remaining_skips: how many scans to skip
201 */
202 struct ksm_rmap_item {
203 struct ksm_rmap_item *rmap_list;
204 union {
205 struct anon_vma *anon_vma; /* when stable */
206 #ifdef CONFIG_NUMA
207 int nid; /* when node of unstable tree */
208 #endif
209 };
210 struct mm_struct *mm;
211 unsigned long address; /* + low bits used for flags below */
212 unsigned int oldchecksum; /* when unstable */
213 rmap_age_t age;
214 rmap_age_t remaining_skips;
215 union {
216 struct rb_node node; /* when node of unstable tree */
217 struct { /* when listed from stable tree */
218 struct ksm_stable_node *head;
219 struct hlist_node hlist;
220 };
221 };
222 };
223
224 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
225 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
226 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
227
228 /* The stable and unstable tree heads */
229 static struct rb_root one_stable_tree[1] = { RB_ROOT };
230 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
231 static struct rb_root *root_stable_tree = one_stable_tree;
232 static struct rb_root *root_unstable_tree = one_unstable_tree;
233
234 /* Recently migrated nodes of stable tree, pending proper placement */
235 static LIST_HEAD(migrate_nodes);
236 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
237
238 #define MM_SLOTS_HASH_BITS 10
239 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
240
241 static struct ksm_mm_slot ksm_mm_head = {
242 .slot.mm_node = LIST_HEAD_INIT(ksm_mm_head.slot.mm_node),
243 };
244 static struct ksm_scan ksm_scan = {
245 .mm_slot = &ksm_mm_head,
246 };
247
248 static struct kmem_cache *rmap_item_cache;
249 static struct kmem_cache *stable_node_cache;
250 static struct kmem_cache *mm_slot_cache;
251
252 /* Default number of pages to scan per batch */
253 #define DEFAULT_PAGES_TO_SCAN 100
254
255 /* The number of pages scanned */
256 static unsigned long ksm_pages_scanned;
257
258 /* The number of nodes in the stable tree */
259 static unsigned long ksm_pages_shared;
260
261 /* The number of page slots additionally sharing those nodes */
262 static unsigned long ksm_pages_sharing;
263
264 /* The number of nodes in the unstable tree */
265 static unsigned long ksm_pages_unshared;
266
267 /* The number of rmap_items in use: to calculate pages_volatile */
268 static unsigned long ksm_rmap_items;
269
270 /* The number of stable_node chains */
271 static unsigned long ksm_stable_node_chains;
272
273 /* The number of stable_node dups linked to the stable_node chains */
274 static unsigned long ksm_stable_node_dups;
275
276 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
277 static unsigned int ksm_stable_node_chains_prune_millisecs = 2000;
278
279 /* Maximum number of page slots sharing a stable node */
280 static int ksm_max_page_sharing = 256;
281
282 /* Number of pages ksmd should scan in one batch */
283 static unsigned int ksm_thread_pages_to_scan = DEFAULT_PAGES_TO_SCAN;
284
285 /* Milliseconds ksmd should sleep between batches */
286 static unsigned int ksm_thread_sleep_millisecs = 20;
287
288 /* Checksum of an empty (zeroed) page */
289 static unsigned int zero_checksum __read_mostly;
290
291 /* Whether to merge empty (zeroed) pages with actual zero pages */
292 static bool ksm_use_zero_pages __read_mostly;
293
294 /* Skip pages that couldn't be de-duplicated previously */
295 /* Default to true at least temporarily, for testing */
296 static bool ksm_smart_scan = true;
297
298 /* The number of zero pages which is placed by KSM */
299 atomic_long_t ksm_zero_pages = ATOMIC_LONG_INIT(0);
300
301 /* The number of pages that have been skipped due to "smart scanning" */
302 static unsigned long ksm_pages_skipped;
303
304 /* Don't scan more than max pages per batch. */
305 static unsigned long ksm_advisor_max_pages_to_scan = 30000;
306
307 /* Min CPU for scanning pages per scan */
308 #define KSM_ADVISOR_MIN_CPU 10
309
310 /* Max CPU for scanning pages per scan */
311 static unsigned int ksm_advisor_max_cpu = 70;
312
313 /* Target scan time in seconds to analyze all KSM candidate pages. */
314 static unsigned long ksm_advisor_target_scan_time = 200;
315
316 /* Exponentially weighted moving average. */
317 #define EWMA_WEIGHT 30
318
319 /**
320 * struct advisor_ctx - metadata for KSM advisor
321 * @start_scan: start time of the current scan
322 * @scan_time: scan time of previous scan
323 * @change: change in percent to pages_to_scan parameter
324 * @cpu_time: cpu time consumed by the ksmd thread in the previous scan
325 */
326 struct advisor_ctx {
327 ktime_t start_scan;
328 unsigned long scan_time;
329 unsigned long change;
330 unsigned long long cpu_time;
331 };
332 static struct advisor_ctx advisor_ctx;
333
334 /* Define different advisor's */
335 enum ksm_advisor_type {
336 KSM_ADVISOR_NONE,
337 KSM_ADVISOR_SCAN_TIME,
338 };
339 static enum ksm_advisor_type ksm_advisor;
340
341 #ifdef CONFIG_SYSFS
342 /*
343 * Only called through the sysfs control interface:
344 */
345
346 /* At least scan this many pages per batch. */
347 static unsigned long ksm_advisor_min_pages_to_scan = 500;
348
349 static void set_advisor_defaults(void)
350 {
351 if (ksm_advisor == KSM_ADVISOR_NONE) {
352 ksm_thread_pages_to_scan = DEFAULT_PAGES_TO_SCAN;
353 } else if (ksm_advisor == KSM_ADVISOR_SCAN_TIME) {
354 advisor_ctx = (const struct advisor_ctx){ 0 };
355 ksm_thread_pages_to_scan = ksm_advisor_min_pages_to_scan;
356 }
357 }
358 #endif /* CONFIG_SYSFS */
359
360 static inline void advisor_start_scan(void)
361 {
362 if (ksm_advisor == KSM_ADVISOR_SCAN_TIME)
363 advisor_ctx.start_scan = ktime_get();
364 }
365
366 /*
367 * Use previous scan time if available, otherwise use current scan time as an
368 * approximation for the previous scan time.
369 */
370 static inline unsigned long prev_scan_time(struct advisor_ctx *ctx,
371 unsigned long scan_time)
372 {
373 return ctx->scan_time ? ctx->scan_time : scan_time;
374 }
375
376 /* Calculate exponential weighted moving average */
377 static unsigned long ewma(unsigned long prev, unsigned long curr)
378 {
379 return ((100 - EWMA_WEIGHT) * prev + EWMA_WEIGHT * curr) / 100;
380 }
381
382 /*
383 * The scan time advisor is based on the current scan rate and the target
384 * scan rate.
385 *
386 * new_pages_to_scan = pages_to_scan * (scan_time / target_scan_time)
387 *
388 * To avoid perturbations it calculates a change factor of previous changes.
389 * A new change factor is calculated for each iteration and it uses an
390 * exponentially weighted moving average. The new pages_to_scan value is
391 * multiplied with that change factor:
392 *
393 * new_pages_to_scan *= change facor
394 *
395 * The new_pages_to_scan value is limited by the cpu min and max values. It
396 * calculates the cpu percent for the last scan and calculates the new
397 * estimated cpu percent cost for the next scan. That value is capped by the
398 * cpu min and max setting.
399 *
400 * In addition the new pages_to_scan value is capped by the max and min
401 * limits.
402 */
403 static void scan_time_advisor(void)
404 {
405 unsigned int cpu_percent;
406 unsigned long cpu_time;
407 unsigned long cpu_time_diff;
408 unsigned long cpu_time_diff_ms;
409 unsigned long pages;
410 unsigned long per_page_cost;
411 unsigned long factor;
412 unsigned long change;
413 unsigned long last_scan_time;
414 unsigned long scan_time;
415
416 /* Convert scan time to seconds */
417 scan_time = div_s64(ktime_ms_delta(ktime_get(), advisor_ctx.start_scan),
418 MSEC_PER_SEC);
419 scan_time = scan_time ? scan_time : 1;
420
421 /* Calculate CPU consumption of ksmd background thread */
422 cpu_time = task_sched_runtime(current);
423 cpu_time_diff = cpu_time - advisor_ctx.cpu_time;
424 cpu_time_diff_ms = cpu_time_diff / 1000 / 1000;
425
426 cpu_percent = (cpu_time_diff_ms * 100) / (scan_time * 1000);
427 cpu_percent = cpu_percent ? cpu_percent : 1;
428 last_scan_time = prev_scan_time(&advisor_ctx, scan_time);
429
430 /* Calculate scan time as percentage of target scan time */
431 factor = ksm_advisor_target_scan_time * 100 / scan_time;
432 factor = factor ? factor : 1;
433
434 /*
435 * Calculate scan time as percentage of last scan time and use
436 * exponentially weighted average to smooth it
437 */
438 change = scan_time * 100 / last_scan_time;
439 change = change ? change : 1;
440 change = ewma(advisor_ctx.change, change);
441
442 /* Calculate new scan rate based on target scan rate. */
443 pages = ksm_thread_pages_to_scan * 100 / factor;
444 /* Update pages_to_scan by weighted change percentage. */
445 pages = pages * change / 100;
446
447 /* Cap new pages_to_scan value */
448 per_page_cost = ksm_thread_pages_to_scan / cpu_percent;
449 per_page_cost = per_page_cost ? per_page_cost : 1;
450
451 pages = min(pages, per_page_cost * ksm_advisor_max_cpu);
452 pages = max(pages, per_page_cost * KSM_ADVISOR_MIN_CPU);
453 pages = min(pages, ksm_advisor_max_pages_to_scan);
454
455 /* Update advisor context */
456 advisor_ctx.change = change;
457 advisor_ctx.scan_time = scan_time;
458 advisor_ctx.cpu_time = cpu_time;
459
460 ksm_thread_pages_to_scan = pages;
461 trace_ksm_advisor(scan_time, pages, cpu_percent);
462 }
463
464 static void advisor_stop_scan(void)
465 {
466 if (ksm_advisor == KSM_ADVISOR_SCAN_TIME)
467 scan_time_advisor();
468 }
469
470 #ifdef CONFIG_NUMA
471 /* Zeroed when merging across nodes is not allowed */
472 static unsigned int ksm_merge_across_nodes = 1;
473 static int ksm_nr_node_ids = 1;
474 #else
475 #define ksm_merge_across_nodes 1U
476 #define ksm_nr_node_ids 1
477 #endif
478
479 #define KSM_RUN_STOP 0
480 #define KSM_RUN_MERGE 1
481 #define KSM_RUN_UNMERGE 2
482 #define KSM_RUN_OFFLINE 4
483 static unsigned long ksm_run = KSM_RUN_STOP;
484 static void wait_while_offlining(void);
485
486 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
487 static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait);
488 static DEFINE_MUTEX(ksm_thread_mutex);
489 static DEFINE_SPINLOCK(ksm_mmlist_lock);
490
491 static int __init ksm_slab_init(void)
492 {
493 rmap_item_cache = KMEM_CACHE(ksm_rmap_item, 0);
494 if (!rmap_item_cache)
495 goto out;
496
497 stable_node_cache = KMEM_CACHE(ksm_stable_node, 0);
498 if (!stable_node_cache)
499 goto out_free1;
500
501 mm_slot_cache = KMEM_CACHE(ksm_mm_slot, 0);
502 if (!mm_slot_cache)
503 goto out_free2;
504
505 return 0;
506
507 out_free2:
508 kmem_cache_destroy(stable_node_cache);
509 out_free1:
510 kmem_cache_destroy(rmap_item_cache);
511 out:
512 return -ENOMEM;
513 }
514
515 static void __init ksm_slab_free(void)
516 {
517 kmem_cache_destroy(mm_slot_cache);
518 kmem_cache_destroy(stable_node_cache);
519 kmem_cache_destroy(rmap_item_cache);
520 mm_slot_cache = NULL;
521 }
522
523 static __always_inline bool is_stable_node_chain(struct ksm_stable_node *chain)
524 {
525 return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
526 }
527
528 static __always_inline bool is_stable_node_dup(struct ksm_stable_node *dup)
529 {
530 return dup->head == STABLE_NODE_DUP_HEAD;
531 }
532
533 static inline void stable_node_chain_add_dup(struct ksm_stable_node *dup,
534 struct ksm_stable_node *chain)
535 {
536 VM_BUG_ON(is_stable_node_dup(dup));
537 dup->head = STABLE_NODE_DUP_HEAD;
538 VM_BUG_ON(!is_stable_node_chain(chain));
539 hlist_add_head(&dup->hlist_dup, &chain->hlist);
540 ksm_stable_node_dups++;
541 }
542
543 static inline void __stable_node_dup_del(struct ksm_stable_node *dup)
544 {
545 VM_BUG_ON(!is_stable_node_dup(dup));
546 hlist_del(&dup->hlist_dup);
547 ksm_stable_node_dups--;
548 }
549
550 static inline void stable_node_dup_del(struct ksm_stable_node *dup)
551 {
552 VM_BUG_ON(is_stable_node_chain(dup));
553 if (is_stable_node_dup(dup))
554 __stable_node_dup_del(dup);
555 else
556 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
557 #ifdef CONFIG_DEBUG_VM
558 dup->head = NULL;
559 #endif
560 }
561
562 static inline struct ksm_rmap_item *alloc_rmap_item(void)
563 {
564 struct ksm_rmap_item *rmap_item;
565
566 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
567 __GFP_NORETRY | __GFP_NOWARN);
568 if (rmap_item)
569 ksm_rmap_items++;
570 return rmap_item;
571 }
572
573 static inline void free_rmap_item(struct ksm_rmap_item *rmap_item)
574 {
575 ksm_rmap_items--;
576 rmap_item->mm->ksm_rmap_items--;
577 rmap_item->mm = NULL; /* debug safety */
578 kmem_cache_free(rmap_item_cache, rmap_item);
579 }
580
581 static inline struct ksm_stable_node *alloc_stable_node(void)
582 {
583 /*
584 * The allocation can take too long with GFP_KERNEL when memory is under
585 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
586 * grants access to memory reserves, helping to avoid this problem.
587 */
588 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
589 }
590
591 static inline void free_stable_node(struct ksm_stable_node *stable_node)
592 {
593 VM_BUG_ON(stable_node->rmap_hlist_len &&
594 !is_stable_node_chain(stable_node));
595 kmem_cache_free(stable_node_cache, stable_node);
596 }
597
598 /*
599 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
600 * page tables after it has passed through ksm_exit() - which, if necessary,
601 * takes mmap_lock briefly to serialize against them. ksm_exit() does not set
602 * a special flag: they can just back out as soon as mm_users goes to zero.
603 * ksm_test_exit() is used throughout to make this test for exit: in some
604 * places for correctness, in some places just to avoid unnecessary work.
605 */
606 static inline bool ksm_test_exit(struct mm_struct *mm)
607 {
608 return atomic_read(&mm->mm_users) == 0;
609 }
610
611 static int break_ksm_pmd_entry(pmd_t *pmd, unsigned long addr, unsigned long next,
612 struct mm_walk *walk)
613 {
614 struct page *page = NULL;
615 spinlock_t *ptl;
616 pte_t *pte;
617 pte_t ptent;
618 int ret;
619
620 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
621 if (!pte)
622 return 0;
623 ptent = ptep_get(pte);
624 if (pte_present(ptent)) {
625 page = vm_normal_page(walk->vma, addr, ptent);
626 } else if (!pte_none(ptent)) {
627 swp_entry_t entry = pte_to_swp_entry(ptent);
628
629 /*
630 * As KSM pages remain KSM pages until freed, no need to wait
631 * here for migration to end.
632 */
633 if (is_migration_entry(entry))
634 page = pfn_swap_entry_to_page(entry);
635 }
636 /* return 1 if the page is an normal ksm page or KSM-placed zero page */
637 ret = (page && PageKsm(page)) || is_ksm_zero_pte(ptent);
638 pte_unmap_unlock(pte, ptl);
639 return ret;
640 }
641
642 static const struct mm_walk_ops break_ksm_ops = {
643 .pmd_entry = break_ksm_pmd_entry,
644 .walk_lock = PGWALK_RDLOCK,
645 };
646
647 static const struct mm_walk_ops break_ksm_lock_vma_ops = {
648 .pmd_entry = break_ksm_pmd_entry,
649 .walk_lock = PGWALK_WRLOCK,
650 };
651
652 /*
653 * We use break_ksm to break COW on a ksm page by triggering unsharing,
654 * such that the ksm page will get replaced by an exclusive anonymous page.
655 *
656 * We take great care only to touch a ksm page, in a VM_MERGEABLE vma,
657 * in case the application has unmapped and remapped mm,addr meanwhile.
658 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
659 * mmap of /dev/mem, where we would not want to touch it.
660 *
661 * FAULT_FLAG_REMOTE/FOLL_REMOTE are because we do this outside the context
662 * of the process that owns 'vma'. We also do not want to enforce
663 * protection keys here anyway.
664 */
665 static int break_ksm(struct vm_area_struct *vma, unsigned long addr, bool lock_vma)
666 {
667 vm_fault_t ret = 0;
668 const struct mm_walk_ops *ops = lock_vma ?
669 &break_ksm_lock_vma_ops : &break_ksm_ops;
670
671 do {
672 int ksm_page;
673
674 cond_resched();
675 ksm_page = walk_page_range_vma(vma, addr, addr + 1, ops, NULL);
676 if (WARN_ON_ONCE(ksm_page < 0))
677 return ksm_page;
678 if (!ksm_page)
679 return 0;
680 ret = handle_mm_fault(vma, addr,
681 FAULT_FLAG_UNSHARE | FAULT_FLAG_REMOTE,
682 NULL);
683 } while (!(ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
684 /*
685 * We must loop until we no longer find a KSM page because
686 * handle_mm_fault() may back out if there's any difficulty e.g. if
687 * pte accessed bit gets updated concurrently.
688 *
689 * VM_FAULT_SIGBUS could occur if we race with truncation of the
690 * backing file, which also invalidates anonymous pages: that's
691 * okay, that truncation will have unmapped the PageKsm for us.
692 *
693 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
694 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
695 * current task has TIF_MEMDIE set, and will be OOM killed on return
696 * to user; and ksmd, having no mm, would never be chosen for that.
697 *
698 * But if the mm is in a limited mem_cgroup, then the fault may fail
699 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
700 * even ksmd can fail in this way - though it's usually breaking ksm
701 * just to undo a merge it made a moment before, so unlikely to oom.
702 *
703 * That's a pity: we might therefore have more kernel pages allocated
704 * than we're counting as nodes in the stable tree; but ksm_do_scan
705 * will retry to break_cow on each pass, so should recover the page
706 * in due course. The important thing is to not let VM_MERGEABLE
707 * be cleared while any such pages might remain in the area.
708 */
709 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
710 }
711
712 static bool vma_ksm_compatible(struct vm_area_struct *vma)
713 {
714 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE | VM_PFNMAP |
715 VM_IO | VM_DONTEXPAND | VM_HUGETLB |
716 VM_MIXEDMAP| VM_DROPPABLE))
717 return false; /* just ignore the advice */
718
719 if (vma_is_dax(vma))
720 return false;
721
722 #ifdef VM_SAO
723 if (vma->vm_flags & VM_SAO)
724 return false;
725 #endif
726 #ifdef VM_SPARC_ADI
727 if (vma->vm_flags & VM_SPARC_ADI)
728 return false;
729 #endif
730
731 return true;
732 }
733
734 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
735 unsigned long addr)
736 {
737 struct vm_area_struct *vma;
738 if (ksm_test_exit(mm))
739 return NULL;
740 vma = vma_lookup(mm, addr);
741 if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
742 return NULL;
743 return vma;
744 }
745
746 static void break_cow(struct ksm_rmap_item *rmap_item)
747 {
748 struct mm_struct *mm = rmap_item->mm;
749 unsigned long addr = rmap_item->address;
750 struct vm_area_struct *vma;
751
752 /*
753 * It is not an accident that whenever we want to break COW
754 * to undo, we also need to drop a reference to the anon_vma.
755 */
756 put_anon_vma(rmap_item->anon_vma);
757
758 mmap_read_lock(mm);
759 vma = find_mergeable_vma(mm, addr);
760 if (vma)
761 break_ksm(vma, addr, false);
762 mmap_read_unlock(mm);
763 }
764
765 static struct page *get_mergeable_page(struct ksm_rmap_item *rmap_item)
766 {
767 struct mm_struct *mm = rmap_item->mm;
768 unsigned long addr = rmap_item->address;
769 struct vm_area_struct *vma;
770 struct page *page;
771
772 mmap_read_lock(mm);
773 vma = find_mergeable_vma(mm, addr);
774 if (!vma)
775 goto out;
776
777 page = follow_page(vma, addr, FOLL_GET);
778 if (IS_ERR_OR_NULL(page))
779 goto out;
780 if (is_zone_device_page(page))
781 goto out_putpage;
782 if (PageAnon(page)) {
783 flush_anon_page(vma, page, addr);
784 flush_dcache_page(page);
785 } else {
786 out_putpage:
787 put_page(page);
788 out:
789 page = NULL;
790 }
791 mmap_read_unlock(mm);
792 return page;
793 }
794
795 /*
796 * This helper is used for getting right index into array of tree roots.
797 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
798 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
799 * every node has its own stable and unstable tree.
800 */
801 static inline int get_kpfn_nid(unsigned long kpfn)
802 {
803 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
804 }
805
806 static struct ksm_stable_node *alloc_stable_node_chain(struct ksm_stable_node *dup,
807 struct rb_root *root)
808 {
809 struct ksm_stable_node *chain = alloc_stable_node();
810 VM_BUG_ON(is_stable_node_chain(dup));
811 if (likely(chain)) {
812 INIT_HLIST_HEAD(&chain->hlist);
813 chain->chain_prune_time = jiffies;
814 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
815 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
816 chain->nid = NUMA_NO_NODE; /* debug */
817 #endif
818 ksm_stable_node_chains++;
819
820 /*
821 * Put the stable node chain in the first dimension of
822 * the stable tree and at the same time remove the old
823 * stable node.
824 */
825 rb_replace_node(&dup->node, &chain->node, root);
826
827 /*
828 * Move the old stable node to the second dimension
829 * queued in the hlist_dup. The invariant is that all
830 * dup stable_nodes in the chain->hlist point to pages
831 * that are write protected and have the exact same
832 * content.
833 */
834 stable_node_chain_add_dup(dup, chain);
835 }
836 return chain;
837 }
838
839 static inline void free_stable_node_chain(struct ksm_stable_node *chain,
840 struct rb_root *root)
841 {
842 rb_erase(&chain->node, root);
843 free_stable_node(chain);
844 ksm_stable_node_chains--;
845 }
846
847 static void remove_node_from_stable_tree(struct ksm_stable_node *stable_node)
848 {
849 struct ksm_rmap_item *rmap_item;
850
851 /* check it's not STABLE_NODE_CHAIN or negative */
852 BUG_ON(stable_node->rmap_hlist_len < 0);
853
854 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
855 if (rmap_item->hlist.next) {
856 ksm_pages_sharing--;
857 trace_ksm_remove_rmap_item(stable_node->kpfn, rmap_item, rmap_item->mm);
858 } else {
859 ksm_pages_shared--;
860 }
861
862 rmap_item->mm->ksm_merging_pages--;
863
864 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
865 stable_node->rmap_hlist_len--;
866 put_anon_vma(rmap_item->anon_vma);
867 rmap_item->address &= PAGE_MASK;
868 cond_resched();
869 }
870
871 /*
872 * We need the second aligned pointer of the migrate_nodes
873 * list_head to stay clear from the rb_parent_color union
874 * (aligned and different than any node) and also different
875 * from &migrate_nodes. This will verify that future list.h changes
876 * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it.
877 */
878 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
879 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
880
881 trace_ksm_remove_ksm_page(stable_node->kpfn);
882 if (stable_node->head == &migrate_nodes)
883 list_del(&stable_node->list);
884 else
885 stable_node_dup_del(stable_node);
886 free_stable_node(stable_node);
887 }
888
889 enum ksm_get_folio_flags {
890 KSM_GET_FOLIO_NOLOCK,
891 KSM_GET_FOLIO_LOCK,
892 KSM_GET_FOLIO_TRYLOCK
893 };
894
895 /*
896 * ksm_get_folio: checks if the page indicated by the stable node
897 * is still its ksm page, despite having held no reference to it.
898 * In which case we can trust the content of the page, and it
899 * returns the gotten page; but if the page has now been zapped,
900 * remove the stale node from the stable tree and return NULL.
901 * But beware, the stable node's page might be being migrated.
902 *
903 * You would expect the stable_node to hold a reference to the ksm page.
904 * But if it increments the page's count, swapping out has to wait for
905 * ksmd to come around again before it can free the page, which may take
906 * seconds or even minutes: much too unresponsive. So instead we use a
907 * "keyhole reference": access to the ksm page from the stable node peeps
908 * out through its keyhole to see if that page still holds the right key,
909 * pointing back to this stable node. This relies on freeing a PageAnon
910 * page to reset its page->mapping to NULL, and relies on no other use of
911 * a page to put something that might look like our key in page->mapping.
912 * is on its way to being freed; but it is an anomaly to bear in mind.
913 */
914 static struct folio *ksm_get_folio(struct ksm_stable_node *stable_node,
915 enum ksm_get_folio_flags flags)
916 {
917 struct folio *folio;
918 void *expected_mapping;
919 unsigned long kpfn;
920
921 expected_mapping = (void *)((unsigned long)stable_node |
922 PAGE_MAPPING_KSM);
923 again:
924 kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
925 folio = pfn_folio(kpfn);
926 if (READ_ONCE(folio->mapping) != expected_mapping)
927 goto stale;
928
929 /*
930 * We cannot do anything with the page while its refcount is 0.
931 * Usually 0 means free, or tail of a higher-order page: in which
932 * case this node is no longer referenced, and should be freed;
933 * however, it might mean that the page is under page_ref_freeze().
934 * The __remove_mapping() case is easy, again the node is now stale;
935 * the same is in reuse_ksm_page() case; but if page is swapcache
936 * in folio_migrate_mapping(), it might still be our page,
937 * in which case it's essential to keep the node.
938 */
939 while (!folio_try_get(folio)) {
940 /*
941 * Another check for page->mapping != expected_mapping would
942 * work here too. We have chosen the !PageSwapCache test to
943 * optimize the common case, when the page is or is about to
944 * be freed: PageSwapCache is cleared (under spin_lock_irq)
945 * in the ref_freeze section of __remove_mapping(); but Anon
946 * folio->mapping reset to NULL later, in free_pages_prepare().
947 */
948 if (!folio_test_swapcache(folio))
949 goto stale;
950 cpu_relax();
951 }
952
953 if (READ_ONCE(folio->mapping) != expected_mapping) {
954 folio_put(folio);
955 goto stale;
956 }
957
958 if (flags == KSM_GET_FOLIO_TRYLOCK) {
959 if (!folio_trylock(folio)) {
960 folio_put(folio);
961 return ERR_PTR(-EBUSY);
962 }
963 } else if (flags == KSM_GET_FOLIO_LOCK)
964 folio_lock(folio);
965
966 if (flags != KSM_GET_FOLIO_NOLOCK) {
967 if (READ_ONCE(folio->mapping) != expected_mapping) {
968 folio_unlock(folio);
969 folio_put(folio);
970 goto stale;
971 }
972 }
973 return folio;
974
975 stale:
976 /*
977 * We come here from above when page->mapping or !PageSwapCache
978 * suggests that the node is stale; but it might be under migration.
979 * We need smp_rmb(), matching the smp_wmb() in folio_migrate_ksm(),
980 * before checking whether node->kpfn has been changed.
981 */
982 smp_rmb();
983 if (READ_ONCE(stable_node->kpfn) != kpfn)
984 goto again;
985 remove_node_from_stable_tree(stable_node);
986 return NULL;
987 }
988
989 /*
990 * Removing rmap_item from stable or unstable tree.
991 * This function will clean the information from the stable/unstable tree.
992 */
993 static void remove_rmap_item_from_tree(struct ksm_rmap_item *rmap_item)
994 {
995 if (rmap_item->address & STABLE_FLAG) {
996 struct ksm_stable_node *stable_node;
997 struct folio *folio;
998
999 stable_node = rmap_item->head;
1000 folio = ksm_get_folio(stable_node, KSM_GET_FOLIO_LOCK);
1001 if (!folio)
1002 goto out;
1003
1004 hlist_del(&rmap_item->hlist);
1005 folio_unlock(folio);
1006 folio_put(folio);
1007
1008 if (!hlist_empty(&stable_node->hlist))
1009 ksm_pages_sharing--;
1010 else
1011 ksm_pages_shared--;
1012
1013 rmap_item->mm->ksm_merging_pages--;
1014
1015 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
1016 stable_node->rmap_hlist_len--;
1017
1018 put_anon_vma(rmap_item->anon_vma);
1019 rmap_item->head = NULL;
1020 rmap_item->address &= PAGE_MASK;
1021
1022 } else if (rmap_item->address & UNSTABLE_FLAG) {
1023 unsigned char age;
1024 /*
1025 * Usually ksmd can and must skip the rb_erase, because
1026 * root_unstable_tree was already reset to RB_ROOT.
1027 * But be careful when an mm is exiting: do the rb_erase
1028 * if this rmap_item was inserted by this scan, rather
1029 * than left over from before.
1030 */
1031 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
1032 BUG_ON(age > 1);
1033 if (!age)
1034 rb_erase(&rmap_item->node,
1035 root_unstable_tree + NUMA(rmap_item->nid));
1036 ksm_pages_unshared--;
1037 rmap_item->address &= PAGE_MASK;
1038 }
1039 out:
1040 cond_resched(); /* we're called from many long loops */
1041 }
1042
1043 static void remove_trailing_rmap_items(struct ksm_rmap_item **rmap_list)
1044 {
1045 while (*rmap_list) {
1046 struct ksm_rmap_item *rmap_item = *rmap_list;
1047 *rmap_list = rmap_item->rmap_list;
1048 remove_rmap_item_from_tree(rmap_item);
1049 free_rmap_item(rmap_item);
1050 }
1051 }
1052
1053 /*
1054 * Though it's very tempting to unmerge rmap_items from stable tree rather
1055 * than check every pte of a given vma, the locking doesn't quite work for
1056 * that - an rmap_item is assigned to the stable tree after inserting ksm
1057 * page and upping mmap_lock. Nor does it fit with the way we skip dup'ing
1058 * rmap_items from parent to child at fork time (so as not to waste time
1059 * if exit comes before the next scan reaches it).
1060 *
1061 * Similarly, although we'd like to remove rmap_items (so updating counts
1062 * and freeing memory) when unmerging an area, it's easier to leave that
1063 * to the next pass of ksmd - consider, for example, how ksmd might be
1064 * in cmp_and_merge_page on one of the rmap_items we would be removing.
1065 */
1066 static int unmerge_ksm_pages(struct vm_area_struct *vma,
1067 unsigned long start, unsigned long end, bool lock_vma)
1068 {
1069 unsigned long addr;
1070 int err = 0;
1071
1072 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
1073 if (ksm_test_exit(vma->vm_mm))
1074 break;
1075 if (signal_pending(current))
1076 err = -ERESTARTSYS;
1077 else
1078 err = break_ksm(vma, addr, lock_vma);
1079 }
1080 return err;
1081 }
1082
1083 static inline struct ksm_stable_node *folio_stable_node(struct folio *folio)
1084 {
1085 return folio_test_ksm(folio) ? folio_raw_mapping(folio) : NULL;
1086 }
1087
1088 static inline struct ksm_stable_node *page_stable_node(struct page *page)
1089 {
1090 return folio_stable_node(page_folio(page));
1091 }
1092
1093 static inline void folio_set_stable_node(struct folio *folio,
1094 struct ksm_stable_node *stable_node)
1095 {
1096 VM_WARN_ON_FOLIO(folio_test_anon(folio) && PageAnonExclusive(&folio->page), folio);
1097 folio->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
1098 }
1099
1100 #ifdef CONFIG_SYSFS
1101 /*
1102 * Only called through the sysfs control interface:
1103 */
1104 static int remove_stable_node(struct ksm_stable_node *stable_node)
1105 {
1106 struct folio *folio;
1107 int err;
1108
1109 folio = ksm_get_folio(stable_node, KSM_GET_FOLIO_LOCK);
1110 if (!folio) {
1111 /*
1112 * ksm_get_folio did remove_node_from_stable_tree itself.
1113 */
1114 return 0;
1115 }
1116
1117 /*
1118 * Page could be still mapped if this races with __mmput() running in
1119 * between ksm_exit() and exit_mmap(). Just refuse to let
1120 * merge_across_nodes/max_page_sharing be switched.
1121 */
1122 err = -EBUSY;
1123 if (!folio_mapped(folio)) {
1124 /*
1125 * The stable node did not yet appear stale to ksm_get_folio(),
1126 * since that allows for an unmapped ksm folio to be recognized
1127 * right up until it is freed; but the node is safe to remove.
1128 * This folio might be in an LRU cache waiting to be freed,
1129 * or it might be in the swapcache (perhaps under writeback),
1130 * or it might have been removed from swapcache a moment ago.
1131 */
1132 folio_set_stable_node(folio, NULL);
1133 remove_node_from_stable_tree(stable_node);
1134 err = 0;
1135 }
1136
1137 folio_unlock(folio);
1138 folio_put(folio);
1139 return err;
1140 }
1141
1142 static int remove_stable_node_chain(struct ksm_stable_node *stable_node,
1143 struct rb_root *root)
1144 {
1145 struct ksm_stable_node *dup;
1146 struct hlist_node *hlist_safe;
1147
1148 if (!is_stable_node_chain(stable_node)) {
1149 VM_BUG_ON(is_stable_node_dup(stable_node));
1150 if (remove_stable_node(stable_node))
1151 return true;
1152 else
1153 return false;
1154 }
1155
1156 hlist_for_each_entry_safe(dup, hlist_safe,
1157 &stable_node->hlist, hlist_dup) {
1158 VM_BUG_ON(!is_stable_node_dup(dup));
1159 if (remove_stable_node(dup))
1160 return true;
1161 }
1162 BUG_ON(!hlist_empty(&stable_node->hlist));
1163 free_stable_node_chain(stable_node, root);
1164 return false;
1165 }
1166
1167 static int remove_all_stable_nodes(void)
1168 {
1169 struct ksm_stable_node *stable_node, *next;
1170 int nid;
1171 int err = 0;
1172
1173 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
1174 while (root_stable_tree[nid].rb_node) {
1175 stable_node = rb_entry(root_stable_tree[nid].rb_node,
1176 struct ksm_stable_node, node);
1177 if (remove_stable_node_chain(stable_node,
1178 root_stable_tree + nid)) {
1179 err = -EBUSY;
1180 break; /* proceed to next nid */
1181 }
1182 cond_resched();
1183 }
1184 }
1185 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
1186 if (remove_stable_node(stable_node))
1187 err = -EBUSY;
1188 cond_resched();
1189 }
1190 return err;
1191 }
1192
1193 static int unmerge_and_remove_all_rmap_items(void)
1194 {
1195 struct ksm_mm_slot *mm_slot;
1196 struct mm_slot *slot;
1197 struct mm_struct *mm;
1198 struct vm_area_struct *vma;
1199 int err = 0;
1200
1201 spin_lock(&ksm_mmlist_lock);
1202 slot = list_entry(ksm_mm_head.slot.mm_node.next,
1203 struct mm_slot, mm_node);
1204 ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
1205 spin_unlock(&ksm_mmlist_lock);
1206
1207 for (mm_slot = ksm_scan.mm_slot; mm_slot != &ksm_mm_head;
1208 mm_slot = ksm_scan.mm_slot) {
1209 VMA_ITERATOR(vmi, mm_slot->slot.mm, 0);
1210
1211 mm = mm_slot->slot.mm;
1212 mmap_read_lock(mm);
1213
1214 /*
1215 * Exit right away if mm is exiting to avoid lockdep issue in
1216 * the maple tree
1217 */
1218 if (ksm_test_exit(mm))
1219 goto mm_exiting;
1220
1221 for_each_vma(vmi, vma) {
1222 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
1223 continue;
1224 err = unmerge_ksm_pages(vma,
1225 vma->vm_start, vma->vm_end, false);
1226 if (err)
1227 goto error;
1228 }
1229
1230 mm_exiting:
1231 remove_trailing_rmap_items(&mm_slot->rmap_list);
1232 mmap_read_unlock(mm);
1233
1234 spin_lock(&ksm_mmlist_lock);
1235 slot = list_entry(mm_slot->slot.mm_node.next,
1236 struct mm_slot, mm_node);
1237 ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
1238 if (ksm_test_exit(mm)) {
1239 hash_del(&mm_slot->slot.hash);
1240 list_del(&mm_slot->slot.mm_node);
1241 spin_unlock(&ksm_mmlist_lock);
1242
1243 mm_slot_free(mm_slot_cache, mm_slot);
1244 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1245 clear_bit(MMF_VM_MERGE_ANY, &mm->flags);
1246 mmdrop(mm);
1247 } else
1248 spin_unlock(&ksm_mmlist_lock);
1249 }
1250
1251 /* Clean up stable nodes, but don't worry if some are still busy */
1252 remove_all_stable_nodes();
1253 ksm_scan.seqnr = 0;
1254 return 0;
1255
1256 error:
1257 mmap_read_unlock(mm);
1258 spin_lock(&ksm_mmlist_lock);
1259 ksm_scan.mm_slot = &ksm_mm_head;
1260 spin_unlock(&ksm_mmlist_lock);
1261 return err;
1262 }
1263 #endif /* CONFIG_SYSFS */
1264
1265 static u32 calc_checksum(struct page *page)
1266 {
1267 u32 checksum;
1268 void *addr = kmap_local_page(page);
1269 checksum = xxhash(addr, PAGE_SIZE, 0);
1270 kunmap_local(addr);
1271 return checksum;
1272 }
1273
1274 static int write_protect_page(struct vm_area_struct *vma, struct folio *folio,
1275 pte_t *orig_pte)
1276 {
1277 struct mm_struct *mm = vma->vm_mm;
1278 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, 0, 0);
1279 int swapped;
1280 int err = -EFAULT;
1281 struct mmu_notifier_range range;
1282 bool anon_exclusive;
1283 pte_t entry;
1284
1285 if (WARN_ON_ONCE(folio_test_large(folio)))
1286 return err;
1287
1288 pvmw.address = page_address_in_vma(&folio->page, vma);
1289 if (pvmw.address == -EFAULT)
1290 goto out;
1291
1292 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, pvmw.address,
1293 pvmw.address + PAGE_SIZE);
1294 mmu_notifier_invalidate_range_start(&range);
1295
1296 if (!page_vma_mapped_walk(&pvmw))
1297 goto out_mn;
1298 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1299 goto out_unlock;
1300
1301 anon_exclusive = PageAnonExclusive(&folio->page);
1302 entry = ptep_get(pvmw.pte);
1303 if (pte_write(entry) || pte_dirty(entry) ||
1304 anon_exclusive || mm_tlb_flush_pending(mm)) {
1305 swapped = folio_test_swapcache(folio);
1306 flush_cache_page(vma, pvmw.address, folio_pfn(folio));
1307 /*
1308 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1309 * take any lock, therefore the check that we are going to make
1310 * with the pagecount against the mapcount is racy and
1311 * O_DIRECT can happen right after the check.
1312 * So we clear the pte and flush the tlb before the check
1313 * this assure us that no O_DIRECT can happen after the check
1314 * or in the middle of the check.
1315 *
1316 * No need to notify as we are downgrading page table to read
1317 * only not changing it to point to a new page.
1318 *
1319 * See Documentation/mm/mmu_notifier.rst
1320 */
1321 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1322 /*
1323 * Check that no O_DIRECT or similar I/O is in progress on the
1324 * page
1325 */
1326 if (folio_mapcount(folio) + 1 + swapped != folio_ref_count(folio)) {
1327 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1328 goto out_unlock;
1329 }
1330
1331 /* See folio_try_share_anon_rmap_pte(): clear PTE first. */
1332 if (anon_exclusive &&
1333 folio_try_share_anon_rmap_pte(folio, &folio->page)) {
1334 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1335 goto out_unlock;
1336 }
1337
1338 if (pte_dirty(entry))
1339 folio_mark_dirty(folio);
1340 entry = pte_mkclean(entry);
1341
1342 if (pte_write(entry))
1343 entry = pte_wrprotect(entry);
1344
1345 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1346 }
1347 *orig_pte = entry;
1348 err = 0;
1349
1350 out_unlock:
1351 page_vma_mapped_walk_done(&pvmw);
1352 out_mn:
1353 mmu_notifier_invalidate_range_end(&range);
1354 out:
1355 return err;
1356 }
1357
1358 /**
1359 * replace_page - replace page in vma by new ksm page
1360 * @vma: vma that holds the pte pointing to page
1361 * @page: the page we are replacing by kpage
1362 * @kpage: the ksm page we replace page by
1363 * @orig_pte: the original value of the pte
1364 *
1365 * Returns 0 on success, -EFAULT on failure.
1366 */
1367 static int replace_page(struct vm_area_struct *vma, struct page *page,
1368 struct page *kpage, pte_t orig_pte)
1369 {
1370 struct folio *kfolio = page_folio(kpage);
1371 struct mm_struct *mm = vma->vm_mm;
1372 struct folio *folio;
1373 pmd_t *pmd;
1374 pmd_t pmde;
1375 pte_t *ptep;
1376 pte_t newpte;
1377 spinlock_t *ptl;
1378 unsigned long addr;
1379 int err = -EFAULT;
1380 struct mmu_notifier_range range;
1381
1382 addr = page_address_in_vma(page, vma);
1383 if (addr == -EFAULT)
1384 goto out;
1385
1386 pmd = mm_find_pmd(mm, addr);
1387 if (!pmd)
1388 goto out;
1389 /*
1390 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
1391 * without holding anon_vma lock for write. So when looking for a
1392 * genuine pmde (in which to find pte), test present and !THP together.
1393 */
1394 pmde = pmdp_get_lockless(pmd);
1395 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
1396 goto out;
1397
1398 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, addr,
1399 addr + PAGE_SIZE);
1400 mmu_notifier_invalidate_range_start(&range);
1401
1402 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1403 if (!ptep)
1404 goto out_mn;
1405 if (!pte_same(ptep_get(ptep), orig_pte)) {
1406 pte_unmap_unlock(ptep, ptl);
1407 goto out_mn;
1408 }
1409 VM_BUG_ON_PAGE(PageAnonExclusive(page), page);
1410 VM_BUG_ON_FOLIO(folio_test_anon(kfolio) && PageAnonExclusive(kpage),
1411 kfolio);
1412
1413 /*
1414 * No need to check ksm_use_zero_pages here: we can only have a
1415 * zero_page here if ksm_use_zero_pages was enabled already.
1416 */
1417 if (!is_zero_pfn(page_to_pfn(kpage))) {
1418 folio_get(kfolio);
1419 folio_add_anon_rmap_pte(kfolio, kpage, vma, addr, RMAP_NONE);
1420 newpte = mk_pte(kpage, vma->vm_page_prot);
1421 } else {
1422 /*
1423 * Use pte_mkdirty to mark the zero page mapped by KSM, and then
1424 * we can easily track all KSM-placed zero pages by checking if
1425 * the dirty bit in zero page's PTE is set.
1426 */
1427 newpte = pte_mkdirty(pte_mkspecial(pfn_pte(page_to_pfn(kpage), vma->vm_page_prot)));
1428 ksm_map_zero_page(mm);
1429 /*
1430 * We're replacing an anonymous page with a zero page, which is
1431 * not anonymous. We need to do proper accounting otherwise we
1432 * will get wrong values in /proc, and a BUG message in dmesg
1433 * when tearing down the mm.
1434 */
1435 dec_mm_counter(mm, MM_ANONPAGES);
1436 }
1437
1438 flush_cache_page(vma, addr, pte_pfn(ptep_get(ptep)));
1439 /*
1440 * No need to notify as we are replacing a read only page with another
1441 * read only page with the same content.
1442 *
1443 * See Documentation/mm/mmu_notifier.rst
1444 */
1445 ptep_clear_flush(vma, addr, ptep);
1446 set_pte_at(mm, addr, ptep, newpte);
1447
1448 folio = page_folio(page);
1449 folio_remove_rmap_pte(folio, page, vma);
1450 if (!folio_mapped(folio))
1451 folio_free_swap(folio);
1452 folio_put(folio);
1453
1454 pte_unmap_unlock(ptep, ptl);
1455 err = 0;
1456 out_mn:
1457 mmu_notifier_invalidate_range_end(&range);
1458 out:
1459 return err;
1460 }
1461
1462 /*
1463 * try_to_merge_one_page - take two pages and merge them into one
1464 * @vma: the vma that holds the pte pointing to page
1465 * @page: the PageAnon page that we want to replace with kpage
1466 * @kpage: the PageKsm page that we want to map instead of page,
1467 * or NULL the first time when we want to use page as kpage.
1468 *
1469 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1470 */
1471 static int try_to_merge_one_page(struct vm_area_struct *vma,
1472 struct page *page, struct page *kpage)
1473 {
1474 pte_t orig_pte = __pte(0);
1475 int err = -EFAULT;
1476
1477 if (page == kpage) /* ksm page forked */
1478 return 0;
1479
1480 if (!PageAnon(page))
1481 goto out;
1482
1483 /*
1484 * We need the page lock to read a stable PageSwapCache in
1485 * write_protect_page(). We use trylock_page() instead of
1486 * lock_page() because we don't want to wait here - we
1487 * prefer to continue scanning and merging different pages,
1488 * then come back to this page when it is unlocked.
1489 */
1490 if (!trylock_page(page))
1491 goto out;
1492
1493 if (PageTransCompound(page)) {
1494 if (split_huge_page(page))
1495 goto out_unlock;
1496 }
1497
1498 /*
1499 * If this anonymous page is mapped only here, its pte may need
1500 * to be write-protected. If it's mapped elsewhere, all of its
1501 * ptes are necessarily already write-protected. But in either
1502 * case, we need to lock and check page_count is not raised.
1503 */
1504 if (write_protect_page(vma, page_folio(page), &orig_pte) == 0) {
1505 if (!kpage) {
1506 /*
1507 * While we hold page lock, upgrade page from
1508 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1509 * stable_tree_insert() will update stable_node.
1510 */
1511 folio_set_stable_node(page_folio(page), NULL);
1512 mark_page_accessed(page);
1513 /*
1514 * Page reclaim just frees a clean page with no dirty
1515 * ptes: make sure that the ksm page would be swapped.
1516 */
1517 if (!PageDirty(page))
1518 SetPageDirty(page);
1519 err = 0;
1520 } else if (pages_identical(page, kpage))
1521 err = replace_page(vma, page, kpage, orig_pte);
1522 }
1523
1524 out_unlock:
1525 unlock_page(page);
1526 out:
1527 return err;
1528 }
1529
1530 /*
1531 * This function returns 0 if the pages were merged or if they are
1532 * no longer merging candidates (e.g., VMA stale), -EFAULT otherwise.
1533 */
1534 static int try_to_merge_with_zero_page(struct ksm_rmap_item *rmap_item,
1535 struct page *page)
1536 {
1537 struct mm_struct *mm = rmap_item->mm;
1538 int err = -EFAULT;
1539
1540 /*
1541 * Same checksum as an empty page. We attempt to merge it with the
1542 * appropriate zero page if the user enabled this via sysfs.
1543 */
1544 if (ksm_use_zero_pages && (rmap_item->oldchecksum == zero_checksum)) {
1545 struct vm_area_struct *vma;
1546
1547 mmap_read_lock(mm);
1548 vma = find_mergeable_vma(mm, rmap_item->address);
1549 if (vma) {
1550 err = try_to_merge_one_page(vma, page,
1551 ZERO_PAGE(rmap_item->address));
1552 trace_ksm_merge_one_page(
1553 page_to_pfn(ZERO_PAGE(rmap_item->address)),
1554 rmap_item, mm, err);
1555 } else {
1556 /*
1557 * If the vma is out of date, we do not need to
1558 * continue.
1559 */
1560 err = 0;
1561 }
1562 mmap_read_unlock(mm);
1563 }
1564
1565 return err;
1566 }
1567
1568 /*
1569 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1570 * but no new kernel page is allocated: kpage must already be a ksm page.
1571 *
1572 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1573 */
1574 static int try_to_merge_with_ksm_page(struct ksm_rmap_item *rmap_item,
1575 struct page *page, struct page *kpage)
1576 {
1577 struct mm_struct *mm = rmap_item->mm;
1578 struct vm_area_struct *vma;
1579 int err = -EFAULT;
1580
1581 mmap_read_lock(mm);
1582 vma = find_mergeable_vma(mm, rmap_item->address);
1583 if (!vma)
1584 goto out;
1585
1586 err = try_to_merge_one_page(vma, page, kpage);
1587 if (err)
1588 goto out;
1589
1590 /* Unstable nid is in union with stable anon_vma: remove first */
1591 remove_rmap_item_from_tree(rmap_item);
1592
1593 /* Must get reference to anon_vma while still holding mmap_lock */
1594 rmap_item->anon_vma = vma->anon_vma;
1595 get_anon_vma(vma->anon_vma);
1596 out:
1597 mmap_read_unlock(mm);
1598 trace_ksm_merge_with_ksm_page(kpage, page_to_pfn(kpage ? kpage : page),
1599 rmap_item, mm, err);
1600 return err;
1601 }
1602
1603 /*
1604 * try_to_merge_two_pages - take two identical pages and prepare them
1605 * to be merged into one page.
1606 *
1607 * This function returns the kpage if we successfully merged two identical
1608 * pages into one ksm page, NULL otherwise.
1609 *
1610 * Note that this function upgrades page to ksm page: if one of the pages
1611 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1612 */
1613 static struct page *try_to_merge_two_pages(struct ksm_rmap_item *rmap_item,
1614 struct page *page,
1615 struct ksm_rmap_item *tree_rmap_item,
1616 struct page *tree_page)
1617 {
1618 int err;
1619
1620 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1621 if (!err) {
1622 err = try_to_merge_with_ksm_page(tree_rmap_item,
1623 tree_page, page);
1624 /*
1625 * If that fails, we have a ksm page with only one pte
1626 * pointing to it: so break it.
1627 */
1628 if (err)
1629 break_cow(rmap_item);
1630 }
1631 return err ? NULL : page;
1632 }
1633
1634 static __always_inline
1635 bool __is_page_sharing_candidate(struct ksm_stable_node *stable_node, int offset)
1636 {
1637 VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1638 /*
1639 * Check that at least one mapping still exists, otherwise
1640 * there's no much point to merge and share with this
1641 * stable_node, as the underlying tree_page of the other
1642 * sharer is going to be freed soon.
1643 */
1644 return stable_node->rmap_hlist_len &&
1645 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1646 }
1647
1648 static __always_inline
1649 bool is_page_sharing_candidate(struct ksm_stable_node *stable_node)
1650 {
1651 return __is_page_sharing_candidate(stable_node, 0);
1652 }
1653
1654 static struct folio *stable_node_dup(struct ksm_stable_node **_stable_node_dup,
1655 struct ksm_stable_node **_stable_node,
1656 struct rb_root *root,
1657 bool prune_stale_stable_nodes)
1658 {
1659 struct ksm_stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1660 struct hlist_node *hlist_safe;
1661 struct folio *folio, *tree_folio = NULL;
1662 int found_rmap_hlist_len;
1663
1664 if (!prune_stale_stable_nodes ||
1665 time_before(jiffies, stable_node->chain_prune_time +
1666 msecs_to_jiffies(
1667 ksm_stable_node_chains_prune_millisecs)))
1668 prune_stale_stable_nodes = false;
1669 else
1670 stable_node->chain_prune_time = jiffies;
1671
1672 hlist_for_each_entry_safe(dup, hlist_safe,
1673 &stable_node->hlist, hlist_dup) {
1674 cond_resched();
1675 /*
1676 * We must walk all stable_node_dup to prune the stale
1677 * stable nodes during lookup.
1678 *
1679 * ksm_get_folio can drop the nodes from the
1680 * stable_node->hlist if they point to freed pages
1681 * (that's why we do a _safe walk). The "dup"
1682 * stable_node parameter itself will be freed from
1683 * under us if it returns NULL.
1684 */
1685 folio = ksm_get_folio(dup, KSM_GET_FOLIO_NOLOCK);
1686 if (!folio)
1687 continue;
1688 /* Pick the best candidate if possible. */
1689 if (!found || (is_page_sharing_candidate(dup) &&
1690 (!is_page_sharing_candidate(found) ||
1691 dup->rmap_hlist_len > found_rmap_hlist_len))) {
1692 if (found)
1693 folio_put(tree_folio);
1694 found = dup;
1695 found_rmap_hlist_len = found->rmap_hlist_len;
1696 tree_folio = folio;
1697 /* skip put_page for found candidate */
1698 if (!prune_stale_stable_nodes &&
1699 is_page_sharing_candidate(found))
1700 break;
1701 continue;
1702 }
1703 folio_put(folio);
1704 }
1705
1706 if (found) {
1707 if (hlist_is_singular_node(&found->hlist_dup, &stable_node->hlist)) {
1708 /*
1709 * If there's not just one entry it would
1710 * corrupt memory, better BUG_ON. In KSM
1711 * context with no lock held it's not even
1712 * fatal.
1713 */
1714 BUG_ON(stable_node->hlist.first->next);
1715
1716 /*
1717 * There's just one entry and it is below the
1718 * deduplication limit so drop the chain.
1719 */
1720 rb_replace_node(&stable_node->node, &found->node,
1721 root);
1722 free_stable_node(stable_node);
1723 ksm_stable_node_chains--;
1724 ksm_stable_node_dups--;
1725 /*
1726 * NOTE: the caller depends on the stable_node
1727 * to be equal to stable_node_dup if the chain
1728 * was collapsed.
1729 */
1730 *_stable_node = found;
1731 /*
1732 * Just for robustness, as stable_node is
1733 * otherwise left as a stable pointer, the
1734 * compiler shall optimize it away at build
1735 * time.
1736 */
1737 stable_node = NULL;
1738 } else if (stable_node->hlist.first != &found->hlist_dup &&
1739 __is_page_sharing_candidate(found, 1)) {
1740 /*
1741 * If the found stable_node dup can accept one
1742 * more future merge (in addition to the one
1743 * that is underway) and is not at the head of
1744 * the chain, put it there so next search will
1745 * be quicker in the !prune_stale_stable_nodes
1746 * case.
1747 *
1748 * NOTE: it would be inaccurate to use nr > 1
1749 * instead of checking the hlist.first pointer
1750 * directly, because in the
1751 * prune_stale_stable_nodes case "nr" isn't
1752 * the position of the found dup in the chain,
1753 * but the total number of dups in the chain.
1754 */
1755 hlist_del(&found->hlist_dup);
1756 hlist_add_head(&found->hlist_dup,
1757 &stable_node->hlist);
1758 }
1759 } else {
1760 /* Its hlist must be empty if no one found. */
1761 free_stable_node_chain(stable_node, root);
1762 }
1763
1764 *_stable_node_dup = found;
1765 return tree_folio;
1766 }
1767
1768 /*
1769 * Like for ksm_get_folio, this function can free the *_stable_node and
1770 * *_stable_node_dup if the returned tree_page is NULL.
1771 *
1772 * It can also free and overwrite *_stable_node with the found
1773 * stable_node_dup if the chain is collapsed (in which case
1774 * *_stable_node will be equal to *_stable_node_dup like if the chain
1775 * never existed). It's up to the caller to verify tree_page is not
1776 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1777 *
1778 * *_stable_node_dup is really a second output parameter of this
1779 * function and will be overwritten in all cases, the caller doesn't
1780 * need to initialize it.
1781 */
1782 static struct folio *__stable_node_chain(struct ksm_stable_node **_stable_node_dup,
1783 struct ksm_stable_node **_stable_node,
1784 struct rb_root *root,
1785 bool prune_stale_stable_nodes)
1786 {
1787 struct ksm_stable_node *stable_node = *_stable_node;
1788
1789 if (!is_stable_node_chain(stable_node)) {
1790 *_stable_node_dup = stable_node;
1791 return ksm_get_folio(stable_node, KSM_GET_FOLIO_NOLOCK);
1792 }
1793 return stable_node_dup(_stable_node_dup, _stable_node, root,
1794 prune_stale_stable_nodes);
1795 }
1796
1797 static __always_inline struct folio *chain_prune(struct ksm_stable_node **s_n_d,
1798 struct ksm_stable_node **s_n,
1799 struct rb_root *root)
1800 {
1801 return __stable_node_chain(s_n_d, s_n, root, true);
1802 }
1803
1804 static __always_inline struct folio *chain(struct ksm_stable_node **s_n_d,
1805 struct ksm_stable_node **s_n,
1806 struct rb_root *root)
1807 {
1808 return __stable_node_chain(s_n_d, s_n, root, false);
1809 }
1810
1811 /*
1812 * stable_tree_search - search for page inside the stable tree
1813 *
1814 * This function checks if there is a page inside the stable tree
1815 * with identical content to the page that we are scanning right now.
1816 *
1817 * This function returns the stable tree node of identical content if found,
1818 * NULL otherwise.
1819 */
1820 static struct page *stable_tree_search(struct page *page)
1821 {
1822 int nid;
1823 struct rb_root *root;
1824 struct rb_node **new;
1825 struct rb_node *parent;
1826 struct ksm_stable_node *stable_node, *stable_node_dup;
1827 struct ksm_stable_node *page_node;
1828 struct folio *folio;
1829
1830 folio = page_folio(page);
1831 page_node = folio_stable_node(folio);
1832 if (page_node && page_node->head != &migrate_nodes) {
1833 /* ksm page forked */
1834 folio_get(folio);
1835 return &folio->page;
1836 }
1837
1838 nid = get_kpfn_nid(folio_pfn(folio));
1839 root = root_stable_tree + nid;
1840 again:
1841 new = &root->rb_node;
1842 parent = NULL;
1843
1844 while (*new) {
1845 struct folio *tree_folio;
1846 int ret;
1847
1848 cond_resched();
1849 stable_node = rb_entry(*new, struct ksm_stable_node, node);
1850 tree_folio = chain_prune(&stable_node_dup, &stable_node, root);
1851 if (!tree_folio) {
1852 /*
1853 * If we walked over a stale stable_node,
1854 * ksm_get_folio() will call rb_erase() and it
1855 * may rebalance the tree from under us. So
1856 * restart the search from scratch. Returning
1857 * NULL would be safe too, but we'd generate
1858 * false negative insertions just because some
1859 * stable_node was stale.
1860 */
1861 goto again;
1862 }
1863
1864 ret = memcmp_pages(page, &tree_folio->page);
1865 folio_put(tree_folio);
1866
1867 parent = *new;
1868 if (ret < 0)
1869 new = &parent->rb_left;
1870 else if (ret > 0)
1871 new = &parent->rb_right;
1872 else {
1873 if (page_node) {
1874 VM_BUG_ON(page_node->head != &migrate_nodes);
1875 /*
1876 * If the mapcount of our migrated KSM folio is
1877 * at most 1, we can merge it with another
1878 * KSM folio where we know that we have space
1879 * for one more mapping without exceeding the
1880 * ksm_max_page_sharing limit: see
1881 * chain_prune(). This way, we can avoid adding
1882 * this stable node to the chain.
1883 */
1884 if (folio_mapcount(folio) > 1)
1885 goto chain_append;
1886 }
1887
1888 if (!is_page_sharing_candidate(stable_node_dup)) {
1889 /*
1890 * If the stable_node is a chain and
1891 * we got a payload match in memcmp
1892 * but we cannot merge the scanned
1893 * page in any of the existing
1894 * stable_node dups because they're
1895 * all full, we need to wait the
1896 * scanned page to find itself a match
1897 * in the unstable tree to create a
1898 * brand new KSM page to add later to
1899 * the dups of this stable_node.
1900 */
1901 return NULL;
1902 }
1903
1904 /*
1905 * Lock and unlock the stable_node's page (which
1906 * might already have been migrated) so that page
1907 * migration is sure to notice its raised count.
1908 * It would be more elegant to return stable_node
1909 * than kpage, but that involves more changes.
1910 */
1911 tree_folio = ksm_get_folio(stable_node_dup,
1912 KSM_GET_FOLIO_TRYLOCK);
1913
1914 if (PTR_ERR(tree_folio) == -EBUSY)
1915 return ERR_PTR(-EBUSY);
1916
1917 if (unlikely(!tree_folio))
1918 /*
1919 * The tree may have been rebalanced,
1920 * so re-evaluate parent and new.
1921 */
1922 goto again;
1923 folio_unlock(tree_folio);
1924
1925 if (get_kpfn_nid(stable_node_dup->kpfn) !=
1926 NUMA(stable_node_dup->nid)) {
1927 folio_put(tree_folio);
1928 goto replace;
1929 }
1930 return &tree_folio->page;
1931 }
1932 }
1933
1934 if (!page_node)
1935 return NULL;
1936
1937 list_del(&page_node->list);
1938 DO_NUMA(page_node->nid = nid);
1939 rb_link_node(&page_node->node, parent, new);
1940 rb_insert_color(&page_node->node, root);
1941 out:
1942 if (is_page_sharing_candidate(page_node)) {
1943 folio_get(folio);
1944 return &folio->page;
1945 } else
1946 return NULL;
1947
1948 replace:
1949 /*
1950 * If stable_node was a chain and chain_prune collapsed it,
1951 * stable_node has been updated to be the new regular
1952 * stable_node. A collapse of the chain is indistinguishable
1953 * from the case there was no chain in the stable
1954 * rbtree. Otherwise stable_node is the chain and
1955 * stable_node_dup is the dup to replace.
1956 */
1957 if (stable_node_dup == stable_node) {
1958 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1959 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1960 /* there is no chain */
1961 if (page_node) {
1962 VM_BUG_ON(page_node->head != &migrate_nodes);
1963 list_del(&page_node->list);
1964 DO_NUMA(page_node->nid = nid);
1965 rb_replace_node(&stable_node_dup->node,
1966 &page_node->node,
1967 root);
1968 if (is_page_sharing_candidate(page_node))
1969 folio_get(folio);
1970 else
1971 folio = NULL;
1972 } else {
1973 rb_erase(&stable_node_dup->node, root);
1974 folio = NULL;
1975 }
1976 } else {
1977 VM_BUG_ON(!is_stable_node_chain(stable_node));
1978 __stable_node_dup_del(stable_node_dup);
1979 if (page_node) {
1980 VM_BUG_ON(page_node->head != &migrate_nodes);
1981 list_del(&page_node->list);
1982 DO_NUMA(page_node->nid = nid);
1983 stable_node_chain_add_dup(page_node, stable_node);
1984 if (is_page_sharing_candidate(page_node))
1985 folio_get(folio);
1986 else
1987 folio = NULL;
1988 } else {
1989 folio = NULL;
1990 }
1991 }
1992 stable_node_dup->head = &migrate_nodes;
1993 list_add(&stable_node_dup->list, stable_node_dup->head);
1994 return &folio->page;
1995
1996 chain_append:
1997 /*
1998 * If stable_node was a chain and chain_prune collapsed it,
1999 * stable_node has been updated to be the new regular
2000 * stable_node. A collapse of the chain is indistinguishable
2001 * from the case there was no chain in the stable
2002 * rbtree. Otherwise stable_node is the chain and
2003 * stable_node_dup is the dup to replace.
2004 */
2005 if (stable_node_dup == stable_node) {
2006 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
2007 /* chain is missing so create it */
2008 stable_node = alloc_stable_node_chain(stable_node_dup,
2009 root);
2010 if (!stable_node)
2011 return NULL;
2012 }
2013 /*
2014 * Add this stable_node dup that was
2015 * migrated to the stable_node chain
2016 * of the current nid for this page
2017 * content.
2018 */
2019 VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
2020 VM_BUG_ON(page_node->head != &migrate_nodes);
2021 list_del(&page_node->list);
2022 DO_NUMA(page_node->nid = nid);
2023 stable_node_chain_add_dup(page_node, stable_node);
2024 goto out;
2025 }
2026
2027 /*
2028 * stable_tree_insert - insert stable tree node pointing to new ksm page
2029 * into the stable tree.
2030 *
2031 * This function returns the stable tree node just allocated on success,
2032 * NULL otherwise.
2033 */
2034 static struct ksm_stable_node *stable_tree_insert(struct folio *kfolio)
2035 {
2036 int nid;
2037 unsigned long kpfn;
2038 struct rb_root *root;
2039 struct rb_node **new;
2040 struct rb_node *parent;
2041 struct ksm_stable_node *stable_node, *stable_node_dup;
2042 bool need_chain = false;
2043
2044 kpfn = folio_pfn(kfolio);
2045 nid = get_kpfn_nid(kpfn);
2046 root = root_stable_tree + nid;
2047 again:
2048 parent = NULL;
2049 new = &root->rb_node;
2050
2051 while (*new) {
2052 struct folio *tree_folio;
2053 int ret;
2054
2055 cond_resched();
2056 stable_node = rb_entry(*new, struct ksm_stable_node, node);
2057 tree_folio = chain(&stable_node_dup, &stable_node, root);
2058 if (!tree_folio) {
2059 /*
2060 * If we walked over a stale stable_node,
2061 * ksm_get_folio() will call rb_erase() and it
2062 * may rebalance the tree from under us. So
2063 * restart the search from scratch. Returning
2064 * NULL would be safe too, but we'd generate
2065 * false negative insertions just because some
2066 * stable_node was stale.
2067 */
2068 goto again;
2069 }
2070
2071 ret = memcmp_pages(&kfolio->page, &tree_folio->page);
2072 folio_put(tree_folio);
2073
2074 parent = *new;
2075 if (ret < 0)
2076 new = &parent->rb_left;
2077 else if (ret > 0)
2078 new = &parent->rb_right;
2079 else {
2080 need_chain = true;
2081 break;
2082 }
2083 }
2084
2085 stable_node_dup = alloc_stable_node();
2086 if (!stable_node_dup)
2087 return NULL;
2088
2089 INIT_HLIST_HEAD(&stable_node_dup->hlist);
2090 stable_node_dup->kpfn = kpfn;
2091 stable_node_dup->rmap_hlist_len = 0;
2092 DO_NUMA(stable_node_dup->nid = nid);
2093 if (!need_chain) {
2094 rb_link_node(&stable_node_dup->node, parent, new);
2095 rb_insert_color(&stable_node_dup->node, root);
2096 } else {
2097 if (!is_stable_node_chain(stable_node)) {
2098 struct ksm_stable_node *orig = stable_node;
2099 /* chain is missing so create it */
2100 stable_node = alloc_stable_node_chain(orig, root);
2101 if (!stable_node) {
2102 free_stable_node(stable_node_dup);
2103 return NULL;
2104 }
2105 }
2106 stable_node_chain_add_dup(stable_node_dup, stable_node);
2107 }
2108
2109 folio_set_stable_node(kfolio, stable_node_dup);
2110
2111 return stable_node_dup;
2112 }
2113
2114 /*
2115 * unstable_tree_search_insert - search for identical page,
2116 * else insert rmap_item into the unstable tree.
2117 *
2118 * This function searches for a page in the unstable tree identical to the
2119 * page currently being scanned; and if no identical page is found in the
2120 * tree, we insert rmap_item as a new object into the unstable tree.
2121 *
2122 * This function returns pointer to rmap_item found to be identical
2123 * to the currently scanned page, NULL otherwise.
2124 *
2125 * This function does both searching and inserting, because they share
2126 * the same walking algorithm in an rbtree.
2127 */
2128 static
2129 struct ksm_rmap_item *unstable_tree_search_insert(struct ksm_rmap_item *rmap_item,
2130 struct page *page,
2131 struct page **tree_pagep)
2132 {
2133 struct rb_node **new;
2134 struct rb_root *root;
2135 struct rb_node *parent = NULL;
2136 int nid;
2137
2138 nid = get_kpfn_nid(page_to_pfn(page));
2139 root = root_unstable_tree + nid;
2140 new = &root->rb_node;
2141
2142 while (*new) {
2143 struct ksm_rmap_item *tree_rmap_item;
2144 struct page *tree_page;
2145 int ret;
2146
2147 cond_resched();
2148 tree_rmap_item = rb_entry(*new, struct ksm_rmap_item, node);
2149 tree_page = get_mergeable_page(tree_rmap_item);
2150 if (!tree_page)
2151 return NULL;
2152
2153 /*
2154 * Don't substitute a ksm page for a forked page.
2155 */
2156 if (page == tree_page) {
2157 put_page(tree_page);
2158 return NULL;
2159 }
2160
2161 ret = memcmp_pages(page, tree_page);
2162
2163 parent = *new;
2164 if (ret < 0) {
2165 put_page(tree_page);
2166 new = &parent->rb_left;
2167 } else if (ret > 0) {
2168 put_page(tree_page);
2169 new = &parent->rb_right;
2170 } else if (!ksm_merge_across_nodes &&
2171 page_to_nid(tree_page) != nid) {
2172 /*
2173 * If tree_page has been migrated to another NUMA node,
2174 * it will be flushed out and put in the right unstable
2175 * tree next time: only merge with it when across_nodes.
2176 */
2177 put_page(tree_page);
2178 return NULL;
2179 } else {
2180 *tree_pagep = tree_page;
2181 return tree_rmap_item;
2182 }
2183 }
2184
2185 rmap_item->address |= UNSTABLE_FLAG;
2186 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
2187 DO_NUMA(rmap_item->nid = nid);
2188 rb_link_node(&rmap_item->node, parent, new);
2189 rb_insert_color(&rmap_item->node, root);
2190
2191 ksm_pages_unshared++;
2192 return NULL;
2193 }
2194
2195 /*
2196 * stable_tree_append - add another rmap_item to the linked list of
2197 * rmap_items hanging off a given node of the stable tree, all sharing
2198 * the same ksm page.
2199 */
2200 static void stable_tree_append(struct ksm_rmap_item *rmap_item,
2201 struct ksm_stable_node *stable_node,
2202 bool max_page_sharing_bypass)
2203 {
2204 /*
2205 * rmap won't find this mapping if we don't insert the
2206 * rmap_item in the right stable_node
2207 * duplicate. page_migration could break later if rmap breaks,
2208 * so we can as well crash here. We really need to check for
2209 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
2210 * for other negative values as an underflow if detected here
2211 * for the first time (and not when decreasing rmap_hlist_len)
2212 * would be sign of memory corruption in the stable_node.
2213 */
2214 BUG_ON(stable_node->rmap_hlist_len < 0);
2215
2216 stable_node->rmap_hlist_len++;
2217 if (!max_page_sharing_bypass)
2218 /* possibly non fatal but unexpected overflow, only warn */
2219 WARN_ON_ONCE(stable_node->rmap_hlist_len >
2220 ksm_max_page_sharing);
2221
2222 rmap_item->head = stable_node;
2223 rmap_item->address |= STABLE_FLAG;
2224 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2225
2226 if (rmap_item->hlist.next)
2227 ksm_pages_sharing++;
2228 else
2229 ksm_pages_shared++;
2230
2231 rmap_item->mm->ksm_merging_pages++;
2232 }
2233
2234 /*
2235 * cmp_and_merge_page - first see if page can be merged into the stable tree;
2236 * if not, compare checksum to previous and if it's the same, see if page can
2237 * be inserted into the unstable tree, or merged with a page already there and
2238 * both transferred to the stable tree.
2239 *
2240 * @page: the page that we are searching identical page to.
2241 * @rmap_item: the reverse mapping into the virtual address of this page
2242 */
2243 static void cmp_and_merge_page(struct page *page, struct ksm_rmap_item *rmap_item)
2244 {
2245 struct ksm_rmap_item *tree_rmap_item;
2246 struct page *tree_page = NULL;
2247 struct ksm_stable_node *stable_node;
2248 struct page *kpage;
2249 unsigned int checksum;
2250 int err;
2251 bool max_page_sharing_bypass = false;
2252
2253 stable_node = page_stable_node(page);
2254 if (stable_node) {
2255 if (stable_node->head != &migrate_nodes &&
2256 get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2257 NUMA(stable_node->nid)) {
2258 stable_node_dup_del(stable_node);
2259 stable_node->head = &migrate_nodes;
2260 list_add(&stable_node->list, stable_node->head);
2261 }
2262 if (stable_node->head != &migrate_nodes &&
2263 rmap_item->head == stable_node)
2264 return;
2265 /*
2266 * If it's a KSM fork, allow it to go over the sharing limit
2267 * without warnings.
2268 */
2269 if (!is_page_sharing_candidate(stable_node))
2270 max_page_sharing_bypass = true;
2271 } else {
2272 remove_rmap_item_from_tree(rmap_item);
2273
2274 /*
2275 * If the hash value of the page has changed from the last time
2276 * we calculated it, this page is changing frequently: therefore we
2277 * don't want to insert it in the unstable tree, and we don't want
2278 * to waste our time searching for something identical to it there.
2279 */
2280 checksum = calc_checksum(page);
2281 if (rmap_item->oldchecksum != checksum) {
2282 rmap_item->oldchecksum = checksum;
2283 return;
2284 }
2285
2286 if (!try_to_merge_with_zero_page(rmap_item, page))
2287 return;
2288 }
2289
2290 /* We first start with searching the page inside the stable tree */
2291 kpage = stable_tree_search(page);
2292 if (kpage == page && rmap_item->head == stable_node) {
2293 put_page(kpage);
2294 return;
2295 }
2296
2297 remove_rmap_item_from_tree(rmap_item);
2298
2299 if (kpage) {
2300 if (PTR_ERR(kpage) == -EBUSY)
2301 return;
2302
2303 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2304 if (!err) {
2305 /*
2306 * The page was successfully merged:
2307 * add its rmap_item to the stable tree.
2308 */
2309 lock_page(kpage);
2310 stable_tree_append(rmap_item, page_stable_node(kpage),
2311 max_page_sharing_bypass);
2312 unlock_page(kpage);
2313 }
2314 put_page(kpage);
2315 return;
2316 }
2317
2318 tree_rmap_item =
2319 unstable_tree_search_insert(rmap_item, page, &tree_page);
2320 if (tree_rmap_item) {
2321 bool split;
2322
2323 kpage = try_to_merge_two_pages(rmap_item, page,
2324 tree_rmap_item, tree_page);
2325 /*
2326 * If both pages we tried to merge belong to the same compound
2327 * page, then we actually ended up increasing the reference
2328 * count of the same compound page twice, and split_huge_page
2329 * failed.
2330 * Here we set a flag if that happened, and we use it later to
2331 * try split_huge_page again. Since we call put_page right
2332 * afterwards, the reference count will be correct and
2333 * split_huge_page should succeed.
2334 */
2335 split = PageTransCompound(page)
2336 && compound_head(page) == compound_head(tree_page);
2337 put_page(tree_page);
2338 if (kpage) {
2339 /*
2340 * The pages were successfully merged: insert new
2341 * node in the stable tree and add both rmap_items.
2342 */
2343 lock_page(kpage);
2344 stable_node = stable_tree_insert(page_folio(kpage));
2345 if (stable_node) {
2346 stable_tree_append(tree_rmap_item, stable_node,
2347 false);
2348 stable_tree_append(rmap_item, stable_node,
2349 false);
2350 }
2351 unlock_page(kpage);
2352
2353 /*
2354 * If we fail to insert the page into the stable tree,
2355 * we will have 2 virtual addresses that are pointing
2356 * to a ksm page left outside the stable tree,
2357 * in which case we need to break_cow on both.
2358 */
2359 if (!stable_node) {
2360 break_cow(tree_rmap_item);
2361 break_cow(rmap_item);
2362 }
2363 } else if (split) {
2364 /*
2365 * We are here if we tried to merge two pages and
2366 * failed because they both belonged to the same
2367 * compound page. We will split the page now, but no
2368 * merging will take place.
2369 * We do not want to add the cost of a full lock; if
2370 * the page is locked, it is better to skip it and
2371 * perhaps try again later.
2372 */
2373 if (!trylock_page(page))
2374 return;
2375 split_huge_page(page);
2376 unlock_page(page);
2377 }
2378 }
2379 }
2380
2381 static struct ksm_rmap_item *get_next_rmap_item(struct ksm_mm_slot *mm_slot,
2382 struct ksm_rmap_item **rmap_list,
2383 unsigned long addr)
2384 {
2385 struct ksm_rmap_item *rmap_item;
2386
2387 while (*rmap_list) {
2388 rmap_item = *rmap_list;
2389 if ((rmap_item->address & PAGE_MASK) == addr)
2390 return rmap_item;
2391 if (rmap_item->address > addr)
2392 break;
2393 *rmap_list = rmap_item->rmap_list;
2394 remove_rmap_item_from_tree(rmap_item);
2395 free_rmap_item(rmap_item);
2396 }
2397
2398 rmap_item = alloc_rmap_item();
2399 if (rmap_item) {
2400 /* It has already been zeroed */
2401 rmap_item->mm = mm_slot->slot.mm;
2402 rmap_item->mm->ksm_rmap_items++;
2403 rmap_item->address = addr;
2404 rmap_item->rmap_list = *rmap_list;
2405 *rmap_list = rmap_item;
2406 }
2407 return rmap_item;
2408 }
2409
2410 /*
2411 * Calculate skip age for the ksm page age. The age determines how often
2412 * de-duplicating has already been tried unsuccessfully. If the age is
2413 * smaller, the scanning of this page is skipped for less scans.
2414 *
2415 * @age: rmap_item age of page
2416 */
2417 static unsigned int skip_age(rmap_age_t age)
2418 {
2419 if (age <= 3)
2420 return 1;
2421 if (age <= 5)
2422 return 2;
2423 if (age <= 8)
2424 return 4;
2425
2426 return 8;
2427 }
2428
2429 /*
2430 * Determines if a page should be skipped for the current scan.
2431 *
2432 * @page: page to check
2433 * @rmap_item: associated rmap_item of page
2434 */
2435 static bool should_skip_rmap_item(struct page *page,
2436 struct ksm_rmap_item *rmap_item)
2437 {
2438 rmap_age_t age;
2439
2440 if (!ksm_smart_scan)
2441 return false;
2442
2443 /*
2444 * Never skip pages that are already KSM; pages cmp_and_merge_page()
2445 * will essentially ignore them, but we still have to process them
2446 * properly.
2447 */
2448 if (PageKsm(page))
2449 return false;
2450
2451 age = rmap_item->age;
2452 if (age != U8_MAX)
2453 rmap_item->age++;
2454
2455 /*
2456 * Smaller ages are not skipped, they need to get a chance to go
2457 * through the different phases of the KSM merging.
2458 */
2459 if (age < 3)
2460 return false;
2461
2462 /*
2463 * Are we still allowed to skip? If not, then don't skip it
2464 * and determine how much more often we are allowed to skip next.
2465 */
2466 if (!rmap_item->remaining_skips) {
2467 rmap_item->remaining_skips = skip_age(age);
2468 return false;
2469 }
2470
2471 /* Skip this page */
2472 ksm_pages_skipped++;
2473 rmap_item->remaining_skips--;
2474 remove_rmap_item_from_tree(rmap_item);
2475 return true;
2476 }
2477
2478 static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page)
2479 {
2480 struct mm_struct *mm;
2481 struct ksm_mm_slot *mm_slot;
2482 struct mm_slot *slot;
2483 struct vm_area_struct *vma;
2484 struct ksm_rmap_item *rmap_item;
2485 struct vma_iterator vmi;
2486 int nid;
2487
2488 if (list_empty(&ksm_mm_head.slot.mm_node))
2489 return NULL;
2490
2491 mm_slot = ksm_scan.mm_slot;
2492 if (mm_slot == &ksm_mm_head) {
2493 advisor_start_scan();
2494 trace_ksm_start_scan(ksm_scan.seqnr, ksm_rmap_items);
2495
2496 /*
2497 * A number of pages can hang around indefinitely in per-cpu
2498 * LRU cache, raised page count preventing write_protect_page
2499 * from merging them. Though it doesn't really matter much,
2500 * it is puzzling to see some stuck in pages_volatile until
2501 * other activity jostles them out, and they also prevented
2502 * LTP's KSM test from succeeding deterministically; so drain
2503 * them here (here rather than on entry to ksm_do_scan(),
2504 * so we don't IPI too often when pages_to_scan is set low).
2505 */
2506 lru_add_drain_all();
2507
2508 /*
2509 * Whereas stale stable_nodes on the stable_tree itself
2510 * get pruned in the regular course of stable_tree_search(),
2511 * those moved out to the migrate_nodes list can accumulate:
2512 * so prune them once before each full scan.
2513 */
2514 if (!ksm_merge_across_nodes) {
2515 struct ksm_stable_node *stable_node, *next;
2516 struct folio *folio;
2517
2518 list_for_each_entry_safe(stable_node, next,
2519 &migrate_nodes, list) {
2520 folio = ksm_get_folio(stable_node,
2521 KSM_GET_FOLIO_NOLOCK);
2522 if (folio)
2523 folio_put(folio);
2524 cond_resched();
2525 }
2526 }
2527
2528 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2529 root_unstable_tree[nid] = RB_ROOT;
2530
2531 spin_lock(&ksm_mmlist_lock);
2532 slot = list_entry(mm_slot->slot.mm_node.next,
2533 struct mm_slot, mm_node);
2534 mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
2535 ksm_scan.mm_slot = mm_slot;
2536 spin_unlock(&ksm_mmlist_lock);
2537 /*
2538 * Although we tested list_empty() above, a racing __ksm_exit
2539 * of the last mm on the list may have removed it since then.
2540 */
2541 if (mm_slot == &ksm_mm_head)
2542 return NULL;
2543 next_mm:
2544 ksm_scan.address = 0;
2545 ksm_scan.rmap_list = &mm_slot->rmap_list;
2546 }
2547
2548 slot = &mm_slot->slot;
2549 mm = slot->mm;
2550 vma_iter_init(&vmi, mm, ksm_scan.address);
2551
2552 mmap_read_lock(mm);
2553 if (ksm_test_exit(mm))
2554 goto no_vmas;
2555
2556 for_each_vma(vmi, vma) {
2557 if (!(vma->vm_flags & VM_MERGEABLE))
2558 continue;
2559 if (ksm_scan.address < vma->vm_start)
2560 ksm_scan.address = vma->vm_start;
2561 if (!vma->anon_vma)
2562 ksm_scan.address = vma->vm_end;
2563
2564 while (ksm_scan.address < vma->vm_end) {
2565 if (ksm_test_exit(mm))
2566 break;
2567 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2568 if (IS_ERR_OR_NULL(*page)) {
2569 ksm_scan.address += PAGE_SIZE;
2570 cond_resched();
2571 continue;
2572 }
2573 if (is_zone_device_page(*page))
2574 goto next_page;
2575 if (PageAnon(*page)) {
2576 flush_anon_page(vma, *page, ksm_scan.address);
2577 flush_dcache_page(*page);
2578 rmap_item = get_next_rmap_item(mm_slot,
2579 ksm_scan.rmap_list, ksm_scan.address);
2580 if (rmap_item) {
2581 ksm_scan.rmap_list =
2582 &rmap_item->rmap_list;
2583
2584 if (should_skip_rmap_item(*page, rmap_item))
2585 goto next_page;
2586
2587 ksm_scan.address += PAGE_SIZE;
2588 } else
2589 put_page(*page);
2590 mmap_read_unlock(mm);
2591 return rmap_item;
2592 }
2593 next_page:
2594 put_page(*page);
2595 ksm_scan.address += PAGE_SIZE;
2596 cond_resched();
2597 }
2598 }
2599
2600 if (ksm_test_exit(mm)) {
2601 no_vmas:
2602 ksm_scan.address = 0;
2603 ksm_scan.rmap_list = &mm_slot->rmap_list;
2604 }
2605 /*
2606 * Nuke all the rmap_items that are above this current rmap:
2607 * because there were no VM_MERGEABLE vmas with such addresses.
2608 */
2609 remove_trailing_rmap_items(ksm_scan.rmap_list);
2610
2611 spin_lock(&ksm_mmlist_lock);
2612 slot = list_entry(mm_slot->slot.mm_node.next,
2613 struct mm_slot, mm_node);
2614 ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
2615 if (ksm_scan.address == 0) {
2616 /*
2617 * We've completed a full scan of all vmas, holding mmap_lock
2618 * throughout, and found no VM_MERGEABLE: so do the same as
2619 * __ksm_exit does to remove this mm from all our lists now.
2620 * This applies either when cleaning up after __ksm_exit
2621 * (but beware: we can reach here even before __ksm_exit),
2622 * or when all VM_MERGEABLE areas have been unmapped (and
2623 * mmap_lock then protects against race with MADV_MERGEABLE).
2624 */
2625 hash_del(&mm_slot->slot.hash);
2626 list_del(&mm_slot->slot.mm_node);
2627 spin_unlock(&ksm_mmlist_lock);
2628
2629 mm_slot_free(mm_slot_cache, mm_slot);
2630 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2631 clear_bit(MMF_VM_MERGE_ANY, &mm->flags);
2632 mmap_read_unlock(mm);
2633 mmdrop(mm);
2634 } else {
2635 mmap_read_unlock(mm);
2636 /*
2637 * mmap_read_unlock(mm) first because after
2638 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2639 * already have been freed under us by __ksm_exit()
2640 * because the "mm_slot" is still hashed and
2641 * ksm_scan.mm_slot doesn't point to it anymore.
2642 */
2643 spin_unlock(&ksm_mmlist_lock);
2644 }
2645
2646 /* Repeat until we've completed scanning the whole list */
2647 mm_slot = ksm_scan.mm_slot;
2648 if (mm_slot != &ksm_mm_head)
2649 goto next_mm;
2650
2651 advisor_stop_scan();
2652
2653 trace_ksm_stop_scan(ksm_scan.seqnr, ksm_rmap_items);
2654 ksm_scan.seqnr++;
2655 return NULL;
2656 }
2657
2658 /**
2659 * ksm_do_scan - the ksm scanner main worker function.
2660 * @scan_npages: number of pages we want to scan before we return.
2661 */
2662 static void ksm_do_scan(unsigned int scan_npages)
2663 {
2664 struct ksm_rmap_item *rmap_item;
2665 struct page *page;
2666
2667 while (scan_npages-- && likely(!freezing(current))) {
2668 cond_resched();
2669 rmap_item = scan_get_next_rmap_item(&page);
2670 if (!rmap_item)
2671 return;
2672 cmp_and_merge_page(page, rmap_item);
2673 put_page(page);
2674 ksm_pages_scanned++;
2675 }
2676 }
2677
2678 static int ksmd_should_run(void)
2679 {
2680 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.slot.mm_node);
2681 }
2682
2683 static int ksm_scan_thread(void *nothing)
2684 {
2685 unsigned int sleep_ms;
2686
2687 set_freezable();
2688 set_user_nice(current, 5);
2689
2690 while (!kthread_should_stop()) {
2691 mutex_lock(&ksm_thread_mutex);
2692 wait_while_offlining();
2693 if (ksmd_should_run())
2694 ksm_do_scan(ksm_thread_pages_to_scan);
2695 mutex_unlock(&ksm_thread_mutex);
2696
2697 if (ksmd_should_run()) {
2698 sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs);
2699 wait_event_freezable_timeout(ksm_iter_wait,
2700 sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs),
2701 msecs_to_jiffies(sleep_ms));
2702 } else {
2703 wait_event_freezable(ksm_thread_wait,
2704 ksmd_should_run() || kthread_should_stop());
2705 }
2706 }
2707 return 0;
2708 }
2709
2710 static void __ksm_add_vma(struct vm_area_struct *vma)
2711 {
2712 unsigned long vm_flags = vma->vm_flags;
2713
2714 if (vm_flags & VM_MERGEABLE)
2715 return;
2716
2717 if (vma_ksm_compatible(vma))
2718 vm_flags_set(vma, VM_MERGEABLE);
2719 }
2720
2721 static int __ksm_del_vma(struct vm_area_struct *vma)
2722 {
2723 int err;
2724
2725 if (!(vma->vm_flags & VM_MERGEABLE))
2726 return 0;
2727
2728 if (vma->anon_vma) {
2729 err = unmerge_ksm_pages(vma, vma->vm_start, vma->vm_end, true);
2730 if (err)
2731 return err;
2732 }
2733
2734 vm_flags_clear(vma, VM_MERGEABLE);
2735 return 0;
2736 }
2737 /**
2738 * ksm_add_vma - Mark vma as mergeable if compatible
2739 *
2740 * @vma: Pointer to vma
2741 */
2742 void ksm_add_vma(struct vm_area_struct *vma)
2743 {
2744 struct mm_struct *mm = vma->vm_mm;
2745
2746 if (test_bit(MMF_VM_MERGE_ANY, &mm->flags))
2747 __ksm_add_vma(vma);
2748 }
2749
2750 static void ksm_add_vmas(struct mm_struct *mm)
2751 {
2752 struct vm_area_struct *vma;
2753
2754 VMA_ITERATOR(vmi, mm, 0);
2755 for_each_vma(vmi, vma)
2756 __ksm_add_vma(vma);
2757 }
2758
2759 static int ksm_del_vmas(struct mm_struct *mm)
2760 {
2761 struct vm_area_struct *vma;
2762 int err;
2763
2764 VMA_ITERATOR(vmi, mm, 0);
2765 for_each_vma(vmi, vma) {
2766 err = __ksm_del_vma(vma);
2767 if (err)
2768 return err;
2769 }
2770 return 0;
2771 }
2772
2773 /**
2774 * ksm_enable_merge_any - Add mm to mm ksm list and enable merging on all
2775 * compatible VMA's
2776 *
2777 * @mm: Pointer to mm
2778 *
2779 * Returns 0 on success, otherwise error code
2780 */
2781 int ksm_enable_merge_any(struct mm_struct *mm)
2782 {
2783 int err;
2784
2785 if (test_bit(MMF_VM_MERGE_ANY, &mm->flags))
2786 return 0;
2787
2788 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2789 err = __ksm_enter(mm);
2790 if (err)
2791 return err;
2792 }
2793
2794 set_bit(MMF_VM_MERGE_ANY, &mm->flags);
2795 ksm_add_vmas(mm);
2796
2797 return 0;
2798 }
2799
2800 /**
2801 * ksm_disable_merge_any - Disable merging on all compatible VMA's of the mm,
2802 * previously enabled via ksm_enable_merge_any().
2803 *
2804 * Disabling merging implies unmerging any merged pages, like setting
2805 * MADV_UNMERGEABLE would. If unmerging fails, the whole operation fails and
2806 * merging on all compatible VMA's remains enabled.
2807 *
2808 * @mm: Pointer to mm
2809 *
2810 * Returns 0 on success, otherwise error code
2811 */
2812 int ksm_disable_merge_any(struct mm_struct *mm)
2813 {
2814 int err;
2815
2816 if (!test_bit(MMF_VM_MERGE_ANY, &mm->flags))
2817 return 0;
2818
2819 err = ksm_del_vmas(mm);
2820 if (err) {
2821 ksm_add_vmas(mm);
2822 return err;
2823 }
2824
2825 clear_bit(MMF_VM_MERGE_ANY, &mm->flags);
2826 return 0;
2827 }
2828
2829 int ksm_disable(struct mm_struct *mm)
2830 {
2831 mmap_assert_write_locked(mm);
2832
2833 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags))
2834 return 0;
2835 if (test_bit(MMF_VM_MERGE_ANY, &mm->flags))
2836 return ksm_disable_merge_any(mm);
2837 return ksm_del_vmas(mm);
2838 }
2839
2840 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2841 unsigned long end, int advice, unsigned long *vm_flags)
2842 {
2843 struct mm_struct *mm = vma->vm_mm;
2844 int err;
2845
2846 switch (advice) {
2847 case MADV_MERGEABLE:
2848 if (vma->vm_flags & VM_MERGEABLE)
2849 return 0;
2850 if (!vma_ksm_compatible(vma))
2851 return 0;
2852
2853 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2854 err = __ksm_enter(mm);
2855 if (err)
2856 return err;
2857 }
2858
2859 *vm_flags |= VM_MERGEABLE;
2860 break;
2861
2862 case MADV_UNMERGEABLE:
2863 if (!(*vm_flags & VM_MERGEABLE))
2864 return 0; /* just ignore the advice */
2865
2866 if (vma->anon_vma) {
2867 err = unmerge_ksm_pages(vma, start, end, true);
2868 if (err)
2869 return err;
2870 }
2871
2872 *vm_flags &= ~VM_MERGEABLE;
2873 break;
2874 }
2875
2876 return 0;
2877 }
2878 EXPORT_SYMBOL_GPL(ksm_madvise);
2879
2880 int __ksm_enter(struct mm_struct *mm)
2881 {
2882 struct ksm_mm_slot *mm_slot;
2883 struct mm_slot *slot;
2884 int needs_wakeup;
2885
2886 mm_slot = mm_slot_alloc(mm_slot_cache);
2887 if (!mm_slot)
2888 return -ENOMEM;
2889
2890 slot = &mm_slot->slot;
2891
2892 /* Check ksm_run too? Would need tighter locking */
2893 needs_wakeup = list_empty(&ksm_mm_head.slot.mm_node);
2894
2895 spin_lock(&ksm_mmlist_lock);
2896 mm_slot_insert(mm_slots_hash, mm, slot);
2897 /*
2898 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2899 * insert just behind the scanning cursor, to let the area settle
2900 * down a little; when fork is followed by immediate exec, we don't
2901 * want ksmd to waste time setting up and tearing down an rmap_list.
2902 *
2903 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2904 * scanning cursor, otherwise KSM pages in newly forked mms will be
2905 * missed: then we might as well insert at the end of the list.
2906 */
2907 if (ksm_run & KSM_RUN_UNMERGE)
2908 list_add_tail(&slot->mm_node, &ksm_mm_head.slot.mm_node);
2909 else
2910 list_add_tail(&slot->mm_node, &ksm_scan.mm_slot->slot.mm_node);
2911 spin_unlock(&ksm_mmlist_lock);
2912
2913 set_bit(MMF_VM_MERGEABLE, &mm->flags);
2914 mmgrab(mm);
2915
2916 if (needs_wakeup)
2917 wake_up_interruptible(&ksm_thread_wait);
2918
2919 trace_ksm_enter(mm);
2920 return 0;
2921 }
2922
2923 void __ksm_exit(struct mm_struct *mm)
2924 {
2925 struct ksm_mm_slot *mm_slot;
2926 struct mm_slot *slot;
2927 int easy_to_free = 0;
2928
2929 /*
2930 * This process is exiting: if it's straightforward (as is the
2931 * case when ksmd was never running), free mm_slot immediately.
2932 * But if it's at the cursor or has rmap_items linked to it, use
2933 * mmap_lock to synchronize with any break_cows before pagetables
2934 * are freed, and leave the mm_slot on the list for ksmd to free.
2935 * Beware: ksm may already have noticed it exiting and freed the slot.
2936 */
2937
2938 spin_lock(&ksm_mmlist_lock);
2939 slot = mm_slot_lookup(mm_slots_hash, mm);
2940 mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
2941 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2942 if (!mm_slot->rmap_list) {
2943 hash_del(&slot->hash);
2944 list_del(&slot->mm_node);
2945 easy_to_free = 1;
2946 } else {
2947 list_move(&slot->mm_node,
2948 &ksm_scan.mm_slot->slot.mm_node);
2949 }
2950 }
2951 spin_unlock(&ksm_mmlist_lock);
2952
2953 if (easy_to_free) {
2954 mm_slot_free(mm_slot_cache, mm_slot);
2955 clear_bit(MMF_VM_MERGE_ANY, &mm->flags);
2956 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2957 mmdrop(mm);
2958 } else if (mm_slot) {
2959 mmap_write_lock(mm);
2960 mmap_write_unlock(mm);
2961 }
2962
2963 trace_ksm_exit(mm);
2964 }
2965
2966 struct folio *ksm_might_need_to_copy(struct folio *folio,
2967 struct vm_area_struct *vma, unsigned long addr)
2968 {
2969 struct page *page = folio_page(folio, 0);
2970 struct anon_vma *anon_vma = folio_anon_vma(folio);
2971 struct folio *new_folio;
2972
2973 if (folio_test_large(folio))
2974 return folio;
2975
2976 if (folio_test_ksm(folio)) {
2977 if (folio_stable_node(folio) &&
2978 !(ksm_run & KSM_RUN_UNMERGE))
2979 return folio; /* no need to copy it */
2980 } else if (!anon_vma) {
2981 return folio; /* no need to copy it */
2982 } else if (folio->index == linear_page_index(vma, addr) &&
2983 anon_vma->root == vma->anon_vma->root) {
2984 return folio; /* still no need to copy it */
2985 }
2986 if (PageHWPoison(page))
2987 return ERR_PTR(-EHWPOISON);
2988 if (!folio_test_uptodate(folio))
2989 return folio; /* let do_swap_page report the error */
2990
2991 new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma, addr, false);
2992 if (new_folio &&
2993 mem_cgroup_charge(new_folio, vma->vm_mm, GFP_KERNEL)) {
2994 folio_put(new_folio);
2995 new_folio = NULL;
2996 }
2997 if (new_folio) {
2998 if (copy_mc_user_highpage(folio_page(new_folio, 0), page,
2999 addr, vma)) {
3000 folio_put(new_folio);
3001 return ERR_PTR(-EHWPOISON);
3002 }
3003 folio_set_dirty(new_folio);
3004 __folio_mark_uptodate(new_folio);
3005 __folio_set_locked(new_folio);
3006 #ifdef CONFIG_SWAP
3007 count_vm_event(KSM_SWPIN_COPY);
3008 #endif
3009 }
3010
3011 return new_folio;
3012 }
3013
3014 void rmap_walk_ksm(struct folio *folio, struct rmap_walk_control *rwc)
3015 {
3016 struct ksm_stable_node *stable_node;
3017 struct ksm_rmap_item *rmap_item;
3018 int search_new_forks = 0;
3019
3020 VM_BUG_ON_FOLIO(!folio_test_ksm(folio), folio);
3021
3022 /*
3023 * Rely on the page lock to protect against concurrent modifications
3024 * to that page's node of the stable tree.
3025 */
3026 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
3027
3028 stable_node = folio_stable_node(folio);
3029 if (!stable_node)
3030 return;
3031 again:
3032 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
3033 struct anon_vma *anon_vma = rmap_item->anon_vma;
3034 struct anon_vma_chain *vmac;
3035 struct vm_area_struct *vma;
3036
3037 cond_resched();
3038 if (!anon_vma_trylock_read(anon_vma)) {
3039 if (rwc->try_lock) {
3040 rwc->contended = true;
3041 return;
3042 }
3043 anon_vma_lock_read(anon_vma);
3044 }
3045 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
3046 0, ULONG_MAX) {
3047 unsigned long addr;
3048
3049 cond_resched();
3050 vma = vmac->vma;
3051
3052 /* Ignore the stable/unstable/sqnr flags */
3053 addr = rmap_item->address & PAGE_MASK;
3054
3055 if (addr < vma->vm_start || addr >= vma->vm_end)
3056 continue;
3057 /*
3058 * Initially we examine only the vma which covers this
3059 * rmap_item; but later, if there is still work to do,
3060 * we examine covering vmas in other mms: in case they
3061 * were forked from the original since ksmd passed.
3062 */
3063 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
3064 continue;
3065
3066 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
3067 continue;
3068
3069 if (!rwc->rmap_one(folio, vma, addr, rwc->arg)) {
3070 anon_vma_unlock_read(anon_vma);
3071 return;
3072 }
3073 if (rwc->done && rwc->done(folio)) {
3074 anon_vma_unlock_read(anon_vma);
3075 return;
3076 }
3077 }
3078 anon_vma_unlock_read(anon_vma);
3079 }
3080 if (!search_new_forks++)
3081 goto again;
3082 }
3083
3084 #ifdef CONFIG_MEMORY_FAILURE
3085 /*
3086 * Collect processes when the error hit an ksm page.
3087 */
3088 void collect_procs_ksm(struct folio *folio, struct page *page,
3089 struct list_head *to_kill, int force_early)
3090 {
3091 struct ksm_stable_node *stable_node;
3092 struct ksm_rmap_item *rmap_item;
3093 struct vm_area_struct *vma;
3094 struct task_struct *tsk;
3095
3096 stable_node = folio_stable_node(folio);
3097 if (!stable_node)
3098 return;
3099 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
3100 struct anon_vma *av = rmap_item->anon_vma;
3101
3102 anon_vma_lock_read(av);
3103 rcu_read_lock();
3104 for_each_process(tsk) {
3105 struct anon_vma_chain *vmac;
3106 unsigned long addr;
3107 struct task_struct *t =
3108 task_early_kill(tsk, force_early);
3109 if (!t)
3110 continue;
3111 anon_vma_interval_tree_foreach(vmac, &av->rb_root, 0,
3112 ULONG_MAX)
3113 {
3114 vma = vmac->vma;
3115 if (vma->vm_mm == t->mm) {
3116 addr = rmap_item->address & PAGE_MASK;
3117 add_to_kill_ksm(t, page, vma, to_kill,
3118 addr);
3119 }
3120 }
3121 }
3122 rcu_read_unlock();
3123 anon_vma_unlock_read(av);
3124 }
3125 }
3126 #endif
3127
3128 #ifdef CONFIG_MIGRATION
3129 void folio_migrate_ksm(struct folio *newfolio, struct folio *folio)
3130 {
3131 struct ksm_stable_node *stable_node;
3132
3133 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
3134 VM_BUG_ON_FOLIO(!folio_test_locked(newfolio), newfolio);
3135 VM_BUG_ON_FOLIO(newfolio->mapping != folio->mapping, newfolio);
3136
3137 stable_node = folio_stable_node(folio);
3138 if (stable_node) {
3139 VM_BUG_ON_FOLIO(stable_node->kpfn != folio_pfn(folio), folio);
3140 stable_node->kpfn = folio_pfn(newfolio);
3141 /*
3142 * newfolio->mapping was set in advance; now we need smp_wmb()
3143 * to make sure that the new stable_node->kpfn is visible
3144 * to ksm_get_folio() before it can see that folio->mapping
3145 * has gone stale (or that folio_test_swapcache has been cleared).
3146 */
3147 smp_wmb();
3148 folio_set_stable_node(folio, NULL);
3149 }
3150 }
3151 #endif /* CONFIG_MIGRATION */
3152
3153 #ifdef CONFIG_MEMORY_HOTREMOVE
3154 static void wait_while_offlining(void)
3155 {
3156 while (ksm_run & KSM_RUN_OFFLINE) {
3157 mutex_unlock(&ksm_thread_mutex);
3158 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
3159 TASK_UNINTERRUPTIBLE);
3160 mutex_lock(&ksm_thread_mutex);
3161 }
3162 }
3163
3164 static bool stable_node_dup_remove_range(struct ksm_stable_node *stable_node,
3165 unsigned long start_pfn,
3166 unsigned long end_pfn)
3167 {
3168 if (stable_node->kpfn >= start_pfn &&
3169 stable_node->kpfn < end_pfn) {
3170 /*
3171 * Don't ksm_get_folio, page has already gone:
3172 * which is why we keep kpfn instead of page*
3173 */
3174 remove_node_from_stable_tree(stable_node);
3175 return true;
3176 }
3177 return false;
3178 }
3179
3180 static bool stable_node_chain_remove_range(struct ksm_stable_node *stable_node,
3181 unsigned long start_pfn,
3182 unsigned long end_pfn,
3183 struct rb_root *root)
3184 {
3185 struct ksm_stable_node *dup;
3186 struct hlist_node *hlist_safe;
3187
3188 if (!is_stable_node_chain(stable_node)) {
3189 VM_BUG_ON(is_stable_node_dup(stable_node));
3190 return stable_node_dup_remove_range(stable_node, start_pfn,
3191 end_pfn);
3192 }
3193
3194 hlist_for_each_entry_safe(dup, hlist_safe,
3195 &stable_node->hlist, hlist_dup) {
3196 VM_BUG_ON(!is_stable_node_dup(dup));
3197 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
3198 }
3199 if (hlist_empty(&stable_node->hlist)) {
3200 free_stable_node_chain(stable_node, root);
3201 return true; /* notify caller that tree was rebalanced */
3202 } else
3203 return false;
3204 }
3205
3206 static void ksm_check_stable_tree(unsigned long start_pfn,
3207 unsigned long end_pfn)
3208 {
3209 struct ksm_stable_node *stable_node, *next;
3210 struct rb_node *node;
3211 int nid;
3212
3213 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
3214 node = rb_first(root_stable_tree + nid);
3215 while (node) {
3216 stable_node = rb_entry(node, struct ksm_stable_node, node);
3217 if (stable_node_chain_remove_range(stable_node,
3218 start_pfn, end_pfn,
3219 root_stable_tree +
3220 nid))
3221 node = rb_first(root_stable_tree + nid);
3222 else
3223 node = rb_next(node);
3224 cond_resched();
3225 }
3226 }
3227 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
3228 if (stable_node->kpfn >= start_pfn &&
3229 stable_node->kpfn < end_pfn)
3230 remove_node_from_stable_tree(stable_node);
3231 cond_resched();
3232 }
3233 }
3234
3235 static int ksm_memory_callback(struct notifier_block *self,
3236 unsigned long action, void *arg)
3237 {
3238 struct memory_notify *mn = arg;
3239
3240 switch (action) {
3241 case MEM_GOING_OFFLINE:
3242 /*
3243 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
3244 * and remove_all_stable_nodes() while memory is going offline:
3245 * it is unsafe for them to touch the stable tree at this time.
3246 * But unmerge_ksm_pages(), rmap lookups and other entry points
3247 * which do not need the ksm_thread_mutex are all safe.
3248 */
3249 mutex_lock(&ksm_thread_mutex);
3250 ksm_run |= KSM_RUN_OFFLINE;
3251 mutex_unlock(&ksm_thread_mutex);
3252 break;
3253
3254 case MEM_OFFLINE:
3255 /*
3256 * Most of the work is done by page migration; but there might
3257 * be a few stable_nodes left over, still pointing to struct
3258 * pages which have been offlined: prune those from the tree,
3259 * otherwise ksm_get_folio() might later try to access a
3260 * non-existent struct page.
3261 */
3262 ksm_check_stable_tree(mn->start_pfn,
3263 mn->start_pfn + mn->nr_pages);
3264 fallthrough;
3265 case MEM_CANCEL_OFFLINE:
3266 mutex_lock(&ksm_thread_mutex);
3267 ksm_run &= ~KSM_RUN_OFFLINE;
3268 mutex_unlock(&ksm_thread_mutex);
3269
3270 smp_mb(); /* wake_up_bit advises this */
3271 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
3272 break;
3273 }
3274 return NOTIFY_OK;
3275 }
3276 #else
3277 static void wait_while_offlining(void)
3278 {
3279 }
3280 #endif /* CONFIG_MEMORY_HOTREMOVE */
3281
3282 #ifdef CONFIG_PROC_FS
3283 long ksm_process_profit(struct mm_struct *mm)
3284 {
3285 return (long)(mm->ksm_merging_pages + mm_ksm_zero_pages(mm)) * PAGE_SIZE -
3286 mm->ksm_rmap_items * sizeof(struct ksm_rmap_item);
3287 }
3288 #endif /* CONFIG_PROC_FS */
3289
3290 #ifdef CONFIG_SYSFS
3291 /*
3292 * This all compiles without CONFIG_SYSFS, but is a waste of space.
3293 */
3294
3295 #define KSM_ATTR_RO(_name) \
3296 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3297 #define KSM_ATTR(_name) \
3298 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3299
3300 static ssize_t sleep_millisecs_show(struct kobject *kobj,
3301 struct kobj_attribute *attr, char *buf)
3302 {
3303 return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs);
3304 }
3305
3306 static ssize_t sleep_millisecs_store(struct kobject *kobj,
3307 struct kobj_attribute *attr,
3308 const char *buf, size_t count)
3309 {
3310 unsigned int msecs;
3311 int err;
3312
3313 err = kstrtouint(buf, 10, &msecs);
3314 if (err)
3315 return -EINVAL;
3316
3317 ksm_thread_sleep_millisecs = msecs;
3318 wake_up_interruptible(&ksm_iter_wait);
3319
3320 return count;
3321 }
3322 KSM_ATTR(sleep_millisecs);
3323
3324 static ssize_t pages_to_scan_show(struct kobject *kobj,
3325 struct kobj_attribute *attr, char *buf)
3326 {
3327 return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan);
3328 }
3329
3330 static ssize_t pages_to_scan_store(struct kobject *kobj,
3331 struct kobj_attribute *attr,
3332 const char *buf, size_t count)
3333 {
3334 unsigned int nr_pages;
3335 int err;
3336
3337 if (ksm_advisor != KSM_ADVISOR_NONE)
3338 return -EINVAL;
3339
3340 err = kstrtouint(buf, 10, &nr_pages);
3341 if (err)
3342 return -EINVAL;
3343
3344 ksm_thread_pages_to_scan = nr_pages;
3345
3346 return count;
3347 }
3348 KSM_ATTR(pages_to_scan);
3349
3350 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
3351 char *buf)
3352 {
3353 return sysfs_emit(buf, "%lu\n", ksm_run);
3354 }
3355
3356 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
3357 const char *buf, size_t count)
3358 {
3359 unsigned int flags;
3360 int err;
3361
3362 err = kstrtouint(buf, 10, &flags);
3363 if (err)
3364 return -EINVAL;
3365 if (flags > KSM_RUN_UNMERGE)
3366 return -EINVAL;
3367
3368 /*
3369 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
3370 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
3371 * breaking COW to free the pages_shared (but leaves mm_slots
3372 * on the list for when ksmd may be set running again).
3373 */
3374
3375 mutex_lock(&ksm_thread_mutex);
3376 wait_while_offlining();
3377 if (ksm_run != flags) {
3378 ksm_run = flags;
3379 if (flags & KSM_RUN_UNMERGE) {
3380 set_current_oom_origin();
3381 err = unmerge_and_remove_all_rmap_items();
3382 clear_current_oom_origin();
3383 if (err) {
3384 ksm_run = KSM_RUN_STOP;
3385 count = err;
3386 }
3387 }
3388 }
3389 mutex_unlock(&ksm_thread_mutex);
3390
3391 if (flags & KSM_RUN_MERGE)
3392 wake_up_interruptible(&ksm_thread_wait);
3393
3394 return count;
3395 }
3396 KSM_ATTR(run);
3397
3398 #ifdef CONFIG_NUMA
3399 static ssize_t merge_across_nodes_show(struct kobject *kobj,
3400 struct kobj_attribute *attr, char *buf)
3401 {
3402 return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes);
3403 }
3404
3405 static ssize_t merge_across_nodes_store(struct kobject *kobj,
3406 struct kobj_attribute *attr,
3407 const char *buf, size_t count)
3408 {
3409 int err;
3410 unsigned long knob;
3411
3412 err = kstrtoul(buf, 10, &knob);
3413 if (err)
3414 return err;
3415 if (knob > 1)
3416 return -EINVAL;
3417
3418 mutex_lock(&ksm_thread_mutex);
3419 wait_while_offlining();
3420 if (ksm_merge_across_nodes != knob) {
3421 if (ksm_pages_shared || remove_all_stable_nodes())
3422 err = -EBUSY;
3423 else if (root_stable_tree == one_stable_tree) {
3424 struct rb_root *buf;
3425 /*
3426 * This is the first time that we switch away from the
3427 * default of merging across nodes: must now allocate
3428 * a buffer to hold as many roots as may be needed.
3429 * Allocate stable and unstable together:
3430 * MAXSMP NODES_SHIFT 10 will use 16kB.
3431 */
3432 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
3433 GFP_KERNEL);
3434 /* Let us assume that RB_ROOT is NULL is zero */
3435 if (!buf)
3436 err = -ENOMEM;
3437 else {
3438 root_stable_tree = buf;
3439 root_unstable_tree = buf + nr_node_ids;
3440 /* Stable tree is empty but not the unstable */
3441 root_unstable_tree[0] = one_unstable_tree[0];
3442 }
3443 }
3444 if (!err) {
3445 ksm_merge_across_nodes = knob;
3446 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
3447 }
3448 }
3449 mutex_unlock(&ksm_thread_mutex);
3450
3451 return err ? err : count;
3452 }
3453 KSM_ATTR(merge_across_nodes);
3454 #endif
3455
3456 static ssize_t use_zero_pages_show(struct kobject *kobj,
3457 struct kobj_attribute *attr, char *buf)
3458 {
3459 return sysfs_emit(buf, "%u\n", ksm_use_zero_pages);
3460 }
3461 static ssize_t use_zero_pages_store(struct kobject *kobj,
3462 struct kobj_attribute *attr,
3463 const char *buf, size_t count)
3464 {
3465 int err;
3466 bool value;
3467
3468 err = kstrtobool(buf, &value);
3469 if (err)
3470 return -EINVAL;
3471
3472 ksm_use_zero_pages = value;
3473
3474 return count;
3475 }
3476 KSM_ATTR(use_zero_pages);
3477
3478 static ssize_t max_page_sharing_show(struct kobject *kobj,
3479 struct kobj_attribute *attr, char *buf)
3480 {
3481 return sysfs_emit(buf, "%u\n", ksm_max_page_sharing);
3482 }
3483
3484 static ssize_t max_page_sharing_store(struct kobject *kobj,
3485 struct kobj_attribute *attr,
3486 const char *buf, size_t count)
3487 {
3488 int err;
3489 int knob;
3490
3491 err = kstrtoint(buf, 10, &knob);
3492 if (err)
3493 return err;
3494 /*
3495 * When a KSM page is created it is shared by 2 mappings. This
3496 * being a signed comparison, it implicitly verifies it's not
3497 * negative.
3498 */
3499 if (knob < 2)
3500 return -EINVAL;
3501
3502 if (READ_ONCE(ksm_max_page_sharing) == knob)
3503 return count;
3504
3505 mutex_lock(&ksm_thread_mutex);
3506 wait_while_offlining();
3507 if (ksm_max_page_sharing != knob) {
3508 if (ksm_pages_shared || remove_all_stable_nodes())
3509 err = -EBUSY;
3510 else
3511 ksm_max_page_sharing = knob;
3512 }
3513 mutex_unlock(&ksm_thread_mutex);
3514
3515 return err ? err : count;
3516 }
3517 KSM_ATTR(max_page_sharing);
3518
3519 static ssize_t pages_scanned_show(struct kobject *kobj,
3520 struct kobj_attribute *attr, char *buf)
3521 {
3522 return sysfs_emit(buf, "%lu\n", ksm_pages_scanned);
3523 }
3524 KSM_ATTR_RO(pages_scanned);
3525
3526 static ssize_t pages_shared_show(struct kobject *kobj,
3527 struct kobj_attribute *attr, char *buf)
3528 {
3529 return sysfs_emit(buf, "%lu\n", ksm_pages_shared);
3530 }
3531 KSM_ATTR_RO(pages_shared);
3532
3533 static ssize_t pages_sharing_show(struct kobject *kobj,
3534 struct kobj_attribute *attr, char *buf)
3535 {
3536 return sysfs_emit(buf, "%lu\n", ksm_pages_sharing);
3537 }
3538 KSM_ATTR_RO(pages_sharing);
3539
3540 static ssize_t pages_unshared_show(struct kobject *kobj,
3541 struct kobj_attribute *attr, char *buf)
3542 {
3543 return sysfs_emit(buf, "%lu\n", ksm_pages_unshared);
3544 }
3545 KSM_ATTR_RO(pages_unshared);
3546
3547 static ssize_t pages_volatile_show(struct kobject *kobj,
3548 struct kobj_attribute *attr, char *buf)
3549 {
3550 long ksm_pages_volatile;
3551
3552 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3553 - ksm_pages_sharing - ksm_pages_unshared;
3554 /*
3555 * It was not worth any locking to calculate that statistic,
3556 * but it might therefore sometimes be negative: conceal that.
3557 */
3558 if (ksm_pages_volatile < 0)
3559 ksm_pages_volatile = 0;
3560 return sysfs_emit(buf, "%ld\n", ksm_pages_volatile);
3561 }
3562 KSM_ATTR_RO(pages_volatile);
3563
3564 static ssize_t pages_skipped_show(struct kobject *kobj,
3565 struct kobj_attribute *attr, char *buf)
3566 {
3567 return sysfs_emit(buf, "%lu\n", ksm_pages_skipped);
3568 }
3569 KSM_ATTR_RO(pages_skipped);
3570
3571 static ssize_t ksm_zero_pages_show(struct kobject *kobj,
3572 struct kobj_attribute *attr, char *buf)
3573 {
3574 return sysfs_emit(buf, "%ld\n", atomic_long_read(&ksm_zero_pages));
3575 }
3576 KSM_ATTR_RO(ksm_zero_pages);
3577
3578 static ssize_t general_profit_show(struct kobject *kobj,
3579 struct kobj_attribute *attr, char *buf)
3580 {
3581 long general_profit;
3582
3583 general_profit = (ksm_pages_sharing + atomic_long_read(&ksm_zero_pages)) * PAGE_SIZE -
3584 ksm_rmap_items * sizeof(struct ksm_rmap_item);
3585
3586 return sysfs_emit(buf, "%ld\n", general_profit);
3587 }
3588 KSM_ATTR_RO(general_profit);
3589
3590 static ssize_t stable_node_dups_show(struct kobject *kobj,
3591 struct kobj_attribute *attr, char *buf)
3592 {
3593 return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups);
3594 }
3595 KSM_ATTR_RO(stable_node_dups);
3596
3597 static ssize_t stable_node_chains_show(struct kobject *kobj,
3598 struct kobj_attribute *attr, char *buf)
3599 {
3600 return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains);
3601 }
3602 KSM_ATTR_RO(stable_node_chains);
3603
3604 static ssize_t
3605 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3606 struct kobj_attribute *attr,
3607 char *buf)
3608 {
3609 return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3610 }
3611
3612 static ssize_t
3613 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3614 struct kobj_attribute *attr,
3615 const char *buf, size_t count)
3616 {
3617 unsigned int msecs;
3618 int err;
3619
3620 err = kstrtouint(buf, 10, &msecs);
3621 if (err)
3622 return -EINVAL;
3623
3624 ksm_stable_node_chains_prune_millisecs = msecs;
3625
3626 return count;
3627 }
3628 KSM_ATTR(stable_node_chains_prune_millisecs);
3629
3630 static ssize_t full_scans_show(struct kobject *kobj,
3631 struct kobj_attribute *attr, char *buf)
3632 {
3633 return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr);
3634 }
3635 KSM_ATTR_RO(full_scans);
3636
3637 static ssize_t smart_scan_show(struct kobject *kobj,
3638 struct kobj_attribute *attr, char *buf)
3639 {
3640 return sysfs_emit(buf, "%u\n", ksm_smart_scan);
3641 }
3642
3643 static ssize_t smart_scan_store(struct kobject *kobj,
3644 struct kobj_attribute *attr,
3645 const char *buf, size_t count)
3646 {
3647 int err;
3648 bool value;
3649
3650 err = kstrtobool(buf, &value);
3651 if (err)
3652 return -EINVAL;
3653
3654 ksm_smart_scan = value;
3655 return count;
3656 }
3657 KSM_ATTR(smart_scan);
3658
3659 static ssize_t advisor_mode_show(struct kobject *kobj,
3660 struct kobj_attribute *attr, char *buf)
3661 {
3662 const char *output;
3663
3664 if (ksm_advisor == KSM_ADVISOR_NONE)
3665 output = "[none] scan-time";
3666 else if (ksm_advisor == KSM_ADVISOR_SCAN_TIME)
3667 output = "none [scan-time]";
3668
3669 return sysfs_emit(buf, "%s\n", output);
3670 }
3671
3672 static ssize_t advisor_mode_store(struct kobject *kobj,
3673 struct kobj_attribute *attr, const char *buf,
3674 size_t count)
3675 {
3676 enum ksm_advisor_type curr_advisor = ksm_advisor;
3677
3678 if (sysfs_streq("scan-time", buf))
3679 ksm_advisor = KSM_ADVISOR_SCAN_TIME;
3680 else if (sysfs_streq("none", buf))
3681 ksm_advisor = KSM_ADVISOR_NONE;
3682 else
3683 return -EINVAL;
3684
3685 /* Set advisor default values */
3686 if (curr_advisor != ksm_advisor)
3687 set_advisor_defaults();
3688
3689 return count;
3690 }
3691 KSM_ATTR(advisor_mode);
3692
3693 static ssize_t advisor_max_cpu_show(struct kobject *kobj,
3694 struct kobj_attribute *attr, char *buf)
3695 {
3696 return sysfs_emit(buf, "%u\n", ksm_advisor_max_cpu);
3697 }
3698
3699 static ssize_t advisor_max_cpu_store(struct kobject *kobj,
3700 struct kobj_attribute *attr,
3701 const char *buf, size_t count)
3702 {
3703 int err;
3704 unsigned long value;
3705
3706 err = kstrtoul(buf, 10, &value);
3707 if (err)
3708 return -EINVAL;
3709
3710 ksm_advisor_max_cpu = value;
3711 return count;
3712 }
3713 KSM_ATTR(advisor_max_cpu);
3714
3715 static ssize_t advisor_min_pages_to_scan_show(struct kobject *kobj,
3716 struct kobj_attribute *attr, char *buf)
3717 {
3718 return sysfs_emit(buf, "%lu\n", ksm_advisor_min_pages_to_scan);
3719 }
3720
3721 static ssize_t advisor_min_pages_to_scan_store(struct kobject *kobj,
3722 struct kobj_attribute *attr,
3723 const char *buf, size_t count)
3724 {
3725 int err;
3726 unsigned long value;
3727
3728 err = kstrtoul(buf, 10, &value);
3729 if (err)
3730 return -EINVAL;
3731
3732 ksm_advisor_min_pages_to_scan = value;
3733 return count;
3734 }
3735 KSM_ATTR(advisor_min_pages_to_scan);
3736
3737 static ssize_t advisor_max_pages_to_scan_show(struct kobject *kobj,
3738 struct kobj_attribute *attr, char *buf)
3739 {
3740 return sysfs_emit(buf, "%lu\n", ksm_advisor_max_pages_to_scan);
3741 }
3742
3743 static ssize_t advisor_max_pages_to_scan_store(struct kobject *kobj,
3744 struct kobj_attribute *attr,
3745 const char *buf, size_t count)
3746 {
3747 int err;
3748 unsigned long value;
3749
3750 err = kstrtoul(buf, 10, &value);
3751 if (err)
3752 return -EINVAL;
3753
3754 ksm_advisor_max_pages_to_scan = value;
3755 return count;
3756 }
3757 KSM_ATTR(advisor_max_pages_to_scan);
3758
3759 static ssize_t advisor_target_scan_time_show(struct kobject *kobj,
3760 struct kobj_attribute *attr, char *buf)
3761 {
3762 return sysfs_emit(buf, "%lu\n", ksm_advisor_target_scan_time);
3763 }
3764
3765 static ssize_t advisor_target_scan_time_store(struct kobject *kobj,
3766 struct kobj_attribute *attr,
3767 const char *buf, size_t count)
3768 {
3769 int err;
3770 unsigned long value;
3771
3772 err = kstrtoul(buf, 10, &value);
3773 if (err)
3774 return -EINVAL;
3775 if (value < 1)
3776 return -EINVAL;
3777
3778 ksm_advisor_target_scan_time = value;
3779 return count;
3780 }
3781 KSM_ATTR(advisor_target_scan_time);
3782
3783 static struct attribute *ksm_attrs[] = {
3784 &sleep_millisecs_attr.attr,
3785 &pages_to_scan_attr.attr,
3786 &run_attr.attr,
3787 &pages_scanned_attr.attr,
3788 &pages_shared_attr.attr,
3789 &pages_sharing_attr.attr,
3790 &pages_unshared_attr.attr,
3791 &pages_volatile_attr.attr,
3792 &pages_skipped_attr.attr,
3793 &ksm_zero_pages_attr.attr,
3794 &full_scans_attr.attr,
3795 #ifdef CONFIG_NUMA
3796 &merge_across_nodes_attr.attr,
3797 #endif
3798 &max_page_sharing_attr.attr,
3799 &stable_node_chains_attr.attr,
3800 &stable_node_dups_attr.attr,
3801 &stable_node_chains_prune_millisecs_attr.attr,
3802 &use_zero_pages_attr.attr,
3803 &general_profit_attr.attr,
3804 &smart_scan_attr.attr,
3805 &advisor_mode_attr.attr,
3806 &advisor_max_cpu_attr.attr,
3807 &advisor_min_pages_to_scan_attr.attr,
3808 &advisor_max_pages_to_scan_attr.attr,
3809 &advisor_target_scan_time_attr.attr,
3810 NULL,
3811 };
3812
3813 static const struct attribute_group ksm_attr_group = {
3814 .attrs = ksm_attrs,
3815 .name = "ksm",
3816 };
3817 #endif /* CONFIG_SYSFS */
3818
3819 static int __init ksm_init(void)
3820 {
3821 struct task_struct *ksm_thread;
3822 int err;
3823
3824 /* The correct value depends on page size and endianness */
3825 zero_checksum = calc_checksum(ZERO_PAGE(0));
3826 /* Default to false for backwards compatibility */
3827 ksm_use_zero_pages = false;
3828
3829 err = ksm_slab_init();
3830 if (err)
3831 goto out;
3832
3833 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3834 if (IS_ERR(ksm_thread)) {
3835 pr_err("ksm: creating kthread failed\n");
3836 err = PTR_ERR(ksm_thread);
3837 goto out_free;
3838 }
3839
3840 #ifdef CONFIG_SYSFS
3841 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3842 if (err) {
3843 pr_err("ksm: register sysfs failed\n");
3844 kthread_stop(ksm_thread);
3845 goto out_free;
3846 }
3847 #else
3848 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
3849
3850 #endif /* CONFIG_SYSFS */
3851
3852 #ifdef CONFIG_MEMORY_HOTREMOVE
3853 /* There is no significance to this priority 100 */
3854 hotplug_memory_notifier(ksm_memory_callback, KSM_CALLBACK_PRI);
3855 #endif
3856 return 0;
3857
3858 out_free:
3859 ksm_slab_free();
3860 out:
3861 return err;
3862 }
3863 subsys_initcall(ksm_init);
3864
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.