17a338472SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 2f8af4da3SHugh Dickins /* 331dbd01fSIzik Eidus * Memory merging support. 431dbd01fSIzik Eidus * 531dbd01fSIzik Eidus * This code enables dynamic sharing of identical pages found in different 631dbd01fSIzik Eidus * memory areas, even if they are not shared by fork() 731dbd01fSIzik Eidus * 836b2528dSIzik Eidus * Copyright (C) 2008-2009 Red Hat, Inc. 931dbd01fSIzik Eidus * Authors: 1031dbd01fSIzik Eidus * Izik Eidus 1131dbd01fSIzik Eidus * Andrea Arcangeli 1231dbd01fSIzik Eidus * Chris Wright 1336b2528dSIzik Eidus * Hugh Dickins 14f8af4da3SHugh Dickins */ 15f8af4da3SHugh Dickins 16f8af4da3SHugh Dickins #include <linux/errno.h> 1731dbd01fSIzik Eidus #include <linux/mm.h> 1836090defSArnd Bergmann #include <linux/mm_inline.h> 1931dbd01fSIzik Eidus #include <linux/fs.h> 20f8af4da3SHugh Dickins #include <linux/mman.h> 2131dbd01fSIzik Eidus #include <linux/sched.h> 226e84f315SIngo Molnar #include <linux/sched/mm.h> 23f7ccbae4SIngo Molnar #include <linux/sched/coredump.h> 2431dbd01fSIzik Eidus #include <linux/rwsem.h> 2531dbd01fSIzik Eidus #include <linux/pagemap.h> 2631dbd01fSIzik Eidus #include <linux/rmap.h> 2731dbd01fSIzik Eidus #include <linux/spinlock.h> 2859e1a2f4STimofey Titovets #include <linux/xxhash.h> 2931dbd01fSIzik Eidus #include <linux/delay.h> 3031dbd01fSIzik Eidus #include <linux/kthread.h> 3131dbd01fSIzik Eidus #include <linux/wait.h> 3231dbd01fSIzik Eidus #include <linux/slab.h> 3331dbd01fSIzik Eidus #include <linux/rbtree.h> 3462b61f61SHugh Dickins #include <linux/memory.h> 3531dbd01fSIzik Eidus #include <linux/mmu_notifier.h> 362c6854fdSIzik Eidus #include <linux/swap.h> 37f8af4da3SHugh Dickins #include <linux/ksm.h> 384ca3a69bSSasha Levin #include <linux/hashtable.h> 39878aee7dSAndrea Arcangeli #include <linux/freezer.h> 4072788c38SDavid Rientjes #include <linux/oom.h> 4190bd6fd3SPetr Holasek #include <linux/numa.h> 42d7c0e68dSDavid Hildenbrand #include <linux/pagewalk.h> 43f8af4da3SHugh Dickins 4431dbd01fSIzik Eidus #include <asm/tlbflush.h> 4573848b46SHugh Dickins #include "internal.h" 4658730ab6SQi Zheng #include "mm_slot.h" 4731dbd01fSIzik Eidus 48739100c8SStefan Roesch #define CREATE_TRACE_POINTS 49739100c8SStefan Roesch #include <trace/events/ksm.h> 50739100c8SStefan Roesch 51e850dcf5SHugh Dickins #ifdef CONFIG_NUMA 52e850dcf5SHugh Dickins #define NUMA(x) (x) 53e850dcf5SHugh Dickins #define DO_NUMA(x) do { (x); } while (0) 54e850dcf5SHugh Dickins #else 55e850dcf5SHugh Dickins #define NUMA(x) (0) 56e850dcf5SHugh Dickins #define DO_NUMA(x) do { } while (0) 57e850dcf5SHugh Dickins #endif 58e850dcf5SHugh Dickins 595a2ca3efSMike Rapoport /** 605a2ca3efSMike Rapoport * DOC: Overview 615a2ca3efSMike Rapoport * 6231dbd01fSIzik Eidus * A few notes about the KSM scanning process, 6331dbd01fSIzik Eidus * to make it easier to understand the data structures below: 6431dbd01fSIzik Eidus * 6531dbd01fSIzik Eidus * In order to reduce excessive scanning, KSM sorts the memory pages by their 6631dbd01fSIzik Eidus * contents into a data structure that holds pointers to the pages' locations. 6731dbd01fSIzik Eidus * 6831dbd01fSIzik Eidus * Since the contents of the pages may change at any moment, KSM cannot just 6931dbd01fSIzik Eidus * insert the pages into a normal sorted tree and expect it to find anything. 7031dbd01fSIzik Eidus * Therefore KSM uses two data structures - the stable and the unstable tree. 7131dbd01fSIzik Eidus * 7231dbd01fSIzik Eidus * The stable tree holds pointers to all the merged pages (ksm pages), sorted 7331dbd01fSIzik Eidus * by their contents. Because each such page is write-protected, searching on 7431dbd01fSIzik Eidus * this tree is fully assured to be working (except when pages are unmapped), 7531dbd01fSIzik Eidus * and therefore this tree is called the stable tree. 7631dbd01fSIzik Eidus * 775a2ca3efSMike Rapoport * The stable tree node includes information required for reverse 785a2ca3efSMike Rapoport * mapping from a KSM page to virtual addresses that map this page. 795a2ca3efSMike Rapoport * 805a2ca3efSMike Rapoport * In order to avoid large latencies of the rmap walks on KSM pages, 815a2ca3efSMike Rapoport * KSM maintains two types of nodes in the stable tree: 825a2ca3efSMike Rapoport * 835a2ca3efSMike Rapoport * * the regular nodes that keep the reverse mapping structures in a 845a2ca3efSMike Rapoport * linked list 855a2ca3efSMike Rapoport * * the "chains" that link nodes ("dups") that represent the same 865a2ca3efSMike Rapoport * write protected memory content, but each "dup" corresponds to a 875a2ca3efSMike Rapoport * different KSM page copy of that content 885a2ca3efSMike Rapoport * 895a2ca3efSMike Rapoport * Internally, the regular nodes, "dups" and "chains" are represented 9021fbd591SQi Zheng * using the same struct ksm_stable_node structure. 915a2ca3efSMike Rapoport * 9231dbd01fSIzik Eidus * In addition to the stable tree, KSM uses a second data structure called the 9331dbd01fSIzik Eidus * unstable tree: this tree holds pointers to pages which have been found to 9431dbd01fSIzik Eidus * be "unchanged for a period of time". The unstable tree sorts these pages 9531dbd01fSIzik Eidus * by their contents, but since they are not write-protected, KSM cannot rely 9631dbd01fSIzik Eidus * upon the unstable tree to work correctly - the unstable tree is liable to 9731dbd01fSIzik Eidus * be corrupted as its contents are modified, and so it is called unstable. 9831dbd01fSIzik Eidus * 9931dbd01fSIzik Eidus * KSM solves this problem by several techniques: 10031dbd01fSIzik Eidus * 10131dbd01fSIzik Eidus * 1) The unstable tree is flushed every time KSM completes scanning all 10231dbd01fSIzik Eidus * memory areas, and then the tree is rebuilt again from the beginning. 10331dbd01fSIzik Eidus * 2) KSM will only insert into the unstable tree, pages whose hash value 10431dbd01fSIzik Eidus * has not changed since the previous scan of all memory areas. 10531dbd01fSIzik Eidus * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the 10631dbd01fSIzik Eidus * colors of the nodes and not on their contents, assuring that even when 10731dbd01fSIzik Eidus * the tree gets "corrupted" it won't get out of balance, so scanning time 10831dbd01fSIzik Eidus * remains the same (also, searching and inserting nodes in an rbtree uses 10931dbd01fSIzik Eidus * the same algorithm, so we have no overhead when we flush and rebuild). 11031dbd01fSIzik Eidus * 4) KSM never flushes the stable tree, which means that even if it were to 11131dbd01fSIzik Eidus * take 10 attempts to find a page in the unstable tree, once it is found, 11231dbd01fSIzik Eidus * it is secured in the stable tree. (When we scan a new page, we first 11331dbd01fSIzik Eidus * compare it against the stable tree, and then against the unstable tree.) 1148fdb3dbfSHugh Dickins * 1158fdb3dbfSHugh Dickins * If the merge_across_nodes tunable is unset, then KSM maintains multiple 1168fdb3dbfSHugh Dickins * stable trees and multiple unstable trees: one of each for each NUMA node. 11731dbd01fSIzik Eidus */ 11831dbd01fSIzik Eidus 11931dbd01fSIzik Eidus /** 12021fbd591SQi Zheng * struct ksm_mm_slot - ksm information per mm that is being scanned 12158730ab6SQi Zheng * @slot: hash lookup from mm to mm_slot 1226514d511SHugh Dickins * @rmap_list: head for this mm_slot's singly-linked list of rmap_items 12331dbd01fSIzik Eidus */ 12421fbd591SQi Zheng struct ksm_mm_slot { 12558730ab6SQi Zheng struct mm_slot slot; 12621fbd591SQi Zheng struct ksm_rmap_item *rmap_list; 12731dbd01fSIzik Eidus }; 12831dbd01fSIzik Eidus 12931dbd01fSIzik Eidus /** 13031dbd01fSIzik Eidus * struct ksm_scan - cursor for scanning 13131dbd01fSIzik Eidus * @mm_slot: the current mm_slot we are scanning 13231dbd01fSIzik Eidus * @address: the next address inside that to be scanned 1336514d511SHugh Dickins * @rmap_list: link to the next rmap to be scanned in the rmap_list 13431dbd01fSIzik Eidus * @seqnr: count of completed full scans (needed when removing unstable node) 13531dbd01fSIzik Eidus * 13631dbd01fSIzik Eidus * There is only the one ksm_scan instance of this cursor structure. 13731dbd01fSIzik Eidus */ 13831dbd01fSIzik Eidus struct ksm_scan { 13921fbd591SQi Zheng struct ksm_mm_slot *mm_slot; 14031dbd01fSIzik Eidus unsigned long address; 14121fbd591SQi Zheng struct ksm_rmap_item **rmap_list; 14231dbd01fSIzik Eidus unsigned long seqnr; 14331dbd01fSIzik Eidus }; 14431dbd01fSIzik Eidus 14531dbd01fSIzik Eidus /** 14621fbd591SQi Zheng * struct ksm_stable_node - node of the stable rbtree 1477b6ba2c7SHugh Dickins * @node: rb node of this ksm page in the stable tree 1484146d2d6SHugh Dickins * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list 1492c653d0eSAndrea Arcangeli * @hlist_dup: linked into the stable_node->hlist with a stable_node chain 1504146d2d6SHugh Dickins * @list: linked into migrate_nodes, pending placement in the proper node tree 1517b6ba2c7SHugh Dickins * @hlist: hlist head of rmap_items using this ksm page 1524146d2d6SHugh Dickins * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid) 1532c653d0eSAndrea Arcangeli * @chain_prune_time: time of the last full garbage collection 1542c653d0eSAndrea Arcangeli * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN 1554146d2d6SHugh Dickins * @nid: NUMA node id of stable tree in which linked (may not match kpfn) 1567b6ba2c7SHugh Dickins */ 15721fbd591SQi Zheng struct ksm_stable_node { 1584146d2d6SHugh Dickins union { 1594146d2d6SHugh Dickins struct rb_node node; /* when node of stable tree */ 1604146d2d6SHugh Dickins struct { /* when listed for migration */ 1614146d2d6SHugh Dickins struct list_head *head; 1622c653d0eSAndrea Arcangeli struct { 1632c653d0eSAndrea Arcangeli struct hlist_node hlist_dup; 1644146d2d6SHugh Dickins struct list_head list; 1654146d2d6SHugh Dickins }; 1664146d2d6SHugh Dickins }; 1672c653d0eSAndrea Arcangeli }; 1687b6ba2c7SHugh Dickins struct hlist_head hlist; 1692c653d0eSAndrea Arcangeli union { 17062b61f61SHugh Dickins unsigned long kpfn; 1712c653d0eSAndrea Arcangeli unsigned long chain_prune_time; 1722c653d0eSAndrea Arcangeli }; 1732c653d0eSAndrea Arcangeli /* 1742c653d0eSAndrea Arcangeli * STABLE_NODE_CHAIN can be any negative number in 1752c653d0eSAndrea Arcangeli * rmap_hlist_len negative range, but better not -1 to be able 1762c653d0eSAndrea Arcangeli * to reliably detect underflows. 1772c653d0eSAndrea Arcangeli */ 1782c653d0eSAndrea Arcangeli #define STABLE_NODE_CHAIN -1024 1792c653d0eSAndrea Arcangeli int rmap_hlist_len; 1804146d2d6SHugh Dickins #ifdef CONFIG_NUMA 1814146d2d6SHugh Dickins int nid; 1824146d2d6SHugh Dickins #endif 1837b6ba2c7SHugh Dickins }; 1847b6ba2c7SHugh Dickins 1857b6ba2c7SHugh Dickins /** 18621fbd591SQi Zheng * struct ksm_rmap_item - reverse mapping item for virtual addresses 1876514d511SHugh Dickins * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list 188db114b83SHugh Dickins * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree 189bc56620bSHugh Dickins * @nid: NUMA node id of unstable tree in which linked (may not match page) 19031dbd01fSIzik Eidus * @mm: the memory structure this rmap_item is pointing into 19131dbd01fSIzik Eidus * @address: the virtual address this rmap_item tracks (+ flags in low bits) 19231dbd01fSIzik Eidus * @oldchecksum: previous checksum of the page at that virtual address 1937b6ba2c7SHugh Dickins * @node: rb node of this rmap_item in the unstable tree 1947b6ba2c7SHugh Dickins * @head: pointer to stable_node heading this list in the stable tree 1957b6ba2c7SHugh Dickins * @hlist: link into hlist of rmap_items hanging off that stable_node 19631dbd01fSIzik Eidus */ 19721fbd591SQi Zheng struct ksm_rmap_item { 19821fbd591SQi Zheng struct ksm_rmap_item *rmap_list; 199bc56620bSHugh Dickins union { 200db114b83SHugh Dickins struct anon_vma *anon_vma; /* when stable */ 201bc56620bSHugh Dickins #ifdef CONFIG_NUMA 202bc56620bSHugh Dickins int nid; /* when node of unstable tree */ 203bc56620bSHugh Dickins #endif 204bc56620bSHugh Dickins }; 20531dbd01fSIzik Eidus struct mm_struct *mm; 20631dbd01fSIzik Eidus unsigned long address; /* + low bits used for flags below */ 20731dbd01fSIzik Eidus unsigned int oldchecksum; /* when unstable */ 20831dbd01fSIzik Eidus union { 2097b6ba2c7SHugh Dickins struct rb_node node; /* when node of unstable tree */ 2107b6ba2c7SHugh Dickins struct { /* when listed from stable tree */ 21121fbd591SQi Zheng struct ksm_stable_node *head; 2127b6ba2c7SHugh Dickins struct hlist_node hlist; 2137b6ba2c7SHugh Dickins }; 21431dbd01fSIzik Eidus }; 21531dbd01fSIzik Eidus }; 21631dbd01fSIzik Eidus 21731dbd01fSIzik Eidus #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ 2187b6ba2c7SHugh Dickins #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ 2197b6ba2c7SHugh Dickins #define STABLE_FLAG 0x200 /* is listed from the stable tree */ 22031dbd01fSIzik Eidus 22131dbd01fSIzik Eidus /* The stable and unstable tree heads */ 222ef53d16cSHugh Dickins static struct rb_root one_stable_tree[1] = { RB_ROOT }; 223ef53d16cSHugh Dickins static struct rb_root one_unstable_tree[1] = { RB_ROOT }; 224ef53d16cSHugh Dickins static struct rb_root *root_stable_tree = one_stable_tree; 225ef53d16cSHugh Dickins static struct rb_root *root_unstable_tree = one_unstable_tree; 22631dbd01fSIzik Eidus 2274146d2d6SHugh Dickins /* Recently migrated nodes of stable tree, pending proper placement */ 2284146d2d6SHugh Dickins static LIST_HEAD(migrate_nodes); 2292c653d0eSAndrea Arcangeli #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev) 2304146d2d6SHugh Dickins 2314ca3a69bSSasha Levin #define MM_SLOTS_HASH_BITS 10 2324ca3a69bSSasha Levin static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 23331dbd01fSIzik Eidus 23421fbd591SQi Zheng static struct ksm_mm_slot ksm_mm_head = { 23558730ab6SQi Zheng .slot.mm_node = LIST_HEAD_INIT(ksm_mm_head.slot.mm_node), 23631dbd01fSIzik Eidus }; 23731dbd01fSIzik Eidus static struct ksm_scan ksm_scan = { 23831dbd01fSIzik Eidus .mm_slot = &ksm_mm_head, 23931dbd01fSIzik Eidus }; 24031dbd01fSIzik Eidus 24131dbd01fSIzik Eidus static struct kmem_cache *rmap_item_cache; 2427b6ba2c7SHugh Dickins static struct kmem_cache *stable_node_cache; 24331dbd01fSIzik Eidus static struct kmem_cache *mm_slot_cache; 24431dbd01fSIzik Eidus 24531dbd01fSIzik Eidus /* The number of nodes in the stable tree */ 246b4028260SHugh Dickins static unsigned long ksm_pages_shared; 24731dbd01fSIzik Eidus 248e178dfdeSHugh Dickins /* The number of page slots additionally sharing those nodes */ 249b4028260SHugh Dickins static unsigned long ksm_pages_sharing; 25031dbd01fSIzik Eidus 251473b0ce4SHugh Dickins /* The number of nodes in the unstable tree */ 252473b0ce4SHugh Dickins static unsigned long ksm_pages_unshared; 253473b0ce4SHugh Dickins 254473b0ce4SHugh Dickins /* The number of rmap_items in use: to calculate pages_volatile */ 255473b0ce4SHugh Dickins static unsigned long ksm_rmap_items; 256473b0ce4SHugh Dickins 2572c653d0eSAndrea Arcangeli /* The number of stable_node chains */ 2582c653d0eSAndrea Arcangeli static unsigned long ksm_stable_node_chains; 2592c653d0eSAndrea Arcangeli 2602c653d0eSAndrea Arcangeli /* The number of stable_node dups linked to the stable_node chains */ 2612c653d0eSAndrea Arcangeli static unsigned long ksm_stable_node_dups; 2622c653d0eSAndrea Arcangeli 2632c653d0eSAndrea Arcangeli /* Delay in pruning stale stable_node_dups in the stable_node_chains */ 264584ff0dfSZhansaya Bagdauletkyzy static unsigned int ksm_stable_node_chains_prune_millisecs = 2000; 2652c653d0eSAndrea Arcangeli 2662c653d0eSAndrea Arcangeli /* Maximum number of page slots sharing a stable node */ 2672c653d0eSAndrea Arcangeli static int ksm_max_page_sharing = 256; 2682c653d0eSAndrea Arcangeli 26931dbd01fSIzik Eidus /* Number of pages ksmd should scan in one batch */ 2702c6854fdSIzik Eidus static unsigned int ksm_thread_pages_to_scan = 100; 27131dbd01fSIzik Eidus 27231dbd01fSIzik Eidus /* Milliseconds ksmd should sleep between batches */ 2732ffd8679SHugh Dickins static unsigned int ksm_thread_sleep_millisecs = 20; 27431dbd01fSIzik Eidus 275e86c59b1SClaudio Imbrenda /* Checksum of an empty (zeroed) page */ 276e86c59b1SClaudio Imbrenda static unsigned int zero_checksum __read_mostly; 277e86c59b1SClaudio Imbrenda 278e86c59b1SClaudio Imbrenda /* Whether to merge empty (zeroed) pages with actual zero pages */ 279e86c59b1SClaudio Imbrenda static bool ksm_use_zero_pages __read_mostly; 280e86c59b1SClaudio Imbrenda 281e850dcf5SHugh Dickins #ifdef CONFIG_NUMA 28290bd6fd3SPetr Holasek /* Zeroed when merging across nodes is not allowed */ 28390bd6fd3SPetr Holasek static unsigned int ksm_merge_across_nodes = 1; 284ef53d16cSHugh Dickins static int ksm_nr_node_ids = 1; 285e850dcf5SHugh Dickins #else 286e850dcf5SHugh Dickins #define ksm_merge_across_nodes 1U 287ef53d16cSHugh Dickins #define ksm_nr_node_ids 1 288e850dcf5SHugh Dickins #endif 28990bd6fd3SPetr Holasek 29031dbd01fSIzik Eidus #define KSM_RUN_STOP 0 29131dbd01fSIzik Eidus #define KSM_RUN_MERGE 1 29231dbd01fSIzik Eidus #define KSM_RUN_UNMERGE 2 293ef4d43a8SHugh Dickins #define KSM_RUN_OFFLINE 4 294ef4d43a8SHugh Dickins static unsigned long ksm_run = KSM_RUN_STOP; 295ef4d43a8SHugh Dickins static void wait_while_offlining(void); 29631dbd01fSIzik Eidus 29731dbd01fSIzik Eidus static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); 298fcf9a0efSKirill Tkhai static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait); 29931dbd01fSIzik Eidus static DEFINE_MUTEX(ksm_thread_mutex); 30031dbd01fSIzik Eidus static DEFINE_SPINLOCK(ksm_mmlist_lock); 30131dbd01fSIzik Eidus 30221fbd591SQi Zheng #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create(#__struct,\ 30331dbd01fSIzik Eidus sizeof(struct __struct), __alignof__(struct __struct),\ 30431dbd01fSIzik Eidus (__flags), NULL) 30531dbd01fSIzik Eidus 30631dbd01fSIzik Eidus static int __init ksm_slab_init(void) 30731dbd01fSIzik Eidus { 30821fbd591SQi Zheng rmap_item_cache = KSM_KMEM_CACHE(ksm_rmap_item, 0); 30931dbd01fSIzik Eidus if (!rmap_item_cache) 31031dbd01fSIzik Eidus goto out; 31131dbd01fSIzik Eidus 31221fbd591SQi Zheng stable_node_cache = KSM_KMEM_CACHE(ksm_stable_node, 0); 3137b6ba2c7SHugh Dickins if (!stable_node_cache) 3147b6ba2c7SHugh Dickins goto out_free1; 3157b6ba2c7SHugh Dickins 31621fbd591SQi Zheng mm_slot_cache = KSM_KMEM_CACHE(ksm_mm_slot, 0); 31731dbd01fSIzik Eidus if (!mm_slot_cache) 3187b6ba2c7SHugh Dickins goto out_free2; 31931dbd01fSIzik Eidus 32031dbd01fSIzik Eidus return 0; 32131dbd01fSIzik Eidus 3227b6ba2c7SHugh Dickins out_free2: 3237b6ba2c7SHugh Dickins kmem_cache_destroy(stable_node_cache); 3247b6ba2c7SHugh Dickins out_free1: 32531dbd01fSIzik Eidus kmem_cache_destroy(rmap_item_cache); 32631dbd01fSIzik Eidus out: 32731dbd01fSIzik Eidus return -ENOMEM; 32831dbd01fSIzik Eidus } 32931dbd01fSIzik Eidus 33031dbd01fSIzik Eidus static void __init ksm_slab_free(void) 33131dbd01fSIzik Eidus { 33231dbd01fSIzik Eidus kmem_cache_destroy(mm_slot_cache); 3337b6ba2c7SHugh Dickins kmem_cache_destroy(stable_node_cache); 33431dbd01fSIzik Eidus kmem_cache_destroy(rmap_item_cache); 33531dbd01fSIzik Eidus mm_slot_cache = NULL; 33631dbd01fSIzik Eidus } 33731dbd01fSIzik Eidus 33821fbd591SQi Zheng static __always_inline bool is_stable_node_chain(struct ksm_stable_node *chain) 3392c653d0eSAndrea Arcangeli { 3402c653d0eSAndrea Arcangeli return chain->rmap_hlist_len == STABLE_NODE_CHAIN; 3412c653d0eSAndrea Arcangeli } 3422c653d0eSAndrea Arcangeli 34321fbd591SQi Zheng static __always_inline bool is_stable_node_dup(struct ksm_stable_node *dup) 3442c653d0eSAndrea Arcangeli { 3452c653d0eSAndrea Arcangeli return dup->head == STABLE_NODE_DUP_HEAD; 3462c653d0eSAndrea Arcangeli } 3472c653d0eSAndrea Arcangeli 34821fbd591SQi Zheng static inline void stable_node_chain_add_dup(struct ksm_stable_node *dup, 34921fbd591SQi Zheng struct ksm_stable_node *chain) 3502c653d0eSAndrea Arcangeli { 3512c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(dup)); 3522c653d0eSAndrea Arcangeli dup->head = STABLE_NODE_DUP_HEAD; 3532c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_chain(chain)); 3542c653d0eSAndrea Arcangeli hlist_add_head(&dup->hlist_dup, &chain->hlist); 3552c653d0eSAndrea Arcangeli ksm_stable_node_dups++; 3562c653d0eSAndrea Arcangeli } 3572c653d0eSAndrea Arcangeli 35821fbd591SQi Zheng static inline void __stable_node_dup_del(struct ksm_stable_node *dup) 3592c653d0eSAndrea Arcangeli { 360b4fecc67SAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(dup)); 3612c653d0eSAndrea Arcangeli hlist_del(&dup->hlist_dup); 3622c653d0eSAndrea Arcangeli ksm_stable_node_dups--; 3632c653d0eSAndrea Arcangeli } 3642c653d0eSAndrea Arcangeli 36521fbd591SQi Zheng static inline void stable_node_dup_del(struct ksm_stable_node *dup) 3662c653d0eSAndrea Arcangeli { 3672c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_chain(dup)); 3682c653d0eSAndrea Arcangeli if (is_stable_node_dup(dup)) 3692c653d0eSAndrea Arcangeli __stable_node_dup_del(dup); 3702c653d0eSAndrea Arcangeli else 3712c653d0eSAndrea Arcangeli rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid)); 3722c653d0eSAndrea Arcangeli #ifdef CONFIG_DEBUG_VM 3732c653d0eSAndrea Arcangeli dup->head = NULL; 3742c653d0eSAndrea Arcangeli #endif 3752c653d0eSAndrea Arcangeli } 3762c653d0eSAndrea Arcangeli 37721fbd591SQi Zheng static inline struct ksm_rmap_item *alloc_rmap_item(void) 37831dbd01fSIzik Eidus { 37921fbd591SQi Zheng struct ksm_rmap_item *rmap_item; 380473b0ce4SHugh Dickins 3815b398e41Szhong jiang rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL | 3825b398e41Szhong jiang __GFP_NORETRY | __GFP_NOWARN); 383473b0ce4SHugh Dickins if (rmap_item) 384473b0ce4SHugh Dickins ksm_rmap_items++; 385473b0ce4SHugh Dickins return rmap_item; 38631dbd01fSIzik Eidus } 38731dbd01fSIzik Eidus 38821fbd591SQi Zheng static inline void free_rmap_item(struct ksm_rmap_item *rmap_item) 38931dbd01fSIzik Eidus { 390473b0ce4SHugh Dickins ksm_rmap_items--; 391cb4df4caSxu xin rmap_item->mm->ksm_rmap_items--; 39231dbd01fSIzik Eidus rmap_item->mm = NULL; /* debug safety */ 39331dbd01fSIzik Eidus kmem_cache_free(rmap_item_cache, rmap_item); 39431dbd01fSIzik Eidus } 39531dbd01fSIzik Eidus 39621fbd591SQi Zheng static inline struct ksm_stable_node *alloc_stable_node(void) 3977b6ba2c7SHugh Dickins { 3986213055fSzhong jiang /* 3996213055fSzhong jiang * The allocation can take too long with GFP_KERNEL when memory is under 4006213055fSzhong jiang * pressure, which may lead to hung task warnings. Adding __GFP_HIGH 4016213055fSzhong jiang * grants access to memory reserves, helping to avoid this problem. 4026213055fSzhong jiang */ 4036213055fSzhong jiang return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH); 4047b6ba2c7SHugh Dickins } 4057b6ba2c7SHugh Dickins 40621fbd591SQi Zheng static inline void free_stable_node(struct ksm_stable_node *stable_node) 4077b6ba2c7SHugh Dickins { 4082c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len && 4092c653d0eSAndrea Arcangeli !is_stable_node_chain(stable_node)); 4107b6ba2c7SHugh Dickins kmem_cache_free(stable_node_cache, stable_node); 4117b6ba2c7SHugh Dickins } 4127b6ba2c7SHugh Dickins 41331dbd01fSIzik Eidus /* 414a913e182SHugh Dickins * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's 415a913e182SHugh Dickins * page tables after it has passed through ksm_exit() - which, if necessary, 416c1e8d7c6SMichel Lespinasse * takes mmap_lock briefly to serialize against them. ksm_exit() does not set 417a913e182SHugh Dickins * a special flag: they can just back out as soon as mm_users goes to zero. 418a913e182SHugh Dickins * ksm_test_exit() is used throughout to make this test for exit: in some 419a913e182SHugh Dickins * places for correctness, in some places just to avoid unnecessary work. 420a913e182SHugh Dickins */ 421a913e182SHugh Dickins static inline bool ksm_test_exit(struct mm_struct *mm) 422a913e182SHugh Dickins { 423a913e182SHugh Dickins return atomic_read(&mm->mm_users) == 0; 424a913e182SHugh Dickins } 425a913e182SHugh Dickins 426d7c0e68dSDavid Hildenbrand static int break_ksm_pmd_entry(pmd_t *pmd, unsigned long addr, unsigned long next, 427d7c0e68dSDavid Hildenbrand struct mm_walk *walk) 428d7c0e68dSDavid Hildenbrand { 429d7c0e68dSDavid Hildenbrand struct page *page = NULL; 430d7c0e68dSDavid Hildenbrand spinlock_t *ptl; 431d7c0e68dSDavid Hildenbrand pte_t *pte; 432*c33c7948SRyan Roberts pte_t ptent; 433d7c0e68dSDavid Hildenbrand int ret; 434d7c0e68dSDavid Hildenbrand 435d7c0e68dSDavid Hildenbrand pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl); 43604dee9e8SHugh Dickins if (!pte) 43704dee9e8SHugh Dickins return 0; 438*c33c7948SRyan Roberts ptent = ptep_get(pte); 439*c33c7948SRyan Roberts if (pte_present(ptent)) { 440*c33c7948SRyan Roberts page = vm_normal_page(walk->vma, addr, ptent); 441*c33c7948SRyan Roberts } else if (!pte_none(ptent)) { 442*c33c7948SRyan Roberts swp_entry_t entry = pte_to_swp_entry(ptent); 443d7c0e68dSDavid Hildenbrand 444d7c0e68dSDavid Hildenbrand /* 445d7c0e68dSDavid Hildenbrand * As KSM pages remain KSM pages until freed, no need to wait 446d7c0e68dSDavid Hildenbrand * here for migration to end. 447d7c0e68dSDavid Hildenbrand */ 448d7c0e68dSDavid Hildenbrand if (is_migration_entry(entry)) 449d7c0e68dSDavid Hildenbrand page = pfn_swap_entry_to_page(entry); 450d7c0e68dSDavid Hildenbrand } 451d7c0e68dSDavid Hildenbrand ret = page && PageKsm(page); 452d7c0e68dSDavid Hildenbrand pte_unmap_unlock(pte, ptl); 453d7c0e68dSDavid Hildenbrand return ret; 454d7c0e68dSDavid Hildenbrand } 455d7c0e68dSDavid Hildenbrand 456d7c0e68dSDavid Hildenbrand static const struct mm_walk_ops break_ksm_ops = { 457d7c0e68dSDavid Hildenbrand .pmd_entry = break_ksm_pmd_entry, 458d7c0e68dSDavid Hildenbrand }; 459d7c0e68dSDavid Hildenbrand 460a913e182SHugh Dickins /* 4616cce3314SDavid Hildenbrand * We use break_ksm to break COW on a ksm page by triggering unsharing, 4626cce3314SDavid Hildenbrand * such that the ksm page will get replaced by an exclusive anonymous page. 46331dbd01fSIzik Eidus * 4646cce3314SDavid Hildenbrand * We take great care only to touch a ksm page, in a VM_MERGEABLE vma, 46531dbd01fSIzik Eidus * in case the application has unmapped and remapped mm,addr meanwhile. 46631dbd01fSIzik Eidus * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP 467bbcd53c9SDavid Hildenbrand * mmap of /dev/mem, where we would not want to touch it. 4681b2ee126SDave Hansen * 4696cce3314SDavid Hildenbrand * FAULT_FLAG_REMOTE/FOLL_REMOTE are because we do this outside the context 4701b2ee126SDave Hansen * of the process that owns 'vma'. We also do not want to enforce 4711b2ee126SDave Hansen * protection keys here anyway. 47231dbd01fSIzik Eidus */ 473d952b791SHugh Dickins static int break_ksm(struct vm_area_struct *vma, unsigned long addr) 47431dbd01fSIzik Eidus { 47550a7ca3cSSouptick Joarder vm_fault_t ret = 0; 47631dbd01fSIzik Eidus 47731dbd01fSIzik Eidus do { 478d7c0e68dSDavid Hildenbrand int ksm_page; 47958f595c6SDavid Hildenbrand 48031dbd01fSIzik Eidus cond_resched(); 481d7c0e68dSDavid Hildenbrand ksm_page = walk_page_range_vma(vma, addr, addr + 1, 482d7c0e68dSDavid Hildenbrand &break_ksm_ops, NULL); 483d7c0e68dSDavid Hildenbrand if (WARN_ON_ONCE(ksm_page < 0)) 484d7c0e68dSDavid Hildenbrand return ksm_page; 48558f595c6SDavid Hildenbrand if (!ksm_page) 48658f595c6SDavid Hildenbrand return 0; 487dcddffd4SKirill A. Shutemov ret = handle_mm_fault(vma, addr, 4886cce3314SDavid Hildenbrand FAULT_FLAG_UNSHARE | FAULT_FLAG_REMOTE, 489bce617edSPeter Xu NULL); 49058f595c6SDavid Hildenbrand } while (!(ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM))); 491d952b791SHugh Dickins /* 49258f595c6SDavid Hildenbrand * We must loop until we no longer find a KSM page because 49358f595c6SDavid Hildenbrand * handle_mm_fault() may back out if there's any difficulty e.g. if 49458f595c6SDavid Hildenbrand * pte accessed bit gets updated concurrently. 495d952b791SHugh Dickins * 496d952b791SHugh Dickins * VM_FAULT_SIGBUS could occur if we race with truncation of the 497d952b791SHugh Dickins * backing file, which also invalidates anonymous pages: that's 498d952b791SHugh Dickins * okay, that truncation will have unmapped the PageKsm for us. 499d952b791SHugh Dickins * 500d952b791SHugh Dickins * VM_FAULT_OOM: at the time of writing (late July 2009), setting 501d952b791SHugh Dickins * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the 502d952b791SHugh Dickins * current task has TIF_MEMDIE set, and will be OOM killed on return 503d952b791SHugh Dickins * to user; and ksmd, having no mm, would never be chosen for that. 504d952b791SHugh Dickins * 505d952b791SHugh Dickins * But if the mm is in a limited mem_cgroup, then the fault may fail 506d952b791SHugh Dickins * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and 507d952b791SHugh Dickins * even ksmd can fail in this way - though it's usually breaking ksm 508d952b791SHugh Dickins * just to undo a merge it made a moment before, so unlikely to oom. 509d952b791SHugh Dickins * 510d952b791SHugh Dickins * That's a pity: we might therefore have more kernel pages allocated 511d952b791SHugh Dickins * than we're counting as nodes in the stable tree; but ksm_do_scan 512d952b791SHugh Dickins * will retry to break_cow on each pass, so should recover the page 513d952b791SHugh Dickins * in due course. The important thing is to not let VM_MERGEABLE 514d952b791SHugh Dickins * be cleared while any such pages might remain in the area. 515d952b791SHugh Dickins */ 516d952b791SHugh Dickins return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; 51731dbd01fSIzik Eidus } 51831dbd01fSIzik Eidus 519d7597f59SStefan Roesch static bool vma_ksm_compatible(struct vm_area_struct *vma) 520d7597f59SStefan Roesch { 521d7597f59SStefan Roesch if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE | VM_PFNMAP | 522d7597f59SStefan Roesch VM_IO | VM_DONTEXPAND | VM_HUGETLB | 523d7597f59SStefan Roesch VM_MIXEDMAP)) 524d7597f59SStefan Roesch return false; /* just ignore the advice */ 525d7597f59SStefan Roesch 526d7597f59SStefan Roesch if (vma_is_dax(vma)) 527d7597f59SStefan Roesch return false; 528d7597f59SStefan Roesch 529d7597f59SStefan Roesch #ifdef VM_SAO 530d7597f59SStefan Roesch if (vma->vm_flags & VM_SAO) 531d7597f59SStefan Roesch return false; 532d7597f59SStefan Roesch #endif 533d7597f59SStefan Roesch #ifdef VM_SPARC_ADI 534d7597f59SStefan Roesch if (vma->vm_flags & VM_SPARC_ADI) 535d7597f59SStefan Roesch return false; 536d7597f59SStefan Roesch #endif 537d7597f59SStefan Roesch 538d7597f59SStefan Roesch return true; 539d7597f59SStefan Roesch } 540d7597f59SStefan Roesch 541ef694222SBob Liu static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm, 542ef694222SBob Liu unsigned long addr) 543ef694222SBob Liu { 544ef694222SBob Liu struct vm_area_struct *vma; 545ef694222SBob Liu if (ksm_test_exit(mm)) 546ef694222SBob Liu return NULL; 547ff69fb81SLiam Howlett vma = vma_lookup(mm, addr); 548ff69fb81SLiam Howlett if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) 549ef694222SBob Liu return NULL; 550ef694222SBob Liu return vma; 551ef694222SBob Liu } 552ef694222SBob Liu 55321fbd591SQi Zheng static void break_cow(struct ksm_rmap_item *rmap_item) 55431dbd01fSIzik Eidus { 5558dd3557aSHugh Dickins struct mm_struct *mm = rmap_item->mm; 5568dd3557aSHugh Dickins unsigned long addr = rmap_item->address; 55731dbd01fSIzik Eidus struct vm_area_struct *vma; 55831dbd01fSIzik Eidus 5594035c07aSHugh Dickins /* 5604035c07aSHugh Dickins * It is not an accident that whenever we want to break COW 5614035c07aSHugh Dickins * to undo, we also need to drop a reference to the anon_vma. 5624035c07aSHugh Dickins */ 5639e60109fSPeter Zijlstra put_anon_vma(rmap_item->anon_vma); 5644035c07aSHugh Dickins 565d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 566ef694222SBob Liu vma = find_mergeable_vma(mm, addr); 567ef694222SBob Liu if (vma) 56831dbd01fSIzik Eidus break_ksm(vma, addr); 569d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 57031dbd01fSIzik Eidus } 57131dbd01fSIzik Eidus 57221fbd591SQi Zheng static struct page *get_mergeable_page(struct ksm_rmap_item *rmap_item) 57331dbd01fSIzik Eidus { 57431dbd01fSIzik Eidus struct mm_struct *mm = rmap_item->mm; 57531dbd01fSIzik Eidus unsigned long addr = rmap_item->address; 57631dbd01fSIzik Eidus struct vm_area_struct *vma; 57731dbd01fSIzik Eidus struct page *page; 57831dbd01fSIzik Eidus 579d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 580ef694222SBob Liu vma = find_mergeable_vma(mm, addr); 581ef694222SBob Liu if (!vma) 58231dbd01fSIzik Eidus goto out; 58331dbd01fSIzik Eidus 58431dbd01fSIzik Eidus page = follow_page(vma, addr, FOLL_GET); 585f7091ed6SHaiyue Wang if (IS_ERR_OR_NULL(page)) 58631dbd01fSIzik Eidus goto out; 587f7091ed6SHaiyue Wang if (is_zone_device_page(page)) 588f7091ed6SHaiyue Wang goto out_putpage; 589f765f540SKirill A. Shutemov if (PageAnon(page)) { 59031dbd01fSIzik Eidus flush_anon_page(vma, page, addr); 59131dbd01fSIzik Eidus flush_dcache_page(page); 59231dbd01fSIzik Eidus } else { 593f7091ed6SHaiyue Wang out_putpage: 59431dbd01fSIzik Eidus put_page(page); 595c8f95ed1SAndrea Arcangeli out: 596c8f95ed1SAndrea Arcangeli page = NULL; 59731dbd01fSIzik Eidus } 598d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 59931dbd01fSIzik Eidus return page; 60031dbd01fSIzik Eidus } 60131dbd01fSIzik Eidus 60290bd6fd3SPetr Holasek /* 60390bd6fd3SPetr Holasek * This helper is used for getting right index into array of tree roots. 60490bd6fd3SPetr Holasek * When merge_across_nodes knob is set to 1, there are only two rb-trees for 60590bd6fd3SPetr Holasek * stable and unstable pages from all nodes with roots in index 0. Otherwise, 60690bd6fd3SPetr Holasek * every node has its own stable and unstable tree. 60790bd6fd3SPetr Holasek */ 60890bd6fd3SPetr Holasek static inline int get_kpfn_nid(unsigned long kpfn) 60990bd6fd3SPetr Holasek { 610d8fc16a8SHugh Dickins return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn)); 61190bd6fd3SPetr Holasek } 61290bd6fd3SPetr Holasek 61321fbd591SQi Zheng static struct ksm_stable_node *alloc_stable_node_chain(struct ksm_stable_node *dup, 6142c653d0eSAndrea Arcangeli struct rb_root *root) 6152c653d0eSAndrea Arcangeli { 61621fbd591SQi Zheng struct ksm_stable_node *chain = alloc_stable_node(); 6172c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_chain(dup)); 6182c653d0eSAndrea Arcangeli if (likely(chain)) { 6192c653d0eSAndrea Arcangeli INIT_HLIST_HEAD(&chain->hlist); 6202c653d0eSAndrea Arcangeli chain->chain_prune_time = jiffies; 6212c653d0eSAndrea Arcangeli chain->rmap_hlist_len = STABLE_NODE_CHAIN; 6222c653d0eSAndrea Arcangeli #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA) 62398fa15f3SAnshuman Khandual chain->nid = NUMA_NO_NODE; /* debug */ 6242c653d0eSAndrea Arcangeli #endif 6252c653d0eSAndrea Arcangeli ksm_stable_node_chains++; 6262c653d0eSAndrea Arcangeli 6272c653d0eSAndrea Arcangeli /* 6282c653d0eSAndrea Arcangeli * Put the stable node chain in the first dimension of 6292c653d0eSAndrea Arcangeli * the stable tree and at the same time remove the old 6302c653d0eSAndrea Arcangeli * stable node. 6312c653d0eSAndrea Arcangeli */ 6322c653d0eSAndrea Arcangeli rb_replace_node(&dup->node, &chain->node, root); 6332c653d0eSAndrea Arcangeli 6342c653d0eSAndrea Arcangeli /* 6352c653d0eSAndrea Arcangeli * Move the old stable node to the second dimension 6362c653d0eSAndrea Arcangeli * queued in the hlist_dup. The invariant is that all 6372c653d0eSAndrea Arcangeli * dup stable_nodes in the chain->hlist point to pages 638457aef94SEthon Paul * that are write protected and have the exact same 6392c653d0eSAndrea Arcangeli * content. 6402c653d0eSAndrea Arcangeli */ 6412c653d0eSAndrea Arcangeli stable_node_chain_add_dup(dup, chain); 6422c653d0eSAndrea Arcangeli } 6432c653d0eSAndrea Arcangeli return chain; 6442c653d0eSAndrea Arcangeli } 6452c653d0eSAndrea Arcangeli 64621fbd591SQi Zheng static inline void free_stable_node_chain(struct ksm_stable_node *chain, 6472c653d0eSAndrea Arcangeli struct rb_root *root) 6482c653d0eSAndrea Arcangeli { 6492c653d0eSAndrea Arcangeli rb_erase(&chain->node, root); 6502c653d0eSAndrea Arcangeli free_stable_node(chain); 6512c653d0eSAndrea Arcangeli ksm_stable_node_chains--; 6522c653d0eSAndrea Arcangeli } 6532c653d0eSAndrea Arcangeli 65421fbd591SQi Zheng static void remove_node_from_stable_tree(struct ksm_stable_node *stable_node) 6554035c07aSHugh Dickins { 65621fbd591SQi Zheng struct ksm_rmap_item *rmap_item; 6574035c07aSHugh Dickins 6582c653d0eSAndrea Arcangeli /* check it's not STABLE_NODE_CHAIN or negative */ 6592c653d0eSAndrea Arcangeli BUG_ON(stable_node->rmap_hlist_len < 0); 6602c653d0eSAndrea Arcangeli 661b67bfe0dSSasha Levin hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { 662739100c8SStefan Roesch if (rmap_item->hlist.next) { 6634035c07aSHugh Dickins ksm_pages_sharing--; 664739100c8SStefan Roesch trace_ksm_remove_rmap_item(stable_node->kpfn, rmap_item, rmap_item->mm); 665739100c8SStefan Roesch } else { 6664035c07aSHugh Dickins ksm_pages_shared--; 667739100c8SStefan Roesch } 66876093853Sxu xin 66976093853Sxu xin rmap_item->mm->ksm_merging_pages--; 67076093853Sxu xin 6712c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len <= 0); 6722c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len--; 6739e60109fSPeter Zijlstra put_anon_vma(rmap_item->anon_vma); 6744035c07aSHugh Dickins rmap_item->address &= PAGE_MASK; 6754035c07aSHugh Dickins cond_resched(); 6764035c07aSHugh Dickins } 6774035c07aSHugh Dickins 6782c653d0eSAndrea Arcangeli /* 6792c653d0eSAndrea Arcangeli * We need the second aligned pointer of the migrate_nodes 6802c653d0eSAndrea Arcangeli * list_head to stay clear from the rb_parent_color union 6812c653d0eSAndrea Arcangeli * (aligned and different than any node) and also different 6822c653d0eSAndrea Arcangeli * from &migrate_nodes. This will verify that future list.h changes 683815f0ddbSNick Desaulniers * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it. 6842c653d0eSAndrea Arcangeli */ 6852c653d0eSAndrea Arcangeli BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes); 6862c653d0eSAndrea Arcangeli BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1); 6872c653d0eSAndrea Arcangeli 688739100c8SStefan Roesch trace_ksm_remove_ksm_page(stable_node->kpfn); 6894146d2d6SHugh Dickins if (stable_node->head == &migrate_nodes) 6904146d2d6SHugh Dickins list_del(&stable_node->list); 6914146d2d6SHugh Dickins else 6922c653d0eSAndrea Arcangeli stable_node_dup_del(stable_node); 6934035c07aSHugh Dickins free_stable_node(stable_node); 6944035c07aSHugh Dickins } 6954035c07aSHugh Dickins 6962cee57d1SYang Shi enum get_ksm_page_flags { 6972cee57d1SYang Shi GET_KSM_PAGE_NOLOCK, 6982cee57d1SYang Shi GET_KSM_PAGE_LOCK, 6992cee57d1SYang Shi GET_KSM_PAGE_TRYLOCK 7002cee57d1SYang Shi }; 7012cee57d1SYang Shi 7024035c07aSHugh Dickins /* 7034035c07aSHugh Dickins * get_ksm_page: checks if the page indicated by the stable node 7044035c07aSHugh Dickins * is still its ksm page, despite having held no reference to it. 7054035c07aSHugh Dickins * In which case we can trust the content of the page, and it 7064035c07aSHugh Dickins * returns the gotten page; but if the page has now been zapped, 7074035c07aSHugh Dickins * remove the stale node from the stable tree and return NULL. 708c8d6553bSHugh Dickins * But beware, the stable node's page might be being migrated. 7094035c07aSHugh Dickins * 7104035c07aSHugh Dickins * You would expect the stable_node to hold a reference to the ksm page. 7114035c07aSHugh Dickins * But if it increments the page's count, swapping out has to wait for 7124035c07aSHugh Dickins * ksmd to come around again before it can free the page, which may take 7134035c07aSHugh Dickins * seconds or even minutes: much too unresponsive. So instead we use a 7144035c07aSHugh Dickins * "keyhole reference": access to the ksm page from the stable node peeps 7154035c07aSHugh Dickins * out through its keyhole to see if that page still holds the right key, 7164035c07aSHugh Dickins * pointing back to this stable node. This relies on freeing a PageAnon 7174035c07aSHugh Dickins * page to reset its page->mapping to NULL, and relies on no other use of 7184035c07aSHugh Dickins * a page to put something that might look like our key in page->mapping. 7194035c07aSHugh Dickins * is on its way to being freed; but it is an anomaly to bear in mind. 7204035c07aSHugh Dickins */ 72121fbd591SQi Zheng static struct page *get_ksm_page(struct ksm_stable_node *stable_node, 7222cee57d1SYang Shi enum get_ksm_page_flags flags) 7234035c07aSHugh Dickins { 7244035c07aSHugh Dickins struct page *page; 7254035c07aSHugh Dickins void *expected_mapping; 726c8d6553bSHugh Dickins unsigned long kpfn; 7274035c07aSHugh Dickins 728bda807d4SMinchan Kim expected_mapping = (void *)((unsigned long)stable_node | 729bda807d4SMinchan Kim PAGE_MAPPING_KSM); 730c8d6553bSHugh Dickins again: 73108df4774SPaul E. McKenney kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */ 732c8d6553bSHugh Dickins page = pfn_to_page(kpfn); 7334db0c3c2SJason Low if (READ_ONCE(page->mapping) != expected_mapping) 7344035c07aSHugh Dickins goto stale; 735c8d6553bSHugh Dickins 736c8d6553bSHugh Dickins /* 737c8d6553bSHugh Dickins * We cannot do anything with the page while its refcount is 0. 738c8d6553bSHugh Dickins * Usually 0 means free, or tail of a higher-order page: in which 739c8d6553bSHugh Dickins * case this node is no longer referenced, and should be freed; 7401c4c3b99SJiang Biao * however, it might mean that the page is under page_ref_freeze(). 741c8d6553bSHugh Dickins * The __remove_mapping() case is easy, again the node is now stale; 74252d1e606SKirill Tkhai * the same is in reuse_ksm_page() case; but if page is swapcache 7439800562fSMatthew Wilcox (Oracle) * in folio_migrate_mapping(), it might still be our page, 74452d1e606SKirill Tkhai * in which case it's essential to keep the node. 745c8d6553bSHugh Dickins */ 746c8d6553bSHugh Dickins while (!get_page_unless_zero(page)) { 747c8d6553bSHugh Dickins /* 748c8d6553bSHugh Dickins * Another check for page->mapping != expected_mapping would 749c8d6553bSHugh Dickins * work here too. We have chosen the !PageSwapCache test to 750c8d6553bSHugh Dickins * optimize the common case, when the page is or is about to 751c8d6553bSHugh Dickins * be freed: PageSwapCache is cleared (under spin_lock_irq) 7521c4c3b99SJiang Biao * in the ref_freeze section of __remove_mapping(); but Anon 753c8d6553bSHugh Dickins * page->mapping reset to NULL later, in free_pages_prepare(). 754c8d6553bSHugh Dickins */ 755c8d6553bSHugh Dickins if (!PageSwapCache(page)) 7564035c07aSHugh Dickins goto stale; 757c8d6553bSHugh Dickins cpu_relax(); 758c8d6553bSHugh Dickins } 759c8d6553bSHugh Dickins 7604db0c3c2SJason Low if (READ_ONCE(page->mapping) != expected_mapping) { 7614035c07aSHugh Dickins put_page(page); 7624035c07aSHugh Dickins goto stale; 7634035c07aSHugh Dickins } 764c8d6553bSHugh Dickins 7652cee57d1SYang Shi if (flags == GET_KSM_PAGE_TRYLOCK) { 7662cee57d1SYang Shi if (!trylock_page(page)) { 7672cee57d1SYang Shi put_page(page); 7682cee57d1SYang Shi return ERR_PTR(-EBUSY); 7692cee57d1SYang Shi } 7702cee57d1SYang Shi } else if (flags == GET_KSM_PAGE_LOCK) 7718aafa6a4SHugh Dickins lock_page(page); 7722cee57d1SYang Shi 7732cee57d1SYang Shi if (flags != GET_KSM_PAGE_NOLOCK) { 7744db0c3c2SJason Low if (READ_ONCE(page->mapping) != expected_mapping) { 7758aafa6a4SHugh Dickins unlock_page(page); 7768aafa6a4SHugh Dickins put_page(page); 7778aafa6a4SHugh Dickins goto stale; 7788aafa6a4SHugh Dickins } 7798aafa6a4SHugh Dickins } 7804035c07aSHugh Dickins return page; 781c8d6553bSHugh Dickins 7824035c07aSHugh Dickins stale: 783c8d6553bSHugh Dickins /* 784c8d6553bSHugh Dickins * We come here from above when page->mapping or !PageSwapCache 785c8d6553bSHugh Dickins * suggests that the node is stale; but it might be under migration. 78619138349SMatthew Wilcox (Oracle) * We need smp_rmb(), matching the smp_wmb() in folio_migrate_ksm(), 787c8d6553bSHugh Dickins * before checking whether node->kpfn has been changed. 788c8d6553bSHugh Dickins */ 789c8d6553bSHugh Dickins smp_rmb(); 7904db0c3c2SJason Low if (READ_ONCE(stable_node->kpfn) != kpfn) 791c8d6553bSHugh Dickins goto again; 7924035c07aSHugh Dickins remove_node_from_stable_tree(stable_node); 7934035c07aSHugh Dickins return NULL; 7944035c07aSHugh Dickins } 7954035c07aSHugh Dickins 79631dbd01fSIzik Eidus /* 79731dbd01fSIzik Eidus * Removing rmap_item from stable or unstable tree. 79831dbd01fSIzik Eidus * This function will clean the information from the stable/unstable tree. 79931dbd01fSIzik Eidus */ 80021fbd591SQi Zheng static void remove_rmap_item_from_tree(struct ksm_rmap_item *rmap_item) 80131dbd01fSIzik Eidus { 8027b6ba2c7SHugh Dickins if (rmap_item->address & STABLE_FLAG) { 80321fbd591SQi Zheng struct ksm_stable_node *stable_node; 8045ad64688SHugh Dickins struct page *page; 80531dbd01fSIzik Eidus 8067b6ba2c7SHugh Dickins stable_node = rmap_item->head; 80762862290SHugh Dickins page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); 8084035c07aSHugh Dickins if (!page) 8094035c07aSHugh Dickins goto out; 8105ad64688SHugh Dickins 8117b6ba2c7SHugh Dickins hlist_del(&rmap_item->hlist); 81262862290SHugh Dickins unlock_page(page); 8135ad64688SHugh Dickins put_page(page); 81408beca44SHugh Dickins 81598666f8aSAndrea Arcangeli if (!hlist_empty(&stable_node->hlist)) 8164035c07aSHugh Dickins ksm_pages_sharing--; 8174035c07aSHugh Dickins else 818b4028260SHugh Dickins ksm_pages_shared--; 81976093853Sxu xin 82076093853Sxu xin rmap_item->mm->ksm_merging_pages--; 82176093853Sxu xin 8222c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len <= 0); 8232c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len--; 82431dbd01fSIzik Eidus 8259e60109fSPeter Zijlstra put_anon_vma(rmap_item->anon_vma); 826c89a384eSMiaohe Lin rmap_item->head = NULL; 82793d17715SHugh Dickins rmap_item->address &= PAGE_MASK; 82831dbd01fSIzik Eidus 8297b6ba2c7SHugh Dickins } else if (rmap_item->address & UNSTABLE_FLAG) { 83031dbd01fSIzik Eidus unsigned char age; 83131dbd01fSIzik Eidus /* 8329ba69294SHugh Dickins * Usually ksmd can and must skip the rb_erase, because 83331dbd01fSIzik Eidus * root_unstable_tree was already reset to RB_ROOT. 8349ba69294SHugh Dickins * But be careful when an mm is exiting: do the rb_erase 8359ba69294SHugh Dickins * if this rmap_item was inserted by this scan, rather 8369ba69294SHugh Dickins * than left over from before. 83731dbd01fSIzik Eidus */ 83831dbd01fSIzik Eidus age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); 839cd551f97SHugh Dickins BUG_ON(age > 1); 84031dbd01fSIzik Eidus if (!age) 84190bd6fd3SPetr Holasek rb_erase(&rmap_item->node, 842ef53d16cSHugh Dickins root_unstable_tree + NUMA(rmap_item->nid)); 84393d17715SHugh Dickins ksm_pages_unshared--; 84431dbd01fSIzik Eidus rmap_item->address &= PAGE_MASK; 84593d17715SHugh Dickins } 8464035c07aSHugh Dickins out: 84731dbd01fSIzik Eidus cond_resched(); /* we're called from many long loops */ 84831dbd01fSIzik Eidus } 84931dbd01fSIzik Eidus 85021fbd591SQi Zheng static void remove_trailing_rmap_items(struct ksm_rmap_item **rmap_list) 85131dbd01fSIzik Eidus { 8526514d511SHugh Dickins while (*rmap_list) { 85321fbd591SQi Zheng struct ksm_rmap_item *rmap_item = *rmap_list; 8546514d511SHugh Dickins *rmap_list = rmap_item->rmap_list; 85531dbd01fSIzik Eidus remove_rmap_item_from_tree(rmap_item); 85631dbd01fSIzik Eidus free_rmap_item(rmap_item); 85731dbd01fSIzik Eidus } 85831dbd01fSIzik Eidus } 85931dbd01fSIzik Eidus 86031dbd01fSIzik Eidus /* 861e850dcf5SHugh Dickins * Though it's very tempting to unmerge rmap_items from stable tree rather 86231dbd01fSIzik Eidus * than check every pte of a given vma, the locking doesn't quite work for 86331dbd01fSIzik Eidus * that - an rmap_item is assigned to the stable tree after inserting ksm 864c1e8d7c6SMichel Lespinasse * page and upping mmap_lock. Nor does it fit with the way we skip dup'ing 86531dbd01fSIzik Eidus * rmap_items from parent to child at fork time (so as not to waste time 86631dbd01fSIzik Eidus * if exit comes before the next scan reaches it). 86781464e30SHugh Dickins * 86881464e30SHugh Dickins * Similarly, although we'd like to remove rmap_items (so updating counts 86981464e30SHugh Dickins * and freeing memory) when unmerging an area, it's easier to leave that 87081464e30SHugh Dickins * to the next pass of ksmd - consider, for example, how ksmd might be 87181464e30SHugh Dickins * in cmp_and_merge_page on one of the rmap_items we would be removing. 87231dbd01fSIzik Eidus */ 873d952b791SHugh Dickins static int unmerge_ksm_pages(struct vm_area_struct *vma, 87431dbd01fSIzik Eidus unsigned long start, unsigned long end) 87531dbd01fSIzik Eidus { 87631dbd01fSIzik Eidus unsigned long addr; 877d952b791SHugh Dickins int err = 0; 87831dbd01fSIzik Eidus 879d952b791SHugh Dickins for (addr = start; addr < end && !err; addr += PAGE_SIZE) { 8809ba69294SHugh Dickins if (ksm_test_exit(vma->vm_mm)) 8819ba69294SHugh Dickins break; 882d952b791SHugh Dickins if (signal_pending(current)) 883d952b791SHugh Dickins err = -ERESTARTSYS; 884d952b791SHugh Dickins else 885d952b791SHugh Dickins err = break_ksm(vma, addr); 886d952b791SHugh Dickins } 887d952b791SHugh Dickins return err; 88831dbd01fSIzik Eidus } 88931dbd01fSIzik Eidus 89021fbd591SQi Zheng static inline struct ksm_stable_node *folio_stable_node(struct folio *folio) 89119138349SMatthew Wilcox (Oracle) { 89219138349SMatthew Wilcox (Oracle) return folio_test_ksm(folio) ? folio_raw_mapping(folio) : NULL; 89319138349SMatthew Wilcox (Oracle) } 89419138349SMatthew Wilcox (Oracle) 89521fbd591SQi Zheng static inline struct ksm_stable_node *page_stable_node(struct page *page) 89688484826SMike Rapoport { 89719138349SMatthew Wilcox (Oracle) return folio_stable_node(page_folio(page)); 89888484826SMike Rapoport } 89988484826SMike Rapoport 90088484826SMike Rapoport static inline void set_page_stable_node(struct page *page, 90121fbd591SQi Zheng struct ksm_stable_node *stable_node) 90288484826SMike Rapoport { 9036c287605SDavid Hildenbrand VM_BUG_ON_PAGE(PageAnon(page) && PageAnonExclusive(page), page); 90488484826SMike Rapoport page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM); 90588484826SMike Rapoport } 90688484826SMike Rapoport 9072ffd8679SHugh Dickins #ifdef CONFIG_SYSFS 9082ffd8679SHugh Dickins /* 9092ffd8679SHugh Dickins * Only called through the sysfs control interface: 9102ffd8679SHugh Dickins */ 91121fbd591SQi Zheng static int remove_stable_node(struct ksm_stable_node *stable_node) 912cbf86cfeSHugh Dickins { 913cbf86cfeSHugh Dickins struct page *page; 914cbf86cfeSHugh Dickins int err; 915cbf86cfeSHugh Dickins 9162cee57d1SYang Shi page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); 917cbf86cfeSHugh Dickins if (!page) { 918cbf86cfeSHugh Dickins /* 919cbf86cfeSHugh Dickins * get_ksm_page did remove_node_from_stable_tree itself. 920cbf86cfeSHugh Dickins */ 921cbf86cfeSHugh Dickins return 0; 922cbf86cfeSHugh Dickins } 923cbf86cfeSHugh Dickins 924cbf86cfeSHugh Dickins /* 9259a63236fSAndrey Ryabinin * Page could be still mapped if this races with __mmput() running in 9269a63236fSAndrey Ryabinin * between ksm_exit() and exit_mmap(). Just refuse to let 9279a63236fSAndrey Ryabinin * merge_across_nodes/max_page_sharing be switched. 9288fdb3dbfSHugh Dickins */ 9298fdb3dbfSHugh Dickins err = -EBUSY; 9309a63236fSAndrey Ryabinin if (!page_mapped(page)) { 9318fdb3dbfSHugh Dickins /* 9328fdb3dbfSHugh Dickins * The stable node did not yet appear stale to get_ksm_page(), 9338fdb3dbfSHugh Dickins * since that allows for an unmapped ksm page to be recognized 9348fdb3dbfSHugh Dickins * right up until it is freed; but the node is safe to remove. 935cbf86cfeSHugh Dickins * This page might be in a pagevec waiting to be freed, 936cbf86cfeSHugh Dickins * or it might be PageSwapCache (perhaps under writeback), 937cbf86cfeSHugh Dickins * or it might have been removed from swapcache a moment ago. 938cbf86cfeSHugh Dickins */ 939cbf86cfeSHugh Dickins set_page_stable_node(page, NULL); 940cbf86cfeSHugh Dickins remove_node_from_stable_tree(stable_node); 941cbf86cfeSHugh Dickins err = 0; 942cbf86cfeSHugh Dickins } 943cbf86cfeSHugh Dickins 944cbf86cfeSHugh Dickins unlock_page(page); 945cbf86cfeSHugh Dickins put_page(page); 946cbf86cfeSHugh Dickins return err; 947cbf86cfeSHugh Dickins } 948cbf86cfeSHugh Dickins 94921fbd591SQi Zheng static int remove_stable_node_chain(struct ksm_stable_node *stable_node, 9502c653d0eSAndrea Arcangeli struct rb_root *root) 9512c653d0eSAndrea Arcangeli { 95221fbd591SQi Zheng struct ksm_stable_node *dup; 9532c653d0eSAndrea Arcangeli struct hlist_node *hlist_safe; 9542c653d0eSAndrea Arcangeli 9552c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 9562c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node)); 9572c653d0eSAndrea Arcangeli if (remove_stable_node(stable_node)) 9582c653d0eSAndrea Arcangeli return true; 9592c653d0eSAndrea Arcangeli else 9602c653d0eSAndrea Arcangeli return false; 9612c653d0eSAndrea Arcangeli } 9622c653d0eSAndrea Arcangeli 9632c653d0eSAndrea Arcangeli hlist_for_each_entry_safe(dup, hlist_safe, 9642c653d0eSAndrea Arcangeli &stable_node->hlist, hlist_dup) { 9652c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(dup)); 9662c653d0eSAndrea Arcangeli if (remove_stable_node(dup)) 9672c653d0eSAndrea Arcangeli return true; 9682c653d0eSAndrea Arcangeli } 9692c653d0eSAndrea Arcangeli BUG_ON(!hlist_empty(&stable_node->hlist)); 9702c653d0eSAndrea Arcangeli free_stable_node_chain(stable_node, root); 9712c653d0eSAndrea Arcangeli return false; 9722c653d0eSAndrea Arcangeli } 9732c653d0eSAndrea Arcangeli 974cbf86cfeSHugh Dickins static int remove_all_stable_nodes(void) 975cbf86cfeSHugh Dickins { 97621fbd591SQi Zheng struct ksm_stable_node *stable_node, *next; 977cbf86cfeSHugh Dickins int nid; 978cbf86cfeSHugh Dickins int err = 0; 979cbf86cfeSHugh Dickins 980ef53d16cSHugh Dickins for (nid = 0; nid < ksm_nr_node_ids; nid++) { 981cbf86cfeSHugh Dickins while (root_stable_tree[nid].rb_node) { 982cbf86cfeSHugh Dickins stable_node = rb_entry(root_stable_tree[nid].rb_node, 98321fbd591SQi Zheng struct ksm_stable_node, node); 9842c653d0eSAndrea Arcangeli if (remove_stable_node_chain(stable_node, 9852c653d0eSAndrea Arcangeli root_stable_tree + nid)) { 986cbf86cfeSHugh Dickins err = -EBUSY; 987cbf86cfeSHugh Dickins break; /* proceed to next nid */ 988cbf86cfeSHugh Dickins } 989cbf86cfeSHugh Dickins cond_resched(); 990cbf86cfeSHugh Dickins } 991cbf86cfeSHugh Dickins } 99203640418SGeliang Tang list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { 9934146d2d6SHugh Dickins if (remove_stable_node(stable_node)) 9944146d2d6SHugh Dickins err = -EBUSY; 9954146d2d6SHugh Dickins cond_resched(); 9964146d2d6SHugh Dickins } 997cbf86cfeSHugh Dickins return err; 998cbf86cfeSHugh Dickins } 999cbf86cfeSHugh Dickins 1000d952b791SHugh Dickins static int unmerge_and_remove_all_rmap_items(void) 100131dbd01fSIzik Eidus { 100221fbd591SQi Zheng struct ksm_mm_slot *mm_slot; 100358730ab6SQi Zheng struct mm_slot *slot; 100431dbd01fSIzik Eidus struct mm_struct *mm; 100531dbd01fSIzik Eidus struct vm_area_struct *vma; 1006d952b791SHugh Dickins int err = 0; 100731dbd01fSIzik Eidus 1008d952b791SHugh Dickins spin_lock(&ksm_mmlist_lock); 100958730ab6SQi Zheng slot = list_entry(ksm_mm_head.slot.mm_node.next, 101058730ab6SQi Zheng struct mm_slot, mm_node); 101158730ab6SQi Zheng ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); 1012d952b791SHugh Dickins spin_unlock(&ksm_mmlist_lock); 1013d952b791SHugh Dickins 1014a5f18ba0SMatthew Wilcox (Oracle) for (mm_slot = ksm_scan.mm_slot; mm_slot != &ksm_mm_head; 1015a5f18ba0SMatthew Wilcox (Oracle) mm_slot = ksm_scan.mm_slot) { 101658730ab6SQi Zheng VMA_ITERATOR(vmi, mm_slot->slot.mm, 0); 1017a5f18ba0SMatthew Wilcox (Oracle) 101858730ab6SQi Zheng mm = mm_slot->slot.mm; 1019d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 10206db504ceSLiam R. Howlett 10216db504ceSLiam R. Howlett /* 10226db504ceSLiam R. Howlett * Exit right away if mm is exiting to avoid lockdep issue in 10236db504ceSLiam R. Howlett * the maple tree 10246db504ceSLiam R. Howlett */ 10259ba69294SHugh Dickins if (ksm_test_exit(mm)) 10266db504ceSLiam R. Howlett goto mm_exiting; 10276db504ceSLiam R. Howlett 10286db504ceSLiam R. Howlett for_each_vma(vmi, vma) { 102931dbd01fSIzik Eidus if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) 103031dbd01fSIzik Eidus continue; 1031d952b791SHugh Dickins err = unmerge_ksm_pages(vma, 1032d952b791SHugh Dickins vma->vm_start, vma->vm_end); 10339ba69294SHugh Dickins if (err) 10349ba69294SHugh Dickins goto error; 1035d952b791SHugh Dickins } 10369ba69294SHugh Dickins 10376db504ceSLiam R. Howlett mm_exiting: 1038420be4edSChengyang Fan remove_trailing_rmap_items(&mm_slot->rmap_list); 1039d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 104031dbd01fSIzik Eidus 104131dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 104258730ab6SQi Zheng slot = list_entry(mm_slot->slot.mm_node.next, 104358730ab6SQi Zheng struct mm_slot, mm_node); 104458730ab6SQi Zheng ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); 10459ba69294SHugh Dickins if (ksm_test_exit(mm)) { 104658730ab6SQi Zheng hash_del(&mm_slot->slot.hash); 104758730ab6SQi Zheng list_del(&mm_slot->slot.mm_node); 104831dbd01fSIzik Eidus spin_unlock(&ksm_mmlist_lock); 10499ba69294SHugh Dickins 105058730ab6SQi Zheng mm_slot_free(mm_slot_cache, mm_slot); 10519ba69294SHugh Dickins clear_bit(MMF_VM_MERGEABLE, &mm->flags); 1052d7597f59SStefan Roesch clear_bit(MMF_VM_MERGE_ANY, &mm->flags); 10539ba69294SHugh Dickins mmdrop(mm); 10547496fea9SZhou Chengming } else 10559ba69294SHugh Dickins spin_unlock(&ksm_mmlist_lock); 105631dbd01fSIzik Eidus } 105731dbd01fSIzik Eidus 1058cbf86cfeSHugh Dickins /* Clean up stable nodes, but don't worry if some are still busy */ 1059cbf86cfeSHugh Dickins remove_all_stable_nodes(); 1060d952b791SHugh Dickins ksm_scan.seqnr = 0; 10619ba69294SHugh Dickins return 0; 10629ba69294SHugh Dickins 10639ba69294SHugh Dickins error: 1064d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1065d952b791SHugh Dickins spin_lock(&ksm_mmlist_lock); 1066d952b791SHugh Dickins ksm_scan.mm_slot = &ksm_mm_head; 1067d952b791SHugh Dickins spin_unlock(&ksm_mmlist_lock); 1068d952b791SHugh Dickins return err; 1069d952b791SHugh Dickins } 10702ffd8679SHugh Dickins #endif /* CONFIG_SYSFS */ 1071d952b791SHugh Dickins 107231dbd01fSIzik Eidus static u32 calc_checksum(struct page *page) 107331dbd01fSIzik Eidus { 107431dbd01fSIzik Eidus u32 checksum; 10759b04c5feSCong Wang void *addr = kmap_atomic(page); 107659e1a2f4STimofey Titovets checksum = xxhash(addr, PAGE_SIZE, 0); 10779b04c5feSCong Wang kunmap_atomic(addr); 107831dbd01fSIzik Eidus return checksum; 107931dbd01fSIzik Eidus } 108031dbd01fSIzik Eidus 108131dbd01fSIzik Eidus static int write_protect_page(struct vm_area_struct *vma, struct page *page, 108231dbd01fSIzik Eidus pte_t *orig_pte) 108331dbd01fSIzik Eidus { 108431dbd01fSIzik Eidus struct mm_struct *mm = vma->vm_mm; 1085eed05e54SMatthew Wilcox (Oracle) DEFINE_PAGE_VMA_WALK(pvmw, page, vma, 0, 0); 108631dbd01fSIzik Eidus int swapped; 108731dbd01fSIzik Eidus int err = -EFAULT; 1088ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 10896c287605SDavid Hildenbrand bool anon_exclusive; 1090*c33c7948SRyan Roberts pte_t entry; 109131dbd01fSIzik Eidus 109236eaff33SKirill A. Shutemov pvmw.address = page_address_in_vma(page, vma); 109336eaff33SKirill A. Shutemov if (pvmw.address == -EFAULT) 109431dbd01fSIzik Eidus goto out; 109531dbd01fSIzik Eidus 109629ad768cSAndrea Arcangeli BUG_ON(PageTransCompound(page)); 10976bdb913fSHaggai Eran 10987d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, pvmw.address, 1099ac46d4f3SJérôme Glisse pvmw.address + PAGE_SIZE); 1100ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 11016bdb913fSHaggai Eran 110236eaff33SKirill A. Shutemov if (!page_vma_mapped_walk(&pvmw)) 11036bdb913fSHaggai Eran goto out_mn; 110436eaff33SKirill A. Shutemov if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?")) 110536eaff33SKirill A. Shutemov goto out_unlock; 110631dbd01fSIzik Eidus 11076c287605SDavid Hildenbrand anon_exclusive = PageAnonExclusive(page); 1108*c33c7948SRyan Roberts entry = ptep_get(pvmw.pte); 1109*c33c7948SRyan Roberts if (pte_write(entry) || pte_dirty(entry) || 11106c287605SDavid Hildenbrand anon_exclusive || mm_tlb_flush_pending(mm)) { 111131dbd01fSIzik Eidus swapped = PageSwapCache(page); 111236eaff33SKirill A. Shutemov flush_cache_page(vma, pvmw.address, page_to_pfn(page)); 111331dbd01fSIzik Eidus /* 111425985edcSLucas De Marchi * Ok this is tricky, when get_user_pages_fast() run it doesn't 111531dbd01fSIzik Eidus * take any lock, therefore the check that we are going to make 1116f0953a1bSIngo Molnar * with the pagecount against the mapcount is racy and 111731dbd01fSIzik Eidus * O_DIRECT can happen right after the check. 111831dbd01fSIzik Eidus * So we clear the pte and flush the tlb before the check 111931dbd01fSIzik Eidus * this assure us that no O_DIRECT can happen after the check 112031dbd01fSIzik Eidus * or in the middle of the check. 11210f10851eSJérôme Glisse * 11220f10851eSJérôme Glisse * No need to notify as we are downgrading page table to read 11230f10851eSJérôme Glisse * only not changing it to point to a new page. 11240f10851eSJérôme Glisse * 1125ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 112631dbd01fSIzik Eidus */ 11270f10851eSJérôme Glisse entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte); 112831dbd01fSIzik Eidus /* 112931dbd01fSIzik Eidus * Check that no O_DIRECT or similar I/O is in progress on the 113031dbd01fSIzik Eidus * page 113131dbd01fSIzik Eidus */ 113231e855eaSHugh Dickins if (page_mapcount(page) + 1 + swapped != page_count(page)) { 113336eaff33SKirill A. Shutemov set_pte_at(mm, pvmw.address, pvmw.pte, entry); 113431dbd01fSIzik Eidus goto out_unlock; 113531dbd01fSIzik Eidus } 11366c287605SDavid Hildenbrand 1137088b8aa5SDavid Hildenbrand /* See page_try_share_anon_rmap(): clear PTE first. */ 11386c287605SDavid Hildenbrand if (anon_exclusive && page_try_share_anon_rmap(page)) { 11396c287605SDavid Hildenbrand set_pte_at(mm, pvmw.address, pvmw.pte, entry); 11406c287605SDavid Hildenbrand goto out_unlock; 11416c287605SDavid Hildenbrand } 11426c287605SDavid Hildenbrand 11434e31635cSHugh Dickins if (pte_dirty(entry)) 11444e31635cSHugh Dickins set_page_dirty(page); 11456a56ccbcSDavid Hildenbrand entry = pte_mkclean(entry); 1146595cd8f2SAneesh Kumar K.V 11476a56ccbcSDavid Hildenbrand if (pte_write(entry)) 11486a56ccbcSDavid Hildenbrand entry = pte_wrprotect(entry); 11496a56ccbcSDavid Hildenbrand 115036eaff33SKirill A. Shutemov set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry); 115131dbd01fSIzik Eidus } 1152*c33c7948SRyan Roberts *orig_pte = entry; 115331dbd01fSIzik Eidus err = 0; 115431dbd01fSIzik Eidus 115531dbd01fSIzik Eidus out_unlock: 115636eaff33SKirill A. Shutemov page_vma_mapped_walk_done(&pvmw); 11576bdb913fSHaggai Eran out_mn: 1158ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 115931dbd01fSIzik Eidus out: 116031dbd01fSIzik Eidus return err; 116131dbd01fSIzik Eidus } 116231dbd01fSIzik Eidus 116331dbd01fSIzik Eidus /** 116431dbd01fSIzik Eidus * replace_page - replace page in vma by new ksm page 11658dd3557aSHugh Dickins * @vma: vma that holds the pte pointing to page 11668dd3557aSHugh Dickins * @page: the page we are replacing by kpage 11678dd3557aSHugh Dickins * @kpage: the ksm page we replace page by 116831dbd01fSIzik Eidus * @orig_pte: the original value of the pte 116931dbd01fSIzik Eidus * 117031dbd01fSIzik Eidus * Returns 0 on success, -EFAULT on failure. 117131dbd01fSIzik Eidus */ 11728dd3557aSHugh Dickins static int replace_page(struct vm_area_struct *vma, struct page *page, 11738dd3557aSHugh Dickins struct page *kpage, pte_t orig_pte) 117431dbd01fSIzik Eidus { 117531dbd01fSIzik Eidus struct mm_struct *mm = vma->vm_mm; 1176b4e6f66eSMatthew Wilcox (Oracle) struct folio *folio; 117731dbd01fSIzik Eidus pmd_t *pmd; 117850722804SZach O'Keefe pmd_t pmde; 117931dbd01fSIzik Eidus pte_t *ptep; 1180e86c59b1SClaudio Imbrenda pte_t newpte; 118131dbd01fSIzik Eidus spinlock_t *ptl; 118231dbd01fSIzik Eidus unsigned long addr; 118331dbd01fSIzik Eidus int err = -EFAULT; 1184ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 118531dbd01fSIzik Eidus 11868dd3557aSHugh Dickins addr = page_address_in_vma(page, vma); 118731dbd01fSIzik Eidus if (addr == -EFAULT) 118831dbd01fSIzik Eidus goto out; 118931dbd01fSIzik Eidus 11906219049aSBob Liu pmd = mm_find_pmd(mm, addr); 11916219049aSBob Liu if (!pmd) 119231dbd01fSIzik Eidus goto out; 119350722804SZach O'Keefe /* 119450722804SZach O'Keefe * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at() 119550722804SZach O'Keefe * without holding anon_vma lock for write. So when looking for a 119650722804SZach O'Keefe * genuine pmde (in which to find pte), test present and !THP together. 119750722804SZach O'Keefe */ 119826e1a0c3SHugh Dickins pmde = pmdp_get_lockless(pmd); 119950722804SZach O'Keefe if (!pmd_present(pmde) || pmd_trans_huge(pmde)) 120050722804SZach O'Keefe goto out; 120131dbd01fSIzik Eidus 12027d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, addr, 12036f4f13e8SJérôme Glisse addr + PAGE_SIZE); 1204ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 12056bdb913fSHaggai Eran 120631dbd01fSIzik Eidus ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); 120704dee9e8SHugh Dickins if (!ptep) 120804dee9e8SHugh Dickins goto out_mn; 1209*c33c7948SRyan Roberts if (!pte_same(ptep_get(ptep), orig_pte)) { 121031dbd01fSIzik Eidus pte_unmap_unlock(ptep, ptl); 12116bdb913fSHaggai Eran goto out_mn; 121231dbd01fSIzik Eidus } 12136c287605SDavid Hildenbrand VM_BUG_ON_PAGE(PageAnonExclusive(page), page); 12146c287605SDavid Hildenbrand VM_BUG_ON_PAGE(PageAnon(kpage) && PageAnonExclusive(kpage), kpage); 121531dbd01fSIzik Eidus 1216e86c59b1SClaudio Imbrenda /* 1217e86c59b1SClaudio Imbrenda * No need to check ksm_use_zero_pages here: we can only have a 1218457aef94SEthon Paul * zero_page here if ksm_use_zero_pages was enabled already. 1219e86c59b1SClaudio Imbrenda */ 1220e86c59b1SClaudio Imbrenda if (!is_zero_pfn(page_to_pfn(kpage))) { 12218dd3557aSHugh Dickins get_page(kpage); 1222f1e2db12SDavid Hildenbrand page_add_anon_rmap(kpage, vma, addr, RMAP_NONE); 1223e86c59b1SClaudio Imbrenda newpte = mk_pte(kpage, vma->vm_page_prot); 1224e86c59b1SClaudio Imbrenda } else { 1225e86c59b1SClaudio Imbrenda newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage), 1226e86c59b1SClaudio Imbrenda vma->vm_page_prot)); 1227a38c015fSClaudio Imbrenda /* 1228a38c015fSClaudio Imbrenda * We're replacing an anonymous page with a zero page, which is 1229a38c015fSClaudio Imbrenda * not anonymous. We need to do proper accounting otherwise we 1230a38c015fSClaudio Imbrenda * will get wrong values in /proc, and a BUG message in dmesg 1231a38c015fSClaudio Imbrenda * when tearing down the mm. 1232a38c015fSClaudio Imbrenda */ 1233a38c015fSClaudio Imbrenda dec_mm_counter(mm, MM_ANONPAGES); 1234e86c59b1SClaudio Imbrenda } 123531dbd01fSIzik Eidus 1236*c33c7948SRyan Roberts flush_cache_page(vma, addr, pte_pfn(ptep_get(ptep))); 12370f10851eSJérôme Glisse /* 12380f10851eSJérôme Glisse * No need to notify as we are replacing a read only page with another 12390f10851eSJérôme Glisse * read only page with the same content. 12400f10851eSJérôme Glisse * 1241ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 12420f10851eSJérôme Glisse */ 12430f10851eSJérôme Glisse ptep_clear_flush(vma, addr, ptep); 1244e86c59b1SClaudio Imbrenda set_pte_at_notify(mm, addr, ptep, newpte); 124531dbd01fSIzik Eidus 1246b4e6f66eSMatthew Wilcox (Oracle) folio = page_folio(page); 1247cea86fe2SHugh Dickins page_remove_rmap(page, vma, false); 1248b4e6f66eSMatthew Wilcox (Oracle) if (!folio_mapped(folio)) 1249b4e6f66eSMatthew Wilcox (Oracle) folio_free_swap(folio); 1250b4e6f66eSMatthew Wilcox (Oracle) folio_put(folio); 125131dbd01fSIzik Eidus 125231dbd01fSIzik Eidus pte_unmap_unlock(ptep, ptl); 125331dbd01fSIzik Eidus err = 0; 12546bdb913fSHaggai Eran out_mn: 1255ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 125631dbd01fSIzik Eidus out: 125731dbd01fSIzik Eidus return err; 125831dbd01fSIzik Eidus } 125931dbd01fSIzik Eidus 126031dbd01fSIzik Eidus /* 126131dbd01fSIzik Eidus * try_to_merge_one_page - take two pages and merge them into one 12628dd3557aSHugh Dickins * @vma: the vma that holds the pte pointing to page 12638dd3557aSHugh Dickins * @page: the PageAnon page that we want to replace with kpage 126480e14822SHugh Dickins * @kpage: the PageKsm page that we want to map instead of page, 126580e14822SHugh Dickins * or NULL the first time when we want to use page as kpage. 126631dbd01fSIzik Eidus * 126731dbd01fSIzik Eidus * This function returns 0 if the pages were merged, -EFAULT otherwise. 126831dbd01fSIzik Eidus */ 126931dbd01fSIzik Eidus static int try_to_merge_one_page(struct vm_area_struct *vma, 12708dd3557aSHugh Dickins struct page *page, struct page *kpage) 127131dbd01fSIzik Eidus { 127231dbd01fSIzik Eidus pte_t orig_pte = __pte(0); 127331dbd01fSIzik Eidus int err = -EFAULT; 127431dbd01fSIzik Eidus 1275db114b83SHugh Dickins if (page == kpage) /* ksm page forked */ 1276db114b83SHugh Dickins return 0; 1277db114b83SHugh Dickins 12788dd3557aSHugh Dickins if (!PageAnon(page)) 127931dbd01fSIzik Eidus goto out; 128031dbd01fSIzik Eidus 128131dbd01fSIzik Eidus /* 128231dbd01fSIzik Eidus * We need the page lock to read a stable PageSwapCache in 128331dbd01fSIzik Eidus * write_protect_page(). We use trylock_page() instead of 128431dbd01fSIzik Eidus * lock_page() because we don't want to wait here - we 128531dbd01fSIzik Eidus * prefer to continue scanning and merging different pages, 128631dbd01fSIzik Eidus * then come back to this page when it is unlocked. 128731dbd01fSIzik Eidus */ 12888dd3557aSHugh Dickins if (!trylock_page(page)) 128931e855eaSHugh Dickins goto out; 1290f765f540SKirill A. Shutemov 1291f765f540SKirill A. Shutemov if (PageTransCompound(page)) { 1292a7306c34SAndrea Arcangeli if (split_huge_page(page)) 1293f765f540SKirill A. Shutemov goto out_unlock; 1294f765f540SKirill A. Shutemov } 1295f765f540SKirill A. Shutemov 129631dbd01fSIzik Eidus /* 129731dbd01fSIzik Eidus * If this anonymous page is mapped only here, its pte may need 129831dbd01fSIzik Eidus * to be write-protected. If it's mapped elsewhere, all of its 129931dbd01fSIzik Eidus * ptes are necessarily already write-protected. But in either 130031dbd01fSIzik Eidus * case, we need to lock and check page_count is not raised. 130131dbd01fSIzik Eidus */ 130280e14822SHugh Dickins if (write_protect_page(vma, page, &orig_pte) == 0) { 130380e14822SHugh Dickins if (!kpage) { 130480e14822SHugh Dickins /* 130580e14822SHugh Dickins * While we hold page lock, upgrade page from 130680e14822SHugh Dickins * PageAnon+anon_vma to PageKsm+NULL stable_node: 130780e14822SHugh Dickins * stable_tree_insert() will update stable_node. 130880e14822SHugh Dickins */ 130980e14822SHugh Dickins set_page_stable_node(page, NULL); 131080e14822SHugh Dickins mark_page_accessed(page); 1311337ed7ebSMinchan Kim /* 1312337ed7ebSMinchan Kim * Page reclaim just frees a clean page with no dirty 1313337ed7ebSMinchan Kim * ptes: make sure that the ksm page would be swapped. 1314337ed7ebSMinchan Kim */ 1315337ed7ebSMinchan Kim if (!PageDirty(page)) 1316337ed7ebSMinchan Kim SetPageDirty(page); 131780e14822SHugh Dickins err = 0; 131880e14822SHugh Dickins } else if (pages_identical(page, kpage)) 13198dd3557aSHugh Dickins err = replace_page(vma, page, kpage, orig_pte); 132080e14822SHugh Dickins } 132131dbd01fSIzik Eidus 1322f765f540SKirill A. Shutemov out_unlock: 13238dd3557aSHugh Dickins unlock_page(page); 132431dbd01fSIzik Eidus out: 132531dbd01fSIzik Eidus return err; 132631dbd01fSIzik Eidus } 132731dbd01fSIzik Eidus 132831dbd01fSIzik Eidus /* 132981464e30SHugh Dickins * try_to_merge_with_ksm_page - like try_to_merge_two_pages, 133081464e30SHugh Dickins * but no new kernel page is allocated: kpage must already be a ksm page. 13318dd3557aSHugh Dickins * 13328dd3557aSHugh Dickins * This function returns 0 if the pages were merged, -EFAULT otherwise. 133381464e30SHugh Dickins */ 133421fbd591SQi Zheng static int try_to_merge_with_ksm_page(struct ksm_rmap_item *rmap_item, 13358dd3557aSHugh Dickins struct page *page, struct page *kpage) 133681464e30SHugh Dickins { 13378dd3557aSHugh Dickins struct mm_struct *mm = rmap_item->mm; 133881464e30SHugh Dickins struct vm_area_struct *vma; 133981464e30SHugh Dickins int err = -EFAULT; 134081464e30SHugh Dickins 1341d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 134285c6e8ddSAndrea Arcangeli vma = find_mergeable_vma(mm, rmap_item->address); 134385c6e8ddSAndrea Arcangeli if (!vma) 13449ba69294SHugh Dickins goto out; 13459ba69294SHugh Dickins 13468dd3557aSHugh Dickins err = try_to_merge_one_page(vma, page, kpage); 1347db114b83SHugh Dickins if (err) 1348db114b83SHugh Dickins goto out; 1349db114b83SHugh Dickins 1350bc56620bSHugh Dickins /* Unstable nid is in union with stable anon_vma: remove first */ 1351bc56620bSHugh Dickins remove_rmap_item_from_tree(rmap_item); 1352bc56620bSHugh Dickins 1353c1e8d7c6SMichel Lespinasse /* Must get reference to anon_vma while still holding mmap_lock */ 13549e60109fSPeter Zijlstra rmap_item->anon_vma = vma->anon_vma; 13559e60109fSPeter Zijlstra get_anon_vma(vma->anon_vma); 135681464e30SHugh Dickins out: 1357d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1358739100c8SStefan Roesch trace_ksm_merge_with_ksm_page(kpage, page_to_pfn(kpage ? kpage : page), 1359739100c8SStefan Roesch rmap_item, mm, err); 136081464e30SHugh Dickins return err; 136181464e30SHugh Dickins } 136281464e30SHugh Dickins 136381464e30SHugh Dickins /* 136431dbd01fSIzik Eidus * try_to_merge_two_pages - take two identical pages and prepare them 136531dbd01fSIzik Eidus * to be merged into one page. 136631dbd01fSIzik Eidus * 13678dd3557aSHugh Dickins * This function returns the kpage if we successfully merged two identical 13688dd3557aSHugh Dickins * pages into one ksm page, NULL otherwise. 136931dbd01fSIzik Eidus * 137080e14822SHugh Dickins * Note that this function upgrades page to ksm page: if one of the pages 137131dbd01fSIzik Eidus * is already a ksm page, try_to_merge_with_ksm_page should be used. 137231dbd01fSIzik Eidus */ 137321fbd591SQi Zheng static struct page *try_to_merge_two_pages(struct ksm_rmap_item *rmap_item, 13748dd3557aSHugh Dickins struct page *page, 137521fbd591SQi Zheng struct ksm_rmap_item *tree_rmap_item, 13768dd3557aSHugh Dickins struct page *tree_page) 137731dbd01fSIzik Eidus { 137880e14822SHugh Dickins int err; 137931dbd01fSIzik Eidus 138080e14822SHugh Dickins err = try_to_merge_with_ksm_page(rmap_item, page, NULL); 138131dbd01fSIzik Eidus if (!err) { 13828dd3557aSHugh Dickins err = try_to_merge_with_ksm_page(tree_rmap_item, 138380e14822SHugh Dickins tree_page, page); 138431dbd01fSIzik Eidus /* 138581464e30SHugh Dickins * If that fails, we have a ksm page with only one pte 138681464e30SHugh Dickins * pointing to it: so break it. 138731dbd01fSIzik Eidus */ 13884035c07aSHugh Dickins if (err) 13898dd3557aSHugh Dickins break_cow(rmap_item); 139031dbd01fSIzik Eidus } 139180e14822SHugh Dickins return err ? NULL : page; 139231dbd01fSIzik Eidus } 139331dbd01fSIzik Eidus 13942c653d0eSAndrea Arcangeli static __always_inline 139521fbd591SQi Zheng bool __is_page_sharing_candidate(struct ksm_stable_node *stable_node, int offset) 13962c653d0eSAndrea Arcangeli { 13972c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len < 0); 13982c653d0eSAndrea Arcangeli /* 13992c653d0eSAndrea Arcangeli * Check that at least one mapping still exists, otherwise 14002c653d0eSAndrea Arcangeli * there's no much point to merge and share with this 14012c653d0eSAndrea Arcangeli * stable_node, as the underlying tree_page of the other 14022c653d0eSAndrea Arcangeli * sharer is going to be freed soon. 14032c653d0eSAndrea Arcangeli */ 14042c653d0eSAndrea Arcangeli return stable_node->rmap_hlist_len && 14052c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len + offset < ksm_max_page_sharing; 14062c653d0eSAndrea Arcangeli } 14072c653d0eSAndrea Arcangeli 14082c653d0eSAndrea Arcangeli static __always_inline 140921fbd591SQi Zheng bool is_page_sharing_candidate(struct ksm_stable_node *stable_node) 14102c653d0eSAndrea Arcangeli { 14112c653d0eSAndrea Arcangeli return __is_page_sharing_candidate(stable_node, 0); 14122c653d0eSAndrea Arcangeli } 14132c653d0eSAndrea Arcangeli 141421fbd591SQi Zheng static struct page *stable_node_dup(struct ksm_stable_node **_stable_node_dup, 141521fbd591SQi Zheng struct ksm_stable_node **_stable_node, 14162c653d0eSAndrea Arcangeli struct rb_root *root, 14172c653d0eSAndrea Arcangeli bool prune_stale_stable_nodes) 14182c653d0eSAndrea Arcangeli { 141921fbd591SQi Zheng struct ksm_stable_node *dup, *found = NULL, *stable_node = *_stable_node; 14202c653d0eSAndrea Arcangeli struct hlist_node *hlist_safe; 14218dc5ffcdSAndrea Arcangeli struct page *_tree_page, *tree_page = NULL; 14222c653d0eSAndrea Arcangeli int nr = 0; 14232c653d0eSAndrea Arcangeli int found_rmap_hlist_len; 14242c653d0eSAndrea Arcangeli 14252c653d0eSAndrea Arcangeli if (!prune_stale_stable_nodes || 14262c653d0eSAndrea Arcangeli time_before(jiffies, stable_node->chain_prune_time + 14272c653d0eSAndrea Arcangeli msecs_to_jiffies( 14282c653d0eSAndrea Arcangeli ksm_stable_node_chains_prune_millisecs))) 14292c653d0eSAndrea Arcangeli prune_stale_stable_nodes = false; 14302c653d0eSAndrea Arcangeli else 14312c653d0eSAndrea Arcangeli stable_node->chain_prune_time = jiffies; 14322c653d0eSAndrea Arcangeli 14332c653d0eSAndrea Arcangeli hlist_for_each_entry_safe(dup, hlist_safe, 14342c653d0eSAndrea Arcangeli &stable_node->hlist, hlist_dup) { 14352c653d0eSAndrea Arcangeli cond_resched(); 14362c653d0eSAndrea Arcangeli /* 14372c653d0eSAndrea Arcangeli * We must walk all stable_node_dup to prune the stale 14382c653d0eSAndrea Arcangeli * stable nodes during lookup. 14392c653d0eSAndrea Arcangeli * 14402c653d0eSAndrea Arcangeli * get_ksm_page can drop the nodes from the 14412c653d0eSAndrea Arcangeli * stable_node->hlist if they point to freed pages 14422c653d0eSAndrea Arcangeli * (that's why we do a _safe walk). The "dup" 14432c653d0eSAndrea Arcangeli * stable_node parameter itself will be freed from 14442c653d0eSAndrea Arcangeli * under us if it returns NULL. 14452c653d0eSAndrea Arcangeli */ 14462cee57d1SYang Shi _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK); 14472c653d0eSAndrea Arcangeli if (!_tree_page) 14482c653d0eSAndrea Arcangeli continue; 14492c653d0eSAndrea Arcangeli nr += 1; 14502c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(dup)) { 14512c653d0eSAndrea Arcangeli if (!found || 14522c653d0eSAndrea Arcangeli dup->rmap_hlist_len > found_rmap_hlist_len) { 14532c653d0eSAndrea Arcangeli if (found) 14548dc5ffcdSAndrea Arcangeli put_page(tree_page); 14552c653d0eSAndrea Arcangeli found = dup; 14562c653d0eSAndrea Arcangeli found_rmap_hlist_len = found->rmap_hlist_len; 14578dc5ffcdSAndrea Arcangeli tree_page = _tree_page; 14582c653d0eSAndrea Arcangeli 14598dc5ffcdSAndrea Arcangeli /* skip put_page for found dup */ 14602c653d0eSAndrea Arcangeli if (!prune_stale_stable_nodes) 14612c653d0eSAndrea Arcangeli break; 14622c653d0eSAndrea Arcangeli continue; 14632c653d0eSAndrea Arcangeli } 14642c653d0eSAndrea Arcangeli } 14652c653d0eSAndrea Arcangeli put_page(_tree_page); 14662c653d0eSAndrea Arcangeli } 14672c653d0eSAndrea Arcangeli 146880b18dfaSAndrea Arcangeli if (found) { 14692c653d0eSAndrea Arcangeli /* 147080b18dfaSAndrea Arcangeli * nr is counting all dups in the chain only if 147180b18dfaSAndrea Arcangeli * prune_stale_stable_nodes is true, otherwise we may 147280b18dfaSAndrea Arcangeli * break the loop at nr == 1 even if there are 147380b18dfaSAndrea Arcangeli * multiple entries. 14742c653d0eSAndrea Arcangeli */ 147580b18dfaSAndrea Arcangeli if (prune_stale_stable_nodes && nr == 1) { 14762c653d0eSAndrea Arcangeli /* 14772c653d0eSAndrea Arcangeli * If there's not just one entry it would 14782c653d0eSAndrea Arcangeli * corrupt memory, better BUG_ON. In KSM 14792c653d0eSAndrea Arcangeli * context with no lock held it's not even 14802c653d0eSAndrea Arcangeli * fatal. 14812c653d0eSAndrea Arcangeli */ 14822c653d0eSAndrea Arcangeli BUG_ON(stable_node->hlist.first->next); 14832c653d0eSAndrea Arcangeli 14842c653d0eSAndrea Arcangeli /* 14852c653d0eSAndrea Arcangeli * There's just one entry and it is below the 14862c653d0eSAndrea Arcangeli * deduplication limit so drop the chain. 14872c653d0eSAndrea Arcangeli */ 14882c653d0eSAndrea Arcangeli rb_replace_node(&stable_node->node, &found->node, 14892c653d0eSAndrea Arcangeli root); 14902c653d0eSAndrea Arcangeli free_stable_node(stable_node); 14912c653d0eSAndrea Arcangeli ksm_stable_node_chains--; 14922c653d0eSAndrea Arcangeli ksm_stable_node_dups--; 1493b4fecc67SAndrea Arcangeli /* 14940ba1d0f7SAndrea Arcangeli * NOTE: the caller depends on the stable_node 14950ba1d0f7SAndrea Arcangeli * to be equal to stable_node_dup if the chain 14960ba1d0f7SAndrea Arcangeli * was collapsed. 1497b4fecc67SAndrea Arcangeli */ 14980ba1d0f7SAndrea Arcangeli *_stable_node = found; 14990ba1d0f7SAndrea Arcangeli /* 1500f0953a1bSIngo Molnar * Just for robustness, as stable_node is 15010ba1d0f7SAndrea Arcangeli * otherwise left as a stable pointer, the 15020ba1d0f7SAndrea Arcangeli * compiler shall optimize it away at build 15030ba1d0f7SAndrea Arcangeli * time. 15040ba1d0f7SAndrea Arcangeli */ 15050ba1d0f7SAndrea Arcangeli stable_node = NULL; 150680b18dfaSAndrea Arcangeli } else if (stable_node->hlist.first != &found->hlist_dup && 150780b18dfaSAndrea Arcangeli __is_page_sharing_candidate(found, 1)) { 15082c653d0eSAndrea Arcangeli /* 150980b18dfaSAndrea Arcangeli * If the found stable_node dup can accept one 151080b18dfaSAndrea Arcangeli * more future merge (in addition to the one 151180b18dfaSAndrea Arcangeli * that is underway) and is not at the head of 151280b18dfaSAndrea Arcangeli * the chain, put it there so next search will 151380b18dfaSAndrea Arcangeli * be quicker in the !prune_stale_stable_nodes 151480b18dfaSAndrea Arcangeli * case. 151580b18dfaSAndrea Arcangeli * 151680b18dfaSAndrea Arcangeli * NOTE: it would be inaccurate to use nr > 1 151780b18dfaSAndrea Arcangeli * instead of checking the hlist.first pointer 151880b18dfaSAndrea Arcangeli * directly, because in the 151980b18dfaSAndrea Arcangeli * prune_stale_stable_nodes case "nr" isn't 152080b18dfaSAndrea Arcangeli * the position of the found dup in the chain, 152180b18dfaSAndrea Arcangeli * but the total number of dups in the chain. 15222c653d0eSAndrea Arcangeli */ 15232c653d0eSAndrea Arcangeli hlist_del(&found->hlist_dup); 15242c653d0eSAndrea Arcangeli hlist_add_head(&found->hlist_dup, 15252c653d0eSAndrea Arcangeli &stable_node->hlist); 15262c653d0eSAndrea Arcangeli } 15272c653d0eSAndrea Arcangeli } 15282c653d0eSAndrea Arcangeli 15298dc5ffcdSAndrea Arcangeli *_stable_node_dup = found; 15308dc5ffcdSAndrea Arcangeli return tree_page; 15312c653d0eSAndrea Arcangeli } 15322c653d0eSAndrea Arcangeli 153321fbd591SQi Zheng static struct ksm_stable_node *stable_node_dup_any(struct ksm_stable_node *stable_node, 15342c653d0eSAndrea Arcangeli struct rb_root *root) 15352c653d0eSAndrea Arcangeli { 15362c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) 15372c653d0eSAndrea Arcangeli return stable_node; 15382c653d0eSAndrea Arcangeli if (hlist_empty(&stable_node->hlist)) { 15392c653d0eSAndrea Arcangeli free_stable_node_chain(stable_node, root); 15402c653d0eSAndrea Arcangeli return NULL; 15412c653d0eSAndrea Arcangeli } 15422c653d0eSAndrea Arcangeli return hlist_entry(stable_node->hlist.first, 15432c653d0eSAndrea Arcangeli typeof(*stable_node), hlist_dup); 15442c653d0eSAndrea Arcangeli } 15452c653d0eSAndrea Arcangeli 15468dc5ffcdSAndrea Arcangeli /* 15478dc5ffcdSAndrea Arcangeli * Like for get_ksm_page, this function can free the *_stable_node and 15488dc5ffcdSAndrea Arcangeli * *_stable_node_dup if the returned tree_page is NULL. 15498dc5ffcdSAndrea Arcangeli * 15508dc5ffcdSAndrea Arcangeli * It can also free and overwrite *_stable_node with the found 15518dc5ffcdSAndrea Arcangeli * stable_node_dup if the chain is collapsed (in which case 15528dc5ffcdSAndrea Arcangeli * *_stable_node will be equal to *_stable_node_dup like if the chain 15538dc5ffcdSAndrea Arcangeli * never existed). It's up to the caller to verify tree_page is not 15548dc5ffcdSAndrea Arcangeli * NULL before dereferencing *_stable_node or *_stable_node_dup. 15558dc5ffcdSAndrea Arcangeli * 15568dc5ffcdSAndrea Arcangeli * *_stable_node_dup is really a second output parameter of this 15578dc5ffcdSAndrea Arcangeli * function and will be overwritten in all cases, the caller doesn't 15588dc5ffcdSAndrea Arcangeli * need to initialize it. 15598dc5ffcdSAndrea Arcangeli */ 156021fbd591SQi Zheng static struct page *__stable_node_chain(struct ksm_stable_node **_stable_node_dup, 156121fbd591SQi Zheng struct ksm_stable_node **_stable_node, 15622c653d0eSAndrea Arcangeli struct rb_root *root, 15632c653d0eSAndrea Arcangeli bool prune_stale_stable_nodes) 15642c653d0eSAndrea Arcangeli { 156521fbd591SQi Zheng struct ksm_stable_node *stable_node = *_stable_node; 15662c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 15672c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(stable_node)) { 15688dc5ffcdSAndrea Arcangeli *_stable_node_dup = stable_node; 15692cee57d1SYang Shi return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK); 15702c653d0eSAndrea Arcangeli } 15718dc5ffcdSAndrea Arcangeli /* 15728dc5ffcdSAndrea Arcangeli * _stable_node_dup set to NULL means the stable_node 15738dc5ffcdSAndrea Arcangeli * reached the ksm_max_page_sharing limit. 15748dc5ffcdSAndrea Arcangeli */ 15758dc5ffcdSAndrea Arcangeli *_stable_node_dup = NULL; 15762c653d0eSAndrea Arcangeli return NULL; 15772c653d0eSAndrea Arcangeli } 15788dc5ffcdSAndrea Arcangeli return stable_node_dup(_stable_node_dup, _stable_node, root, 15792c653d0eSAndrea Arcangeli prune_stale_stable_nodes); 15802c653d0eSAndrea Arcangeli } 15812c653d0eSAndrea Arcangeli 158221fbd591SQi Zheng static __always_inline struct page *chain_prune(struct ksm_stable_node **s_n_d, 158321fbd591SQi Zheng struct ksm_stable_node **s_n, 15842c653d0eSAndrea Arcangeli struct rb_root *root) 15852c653d0eSAndrea Arcangeli { 15868dc5ffcdSAndrea Arcangeli return __stable_node_chain(s_n_d, s_n, root, true); 15872c653d0eSAndrea Arcangeli } 15882c653d0eSAndrea Arcangeli 158921fbd591SQi Zheng static __always_inline struct page *chain(struct ksm_stable_node **s_n_d, 159021fbd591SQi Zheng struct ksm_stable_node *s_n, 15912c653d0eSAndrea Arcangeli struct rb_root *root) 15922c653d0eSAndrea Arcangeli { 159321fbd591SQi Zheng struct ksm_stable_node *old_stable_node = s_n; 15948dc5ffcdSAndrea Arcangeli struct page *tree_page; 15958dc5ffcdSAndrea Arcangeli 15968dc5ffcdSAndrea Arcangeli tree_page = __stable_node_chain(s_n_d, &s_n, root, false); 15978dc5ffcdSAndrea Arcangeli /* not pruning dups so s_n cannot have changed */ 15988dc5ffcdSAndrea Arcangeli VM_BUG_ON(s_n != old_stable_node); 15998dc5ffcdSAndrea Arcangeli return tree_page; 16002c653d0eSAndrea Arcangeli } 16012c653d0eSAndrea Arcangeli 160231dbd01fSIzik Eidus /* 16038dd3557aSHugh Dickins * stable_tree_search - search for page inside the stable tree 160431dbd01fSIzik Eidus * 160531dbd01fSIzik Eidus * This function checks if there is a page inside the stable tree 160631dbd01fSIzik Eidus * with identical content to the page that we are scanning right now. 160731dbd01fSIzik Eidus * 16087b6ba2c7SHugh Dickins * This function returns the stable tree node of identical content if found, 160931dbd01fSIzik Eidus * NULL otherwise. 161031dbd01fSIzik Eidus */ 161162b61f61SHugh Dickins static struct page *stable_tree_search(struct page *page) 161231dbd01fSIzik Eidus { 161390bd6fd3SPetr Holasek int nid; 1614ef53d16cSHugh Dickins struct rb_root *root; 16154146d2d6SHugh Dickins struct rb_node **new; 16164146d2d6SHugh Dickins struct rb_node *parent; 161721fbd591SQi Zheng struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; 161821fbd591SQi Zheng struct ksm_stable_node *page_node; 161931dbd01fSIzik Eidus 16204146d2d6SHugh Dickins page_node = page_stable_node(page); 16214146d2d6SHugh Dickins if (page_node && page_node->head != &migrate_nodes) { 16224146d2d6SHugh Dickins /* ksm page forked */ 162308beca44SHugh Dickins get_page(page); 162462b61f61SHugh Dickins return page; 162508beca44SHugh Dickins } 162608beca44SHugh Dickins 162790bd6fd3SPetr Holasek nid = get_kpfn_nid(page_to_pfn(page)); 1628ef53d16cSHugh Dickins root = root_stable_tree + nid; 16294146d2d6SHugh Dickins again: 1630ef53d16cSHugh Dickins new = &root->rb_node; 16314146d2d6SHugh Dickins parent = NULL; 163290bd6fd3SPetr Holasek 16334146d2d6SHugh Dickins while (*new) { 16344035c07aSHugh Dickins struct page *tree_page; 163531dbd01fSIzik Eidus int ret; 163631dbd01fSIzik Eidus 163731dbd01fSIzik Eidus cond_resched(); 163821fbd591SQi Zheng stable_node = rb_entry(*new, struct ksm_stable_node, node); 16392c653d0eSAndrea Arcangeli stable_node_any = NULL; 16408dc5ffcdSAndrea Arcangeli tree_page = chain_prune(&stable_node_dup, &stable_node, root); 1641b4fecc67SAndrea Arcangeli /* 1642b4fecc67SAndrea Arcangeli * NOTE: stable_node may have been freed by 1643b4fecc67SAndrea Arcangeli * chain_prune() if the returned stable_node_dup is 1644b4fecc67SAndrea Arcangeli * not NULL. stable_node_dup may have been inserted in 1645b4fecc67SAndrea Arcangeli * the rbtree instead as a regular stable_node (in 1646b4fecc67SAndrea Arcangeli * order to collapse the stable_node chain if a single 16470ba1d0f7SAndrea Arcangeli * stable_node dup was found in it). In such case the 16483413b2c8SJulia Lawall * stable_node is overwritten by the callee to point 16490ba1d0f7SAndrea Arcangeli * to the stable_node_dup that was collapsed in the 16500ba1d0f7SAndrea Arcangeli * stable rbtree and stable_node will be equal to 16510ba1d0f7SAndrea Arcangeli * stable_node_dup like if the chain never existed. 1652b4fecc67SAndrea Arcangeli */ 16532c653d0eSAndrea Arcangeli if (!stable_node_dup) { 16542c653d0eSAndrea Arcangeli /* 16552c653d0eSAndrea Arcangeli * Either all stable_node dups were full in 16562c653d0eSAndrea Arcangeli * this stable_node chain, or this chain was 16572c653d0eSAndrea Arcangeli * empty and should be rb_erased. 16582c653d0eSAndrea Arcangeli */ 16592c653d0eSAndrea Arcangeli stable_node_any = stable_node_dup_any(stable_node, 16602c653d0eSAndrea Arcangeli root); 16612c653d0eSAndrea Arcangeli if (!stable_node_any) { 16622c653d0eSAndrea Arcangeli /* rb_erase just run */ 16632c653d0eSAndrea Arcangeli goto again; 16642c653d0eSAndrea Arcangeli } 16652c653d0eSAndrea Arcangeli /* 16662c653d0eSAndrea Arcangeli * Take any of the stable_node dups page of 16672c653d0eSAndrea Arcangeli * this stable_node chain to let the tree walk 16682c653d0eSAndrea Arcangeli * continue. All KSM pages belonging to the 16692c653d0eSAndrea Arcangeli * stable_node dups in a stable_node chain 16702c653d0eSAndrea Arcangeli * have the same content and they're 1671457aef94SEthon Paul * write protected at all times. Any will work 16722c653d0eSAndrea Arcangeli * fine to continue the walk. 16732c653d0eSAndrea Arcangeli */ 16742cee57d1SYang Shi tree_page = get_ksm_page(stable_node_any, 16752cee57d1SYang Shi GET_KSM_PAGE_NOLOCK); 16762c653d0eSAndrea Arcangeli } 16772c653d0eSAndrea Arcangeli VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); 1678f2e5ff85SAndrea Arcangeli if (!tree_page) { 1679f2e5ff85SAndrea Arcangeli /* 1680f2e5ff85SAndrea Arcangeli * If we walked over a stale stable_node, 1681f2e5ff85SAndrea Arcangeli * get_ksm_page() will call rb_erase() and it 1682f2e5ff85SAndrea Arcangeli * may rebalance the tree from under us. So 1683f2e5ff85SAndrea Arcangeli * restart the search from scratch. Returning 1684f2e5ff85SAndrea Arcangeli * NULL would be safe too, but we'd generate 1685f2e5ff85SAndrea Arcangeli * false negative insertions just because some 1686f2e5ff85SAndrea Arcangeli * stable_node was stale. 1687f2e5ff85SAndrea Arcangeli */ 1688f2e5ff85SAndrea Arcangeli goto again; 1689f2e5ff85SAndrea Arcangeli } 169031dbd01fSIzik Eidus 16914035c07aSHugh Dickins ret = memcmp_pages(page, tree_page); 1692c8d6553bSHugh Dickins put_page(tree_page); 169331dbd01fSIzik Eidus 16944146d2d6SHugh Dickins parent = *new; 1695c8d6553bSHugh Dickins if (ret < 0) 16964146d2d6SHugh Dickins new = &parent->rb_left; 1697c8d6553bSHugh Dickins else if (ret > 0) 16984146d2d6SHugh Dickins new = &parent->rb_right; 1699c8d6553bSHugh Dickins else { 17002c653d0eSAndrea Arcangeli if (page_node) { 17012c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 17022c653d0eSAndrea Arcangeli /* 17032c653d0eSAndrea Arcangeli * Test if the migrated page should be merged 17042c653d0eSAndrea Arcangeli * into a stable node dup. If the mapcount is 17052c653d0eSAndrea Arcangeli * 1 we can migrate it with another KSM page 17062c653d0eSAndrea Arcangeli * without adding it to the chain. 17072c653d0eSAndrea Arcangeli */ 17082c653d0eSAndrea Arcangeli if (page_mapcount(page) > 1) 17092c653d0eSAndrea Arcangeli goto chain_append; 17102c653d0eSAndrea Arcangeli } 17112c653d0eSAndrea Arcangeli 17122c653d0eSAndrea Arcangeli if (!stable_node_dup) { 17132c653d0eSAndrea Arcangeli /* 17142c653d0eSAndrea Arcangeli * If the stable_node is a chain and 17152c653d0eSAndrea Arcangeli * we got a payload match in memcmp 17162c653d0eSAndrea Arcangeli * but we cannot merge the scanned 17172c653d0eSAndrea Arcangeli * page in any of the existing 17182c653d0eSAndrea Arcangeli * stable_node dups because they're 17192c653d0eSAndrea Arcangeli * all full, we need to wait the 17202c653d0eSAndrea Arcangeli * scanned page to find itself a match 17212c653d0eSAndrea Arcangeli * in the unstable tree to create a 17222c653d0eSAndrea Arcangeli * brand new KSM page to add later to 17232c653d0eSAndrea Arcangeli * the dups of this stable_node. 17242c653d0eSAndrea Arcangeli */ 17252c653d0eSAndrea Arcangeli return NULL; 17262c653d0eSAndrea Arcangeli } 17272c653d0eSAndrea Arcangeli 1728c8d6553bSHugh Dickins /* 1729c8d6553bSHugh Dickins * Lock and unlock the stable_node's page (which 1730c8d6553bSHugh Dickins * might already have been migrated) so that page 1731c8d6553bSHugh Dickins * migration is sure to notice its raised count. 1732c8d6553bSHugh Dickins * It would be more elegant to return stable_node 1733c8d6553bSHugh Dickins * than kpage, but that involves more changes. 1734c8d6553bSHugh Dickins */ 17352cee57d1SYang Shi tree_page = get_ksm_page(stable_node_dup, 17362cee57d1SYang Shi GET_KSM_PAGE_TRYLOCK); 17372cee57d1SYang Shi 17382cee57d1SYang Shi if (PTR_ERR(tree_page) == -EBUSY) 17392cee57d1SYang Shi return ERR_PTR(-EBUSY); 17402cee57d1SYang Shi 17412c653d0eSAndrea Arcangeli if (unlikely(!tree_page)) 17422c653d0eSAndrea Arcangeli /* 17432c653d0eSAndrea Arcangeli * The tree may have been rebalanced, 17442c653d0eSAndrea Arcangeli * so re-evaluate parent and new. 17452c653d0eSAndrea Arcangeli */ 17462c653d0eSAndrea Arcangeli goto again; 1747c8d6553bSHugh Dickins unlock_page(tree_page); 17482c653d0eSAndrea Arcangeli 17492c653d0eSAndrea Arcangeli if (get_kpfn_nid(stable_node_dup->kpfn) != 17502c653d0eSAndrea Arcangeli NUMA(stable_node_dup->nid)) { 17514146d2d6SHugh Dickins put_page(tree_page); 17524146d2d6SHugh Dickins goto replace; 17534146d2d6SHugh Dickins } 175462b61f61SHugh Dickins return tree_page; 175531dbd01fSIzik Eidus } 1756c8d6553bSHugh Dickins } 175731dbd01fSIzik Eidus 17584146d2d6SHugh Dickins if (!page_node) 175931dbd01fSIzik Eidus return NULL; 17604146d2d6SHugh Dickins 17614146d2d6SHugh Dickins list_del(&page_node->list); 17624146d2d6SHugh Dickins DO_NUMA(page_node->nid = nid); 17634146d2d6SHugh Dickins rb_link_node(&page_node->node, parent, new); 1764ef53d16cSHugh Dickins rb_insert_color(&page_node->node, root); 17652c653d0eSAndrea Arcangeli out: 17662c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(page_node)) { 17674146d2d6SHugh Dickins get_page(page); 17684146d2d6SHugh Dickins return page; 17692c653d0eSAndrea Arcangeli } else 17702c653d0eSAndrea Arcangeli return NULL; 17714146d2d6SHugh Dickins 17724146d2d6SHugh Dickins replace: 1773b4fecc67SAndrea Arcangeli /* 1774b4fecc67SAndrea Arcangeli * If stable_node was a chain and chain_prune collapsed it, 17750ba1d0f7SAndrea Arcangeli * stable_node has been updated to be the new regular 17760ba1d0f7SAndrea Arcangeli * stable_node. A collapse of the chain is indistinguishable 17770ba1d0f7SAndrea Arcangeli * from the case there was no chain in the stable 17780ba1d0f7SAndrea Arcangeli * rbtree. Otherwise stable_node is the chain and 17790ba1d0f7SAndrea Arcangeli * stable_node_dup is the dup to replace. 1780b4fecc67SAndrea Arcangeli */ 17810ba1d0f7SAndrea Arcangeli if (stable_node_dup == stable_node) { 1782b4fecc67SAndrea Arcangeli VM_BUG_ON(is_stable_node_chain(stable_node_dup)); 1783b4fecc67SAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node_dup)); 17842c653d0eSAndrea Arcangeli /* there is no chain */ 17854146d2d6SHugh Dickins if (page_node) { 17862c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 17874146d2d6SHugh Dickins list_del(&page_node->list); 17884146d2d6SHugh Dickins DO_NUMA(page_node->nid = nid); 1789b4fecc67SAndrea Arcangeli rb_replace_node(&stable_node_dup->node, 1790b4fecc67SAndrea Arcangeli &page_node->node, 17912c653d0eSAndrea Arcangeli root); 17922c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(page_node)) 17934146d2d6SHugh Dickins get_page(page); 17942c653d0eSAndrea Arcangeli else 17952c653d0eSAndrea Arcangeli page = NULL; 17964146d2d6SHugh Dickins } else { 1797b4fecc67SAndrea Arcangeli rb_erase(&stable_node_dup->node, root); 17984146d2d6SHugh Dickins page = NULL; 17994146d2d6SHugh Dickins } 18002c653d0eSAndrea Arcangeli } else { 18012c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_chain(stable_node)); 18022c653d0eSAndrea Arcangeli __stable_node_dup_del(stable_node_dup); 18032c653d0eSAndrea Arcangeli if (page_node) { 18042c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 18052c653d0eSAndrea Arcangeli list_del(&page_node->list); 18062c653d0eSAndrea Arcangeli DO_NUMA(page_node->nid = nid); 18072c653d0eSAndrea Arcangeli stable_node_chain_add_dup(page_node, stable_node); 18082c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(page_node)) 18092c653d0eSAndrea Arcangeli get_page(page); 18102c653d0eSAndrea Arcangeli else 18112c653d0eSAndrea Arcangeli page = NULL; 18122c653d0eSAndrea Arcangeli } else { 18132c653d0eSAndrea Arcangeli page = NULL; 18142c653d0eSAndrea Arcangeli } 18152c653d0eSAndrea Arcangeli } 18162c653d0eSAndrea Arcangeli stable_node_dup->head = &migrate_nodes; 18172c653d0eSAndrea Arcangeli list_add(&stable_node_dup->list, stable_node_dup->head); 18184146d2d6SHugh Dickins return page; 18192c653d0eSAndrea Arcangeli 18202c653d0eSAndrea Arcangeli chain_append: 18212c653d0eSAndrea Arcangeli /* stable_node_dup could be null if it reached the limit */ 18222c653d0eSAndrea Arcangeli if (!stable_node_dup) 18232c653d0eSAndrea Arcangeli stable_node_dup = stable_node_any; 1824b4fecc67SAndrea Arcangeli /* 1825b4fecc67SAndrea Arcangeli * If stable_node was a chain and chain_prune collapsed it, 18260ba1d0f7SAndrea Arcangeli * stable_node has been updated to be the new regular 18270ba1d0f7SAndrea Arcangeli * stable_node. A collapse of the chain is indistinguishable 18280ba1d0f7SAndrea Arcangeli * from the case there was no chain in the stable 18290ba1d0f7SAndrea Arcangeli * rbtree. Otherwise stable_node is the chain and 18300ba1d0f7SAndrea Arcangeli * stable_node_dup is the dup to replace. 1831b4fecc67SAndrea Arcangeli */ 18320ba1d0f7SAndrea Arcangeli if (stable_node_dup == stable_node) { 1833b4fecc67SAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node_dup)); 18342c653d0eSAndrea Arcangeli /* chain is missing so create it */ 18352c653d0eSAndrea Arcangeli stable_node = alloc_stable_node_chain(stable_node_dup, 18362c653d0eSAndrea Arcangeli root); 18372c653d0eSAndrea Arcangeli if (!stable_node) 18382c653d0eSAndrea Arcangeli return NULL; 18392c653d0eSAndrea Arcangeli } 18402c653d0eSAndrea Arcangeli /* 18412c653d0eSAndrea Arcangeli * Add this stable_node dup that was 18422c653d0eSAndrea Arcangeli * migrated to the stable_node chain 18432c653d0eSAndrea Arcangeli * of the current nid for this page 18442c653d0eSAndrea Arcangeli * content. 18452c653d0eSAndrea Arcangeli */ 1846b4fecc67SAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(stable_node_dup)); 18472c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 18482c653d0eSAndrea Arcangeli list_del(&page_node->list); 18492c653d0eSAndrea Arcangeli DO_NUMA(page_node->nid = nid); 18502c653d0eSAndrea Arcangeli stable_node_chain_add_dup(page_node, stable_node); 18512c653d0eSAndrea Arcangeli goto out; 185231dbd01fSIzik Eidus } 185331dbd01fSIzik Eidus 185431dbd01fSIzik Eidus /* 1855e850dcf5SHugh Dickins * stable_tree_insert - insert stable tree node pointing to new ksm page 185631dbd01fSIzik Eidus * into the stable tree. 185731dbd01fSIzik Eidus * 18587b6ba2c7SHugh Dickins * This function returns the stable tree node just allocated on success, 18597b6ba2c7SHugh Dickins * NULL otherwise. 186031dbd01fSIzik Eidus */ 186121fbd591SQi Zheng static struct ksm_stable_node *stable_tree_insert(struct page *kpage) 186231dbd01fSIzik Eidus { 186390bd6fd3SPetr Holasek int nid; 186490bd6fd3SPetr Holasek unsigned long kpfn; 1865ef53d16cSHugh Dickins struct rb_root *root; 186690bd6fd3SPetr Holasek struct rb_node **new; 1867f2e5ff85SAndrea Arcangeli struct rb_node *parent; 186821fbd591SQi Zheng struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; 18692c653d0eSAndrea Arcangeli bool need_chain = false; 187031dbd01fSIzik Eidus 187190bd6fd3SPetr Holasek kpfn = page_to_pfn(kpage); 187290bd6fd3SPetr Holasek nid = get_kpfn_nid(kpfn); 1873ef53d16cSHugh Dickins root = root_stable_tree + nid; 1874f2e5ff85SAndrea Arcangeli again: 1875f2e5ff85SAndrea Arcangeli parent = NULL; 1876ef53d16cSHugh Dickins new = &root->rb_node; 187790bd6fd3SPetr Holasek 187831dbd01fSIzik Eidus while (*new) { 18794035c07aSHugh Dickins struct page *tree_page; 188031dbd01fSIzik Eidus int ret; 188131dbd01fSIzik Eidus 188231dbd01fSIzik Eidus cond_resched(); 188321fbd591SQi Zheng stable_node = rb_entry(*new, struct ksm_stable_node, node); 18842c653d0eSAndrea Arcangeli stable_node_any = NULL; 18858dc5ffcdSAndrea Arcangeli tree_page = chain(&stable_node_dup, stable_node, root); 18862c653d0eSAndrea Arcangeli if (!stable_node_dup) { 18872c653d0eSAndrea Arcangeli /* 18882c653d0eSAndrea Arcangeli * Either all stable_node dups were full in 18892c653d0eSAndrea Arcangeli * this stable_node chain, or this chain was 18902c653d0eSAndrea Arcangeli * empty and should be rb_erased. 18912c653d0eSAndrea Arcangeli */ 18922c653d0eSAndrea Arcangeli stable_node_any = stable_node_dup_any(stable_node, 18932c653d0eSAndrea Arcangeli root); 18942c653d0eSAndrea Arcangeli if (!stable_node_any) { 18952c653d0eSAndrea Arcangeli /* rb_erase just run */ 18962c653d0eSAndrea Arcangeli goto again; 18972c653d0eSAndrea Arcangeli } 18982c653d0eSAndrea Arcangeli /* 18992c653d0eSAndrea Arcangeli * Take any of the stable_node dups page of 19002c653d0eSAndrea Arcangeli * this stable_node chain to let the tree walk 19012c653d0eSAndrea Arcangeli * continue. All KSM pages belonging to the 19022c653d0eSAndrea Arcangeli * stable_node dups in a stable_node chain 19032c653d0eSAndrea Arcangeli * have the same content and they're 1904457aef94SEthon Paul * write protected at all times. Any will work 19052c653d0eSAndrea Arcangeli * fine to continue the walk. 19062c653d0eSAndrea Arcangeli */ 19072cee57d1SYang Shi tree_page = get_ksm_page(stable_node_any, 19082cee57d1SYang Shi GET_KSM_PAGE_NOLOCK); 19092c653d0eSAndrea Arcangeli } 19102c653d0eSAndrea Arcangeli VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); 1911f2e5ff85SAndrea Arcangeli if (!tree_page) { 1912f2e5ff85SAndrea Arcangeli /* 1913f2e5ff85SAndrea Arcangeli * If we walked over a stale stable_node, 1914f2e5ff85SAndrea Arcangeli * get_ksm_page() will call rb_erase() and it 1915f2e5ff85SAndrea Arcangeli * may rebalance the tree from under us. So 1916f2e5ff85SAndrea Arcangeli * restart the search from scratch. Returning 1917f2e5ff85SAndrea Arcangeli * NULL would be safe too, but we'd generate 1918f2e5ff85SAndrea Arcangeli * false negative insertions just because some 1919f2e5ff85SAndrea Arcangeli * stable_node was stale. 1920f2e5ff85SAndrea Arcangeli */ 1921f2e5ff85SAndrea Arcangeli goto again; 1922f2e5ff85SAndrea Arcangeli } 192331dbd01fSIzik Eidus 19244035c07aSHugh Dickins ret = memcmp_pages(kpage, tree_page); 19254035c07aSHugh Dickins put_page(tree_page); 192631dbd01fSIzik Eidus 192731dbd01fSIzik Eidus parent = *new; 192831dbd01fSIzik Eidus if (ret < 0) 192931dbd01fSIzik Eidus new = &parent->rb_left; 193031dbd01fSIzik Eidus else if (ret > 0) 193131dbd01fSIzik Eidus new = &parent->rb_right; 193231dbd01fSIzik Eidus else { 19332c653d0eSAndrea Arcangeli need_chain = true; 19342c653d0eSAndrea Arcangeli break; 193531dbd01fSIzik Eidus } 193631dbd01fSIzik Eidus } 193731dbd01fSIzik Eidus 19382c653d0eSAndrea Arcangeli stable_node_dup = alloc_stable_node(); 19392c653d0eSAndrea Arcangeli if (!stable_node_dup) 19407b6ba2c7SHugh Dickins return NULL; 194131dbd01fSIzik Eidus 19422c653d0eSAndrea Arcangeli INIT_HLIST_HEAD(&stable_node_dup->hlist); 19432c653d0eSAndrea Arcangeli stable_node_dup->kpfn = kpfn; 19442c653d0eSAndrea Arcangeli set_page_stable_node(kpage, stable_node_dup); 19452c653d0eSAndrea Arcangeli stable_node_dup->rmap_hlist_len = 0; 19462c653d0eSAndrea Arcangeli DO_NUMA(stable_node_dup->nid = nid); 19472c653d0eSAndrea Arcangeli if (!need_chain) { 19482c653d0eSAndrea Arcangeli rb_link_node(&stable_node_dup->node, parent, new); 19492c653d0eSAndrea Arcangeli rb_insert_color(&stable_node_dup->node, root); 19502c653d0eSAndrea Arcangeli } else { 19512c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 195221fbd591SQi Zheng struct ksm_stable_node *orig = stable_node; 19532c653d0eSAndrea Arcangeli /* chain is missing so create it */ 19542c653d0eSAndrea Arcangeli stable_node = alloc_stable_node_chain(orig, root); 19552c653d0eSAndrea Arcangeli if (!stable_node) { 19562c653d0eSAndrea Arcangeli free_stable_node(stable_node_dup); 19572c653d0eSAndrea Arcangeli return NULL; 19582c653d0eSAndrea Arcangeli } 19592c653d0eSAndrea Arcangeli } 19602c653d0eSAndrea Arcangeli stable_node_chain_add_dup(stable_node_dup, stable_node); 19612c653d0eSAndrea Arcangeli } 196208beca44SHugh Dickins 19632c653d0eSAndrea Arcangeli return stable_node_dup; 196431dbd01fSIzik Eidus } 196531dbd01fSIzik Eidus 196631dbd01fSIzik Eidus /* 19678dd3557aSHugh Dickins * unstable_tree_search_insert - search for identical page, 19688dd3557aSHugh Dickins * else insert rmap_item into the unstable tree. 196931dbd01fSIzik Eidus * 197031dbd01fSIzik Eidus * This function searches for a page in the unstable tree identical to the 197131dbd01fSIzik Eidus * page currently being scanned; and if no identical page is found in the 197231dbd01fSIzik Eidus * tree, we insert rmap_item as a new object into the unstable tree. 197331dbd01fSIzik Eidus * 197431dbd01fSIzik Eidus * This function returns pointer to rmap_item found to be identical 197531dbd01fSIzik Eidus * to the currently scanned page, NULL otherwise. 197631dbd01fSIzik Eidus * 197731dbd01fSIzik Eidus * This function does both searching and inserting, because they share 197831dbd01fSIzik Eidus * the same walking algorithm in an rbtree. 197931dbd01fSIzik Eidus */ 19808dd3557aSHugh Dickins static 198121fbd591SQi Zheng struct ksm_rmap_item *unstable_tree_search_insert(struct ksm_rmap_item *rmap_item, 19828dd3557aSHugh Dickins struct page *page, 19838dd3557aSHugh Dickins struct page **tree_pagep) 198431dbd01fSIzik Eidus { 198590bd6fd3SPetr Holasek struct rb_node **new; 198690bd6fd3SPetr Holasek struct rb_root *root; 198731dbd01fSIzik Eidus struct rb_node *parent = NULL; 198890bd6fd3SPetr Holasek int nid; 198990bd6fd3SPetr Holasek 199090bd6fd3SPetr Holasek nid = get_kpfn_nid(page_to_pfn(page)); 1991ef53d16cSHugh Dickins root = root_unstable_tree + nid; 199290bd6fd3SPetr Holasek new = &root->rb_node; 199331dbd01fSIzik Eidus 199431dbd01fSIzik Eidus while (*new) { 199521fbd591SQi Zheng struct ksm_rmap_item *tree_rmap_item; 19968dd3557aSHugh Dickins struct page *tree_page; 199731dbd01fSIzik Eidus int ret; 199831dbd01fSIzik Eidus 1999d178f27fSHugh Dickins cond_resched(); 200021fbd591SQi Zheng tree_rmap_item = rb_entry(*new, struct ksm_rmap_item, node); 20018dd3557aSHugh Dickins tree_page = get_mergeable_page(tree_rmap_item); 2002c8f95ed1SAndrea Arcangeli if (!tree_page) 200331dbd01fSIzik Eidus return NULL; 200431dbd01fSIzik Eidus 200531dbd01fSIzik Eidus /* 20068dd3557aSHugh Dickins * Don't substitute a ksm page for a forked page. 200731dbd01fSIzik Eidus */ 20088dd3557aSHugh Dickins if (page == tree_page) { 20098dd3557aSHugh Dickins put_page(tree_page); 201031dbd01fSIzik Eidus return NULL; 201131dbd01fSIzik Eidus } 201231dbd01fSIzik Eidus 20138dd3557aSHugh Dickins ret = memcmp_pages(page, tree_page); 201431dbd01fSIzik Eidus 201531dbd01fSIzik Eidus parent = *new; 201631dbd01fSIzik Eidus if (ret < 0) { 20178dd3557aSHugh Dickins put_page(tree_page); 201831dbd01fSIzik Eidus new = &parent->rb_left; 201931dbd01fSIzik Eidus } else if (ret > 0) { 20208dd3557aSHugh Dickins put_page(tree_page); 202131dbd01fSIzik Eidus new = &parent->rb_right; 2022b599cbdfSHugh Dickins } else if (!ksm_merge_across_nodes && 2023b599cbdfSHugh Dickins page_to_nid(tree_page) != nid) { 2024b599cbdfSHugh Dickins /* 2025b599cbdfSHugh Dickins * If tree_page has been migrated to another NUMA node, 2026b599cbdfSHugh Dickins * it will be flushed out and put in the right unstable 2027b599cbdfSHugh Dickins * tree next time: only merge with it when across_nodes. 2028b599cbdfSHugh Dickins */ 2029b599cbdfSHugh Dickins put_page(tree_page); 2030b599cbdfSHugh Dickins return NULL; 203131dbd01fSIzik Eidus } else { 20328dd3557aSHugh Dickins *tree_pagep = tree_page; 203331dbd01fSIzik Eidus return tree_rmap_item; 203431dbd01fSIzik Eidus } 203531dbd01fSIzik Eidus } 203631dbd01fSIzik Eidus 20377b6ba2c7SHugh Dickins rmap_item->address |= UNSTABLE_FLAG; 203831dbd01fSIzik Eidus rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); 2039e850dcf5SHugh Dickins DO_NUMA(rmap_item->nid = nid); 204031dbd01fSIzik Eidus rb_link_node(&rmap_item->node, parent, new); 204190bd6fd3SPetr Holasek rb_insert_color(&rmap_item->node, root); 204231dbd01fSIzik Eidus 2043473b0ce4SHugh Dickins ksm_pages_unshared++; 204431dbd01fSIzik Eidus return NULL; 204531dbd01fSIzik Eidus } 204631dbd01fSIzik Eidus 204731dbd01fSIzik Eidus /* 204831dbd01fSIzik Eidus * stable_tree_append - add another rmap_item to the linked list of 204931dbd01fSIzik Eidus * rmap_items hanging off a given node of the stable tree, all sharing 205031dbd01fSIzik Eidus * the same ksm page. 205131dbd01fSIzik Eidus */ 205221fbd591SQi Zheng static void stable_tree_append(struct ksm_rmap_item *rmap_item, 205321fbd591SQi Zheng struct ksm_stable_node *stable_node, 20542c653d0eSAndrea Arcangeli bool max_page_sharing_bypass) 205531dbd01fSIzik Eidus { 20562c653d0eSAndrea Arcangeli /* 20572c653d0eSAndrea Arcangeli * rmap won't find this mapping if we don't insert the 20582c653d0eSAndrea Arcangeli * rmap_item in the right stable_node 20592c653d0eSAndrea Arcangeli * duplicate. page_migration could break later if rmap breaks, 20602c653d0eSAndrea Arcangeli * so we can as well crash here. We really need to check for 20612c653d0eSAndrea Arcangeli * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check 2062457aef94SEthon Paul * for other negative values as an underflow if detected here 20632c653d0eSAndrea Arcangeli * for the first time (and not when decreasing rmap_hlist_len) 20642c653d0eSAndrea Arcangeli * would be sign of memory corruption in the stable_node. 20652c653d0eSAndrea Arcangeli */ 20662c653d0eSAndrea Arcangeli BUG_ON(stable_node->rmap_hlist_len < 0); 20672c653d0eSAndrea Arcangeli 20682c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len++; 20692c653d0eSAndrea Arcangeli if (!max_page_sharing_bypass) 20702c653d0eSAndrea Arcangeli /* possibly non fatal but unexpected overflow, only warn */ 20712c653d0eSAndrea Arcangeli WARN_ON_ONCE(stable_node->rmap_hlist_len > 20722c653d0eSAndrea Arcangeli ksm_max_page_sharing); 20732c653d0eSAndrea Arcangeli 20747b6ba2c7SHugh Dickins rmap_item->head = stable_node; 207531dbd01fSIzik Eidus rmap_item->address |= STABLE_FLAG; 20767b6ba2c7SHugh Dickins hlist_add_head(&rmap_item->hlist, &stable_node->hlist); 2077e178dfdeSHugh Dickins 20787b6ba2c7SHugh Dickins if (rmap_item->hlist.next) 2079e178dfdeSHugh Dickins ksm_pages_sharing++; 20807b6ba2c7SHugh Dickins else 20817b6ba2c7SHugh Dickins ksm_pages_shared++; 208276093853Sxu xin 208376093853Sxu xin rmap_item->mm->ksm_merging_pages++; 208431dbd01fSIzik Eidus } 208531dbd01fSIzik Eidus 208631dbd01fSIzik Eidus /* 208781464e30SHugh Dickins * cmp_and_merge_page - first see if page can be merged into the stable tree; 208881464e30SHugh Dickins * if not, compare checksum to previous and if it's the same, see if page can 208981464e30SHugh Dickins * be inserted into the unstable tree, or merged with a page already there and 209081464e30SHugh Dickins * both transferred to the stable tree. 209131dbd01fSIzik Eidus * 209231dbd01fSIzik Eidus * @page: the page that we are searching identical page to. 209331dbd01fSIzik Eidus * @rmap_item: the reverse mapping into the virtual address of this page 209431dbd01fSIzik Eidus */ 209521fbd591SQi Zheng static void cmp_and_merge_page(struct page *page, struct ksm_rmap_item *rmap_item) 209631dbd01fSIzik Eidus { 20974b22927fSKirill Tkhai struct mm_struct *mm = rmap_item->mm; 209821fbd591SQi Zheng struct ksm_rmap_item *tree_rmap_item; 20998dd3557aSHugh Dickins struct page *tree_page = NULL; 210021fbd591SQi Zheng struct ksm_stable_node *stable_node; 21018dd3557aSHugh Dickins struct page *kpage; 210231dbd01fSIzik Eidus unsigned int checksum; 210331dbd01fSIzik Eidus int err; 21042c653d0eSAndrea Arcangeli bool max_page_sharing_bypass = false; 210531dbd01fSIzik Eidus 21064146d2d6SHugh Dickins stable_node = page_stable_node(page); 21074146d2d6SHugh Dickins if (stable_node) { 21084146d2d6SHugh Dickins if (stable_node->head != &migrate_nodes && 21092c653d0eSAndrea Arcangeli get_kpfn_nid(READ_ONCE(stable_node->kpfn)) != 21102c653d0eSAndrea Arcangeli NUMA(stable_node->nid)) { 21112c653d0eSAndrea Arcangeli stable_node_dup_del(stable_node); 21124146d2d6SHugh Dickins stable_node->head = &migrate_nodes; 21134146d2d6SHugh Dickins list_add(&stable_node->list, stable_node->head); 21144146d2d6SHugh Dickins } 21154146d2d6SHugh Dickins if (stable_node->head != &migrate_nodes && 21164146d2d6SHugh Dickins rmap_item->head == stable_node) 21174146d2d6SHugh Dickins return; 21182c653d0eSAndrea Arcangeli /* 21192c653d0eSAndrea Arcangeli * If it's a KSM fork, allow it to go over the sharing limit 21202c653d0eSAndrea Arcangeli * without warnings. 21212c653d0eSAndrea Arcangeli */ 21222c653d0eSAndrea Arcangeli if (!is_page_sharing_candidate(stable_node)) 21232c653d0eSAndrea Arcangeli max_page_sharing_bypass = true; 21244146d2d6SHugh Dickins } 212531dbd01fSIzik Eidus 212631dbd01fSIzik Eidus /* We first start with searching the page inside the stable tree */ 212762b61f61SHugh Dickins kpage = stable_tree_search(page); 21284146d2d6SHugh Dickins if (kpage == page && rmap_item->head == stable_node) { 21294146d2d6SHugh Dickins put_page(kpage); 21304146d2d6SHugh Dickins return; 21314146d2d6SHugh Dickins } 21324146d2d6SHugh Dickins 21334146d2d6SHugh Dickins remove_rmap_item_from_tree(rmap_item); 21344146d2d6SHugh Dickins 213562b61f61SHugh Dickins if (kpage) { 21362cee57d1SYang Shi if (PTR_ERR(kpage) == -EBUSY) 21372cee57d1SYang Shi return; 21382cee57d1SYang Shi 213908beca44SHugh Dickins err = try_to_merge_with_ksm_page(rmap_item, page, kpage); 214031dbd01fSIzik Eidus if (!err) { 214131dbd01fSIzik Eidus /* 214231dbd01fSIzik Eidus * The page was successfully merged: 214331dbd01fSIzik Eidus * add its rmap_item to the stable tree. 214431dbd01fSIzik Eidus */ 21455ad64688SHugh Dickins lock_page(kpage); 21462c653d0eSAndrea Arcangeli stable_tree_append(rmap_item, page_stable_node(kpage), 21472c653d0eSAndrea Arcangeli max_page_sharing_bypass); 21485ad64688SHugh Dickins unlock_page(kpage); 214931dbd01fSIzik Eidus } 21508dd3557aSHugh Dickins put_page(kpage); 215131dbd01fSIzik Eidus return; 215231dbd01fSIzik Eidus } 215331dbd01fSIzik Eidus 215431dbd01fSIzik Eidus /* 21554035c07aSHugh Dickins * If the hash value of the page has changed from the last time 21564035c07aSHugh Dickins * we calculated it, this page is changing frequently: therefore we 21574035c07aSHugh Dickins * don't want to insert it in the unstable tree, and we don't want 21584035c07aSHugh Dickins * to waste our time searching for something identical to it there. 215931dbd01fSIzik Eidus */ 216031dbd01fSIzik Eidus checksum = calc_checksum(page); 216131dbd01fSIzik Eidus if (rmap_item->oldchecksum != checksum) { 216231dbd01fSIzik Eidus rmap_item->oldchecksum = checksum; 216331dbd01fSIzik Eidus return; 216431dbd01fSIzik Eidus } 216531dbd01fSIzik Eidus 2166e86c59b1SClaudio Imbrenda /* 2167e86c59b1SClaudio Imbrenda * Same checksum as an empty page. We attempt to merge it with the 2168e86c59b1SClaudio Imbrenda * appropriate zero page if the user enabled this via sysfs. 2169e86c59b1SClaudio Imbrenda */ 2170e86c59b1SClaudio Imbrenda if (ksm_use_zero_pages && (checksum == zero_checksum)) { 2171e86c59b1SClaudio Imbrenda struct vm_area_struct *vma; 2172e86c59b1SClaudio Imbrenda 2173d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 21744b22927fSKirill Tkhai vma = find_mergeable_vma(mm, rmap_item->address); 217556df70a6SMuchun Song if (vma) { 2176e86c59b1SClaudio Imbrenda err = try_to_merge_one_page(vma, page, 2177e86c59b1SClaudio Imbrenda ZERO_PAGE(rmap_item->address)); 2178739100c8SStefan Roesch trace_ksm_merge_one_page( 2179739100c8SStefan Roesch page_to_pfn(ZERO_PAGE(rmap_item->address)), 2180739100c8SStefan Roesch rmap_item, mm, err); 218156df70a6SMuchun Song } else { 218256df70a6SMuchun Song /* 218356df70a6SMuchun Song * If the vma is out of date, we do not need to 218456df70a6SMuchun Song * continue. 218556df70a6SMuchun Song */ 218656df70a6SMuchun Song err = 0; 218756df70a6SMuchun Song } 2188d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 2189e86c59b1SClaudio Imbrenda /* 2190e86c59b1SClaudio Imbrenda * In case of failure, the page was not really empty, so we 2191e86c59b1SClaudio Imbrenda * need to continue. Otherwise we're done. 2192e86c59b1SClaudio Imbrenda */ 2193e86c59b1SClaudio Imbrenda if (!err) 2194e86c59b1SClaudio Imbrenda return; 2195e86c59b1SClaudio Imbrenda } 21968dd3557aSHugh Dickins tree_rmap_item = 21978dd3557aSHugh Dickins unstable_tree_search_insert(rmap_item, page, &tree_page); 219831dbd01fSIzik Eidus if (tree_rmap_item) { 219977da2ba0SClaudio Imbrenda bool split; 220077da2ba0SClaudio Imbrenda 22018dd3557aSHugh Dickins kpage = try_to_merge_two_pages(rmap_item, page, 22028dd3557aSHugh Dickins tree_rmap_item, tree_page); 220377da2ba0SClaudio Imbrenda /* 220477da2ba0SClaudio Imbrenda * If both pages we tried to merge belong to the same compound 220577da2ba0SClaudio Imbrenda * page, then we actually ended up increasing the reference 220677da2ba0SClaudio Imbrenda * count of the same compound page twice, and split_huge_page 220777da2ba0SClaudio Imbrenda * failed. 220877da2ba0SClaudio Imbrenda * Here we set a flag if that happened, and we use it later to 220977da2ba0SClaudio Imbrenda * try split_huge_page again. Since we call put_page right 221077da2ba0SClaudio Imbrenda * afterwards, the reference count will be correct and 221177da2ba0SClaudio Imbrenda * split_huge_page should succeed. 221277da2ba0SClaudio Imbrenda */ 221377da2ba0SClaudio Imbrenda split = PageTransCompound(page) 221477da2ba0SClaudio Imbrenda && compound_head(page) == compound_head(tree_page); 22158dd3557aSHugh Dickins put_page(tree_page); 22168dd3557aSHugh Dickins if (kpage) { 2217bc56620bSHugh Dickins /* 2218bc56620bSHugh Dickins * The pages were successfully merged: insert new 2219bc56620bSHugh Dickins * node in the stable tree and add both rmap_items. 2220bc56620bSHugh Dickins */ 22215ad64688SHugh Dickins lock_page(kpage); 22227b6ba2c7SHugh Dickins stable_node = stable_tree_insert(kpage); 22237b6ba2c7SHugh Dickins if (stable_node) { 22242c653d0eSAndrea Arcangeli stable_tree_append(tree_rmap_item, stable_node, 22252c653d0eSAndrea Arcangeli false); 22262c653d0eSAndrea Arcangeli stable_tree_append(rmap_item, stable_node, 22272c653d0eSAndrea Arcangeli false); 22287b6ba2c7SHugh Dickins } 22295ad64688SHugh Dickins unlock_page(kpage); 22307b6ba2c7SHugh Dickins 223131dbd01fSIzik Eidus /* 223231dbd01fSIzik Eidus * If we fail to insert the page into the stable tree, 223331dbd01fSIzik Eidus * we will have 2 virtual addresses that are pointing 223431dbd01fSIzik Eidus * to a ksm page left outside the stable tree, 223531dbd01fSIzik Eidus * in which case we need to break_cow on both. 223631dbd01fSIzik Eidus */ 22377b6ba2c7SHugh Dickins if (!stable_node) { 22388dd3557aSHugh Dickins break_cow(tree_rmap_item); 22398dd3557aSHugh Dickins break_cow(rmap_item); 224031dbd01fSIzik Eidus } 224177da2ba0SClaudio Imbrenda } else if (split) { 224277da2ba0SClaudio Imbrenda /* 224377da2ba0SClaudio Imbrenda * We are here if we tried to merge two pages and 224477da2ba0SClaudio Imbrenda * failed because they both belonged to the same 224577da2ba0SClaudio Imbrenda * compound page. We will split the page now, but no 224677da2ba0SClaudio Imbrenda * merging will take place. 224777da2ba0SClaudio Imbrenda * We do not want to add the cost of a full lock; if 224877da2ba0SClaudio Imbrenda * the page is locked, it is better to skip it and 224977da2ba0SClaudio Imbrenda * perhaps try again later. 225077da2ba0SClaudio Imbrenda */ 225177da2ba0SClaudio Imbrenda if (!trylock_page(page)) 225277da2ba0SClaudio Imbrenda return; 225377da2ba0SClaudio Imbrenda split_huge_page(page); 225477da2ba0SClaudio Imbrenda unlock_page(page); 225531dbd01fSIzik Eidus } 225631dbd01fSIzik Eidus } 225731dbd01fSIzik Eidus } 225831dbd01fSIzik Eidus 225921fbd591SQi Zheng static struct ksm_rmap_item *get_next_rmap_item(struct ksm_mm_slot *mm_slot, 226021fbd591SQi Zheng struct ksm_rmap_item **rmap_list, 226131dbd01fSIzik Eidus unsigned long addr) 226231dbd01fSIzik Eidus { 226321fbd591SQi Zheng struct ksm_rmap_item *rmap_item; 226431dbd01fSIzik Eidus 22656514d511SHugh Dickins while (*rmap_list) { 22666514d511SHugh Dickins rmap_item = *rmap_list; 226793d17715SHugh Dickins if ((rmap_item->address & PAGE_MASK) == addr) 226831dbd01fSIzik Eidus return rmap_item; 226931dbd01fSIzik Eidus if (rmap_item->address > addr) 227031dbd01fSIzik Eidus break; 22716514d511SHugh Dickins *rmap_list = rmap_item->rmap_list; 227231dbd01fSIzik Eidus remove_rmap_item_from_tree(rmap_item); 227331dbd01fSIzik Eidus free_rmap_item(rmap_item); 227431dbd01fSIzik Eidus } 227531dbd01fSIzik Eidus 227631dbd01fSIzik Eidus rmap_item = alloc_rmap_item(); 227731dbd01fSIzik Eidus if (rmap_item) { 227831dbd01fSIzik Eidus /* It has already been zeroed */ 227958730ab6SQi Zheng rmap_item->mm = mm_slot->slot.mm; 2280cb4df4caSxu xin rmap_item->mm->ksm_rmap_items++; 228131dbd01fSIzik Eidus rmap_item->address = addr; 22826514d511SHugh Dickins rmap_item->rmap_list = *rmap_list; 22836514d511SHugh Dickins *rmap_list = rmap_item; 228431dbd01fSIzik Eidus } 228531dbd01fSIzik Eidus return rmap_item; 228631dbd01fSIzik Eidus } 228731dbd01fSIzik Eidus 228821fbd591SQi Zheng static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page) 228931dbd01fSIzik Eidus { 229031dbd01fSIzik Eidus struct mm_struct *mm; 229158730ab6SQi Zheng struct ksm_mm_slot *mm_slot; 229258730ab6SQi Zheng struct mm_slot *slot; 229331dbd01fSIzik Eidus struct vm_area_struct *vma; 229421fbd591SQi Zheng struct ksm_rmap_item *rmap_item; 2295a5f18ba0SMatthew Wilcox (Oracle) struct vma_iterator vmi; 229690bd6fd3SPetr Holasek int nid; 229731dbd01fSIzik Eidus 229858730ab6SQi Zheng if (list_empty(&ksm_mm_head.slot.mm_node)) 229931dbd01fSIzik Eidus return NULL; 230031dbd01fSIzik Eidus 230158730ab6SQi Zheng mm_slot = ksm_scan.mm_slot; 230258730ab6SQi Zheng if (mm_slot == &ksm_mm_head) { 2303739100c8SStefan Roesch trace_ksm_start_scan(ksm_scan.seqnr, ksm_rmap_items); 2304739100c8SStefan Roesch 23052919bfd0SHugh Dickins /* 23062919bfd0SHugh Dickins * A number of pages can hang around indefinitely on per-cpu 23072919bfd0SHugh Dickins * pagevecs, raised page count preventing write_protect_page 23082919bfd0SHugh Dickins * from merging them. Though it doesn't really matter much, 23092919bfd0SHugh Dickins * it is puzzling to see some stuck in pages_volatile until 23102919bfd0SHugh Dickins * other activity jostles them out, and they also prevented 23112919bfd0SHugh Dickins * LTP's KSM test from succeeding deterministically; so drain 23122919bfd0SHugh Dickins * them here (here rather than on entry to ksm_do_scan(), 23132919bfd0SHugh Dickins * so we don't IPI too often when pages_to_scan is set low). 23142919bfd0SHugh Dickins */ 23152919bfd0SHugh Dickins lru_add_drain_all(); 23162919bfd0SHugh Dickins 23174146d2d6SHugh Dickins /* 23184146d2d6SHugh Dickins * Whereas stale stable_nodes on the stable_tree itself 23194146d2d6SHugh Dickins * get pruned in the regular course of stable_tree_search(), 23204146d2d6SHugh Dickins * those moved out to the migrate_nodes list can accumulate: 23214146d2d6SHugh Dickins * so prune them once before each full scan. 23224146d2d6SHugh Dickins */ 23234146d2d6SHugh Dickins if (!ksm_merge_across_nodes) { 232421fbd591SQi Zheng struct ksm_stable_node *stable_node, *next; 23254146d2d6SHugh Dickins struct page *page; 23264146d2d6SHugh Dickins 232703640418SGeliang Tang list_for_each_entry_safe(stable_node, next, 232803640418SGeliang Tang &migrate_nodes, list) { 23292cee57d1SYang Shi page = get_ksm_page(stable_node, 23302cee57d1SYang Shi GET_KSM_PAGE_NOLOCK); 23314146d2d6SHugh Dickins if (page) 23324146d2d6SHugh Dickins put_page(page); 23334146d2d6SHugh Dickins cond_resched(); 23344146d2d6SHugh Dickins } 23354146d2d6SHugh Dickins } 23364146d2d6SHugh Dickins 2337ef53d16cSHugh Dickins for (nid = 0; nid < ksm_nr_node_ids; nid++) 233890bd6fd3SPetr Holasek root_unstable_tree[nid] = RB_ROOT; 233931dbd01fSIzik Eidus 234031dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 234158730ab6SQi Zheng slot = list_entry(mm_slot->slot.mm_node.next, 234258730ab6SQi Zheng struct mm_slot, mm_node); 234358730ab6SQi Zheng mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); 234458730ab6SQi Zheng ksm_scan.mm_slot = mm_slot; 234531dbd01fSIzik Eidus spin_unlock(&ksm_mmlist_lock); 23462b472611SHugh Dickins /* 23472b472611SHugh Dickins * Although we tested list_empty() above, a racing __ksm_exit 23482b472611SHugh Dickins * of the last mm on the list may have removed it since then. 23492b472611SHugh Dickins */ 235058730ab6SQi Zheng if (mm_slot == &ksm_mm_head) 23512b472611SHugh Dickins return NULL; 235231dbd01fSIzik Eidus next_mm: 235331dbd01fSIzik Eidus ksm_scan.address = 0; 235458730ab6SQi Zheng ksm_scan.rmap_list = &mm_slot->rmap_list; 235531dbd01fSIzik Eidus } 235631dbd01fSIzik Eidus 235758730ab6SQi Zheng slot = &mm_slot->slot; 235831dbd01fSIzik Eidus mm = slot->mm; 2359a5f18ba0SMatthew Wilcox (Oracle) vma_iter_init(&vmi, mm, ksm_scan.address); 2360a5f18ba0SMatthew Wilcox (Oracle) 2361d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 23629ba69294SHugh Dickins if (ksm_test_exit(mm)) 2363a5f18ba0SMatthew Wilcox (Oracle) goto no_vmas; 23649ba69294SHugh Dickins 2365a5f18ba0SMatthew Wilcox (Oracle) for_each_vma(vmi, vma) { 236631dbd01fSIzik Eidus if (!(vma->vm_flags & VM_MERGEABLE)) 236731dbd01fSIzik Eidus continue; 236831dbd01fSIzik Eidus if (ksm_scan.address < vma->vm_start) 236931dbd01fSIzik Eidus ksm_scan.address = vma->vm_start; 237031dbd01fSIzik Eidus if (!vma->anon_vma) 237131dbd01fSIzik Eidus ksm_scan.address = vma->vm_end; 237231dbd01fSIzik Eidus 237331dbd01fSIzik Eidus while (ksm_scan.address < vma->vm_end) { 23749ba69294SHugh Dickins if (ksm_test_exit(mm)) 23759ba69294SHugh Dickins break; 237631dbd01fSIzik Eidus *page = follow_page(vma, ksm_scan.address, FOLL_GET); 2377f7091ed6SHaiyue Wang if (IS_ERR_OR_NULL(*page)) { 237821ae5b01SAndrea Arcangeli ksm_scan.address += PAGE_SIZE; 237921ae5b01SAndrea Arcangeli cond_resched(); 238021ae5b01SAndrea Arcangeli continue; 238121ae5b01SAndrea Arcangeli } 2382f7091ed6SHaiyue Wang if (is_zone_device_page(*page)) 2383f7091ed6SHaiyue Wang goto next_page; 2384f765f540SKirill A. Shutemov if (PageAnon(*page)) { 238531dbd01fSIzik Eidus flush_anon_page(vma, *page, ksm_scan.address); 238631dbd01fSIzik Eidus flush_dcache_page(*page); 238758730ab6SQi Zheng rmap_item = get_next_rmap_item(mm_slot, 23886514d511SHugh Dickins ksm_scan.rmap_list, ksm_scan.address); 238931dbd01fSIzik Eidus if (rmap_item) { 23906514d511SHugh Dickins ksm_scan.rmap_list = 23916514d511SHugh Dickins &rmap_item->rmap_list; 239231dbd01fSIzik Eidus ksm_scan.address += PAGE_SIZE; 239331dbd01fSIzik Eidus } else 239431dbd01fSIzik Eidus put_page(*page); 2395d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 239631dbd01fSIzik Eidus return rmap_item; 239731dbd01fSIzik Eidus } 2398f7091ed6SHaiyue Wang next_page: 239931dbd01fSIzik Eidus put_page(*page); 240031dbd01fSIzik Eidus ksm_scan.address += PAGE_SIZE; 240131dbd01fSIzik Eidus cond_resched(); 240231dbd01fSIzik Eidus } 240331dbd01fSIzik Eidus } 240431dbd01fSIzik Eidus 24059ba69294SHugh Dickins if (ksm_test_exit(mm)) { 2406a5f18ba0SMatthew Wilcox (Oracle) no_vmas: 24079ba69294SHugh Dickins ksm_scan.address = 0; 240858730ab6SQi Zheng ksm_scan.rmap_list = &mm_slot->rmap_list; 24099ba69294SHugh Dickins } 241031dbd01fSIzik Eidus /* 241131dbd01fSIzik Eidus * Nuke all the rmap_items that are above this current rmap: 241231dbd01fSIzik Eidus * because there were no VM_MERGEABLE vmas with such addresses. 241331dbd01fSIzik Eidus */ 2414420be4edSChengyang Fan remove_trailing_rmap_items(ksm_scan.rmap_list); 241531dbd01fSIzik Eidus 241631dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 241758730ab6SQi Zheng slot = list_entry(mm_slot->slot.mm_node.next, 241858730ab6SQi Zheng struct mm_slot, mm_node); 241958730ab6SQi Zheng ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); 2420cd551f97SHugh Dickins if (ksm_scan.address == 0) { 2421cd551f97SHugh Dickins /* 2422c1e8d7c6SMichel Lespinasse * We've completed a full scan of all vmas, holding mmap_lock 2423cd551f97SHugh Dickins * throughout, and found no VM_MERGEABLE: so do the same as 2424cd551f97SHugh Dickins * __ksm_exit does to remove this mm from all our lists now. 24259ba69294SHugh Dickins * This applies either when cleaning up after __ksm_exit 24269ba69294SHugh Dickins * (but beware: we can reach here even before __ksm_exit), 24279ba69294SHugh Dickins * or when all VM_MERGEABLE areas have been unmapped (and 2428c1e8d7c6SMichel Lespinasse * mmap_lock then protects against race with MADV_MERGEABLE). 2429cd551f97SHugh Dickins */ 243058730ab6SQi Zheng hash_del(&mm_slot->slot.hash); 243158730ab6SQi Zheng list_del(&mm_slot->slot.mm_node); 24329ba69294SHugh Dickins spin_unlock(&ksm_mmlist_lock); 24339ba69294SHugh Dickins 243458730ab6SQi Zheng mm_slot_free(mm_slot_cache, mm_slot); 2435cd551f97SHugh Dickins clear_bit(MMF_VM_MERGEABLE, &mm->flags); 2436d7597f59SStefan Roesch clear_bit(MMF_VM_MERGE_ANY, &mm->flags); 2437d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 24389ba69294SHugh Dickins mmdrop(mm); 24399ba69294SHugh Dickins } else { 2440d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 24417496fea9SZhou Chengming /* 24423e4e28c5SMichel Lespinasse * mmap_read_unlock(mm) first because after 24437496fea9SZhou Chengming * spin_unlock(&ksm_mmlist_lock) run, the "mm" may 24447496fea9SZhou Chengming * already have been freed under us by __ksm_exit() 24457496fea9SZhou Chengming * because the "mm_slot" is still hashed and 24467496fea9SZhou Chengming * ksm_scan.mm_slot doesn't point to it anymore. 24477496fea9SZhou Chengming */ 24487496fea9SZhou Chengming spin_unlock(&ksm_mmlist_lock); 24499ba69294SHugh Dickins } 245031dbd01fSIzik Eidus 245131dbd01fSIzik Eidus /* Repeat until we've completed scanning the whole list */ 245258730ab6SQi Zheng mm_slot = ksm_scan.mm_slot; 245358730ab6SQi Zheng if (mm_slot != &ksm_mm_head) 245431dbd01fSIzik Eidus goto next_mm; 245531dbd01fSIzik Eidus 2456739100c8SStefan Roesch trace_ksm_stop_scan(ksm_scan.seqnr, ksm_rmap_items); 245731dbd01fSIzik Eidus ksm_scan.seqnr++; 245831dbd01fSIzik Eidus return NULL; 245931dbd01fSIzik Eidus } 246031dbd01fSIzik Eidus 246131dbd01fSIzik Eidus /** 246231dbd01fSIzik Eidus * ksm_do_scan - the ksm scanner main worker function. 2463b7701a5fSMike Rapoport * @scan_npages: number of pages we want to scan before we return. 246431dbd01fSIzik Eidus */ 246531dbd01fSIzik Eidus static void ksm_do_scan(unsigned int scan_npages) 246631dbd01fSIzik Eidus { 246721fbd591SQi Zheng struct ksm_rmap_item *rmap_item; 24683f649ab7SKees Cook struct page *page; 246931dbd01fSIzik Eidus 2470878aee7dSAndrea Arcangeli while (scan_npages-- && likely(!freezing(current))) { 247131dbd01fSIzik Eidus cond_resched(); 247231dbd01fSIzik Eidus rmap_item = scan_get_next_rmap_item(&page); 247331dbd01fSIzik Eidus if (!rmap_item) 247431dbd01fSIzik Eidus return; 247531dbd01fSIzik Eidus cmp_and_merge_page(page, rmap_item); 247631dbd01fSIzik Eidus put_page(page); 247731dbd01fSIzik Eidus } 247831dbd01fSIzik Eidus } 247931dbd01fSIzik Eidus 24806e158384SHugh Dickins static int ksmd_should_run(void) 24816e158384SHugh Dickins { 248258730ab6SQi Zheng return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.slot.mm_node); 24836e158384SHugh Dickins } 24846e158384SHugh Dickins 248531dbd01fSIzik Eidus static int ksm_scan_thread(void *nothing) 248631dbd01fSIzik Eidus { 2487fcf9a0efSKirill Tkhai unsigned int sleep_ms; 2488fcf9a0efSKirill Tkhai 2489878aee7dSAndrea Arcangeli set_freezable(); 2490339aa624SIzik Eidus set_user_nice(current, 5); 249131dbd01fSIzik Eidus 249231dbd01fSIzik Eidus while (!kthread_should_stop()) { 249331dbd01fSIzik Eidus mutex_lock(&ksm_thread_mutex); 2494ef4d43a8SHugh Dickins wait_while_offlining(); 24956e158384SHugh Dickins if (ksmd_should_run()) 249631dbd01fSIzik Eidus ksm_do_scan(ksm_thread_pages_to_scan); 249731dbd01fSIzik Eidus mutex_unlock(&ksm_thread_mutex); 24986e158384SHugh Dickins 2499878aee7dSAndrea Arcangeli try_to_freeze(); 2500878aee7dSAndrea Arcangeli 25016e158384SHugh Dickins if (ksmd_should_run()) { 2502fcf9a0efSKirill Tkhai sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs); 2503fcf9a0efSKirill Tkhai wait_event_interruptible_timeout(ksm_iter_wait, 2504fcf9a0efSKirill Tkhai sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs), 2505fcf9a0efSKirill Tkhai msecs_to_jiffies(sleep_ms)); 250631dbd01fSIzik Eidus } else { 2507878aee7dSAndrea Arcangeli wait_event_freezable(ksm_thread_wait, 25086e158384SHugh Dickins ksmd_should_run() || kthread_should_stop()); 250931dbd01fSIzik Eidus } 251031dbd01fSIzik Eidus } 251131dbd01fSIzik Eidus return 0; 251231dbd01fSIzik Eidus } 251331dbd01fSIzik Eidus 2514d7597f59SStefan Roesch static void __ksm_add_vma(struct vm_area_struct *vma) 2515d7597f59SStefan Roesch { 2516d7597f59SStefan Roesch unsigned long vm_flags = vma->vm_flags; 2517d7597f59SStefan Roesch 2518d7597f59SStefan Roesch if (vm_flags & VM_MERGEABLE) 2519d7597f59SStefan Roesch return; 2520d7597f59SStefan Roesch 2521d7597f59SStefan Roesch if (vma_ksm_compatible(vma)) 2522d7597f59SStefan Roesch vm_flags_set(vma, VM_MERGEABLE); 2523d7597f59SStefan Roesch } 2524d7597f59SStefan Roesch 252524139c07SDavid Hildenbrand static int __ksm_del_vma(struct vm_area_struct *vma) 252624139c07SDavid Hildenbrand { 252724139c07SDavid Hildenbrand int err; 252824139c07SDavid Hildenbrand 252924139c07SDavid Hildenbrand if (!(vma->vm_flags & VM_MERGEABLE)) 253024139c07SDavid Hildenbrand return 0; 253124139c07SDavid Hildenbrand 253224139c07SDavid Hildenbrand if (vma->anon_vma) { 253324139c07SDavid Hildenbrand err = unmerge_ksm_pages(vma, vma->vm_start, vma->vm_end); 253424139c07SDavid Hildenbrand if (err) 253524139c07SDavid Hildenbrand return err; 253624139c07SDavid Hildenbrand } 253724139c07SDavid Hildenbrand 253824139c07SDavid Hildenbrand vm_flags_clear(vma, VM_MERGEABLE); 253924139c07SDavid Hildenbrand return 0; 254024139c07SDavid Hildenbrand } 2541d7597f59SStefan Roesch /** 2542d7597f59SStefan Roesch * ksm_add_vma - Mark vma as mergeable if compatible 2543d7597f59SStefan Roesch * 2544d7597f59SStefan Roesch * @vma: Pointer to vma 2545d7597f59SStefan Roesch */ 2546d7597f59SStefan Roesch void ksm_add_vma(struct vm_area_struct *vma) 2547d7597f59SStefan Roesch { 2548d7597f59SStefan Roesch struct mm_struct *mm = vma->vm_mm; 2549d7597f59SStefan Roesch 2550d7597f59SStefan Roesch if (test_bit(MMF_VM_MERGE_ANY, &mm->flags)) 2551d7597f59SStefan Roesch __ksm_add_vma(vma); 2552d7597f59SStefan Roesch } 2553d7597f59SStefan Roesch 2554d7597f59SStefan Roesch static void ksm_add_vmas(struct mm_struct *mm) 2555d7597f59SStefan Roesch { 2556d7597f59SStefan Roesch struct vm_area_struct *vma; 2557d7597f59SStefan Roesch 2558d7597f59SStefan Roesch VMA_ITERATOR(vmi, mm, 0); 2559d7597f59SStefan Roesch for_each_vma(vmi, vma) 2560d7597f59SStefan Roesch __ksm_add_vma(vma); 2561d7597f59SStefan Roesch } 2562d7597f59SStefan Roesch 256324139c07SDavid Hildenbrand static int ksm_del_vmas(struct mm_struct *mm) 256424139c07SDavid Hildenbrand { 256524139c07SDavid Hildenbrand struct vm_area_struct *vma; 256624139c07SDavid Hildenbrand int err; 256724139c07SDavid Hildenbrand 256824139c07SDavid Hildenbrand VMA_ITERATOR(vmi, mm, 0); 256924139c07SDavid Hildenbrand for_each_vma(vmi, vma) { 257024139c07SDavid Hildenbrand err = __ksm_del_vma(vma); 257124139c07SDavid Hildenbrand if (err) 257224139c07SDavid Hildenbrand return err; 257324139c07SDavid Hildenbrand } 257424139c07SDavid Hildenbrand return 0; 257524139c07SDavid Hildenbrand } 257624139c07SDavid Hildenbrand 2577d7597f59SStefan Roesch /** 2578d7597f59SStefan Roesch * ksm_enable_merge_any - Add mm to mm ksm list and enable merging on all 2579d7597f59SStefan Roesch * compatible VMA's 2580d7597f59SStefan Roesch * 2581d7597f59SStefan Roesch * @mm: Pointer to mm 2582d7597f59SStefan Roesch * 2583d7597f59SStefan Roesch * Returns 0 on success, otherwise error code 2584d7597f59SStefan Roesch */ 2585d7597f59SStefan Roesch int ksm_enable_merge_any(struct mm_struct *mm) 2586d7597f59SStefan Roesch { 2587d7597f59SStefan Roesch int err; 2588d7597f59SStefan Roesch 2589d7597f59SStefan Roesch if (test_bit(MMF_VM_MERGE_ANY, &mm->flags)) 2590d7597f59SStefan Roesch return 0; 2591d7597f59SStefan Roesch 2592d7597f59SStefan Roesch if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { 2593d7597f59SStefan Roesch err = __ksm_enter(mm); 2594d7597f59SStefan Roesch if (err) 2595d7597f59SStefan Roesch return err; 2596d7597f59SStefan Roesch } 2597d7597f59SStefan Roesch 2598d7597f59SStefan Roesch set_bit(MMF_VM_MERGE_ANY, &mm->flags); 2599d7597f59SStefan Roesch ksm_add_vmas(mm); 2600d7597f59SStefan Roesch 2601d7597f59SStefan Roesch return 0; 2602d7597f59SStefan Roesch } 2603d7597f59SStefan Roesch 260424139c07SDavid Hildenbrand /** 260524139c07SDavid Hildenbrand * ksm_disable_merge_any - Disable merging on all compatible VMA's of the mm, 260624139c07SDavid Hildenbrand * previously enabled via ksm_enable_merge_any(). 260724139c07SDavid Hildenbrand * 260824139c07SDavid Hildenbrand * Disabling merging implies unmerging any merged pages, like setting 260924139c07SDavid Hildenbrand * MADV_UNMERGEABLE would. If unmerging fails, the whole operation fails and 261024139c07SDavid Hildenbrand * merging on all compatible VMA's remains enabled. 261124139c07SDavid Hildenbrand * 261224139c07SDavid Hildenbrand * @mm: Pointer to mm 261324139c07SDavid Hildenbrand * 261424139c07SDavid Hildenbrand * Returns 0 on success, otherwise error code 261524139c07SDavid Hildenbrand */ 261624139c07SDavid Hildenbrand int ksm_disable_merge_any(struct mm_struct *mm) 261724139c07SDavid Hildenbrand { 261824139c07SDavid Hildenbrand int err; 261924139c07SDavid Hildenbrand 262024139c07SDavid Hildenbrand if (!test_bit(MMF_VM_MERGE_ANY, &mm->flags)) 262124139c07SDavid Hildenbrand return 0; 262224139c07SDavid Hildenbrand 262324139c07SDavid Hildenbrand err = ksm_del_vmas(mm); 262424139c07SDavid Hildenbrand if (err) { 262524139c07SDavid Hildenbrand ksm_add_vmas(mm); 262624139c07SDavid Hildenbrand return err; 262724139c07SDavid Hildenbrand } 262824139c07SDavid Hildenbrand 262924139c07SDavid Hildenbrand clear_bit(MMF_VM_MERGE_ANY, &mm->flags); 263024139c07SDavid Hildenbrand return 0; 263124139c07SDavid Hildenbrand } 263224139c07SDavid Hildenbrand 26332c281f54SDavid Hildenbrand int ksm_disable(struct mm_struct *mm) 26342c281f54SDavid Hildenbrand { 26352c281f54SDavid Hildenbrand mmap_assert_write_locked(mm); 26362c281f54SDavid Hildenbrand 26372c281f54SDavid Hildenbrand if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) 26382c281f54SDavid Hildenbrand return 0; 26392c281f54SDavid Hildenbrand if (test_bit(MMF_VM_MERGE_ANY, &mm->flags)) 26402c281f54SDavid Hildenbrand return ksm_disable_merge_any(mm); 26412c281f54SDavid Hildenbrand return ksm_del_vmas(mm); 26422c281f54SDavid Hildenbrand } 26432c281f54SDavid Hildenbrand 2644f8af4da3SHugh Dickins int ksm_madvise(struct vm_area_struct *vma, unsigned long start, 2645f8af4da3SHugh Dickins unsigned long end, int advice, unsigned long *vm_flags) 2646f8af4da3SHugh Dickins { 2647f8af4da3SHugh Dickins struct mm_struct *mm = vma->vm_mm; 2648d952b791SHugh Dickins int err; 2649f8af4da3SHugh Dickins 2650f8af4da3SHugh Dickins switch (advice) { 2651f8af4da3SHugh Dickins case MADV_MERGEABLE: 2652d7597f59SStefan Roesch if (vma->vm_flags & VM_MERGEABLE) 2653e1fb4a08SDave Jiang return 0; 2654d7597f59SStefan Roesch if (!vma_ksm_compatible(vma)) 265512564485SShawn Anastasio return 0; 2656cc2383ecSKonstantin Khlebnikov 2657d952b791SHugh Dickins if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { 2658d952b791SHugh Dickins err = __ksm_enter(mm); 2659d952b791SHugh Dickins if (err) 2660d952b791SHugh Dickins return err; 2661d952b791SHugh Dickins } 2662f8af4da3SHugh Dickins 2663f8af4da3SHugh Dickins *vm_flags |= VM_MERGEABLE; 2664f8af4da3SHugh Dickins break; 2665f8af4da3SHugh Dickins 2666f8af4da3SHugh Dickins case MADV_UNMERGEABLE: 2667f8af4da3SHugh Dickins if (!(*vm_flags & VM_MERGEABLE)) 2668f8af4da3SHugh Dickins return 0; /* just ignore the advice */ 2669f8af4da3SHugh Dickins 2670d952b791SHugh Dickins if (vma->anon_vma) { 2671d952b791SHugh Dickins err = unmerge_ksm_pages(vma, start, end); 2672d952b791SHugh Dickins if (err) 2673d952b791SHugh Dickins return err; 2674d952b791SHugh Dickins } 2675f8af4da3SHugh Dickins 2676f8af4da3SHugh Dickins *vm_flags &= ~VM_MERGEABLE; 2677f8af4da3SHugh Dickins break; 2678f8af4da3SHugh Dickins } 2679f8af4da3SHugh Dickins 2680f8af4da3SHugh Dickins return 0; 2681f8af4da3SHugh Dickins } 268233cf1707SBharata B Rao EXPORT_SYMBOL_GPL(ksm_madvise); 2683f8af4da3SHugh Dickins 2684f8af4da3SHugh Dickins int __ksm_enter(struct mm_struct *mm) 2685f8af4da3SHugh Dickins { 268621fbd591SQi Zheng struct ksm_mm_slot *mm_slot; 268758730ab6SQi Zheng struct mm_slot *slot; 26886e158384SHugh Dickins int needs_wakeup; 26896e158384SHugh Dickins 269058730ab6SQi Zheng mm_slot = mm_slot_alloc(mm_slot_cache); 269131dbd01fSIzik Eidus if (!mm_slot) 269231dbd01fSIzik Eidus return -ENOMEM; 269331dbd01fSIzik Eidus 269458730ab6SQi Zheng slot = &mm_slot->slot; 269558730ab6SQi Zheng 26966e158384SHugh Dickins /* Check ksm_run too? Would need tighter locking */ 269758730ab6SQi Zheng needs_wakeup = list_empty(&ksm_mm_head.slot.mm_node); 26986e158384SHugh Dickins 269931dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 270058730ab6SQi Zheng mm_slot_insert(mm_slots_hash, mm, slot); 270131dbd01fSIzik Eidus /* 2702cbf86cfeSHugh Dickins * When KSM_RUN_MERGE (or KSM_RUN_STOP), 2703cbf86cfeSHugh Dickins * insert just behind the scanning cursor, to let the area settle 270431dbd01fSIzik Eidus * down a little; when fork is followed by immediate exec, we don't 270531dbd01fSIzik Eidus * want ksmd to waste time setting up and tearing down an rmap_list. 2706cbf86cfeSHugh Dickins * 2707cbf86cfeSHugh Dickins * But when KSM_RUN_UNMERGE, it's important to insert ahead of its 2708cbf86cfeSHugh Dickins * scanning cursor, otherwise KSM pages in newly forked mms will be 2709cbf86cfeSHugh Dickins * missed: then we might as well insert at the end of the list. 271031dbd01fSIzik Eidus */ 2711cbf86cfeSHugh Dickins if (ksm_run & KSM_RUN_UNMERGE) 271258730ab6SQi Zheng list_add_tail(&slot->mm_node, &ksm_mm_head.slot.mm_node); 2713cbf86cfeSHugh Dickins else 271458730ab6SQi Zheng list_add_tail(&slot->mm_node, &ksm_scan.mm_slot->slot.mm_node); 271531dbd01fSIzik Eidus spin_unlock(&ksm_mmlist_lock); 271631dbd01fSIzik Eidus 2717f8af4da3SHugh Dickins set_bit(MMF_VM_MERGEABLE, &mm->flags); 2718f1f10076SVegard Nossum mmgrab(mm); 27196e158384SHugh Dickins 27206e158384SHugh Dickins if (needs_wakeup) 27216e158384SHugh Dickins wake_up_interruptible(&ksm_thread_wait); 27226e158384SHugh Dickins 2723739100c8SStefan Roesch trace_ksm_enter(mm); 2724f8af4da3SHugh Dickins return 0; 2725f8af4da3SHugh Dickins } 2726f8af4da3SHugh Dickins 27271c2fb7a4SAndrea Arcangeli void __ksm_exit(struct mm_struct *mm) 2728f8af4da3SHugh Dickins { 272921fbd591SQi Zheng struct ksm_mm_slot *mm_slot; 273058730ab6SQi Zheng struct mm_slot *slot; 27319ba69294SHugh Dickins int easy_to_free = 0; 2732cd551f97SHugh Dickins 273331dbd01fSIzik Eidus /* 27349ba69294SHugh Dickins * This process is exiting: if it's straightforward (as is the 27359ba69294SHugh Dickins * case when ksmd was never running), free mm_slot immediately. 27369ba69294SHugh Dickins * But if it's at the cursor or has rmap_items linked to it, use 2737c1e8d7c6SMichel Lespinasse * mmap_lock to synchronize with any break_cows before pagetables 27389ba69294SHugh Dickins * are freed, and leave the mm_slot on the list for ksmd to free. 27399ba69294SHugh Dickins * Beware: ksm may already have noticed it exiting and freed the slot. 274031dbd01fSIzik Eidus */ 27419ba69294SHugh Dickins 2742cd551f97SHugh Dickins spin_lock(&ksm_mmlist_lock); 274358730ab6SQi Zheng slot = mm_slot_lookup(mm_slots_hash, mm); 274458730ab6SQi Zheng mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); 27459ba69294SHugh Dickins if (mm_slot && ksm_scan.mm_slot != mm_slot) { 27466514d511SHugh Dickins if (!mm_slot->rmap_list) { 274758730ab6SQi Zheng hash_del(&slot->hash); 274858730ab6SQi Zheng list_del(&slot->mm_node); 27499ba69294SHugh Dickins easy_to_free = 1; 27509ba69294SHugh Dickins } else { 275158730ab6SQi Zheng list_move(&slot->mm_node, 275258730ab6SQi Zheng &ksm_scan.mm_slot->slot.mm_node); 27539ba69294SHugh Dickins } 27549ba69294SHugh Dickins } 2755cd551f97SHugh Dickins spin_unlock(&ksm_mmlist_lock); 2756cd551f97SHugh Dickins 27579ba69294SHugh Dickins if (easy_to_free) { 275858730ab6SQi Zheng mm_slot_free(mm_slot_cache, mm_slot); 2759d7597f59SStefan Roesch clear_bit(MMF_VM_MERGE_ANY, &mm->flags); 2760cd551f97SHugh Dickins clear_bit(MMF_VM_MERGEABLE, &mm->flags); 27619ba69294SHugh Dickins mmdrop(mm); 27629ba69294SHugh Dickins } else if (mm_slot) { 2763d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 2764d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 27659ba69294SHugh Dickins } 2766739100c8SStefan Roesch 2767739100c8SStefan Roesch trace_ksm_exit(mm); 2768f8af4da3SHugh Dickins } 276931dbd01fSIzik Eidus 2770cbf86cfeSHugh Dickins struct page *ksm_might_need_to_copy(struct page *page, 27715ad64688SHugh Dickins struct vm_area_struct *vma, unsigned long address) 27725ad64688SHugh Dickins { 2773e05b3453SMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 2774e05b3453SMatthew Wilcox (Oracle) struct anon_vma *anon_vma = folio_anon_vma(folio); 27755ad64688SHugh Dickins struct page *new_page; 27765ad64688SHugh Dickins 2777cbf86cfeSHugh Dickins if (PageKsm(page)) { 2778cbf86cfeSHugh Dickins if (page_stable_node(page) && 2779cbf86cfeSHugh Dickins !(ksm_run & KSM_RUN_UNMERGE)) 2780cbf86cfeSHugh Dickins return page; /* no need to copy it */ 2781cbf86cfeSHugh Dickins } else if (!anon_vma) { 2782cbf86cfeSHugh Dickins return page; /* no need to copy it */ 2783e1c63e11SNanyong Sun } else if (page->index == linear_page_index(vma, address) && 2784e1c63e11SNanyong Sun anon_vma->root == vma->anon_vma->root) { 2785cbf86cfeSHugh Dickins return page; /* still no need to copy it */ 2786cbf86cfeSHugh Dickins } 2787cbf86cfeSHugh Dickins if (!PageUptodate(page)) 2788cbf86cfeSHugh Dickins return page; /* let do_swap_page report the error */ 2789cbf86cfeSHugh Dickins 27905ad64688SHugh Dickins new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); 27918f425e4eSMatthew Wilcox (Oracle) if (new_page && 27928f425e4eSMatthew Wilcox (Oracle) mem_cgroup_charge(page_folio(new_page), vma->vm_mm, GFP_KERNEL)) { 279362fdb163SHugh Dickins put_page(new_page); 279462fdb163SHugh Dickins new_page = NULL; 279562fdb163SHugh Dickins } 27965ad64688SHugh Dickins if (new_page) { 27976b970599SKefeng Wang if (copy_mc_user_highpage(new_page, page, address, vma)) { 27986b970599SKefeng Wang put_page(new_page); 27996b970599SKefeng Wang memory_failure_queue(page_to_pfn(page), 0); 28006b970599SKefeng Wang return ERR_PTR(-EHWPOISON); 28016b970599SKefeng Wang } 28025ad64688SHugh Dickins SetPageDirty(new_page); 28035ad64688SHugh Dickins __SetPageUptodate(new_page); 280448c935adSKirill A. Shutemov __SetPageLocked(new_page); 28054d45c3afSYang Yang #ifdef CONFIG_SWAP 28064d45c3afSYang Yang count_vm_event(KSM_SWPIN_COPY); 28074d45c3afSYang Yang #endif 28085ad64688SHugh Dickins } 28095ad64688SHugh Dickins 28105ad64688SHugh Dickins return new_page; 28115ad64688SHugh Dickins } 28125ad64688SHugh Dickins 28136d4675e6SMinchan Kim void rmap_walk_ksm(struct folio *folio, struct rmap_walk_control *rwc) 2814e9995ef9SHugh Dickins { 281521fbd591SQi Zheng struct ksm_stable_node *stable_node; 281621fbd591SQi Zheng struct ksm_rmap_item *rmap_item; 2817e9995ef9SHugh Dickins int search_new_forks = 0; 2818e9995ef9SHugh Dickins 28192f031c6fSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_ksm(folio), folio); 28209f32624bSJoonsoo Kim 28219f32624bSJoonsoo Kim /* 28229f32624bSJoonsoo Kim * Rely on the page lock to protect against concurrent modifications 28239f32624bSJoonsoo Kim * to that page's node of the stable tree. 28249f32624bSJoonsoo Kim */ 28252f031c6fSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 2826e9995ef9SHugh Dickins 28272f031c6fSMatthew Wilcox (Oracle) stable_node = folio_stable_node(folio); 2828e9995ef9SHugh Dickins if (!stable_node) 28291df631aeSMinchan Kim return; 2830e9995ef9SHugh Dickins again: 2831b67bfe0dSSasha Levin hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { 2832e9995ef9SHugh Dickins struct anon_vma *anon_vma = rmap_item->anon_vma; 28335beb4930SRik van Riel struct anon_vma_chain *vmac; 2834e9995ef9SHugh Dickins struct vm_area_struct *vma; 2835e9995ef9SHugh Dickins 2836ad12695fSAndrea Arcangeli cond_resched(); 28376d4675e6SMinchan Kim if (!anon_vma_trylock_read(anon_vma)) { 28386d4675e6SMinchan Kim if (rwc->try_lock) { 28396d4675e6SMinchan Kim rwc->contended = true; 28406d4675e6SMinchan Kim return; 28416d4675e6SMinchan Kim } 2842b6b19f25SHugh Dickins anon_vma_lock_read(anon_vma); 28436d4675e6SMinchan Kim } 2844bf181b9fSMichel Lespinasse anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, 2845bf181b9fSMichel Lespinasse 0, ULONG_MAX) { 28461105a2fcSJia He unsigned long addr; 28471105a2fcSJia He 2848ad12695fSAndrea Arcangeli cond_resched(); 28495beb4930SRik van Riel vma = vmac->vma; 28501105a2fcSJia He 28511105a2fcSJia He /* Ignore the stable/unstable/sqnr flags */ 2852cd7fae26SMiaohe Lin addr = rmap_item->address & PAGE_MASK; 28531105a2fcSJia He 28541105a2fcSJia He if (addr < vma->vm_start || addr >= vma->vm_end) 2855e9995ef9SHugh Dickins continue; 2856e9995ef9SHugh Dickins /* 2857e9995ef9SHugh Dickins * Initially we examine only the vma which covers this 2858e9995ef9SHugh Dickins * rmap_item; but later, if there is still work to do, 2859e9995ef9SHugh Dickins * we examine covering vmas in other mms: in case they 2860e9995ef9SHugh Dickins * were forked from the original since ksmd passed. 2861e9995ef9SHugh Dickins */ 2862e9995ef9SHugh Dickins if ((rmap_item->mm == vma->vm_mm) == search_new_forks) 2863e9995ef9SHugh Dickins continue; 2864e9995ef9SHugh Dickins 28650dd1c7bbSJoonsoo Kim if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) 28660dd1c7bbSJoonsoo Kim continue; 28670dd1c7bbSJoonsoo Kim 28682f031c6fSMatthew Wilcox (Oracle) if (!rwc->rmap_one(folio, vma, addr, rwc->arg)) { 2869b6b19f25SHugh Dickins anon_vma_unlock_read(anon_vma); 28701df631aeSMinchan Kim return; 2871e9995ef9SHugh Dickins } 28722f031c6fSMatthew Wilcox (Oracle) if (rwc->done && rwc->done(folio)) { 28730dd1c7bbSJoonsoo Kim anon_vma_unlock_read(anon_vma); 28741df631aeSMinchan Kim return; 28750dd1c7bbSJoonsoo Kim } 2876e9995ef9SHugh Dickins } 2877b6b19f25SHugh Dickins anon_vma_unlock_read(anon_vma); 2878e9995ef9SHugh Dickins } 2879e9995ef9SHugh Dickins if (!search_new_forks++) 2880e9995ef9SHugh Dickins goto again; 2881e9995ef9SHugh Dickins } 2882e9995ef9SHugh Dickins 28834248d008SLonglong Xia #ifdef CONFIG_MEMORY_FAILURE 28844248d008SLonglong Xia /* 28854248d008SLonglong Xia * Collect processes when the error hit an ksm page. 28864248d008SLonglong Xia */ 28874248d008SLonglong Xia void collect_procs_ksm(struct page *page, struct list_head *to_kill, 28884248d008SLonglong Xia int force_early) 28894248d008SLonglong Xia { 28904248d008SLonglong Xia struct ksm_stable_node *stable_node; 28914248d008SLonglong Xia struct ksm_rmap_item *rmap_item; 28924248d008SLonglong Xia struct folio *folio = page_folio(page); 28934248d008SLonglong Xia struct vm_area_struct *vma; 28944248d008SLonglong Xia struct task_struct *tsk; 28954248d008SLonglong Xia 28964248d008SLonglong Xia stable_node = folio_stable_node(folio); 28974248d008SLonglong Xia if (!stable_node) 28984248d008SLonglong Xia return; 28994248d008SLonglong Xia hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { 29004248d008SLonglong Xia struct anon_vma *av = rmap_item->anon_vma; 29014248d008SLonglong Xia 29024248d008SLonglong Xia anon_vma_lock_read(av); 29034248d008SLonglong Xia read_lock(&tasklist_lock); 29044248d008SLonglong Xia for_each_process(tsk) { 29054248d008SLonglong Xia struct anon_vma_chain *vmac; 29064248d008SLonglong Xia unsigned long addr; 29074248d008SLonglong Xia struct task_struct *t = 29084248d008SLonglong Xia task_early_kill(tsk, force_early); 29094248d008SLonglong Xia if (!t) 29104248d008SLonglong Xia continue; 29114248d008SLonglong Xia anon_vma_interval_tree_foreach(vmac, &av->rb_root, 0, 29124248d008SLonglong Xia ULONG_MAX) 29134248d008SLonglong Xia { 29144248d008SLonglong Xia vma = vmac->vma; 29154248d008SLonglong Xia if (vma->vm_mm == t->mm) { 29164248d008SLonglong Xia addr = rmap_item->address & PAGE_MASK; 29174248d008SLonglong Xia add_to_kill_ksm(t, page, vma, to_kill, 29184248d008SLonglong Xia addr); 29194248d008SLonglong Xia } 29204248d008SLonglong Xia } 29214248d008SLonglong Xia } 29224248d008SLonglong Xia read_unlock(&tasklist_lock); 29234248d008SLonglong Xia anon_vma_unlock_read(av); 29244248d008SLonglong Xia } 29254248d008SLonglong Xia } 29264248d008SLonglong Xia #endif 29274248d008SLonglong Xia 292852629506SJoonsoo Kim #ifdef CONFIG_MIGRATION 292919138349SMatthew Wilcox (Oracle) void folio_migrate_ksm(struct folio *newfolio, struct folio *folio) 2930e9995ef9SHugh Dickins { 293121fbd591SQi Zheng struct ksm_stable_node *stable_node; 2932e9995ef9SHugh Dickins 293319138349SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 293419138349SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(newfolio), newfolio); 293519138349SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(newfolio->mapping != folio->mapping, newfolio); 2936e9995ef9SHugh Dickins 293719138349SMatthew Wilcox (Oracle) stable_node = folio_stable_node(folio); 2938e9995ef9SHugh Dickins if (stable_node) { 293919138349SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(stable_node->kpfn != folio_pfn(folio), folio); 294019138349SMatthew Wilcox (Oracle) stable_node->kpfn = folio_pfn(newfolio); 2941c8d6553bSHugh Dickins /* 294219138349SMatthew Wilcox (Oracle) * newfolio->mapping was set in advance; now we need smp_wmb() 2943c8d6553bSHugh Dickins * to make sure that the new stable_node->kpfn is visible 294419138349SMatthew Wilcox (Oracle) * to get_ksm_page() before it can see that folio->mapping 294519138349SMatthew Wilcox (Oracle) * has gone stale (or that folio_test_swapcache has been cleared). 2946c8d6553bSHugh Dickins */ 2947c8d6553bSHugh Dickins smp_wmb(); 294819138349SMatthew Wilcox (Oracle) set_page_stable_node(&folio->page, NULL); 2949e9995ef9SHugh Dickins } 2950e9995ef9SHugh Dickins } 2951e9995ef9SHugh Dickins #endif /* CONFIG_MIGRATION */ 2952e9995ef9SHugh Dickins 295362b61f61SHugh Dickins #ifdef CONFIG_MEMORY_HOTREMOVE 2954ef4d43a8SHugh Dickins static void wait_while_offlining(void) 2955ef4d43a8SHugh Dickins { 2956ef4d43a8SHugh Dickins while (ksm_run & KSM_RUN_OFFLINE) { 2957ef4d43a8SHugh Dickins mutex_unlock(&ksm_thread_mutex); 2958ef4d43a8SHugh Dickins wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE), 295974316201SNeilBrown TASK_UNINTERRUPTIBLE); 2960ef4d43a8SHugh Dickins mutex_lock(&ksm_thread_mutex); 2961ef4d43a8SHugh Dickins } 2962ef4d43a8SHugh Dickins } 2963ef4d43a8SHugh Dickins 296421fbd591SQi Zheng static bool stable_node_dup_remove_range(struct ksm_stable_node *stable_node, 29652c653d0eSAndrea Arcangeli unsigned long start_pfn, 29662c653d0eSAndrea Arcangeli unsigned long end_pfn) 29672c653d0eSAndrea Arcangeli { 29682c653d0eSAndrea Arcangeli if (stable_node->kpfn >= start_pfn && 29692c653d0eSAndrea Arcangeli stable_node->kpfn < end_pfn) { 29702c653d0eSAndrea Arcangeli /* 29712c653d0eSAndrea Arcangeli * Don't get_ksm_page, page has already gone: 29722c653d0eSAndrea Arcangeli * which is why we keep kpfn instead of page* 29732c653d0eSAndrea Arcangeli */ 29742c653d0eSAndrea Arcangeli remove_node_from_stable_tree(stable_node); 29752c653d0eSAndrea Arcangeli return true; 29762c653d0eSAndrea Arcangeli } 29772c653d0eSAndrea Arcangeli return false; 29782c653d0eSAndrea Arcangeli } 29792c653d0eSAndrea Arcangeli 298021fbd591SQi Zheng static bool stable_node_chain_remove_range(struct ksm_stable_node *stable_node, 29812c653d0eSAndrea Arcangeli unsigned long start_pfn, 29822c653d0eSAndrea Arcangeli unsigned long end_pfn, 29832c653d0eSAndrea Arcangeli struct rb_root *root) 29842c653d0eSAndrea Arcangeli { 298521fbd591SQi Zheng struct ksm_stable_node *dup; 29862c653d0eSAndrea Arcangeli struct hlist_node *hlist_safe; 29872c653d0eSAndrea Arcangeli 29882c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 29892c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node)); 29902c653d0eSAndrea Arcangeli return stable_node_dup_remove_range(stable_node, start_pfn, 29912c653d0eSAndrea Arcangeli end_pfn); 29922c653d0eSAndrea Arcangeli } 29932c653d0eSAndrea Arcangeli 29942c653d0eSAndrea Arcangeli hlist_for_each_entry_safe(dup, hlist_safe, 29952c653d0eSAndrea Arcangeli &stable_node->hlist, hlist_dup) { 29962c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(dup)); 29972c653d0eSAndrea Arcangeli stable_node_dup_remove_range(dup, start_pfn, end_pfn); 29982c653d0eSAndrea Arcangeli } 29992c653d0eSAndrea Arcangeli if (hlist_empty(&stable_node->hlist)) { 30002c653d0eSAndrea Arcangeli free_stable_node_chain(stable_node, root); 30012c653d0eSAndrea Arcangeli return true; /* notify caller that tree was rebalanced */ 30022c653d0eSAndrea Arcangeli } else 30032c653d0eSAndrea Arcangeli return false; 30042c653d0eSAndrea Arcangeli } 30052c653d0eSAndrea Arcangeli 3006ee0ea59cSHugh Dickins static void ksm_check_stable_tree(unsigned long start_pfn, 300762b61f61SHugh Dickins unsigned long end_pfn) 300862b61f61SHugh Dickins { 300921fbd591SQi Zheng struct ksm_stable_node *stable_node, *next; 301062b61f61SHugh Dickins struct rb_node *node; 301190bd6fd3SPetr Holasek int nid; 301262b61f61SHugh Dickins 3013ef53d16cSHugh Dickins for (nid = 0; nid < ksm_nr_node_ids; nid++) { 3014ef53d16cSHugh Dickins node = rb_first(root_stable_tree + nid); 3015ee0ea59cSHugh Dickins while (node) { 301621fbd591SQi Zheng stable_node = rb_entry(node, struct ksm_stable_node, node); 30172c653d0eSAndrea Arcangeli if (stable_node_chain_remove_range(stable_node, 30182c653d0eSAndrea Arcangeli start_pfn, end_pfn, 30192c653d0eSAndrea Arcangeli root_stable_tree + 30202c653d0eSAndrea Arcangeli nid)) 3021ef53d16cSHugh Dickins node = rb_first(root_stable_tree + nid); 30222c653d0eSAndrea Arcangeli else 3023ee0ea59cSHugh Dickins node = rb_next(node); 3024ee0ea59cSHugh Dickins cond_resched(); 302562b61f61SHugh Dickins } 3026ee0ea59cSHugh Dickins } 302703640418SGeliang Tang list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { 30284146d2d6SHugh Dickins if (stable_node->kpfn >= start_pfn && 30294146d2d6SHugh Dickins stable_node->kpfn < end_pfn) 30304146d2d6SHugh Dickins remove_node_from_stable_tree(stable_node); 30314146d2d6SHugh Dickins cond_resched(); 30324146d2d6SHugh Dickins } 303362b61f61SHugh Dickins } 303462b61f61SHugh Dickins 303562b61f61SHugh Dickins static int ksm_memory_callback(struct notifier_block *self, 303662b61f61SHugh Dickins unsigned long action, void *arg) 303762b61f61SHugh Dickins { 303862b61f61SHugh Dickins struct memory_notify *mn = arg; 303962b61f61SHugh Dickins 304062b61f61SHugh Dickins switch (action) { 304162b61f61SHugh Dickins case MEM_GOING_OFFLINE: 304262b61f61SHugh Dickins /* 3043ef4d43a8SHugh Dickins * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items() 3044ef4d43a8SHugh Dickins * and remove_all_stable_nodes() while memory is going offline: 3045ef4d43a8SHugh Dickins * it is unsafe for them to touch the stable tree at this time. 3046ef4d43a8SHugh Dickins * But unmerge_ksm_pages(), rmap lookups and other entry points 3047ef4d43a8SHugh Dickins * which do not need the ksm_thread_mutex are all safe. 304862b61f61SHugh Dickins */ 3049ef4d43a8SHugh Dickins mutex_lock(&ksm_thread_mutex); 3050ef4d43a8SHugh Dickins ksm_run |= KSM_RUN_OFFLINE; 3051ef4d43a8SHugh Dickins mutex_unlock(&ksm_thread_mutex); 305262b61f61SHugh Dickins break; 305362b61f61SHugh Dickins 305462b61f61SHugh Dickins case MEM_OFFLINE: 305562b61f61SHugh Dickins /* 305662b61f61SHugh Dickins * Most of the work is done by page migration; but there might 305762b61f61SHugh Dickins * be a few stable_nodes left over, still pointing to struct 3058ee0ea59cSHugh Dickins * pages which have been offlined: prune those from the tree, 3059ee0ea59cSHugh Dickins * otherwise get_ksm_page() might later try to access a 3060ee0ea59cSHugh Dickins * non-existent struct page. 306162b61f61SHugh Dickins */ 3062ee0ea59cSHugh Dickins ksm_check_stable_tree(mn->start_pfn, 3063ee0ea59cSHugh Dickins mn->start_pfn + mn->nr_pages); 3064e4a9bc58SJoe Perches fallthrough; 306562b61f61SHugh Dickins case MEM_CANCEL_OFFLINE: 3066ef4d43a8SHugh Dickins mutex_lock(&ksm_thread_mutex); 3067ef4d43a8SHugh Dickins ksm_run &= ~KSM_RUN_OFFLINE; 306862b61f61SHugh Dickins mutex_unlock(&ksm_thread_mutex); 3069ef4d43a8SHugh Dickins 3070ef4d43a8SHugh Dickins smp_mb(); /* wake_up_bit advises this */ 3071ef4d43a8SHugh Dickins wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE)); 307262b61f61SHugh Dickins break; 307362b61f61SHugh Dickins } 307462b61f61SHugh Dickins return NOTIFY_OK; 307562b61f61SHugh Dickins } 3076ef4d43a8SHugh Dickins #else 3077ef4d43a8SHugh Dickins static void wait_while_offlining(void) 3078ef4d43a8SHugh Dickins { 3079ef4d43a8SHugh Dickins } 308062b61f61SHugh Dickins #endif /* CONFIG_MEMORY_HOTREMOVE */ 308162b61f61SHugh Dickins 3082d21077fbSStefan Roesch #ifdef CONFIG_PROC_FS 3083d21077fbSStefan Roesch long ksm_process_profit(struct mm_struct *mm) 3084d21077fbSStefan Roesch { 3085d21077fbSStefan Roesch return mm->ksm_merging_pages * PAGE_SIZE - 3086d21077fbSStefan Roesch mm->ksm_rmap_items * sizeof(struct ksm_rmap_item); 3087d21077fbSStefan Roesch } 3088d21077fbSStefan Roesch #endif /* CONFIG_PROC_FS */ 3089d21077fbSStefan Roesch 30902ffd8679SHugh Dickins #ifdef CONFIG_SYSFS 30912ffd8679SHugh Dickins /* 30922ffd8679SHugh Dickins * This all compiles without CONFIG_SYSFS, but is a waste of space. 30932ffd8679SHugh Dickins */ 30942ffd8679SHugh Dickins 309531dbd01fSIzik Eidus #define KSM_ATTR_RO(_name) \ 309631dbd01fSIzik Eidus static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 309731dbd01fSIzik Eidus #define KSM_ATTR(_name) \ 30981bad2e5cSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 309931dbd01fSIzik Eidus 310031dbd01fSIzik Eidus static ssize_t sleep_millisecs_show(struct kobject *kobj, 310131dbd01fSIzik Eidus struct kobj_attribute *attr, char *buf) 310231dbd01fSIzik Eidus { 3103ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs); 310431dbd01fSIzik Eidus } 310531dbd01fSIzik Eidus 310631dbd01fSIzik Eidus static ssize_t sleep_millisecs_store(struct kobject *kobj, 310731dbd01fSIzik Eidus struct kobj_attribute *attr, 310831dbd01fSIzik Eidus const char *buf, size_t count) 310931dbd01fSIzik Eidus { 3110dfefd226SAlexey Dobriyan unsigned int msecs; 311131dbd01fSIzik Eidus int err; 311231dbd01fSIzik Eidus 3113dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &msecs); 3114dfefd226SAlexey Dobriyan if (err) 311531dbd01fSIzik Eidus return -EINVAL; 311631dbd01fSIzik Eidus 311731dbd01fSIzik Eidus ksm_thread_sleep_millisecs = msecs; 3118fcf9a0efSKirill Tkhai wake_up_interruptible(&ksm_iter_wait); 311931dbd01fSIzik Eidus 312031dbd01fSIzik Eidus return count; 312131dbd01fSIzik Eidus } 312231dbd01fSIzik Eidus KSM_ATTR(sleep_millisecs); 312331dbd01fSIzik Eidus 312431dbd01fSIzik Eidus static ssize_t pages_to_scan_show(struct kobject *kobj, 312531dbd01fSIzik Eidus struct kobj_attribute *attr, char *buf) 312631dbd01fSIzik Eidus { 3127ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan); 312831dbd01fSIzik Eidus } 312931dbd01fSIzik Eidus 313031dbd01fSIzik Eidus static ssize_t pages_to_scan_store(struct kobject *kobj, 313131dbd01fSIzik Eidus struct kobj_attribute *attr, 313231dbd01fSIzik Eidus const char *buf, size_t count) 313331dbd01fSIzik Eidus { 3134dfefd226SAlexey Dobriyan unsigned int nr_pages; 313531dbd01fSIzik Eidus int err; 313631dbd01fSIzik Eidus 3137dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &nr_pages); 3138dfefd226SAlexey Dobriyan if (err) 313931dbd01fSIzik Eidus return -EINVAL; 314031dbd01fSIzik Eidus 314131dbd01fSIzik Eidus ksm_thread_pages_to_scan = nr_pages; 314231dbd01fSIzik Eidus 314331dbd01fSIzik Eidus return count; 314431dbd01fSIzik Eidus } 314531dbd01fSIzik Eidus KSM_ATTR(pages_to_scan); 314631dbd01fSIzik Eidus 314731dbd01fSIzik Eidus static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, 314831dbd01fSIzik Eidus char *buf) 314931dbd01fSIzik Eidus { 3150ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", ksm_run); 315131dbd01fSIzik Eidus } 315231dbd01fSIzik Eidus 315331dbd01fSIzik Eidus static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, 315431dbd01fSIzik Eidus const char *buf, size_t count) 315531dbd01fSIzik Eidus { 3156dfefd226SAlexey Dobriyan unsigned int flags; 315731dbd01fSIzik Eidus int err; 315831dbd01fSIzik Eidus 3159dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &flags); 3160dfefd226SAlexey Dobriyan if (err) 316131dbd01fSIzik Eidus return -EINVAL; 316231dbd01fSIzik Eidus if (flags > KSM_RUN_UNMERGE) 316331dbd01fSIzik Eidus return -EINVAL; 316431dbd01fSIzik Eidus 316531dbd01fSIzik Eidus /* 316631dbd01fSIzik Eidus * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. 316731dbd01fSIzik Eidus * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, 3168d0f209f6SHugh Dickins * breaking COW to free the pages_shared (but leaves mm_slots 3169d0f209f6SHugh Dickins * on the list for when ksmd may be set running again). 317031dbd01fSIzik Eidus */ 317131dbd01fSIzik Eidus 317231dbd01fSIzik Eidus mutex_lock(&ksm_thread_mutex); 3173ef4d43a8SHugh Dickins wait_while_offlining(); 317431dbd01fSIzik Eidus if (ksm_run != flags) { 317531dbd01fSIzik Eidus ksm_run = flags; 3176d952b791SHugh Dickins if (flags & KSM_RUN_UNMERGE) { 3177e1e12d2fSDavid Rientjes set_current_oom_origin(); 3178d952b791SHugh Dickins err = unmerge_and_remove_all_rmap_items(); 3179e1e12d2fSDavid Rientjes clear_current_oom_origin(); 3180d952b791SHugh Dickins if (err) { 3181d952b791SHugh Dickins ksm_run = KSM_RUN_STOP; 3182d952b791SHugh Dickins count = err; 3183d952b791SHugh Dickins } 3184d952b791SHugh Dickins } 318531dbd01fSIzik Eidus } 318631dbd01fSIzik Eidus mutex_unlock(&ksm_thread_mutex); 318731dbd01fSIzik Eidus 318831dbd01fSIzik Eidus if (flags & KSM_RUN_MERGE) 318931dbd01fSIzik Eidus wake_up_interruptible(&ksm_thread_wait); 319031dbd01fSIzik Eidus 319131dbd01fSIzik Eidus return count; 319231dbd01fSIzik Eidus } 319331dbd01fSIzik Eidus KSM_ATTR(run); 319431dbd01fSIzik Eidus 319590bd6fd3SPetr Holasek #ifdef CONFIG_NUMA 319690bd6fd3SPetr Holasek static ssize_t merge_across_nodes_show(struct kobject *kobj, 319790bd6fd3SPetr Holasek struct kobj_attribute *attr, char *buf) 319890bd6fd3SPetr Holasek { 3199ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes); 320090bd6fd3SPetr Holasek } 320190bd6fd3SPetr Holasek 320290bd6fd3SPetr Holasek static ssize_t merge_across_nodes_store(struct kobject *kobj, 320390bd6fd3SPetr Holasek struct kobj_attribute *attr, 320490bd6fd3SPetr Holasek const char *buf, size_t count) 320590bd6fd3SPetr Holasek { 320690bd6fd3SPetr Holasek int err; 320790bd6fd3SPetr Holasek unsigned long knob; 320890bd6fd3SPetr Holasek 320990bd6fd3SPetr Holasek err = kstrtoul(buf, 10, &knob); 321090bd6fd3SPetr Holasek if (err) 321190bd6fd3SPetr Holasek return err; 321290bd6fd3SPetr Holasek if (knob > 1) 321390bd6fd3SPetr Holasek return -EINVAL; 321490bd6fd3SPetr Holasek 321590bd6fd3SPetr Holasek mutex_lock(&ksm_thread_mutex); 3216ef4d43a8SHugh Dickins wait_while_offlining(); 321790bd6fd3SPetr Holasek if (ksm_merge_across_nodes != knob) { 3218cbf86cfeSHugh Dickins if (ksm_pages_shared || remove_all_stable_nodes()) 321990bd6fd3SPetr Holasek err = -EBUSY; 3220ef53d16cSHugh Dickins else if (root_stable_tree == one_stable_tree) { 3221ef53d16cSHugh Dickins struct rb_root *buf; 3222ef53d16cSHugh Dickins /* 3223ef53d16cSHugh Dickins * This is the first time that we switch away from the 3224ef53d16cSHugh Dickins * default of merging across nodes: must now allocate 3225ef53d16cSHugh Dickins * a buffer to hold as many roots as may be needed. 3226ef53d16cSHugh Dickins * Allocate stable and unstable together: 3227ef53d16cSHugh Dickins * MAXSMP NODES_SHIFT 10 will use 16kB. 3228ef53d16cSHugh Dickins */ 3229bafe1e14SJoe Perches buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf), 3230bafe1e14SJoe Perches GFP_KERNEL); 3231ef53d16cSHugh Dickins /* Let us assume that RB_ROOT is NULL is zero */ 3232ef53d16cSHugh Dickins if (!buf) 3233ef53d16cSHugh Dickins err = -ENOMEM; 3234ef53d16cSHugh Dickins else { 3235ef53d16cSHugh Dickins root_stable_tree = buf; 3236ef53d16cSHugh Dickins root_unstable_tree = buf + nr_node_ids; 3237ef53d16cSHugh Dickins /* Stable tree is empty but not the unstable */ 3238ef53d16cSHugh Dickins root_unstable_tree[0] = one_unstable_tree[0]; 3239ef53d16cSHugh Dickins } 3240ef53d16cSHugh Dickins } 3241ef53d16cSHugh Dickins if (!err) { 324290bd6fd3SPetr Holasek ksm_merge_across_nodes = knob; 3243ef53d16cSHugh Dickins ksm_nr_node_ids = knob ? 1 : nr_node_ids; 3244ef53d16cSHugh Dickins } 324590bd6fd3SPetr Holasek } 324690bd6fd3SPetr Holasek mutex_unlock(&ksm_thread_mutex); 324790bd6fd3SPetr Holasek 324890bd6fd3SPetr Holasek return err ? err : count; 324990bd6fd3SPetr Holasek } 325090bd6fd3SPetr Holasek KSM_ATTR(merge_across_nodes); 325190bd6fd3SPetr Holasek #endif 325290bd6fd3SPetr Holasek 3253e86c59b1SClaudio Imbrenda static ssize_t use_zero_pages_show(struct kobject *kobj, 3254e86c59b1SClaudio Imbrenda struct kobj_attribute *attr, char *buf) 3255e86c59b1SClaudio Imbrenda { 3256ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", ksm_use_zero_pages); 3257e86c59b1SClaudio Imbrenda } 3258e86c59b1SClaudio Imbrenda static ssize_t use_zero_pages_store(struct kobject *kobj, 3259e86c59b1SClaudio Imbrenda struct kobj_attribute *attr, 3260e86c59b1SClaudio Imbrenda const char *buf, size_t count) 3261e86c59b1SClaudio Imbrenda { 3262e86c59b1SClaudio Imbrenda int err; 3263e86c59b1SClaudio Imbrenda bool value; 3264e86c59b1SClaudio Imbrenda 3265e86c59b1SClaudio Imbrenda err = kstrtobool(buf, &value); 3266e86c59b1SClaudio Imbrenda if (err) 3267e86c59b1SClaudio Imbrenda return -EINVAL; 3268e86c59b1SClaudio Imbrenda 3269e86c59b1SClaudio Imbrenda ksm_use_zero_pages = value; 3270e86c59b1SClaudio Imbrenda 3271e86c59b1SClaudio Imbrenda return count; 3272e86c59b1SClaudio Imbrenda } 3273e86c59b1SClaudio Imbrenda KSM_ATTR(use_zero_pages); 3274e86c59b1SClaudio Imbrenda 32752c653d0eSAndrea Arcangeli static ssize_t max_page_sharing_show(struct kobject *kobj, 32762c653d0eSAndrea Arcangeli struct kobj_attribute *attr, char *buf) 32772c653d0eSAndrea Arcangeli { 3278ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", ksm_max_page_sharing); 32792c653d0eSAndrea Arcangeli } 32802c653d0eSAndrea Arcangeli 32812c653d0eSAndrea Arcangeli static ssize_t max_page_sharing_store(struct kobject *kobj, 32822c653d0eSAndrea Arcangeli struct kobj_attribute *attr, 32832c653d0eSAndrea Arcangeli const char *buf, size_t count) 32842c653d0eSAndrea Arcangeli { 32852c653d0eSAndrea Arcangeli int err; 32862c653d0eSAndrea Arcangeli int knob; 32872c653d0eSAndrea Arcangeli 32882c653d0eSAndrea Arcangeli err = kstrtoint(buf, 10, &knob); 32892c653d0eSAndrea Arcangeli if (err) 32902c653d0eSAndrea Arcangeli return err; 32912c653d0eSAndrea Arcangeli /* 32922c653d0eSAndrea Arcangeli * When a KSM page is created it is shared by 2 mappings. This 32932c653d0eSAndrea Arcangeli * being a signed comparison, it implicitly verifies it's not 32942c653d0eSAndrea Arcangeli * negative. 32952c653d0eSAndrea Arcangeli */ 32962c653d0eSAndrea Arcangeli if (knob < 2) 32972c653d0eSAndrea Arcangeli return -EINVAL; 32982c653d0eSAndrea Arcangeli 32992c653d0eSAndrea Arcangeli if (READ_ONCE(ksm_max_page_sharing) == knob) 33002c653d0eSAndrea Arcangeli return count; 33012c653d0eSAndrea Arcangeli 33022c653d0eSAndrea Arcangeli mutex_lock(&ksm_thread_mutex); 33032c653d0eSAndrea Arcangeli wait_while_offlining(); 33042c653d0eSAndrea Arcangeli if (ksm_max_page_sharing != knob) { 33052c653d0eSAndrea Arcangeli if (ksm_pages_shared || remove_all_stable_nodes()) 33062c653d0eSAndrea Arcangeli err = -EBUSY; 33072c653d0eSAndrea Arcangeli else 33082c653d0eSAndrea Arcangeli ksm_max_page_sharing = knob; 33092c653d0eSAndrea Arcangeli } 33102c653d0eSAndrea Arcangeli mutex_unlock(&ksm_thread_mutex); 33112c653d0eSAndrea Arcangeli 33122c653d0eSAndrea Arcangeli return err ? err : count; 33132c653d0eSAndrea Arcangeli } 33142c653d0eSAndrea Arcangeli KSM_ATTR(max_page_sharing); 33152c653d0eSAndrea Arcangeli 3316b4028260SHugh Dickins static ssize_t pages_shared_show(struct kobject *kobj, 3317b4028260SHugh Dickins struct kobj_attribute *attr, char *buf) 3318b4028260SHugh Dickins { 3319ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", ksm_pages_shared); 3320b4028260SHugh Dickins } 3321b4028260SHugh Dickins KSM_ATTR_RO(pages_shared); 3322b4028260SHugh Dickins 3323b4028260SHugh Dickins static ssize_t pages_sharing_show(struct kobject *kobj, 3324b4028260SHugh Dickins struct kobj_attribute *attr, char *buf) 3325b4028260SHugh Dickins { 3326ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", ksm_pages_sharing); 3327b4028260SHugh Dickins } 3328b4028260SHugh Dickins KSM_ATTR_RO(pages_sharing); 3329b4028260SHugh Dickins 3330473b0ce4SHugh Dickins static ssize_t pages_unshared_show(struct kobject *kobj, 3331473b0ce4SHugh Dickins struct kobj_attribute *attr, char *buf) 3332473b0ce4SHugh Dickins { 3333ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", ksm_pages_unshared); 3334473b0ce4SHugh Dickins } 3335473b0ce4SHugh Dickins KSM_ATTR_RO(pages_unshared); 3336473b0ce4SHugh Dickins 3337473b0ce4SHugh Dickins static ssize_t pages_volatile_show(struct kobject *kobj, 3338473b0ce4SHugh Dickins struct kobj_attribute *attr, char *buf) 3339473b0ce4SHugh Dickins { 3340473b0ce4SHugh Dickins long ksm_pages_volatile; 3341473b0ce4SHugh Dickins 3342473b0ce4SHugh Dickins ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared 3343473b0ce4SHugh Dickins - ksm_pages_sharing - ksm_pages_unshared; 3344473b0ce4SHugh Dickins /* 3345473b0ce4SHugh Dickins * It was not worth any locking to calculate that statistic, 3346473b0ce4SHugh Dickins * but it might therefore sometimes be negative: conceal that. 3347473b0ce4SHugh Dickins */ 3348473b0ce4SHugh Dickins if (ksm_pages_volatile < 0) 3349473b0ce4SHugh Dickins ksm_pages_volatile = 0; 3350ae7a927dSJoe Perches return sysfs_emit(buf, "%ld\n", ksm_pages_volatile); 3351473b0ce4SHugh Dickins } 3352473b0ce4SHugh Dickins KSM_ATTR_RO(pages_volatile); 3353473b0ce4SHugh Dickins 3354d21077fbSStefan Roesch static ssize_t general_profit_show(struct kobject *kobj, 3355d21077fbSStefan Roesch struct kobj_attribute *attr, char *buf) 3356d21077fbSStefan Roesch { 3357d21077fbSStefan Roesch long general_profit; 3358d21077fbSStefan Roesch 3359d21077fbSStefan Roesch general_profit = ksm_pages_sharing * PAGE_SIZE - 3360d21077fbSStefan Roesch ksm_rmap_items * sizeof(struct ksm_rmap_item); 3361d21077fbSStefan Roesch 3362d21077fbSStefan Roesch return sysfs_emit(buf, "%ld\n", general_profit); 3363d21077fbSStefan Roesch } 3364d21077fbSStefan Roesch KSM_ATTR_RO(general_profit); 3365d21077fbSStefan Roesch 33662c653d0eSAndrea Arcangeli static ssize_t stable_node_dups_show(struct kobject *kobj, 33672c653d0eSAndrea Arcangeli struct kobj_attribute *attr, char *buf) 33682c653d0eSAndrea Arcangeli { 3369ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups); 33702c653d0eSAndrea Arcangeli } 33712c653d0eSAndrea Arcangeli KSM_ATTR_RO(stable_node_dups); 33722c653d0eSAndrea Arcangeli 33732c653d0eSAndrea Arcangeli static ssize_t stable_node_chains_show(struct kobject *kobj, 33742c653d0eSAndrea Arcangeli struct kobj_attribute *attr, char *buf) 33752c653d0eSAndrea Arcangeli { 3376ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains); 33772c653d0eSAndrea Arcangeli } 33782c653d0eSAndrea Arcangeli KSM_ATTR_RO(stable_node_chains); 33792c653d0eSAndrea Arcangeli 33802c653d0eSAndrea Arcangeli static ssize_t 33812c653d0eSAndrea Arcangeli stable_node_chains_prune_millisecs_show(struct kobject *kobj, 33822c653d0eSAndrea Arcangeli struct kobj_attribute *attr, 33832c653d0eSAndrea Arcangeli char *buf) 33842c653d0eSAndrea Arcangeli { 3385ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs); 33862c653d0eSAndrea Arcangeli } 33872c653d0eSAndrea Arcangeli 33882c653d0eSAndrea Arcangeli static ssize_t 33892c653d0eSAndrea Arcangeli stable_node_chains_prune_millisecs_store(struct kobject *kobj, 33902c653d0eSAndrea Arcangeli struct kobj_attribute *attr, 33912c653d0eSAndrea Arcangeli const char *buf, size_t count) 33922c653d0eSAndrea Arcangeli { 3393584ff0dfSZhansaya Bagdauletkyzy unsigned int msecs; 33942c653d0eSAndrea Arcangeli int err; 33952c653d0eSAndrea Arcangeli 3396584ff0dfSZhansaya Bagdauletkyzy err = kstrtouint(buf, 10, &msecs); 3397584ff0dfSZhansaya Bagdauletkyzy if (err) 33982c653d0eSAndrea Arcangeli return -EINVAL; 33992c653d0eSAndrea Arcangeli 34002c653d0eSAndrea Arcangeli ksm_stable_node_chains_prune_millisecs = msecs; 34012c653d0eSAndrea Arcangeli 34022c653d0eSAndrea Arcangeli return count; 34032c653d0eSAndrea Arcangeli } 34042c653d0eSAndrea Arcangeli KSM_ATTR(stable_node_chains_prune_millisecs); 34052c653d0eSAndrea Arcangeli 3406473b0ce4SHugh Dickins static ssize_t full_scans_show(struct kobject *kobj, 3407473b0ce4SHugh Dickins struct kobj_attribute *attr, char *buf) 3408473b0ce4SHugh Dickins { 3409ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr); 3410473b0ce4SHugh Dickins } 3411473b0ce4SHugh Dickins KSM_ATTR_RO(full_scans); 3412473b0ce4SHugh Dickins 341331dbd01fSIzik Eidus static struct attribute *ksm_attrs[] = { 341431dbd01fSIzik Eidus &sleep_millisecs_attr.attr, 341531dbd01fSIzik Eidus &pages_to_scan_attr.attr, 341631dbd01fSIzik Eidus &run_attr.attr, 3417b4028260SHugh Dickins &pages_shared_attr.attr, 3418b4028260SHugh Dickins &pages_sharing_attr.attr, 3419473b0ce4SHugh Dickins &pages_unshared_attr.attr, 3420473b0ce4SHugh Dickins &pages_volatile_attr.attr, 3421473b0ce4SHugh Dickins &full_scans_attr.attr, 342290bd6fd3SPetr Holasek #ifdef CONFIG_NUMA 342390bd6fd3SPetr Holasek &merge_across_nodes_attr.attr, 342490bd6fd3SPetr Holasek #endif 34252c653d0eSAndrea Arcangeli &max_page_sharing_attr.attr, 34262c653d0eSAndrea Arcangeli &stable_node_chains_attr.attr, 34272c653d0eSAndrea Arcangeli &stable_node_dups_attr.attr, 34282c653d0eSAndrea Arcangeli &stable_node_chains_prune_millisecs_attr.attr, 3429e86c59b1SClaudio Imbrenda &use_zero_pages_attr.attr, 3430d21077fbSStefan Roesch &general_profit_attr.attr, 343131dbd01fSIzik Eidus NULL, 343231dbd01fSIzik Eidus }; 343331dbd01fSIzik Eidus 3434f907c26aSArvind Yadav static const struct attribute_group ksm_attr_group = { 343531dbd01fSIzik Eidus .attrs = ksm_attrs, 343631dbd01fSIzik Eidus .name = "ksm", 343731dbd01fSIzik Eidus }; 34382ffd8679SHugh Dickins #endif /* CONFIG_SYSFS */ 343931dbd01fSIzik Eidus 344031dbd01fSIzik Eidus static int __init ksm_init(void) 344131dbd01fSIzik Eidus { 344231dbd01fSIzik Eidus struct task_struct *ksm_thread; 344331dbd01fSIzik Eidus int err; 344431dbd01fSIzik Eidus 3445e86c59b1SClaudio Imbrenda /* The correct value depends on page size and endianness */ 3446e86c59b1SClaudio Imbrenda zero_checksum = calc_checksum(ZERO_PAGE(0)); 3447e86c59b1SClaudio Imbrenda /* Default to false for backwards compatibility */ 3448e86c59b1SClaudio Imbrenda ksm_use_zero_pages = false; 3449e86c59b1SClaudio Imbrenda 345031dbd01fSIzik Eidus err = ksm_slab_init(); 345131dbd01fSIzik Eidus if (err) 345231dbd01fSIzik Eidus goto out; 345331dbd01fSIzik Eidus 345431dbd01fSIzik Eidus ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); 345531dbd01fSIzik Eidus if (IS_ERR(ksm_thread)) { 345625acde31SPaul McQuade pr_err("ksm: creating kthread failed\n"); 345731dbd01fSIzik Eidus err = PTR_ERR(ksm_thread); 3458d9f8984cSLai Jiangshan goto out_free; 345931dbd01fSIzik Eidus } 346031dbd01fSIzik Eidus 34612ffd8679SHugh Dickins #ifdef CONFIG_SYSFS 346231dbd01fSIzik Eidus err = sysfs_create_group(mm_kobj, &ksm_attr_group); 346331dbd01fSIzik Eidus if (err) { 346425acde31SPaul McQuade pr_err("ksm: register sysfs failed\n"); 34652ffd8679SHugh Dickins kthread_stop(ksm_thread); 3466d9f8984cSLai Jiangshan goto out_free; 346731dbd01fSIzik Eidus } 3468c73602adSHugh Dickins #else 3469c73602adSHugh Dickins ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ 3470c73602adSHugh Dickins 34712ffd8679SHugh Dickins #endif /* CONFIG_SYSFS */ 347231dbd01fSIzik Eidus 347362b61f61SHugh Dickins #ifdef CONFIG_MEMORY_HOTREMOVE 3474ef4d43a8SHugh Dickins /* There is no significance to this priority 100 */ 34751eeaa4fdSLiu Shixin hotplug_memory_notifier(ksm_memory_callback, KSM_CALLBACK_PRI); 347662b61f61SHugh Dickins #endif 347731dbd01fSIzik Eidus return 0; 347831dbd01fSIzik Eidus 3479d9f8984cSLai Jiangshan out_free: 348031dbd01fSIzik Eidus ksm_slab_free(); 348131dbd01fSIzik Eidus out: 348231dbd01fSIzik Eidus return err; 348331dbd01fSIzik Eidus } 3484a64fb3cdSPaul Gortmaker subsys_initcall(ksm_init); 3485