1 // SPDX-License-Identifier: GPL-2.0-only
3 * Memory merging support.
5 * This code enables dynamic sharing of identical pages found in different
6 * memory areas, even if they are not shared by fork()
8 * Copyright (C) 2008-2009 Red Hat, Inc.
16 #include <linux/errno.h>
18 #include <linux/mm_inline.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>
45 #include <asm/tlbflush.h>
49 #define CREATE_TRACE_POINTS
50 #include <trace/events/ksm.h>
54 #define DO_NUMA(x) do { (x); } while (0)
57 #define DO_NUMA(x) do { } while (0)
60 typedef u8 rmap_age_t;
65 * A few notes about the KSM scanning process,
66 * to make it easier to understand the data structures below:
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.
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.
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.
80 * The stable tree node includes information required for reverse
81 * mapping from a KSM page to virtual addresses that map this page.
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:
86 * * the regular nodes that keep the reverse mapping structures in a
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
92 * Internally, the regular nodes, "dups" and "chains" are represented
93 * using the same struct ksm_stable_node structure.
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.
102 * KSM solves this problem by several techniques:
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.)
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.
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
129 struct ksm_rmap_item *rmap_list;
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)
139 * There is only the one ksm_scan instance of this cursor structure.
142 struct ksm_mm_slot *mm_slot;
143 unsigned long address;
144 struct ksm_rmap_item **rmap_list;
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)
160 struct ksm_stable_node {
162 struct rb_node node; /* when node of stable tree */
163 struct { /* when listed for migration */
164 struct list_head *head;
166 struct hlist_node hlist_dup;
167 struct list_head list;
171 struct hlist_head hlist;
174 unsigned long chain_prune_time;
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.
181 #define STABLE_NODE_CHAIN -1024
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
202 struct ksm_rmap_item {
203 struct ksm_rmap_item *rmap_list;
205 struct anon_vma *anon_vma; /* when stable */
207 int nid; /* when node of unstable tree */
210 struct mm_struct *mm;
211 unsigned long address; /* + low bits used for flags below */
212 unsigned int oldchecksum; /* when unstable */
214 rmap_age_t remaining_skips;
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;
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 */
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;
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)
238 #define MM_SLOTS_HASH_BITS 10
239 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
241 static struct ksm_mm_slot ksm_mm_head = {
242 .slot.mm_node = LIST_HEAD_INIT(ksm_mm_head.slot.mm_node),
244 static struct ksm_scan ksm_scan = {
245 .mm_slot = &ksm_mm_head,
248 static struct kmem_cache *rmap_item_cache;
249 static struct kmem_cache *stable_node_cache;
250 static struct kmem_cache *mm_slot_cache;
252 /* Default number of pages to scan per batch */
253 #define DEFAULT_PAGES_TO_SCAN 100
255 /* The number of pages scanned */
256 static unsigned long ksm_pages_scanned;
258 /* The number of nodes in the stable tree */
259 static unsigned long ksm_pages_shared;
261 /* The number of page slots additionally sharing those nodes */
262 static unsigned long ksm_pages_sharing;
264 /* The number of nodes in the unstable tree */
265 static unsigned long ksm_pages_unshared;
267 /* The number of rmap_items in use: to calculate pages_volatile */
268 static unsigned long ksm_rmap_items;
270 /* The number of stable_node chains */
271 static unsigned long ksm_stable_node_chains;
273 /* The number of stable_node dups linked to the stable_node chains */
274 static unsigned long ksm_stable_node_dups;
276 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
277 static unsigned int ksm_stable_node_chains_prune_millisecs = 2000;
279 /* Maximum number of page slots sharing a stable node */
280 static int ksm_max_page_sharing = 256;
282 /* Number of pages ksmd should scan in one batch */
283 static unsigned int ksm_thread_pages_to_scan = DEFAULT_PAGES_TO_SCAN;
285 /* Milliseconds ksmd should sleep between batches */
286 static unsigned int ksm_thread_sleep_millisecs = 20;
288 /* Checksum of an empty (zeroed) page */
289 static unsigned int zero_checksum __read_mostly;
291 /* Whether to merge empty (zeroed) pages with actual zero pages */
292 static bool ksm_use_zero_pages __read_mostly;
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;
298 /* The number of zero pages which is placed by KSM */
299 unsigned long ksm_zero_pages;
301 /* The number of pages that have been skipped due to "smart scanning" */
302 static unsigned long ksm_pages_skipped;
304 /* Don't scan more than max pages per batch. */
305 static unsigned long ksm_advisor_max_pages_to_scan = 30000;
307 /* Min CPU for scanning pages per scan */
308 #define KSM_ADVISOR_MIN_CPU 10
310 /* Max CPU for scanning pages per scan */
311 static unsigned int ksm_advisor_max_cpu = 70;
313 /* Target scan time in seconds to analyze all KSM candidate pages. */
314 static unsigned long ksm_advisor_target_scan_time = 200;
316 /* Exponentially weighted moving average. */
317 #define EWMA_WEIGHT 30
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
328 unsigned long scan_time;
329 unsigned long change;
330 unsigned long long cpu_time;
332 static struct advisor_ctx advisor_ctx;
334 /* Define different advisor's */
335 enum ksm_advisor_type {
337 KSM_ADVISOR_SCAN_TIME,
339 static enum ksm_advisor_type ksm_advisor;
343 * Only called through the sysfs control interface:
346 /* At least scan this many pages per batch. */
347 static unsigned long ksm_advisor_min_pages_to_scan = 500;
349 static void set_advisor_defaults(void)
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;
358 #endif /* CONFIG_SYSFS */
360 static inline void advisor_start_scan(void)
362 if (ksm_advisor == KSM_ADVISOR_SCAN_TIME)
363 advisor_ctx.start_scan = ktime_get();
367 * Use previous scan time if available, otherwise use current scan time as an
368 * approximation for the previous scan time.
370 static inline unsigned long prev_scan_time(struct advisor_ctx *ctx,
371 unsigned long scan_time)
373 return ctx->scan_time ? ctx->scan_time : scan_time;
376 /* Calculate exponential weighted moving average */
377 static unsigned long ewma(unsigned long prev, unsigned long curr)
379 return ((100 - EWMA_WEIGHT) * prev + EWMA_WEIGHT * curr) / 100;
383 * The scan time advisor is based on the current scan rate and the target
386 * new_pages_to_scan = pages_to_scan * (scan_time / target_scan_time)
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:
393 * new_pages_to_scan *= change facor
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.
400 * In addition the new pages_to_scan value is capped by the max and min
403 static void scan_time_advisor(void)
405 unsigned int cpu_percent;
406 unsigned long cpu_time;
407 unsigned long cpu_time_diff;
408 unsigned long cpu_time_diff_ms;
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;
416 /* Convert scan time to seconds */
417 scan_time = div_s64(ktime_ms_delta(ktime_get(), advisor_ctx.start_scan),
419 scan_time = scan_time ? scan_time : 1;
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;
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);
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;
435 * Calculate scan time as percentage of last scan time and use
436 * exponentially weighted average to smooth it
438 change = scan_time * 100 / last_scan_time;
439 change = change ? change : 1;
440 change = ewma(advisor_ctx.change, change);
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;
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;
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);
455 /* Update advisor context */
456 advisor_ctx.change = change;
457 advisor_ctx.scan_time = scan_time;
458 advisor_ctx.cpu_time = cpu_time;
460 ksm_thread_pages_to_scan = pages;
461 trace_ksm_advisor(scan_time, pages, cpu_percent);
464 static void advisor_stop_scan(void)
466 if (ksm_advisor == KSM_ADVISOR_SCAN_TIME)
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;
475 #define ksm_merge_across_nodes 1U
476 #define ksm_nr_node_ids 1
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);
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);
491 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create(#__struct,\
492 sizeof(struct __struct), __alignof__(struct __struct),\
495 static int __init ksm_slab_init(void)
497 rmap_item_cache = KSM_KMEM_CACHE(ksm_rmap_item, 0);
498 if (!rmap_item_cache)
501 stable_node_cache = KSM_KMEM_CACHE(ksm_stable_node, 0);
502 if (!stable_node_cache)
505 mm_slot_cache = KSM_KMEM_CACHE(ksm_mm_slot, 0);
512 kmem_cache_destroy(stable_node_cache);
514 kmem_cache_destroy(rmap_item_cache);
519 static void __init ksm_slab_free(void)
521 kmem_cache_destroy(mm_slot_cache);
522 kmem_cache_destroy(stable_node_cache);
523 kmem_cache_destroy(rmap_item_cache);
524 mm_slot_cache = NULL;
527 static __always_inline bool is_stable_node_chain(struct ksm_stable_node *chain)
529 return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
532 static __always_inline bool is_stable_node_dup(struct ksm_stable_node *dup)
534 return dup->head == STABLE_NODE_DUP_HEAD;
537 static inline void stable_node_chain_add_dup(struct ksm_stable_node *dup,
538 struct ksm_stable_node *chain)
540 VM_BUG_ON(is_stable_node_dup(dup));
541 dup->head = STABLE_NODE_DUP_HEAD;
542 VM_BUG_ON(!is_stable_node_chain(chain));
543 hlist_add_head(&dup->hlist_dup, &chain->hlist);
544 ksm_stable_node_dups++;
547 static inline void __stable_node_dup_del(struct ksm_stable_node *dup)
549 VM_BUG_ON(!is_stable_node_dup(dup));
550 hlist_del(&dup->hlist_dup);
551 ksm_stable_node_dups--;
554 static inline void stable_node_dup_del(struct ksm_stable_node *dup)
556 VM_BUG_ON(is_stable_node_chain(dup));
557 if (is_stable_node_dup(dup))
558 __stable_node_dup_del(dup);
560 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
561 #ifdef CONFIG_DEBUG_VM
566 static inline struct ksm_rmap_item *alloc_rmap_item(void)
568 struct ksm_rmap_item *rmap_item;
570 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
571 __GFP_NORETRY | __GFP_NOWARN);
577 static inline void free_rmap_item(struct ksm_rmap_item *rmap_item)
580 rmap_item->mm->ksm_rmap_items--;
581 rmap_item->mm = NULL; /* debug safety */
582 kmem_cache_free(rmap_item_cache, rmap_item);
585 static inline struct ksm_stable_node *alloc_stable_node(void)
588 * The allocation can take too long with GFP_KERNEL when memory is under
589 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
590 * grants access to memory reserves, helping to avoid this problem.
592 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
595 static inline void free_stable_node(struct ksm_stable_node *stable_node)
597 VM_BUG_ON(stable_node->rmap_hlist_len &&
598 !is_stable_node_chain(stable_node));
599 kmem_cache_free(stable_node_cache, stable_node);
603 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
604 * page tables after it has passed through ksm_exit() - which, if necessary,
605 * takes mmap_lock briefly to serialize against them. ksm_exit() does not set
606 * a special flag: they can just back out as soon as mm_users goes to zero.
607 * ksm_test_exit() is used throughout to make this test for exit: in some
608 * places for correctness, in some places just to avoid unnecessary work.
610 static inline bool ksm_test_exit(struct mm_struct *mm)
612 return atomic_read(&mm->mm_users) == 0;
615 static int break_ksm_pmd_entry(pmd_t *pmd, unsigned long addr, unsigned long next,
616 struct mm_walk *walk)
618 struct page *page = NULL;
624 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
627 ptent = ptep_get(pte);
628 if (pte_present(ptent)) {
629 page = vm_normal_page(walk->vma, addr, ptent);
630 } else if (!pte_none(ptent)) {
631 swp_entry_t entry = pte_to_swp_entry(ptent);
634 * As KSM pages remain KSM pages until freed, no need to wait
635 * here for migration to end.
637 if (is_migration_entry(entry))
638 page = pfn_swap_entry_to_page(entry);
640 /* return 1 if the page is an normal ksm page or KSM-placed zero page */
641 ret = (page && PageKsm(page)) || is_ksm_zero_pte(ptent);
642 pte_unmap_unlock(pte, ptl);
646 static const struct mm_walk_ops break_ksm_ops = {
647 .pmd_entry = break_ksm_pmd_entry,
648 .walk_lock = PGWALK_RDLOCK,
651 static const struct mm_walk_ops break_ksm_lock_vma_ops = {
652 .pmd_entry = break_ksm_pmd_entry,
653 .walk_lock = PGWALK_WRLOCK,
657 * We use break_ksm to break COW on a ksm page by triggering unsharing,
658 * such that the ksm page will get replaced by an exclusive anonymous page.
660 * We take great care only to touch a ksm page, in a VM_MERGEABLE vma,
661 * in case the application has unmapped and remapped mm,addr meanwhile.
662 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
663 * mmap of /dev/mem, where we would not want to touch it.
665 * FAULT_FLAG_REMOTE/FOLL_REMOTE are because we do this outside the context
666 * of the process that owns 'vma'. We also do not want to enforce
667 * protection keys here anyway.
669 static int break_ksm(struct vm_area_struct *vma, unsigned long addr, bool lock_vma)
672 const struct mm_walk_ops *ops = lock_vma ?
673 &break_ksm_lock_vma_ops : &break_ksm_ops;
679 ksm_page = walk_page_range_vma(vma, addr, addr + 1, ops, NULL);
680 if (WARN_ON_ONCE(ksm_page < 0))
684 ret = handle_mm_fault(vma, addr,
685 FAULT_FLAG_UNSHARE | FAULT_FLAG_REMOTE,
687 } while (!(ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
689 * We must loop until we no longer find a KSM page because
690 * handle_mm_fault() may back out if there's any difficulty e.g. if
691 * pte accessed bit gets updated concurrently.
693 * VM_FAULT_SIGBUS could occur if we race with truncation of the
694 * backing file, which also invalidates anonymous pages: that's
695 * okay, that truncation will have unmapped the PageKsm for us.
697 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
698 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
699 * current task has TIF_MEMDIE set, and will be OOM killed on return
700 * to user; and ksmd, having no mm, would never be chosen for that.
702 * But if the mm is in a limited mem_cgroup, then the fault may fail
703 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
704 * even ksmd can fail in this way - though it's usually breaking ksm
705 * just to undo a merge it made a moment before, so unlikely to oom.
707 * That's a pity: we might therefore have more kernel pages allocated
708 * than we're counting as nodes in the stable tree; but ksm_do_scan
709 * will retry to break_cow on each pass, so should recover the page
710 * in due course. The important thing is to not let VM_MERGEABLE
711 * be cleared while any such pages might remain in the area.
713 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
716 static bool vma_ksm_compatible(struct vm_area_struct *vma)
718 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE | VM_PFNMAP |
719 VM_IO | VM_DONTEXPAND | VM_HUGETLB |
721 return false; /* just ignore the advice */
727 if (vma->vm_flags & VM_SAO)
731 if (vma->vm_flags & VM_SPARC_ADI)
738 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
741 struct vm_area_struct *vma;
742 if (ksm_test_exit(mm))
744 vma = vma_lookup(mm, addr);
745 if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
750 static void break_cow(struct ksm_rmap_item *rmap_item)
752 struct mm_struct *mm = rmap_item->mm;
753 unsigned long addr = rmap_item->address;
754 struct vm_area_struct *vma;
757 * It is not an accident that whenever we want to break COW
758 * to undo, we also need to drop a reference to the anon_vma.
760 put_anon_vma(rmap_item->anon_vma);
763 vma = find_mergeable_vma(mm, addr);
765 break_ksm(vma, addr, false);
766 mmap_read_unlock(mm);
769 static struct page *get_mergeable_page(struct ksm_rmap_item *rmap_item)
771 struct mm_struct *mm = rmap_item->mm;
772 unsigned long addr = rmap_item->address;
773 struct vm_area_struct *vma;
777 vma = find_mergeable_vma(mm, addr);
781 page = follow_page(vma, addr, FOLL_GET);
782 if (IS_ERR_OR_NULL(page))
784 if (is_zone_device_page(page))
786 if (PageAnon(page)) {
787 flush_anon_page(vma, page, addr);
788 flush_dcache_page(page);
795 mmap_read_unlock(mm);
800 * This helper is used for getting right index into array of tree roots.
801 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
802 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
803 * every node has its own stable and unstable tree.
805 static inline int get_kpfn_nid(unsigned long kpfn)
807 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
810 static struct ksm_stable_node *alloc_stable_node_chain(struct ksm_stable_node *dup,
811 struct rb_root *root)
813 struct ksm_stable_node *chain = alloc_stable_node();
814 VM_BUG_ON(is_stable_node_chain(dup));
816 INIT_HLIST_HEAD(&chain->hlist);
817 chain->chain_prune_time = jiffies;
818 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
819 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
820 chain->nid = NUMA_NO_NODE; /* debug */
822 ksm_stable_node_chains++;
825 * Put the stable node chain in the first dimension of
826 * the stable tree and at the same time remove the old
829 rb_replace_node(&dup->node, &chain->node, root);
832 * Move the old stable node to the second dimension
833 * queued in the hlist_dup. The invariant is that all
834 * dup stable_nodes in the chain->hlist point to pages
835 * that are write protected and have the exact same
838 stable_node_chain_add_dup(dup, chain);
843 static inline void free_stable_node_chain(struct ksm_stable_node *chain,
844 struct rb_root *root)
846 rb_erase(&chain->node, root);
847 free_stable_node(chain);
848 ksm_stable_node_chains--;
851 static void remove_node_from_stable_tree(struct ksm_stable_node *stable_node)
853 struct ksm_rmap_item *rmap_item;
855 /* check it's not STABLE_NODE_CHAIN or negative */
856 BUG_ON(stable_node->rmap_hlist_len < 0);
858 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
859 if (rmap_item->hlist.next) {
861 trace_ksm_remove_rmap_item(stable_node->kpfn, rmap_item, rmap_item->mm);
866 rmap_item->mm->ksm_merging_pages--;
868 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
869 stable_node->rmap_hlist_len--;
870 put_anon_vma(rmap_item->anon_vma);
871 rmap_item->address &= PAGE_MASK;
876 * We need the second aligned pointer of the migrate_nodes
877 * list_head to stay clear from the rb_parent_color union
878 * (aligned and different than any node) and also different
879 * from &migrate_nodes. This will verify that future list.h changes
880 * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it.
882 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
883 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
885 trace_ksm_remove_ksm_page(stable_node->kpfn);
886 if (stable_node->head == &migrate_nodes)
887 list_del(&stable_node->list);
889 stable_node_dup_del(stable_node);
890 free_stable_node(stable_node);
893 enum get_ksm_page_flags {
900 * ksm_get_folio: checks if the page indicated by the stable node
901 * is still its ksm page, despite having held no reference to it.
902 * In which case we can trust the content of the page, and it
903 * returns the gotten page; but if the page has now been zapped,
904 * remove the stale node from the stable tree and return NULL.
905 * But beware, the stable node's page might be being migrated.
907 * You would expect the stable_node to hold a reference to the ksm page.
908 * But if it increments the page's count, swapping out has to wait for
909 * ksmd to come around again before it can free the page, which may take
910 * seconds or even minutes: much too unresponsive. So instead we use a
911 * "keyhole reference": access to the ksm page from the stable node peeps
912 * out through its keyhole to see if that page still holds the right key,
913 * pointing back to this stable node. This relies on freeing a PageAnon
914 * page to reset its page->mapping to NULL, and relies on no other use of
915 * a page to put something that might look like our key in page->mapping.
916 * is on its way to being freed; but it is an anomaly to bear in mind.
918 static struct folio *ksm_get_folio(struct ksm_stable_node *stable_node,
919 enum get_ksm_page_flags flags)
922 void *expected_mapping;
925 expected_mapping = (void *)((unsigned long)stable_node |
928 kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
929 folio = pfn_folio(kpfn);
930 if (READ_ONCE(folio->mapping) != expected_mapping)
934 * We cannot do anything with the page while its refcount is 0.
935 * Usually 0 means free, or tail of a higher-order page: in which
936 * case this node is no longer referenced, and should be freed;
937 * however, it might mean that the page is under page_ref_freeze().
938 * The __remove_mapping() case is easy, again the node is now stale;
939 * the same is in reuse_ksm_page() case; but if page is swapcache
940 * in folio_migrate_mapping(), it might still be our page,
941 * in which case it's essential to keep the node.
943 while (!folio_try_get(folio)) {
945 * Another check for page->mapping != expected_mapping would
946 * work here too. We have chosen the !PageSwapCache test to
947 * optimize the common case, when the page is or is about to
948 * be freed: PageSwapCache is cleared (under spin_lock_irq)
949 * in the ref_freeze section of __remove_mapping(); but Anon
950 * folio->mapping reset to NULL later, in free_pages_prepare().
952 if (!folio_test_swapcache(folio))
957 if (READ_ONCE(folio->mapping) != expected_mapping) {
962 if (flags == GET_KSM_PAGE_TRYLOCK) {
963 if (!folio_trylock(folio)) {
965 return ERR_PTR(-EBUSY);
967 } else if (flags == GET_KSM_PAGE_LOCK)
970 if (flags != GET_KSM_PAGE_NOLOCK) {
971 if (READ_ONCE(folio->mapping) != expected_mapping) {
981 * We come here from above when page->mapping or !PageSwapCache
982 * suggests that the node is stale; but it might be under migration.
983 * We need smp_rmb(), matching the smp_wmb() in folio_migrate_ksm(),
984 * before checking whether node->kpfn has been changed.
987 if (READ_ONCE(stable_node->kpfn) != kpfn)
989 remove_node_from_stable_tree(stable_node);
993 static struct page *get_ksm_page(struct ksm_stable_node *stable_node,
994 enum get_ksm_page_flags flags)
996 struct folio *folio = ksm_get_folio(stable_node, flags);
1002 * Removing rmap_item from stable or unstable tree.
1003 * This function will clean the information from the stable/unstable tree.
1005 static void remove_rmap_item_from_tree(struct ksm_rmap_item *rmap_item)
1007 if (rmap_item->address & STABLE_FLAG) {
1008 struct ksm_stable_node *stable_node;
1009 struct folio *folio;
1011 stable_node = rmap_item->head;
1012 folio = ksm_get_folio(stable_node, GET_KSM_PAGE_LOCK);
1016 hlist_del(&rmap_item->hlist);
1017 folio_unlock(folio);
1020 if (!hlist_empty(&stable_node->hlist))
1021 ksm_pages_sharing--;
1025 rmap_item->mm->ksm_merging_pages--;
1027 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
1028 stable_node->rmap_hlist_len--;
1030 put_anon_vma(rmap_item->anon_vma);
1031 rmap_item->head = NULL;
1032 rmap_item->address &= PAGE_MASK;
1034 } else if (rmap_item->address & UNSTABLE_FLAG) {
1037 * Usually ksmd can and must skip the rb_erase, because
1038 * root_unstable_tree was already reset to RB_ROOT.
1039 * But be careful when an mm is exiting: do the rb_erase
1040 * if this rmap_item was inserted by this scan, rather
1041 * than left over from before.
1043 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
1046 rb_erase(&rmap_item->node,
1047 root_unstable_tree + NUMA(rmap_item->nid));
1048 ksm_pages_unshared--;
1049 rmap_item->address &= PAGE_MASK;
1052 cond_resched(); /* we're called from many long loops */
1055 static void remove_trailing_rmap_items(struct ksm_rmap_item **rmap_list)
1057 while (*rmap_list) {
1058 struct ksm_rmap_item *rmap_item = *rmap_list;
1059 *rmap_list = rmap_item->rmap_list;
1060 remove_rmap_item_from_tree(rmap_item);
1061 free_rmap_item(rmap_item);
1066 * Though it's very tempting to unmerge rmap_items from stable tree rather
1067 * than check every pte of a given vma, the locking doesn't quite work for
1068 * that - an rmap_item is assigned to the stable tree after inserting ksm
1069 * page and upping mmap_lock. Nor does it fit with the way we skip dup'ing
1070 * rmap_items from parent to child at fork time (so as not to waste time
1071 * if exit comes before the next scan reaches it).
1073 * Similarly, although we'd like to remove rmap_items (so updating counts
1074 * and freeing memory) when unmerging an area, it's easier to leave that
1075 * to the next pass of ksmd - consider, for example, how ksmd might be
1076 * in cmp_and_merge_page on one of the rmap_items we would be removing.
1078 static int unmerge_ksm_pages(struct vm_area_struct *vma,
1079 unsigned long start, unsigned long end, bool lock_vma)
1084 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
1085 if (ksm_test_exit(vma->vm_mm))
1087 if (signal_pending(current))
1090 err = break_ksm(vma, addr, lock_vma);
1095 static inline struct ksm_stable_node *folio_stable_node(struct folio *folio)
1097 return folio_test_ksm(folio) ? folio_raw_mapping(folio) : NULL;
1100 static inline struct ksm_stable_node *page_stable_node(struct page *page)
1102 return folio_stable_node(page_folio(page));
1105 static inline void set_page_stable_node(struct page *page,
1106 struct ksm_stable_node *stable_node)
1108 VM_BUG_ON_PAGE(PageAnon(page) && PageAnonExclusive(page), page);
1109 page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
1112 static inline void folio_set_stable_node(struct folio *folio,
1113 struct ksm_stable_node *stable_node)
1115 set_page_stable_node(&folio->page, stable_node);
1120 * Only called through the sysfs control interface:
1122 static int remove_stable_node(struct ksm_stable_node *stable_node)
1124 struct folio *folio;
1127 folio = ksm_get_folio(stable_node, GET_KSM_PAGE_LOCK);
1130 * ksm_get_folio did remove_node_from_stable_tree itself.
1136 * Page could be still mapped if this races with __mmput() running in
1137 * between ksm_exit() and exit_mmap(). Just refuse to let
1138 * merge_across_nodes/max_page_sharing be switched.
1141 if (!folio_mapped(folio)) {
1143 * The stable node did not yet appear stale to ksm_get_folio(),
1144 * since that allows for an unmapped ksm folio to be recognized
1145 * right up until it is freed; but the node is safe to remove.
1146 * This folio might be in an LRU cache waiting to be freed,
1147 * or it might be in the swapcache (perhaps under writeback),
1148 * or it might have been removed from swapcache a moment ago.
1150 folio_set_stable_node(folio, NULL);
1151 remove_node_from_stable_tree(stable_node);
1155 folio_unlock(folio);
1160 static int remove_stable_node_chain(struct ksm_stable_node *stable_node,
1161 struct rb_root *root)
1163 struct ksm_stable_node *dup;
1164 struct hlist_node *hlist_safe;
1166 if (!is_stable_node_chain(stable_node)) {
1167 VM_BUG_ON(is_stable_node_dup(stable_node));
1168 if (remove_stable_node(stable_node))
1174 hlist_for_each_entry_safe(dup, hlist_safe,
1175 &stable_node->hlist, hlist_dup) {
1176 VM_BUG_ON(!is_stable_node_dup(dup));
1177 if (remove_stable_node(dup))
1180 BUG_ON(!hlist_empty(&stable_node->hlist));
1181 free_stable_node_chain(stable_node, root);
1185 static int remove_all_stable_nodes(void)
1187 struct ksm_stable_node *stable_node, *next;
1191 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
1192 while (root_stable_tree[nid].rb_node) {
1193 stable_node = rb_entry(root_stable_tree[nid].rb_node,
1194 struct ksm_stable_node, node);
1195 if (remove_stable_node_chain(stable_node,
1196 root_stable_tree + nid)) {
1198 break; /* proceed to next nid */
1203 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
1204 if (remove_stable_node(stable_node))
1211 static int unmerge_and_remove_all_rmap_items(void)
1213 struct ksm_mm_slot *mm_slot;
1214 struct mm_slot *slot;
1215 struct mm_struct *mm;
1216 struct vm_area_struct *vma;
1219 spin_lock(&ksm_mmlist_lock);
1220 slot = list_entry(ksm_mm_head.slot.mm_node.next,
1221 struct mm_slot, mm_node);
1222 ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
1223 spin_unlock(&ksm_mmlist_lock);
1225 for (mm_slot = ksm_scan.mm_slot; mm_slot != &ksm_mm_head;
1226 mm_slot = ksm_scan.mm_slot) {
1227 VMA_ITERATOR(vmi, mm_slot->slot.mm, 0);
1229 mm = mm_slot->slot.mm;
1233 * Exit right away if mm is exiting to avoid lockdep issue in
1236 if (ksm_test_exit(mm))
1239 for_each_vma(vmi, vma) {
1240 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
1242 err = unmerge_ksm_pages(vma,
1243 vma->vm_start, vma->vm_end, false);
1249 remove_trailing_rmap_items(&mm_slot->rmap_list);
1250 mmap_read_unlock(mm);
1252 spin_lock(&ksm_mmlist_lock);
1253 slot = list_entry(mm_slot->slot.mm_node.next,
1254 struct mm_slot, mm_node);
1255 ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
1256 if (ksm_test_exit(mm)) {
1257 hash_del(&mm_slot->slot.hash);
1258 list_del(&mm_slot->slot.mm_node);
1259 spin_unlock(&ksm_mmlist_lock);
1261 mm_slot_free(mm_slot_cache, mm_slot);
1262 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1263 clear_bit(MMF_VM_MERGE_ANY, &mm->flags);
1266 spin_unlock(&ksm_mmlist_lock);
1269 /* Clean up stable nodes, but don't worry if some are still busy */
1270 remove_all_stable_nodes();
1275 mmap_read_unlock(mm);
1276 spin_lock(&ksm_mmlist_lock);
1277 ksm_scan.mm_slot = &ksm_mm_head;
1278 spin_unlock(&ksm_mmlist_lock);
1281 #endif /* CONFIG_SYSFS */
1283 static u32 calc_checksum(struct page *page)
1286 void *addr = kmap_local_page(page);
1287 checksum = xxhash(addr, PAGE_SIZE, 0);
1292 static int write_protect_page(struct vm_area_struct *vma, struct folio *folio,
1295 struct mm_struct *mm = vma->vm_mm;
1296 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, 0, 0);
1299 struct mmu_notifier_range range;
1300 bool anon_exclusive;
1303 if (WARN_ON_ONCE(folio_test_large(folio)))
1306 pvmw.address = page_address_in_vma(&folio->page, vma);
1307 if (pvmw.address == -EFAULT)
1310 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, pvmw.address,
1311 pvmw.address + PAGE_SIZE);
1312 mmu_notifier_invalidate_range_start(&range);
1314 if (!page_vma_mapped_walk(&pvmw))
1316 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1319 anon_exclusive = PageAnonExclusive(&folio->page);
1320 entry = ptep_get(pvmw.pte);
1321 if (pte_write(entry) || pte_dirty(entry) ||
1322 anon_exclusive || mm_tlb_flush_pending(mm)) {
1323 swapped = folio_test_swapcache(folio);
1324 flush_cache_page(vma, pvmw.address, folio_pfn(folio));
1326 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1327 * take any lock, therefore the check that we are going to make
1328 * with the pagecount against the mapcount is racy and
1329 * O_DIRECT can happen right after the check.
1330 * So we clear the pte and flush the tlb before the check
1331 * this assure us that no O_DIRECT can happen after the check
1332 * or in the middle of the check.
1334 * No need to notify as we are downgrading page table to read
1335 * only not changing it to point to a new page.
1337 * See Documentation/mm/mmu_notifier.rst
1339 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1341 * Check that no O_DIRECT or similar I/O is in progress on the
1344 if (folio_mapcount(folio) + 1 + swapped != folio_ref_count(folio)) {
1345 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1349 /* See folio_try_share_anon_rmap_pte(): clear PTE first. */
1350 if (anon_exclusive &&
1351 folio_try_share_anon_rmap_pte(folio, &folio->page)) {
1352 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1356 if (pte_dirty(entry))
1357 folio_mark_dirty(folio);
1358 entry = pte_mkclean(entry);
1360 if (pte_write(entry))
1361 entry = pte_wrprotect(entry);
1363 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1369 page_vma_mapped_walk_done(&pvmw);
1371 mmu_notifier_invalidate_range_end(&range);
1377 * replace_page - replace page in vma by new ksm page
1378 * @vma: vma that holds the pte pointing to page
1379 * @page: the page we are replacing by kpage
1380 * @kpage: the ksm page we replace page by
1381 * @orig_pte: the original value of the pte
1383 * Returns 0 on success, -EFAULT on failure.
1385 static int replace_page(struct vm_area_struct *vma, struct page *page,
1386 struct page *kpage, pte_t orig_pte)
1388 struct folio *kfolio = page_folio(kpage);
1389 struct mm_struct *mm = vma->vm_mm;
1390 struct folio *folio;
1398 struct mmu_notifier_range range;
1400 addr = page_address_in_vma(page, vma);
1401 if (addr == -EFAULT)
1404 pmd = mm_find_pmd(mm, addr);
1408 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
1409 * without holding anon_vma lock for write. So when looking for a
1410 * genuine pmde (in which to find pte), test present and !THP together.
1412 pmde = pmdp_get_lockless(pmd);
1413 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
1416 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, addr,
1418 mmu_notifier_invalidate_range_start(&range);
1420 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1423 if (!pte_same(ptep_get(ptep), orig_pte)) {
1424 pte_unmap_unlock(ptep, ptl);
1427 VM_BUG_ON_PAGE(PageAnonExclusive(page), page);
1428 VM_BUG_ON_FOLIO(folio_test_anon(kfolio) && PageAnonExclusive(kpage),
1432 * No need to check ksm_use_zero_pages here: we can only have a
1433 * zero_page here if ksm_use_zero_pages was enabled already.
1435 if (!is_zero_pfn(page_to_pfn(kpage))) {
1437 folio_add_anon_rmap_pte(kfolio, kpage, vma, addr, RMAP_NONE);
1438 newpte = mk_pte(kpage, vma->vm_page_prot);
1441 * Use pte_mkdirty to mark the zero page mapped by KSM, and then
1442 * we can easily track all KSM-placed zero pages by checking if
1443 * the dirty bit in zero page's PTE is set.
1445 newpte = pte_mkdirty(pte_mkspecial(pfn_pte(page_to_pfn(kpage), vma->vm_page_prot)));
1447 mm->ksm_zero_pages++;
1449 * We're replacing an anonymous page with a zero page, which is
1450 * not anonymous. We need to do proper accounting otherwise we
1451 * will get wrong values in /proc, and a BUG message in dmesg
1452 * when tearing down the mm.
1454 dec_mm_counter(mm, MM_ANONPAGES);
1457 flush_cache_page(vma, addr, pte_pfn(ptep_get(ptep)));
1459 * No need to notify as we are replacing a read only page with another
1460 * read only page with the same content.
1462 * See Documentation/mm/mmu_notifier.rst
1464 ptep_clear_flush(vma, addr, ptep);
1465 set_pte_at_notify(mm, addr, ptep, newpte);
1467 folio = page_folio(page);
1468 folio_remove_rmap_pte(folio, page, vma);
1469 if (!folio_mapped(folio))
1470 folio_free_swap(folio);
1473 pte_unmap_unlock(ptep, ptl);
1476 mmu_notifier_invalidate_range_end(&range);
1482 * try_to_merge_one_page - take two pages and merge them into one
1483 * @vma: the vma that holds the pte pointing to page
1484 * @page: the PageAnon page that we want to replace with kpage
1485 * @kpage: the PageKsm page that we want to map instead of page,
1486 * or NULL the first time when we want to use page as kpage.
1488 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1490 static int try_to_merge_one_page(struct vm_area_struct *vma,
1491 struct page *page, struct page *kpage)
1493 pte_t orig_pte = __pte(0);
1496 if (page == kpage) /* ksm page forked */
1499 if (!PageAnon(page))
1503 * We need the page lock to read a stable PageSwapCache in
1504 * write_protect_page(). We use trylock_page() instead of
1505 * lock_page() because we don't want to wait here - we
1506 * prefer to continue scanning and merging different pages,
1507 * then come back to this page when it is unlocked.
1509 if (!trylock_page(page))
1512 if (PageTransCompound(page)) {
1513 if (split_huge_page(page))
1518 * If this anonymous page is mapped only here, its pte may need
1519 * to be write-protected. If it's mapped elsewhere, all of its
1520 * ptes are necessarily already write-protected. But in either
1521 * case, we need to lock and check page_count is not raised.
1523 if (write_protect_page(vma, page_folio(page), &orig_pte) == 0) {
1526 * While we hold page lock, upgrade page from
1527 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1528 * stable_tree_insert() will update stable_node.
1530 set_page_stable_node(page, NULL);
1531 mark_page_accessed(page);
1533 * Page reclaim just frees a clean page with no dirty
1534 * ptes: make sure that the ksm page would be swapped.
1536 if (!PageDirty(page))
1539 } else if (pages_identical(page, kpage))
1540 err = replace_page(vma, page, kpage, orig_pte);
1550 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1551 * but no new kernel page is allocated: kpage must already be a ksm page.
1553 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1555 static int try_to_merge_with_ksm_page(struct ksm_rmap_item *rmap_item,
1556 struct page *page, struct page *kpage)
1558 struct mm_struct *mm = rmap_item->mm;
1559 struct vm_area_struct *vma;
1563 vma = find_mergeable_vma(mm, rmap_item->address);
1567 err = try_to_merge_one_page(vma, page, kpage);
1571 /* Unstable nid is in union with stable anon_vma: remove first */
1572 remove_rmap_item_from_tree(rmap_item);
1574 /* Must get reference to anon_vma while still holding mmap_lock */
1575 rmap_item->anon_vma = vma->anon_vma;
1576 get_anon_vma(vma->anon_vma);
1578 mmap_read_unlock(mm);
1579 trace_ksm_merge_with_ksm_page(kpage, page_to_pfn(kpage ? kpage : page),
1580 rmap_item, mm, err);
1585 * try_to_merge_two_pages - take two identical pages and prepare them
1586 * to be merged into one page.
1588 * This function returns the kpage if we successfully merged two identical
1589 * pages into one ksm page, NULL otherwise.
1591 * Note that this function upgrades page to ksm page: if one of the pages
1592 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1594 static struct page *try_to_merge_two_pages(struct ksm_rmap_item *rmap_item,
1596 struct ksm_rmap_item *tree_rmap_item,
1597 struct page *tree_page)
1601 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1603 err = try_to_merge_with_ksm_page(tree_rmap_item,
1606 * If that fails, we have a ksm page with only one pte
1607 * pointing to it: so break it.
1610 break_cow(rmap_item);
1612 return err ? NULL : page;
1615 static __always_inline
1616 bool __is_page_sharing_candidate(struct ksm_stable_node *stable_node, int offset)
1618 VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1620 * Check that at least one mapping still exists, otherwise
1621 * there's no much point to merge and share with this
1622 * stable_node, as the underlying tree_page of the other
1623 * sharer is going to be freed soon.
1625 return stable_node->rmap_hlist_len &&
1626 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1629 static __always_inline
1630 bool is_page_sharing_candidate(struct ksm_stable_node *stable_node)
1632 return __is_page_sharing_candidate(stable_node, 0);
1635 static struct page *stable_node_dup(struct ksm_stable_node **_stable_node_dup,
1636 struct ksm_stable_node **_stable_node,
1637 struct rb_root *root,
1638 bool prune_stale_stable_nodes)
1640 struct ksm_stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1641 struct hlist_node *hlist_safe;
1642 struct folio *folio, *tree_folio = NULL;
1644 int found_rmap_hlist_len;
1646 if (!prune_stale_stable_nodes ||
1647 time_before(jiffies, stable_node->chain_prune_time +
1649 ksm_stable_node_chains_prune_millisecs)))
1650 prune_stale_stable_nodes = false;
1652 stable_node->chain_prune_time = jiffies;
1654 hlist_for_each_entry_safe(dup, hlist_safe,
1655 &stable_node->hlist, hlist_dup) {
1658 * We must walk all stable_node_dup to prune the stale
1659 * stable nodes during lookup.
1661 * ksm_get_folio can drop the nodes from the
1662 * stable_node->hlist if they point to freed pages
1663 * (that's why we do a _safe walk). The "dup"
1664 * stable_node parameter itself will be freed from
1665 * under us if it returns NULL.
1667 folio = ksm_get_folio(dup, GET_KSM_PAGE_NOLOCK);
1671 if (is_page_sharing_candidate(dup)) {
1673 dup->rmap_hlist_len > found_rmap_hlist_len) {
1675 folio_put(tree_folio);
1677 found_rmap_hlist_len = found->rmap_hlist_len;
1680 /* skip put_page for found dup */
1681 if (!prune_stale_stable_nodes)
1691 * nr is counting all dups in the chain only if
1692 * prune_stale_stable_nodes is true, otherwise we may
1693 * break the loop at nr == 1 even if there are
1696 if (prune_stale_stable_nodes && nr == 1) {
1698 * If there's not just one entry it would
1699 * corrupt memory, better BUG_ON. In KSM
1700 * context with no lock held it's not even
1703 BUG_ON(stable_node->hlist.first->next);
1706 * There's just one entry and it is below the
1707 * deduplication limit so drop the chain.
1709 rb_replace_node(&stable_node->node, &found->node,
1711 free_stable_node(stable_node);
1712 ksm_stable_node_chains--;
1713 ksm_stable_node_dups--;
1715 * NOTE: the caller depends on the stable_node
1716 * to be equal to stable_node_dup if the chain
1719 *_stable_node = found;
1721 * Just for robustness, as stable_node is
1722 * otherwise left as a stable pointer, the
1723 * compiler shall optimize it away at build
1727 } else if (stable_node->hlist.first != &found->hlist_dup &&
1728 __is_page_sharing_candidate(found, 1)) {
1730 * If the found stable_node dup can accept one
1731 * more future merge (in addition to the one
1732 * that is underway) and is not at the head of
1733 * the chain, put it there so next search will
1734 * be quicker in the !prune_stale_stable_nodes
1737 * NOTE: it would be inaccurate to use nr > 1
1738 * instead of checking the hlist.first pointer
1739 * directly, because in the
1740 * prune_stale_stable_nodes case "nr" isn't
1741 * the position of the found dup in the chain,
1742 * but the total number of dups in the chain.
1744 hlist_del(&found->hlist_dup);
1745 hlist_add_head(&found->hlist_dup,
1746 &stable_node->hlist);
1750 *_stable_node_dup = found;
1751 return &tree_folio->page;
1754 static struct ksm_stable_node *stable_node_dup_any(struct ksm_stable_node *stable_node,
1755 struct rb_root *root)
1757 if (!is_stable_node_chain(stable_node))
1759 if (hlist_empty(&stable_node->hlist)) {
1760 free_stable_node_chain(stable_node, root);
1763 return hlist_entry(stable_node->hlist.first,
1764 typeof(*stable_node), hlist_dup);
1768 * Like for get_ksm_page, this function can free the *_stable_node and
1769 * *_stable_node_dup if the returned tree_page is NULL.
1771 * It can also free and overwrite *_stable_node with the found
1772 * stable_node_dup if the chain is collapsed (in which case
1773 * *_stable_node will be equal to *_stable_node_dup like if the chain
1774 * never existed). It's up to the caller to verify tree_page is not
1775 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1777 * *_stable_node_dup is really a second output parameter of this
1778 * function and will be overwritten in all cases, the caller doesn't
1779 * need to initialize it.
1781 static struct page *__stable_node_chain(struct ksm_stable_node **_stable_node_dup,
1782 struct ksm_stable_node **_stable_node,
1783 struct rb_root *root,
1784 bool prune_stale_stable_nodes)
1786 struct ksm_stable_node *stable_node = *_stable_node;
1787 if (!is_stable_node_chain(stable_node)) {
1788 if (is_page_sharing_candidate(stable_node)) {
1789 *_stable_node_dup = stable_node;
1790 return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK);
1793 * _stable_node_dup set to NULL means the stable_node
1794 * reached the ksm_max_page_sharing limit.
1796 *_stable_node_dup = NULL;
1799 return stable_node_dup(_stable_node_dup, _stable_node, root,
1800 prune_stale_stable_nodes);
1803 static __always_inline struct page *chain_prune(struct ksm_stable_node **s_n_d,
1804 struct ksm_stable_node **s_n,
1805 struct rb_root *root)
1807 return __stable_node_chain(s_n_d, s_n, root, true);
1810 static __always_inline struct page *chain(struct ksm_stable_node **s_n_d,
1811 struct ksm_stable_node *s_n,
1812 struct rb_root *root)
1814 struct ksm_stable_node *old_stable_node = s_n;
1815 struct page *tree_page;
1817 tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1818 /* not pruning dups so s_n cannot have changed */
1819 VM_BUG_ON(s_n != old_stable_node);
1824 * stable_tree_search - search for page inside the stable tree
1826 * This function checks if there is a page inside the stable tree
1827 * with identical content to the page that we are scanning right now.
1829 * This function returns the stable tree node of identical content if found,
1832 static struct page *stable_tree_search(struct page *page)
1835 struct rb_root *root;
1836 struct rb_node **new;
1837 struct rb_node *parent;
1838 struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any;
1839 struct ksm_stable_node *page_node;
1841 page_node = page_stable_node(page);
1842 if (page_node && page_node->head != &migrate_nodes) {
1843 /* ksm page forked */
1848 nid = get_kpfn_nid(page_to_pfn(page));
1849 root = root_stable_tree + nid;
1851 new = &root->rb_node;
1855 struct page *tree_page;
1859 stable_node = rb_entry(*new, struct ksm_stable_node, node);
1860 stable_node_any = NULL;
1861 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1863 * NOTE: stable_node may have been freed by
1864 * chain_prune() if the returned stable_node_dup is
1865 * not NULL. stable_node_dup may have been inserted in
1866 * the rbtree instead as a regular stable_node (in
1867 * order to collapse the stable_node chain if a single
1868 * stable_node dup was found in it). In such case the
1869 * stable_node is overwritten by the callee to point
1870 * to the stable_node_dup that was collapsed in the
1871 * stable rbtree and stable_node will be equal to
1872 * stable_node_dup like if the chain never existed.
1874 if (!stable_node_dup) {
1876 * Either all stable_node dups were full in
1877 * this stable_node chain, or this chain was
1878 * empty and should be rb_erased.
1880 stable_node_any = stable_node_dup_any(stable_node,
1882 if (!stable_node_any) {
1883 /* rb_erase just run */
1887 * Take any of the stable_node dups page of
1888 * this stable_node chain to let the tree walk
1889 * continue. All KSM pages belonging to the
1890 * stable_node dups in a stable_node chain
1891 * have the same content and they're
1892 * write protected at all times. Any will work
1893 * fine to continue the walk.
1895 tree_page = get_ksm_page(stable_node_any,
1896 GET_KSM_PAGE_NOLOCK);
1898 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1901 * If we walked over a stale stable_node,
1902 * get_ksm_page() will call rb_erase() and it
1903 * may rebalance the tree from under us. So
1904 * restart the search from scratch. Returning
1905 * NULL would be safe too, but we'd generate
1906 * false negative insertions just because some
1907 * stable_node was stale.
1912 ret = memcmp_pages(page, tree_page);
1913 put_page(tree_page);
1917 new = &parent->rb_left;
1919 new = &parent->rb_right;
1922 VM_BUG_ON(page_node->head != &migrate_nodes);
1924 * Test if the migrated page should be merged
1925 * into a stable node dup. If the mapcount is
1926 * 1 we can migrate it with another KSM page
1927 * without adding it to the chain.
1929 if (page_mapcount(page) > 1)
1933 if (!stable_node_dup) {
1935 * If the stable_node is a chain and
1936 * we got a payload match in memcmp
1937 * but we cannot merge the scanned
1938 * page in any of the existing
1939 * stable_node dups because they're
1940 * all full, we need to wait the
1941 * scanned page to find itself a match
1942 * in the unstable tree to create a
1943 * brand new KSM page to add later to
1944 * the dups of this stable_node.
1950 * Lock and unlock the stable_node's page (which
1951 * might already have been migrated) so that page
1952 * migration is sure to notice its raised count.
1953 * It would be more elegant to return stable_node
1954 * than kpage, but that involves more changes.
1956 tree_page = get_ksm_page(stable_node_dup,
1957 GET_KSM_PAGE_TRYLOCK);
1959 if (PTR_ERR(tree_page) == -EBUSY)
1960 return ERR_PTR(-EBUSY);
1962 if (unlikely(!tree_page))
1964 * The tree may have been rebalanced,
1965 * so re-evaluate parent and new.
1968 unlock_page(tree_page);
1970 if (get_kpfn_nid(stable_node_dup->kpfn) !=
1971 NUMA(stable_node_dup->nid)) {
1972 put_page(tree_page);
1982 list_del(&page_node->list);
1983 DO_NUMA(page_node->nid = nid);
1984 rb_link_node(&page_node->node, parent, new);
1985 rb_insert_color(&page_node->node, root);
1987 if (is_page_sharing_candidate(page_node)) {
1995 * If stable_node was a chain and chain_prune collapsed it,
1996 * stable_node has been updated to be the new regular
1997 * stable_node. A collapse of the chain is indistinguishable
1998 * from the case there was no chain in the stable
1999 * rbtree. Otherwise stable_node is the chain and
2000 * stable_node_dup is the dup to replace.
2002 if (stable_node_dup == stable_node) {
2003 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
2004 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
2005 /* there is no chain */
2007 VM_BUG_ON(page_node->head != &migrate_nodes);
2008 list_del(&page_node->list);
2009 DO_NUMA(page_node->nid = nid);
2010 rb_replace_node(&stable_node_dup->node,
2013 if (is_page_sharing_candidate(page_node))
2018 rb_erase(&stable_node_dup->node, root);
2022 VM_BUG_ON(!is_stable_node_chain(stable_node));
2023 __stable_node_dup_del(stable_node_dup);
2025 VM_BUG_ON(page_node->head != &migrate_nodes);
2026 list_del(&page_node->list);
2027 DO_NUMA(page_node->nid = nid);
2028 stable_node_chain_add_dup(page_node, stable_node);
2029 if (is_page_sharing_candidate(page_node))
2037 stable_node_dup->head = &migrate_nodes;
2038 list_add(&stable_node_dup->list, stable_node_dup->head);
2042 /* stable_node_dup could be null if it reached the limit */
2043 if (!stable_node_dup)
2044 stable_node_dup = stable_node_any;
2046 * If stable_node was a chain and chain_prune collapsed it,
2047 * stable_node has been updated to be the new regular
2048 * stable_node. A collapse of the chain is indistinguishable
2049 * from the case there was no chain in the stable
2050 * rbtree. Otherwise stable_node is the chain and
2051 * stable_node_dup is the dup to replace.
2053 if (stable_node_dup == stable_node) {
2054 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
2055 /* chain is missing so create it */
2056 stable_node = alloc_stable_node_chain(stable_node_dup,
2062 * Add this stable_node dup that was
2063 * migrated to the stable_node chain
2064 * of the current nid for this page
2067 VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
2068 VM_BUG_ON(page_node->head != &migrate_nodes);
2069 list_del(&page_node->list);
2070 DO_NUMA(page_node->nid = nid);
2071 stable_node_chain_add_dup(page_node, stable_node);
2076 * stable_tree_insert - insert stable tree node pointing to new ksm page
2077 * into the stable tree.
2079 * This function returns the stable tree node just allocated on success,
2082 static struct ksm_stable_node *stable_tree_insert(struct page *kpage)
2086 struct rb_root *root;
2087 struct rb_node **new;
2088 struct rb_node *parent;
2089 struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any;
2090 bool need_chain = false;
2092 kpfn = page_to_pfn(kpage);
2093 nid = get_kpfn_nid(kpfn);
2094 root = root_stable_tree + nid;
2097 new = &root->rb_node;
2100 struct page *tree_page;
2104 stable_node = rb_entry(*new, struct ksm_stable_node, node);
2105 stable_node_any = NULL;
2106 tree_page = chain(&stable_node_dup, stable_node, root);
2107 if (!stable_node_dup) {
2109 * Either all stable_node dups were full in
2110 * this stable_node chain, or this chain was
2111 * empty and should be rb_erased.
2113 stable_node_any = stable_node_dup_any(stable_node,
2115 if (!stable_node_any) {
2116 /* rb_erase just run */
2120 * Take any of the stable_node dups page of
2121 * this stable_node chain to let the tree walk
2122 * continue. All KSM pages belonging to the
2123 * stable_node dups in a stable_node chain
2124 * have the same content and they're
2125 * write protected at all times. Any will work
2126 * fine to continue the walk.
2128 tree_page = get_ksm_page(stable_node_any,
2129 GET_KSM_PAGE_NOLOCK);
2131 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
2134 * If we walked over a stale stable_node,
2135 * get_ksm_page() will call rb_erase() and it
2136 * may rebalance the tree from under us. So
2137 * restart the search from scratch. Returning
2138 * NULL would be safe too, but we'd generate
2139 * false negative insertions just because some
2140 * stable_node was stale.
2145 ret = memcmp_pages(kpage, tree_page);
2146 put_page(tree_page);
2150 new = &parent->rb_left;
2152 new = &parent->rb_right;
2159 stable_node_dup = alloc_stable_node();
2160 if (!stable_node_dup)
2163 INIT_HLIST_HEAD(&stable_node_dup->hlist);
2164 stable_node_dup->kpfn = kpfn;
2165 set_page_stable_node(kpage, stable_node_dup);
2166 stable_node_dup->rmap_hlist_len = 0;
2167 DO_NUMA(stable_node_dup->nid = nid);
2169 rb_link_node(&stable_node_dup->node, parent, new);
2170 rb_insert_color(&stable_node_dup->node, root);
2172 if (!is_stable_node_chain(stable_node)) {
2173 struct ksm_stable_node *orig = stable_node;
2174 /* chain is missing so create it */
2175 stable_node = alloc_stable_node_chain(orig, root);
2177 free_stable_node(stable_node_dup);
2181 stable_node_chain_add_dup(stable_node_dup, stable_node);
2184 return stable_node_dup;
2188 * unstable_tree_search_insert - search for identical page,
2189 * else insert rmap_item into the unstable tree.
2191 * This function searches for a page in the unstable tree identical to the
2192 * page currently being scanned; and if no identical page is found in the
2193 * tree, we insert rmap_item as a new object into the unstable tree.
2195 * This function returns pointer to rmap_item found to be identical
2196 * to the currently scanned page, NULL otherwise.
2198 * This function does both searching and inserting, because they share
2199 * the same walking algorithm in an rbtree.
2202 struct ksm_rmap_item *unstable_tree_search_insert(struct ksm_rmap_item *rmap_item,
2204 struct page **tree_pagep)
2206 struct rb_node **new;
2207 struct rb_root *root;
2208 struct rb_node *parent = NULL;
2211 nid = get_kpfn_nid(page_to_pfn(page));
2212 root = root_unstable_tree + nid;
2213 new = &root->rb_node;
2216 struct ksm_rmap_item *tree_rmap_item;
2217 struct page *tree_page;
2221 tree_rmap_item = rb_entry(*new, struct ksm_rmap_item, node);
2222 tree_page = get_mergeable_page(tree_rmap_item);
2227 * Don't substitute a ksm page for a forked page.
2229 if (page == tree_page) {
2230 put_page(tree_page);
2234 ret = memcmp_pages(page, tree_page);
2238 put_page(tree_page);
2239 new = &parent->rb_left;
2240 } else if (ret > 0) {
2241 put_page(tree_page);
2242 new = &parent->rb_right;
2243 } else if (!ksm_merge_across_nodes &&
2244 page_to_nid(tree_page) != nid) {
2246 * If tree_page has been migrated to another NUMA node,
2247 * it will be flushed out and put in the right unstable
2248 * tree next time: only merge with it when across_nodes.
2250 put_page(tree_page);
2253 *tree_pagep = tree_page;
2254 return tree_rmap_item;
2258 rmap_item->address |= UNSTABLE_FLAG;
2259 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
2260 DO_NUMA(rmap_item->nid = nid);
2261 rb_link_node(&rmap_item->node, parent, new);
2262 rb_insert_color(&rmap_item->node, root);
2264 ksm_pages_unshared++;
2269 * stable_tree_append - add another rmap_item to the linked list of
2270 * rmap_items hanging off a given node of the stable tree, all sharing
2271 * the same ksm page.
2273 static void stable_tree_append(struct ksm_rmap_item *rmap_item,
2274 struct ksm_stable_node *stable_node,
2275 bool max_page_sharing_bypass)
2278 * rmap won't find this mapping if we don't insert the
2279 * rmap_item in the right stable_node
2280 * duplicate. page_migration could break later if rmap breaks,
2281 * so we can as well crash here. We really need to check for
2282 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
2283 * for other negative values as an underflow if detected here
2284 * for the first time (and not when decreasing rmap_hlist_len)
2285 * would be sign of memory corruption in the stable_node.
2287 BUG_ON(stable_node->rmap_hlist_len < 0);
2289 stable_node->rmap_hlist_len++;
2290 if (!max_page_sharing_bypass)
2291 /* possibly non fatal but unexpected overflow, only warn */
2292 WARN_ON_ONCE(stable_node->rmap_hlist_len >
2293 ksm_max_page_sharing);
2295 rmap_item->head = stable_node;
2296 rmap_item->address |= STABLE_FLAG;
2297 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2299 if (rmap_item->hlist.next)
2300 ksm_pages_sharing++;
2304 rmap_item->mm->ksm_merging_pages++;
2308 * cmp_and_merge_page - first see if page can be merged into the stable tree;
2309 * if not, compare checksum to previous and if it's the same, see if page can
2310 * be inserted into the unstable tree, or merged with a page already there and
2311 * both transferred to the stable tree.
2313 * @page: the page that we are searching identical page to.
2314 * @rmap_item: the reverse mapping into the virtual address of this page
2316 static void cmp_and_merge_page(struct page *page, struct ksm_rmap_item *rmap_item)
2318 struct mm_struct *mm = rmap_item->mm;
2319 struct ksm_rmap_item *tree_rmap_item;
2320 struct page *tree_page = NULL;
2321 struct ksm_stable_node *stable_node;
2323 unsigned int checksum;
2325 bool max_page_sharing_bypass = false;
2327 stable_node = page_stable_node(page);
2329 if (stable_node->head != &migrate_nodes &&
2330 get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2331 NUMA(stable_node->nid)) {
2332 stable_node_dup_del(stable_node);
2333 stable_node->head = &migrate_nodes;
2334 list_add(&stable_node->list, stable_node->head);
2336 if (stable_node->head != &migrate_nodes &&
2337 rmap_item->head == stable_node)
2340 * If it's a KSM fork, allow it to go over the sharing limit
2343 if (!is_page_sharing_candidate(stable_node))
2344 max_page_sharing_bypass = true;
2347 /* We first start with searching the page inside the stable tree */
2348 kpage = stable_tree_search(page);
2349 if (kpage == page && rmap_item->head == stable_node) {
2354 remove_rmap_item_from_tree(rmap_item);
2357 if (PTR_ERR(kpage) == -EBUSY)
2360 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2363 * The page was successfully merged:
2364 * add its rmap_item to the stable tree.
2367 stable_tree_append(rmap_item, page_stable_node(kpage),
2368 max_page_sharing_bypass);
2376 * If the hash value of the page has changed from the last time
2377 * we calculated it, this page is changing frequently: therefore we
2378 * don't want to insert it in the unstable tree, and we don't want
2379 * to waste our time searching for something identical to it there.
2381 checksum = calc_checksum(page);
2382 if (rmap_item->oldchecksum != checksum) {
2383 rmap_item->oldchecksum = checksum;
2388 * Same checksum as an empty page. We attempt to merge it with the
2389 * appropriate zero page if the user enabled this via sysfs.
2391 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2392 struct vm_area_struct *vma;
2395 vma = find_mergeable_vma(mm, rmap_item->address);
2397 err = try_to_merge_one_page(vma, page,
2398 ZERO_PAGE(rmap_item->address));
2399 trace_ksm_merge_one_page(
2400 page_to_pfn(ZERO_PAGE(rmap_item->address)),
2401 rmap_item, mm, err);
2404 * If the vma is out of date, we do not need to
2409 mmap_read_unlock(mm);
2411 * In case of failure, the page was not really empty, so we
2412 * need to continue. Otherwise we're done.
2418 unstable_tree_search_insert(rmap_item, page, &tree_page);
2419 if (tree_rmap_item) {
2422 kpage = try_to_merge_two_pages(rmap_item, page,
2423 tree_rmap_item, tree_page);
2425 * If both pages we tried to merge belong to the same compound
2426 * page, then we actually ended up increasing the reference
2427 * count of the same compound page twice, and split_huge_page
2429 * Here we set a flag if that happened, and we use it later to
2430 * try split_huge_page again. Since we call put_page right
2431 * afterwards, the reference count will be correct and
2432 * split_huge_page should succeed.
2434 split = PageTransCompound(page)
2435 && compound_head(page) == compound_head(tree_page);
2436 put_page(tree_page);
2439 * The pages were successfully merged: insert new
2440 * node in the stable tree and add both rmap_items.
2443 stable_node = stable_tree_insert(kpage);
2445 stable_tree_append(tree_rmap_item, stable_node,
2447 stable_tree_append(rmap_item, stable_node,
2453 * If we fail to insert the page into the stable tree,
2454 * we will have 2 virtual addresses that are pointing
2455 * to a ksm page left outside the stable tree,
2456 * in which case we need to break_cow on both.
2459 break_cow(tree_rmap_item);
2460 break_cow(rmap_item);
2464 * We are here if we tried to merge two pages and
2465 * failed because they both belonged to the same
2466 * compound page. We will split the page now, but no
2467 * merging will take place.
2468 * We do not want to add the cost of a full lock; if
2469 * the page is locked, it is better to skip it and
2470 * perhaps try again later.
2472 if (!trylock_page(page))
2474 split_huge_page(page);
2480 static struct ksm_rmap_item *get_next_rmap_item(struct ksm_mm_slot *mm_slot,
2481 struct ksm_rmap_item **rmap_list,
2484 struct ksm_rmap_item *rmap_item;
2486 while (*rmap_list) {
2487 rmap_item = *rmap_list;
2488 if ((rmap_item->address & PAGE_MASK) == addr)
2490 if (rmap_item->address > addr)
2492 *rmap_list = rmap_item->rmap_list;
2493 remove_rmap_item_from_tree(rmap_item);
2494 free_rmap_item(rmap_item);
2497 rmap_item = alloc_rmap_item();
2499 /* It has already been zeroed */
2500 rmap_item->mm = mm_slot->slot.mm;
2501 rmap_item->mm->ksm_rmap_items++;
2502 rmap_item->address = addr;
2503 rmap_item->rmap_list = *rmap_list;
2504 *rmap_list = rmap_item;
2510 * Calculate skip age for the ksm page age. The age determines how often
2511 * de-duplicating has already been tried unsuccessfully. If the age is
2512 * smaller, the scanning of this page is skipped for less scans.
2514 * @age: rmap_item age of page
2516 static unsigned int skip_age(rmap_age_t age)
2529 * Determines if a page should be skipped for the current scan.
2531 * @page: page to check
2532 * @rmap_item: associated rmap_item of page
2534 static bool should_skip_rmap_item(struct page *page,
2535 struct ksm_rmap_item *rmap_item)
2539 if (!ksm_smart_scan)
2543 * Never skip pages that are already KSM; pages cmp_and_merge_page()
2544 * will essentially ignore them, but we still have to process them
2550 age = rmap_item->age;
2555 * Smaller ages are not skipped, they need to get a chance to go
2556 * through the different phases of the KSM merging.
2562 * Are we still allowed to skip? If not, then don't skip it
2563 * and determine how much more often we are allowed to skip next.
2565 if (!rmap_item->remaining_skips) {
2566 rmap_item->remaining_skips = skip_age(age);
2570 /* Skip this page */
2571 ksm_pages_skipped++;
2572 rmap_item->remaining_skips--;
2573 remove_rmap_item_from_tree(rmap_item);
2577 static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page)
2579 struct mm_struct *mm;
2580 struct ksm_mm_slot *mm_slot;
2581 struct mm_slot *slot;
2582 struct vm_area_struct *vma;
2583 struct ksm_rmap_item *rmap_item;
2584 struct vma_iterator vmi;
2587 if (list_empty(&ksm_mm_head.slot.mm_node))
2590 mm_slot = ksm_scan.mm_slot;
2591 if (mm_slot == &ksm_mm_head) {
2592 advisor_start_scan();
2593 trace_ksm_start_scan(ksm_scan.seqnr, ksm_rmap_items);
2596 * A number of pages can hang around indefinitely in per-cpu
2597 * LRU cache, raised page count preventing write_protect_page
2598 * from merging them. Though it doesn't really matter much,
2599 * it is puzzling to see some stuck in pages_volatile until
2600 * other activity jostles them out, and they also prevented
2601 * LTP's KSM test from succeeding deterministically; so drain
2602 * them here (here rather than on entry to ksm_do_scan(),
2603 * so we don't IPI too often when pages_to_scan is set low).
2605 lru_add_drain_all();
2608 * Whereas stale stable_nodes on the stable_tree itself
2609 * get pruned in the regular course of stable_tree_search(),
2610 * those moved out to the migrate_nodes list can accumulate:
2611 * so prune them once before each full scan.
2613 if (!ksm_merge_across_nodes) {
2614 struct ksm_stable_node *stable_node, *next;
2615 struct folio *folio;
2617 list_for_each_entry_safe(stable_node, next,
2618 &migrate_nodes, list) {
2619 folio = ksm_get_folio(stable_node,
2620 GET_KSM_PAGE_NOLOCK);
2627 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2628 root_unstable_tree[nid] = RB_ROOT;
2630 spin_lock(&ksm_mmlist_lock);
2631 slot = list_entry(mm_slot->slot.mm_node.next,
2632 struct mm_slot, mm_node);
2633 mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
2634 ksm_scan.mm_slot = mm_slot;
2635 spin_unlock(&ksm_mmlist_lock);
2637 * Although we tested list_empty() above, a racing __ksm_exit
2638 * of the last mm on the list may have removed it since then.
2640 if (mm_slot == &ksm_mm_head)
2643 ksm_scan.address = 0;
2644 ksm_scan.rmap_list = &mm_slot->rmap_list;
2647 slot = &mm_slot->slot;
2649 vma_iter_init(&vmi, mm, ksm_scan.address);
2652 if (ksm_test_exit(mm))
2655 for_each_vma(vmi, vma) {
2656 if (!(vma->vm_flags & VM_MERGEABLE))
2658 if (ksm_scan.address < vma->vm_start)
2659 ksm_scan.address = vma->vm_start;
2661 ksm_scan.address = vma->vm_end;
2663 while (ksm_scan.address < vma->vm_end) {
2664 if (ksm_test_exit(mm))
2666 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2667 if (IS_ERR_OR_NULL(*page)) {
2668 ksm_scan.address += PAGE_SIZE;
2672 if (is_zone_device_page(*page))
2674 if (PageAnon(*page)) {
2675 flush_anon_page(vma, *page, ksm_scan.address);
2676 flush_dcache_page(*page);
2677 rmap_item = get_next_rmap_item(mm_slot,
2678 ksm_scan.rmap_list, ksm_scan.address);
2680 ksm_scan.rmap_list =
2681 &rmap_item->rmap_list;
2683 if (should_skip_rmap_item(*page, rmap_item))
2686 ksm_scan.address += PAGE_SIZE;
2689 mmap_read_unlock(mm);
2694 ksm_scan.address += PAGE_SIZE;
2699 if (ksm_test_exit(mm)) {
2701 ksm_scan.address = 0;
2702 ksm_scan.rmap_list = &mm_slot->rmap_list;
2705 * Nuke all the rmap_items that are above this current rmap:
2706 * because there were no VM_MERGEABLE vmas with such addresses.
2708 remove_trailing_rmap_items(ksm_scan.rmap_list);
2710 spin_lock(&ksm_mmlist_lock);
2711 slot = list_entry(mm_slot->slot.mm_node.next,
2712 struct mm_slot, mm_node);
2713 ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
2714 if (ksm_scan.address == 0) {
2716 * We've completed a full scan of all vmas, holding mmap_lock
2717 * throughout, and found no VM_MERGEABLE: so do the same as
2718 * __ksm_exit does to remove this mm from all our lists now.
2719 * This applies either when cleaning up after __ksm_exit
2720 * (but beware: we can reach here even before __ksm_exit),
2721 * or when all VM_MERGEABLE areas have been unmapped (and
2722 * mmap_lock then protects against race with MADV_MERGEABLE).
2724 hash_del(&mm_slot->slot.hash);
2725 list_del(&mm_slot->slot.mm_node);
2726 spin_unlock(&ksm_mmlist_lock);
2728 mm_slot_free(mm_slot_cache, mm_slot);
2729 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2730 clear_bit(MMF_VM_MERGE_ANY, &mm->flags);
2731 mmap_read_unlock(mm);
2734 mmap_read_unlock(mm);
2736 * mmap_read_unlock(mm) first because after
2737 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2738 * already have been freed under us by __ksm_exit()
2739 * because the "mm_slot" is still hashed and
2740 * ksm_scan.mm_slot doesn't point to it anymore.
2742 spin_unlock(&ksm_mmlist_lock);
2745 /* Repeat until we've completed scanning the whole list */
2746 mm_slot = ksm_scan.mm_slot;
2747 if (mm_slot != &ksm_mm_head)
2750 advisor_stop_scan();
2752 trace_ksm_stop_scan(ksm_scan.seqnr, ksm_rmap_items);
2758 * ksm_do_scan - the ksm scanner main worker function.
2759 * @scan_npages: number of pages we want to scan before we return.
2761 static void ksm_do_scan(unsigned int scan_npages)
2763 struct ksm_rmap_item *rmap_item;
2765 unsigned int npages = scan_npages;
2767 while (npages-- && likely(!freezing(current))) {
2769 rmap_item = scan_get_next_rmap_item(&page);
2772 cmp_and_merge_page(page, rmap_item);
2776 ksm_pages_scanned += scan_npages - npages;
2779 static int ksmd_should_run(void)
2781 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.slot.mm_node);
2784 static int ksm_scan_thread(void *nothing)
2786 unsigned int sleep_ms;
2789 set_user_nice(current, 5);
2791 while (!kthread_should_stop()) {
2792 mutex_lock(&ksm_thread_mutex);
2793 wait_while_offlining();
2794 if (ksmd_should_run())
2795 ksm_do_scan(ksm_thread_pages_to_scan);
2796 mutex_unlock(&ksm_thread_mutex);
2798 if (ksmd_should_run()) {
2799 sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs);
2800 wait_event_freezable_timeout(ksm_iter_wait,
2801 sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs),
2802 msecs_to_jiffies(sleep_ms));
2804 wait_event_freezable(ksm_thread_wait,
2805 ksmd_should_run() || kthread_should_stop());
2811 static void __ksm_add_vma(struct vm_area_struct *vma)
2813 unsigned long vm_flags = vma->vm_flags;
2815 if (vm_flags & VM_MERGEABLE)
2818 if (vma_ksm_compatible(vma))
2819 vm_flags_set(vma, VM_MERGEABLE);
2822 static int __ksm_del_vma(struct vm_area_struct *vma)
2826 if (!(vma->vm_flags & VM_MERGEABLE))
2829 if (vma->anon_vma) {
2830 err = unmerge_ksm_pages(vma, vma->vm_start, vma->vm_end, true);
2835 vm_flags_clear(vma, VM_MERGEABLE);
2839 * ksm_add_vma - Mark vma as mergeable if compatible
2841 * @vma: Pointer to vma
2843 void ksm_add_vma(struct vm_area_struct *vma)
2845 struct mm_struct *mm = vma->vm_mm;
2847 if (test_bit(MMF_VM_MERGE_ANY, &mm->flags))
2851 static void ksm_add_vmas(struct mm_struct *mm)
2853 struct vm_area_struct *vma;
2855 VMA_ITERATOR(vmi, mm, 0);
2856 for_each_vma(vmi, vma)
2860 static int ksm_del_vmas(struct mm_struct *mm)
2862 struct vm_area_struct *vma;
2865 VMA_ITERATOR(vmi, mm, 0);
2866 for_each_vma(vmi, vma) {
2867 err = __ksm_del_vma(vma);
2875 * ksm_enable_merge_any - Add mm to mm ksm list and enable merging on all
2878 * @mm: Pointer to mm
2880 * Returns 0 on success, otherwise error code
2882 int ksm_enable_merge_any(struct mm_struct *mm)
2886 if (test_bit(MMF_VM_MERGE_ANY, &mm->flags))
2889 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2890 err = __ksm_enter(mm);
2895 set_bit(MMF_VM_MERGE_ANY, &mm->flags);
2902 * ksm_disable_merge_any - Disable merging on all compatible VMA's of the mm,
2903 * previously enabled via ksm_enable_merge_any().
2905 * Disabling merging implies unmerging any merged pages, like setting
2906 * MADV_UNMERGEABLE would. If unmerging fails, the whole operation fails and
2907 * merging on all compatible VMA's remains enabled.
2909 * @mm: Pointer to mm
2911 * Returns 0 on success, otherwise error code
2913 int ksm_disable_merge_any(struct mm_struct *mm)
2917 if (!test_bit(MMF_VM_MERGE_ANY, &mm->flags))
2920 err = ksm_del_vmas(mm);
2926 clear_bit(MMF_VM_MERGE_ANY, &mm->flags);
2930 int ksm_disable(struct mm_struct *mm)
2932 mmap_assert_write_locked(mm);
2934 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags))
2936 if (test_bit(MMF_VM_MERGE_ANY, &mm->flags))
2937 return ksm_disable_merge_any(mm);
2938 return ksm_del_vmas(mm);
2941 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2942 unsigned long end, int advice, unsigned long *vm_flags)
2944 struct mm_struct *mm = vma->vm_mm;
2948 case MADV_MERGEABLE:
2949 if (vma->vm_flags & VM_MERGEABLE)
2951 if (!vma_ksm_compatible(vma))
2954 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2955 err = __ksm_enter(mm);
2960 *vm_flags |= VM_MERGEABLE;
2963 case MADV_UNMERGEABLE:
2964 if (!(*vm_flags & VM_MERGEABLE))
2965 return 0; /* just ignore the advice */
2967 if (vma->anon_vma) {
2968 err = unmerge_ksm_pages(vma, start, end, true);
2973 *vm_flags &= ~VM_MERGEABLE;
2979 EXPORT_SYMBOL_GPL(ksm_madvise);
2981 int __ksm_enter(struct mm_struct *mm)
2983 struct ksm_mm_slot *mm_slot;
2984 struct mm_slot *slot;
2987 mm_slot = mm_slot_alloc(mm_slot_cache);
2991 slot = &mm_slot->slot;
2993 /* Check ksm_run too? Would need tighter locking */
2994 needs_wakeup = list_empty(&ksm_mm_head.slot.mm_node);
2996 spin_lock(&ksm_mmlist_lock);
2997 mm_slot_insert(mm_slots_hash, mm, slot);
2999 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
3000 * insert just behind the scanning cursor, to let the area settle
3001 * down a little; when fork is followed by immediate exec, we don't
3002 * want ksmd to waste time setting up and tearing down an rmap_list.
3004 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
3005 * scanning cursor, otherwise KSM pages in newly forked mms will be
3006 * missed: then we might as well insert at the end of the list.
3008 if (ksm_run & KSM_RUN_UNMERGE)
3009 list_add_tail(&slot->mm_node, &ksm_mm_head.slot.mm_node);
3011 list_add_tail(&slot->mm_node, &ksm_scan.mm_slot->slot.mm_node);
3012 spin_unlock(&ksm_mmlist_lock);
3014 set_bit(MMF_VM_MERGEABLE, &mm->flags);
3018 wake_up_interruptible(&ksm_thread_wait);
3020 trace_ksm_enter(mm);
3024 void __ksm_exit(struct mm_struct *mm)
3026 struct ksm_mm_slot *mm_slot;
3027 struct mm_slot *slot;
3028 int easy_to_free = 0;
3031 * This process is exiting: if it's straightforward (as is the
3032 * case when ksmd was never running), free mm_slot immediately.
3033 * But if it's at the cursor or has rmap_items linked to it, use
3034 * mmap_lock to synchronize with any break_cows before pagetables
3035 * are freed, and leave the mm_slot on the list for ksmd to free.
3036 * Beware: ksm may already have noticed it exiting and freed the slot.
3039 spin_lock(&ksm_mmlist_lock);
3040 slot = mm_slot_lookup(mm_slots_hash, mm);
3041 mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
3042 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
3043 if (!mm_slot->rmap_list) {
3044 hash_del(&slot->hash);
3045 list_del(&slot->mm_node);
3048 list_move(&slot->mm_node,
3049 &ksm_scan.mm_slot->slot.mm_node);
3052 spin_unlock(&ksm_mmlist_lock);
3055 mm_slot_free(mm_slot_cache, mm_slot);
3056 clear_bit(MMF_VM_MERGE_ANY, &mm->flags);
3057 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
3059 } else if (mm_slot) {
3060 mmap_write_lock(mm);
3061 mmap_write_unlock(mm);
3067 struct folio *ksm_might_need_to_copy(struct folio *folio,
3068 struct vm_area_struct *vma, unsigned long addr)
3070 struct page *page = folio_page(folio, 0);
3071 struct anon_vma *anon_vma = folio_anon_vma(folio);
3072 struct folio *new_folio;
3074 if (folio_test_large(folio))
3077 if (folio_test_ksm(folio)) {
3078 if (folio_stable_node(folio) &&
3079 !(ksm_run & KSM_RUN_UNMERGE))
3080 return folio; /* no need to copy it */
3081 } else if (!anon_vma) {
3082 return folio; /* no need to copy it */
3083 } else if (folio->index == linear_page_index(vma, addr) &&
3084 anon_vma->root == vma->anon_vma->root) {
3085 return folio; /* still no need to copy it */
3087 if (PageHWPoison(page))
3088 return ERR_PTR(-EHWPOISON);
3089 if (!folio_test_uptodate(folio))
3090 return folio; /* let do_swap_page report the error */
3092 new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma, addr, false);
3094 mem_cgroup_charge(new_folio, vma->vm_mm, GFP_KERNEL)) {
3095 folio_put(new_folio);
3099 if (copy_mc_user_highpage(folio_page(new_folio, 0), page,
3101 folio_put(new_folio);
3102 memory_failure_queue(folio_pfn(folio), 0);
3103 return ERR_PTR(-EHWPOISON);
3105 folio_set_dirty(new_folio);
3106 __folio_mark_uptodate(new_folio);
3107 __folio_set_locked(new_folio);
3109 count_vm_event(KSM_SWPIN_COPY);
3116 void rmap_walk_ksm(struct folio *folio, struct rmap_walk_control *rwc)
3118 struct ksm_stable_node *stable_node;
3119 struct ksm_rmap_item *rmap_item;
3120 int search_new_forks = 0;
3122 VM_BUG_ON_FOLIO(!folio_test_ksm(folio), folio);
3125 * Rely on the page lock to protect against concurrent modifications
3126 * to that page's node of the stable tree.
3128 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
3130 stable_node = folio_stable_node(folio);
3134 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
3135 struct anon_vma *anon_vma = rmap_item->anon_vma;
3136 struct anon_vma_chain *vmac;
3137 struct vm_area_struct *vma;
3140 if (!anon_vma_trylock_read(anon_vma)) {
3141 if (rwc->try_lock) {
3142 rwc->contended = true;
3145 anon_vma_lock_read(anon_vma);
3147 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
3154 /* Ignore the stable/unstable/sqnr flags */
3155 addr = rmap_item->address & PAGE_MASK;
3157 if (addr < vma->vm_start || addr >= vma->vm_end)
3160 * Initially we examine only the vma which covers this
3161 * rmap_item; but later, if there is still work to do,
3162 * we examine covering vmas in other mms: in case they
3163 * were forked from the original since ksmd passed.
3165 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
3168 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
3171 if (!rwc->rmap_one(folio, vma, addr, rwc->arg)) {
3172 anon_vma_unlock_read(anon_vma);
3175 if (rwc->done && rwc->done(folio)) {
3176 anon_vma_unlock_read(anon_vma);
3180 anon_vma_unlock_read(anon_vma);
3182 if (!search_new_forks++)
3186 #ifdef CONFIG_MEMORY_FAILURE
3188 * Collect processes when the error hit an ksm page.
3190 void collect_procs_ksm(struct page *page, struct list_head *to_kill,
3193 struct ksm_stable_node *stable_node;
3194 struct ksm_rmap_item *rmap_item;
3195 struct folio *folio = page_folio(page);
3196 struct vm_area_struct *vma;
3197 struct task_struct *tsk;
3199 stable_node = folio_stable_node(folio);
3202 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
3203 struct anon_vma *av = rmap_item->anon_vma;
3205 anon_vma_lock_read(av);
3207 for_each_process(tsk) {
3208 struct anon_vma_chain *vmac;
3210 struct task_struct *t =
3211 task_early_kill(tsk, force_early);
3214 anon_vma_interval_tree_foreach(vmac, &av->rb_root, 0,
3218 if (vma->vm_mm == t->mm) {
3219 addr = rmap_item->address & PAGE_MASK;
3220 add_to_kill_ksm(t, page, vma, to_kill,
3226 anon_vma_unlock_read(av);
3231 #ifdef CONFIG_MIGRATION
3232 void folio_migrate_ksm(struct folio *newfolio, struct folio *folio)
3234 struct ksm_stable_node *stable_node;
3236 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
3237 VM_BUG_ON_FOLIO(!folio_test_locked(newfolio), newfolio);
3238 VM_BUG_ON_FOLIO(newfolio->mapping != folio->mapping, newfolio);
3240 stable_node = folio_stable_node(folio);
3242 VM_BUG_ON_FOLIO(stable_node->kpfn != folio_pfn(folio), folio);
3243 stable_node->kpfn = folio_pfn(newfolio);
3245 * newfolio->mapping was set in advance; now we need smp_wmb()
3246 * to make sure that the new stable_node->kpfn is visible
3247 * to get_ksm_page() before it can see that folio->mapping
3248 * has gone stale (or that folio_test_swapcache has been cleared).
3251 folio_set_stable_node(folio, NULL);
3254 #endif /* CONFIG_MIGRATION */
3256 #ifdef CONFIG_MEMORY_HOTREMOVE
3257 static void wait_while_offlining(void)
3259 while (ksm_run & KSM_RUN_OFFLINE) {
3260 mutex_unlock(&ksm_thread_mutex);
3261 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
3262 TASK_UNINTERRUPTIBLE);
3263 mutex_lock(&ksm_thread_mutex);
3267 static bool stable_node_dup_remove_range(struct ksm_stable_node *stable_node,
3268 unsigned long start_pfn,
3269 unsigned long end_pfn)
3271 if (stable_node->kpfn >= start_pfn &&
3272 stable_node->kpfn < end_pfn) {
3274 * Don't get_ksm_page, page has already gone:
3275 * which is why we keep kpfn instead of page*
3277 remove_node_from_stable_tree(stable_node);
3283 static bool stable_node_chain_remove_range(struct ksm_stable_node *stable_node,
3284 unsigned long start_pfn,
3285 unsigned long end_pfn,
3286 struct rb_root *root)
3288 struct ksm_stable_node *dup;
3289 struct hlist_node *hlist_safe;
3291 if (!is_stable_node_chain(stable_node)) {
3292 VM_BUG_ON(is_stable_node_dup(stable_node));
3293 return stable_node_dup_remove_range(stable_node, start_pfn,
3297 hlist_for_each_entry_safe(dup, hlist_safe,
3298 &stable_node->hlist, hlist_dup) {
3299 VM_BUG_ON(!is_stable_node_dup(dup));
3300 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
3302 if (hlist_empty(&stable_node->hlist)) {
3303 free_stable_node_chain(stable_node, root);
3304 return true; /* notify caller that tree was rebalanced */
3309 static void ksm_check_stable_tree(unsigned long start_pfn,
3310 unsigned long end_pfn)
3312 struct ksm_stable_node *stable_node, *next;
3313 struct rb_node *node;
3316 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
3317 node = rb_first(root_stable_tree + nid);
3319 stable_node = rb_entry(node, struct ksm_stable_node, node);
3320 if (stable_node_chain_remove_range(stable_node,
3324 node = rb_first(root_stable_tree + nid);
3326 node = rb_next(node);
3330 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
3331 if (stable_node->kpfn >= start_pfn &&
3332 stable_node->kpfn < end_pfn)
3333 remove_node_from_stable_tree(stable_node);
3338 static int ksm_memory_callback(struct notifier_block *self,
3339 unsigned long action, void *arg)
3341 struct memory_notify *mn = arg;
3344 case MEM_GOING_OFFLINE:
3346 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
3347 * and remove_all_stable_nodes() while memory is going offline:
3348 * it is unsafe for them to touch the stable tree at this time.
3349 * But unmerge_ksm_pages(), rmap lookups and other entry points
3350 * which do not need the ksm_thread_mutex are all safe.
3352 mutex_lock(&ksm_thread_mutex);
3353 ksm_run |= KSM_RUN_OFFLINE;
3354 mutex_unlock(&ksm_thread_mutex);
3359 * Most of the work is done by page migration; but there might
3360 * be a few stable_nodes left over, still pointing to struct
3361 * pages which have been offlined: prune those from the tree,
3362 * otherwise get_ksm_page() might later try to access a
3363 * non-existent struct page.
3365 ksm_check_stable_tree(mn->start_pfn,
3366 mn->start_pfn + mn->nr_pages);
3368 case MEM_CANCEL_OFFLINE:
3369 mutex_lock(&ksm_thread_mutex);
3370 ksm_run &= ~KSM_RUN_OFFLINE;
3371 mutex_unlock(&ksm_thread_mutex);
3373 smp_mb(); /* wake_up_bit advises this */
3374 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
3380 static void wait_while_offlining(void)
3383 #endif /* CONFIG_MEMORY_HOTREMOVE */
3385 #ifdef CONFIG_PROC_FS
3386 long ksm_process_profit(struct mm_struct *mm)
3388 return (long)(mm->ksm_merging_pages + mm->ksm_zero_pages) * PAGE_SIZE -
3389 mm->ksm_rmap_items * sizeof(struct ksm_rmap_item);
3391 #endif /* CONFIG_PROC_FS */
3395 * This all compiles without CONFIG_SYSFS, but is a waste of space.
3398 #define KSM_ATTR_RO(_name) \
3399 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3400 #define KSM_ATTR(_name) \
3401 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3403 static ssize_t sleep_millisecs_show(struct kobject *kobj,
3404 struct kobj_attribute *attr, char *buf)
3406 return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs);
3409 static ssize_t sleep_millisecs_store(struct kobject *kobj,
3410 struct kobj_attribute *attr,
3411 const char *buf, size_t count)
3416 err = kstrtouint(buf, 10, &msecs);
3420 ksm_thread_sleep_millisecs = msecs;
3421 wake_up_interruptible(&ksm_iter_wait);
3425 KSM_ATTR(sleep_millisecs);
3427 static ssize_t pages_to_scan_show(struct kobject *kobj,
3428 struct kobj_attribute *attr, char *buf)
3430 return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan);
3433 static ssize_t pages_to_scan_store(struct kobject *kobj,
3434 struct kobj_attribute *attr,
3435 const char *buf, size_t count)
3437 unsigned int nr_pages;
3440 if (ksm_advisor != KSM_ADVISOR_NONE)
3443 err = kstrtouint(buf, 10, &nr_pages);
3447 ksm_thread_pages_to_scan = nr_pages;
3451 KSM_ATTR(pages_to_scan);
3453 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
3456 return sysfs_emit(buf, "%lu\n", ksm_run);
3459 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
3460 const char *buf, size_t count)
3465 err = kstrtouint(buf, 10, &flags);
3468 if (flags > KSM_RUN_UNMERGE)
3472 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
3473 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
3474 * breaking COW to free the pages_shared (but leaves mm_slots
3475 * on the list for when ksmd may be set running again).
3478 mutex_lock(&ksm_thread_mutex);
3479 wait_while_offlining();
3480 if (ksm_run != flags) {
3482 if (flags & KSM_RUN_UNMERGE) {
3483 set_current_oom_origin();
3484 err = unmerge_and_remove_all_rmap_items();
3485 clear_current_oom_origin();
3487 ksm_run = KSM_RUN_STOP;
3492 mutex_unlock(&ksm_thread_mutex);
3494 if (flags & KSM_RUN_MERGE)
3495 wake_up_interruptible(&ksm_thread_wait);
3502 static ssize_t merge_across_nodes_show(struct kobject *kobj,
3503 struct kobj_attribute *attr, char *buf)
3505 return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes);
3508 static ssize_t merge_across_nodes_store(struct kobject *kobj,
3509 struct kobj_attribute *attr,
3510 const char *buf, size_t count)
3515 err = kstrtoul(buf, 10, &knob);
3521 mutex_lock(&ksm_thread_mutex);
3522 wait_while_offlining();
3523 if (ksm_merge_across_nodes != knob) {
3524 if (ksm_pages_shared || remove_all_stable_nodes())
3526 else if (root_stable_tree == one_stable_tree) {
3527 struct rb_root *buf;
3529 * This is the first time that we switch away from the
3530 * default of merging across nodes: must now allocate
3531 * a buffer to hold as many roots as may be needed.
3532 * Allocate stable and unstable together:
3533 * MAXSMP NODES_SHIFT 10 will use 16kB.
3535 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
3537 /* Let us assume that RB_ROOT is NULL is zero */
3541 root_stable_tree = buf;
3542 root_unstable_tree = buf + nr_node_ids;
3543 /* Stable tree is empty but not the unstable */
3544 root_unstable_tree[0] = one_unstable_tree[0];
3548 ksm_merge_across_nodes = knob;
3549 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
3552 mutex_unlock(&ksm_thread_mutex);
3554 return err ? err : count;
3556 KSM_ATTR(merge_across_nodes);
3559 static ssize_t use_zero_pages_show(struct kobject *kobj,
3560 struct kobj_attribute *attr, char *buf)
3562 return sysfs_emit(buf, "%u\n", ksm_use_zero_pages);
3564 static ssize_t use_zero_pages_store(struct kobject *kobj,
3565 struct kobj_attribute *attr,
3566 const char *buf, size_t count)
3571 err = kstrtobool(buf, &value);
3575 ksm_use_zero_pages = value;
3579 KSM_ATTR(use_zero_pages);
3581 static ssize_t max_page_sharing_show(struct kobject *kobj,
3582 struct kobj_attribute *attr, char *buf)
3584 return sysfs_emit(buf, "%u\n", ksm_max_page_sharing);
3587 static ssize_t max_page_sharing_store(struct kobject *kobj,
3588 struct kobj_attribute *attr,
3589 const char *buf, size_t count)
3594 err = kstrtoint(buf, 10, &knob);
3598 * When a KSM page is created it is shared by 2 mappings. This
3599 * being a signed comparison, it implicitly verifies it's not
3605 if (READ_ONCE(ksm_max_page_sharing) == knob)
3608 mutex_lock(&ksm_thread_mutex);
3609 wait_while_offlining();
3610 if (ksm_max_page_sharing != knob) {
3611 if (ksm_pages_shared || remove_all_stable_nodes())
3614 ksm_max_page_sharing = knob;
3616 mutex_unlock(&ksm_thread_mutex);
3618 return err ? err : count;
3620 KSM_ATTR(max_page_sharing);
3622 static ssize_t pages_scanned_show(struct kobject *kobj,
3623 struct kobj_attribute *attr, char *buf)
3625 return sysfs_emit(buf, "%lu\n", ksm_pages_scanned);
3627 KSM_ATTR_RO(pages_scanned);
3629 static ssize_t pages_shared_show(struct kobject *kobj,
3630 struct kobj_attribute *attr, char *buf)
3632 return sysfs_emit(buf, "%lu\n", ksm_pages_shared);
3634 KSM_ATTR_RO(pages_shared);
3636 static ssize_t pages_sharing_show(struct kobject *kobj,
3637 struct kobj_attribute *attr, char *buf)
3639 return sysfs_emit(buf, "%lu\n", ksm_pages_sharing);
3641 KSM_ATTR_RO(pages_sharing);
3643 static ssize_t pages_unshared_show(struct kobject *kobj,
3644 struct kobj_attribute *attr, char *buf)
3646 return sysfs_emit(buf, "%lu\n", ksm_pages_unshared);
3648 KSM_ATTR_RO(pages_unshared);
3650 static ssize_t pages_volatile_show(struct kobject *kobj,
3651 struct kobj_attribute *attr, char *buf)
3653 long ksm_pages_volatile;
3655 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3656 - ksm_pages_sharing - ksm_pages_unshared;
3658 * It was not worth any locking to calculate that statistic,
3659 * but it might therefore sometimes be negative: conceal that.
3661 if (ksm_pages_volatile < 0)
3662 ksm_pages_volatile = 0;
3663 return sysfs_emit(buf, "%ld\n", ksm_pages_volatile);
3665 KSM_ATTR_RO(pages_volatile);
3667 static ssize_t pages_skipped_show(struct kobject *kobj,
3668 struct kobj_attribute *attr, char *buf)
3670 return sysfs_emit(buf, "%lu\n", ksm_pages_skipped);
3672 KSM_ATTR_RO(pages_skipped);
3674 static ssize_t ksm_zero_pages_show(struct kobject *kobj,
3675 struct kobj_attribute *attr, char *buf)
3677 return sysfs_emit(buf, "%ld\n", ksm_zero_pages);
3679 KSM_ATTR_RO(ksm_zero_pages);
3681 static ssize_t general_profit_show(struct kobject *kobj,
3682 struct kobj_attribute *attr, char *buf)
3684 long general_profit;
3686 general_profit = (ksm_pages_sharing + ksm_zero_pages) * PAGE_SIZE -
3687 ksm_rmap_items * sizeof(struct ksm_rmap_item);
3689 return sysfs_emit(buf, "%ld\n", general_profit);
3691 KSM_ATTR_RO(general_profit);
3693 static ssize_t stable_node_dups_show(struct kobject *kobj,
3694 struct kobj_attribute *attr, char *buf)
3696 return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups);
3698 KSM_ATTR_RO(stable_node_dups);
3700 static ssize_t stable_node_chains_show(struct kobject *kobj,
3701 struct kobj_attribute *attr, char *buf)
3703 return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains);
3705 KSM_ATTR_RO(stable_node_chains);
3708 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3709 struct kobj_attribute *attr,
3712 return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3716 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3717 struct kobj_attribute *attr,
3718 const char *buf, size_t count)
3723 err = kstrtouint(buf, 10, &msecs);
3727 ksm_stable_node_chains_prune_millisecs = msecs;
3731 KSM_ATTR(stable_node_chains_prune_millisecs);
3733 static ssize_t full_scans_show(struct kobject *kobj,
3734 struct kobj_attribute *attr, char *buf)
3736 return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr);
3738 KSM_ATTR_RO(full_scans);
3740 static ssize_t smart_scan_show(struct kobject *kobj,
3741 struct kobj_attribute *attr, char *buf)
3743 return sysfs_emit(buf, "%u\n", ksm_smart_scan);
3746 static ssize_t smart_scan_store(struct kobject *kobj,
3747 struct kobj_attribute *attr,
3748 const char *buf, size_t count)
3753 err = kstrtobool(buf, &value);
3757 ksm_smart_scan = value;
3760 KSM_ATTR(smart_scan);
3762 static ssize_t advisor_mode_show(struct kobject *kobj,
3763 struct kobj_attribute *attr, char *buf)
3767 if (ksm_advisor == KSM_ADVISOR_NONE)
3768 output = "[none] scan-time";
3769 else if (ksm_advisor == KSM_ADVISOR_SCAN_TIME)
3770 output = "none [scan-time]";
3772 return sysfs_emit(buf, "%s\n", output);
3775 static ssize_t advisor_mode_store(struct kobject *kobj,
3776 struct kobj_attribute *attr, const char *buf,
3779 enum ksm_advisor_type curr_advisor = ksm_advisor;
3781 if (sysfs_streq("scan-time", buf))
3782 ksm_advisor = KSM_ADVISOR_SCAN_TIME;
3783 else if (sysfs_streq("none", buf))
3784 ksm_advisor = KSM_ADVISOR_NONE;
3788 /* Set advisor default values */
3789 if (curr_advisor != ksm_advisor)
3790 set_advisor_defaults();
3794 KSM_ATTR(advisor_mode);
3796 static ssize_t advisor_max_cpu_show(struct kobject *kobj,
3797 struct kobj_attribute *attr, char *buf)
3799 return sysfs_emit(buf, "%u\n", ksm_advisor_max_cpu);
3802 static ssize_t advisor_max_cpu_store(struct kobject *kobj,
3803 struct kobj_attribute *attr,
3804 const char *buf, size_t count)
3807 unsigned long value;
3809 err = kstrtoul(buf, 10, &value);
3813 ksm_advisor_max_cpu = value;
3816 KSM_ATTR(advisor_max_cpu);
3818 static ssize_t advisor_min_pages_to_scan_show(struct kobject *kobj,
3819 struct kobj_attribute *attr, char *buf)
3821 return sysfs_emit(buf, "%lu\n", ksm_advisor_min_pages_to_scan);
3824 static ssize_t advisor_min_pages_to_scan_store(struct kobject *kobj,
3825 struct kobj_attribute *attr,
3826 const char *buf, size_t count)
3829 unsigned long value;
3831 err = kstrtoul(buf, 10, &value);
3835 ksm_advisor_min_pages_to_scan = value;
3838 KSM_ATTR(advisor_min_pages_to_scan);
3840 static ssize_t advisor_max_pages_to_scan_show(struct kobject *kobj,
3841 struct kobj_attribute *attr, char *buf)
3843 return sysfs_emit(buf, "%lu\n", ksm_advisor_max_pages_to_scan);
3846 static ssize_t advisor_max_pages_to_scan_store(struct kobject *kobj,
3847 struct kobj_attribute *attr,
3848 const char *buf, size_t count)
3851 unsigned long value;
3853 err = kstrtoul(buf, 10, &value);
3857 ksm_advisor_max_pages_to_scan = value;
3860 KSM_ATTR(advisor_max_pages_to_scan);
3862 static ssize_t advisor_target_scan_time_show(struct kobject *kobj,
3863 struct kobj_attribute *attr, char *buf)
3865 return sysfs_emit(buf, "%lu\n", ksm_advisor_target_scan_time);
3868 static ssize_t advisor_target_scan_time_store(struct kobject *kobj,
3869 struct kobj_attribute *attr,
3870 const char *buf, size_t count)
3873 unsigned long value;
3875 err = kstrtoul(buf, 10, &value);
3881 ksm_advisor_target_scan_time = value;
3884 KSM_ATTR(advisor_target_scan_time);
3886 static struct attribute *ksm_attrs[] = {
3887 &sleep_millisecs_attr.attr,
3888 &pages_to_scan_attr.attr,
3890 &pages_scanned_attr.attr,
3891 &pages_shared_attr.attr,
3892 &pages_sharing_attr.attr,
3893 &pages_unshared_attr.attr,
3894 &pages_volatile_attr.attr,
3895 &pages_skipped_attr.attr,
3896 &ksm_zero_pages_attr.attr,
3897 &full_scans_attr.attr,
3899 &merge_across_nodes_attr.attr,
3901 &max_page_sharing_attr.attr,
3902 &stable_node_chains_attr.attr,
3903 &stable_node_dups_attr.attr,
3904 &stable_node_chains_prune_millisecs_attr.attr,
3905 &use_zero_pages_attr.attr,
3906 &general_profit_attr.attr,
3907 &smart_scan_attr.attr,
3908 &advisor_mode_attr.attr,
3909 &advisor_max_cpu_attr.attr,
3910 &advisor_min_pages_to_scan_attr.attr,
3911 &advisor_max_pages_to_scan_attr.attr,
3912 &advisor_target_scan_time_attr.attr,
3916 static const struct attribute_group ksm_attr_group = {
3920 #endif /* CONFIG_SYSFS */
3922 static int __init ksm_init(void)
3924 struct task_struct *ksm_thread;
3927 /* The correct value depends on page size and endianness */
3928 zero_checksum = calc_checksum(ZERO_PAGE(0));
3929 /* Default to false for backwards compatibility */
3930 ksm_use_zero_pages = false;
3932 err = ksm_slab_init();
3936 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3937 if (IS_ERR(ksm_thread)) {
3938 pr_err("ksm: creating kthread failed\n");
3939 err = PTR_ERR(ksm_thread);
3944 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3946 pr_err("ksm: register sysfs failed\n");
3947 kthread_stop(ksm_thread);
3951 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
3953 #endif /* CONFIG_SYSFS */
3955 #ifdef CONFIG_MEMORY_HOTREMOVE
3956 /* There is no significance to this priority 100 */
3957 hotplug_memory_notifier(ksm_memory_callback, KSM_CALLBACK_PRI);
3966 subsys_initcall(ksm_init);