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> 42f8af4da3SHugh Dickins 4331dbd01fSIzik Eidus #include <asm/tlbflush.h> 4473848b46SHugh Dickins #include "internal.h" 4558730ab6SQi Zheng #include "mm_slot.h" 4631dbd01fSIzik Eidus 47e850dcf5SHugh Dickins #ifdef CONFIG_NUMA 48e850dcf5SHugh Dickins #define NUMA(x) (x) 49e850dcf5SHugh Dickins #define DO_NUMA(x) do { (x); } while (0) 50e850dcf5SHugh Dickins #else 51e850dcf5SHugh Dickins #define NUMA(x) (0) 52e850dcf5SHugh Dickins #define DO_NUMA(x) do { } while (0) 53e850dcf5SHugh Dickins #endif 54e850dcf5SHugh Dickins 555a2ca3efSMike Rapoport /** 565a2ca3efSMike Rapoport * DOC: Overview 575a2ca3efSMike Rapoport * 5831dbd01fSIzik Eidus * A few notes about the KSM scanning process, 5931dbd01fSIzik Eidus * to make it easier to understand the data structures below: 6031dbd01fSIzik Eidus * 6131dbd01fSIzik Eidus * In order to reduce excessive scanning, KSM sorts the memory pages by their 6231dbd01fSIzik Eidus * contents into a data structure that holds pointers to the pages' locations. 6331dbd01fSIzik Eidus * 6431dbd01fSIzik Eidus * Since the contents of the pages may change at any moment, KSM cannot just 6531dbd01fSIzik Eidus * insert the pages into a normal sorted tree and expect it to find anything. 6631dbd01fSIzik Eidus * Therefore KSM uses two data structures - the stable and the unstable tree. 6731dbd01fSIzik Eidus * 6831dbd01fSIzik Eidus * The stable tree holds pointers to all the merged pages (ksm pages), sorted 6931dbd01fSIzik Eidus * by their contents. Because each such page is write-protected, searching on 7031dbd01fSIzik Eidus * this tree is fully assured to be working (except when pages are unmapped), 7131dbd01fSIzik Eidus * and therefore this tree is called the stable tree. 7231dbd01fSIzik Eidus * 735a2ca3efSMike Rapoport * The stable tree node includes information required for reverse 745a2ca3efSMike Rapoport * mapping from a KSM page to virtual addresses that map this page. 755a2ca3efSMike Rapoport * 765a2ca3efSMike Rapoport * In order to avoid large latencies of the rmap walks on KSM pages, 775a2ca3efSMike Rapoport * KSM maintains two types of nodes in the stable tree: 785a2ca3efSMike Rapoport * 795a2ca3efSMike Rapoport * * the regular nodes that keep the reverse mapping structures in a 805a2ca3efSMike Rapoport * linked list 815a2ca3efSMike Rapoport * * the "chains" that link nodes ("dups") that represent the same 825a2ca3efSMike Rapoport * write protected memory content, but each "dup" corresponds to a 835a2ca3efSMike Rapoport * different KSM page copy of that content 845a2ca3efSMike Rapoport * 855a2ca3efSMike Rapoport * Internally, the regular nodes, "dups" and "chains" are represented 8621fbd591SQi Zheng * using the same struct ksm_stable_node structure. 875a2ca3efSMike Rapoport * 8831dbd01fSIzik Eidus * In addition to the stable tree, KSM uses a second data structure called the 8931dbd01fSIzik Eidus * unstable tree: this tree holds pointers to pages which have been found to 9031dbd01fSIzik Eidus * be "unchanged for a period of time". The unstable tree sorts these pages 9131dbd01fSIzik Eidus * by their contents, but since they are not write-protected, KSM cannot rely 9231dbd01fSIzik Eidus * upon the unstable tree to work correctly - the unstable tree is liable to 9331dbd01fSIzik Eidus * be corrupted as its contents are modified, and so it is called unstable. 9431dbd01fSIzik Eidus * 9531dbd01fSIzik Eidus * KSM solves this problem by several techniques: 9631dbd01fSIzik Eidus * 9731dbd01fSIzik Eidus * 1) The unstable tree is flushed every time KSM completes scanning all 9831dbd01fSIzik Eidus * memory areas, and then the tree is rebuilt again from the beginning. 9931dbd01fSIzik Eidus * 2) KSM will only insert into the unstable tree, pages whose hash value 10031dbd01fSIzik Eidus * has not changed since the previous scan of all memory areas. 10131dbd01fSIzik Eidus * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the 10231dbd01fSIzik Eidus * colors of the nodes and not on their contents, assuring that even when 10331dbd01fSIzik Eidus * the tree gets "corrupted" it won't get out of balance, so scanning time 10431dbd01fSIzik Eidus * remains the same (also, searching and inserting nodes in an rbtree uses 10531dbd01fSIzik Eidus * the same algorithm, so we have no overhead when we flush and rebuild). 10631dbd01fSIzik Eidus * 4) KSM never flushes the stable tree, which means that even if it were to 10731dbd01fSIzik Eidus * take 10 attempts to find a page in the unstable tree, once it is found, 10831dbd01fSIzik Eidus * it is secured in the stable tree. (When we scan a new page, we first 10931dbd01fSIzik Eidus * compare it against the stable tree, and then against the unstable tree.) 1108fdb3dbfSHugh Dickins * 1118fdb3dbfSHugh Dickins * If the merge_across_nodes tunable is unset, then KSM maintains multiple 1128fdb3dbfSHugh Dickins * stable trees and multiple unstable trees: one of each for each NUMA node. 11331dbd01fSIzik Eidus */ 11431dbd01fSIzik Eidus 11531dbd01fSIzik Eidus /** 11621fbd591SQi Zheng * struct ksm_mm_slot - ksm information per mm that is being scanned 11758730ab6SQi Zheng * @slot: hash lookup from mm to mm_slot 1186514d511SHugh Dickins * @rmap_list: head for this mm_slot's singly-linked list of rmap_items 11931dbd01fSIzik Eidus */ 12021fbd591SQi Zheng struct ksm_mm_slot { 12158730ab6SQi Zheng struct mm_slot slot; 12221fbd591SQi Zheng struct ksm_rmap_item *rmap_list; 12331dbd01fSIzik Eidus }; 12431dbd01fSIzik Eidus 12531dbd01fSIzik Eidus /** 12631dbd01fSIzik Eidus * struct ksm_scan - cursor for scanning 12731dbd01fSIzik Eidus * @mm_slot: the current mm_slot we are scanning 12831dbd01fSIzik Eidus * @address: the next address inside that to be scanned 1296514d511SHugh Dickins * @rmap_list: link to the next rmap to be scanned in the rmap_list 13031dbd01fSIzik Eidus * @seqnr: count of completed full scans (needed when removing unstable node) 13131dbd01fSIzik Eidus * 13231dbd01fSIzik Eidus * There is only the one ksm_scan instance of this cursor structure. 13331dbd01fSIzik Eidus */ 13431dbd01fSIzik Eidus struct ksm_scan { 13521fbd591SQi Zheng struct ksm_mm_slot *mm_slot; 13631dbd01fSIzik Eidus unsigned long address; 13721fbd591SQi Zheng struct ksm_rmap_item **rmap_list; 13831dbd01fSIzik Eidus unsigned long seqnr; 13931dbd01fSIzik Eidus }; 14031dbd01fSIzik Eidus 14131dbd01fSIzik Eidus /** 14221fbd591SQi Zheng * struct ksm_stable_node - node of the stable rbtree 1437b6ba2c7SHugh Dickins * @node: rb node of this ksm page in the stable tree 1444146d2d6SHugh Dickins * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list 1452c653d0eSAndrea Arcangeli * @hlist_dup: linked into the stable_node->hlist with a stable_node chain 1464146d2d6SHugh Dickins * @list: linked into migrate_nodes, pending placement in the proper node tree 1477b6ba2c7SHugh Dickins * @hlist: hlist head of rmap_items using this ksm page 1484146d2d6SHugh Dickins * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid) 1492c653d0eSAndrea Arcangeli * @chain_prune_time: time of the last full garbage collection 1502c653d0eSAndrea Arcangeli * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN 1514146d2d6SHugh Dickins * @nid: NUMA node id of stable tree in which linked (may not match kpfn) 1527b6ba2c7SHugh Dickins */ 15321fbd591SQi Zheng struct ksm_stable_node { 1544146d2d6SHugh Dickins union { 1554146d2d6SHugh Dickins struct rb_node node; /* when node of stable tree */ 1564146d2d6SHugh Dickins struct { /* when listed for migration */ 1574146d2d6SHugh Dickins struct list_head *head; 1582c653d0eSAndrea Arcangeli struct { 1592c653d0eSAndrea Arcangeli struct hlist_node hlist_dup; 1604146d2d6SHugh Dickins struct list_head list; 1614146d2d6SHugh Dickins }; 1624146d2d6SHugh Dickins }; 1632c653d0eSAndrea Arcangeli }; 1647b6ba2c7SHugh Dickins struct hlist_head hlist; 1652c653d0eSAndrea Arcangeli union { 16662b61f61SHugh Dickins unsigned long kpfn; 1672c653d0eSAndrea Arcangeli unsigned long chain_prune_time; 1682c653d0eSAndrea Arcangeli }; 1692c653d0eSAndrea Arcangeli /* 1702c653d0eSAndrea Arcangeli * STABLE_NODE_CHAIN can be any negative number in 1712c653d0eSAndrea Arcangeli * rmap_hlist_len negative range, but better not -1 to be able 1722c653d0eSAndrea Arcangeli * to reliably detect underflows. 1732c653d0eSAndrea Arcangeli */ 1742c653d0eSAndrea Arcangeli #define STABLE_NODE_CHAIN -1024 1752c653d0eSAndrea Arcangeli int rmap_hlist_len; 1764146d2d6SHugh Dickins #ifdef CONFIG_NUMA 1774146d2d6SHugh Dickins int nid; 1784146d2d6SHugh Dickins #endif 1797b6ba2c7SHugh Dickins }; 1807b6ba2c7SHugh Dickins 1817b6ba2c7SHugh Dickins /** 18221fbd591SQi Zheng * struct ksm_rmap_item - reverse mapping item for virtual addresses 1836514d511SHugh Dickins * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list 184db114b83SHugh Dickins * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree 185bc56620bSHugh Dickins * @nid: NUMA node id of unstable tree in which linked (may not match page) 18631dbd01fSIzik Eidus * @mm: the memory structure this rmap_item is pointing into 18731dbd01fSIzik Eidus * @address: the virtual address this rmap_item tracks (+ flags in low bits) 18831dbd01fSIzik Eidus * @oldchecksum: previous checksum of the page at that virtual address 1897b6ba2c7SHugh Dickins * @node: rb node of this rmap_item in the unstable tree 1907b6ba2c7SHugh Dickins * @head: pointer to stable_node heading this list in the stable tree 1917b6ba2c7SHugh Dickins * @hlist: link into hlist of rmap_items hanging off that stable_node 19231dbd01fSIzik Eidus */ 19321fbd591SQi Zheng struct ksm_rmap_item { 19421fbd591SQi Zheng struct ksm_rmap_item *rmap_list; 195bc56620bSHugh Dickins union { 196db114b83SHugh Dickins struct anon_vma *anon_vma; /* when stable */ 197bc56620bSHugh Dickins #ifdef CONFIG_NUMA 198bc56620bSHugh Dickins int nid; /* when node of unstable tree */ 199bc56620bSHugh Dickins #endif 200bc56620bSHugh Dickins }; 20131dbd01fSIzik Eidus struct mm_struct *mm; 20231dbd01fSIzik Eidus unsigned long address; /* + low bits used for flags below */ 20331dbd01fSIzik Eidus unsigned int oldchecksum; /* when unstable */ 20431dbd01fSIzik Eidus union { 2057b6ba2c7SHugh Dickins struct rb_node node; /* when node of unstable tree */ 2067b6ba2c7SHugh Dickins struct { /* when listed from stable tree */ 20721fbd591SQi Zheng struct ksm_stable_node *head; 2087b6ba2c7SHugh Dickins struct hlist_node hlist; 2097b6ba2c7SHugh Dickins }; 21031dbd01fSIzik Eidus }; 21131dbd01fSIzik Eidus }; 21231dbd01fSIzik Eidus 21331dbd01fSIzik Eidus #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ 2147b6ba2c7SHugh Dickins #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ 2157b6ba2c7SHugh Dickins #define STABLE_FLAG 0x200 /* is listed from the stable tree */ 21631dbd01fSIzik Eidus 21731dbd01fSIzik Eidus /* The stable and unstable tree heads */ 218ef53d16cSHugh Dickins static struct rb_root one_stable_tree[1] = { RB_ROOT }; 219ef53d16cSHugh Dickins static struct rb_root one_unstable_tree[1] = { RB_ROOT }; 220ef53d16cSHugh Dickins static struct rb_root *root_stable_tree = one_stable_tree; 221ef53d16cSHugh Dickins static struct rb_root *root_unstable_tree = one_unstable_tree; 22231dbd01fSIzik Eidus 2234146d2d6SHugh Dickins /* Recently migrated nodes of stable tree, pending proper placement */ 2244146d2d6SHugh Dickins static LIST_HEAD(migrate_nodes); 2252c653d0eSAndrea Arcangeli #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev) 2264146d2d6SHugh Dickins 2274ca3a69bSSasha Levin #define MM_SLOTS_HASH_BITS 10 2284ca3a69bSSasha Levin static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 22931dbd01fSIzik Eidus 23021fbd591SQi Zheng static struct ksm_mm_slot ksm_mm_head = { 23158730ab6SQi Zheng .slot.mm_node = LIST_HEAD_INIT(ksm_mm_head.slot.mm_node), 23231dbd01fSIzik Eidus }; 23331dbd01fSIzik Eidus static struct ksm_scan ksm_scan = { 23431dbd01fSIzik Eidus .mm_slot = &ksm_mm_head, 23531dbd01fSIzik Eidus }; 23631dbd01fSIzik Eidus 23731dbd01fSIzik Eidus static struct kmem_cache *rmap_item_cache; 2387b6ba2c7SHugh Dickins static struct kmem_cache *stable_node_cache; 23931dbd01fSIzik Eidus static struct kmem_cache *mm_slot_cache; 24031dbd01fSIzik Eidus 24131dbd01fSIzik Eidus /* The number of nodes in the stable tree */ 242b4028260SHugh Dickins static unsigned long ksm_pages_shared; 24331dbd01fSIzik Eidus 244e178dfdeSHugh Dickins /* The number of page slots additionally sharing those nodes */ 245b4028260SHugh Dickins static unsigned long ksm_pages_sharing; 24631dbd01fSIzik Eidus 247473b0ce4SHugh Dickins /* The number of nodes in the unstable tree */ 248473b0ce4SHugh Dickins static unsigned long ksm_pages_unshared; 249473b0ce4SHugh Dickins 250473b0ce4SHugh Dickins /* The number of rmap_items in use: to calculate pages_volatile */ 251473b0ce4SHugh Dickins static unsigned long ksm_rmap_items; 252473b0ce4SHugh Dickins 2532c653d0eSAndrea Arcangeli /* The number of stable_node chains */ 2542c653d0eSAndrea Arcangeli static unsigned long ksm_stable_node_chains; 2552c653d0eSAndrea Arcangeli 2562c653d0eSAndrea Arcangeli /* The number of stable_node dups linked to the stable_node chains */ 2572c653d0eSAndrea Arcangeli static unsigned long ksm_stable_node_dups; 2582c653d0eSAndrea Arcangeli 2592c653d0eSAndrea Arcangeli /* Delay in pruning stale stable_node_dups in the stable_node_chains */ 260584ff0dfSZhansaya Bagdauletkyzy static unsigned int ksm_stable_node_chains_prune_millisecs = 2000; 2612c653d0eSAndrea Arcangeli 2622c653d0eSAndrea Arcangeli /* Maximum number of page slots sharing a stable node */ 2632c653d0eSAndrea Arcangeli static int ksm_max_page_sharing = 256; 2642c653d0eSAndrea Arcangeli 26531dbd01fSIzik Eidus /* Number of pages ksmd should scan in one batch */ 2662c6854fdSIzik Eidus static unsigned int ksm_thread_pages_to_scan = 100; 26731dbd01fSIzik Eidus 26831dbd01fSIzik Eidus /* Milliseconds ksmd should sleep between batches */ 2692ffd8679SHugh Dickins static unsigned int ksm_thread_sleep_millisecs = 20; 27031dbd01fSIzik Eidus 271e86c59b1SClaudio Imbrenda /* Checksum of an empty (zeroed) page */ 272e86c59b1SClaudio Imbrenda static unsigned int zero_checksum __read_mostly; 273e86c59b1SClaudio Imbrenda 274e86c59b1SClaudio Imbrenda /* Whether to merge empty (zeroed) pages with actual zero pages */ 275e86c59b1SClaudio Imbrenda static bool ksm_use_zero_pages __read_mostly; 276e86c59b1SClaudio Imbrenda 277e850dcf5SHugh Dickins #ifdef CONFIG_NUMA 27890bd6fd3SPetr Holasek /* Zeroed when merging across nodes is not allowed */ 27990bd6fd3SPetr Holasek static unsigned int ksm_merge_across_nodes = 1; 280ef53d16cSHugh Dickins static int ksm_nr_node_ids = 1; 281e850dcf5SHugh Dickins #else 282e850dcf5SHugh Dickins #define ksm_merge_across_nodes 1U 283ef53d16cSHugh Dickins #define ksm_nr_node_ids 1 284e850dcf5SHugh Dickins #endif 28590bd6fd3SPetr Holasek 28631dbd01fSIzik Eidus #define KSM_RUN_STOP 0 28731dbd01fSIzik Eidus #define KSM_RUN_MERGE 1 28831dbd01fSIzik Eidus #define KSM_RUN_UNMERGE 2 289ef4d43a8SHugh Dickins #define KSM_RUN_OFFLINE 4 290ef4d43a8SHugh Dickins static unsigned long ksm_run = KSM_RUN_STOP; 291ef4d43a8SHugh Dickins static void wait_while_offlining(void); 29231dbd01fSIzik Eidus 29331dbd01fSIzik Eidus static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); 294fcf9a0efSKirill Tkhai static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait); 29531dbd01fSIzik Eidus static DEFINE_MUTEX(ksm_thread_mutex); 29631dbd01fSIzik Eidus static DEFINE_SPINLOCK(ksm_mmlist_lock); 29731dbd01fSIzik Eidus 29821fbd591SQi Zheng #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create(#__struct,\ 29931dbd01fSIzik Eidus sizeof(struct __struct), __alignof__(struct __struct),\ 30031dbd01fSIzik Eidus (__flags), NULL) 30131dbd01fSIzik Eidus 30231dbd01fSIzik Eidus static int __init ksm_slab_init(void) 30331dbd01fSIzik Eidus { 30421fbd591SQi Zheng rmap_item_cache = KSM_KMEM_CACHE(ksm_rmap_item, 0); 30531dbd01fSIzik Eidus if (!rmap_item_cache) 30631dbd01fSIzik Eidus goto out; 30731dbd01fSIzik Eidus 30821fbd591SQi Zheng stable_node_cache = KSM_KMEM_CACHE(ksm_stable_node, 0); 3097b6ba2c7SHugh Dickins if (!stable_node_cache) 3107b6ba2c7SHugh Dickins goto out_free1; 3117b6ba2c7SHugh Dickins 31221fbd591SQi Zheng mm_slot_cache = KSM_KMEM_CACHE(ksm_mm_slot, 0); 31331dbd01fSIzik Eidus if (!mm_slot_cache) 3147b6ba2c7SHugh Dickins goto out_free2; 31531dbd01fSIzik Eidus 31631dbd01fSIzik Eidus return 0; 31731dbd01fSIzik Eidus 3187b6ba2c7SHugh Dickins out_free2: 3197b6ba2c7SHugh Dickins kmem_cache_destroy(stable_node_cache); 3207b6ba2c7SHugh Dickins out_free1: 32131dbd01fSIzik Eidus kmem_cache_destroy(rmap_item_cache); 32231dbd01fSIzik Eidus out: 32331dbd01fSIzik Eidus return -ENOMEM; 32431dbd01fSIzik Eidus } 32531dbd01fSIzik Eidus 32631dbd01fSIzik Eidus static void __init ksm_slab_free(void) 32731dbd01fSIzik Eidus { 32831dbd01fSIzik Eidus kmem_cache_destroy(mm_slot_cache); 3297b6ba2c7SHugh Dickins kmem_cache_destroy(stable_node_cache); 33031dbd01fSIzik Eidus kmem_cache_destroy(rmap_item_cache); 33131dbd01fSIzik Eidus mm_slot_cache = NULL; 33231dbd01fSIzik Eidus } 33331dbd01fSIzik Eidus 33421fbd591SQi Zheng static __always_inline bool is_stable_node_chain(struct ksm_stable_node *chain) 3352c653d0eSAndrea Arcangeli { 3362c653d0eSAndrea Arcangeli return chain->rmap_hlist_len == STABLE_NODE_CHAIN; 3372c653d0eSAndrea Arcangeli } 3382c653d0eSAndrea Arcangeli 33921fbd591SQi Zheng static __always_inline bool is_stable_node_dup(struct ksm_stable_node *dup) 3402c653d0eSAndrea Arcangeli { 3412c653d0eSAndrea Arcangeli return dup->head == STABLE_NODE_DUP_HEAD; 3422c653d0eSAndrea Arcangeli } 3432c653d0eSAndrea Arcangeli 34421fbd591SQi Zheng static inline void stable_node_chain_add_dup(struct ksm_stable_node *dup, 34521fbd591SQi Zheng struct ksm_stable_node *chain) 3462c653d0eSAndrea Arcangeli { 3472c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(dup)); 3482c653d0eSAndrea Arcangeli dup->head = STABLE_NODE_DUP_HEAD; 3492c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_chain(chain)); 3502c653d0eSAndrea Arcangeli hlist_add_head(&dup->hlist_dup, &chain->hlist); 3512c653d0eSAndrea Arcangeli ksm_stable_node_dups++; 3522c653d0eSAndrea Arcangeli } 3532c653d0eSAndrea Arcangeli 35421fbd591SQi Zheng static inline void __stable_node_dup_del(struct ksm_stable_node *dup) 3552c653d0eSAndrea Arcangeli { 356b4fecc67SAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(dup)); 3572c653d0eSAndrea Arcangeli hlist_del(&dup->hlist_dup); 3582c653d0eSAndrea Arcangeli ksm_stable_node_dups--; 3592c653d0eSAndrea Arcangeli } 3602c653d0eSAndrea Arcangeli 36121fbd591SQi Zheng static inline void stable_node_dup_del(struct ksm_stable_node *dup) 3622c653d0eSAndrea Arcangeli { 3632c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_chain(dup)); 3642c653d0eSAndrea Arcangeli if (is_stable_node_dup(dup)) 3652c653d0eSAndrea Arcangeli __stable_node_dup_del(dup); 3662c653d0eSAndrea Arcangeli else 3672c653d0eSAndrea Arcangeli rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid)); 3682c653d0eSAndrea Arcangeli #ifdef CONFIG_DEBUG_VM 3692c653d0eSAndrea Arcangeli dup->head = NULL; 3702c653d0eSAndrea Arcangeli #endif 3712c653d0eSAndrea Arcangeli } 3722c653d0eSAndrea Arcangeli 37321fbd591SQi Zheng static inline struct ksm_rmap_item *alloc_rmap_item(void) 37431dbd01fSIzik Eidus { 37521fbd591SQi Zheng struct ksm_rmap_item *rmap_item; 376473b0ce4SHugh Dickins 3775b398e41Szhong jiang rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL | 3785b398e41Szhong jiang __GFP_NORETRY | __GFP_NOWARN); 379473b0ce4SHugh Dickins if (rmap_item) 380473b0ce4SHugh Dickins ksm_rmap_items++; 381473b0ce4SHugh Dickins return rmap_item; 38231dbd01fSIzik Eidus } 38331dbd01fSIzik Eidus 38421fbd591SQi Zheng static inline void free_rmap_item(struct ksm_rmap_item *rmap_item) 38531dbd01fSIzik Eidus { 386473b0ce4SHugh Dickins ksm_rmap_items--; 387cb4df4caSxu xin rmap_item->mm->ksm_rmap_items--; 38831dbd01fSIzik Eidus rmap_item->mm = NULL; /* debug safety */ 38931dbd01fSIzik Eidus kmem_cache_free(rmap_item_cache, rmap_item); 39031dbd01fSIzik Eidus } 39131dbd01fSIzik Eidus 39221fbd591SQi Zheng static inline struct ksm_stable_node *alloc_stable_node(void) 3937b6ba2c7SHugh Dickins { 3946213055fSzhong jiang /* 3956213055fSzhong jiang * The allocation can take too long with GFP_KERNEL when memory is under 3966213055fSzhong jiang * pressure, which may lead to hung task warnings. Adding __GFP_HIGH 3976213055fSzhong jiang * grants access to memory reserves, helping to avoid this problem. 3986213055fSzhong jiang */ 3996213055fSzhong jiang return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH); 4007b6ba2c7SHugh Dickins } 4017b6ba2c7SHugh Dickins 40221fbd591SQi Zheng static inline void free_stable_node(struct ksm_stable_node *stable_node) 4037b6ba2c7SHugh Dickins { 4042c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len && 4052c653d0eSAndrea Arcangeli !is_stable_node_chain(stable_node)); 4067b6ba2c7SHugh Dickins kmem_cache_free(stable_node_cache, stable_node); 4077b6ba2c7SHugh Dickins } 4087b6ba2c7SHugh Dickins 40931dbd01fSIzik Eidus /* 410a913e182SHugh Dickins * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's 411a913e182SHugh Dickins * page tables after it has passed through ksm_exit() - which, if necessary, 412c1e8d7c6SMichel Lespinasse * takes mmap_lock briefly to serialize against them. ksm_exit() does not set 413a913e182SHugh Dickins * a special flag: they can just back out as soon as mm_users goes to zero. 414a913e182SHugh Dickins * ksm_test_exit() is used throughout to make this test for exit: in some 415a913e182SHugh Dickins * places for correctness, in some places just to avoid unnecessary work. 416a913e182SHugh Dickins */ 417a913e182SHugh Dickins static inline bool ksm_test_exit(struct mm_struct *mm) 418a913e182SHugh Dickins { 419a913e182SHugh Dickins return atomic_read(&mm->mm_users) == 0; 420a913e182SHugh Dickins } 421a913e182SHugh Dickins 422a913e182SHugh Dickins /* 42331dbd01fSIzik Eidus * We use break_ksm to break COW on a ksm page: it's a stripped down 42431dbd01fSIzik Eidus * 4257a9547fdSLi Chen * if (get_user_pages(addr, 1, FOLL_WRITE, &page, NULL) == 1) 42631dbd01fSIzik Eidus * put_page(page); 42731dbd01fSIzik Eidus * 42831dbd01fSIzik Eidus * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma, 42931dbd01fSIzik Eidus * in case the application has unmapped and remapped mm,addr meanwhile. 43031dbd01fSIzik Eidus * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP 431bbcd53c9SDavid Hildenbrand * mmap of /dev/mem, where we would not want to touch it. 4321b2ee126SDave Hansen * 4331b2ee126SDave Hansen * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context 4341b2ee126SDave Hansen * of the process that owns 'vma'. We also do not want to enforce 4351b2ee126SDave Hansen * protection keys here anyway. 43631dbd01fSIzik Eidus */ 437d952b791SHugh Dickins static int break_ksm(struct vm_area_struct *vma, unsigned long addr) 43831dbd01fSIzik Eidus { 43931dbd01fSIzik Eidus struct page *page; 44050a7ca3cSSouptick Joarder vm_fault_t ret = 0; 44131dbd01fSIzik Eidus 44231dbd01fSIzik Eidus do { 44331dbd01fSIzik Eidus cond_resched(); 4441b2ee126SDave Hansen page = follow_page(vma, addr, 4451b2ee126SDave Hansen FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE); 446f7091ed6SHaiyue Wang if (IS_ERR_OR_NULL(page)) 44731dbd01fSIzik Eidus break; 44831dbd01fSIzik Eidus if (PageKsm(page)) 449dcddffd4SKirill A. Shutemov ret = handle_mm_fault(vma, addr, 450bce617edSPeter Xu FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE, 451bce617edSPeter Xu NULL); 45231dbd01fSIzik Eidus else 45331dbd01fSIzik Eidus ret = VM_FAULT_WRITE; 45431dbd01fSIzik Eidus put_page(page); 45533692f27SLinus Torvalds } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM))); 456d952b791SHugh Dickins /* 457d952b791SHugh Dickins * We must loop because handle_mm_fault() may back out if there's 458d952b791SHugh Dickins * any difficulty e.g. if pte accessed bit gets updated concurrently. 459d952b791SHugh Dickins * 460d952b791SHugh Dickins * VM_FAULT_WRITE is what we have been hoping for: it indicates that 461d952b791SHugh Dickins * COW has been broken, even if the vma does not permit VM_WRITE; 462d952b791SHugh Dickins * but note that a concurrent fault might break PageKsm for us. 463d952b791SHugh Dickins * 464d952b791SHugh Dickins * VM_FAULT_SIGBUS could occur if we race with truncation of the 465d952b791SHugh Dickins * backing file, which also invalidates anonymous pages: that's 466d952b791SHugh Dickins * okay, that truncation will have unmapped the PageKsm for us. 467d952b791SHugh Dickins * 468d952b791SHugh Dickins * VM_FAULT_OOM: at the time of writing (late July 2009), setting 469d952b791SHugh Dickins * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the 470d952b791SHugh Dickins * current task has TIF_MEMDIE set, and will be OOM killed on return 471d952b791SHugh Dickins * to user; and ksmd, having no mm, would never be chosen for that. 472d952b791SHugh Dickins * 473d952b791SHugh Dickins * But if the mm is in a limited mem_cgroup, then the fault may fail 474d952b791SHugh Dickins * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and 475d952b791SHugh Dickins * even ksmd can fail in this way - though it's usually breaking ksm 476d952b791SHugh Dickins * just to undo a merge it made a moment before, so unlikely to oom. 477d952b791SHugh Dickins * 478d952b791SHugh Dickins * That's a pity: we might therefore have more kernel pages allocated 479d952b791SHugh Dickins * than we're counting as nodes in the stable tree; but ksm_do_scan 480d952b791SHugh Dickins * will retry to break_cow on each pass, so should recover the page 481d952b791SHugh Dickins * in due course. The important thing is to not let VM_MERGEABLE 482d952b791SHugh Dickins * be cleared while any such pages might remain in the area. 483d952b791SHugh Dickins */ 484d952b791SHugh Dickins return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; 48531dbd01fSIzik Eidus } 48631dbd01fSIzik Eidus 487ef694222SBob Liu static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm, 488ef694222SBob Liu unsigned long addr) 489ef694222SBob Liu { 490ef694222SBob Liu struct vm_area_struct *vma; 491ef694222SBob Liu if (ksm_test_exit(mm)) 492ef694222SBob Liu return NULL; 493ff69fb81SLiam Howlett vma = vma_lookup(mm, addr); 494ff69fb81SLiam Howlett if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) 495ef694222SBob Liu return NULL; 496ef694222SBob Liu return vma; 497ef694222SBob Liu } 498ef694222SBob Liu 49921fbd591SQi Zheng static void break_cow(struct ksm_rmap_item *rmap_item) 50031dbd01fSIzik Eidus { 5018dd3557aSHugh Dickins struct mm_struct *mm = rmap_item->mm; 5028dd3557aSHugh Dickins unsigned long addr = rmap_item->address; 50331dbd01fSIzik Eidus struct vm_area_struct *vma; 50431dbd01fSIzik Eidus 5054035c07aSHugh Dickins /* 5064035c07aSHugh Dickins * It is not an accident that whenever we want to break COW 5074035c07aSHugh Dickins * to undo, we also need to drop a reference to the anon_vma. 5084035c07aSHugh Dickins */ 5099e60109fSPeter Zijlstra put_anon_vma(rmap_item->anon_vma); 5104035c07aSHugh Dickins 511d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 512ef694222SBob Liu vma = find_mergeable_vma(mm, addr); 513ef694222SBob Liu if (vma) 51431dbd01fSIzik Eidus break_ksm(vma, addr); 515d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 51631dbd01fSIzik Eidus } 51731dbd01fSIzik Eidus 51821fbd591SQi Zheng static struct page *get_mergeable_page(struct ksm_rmap_item *rmap_item) 51931dbd01fSIzik Eidus { 52031dbd01fSIzik Eidus struct mm_struct *mm = rmap_item->mm; 52131dbd01fSIzik Eidus unsigned long addr = rmap_item->address; 52231dbd01fSIzik Eidus struct vm_area_struct *vma; 52331dbd01fSIzik Eidus struct page *page; 52431dbd01fSIzik Eidus 525d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 526ef694222SBob Liu vma = find_mergeable_vma(mm, addr); 527ef694222SBob Liu if (!vma) 52831dbd01fSIzik Eidus goto out; 52931dbd01fSIzik Eidus 53031dbd01fSIzik Eidus page = follow_page(vma, addr, FOLL_GET); 531f7091ed6SHaiyue Wang if (IS_ERR_OR_NULL(page)) 53231dbd01fSIzik Eidus goto out; 533f7091ed6SHaiyue Wang if (is_zone_device_page(page)) 534f7091ed6SHaiyue Wang goto out_putpage; 535f765f540SKirill A. Shutemov if (PageAnon(page)) { 53631dbd01fSIzik Eidus flush_anon_page(vma, page, addr); 53731dbd01fSIzik Eidus flush_dcache_page(page); 53831dbd01fSIzik Eidus } else { 539f7091ed6SHaiyue Wang out_putpage: 54031dbd01fSIzik Eidus put_page(page); 541c8f95ed1SAndrea Arcangeli out: 542c8f95ed1SAndrea Arcangeli page = NULL; 54331dbd01fSIzik Eidus } 544d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 54531dbd01fSIzik Eidus return page; 54631dbd01fSIzik Eidus } 54731dbd01fSIzik Eidus 54890bd6fd3SPetr Holasek /* 54990bd6fd3SPetr Holasek * This helper is used for getting right index into array of tree roots. 55090bd6fd3SPetr Holasek * When merge_across_nodes knob is set to 1, there are only two rb-trees for 55190bd6fd3SPetr Holasek * stable and unstable pages from all nodes with roots in index 0. Otherwise, 55290bd6fd3SPetr Holasek * every node has its own stable and unstable tree. 55390bd6fd3SPetr Holasek */ 55490bd6fd3SPetr Holasek static inline int get_kpfn_nid(unsigned long kpfn) 55590bd6fd3SPetr Holasek { 556d8fc16a8SHugh Dickins return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn)); 55790bd6fd3SPetr Holasek } 55890bd6fd3SPetr Holasek 55921fbd591SQi Zheng static struct ksm_stable_node *alloc_stable_node_chain(struct ksm_stable_node *dup, 5602c653d0eSAndrea Arcangeli struct rb_root *root) 5612c653d0eSAndrea Arcangeli { 56221fbd591SQi Zheng struct ksm_stable_node *chain = alloc_stable_node(); 5632c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_chain(dup)); 5642c653d0eSAndrea Arcangeli if (likely(chain)) { 5652c653d0eSAndrea Arcangeli INIT_HLIST_HEAD(&chain->hlist); 5662c653d0eSAndrea Arcangeli chain->chain_prune_time = jiffies; 5672c653d0eSAndrea Arcangeli chain->rmap_hlist_len = STABLE_NODE_CHAIN; 5682c653d0eSAndrea Arcangeli #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA) 56998fa15f3SAnshuman Khandual chain->nid = NUMA_NO_NODE; /* debug */ 5702c653d0eSAndrea Arcangeli #endif 5712c653d0eSAndrea Arcangeli ksm_stable_node_chains++; 5722c653d0eSAndrea Arcangeli 5732c653d0eSAndrea Arcangeli /* 5742c653d0eSAndrea Arcangeli * Put the stable node chain in the first dimension of 5752c653d0eSAndrea Arcangeli * the stable tree and at the same time remove the old 5762c653d0eSAndrea Arcangeli * stable node. 5772c653d0eSAndrea Arcangeli */ 5782c653d0eSAndrea Arcangeli rb_replace_node(&dup->node, &chain->node, root); 5792c653d0eSAndrea Arcangeli 5802c653d0eSAndrea Arcangeli /* 5812c653d0eSAndrea Arcangeli * Move the old stable node to the second dimension 5822c653d0eSAndrea Arcangeli * queued in the hlist_dup. The invariant is that all 5832c653d0eSAndrea Arcangeli * dup stable_nodes in the chain->hlist point to pages 584457aef94SEthon Paul * that are write protected and have the exact same 5852c653d0eSAndrea Arcangeli * content. 5862c653d0eSAndrea Arcangeli */ 5872c653d0eSAndrea Arcangeli stable_node_chain_add_dup(dup, chain); 5882c653d0eSAndrea Arcangeli } 5892c653d0eSAndrea Arcangeli return chain; 5902c653d0eSAndrea Arcangeli } 5912c653d0eSAndrea Arcangeli 59221fbd591SQi Zheng static inline void free_stable_node_chain(struct ksm_stable_node *chain, 5932c653d0eSAndrea Arcangeli struct rb_root *root) 5942c653d0eSAndrea Arcangeli { 5952c653d0eSAndrea Arcangeli rb_erase(&chain->node, root); 5962c653d0eSAndrea Arcangeli free_stable_node(chain); 5972c653d0eSAndrea Arcangeli ksm_stable_node_chains--; 5982c653d0eSAndrea Arcangeli } 5992c653d0eSAndrea Arcangeli 60021fbd591SQi Zheng static void remove_node_from_stable_tree(struct ksm_stable_node *stable_node) 6014035c07aSHugh Dickins { 60221fbd591SQi Zheng struct ksm_rmap_item *rmap_item; 6034035c07aSHugh Dickins 6042c653d0eSAndrea Arcangeli /* check it's not STABLE_NODE_CHAIN or negative */ 6052c653d0eSAndrea Arcangeli BUG_ON(stable_node->rmap_hlist_len < 0); 6062c653d0eSAndrea Arcangeli 607b67bfe0dSSasha Levin hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { 6084035c07aSHugh Dickins if (rmap_item->hlist.next) 6094035c07aSHugh Dickins ksm_pages_sharing--; 6104035c07aSHugh Dickins else 6114035c07aSHugh Dickins ksm_pages_shared--; 61276093853Sxu xin 61376093853Sxu xin rmap_item->mm->ksm_merging_pages--; 61476093853Sxu xin 6152c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len <= 0); 6162c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len--; 6179e60109fSPeter Zijlstra put_anon_vma(rmap_item->anon_vma); 6184035c07aSHugh Dickins rmap_item->address &= PAGE_MASK; 6194035c07aSHugh Dickins cond_resched(); 6204035c07aSHugh Dickins } 6214035c07aSHugh Dickins 6222c653d0eSAndrea Arcangeli /* 6232c653d0eSAndrea Arcangeli * We need the second aligned pointer of the migrate_nodes 6242c653d0eSAndrea Arcangeli * list_head to stay clear from the rb_parent_color union 6252c653d0eSAndrea Arcangeli * (aligned and different than any node) and also different 6262c653d0eSAndrea Arcangeli * from &migrate_nodes. This will verify that future list.h changes 627815f0ddbSNick Desaulniers * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it. 6282c653d0eSAndrea Arcangeli */ 6292c653d0eSAndrea Arcangeli BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes); 6302c653d0eSAndrea Arcangeli BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1); 6312c653d0eSAndrea Arcangeli 6324146d2d6SHugh Dickins if (stable_node->head == &migrate_nodes) 6334146d2d6SHugh Dickins list_del(&stable_node->list); 6344146d2d6SHugh Dickins else 6352c653d0eSAndrea Arcangeli stable_node_dup_del(stable_node); 6364035c07aSHugh Dickins free_stable_node(stable_node); 6374035c07aSHugh Dickins } 6384035c07aSHugh Dickins 6392cee57d1SYang Shi enum get_ksm_page_flags { 6402cee57d1SYang Shi GET_KSM_PAGE_NOLOCK, 6412cee57d1SYang Shi GET_KSM_PAGE_LOCK, 6422cee57d1SYang Shi GET_KSM_PAGE_TRYLOCK 6432cee57d1SYang Shi }; 6442cee57d1SYang Shi 6454035c07aSHugh Dickins /* 6464035c07aSHugh Dickins * get_ksm_page: checks if the page indicated by the stable node 6474035c07aSHugh Dickins * is still its ksm page, despite having held no reference to it. 6484035c07aSHugh Dickins * In which case we can trust the content of the page, and it 6494035c07aSHugh Dickins * returns the gotten page; but if the page has now been zapped, 6504035c07aSHugh Dickins * remove the stale node from the stable tree and return NULL. 651c8d6553bSHugh Dickins * But beware, the stable node's page might be being migrated. 6524035c07aSHugh Dickins * 6534035c07aSHugh Dickins * You would expect the stable_node to hold a reference to the ksm page. 6544035c07aSHugh Dickins * But if it increments the page's count, swapping out has to wait for 6554035c07aSHugh Dickins * ksmd to come around again before it can free the page, which may take 6564035c07aSHugh Dickins * seconds or even minutes: much too unresponsive. So instead we use a 6574035c07aSHugh Dickins * "keyhole reference": access to the ksm page from the stable node peeps 6584035c07aSHugh Dickins * out through its keyhole to see if that page still holds the right key, 6594035c07aSHugh Dickins * pointing back to this stable node. This relies on freeing a PageAnon 6604035c07aSHugh Dickins * page to reset its page->mapping to NULL, and relies on no other use of 6614035c07aSHugh Dickins * a page to put something that might look like our key in page->mapping. 6624035c07aSHugh Dickins * is on its way to being freed; but it is an anomaly to bear in mind. 6634035c07aSHugh Dickins */ 66421fbd591SQi Zheng static struct page *get_ksm_page(struct ksm_stable_node *stable_node, 6652cee57d1SYang Shi enum get_ksm_page_flags flags) 6664035c07aSHugh Dickins { 6674035c07aSHugh Dickins struct page *page; 6684035c07aSHugh Dickins void *expected_mapping; 669c8d6553bSHugh Dickins unsigned long kpfn; 6704035c07aSHugh Dickins 671bda807d4SMinchan Kim expected_mapping = (void *)((unsigned long)stable_node | 672bda807d4SMinchan Kim PAGE_MAPPING_KSM); 673c8d6553bSHugh Dickins again: 67408df4774SPaul E. McKenney kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */ 675c8d6553bSHugh Dickins page = pfn_to_page(kpfn); 6764db0c3c2SJason Low if (READ_ONCE(page->mapping) != expected_mapping) 6774035c07aSHugh Dickins goto stale; 678c8d6553bSHugh Dickins 679c8d6553bSHugh Dickins /* 680c8d6553bSHugh Dickins * We cannot do anything with the page while its refcount is 0. 681c8d6553bSHugh Dickins * Usually 0 means free, or tail of a higher-order page: in which 682c8d6553bSHugh Dickins * case this node is no longer referenced, and should be freed; 6831c4c3b99SJiang Biao * however, it might mean that the page is under page_ref_freeze(). 684c8d6553bSHugh Dickins * The __remove_mapping() case is easy, again the node is now stale; 68552d1e606SKirill Tkhai * the same is in reuse_ksm_page() case; but if page is swapcache 6869800562fSMatthew Wilcox (Oracle) * in folio_migrate_mapping(), it might still be our page, 68752d1e606SKirill Tkhai * in which case it's essential to keep the node. 688c8d6553bSHugh Dickins */ 689c8d6553bSHugh Dickins while (!get_page_unless_zero(page)) { 690c8d6553bSHugh Dickins /* 691c8d6553bSHugh Dickins * Another check for page->mapping != expected_mapping would 692c8d6553bSHugh Dickins * work here too. We have chosen the !PageSwapCache test to 693c8d6553bSHugh Dickins * optimize the common case, when the page is or is about to 694c8d6553bSHugh Dickins * be freed: PageSwapCache is cleared (under spin_lock_irq) 6951c4c3b99SJiang Biao * in the ref_freeze section of __remove_mapping(); but Anon 696c8d6553bSHugh Dickins * page->mapping reset to NULL later, in free_pages_prepare(). 697c8d6553bSHugh Dickins */ 698c8d6553bSHugh Dickins if (!PageSwapCache(page)) 6994035c07aSHugh Dickins goto stale; 700c8d6553bSHugh Dickins cpu_relax(); 701c8d6553bSHugh Dickins } 702c8d6553bSHugh Dickins 7034db0c3c2SJason Low if (READ_ONCE(page->mapping) != expected_mapping) { 7044035c07aSHugh Dickins put_page(page); 7054035c07aSHugh Dickins goto stale; 7064035c07aSHugh Dickins } 707c8d6553bSHugh Dickins 7082cee57d1SYang Shi if (flags == GET_KSM_PAGE_TRYLOCK) { 7092cee57d1SYang Shi if (!trylock_page(page)) { 7102cee57d1SYang Shi put_page(page); 7112cee57d1SYang Shi return ERR_PTR(-EBUSY); 7122cee57d1SYang Shi } 7132cee57d1SYang Shi } else if (flags == GET_KSM_PAGE_LOCK) 7148aafa6a4SHugh Dickins lock_page(page); 7152cee57d1SYang Shi 7162cee57d1SYang Shi if (flags != GET_KSM_PAGE_NOLOCK) { 7174db0c3c2SJason Low if (READ_ONCE(page->mapping) != expected_mapping) { 7188aafa6a4SHugh Dickins unlock_page(page); 7198aafa6a4SHugh Dickins put_page(page); 7208aafa6a4SHugh Dickins goto stale; 7218aafa6a4SHugh Dickins } 7228aafa6a4SHugh Dickins } 7234035c07aSHugh Dickins return page; 724c8d6553bSHugh Dickins 7254035c07aSHugh Dickins stale: 726c8d6553bSHugh Dickins /* 727c8d6553bSHugh Dickins * We come here from above when page->mapping or !PageSwapCache 728c8d6553bSHugh Dickins * suggests that the node is stale; but it might be under migration. 72919138349SMatthew Wilcox (Oracle) * We need smp_rmb(), matching the smp_wmb() in folio_migrate_ksm(), 730c8d6553bSHugh Dickins * before checking whether node->kpfn has been changed. 731c8d6553bSHugh Dickins */ 732c8d6553bSHugh Dickins smp_rmb(); 7334db0c3c2SJason Low if (READ_ONCE(stable_node->kpfn) != kpfn) 734c8d6553bSHugh Dickins goto again; 7354035c07aSHugh Dickins remove_node_from_stable_tree(stable_node); 7364035c07aSHugh Dickins return NULL; 7374035c07aSHugh Dickins } 7384035c07aSHugh Dickins 73931dbd01fSIzik Eidus /* 74031dbd01fSIzik Eidus * Removing rmap_item from stable or unstable tree. 74131dbd01fSIzik Eidus * This function will clean the information from the stable/unstable tree. 74231dbd01fSIzik Eidus */ 74321fbd591SQi Zheng static void remove_rmap_item_from_tree(struct ksm_rmap_item *rmap_item) 74431dbd01fSIzik Eidus { 7457b6ba2c7SHugh Dickins if (rmap_item->address & STABLE_FLAG) { 74621fbd591SQi Zheng struct ksm_stable_node *stable_node; 7475ad64688SHugh Dickins struct page *page; 74831dbd01fSIzik Eidus 7497b6ba2c7SHugh Dickins stable_node = rmap_item->head; 75062862290SHugh Dickins page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); 7514035c07aSHugh Dickins if (!page) 7524035c07aSHugh Dickins goto out; 7535ad64688SHugh Dickins 7547b6ba2c7SHugh Dickins hlist_del(&rmap_item->hlist); 75562862290SHugh Dickins unlock_page(page); 7565ad64688SHugh Dickins put_page(page); 75708beca44SHugh Dickins 75898666f8aSAndrea Arcangeli if (!hlist_empty(&stable_node->hlist)) 7594035c07aSHugh Dickins ksm_pages_sharing--; 7604035c07aSHugh Dickins else 761b4028260SHugh Dickins ksm_pages_shared--; 76276093853Sxu xin 76376093853Sxu xin rmap_item->mm->ksm_merging_pages--; 76476093853Sxu xin 7652c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len <= 0); 7662c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len--; 76731dbd01fSIzik Eidus 7689e60109fSPeter Zijlstra put_anon_vma(rmap_item->anon_vma); 769c89a384eSMiaohe Lin rmap_item->head = NULL; 77093d17715SHugh Dickins rmap_item->address &= PAGE_MASK; 77131dbd01fSIzik Eidus 7727b6ba2c7SHugh Dickins } else if (rmap_item->address & UNSTABLE_FLAG) { 77331dbd01fSIzik Eidus unsigned char age; 77431dbd01fSIzik Eidus /* 7759ba69294SHugh Dickins * Usually ksmd can and must skip the rb_erase, because 77631dbd01fSIzik Eidus * root_unstable_tree was already reset to RB_ROOT. 7779ba69294SHugh Dickins * But be careful when an mm is exiting: do the rb_erase 7789ba69294SHugh Dickins * if this rmap_item was inserted by this scan, rather 7799ba69294SHugh Dickins * than left over from before. 78031dbd01fSIzik Eidus */ 78131dbd01fSIzik Eidus age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); 782cd551f97SHugh Dickins BUG_ON(age > 1); 78331dbd01fSIzik Eidus if (!age) 78490bd6fd3SPetr Holasek rb_erase(&rmap_item->node, 785ef53d16cSHugh Dickins root_unstable_tree + NUMA(rmap_item->nid)); 78693d17715SHugh Dickins ksm_pages_unshared--; 78731dbd01fSIzik Eidus rmap_item->address &= PAGE_MASK; 78893d17715SHugh Dickins } 7894035c07aSHugh Dickins out: 79031dbd01fSIzik Eidus cond_resched(); /* we're called from many long loops */ 79131dbd01fSIzik Eidus } 79231dbd01fSIzik Eidus 79321fbd591SQi Zheng static void remove_trailing_rmap_items(struct ksm_rmap_item **rmap_list) 79431dbd01fSIzik Eidus { 7956514d511SHugh Dickins while (*rmap_list) { 79621fbd591SQi Zheng struct ksm_rmap_item *rmap_item = *rmap_list; 7976514d511SHugh Dickins *rmap_list = rmap_item->rmap_list; 79831dbd01fSIzik Eidus remove_rmap_item_from_tree(rmap_item); 79931dbd01fSIzik Eidus free_rmap_item(rmap_item); 80031dbd01fSIzik Eidus } 80131dbd01fSIzik Eidus } 80231dbd01fSIzik Eidus 80331dbd01fSIzik Eidus /* 804e850dcf5SHugh Dickins * Though it's very tempting to unmerge rmap_items from stable tree rather 80531dbd01fSIzik Eidus * than check every pte of a given vma, the locking doesn't quite work for 80631dbd01fSIzik Eidus * that - an rmap_item is assigned to the stable tree after inserting ksm 807c1e8d7c6SMichel Lespinasse * page and upping mmap_lock. Nor does it fit with the way we skip dup'ing 80831dbd01fSIzik Eidus * rmap_items from parent to child at fork time (so as not to waste time 80931dbd01fSIzik Eidus * if exit comes before the next scan reaches it). 81081464e30SHugh Dickins * 81181464e30SHugh Dickins * Similarly, although we'd like to remove rmap_items (so updating counts 81281464e30SHugh Dickins * and freeing memory) when unmerging an area, it's easier to leave that 81381464e30SHugh Dickins * to the next pass of ksmd - consider, for example, how ksmd might be 81481464e30SHugh Dickins * in cmp_and_merge_page on one of the rmap_items we would be removing. 81531dbd01fSIzik Eidus */ 816d952b791SHugh Dickins static int unmerge_ksm_pages(struct vm_area_struct *vma, 81731dbd01fSIzik Eidus unsigned long start, unsigned long end) 81831dbd01fSIzik Eidus { 81931dbd01fSIzik Eidus unsigned long addr; 820d952b791SHugh Dickins int err = 0; 82131dbd01fSIzik Eidus 822d952b791SHugh Dickins for (addr = start; addr < end && !err; addr += PAGE_SIZE) { 8239ba69294SHugh Dickins if (ksm_test_exit(vma->vm_mm)) 8249ba69294SHugh Dickins break; 825d952b791SHugh Dickins if (signal_pending(current)) 826d952b791SHugh Dickins err = -ERESTARTSYS; 827d952b791SHugh Dickins else 828d952b791SHugh Dickins err = break_ksm(vma, addr); 829d952b791SHugh Dickins } 830d952b791SHugh Dickins return err; 83131dbd01fSIzik Eidus } 83231dbd01fSIzik Eidus 83321fbd591SQi Zheng static inline struct ksm_stable_node *folio_stable_node(struct folio *folio) 83419138349SMatthew Wilcox (Oracle) { 83519138349SMatthew Wilcox (Oracle) return folio_test_ksm(folio) ? folio_raw_mapping(folio) : NULL; 83619138349SMatthew Wilcox (Oracle) } 83719138349SMatthew Wilcox (Oracle) 83821fbd591SQi Zheng static inline struct ksm_stable_node *page_stable_node(struct page *page) 83988484826SMike Rapoport { 84019138349SMatthew Wilcox (Oracle) return folio_stable_node(page_folio(page)); 84188484826SMike Rapoport } 84288484826SMike Rapoport 84388484826SMike Rapoport static inline void set_page_stable_node(struct page *page, 84421fbd591SQi Zheng struct ksm_stable_node *stable_node) 84588484826SMike Rapoport { 8466c287605SDavid Hildenbrand VM_BUG_ON_PAGE(PageAnon(page) && PageAnonExclusive(page), page); 84788484826SMike Rapoport page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM); 84888484826SMike Rapoport } 84988484826SMike Rapoport 8502ffd8679SHugh Dickins #ifdef CONFIG_SYSFS 8512ffd8679SHugh Dickins /* 8522ffd8679SHugh Dickins * Only called through the sysfs control interface: 8532ffd8679SHugh Dickins */ 85421fbd591SQi Zheng static int remove_stable_node(struct ksm_stable_node *stable_node) 855cbf86cfeSHugh Dickins { 856cbf86cfeSHugh Dickins struct page *page; 857cbf86cfeSHugh Dickins int err; 858cbf86cfeSHugh Dickins 8592cee57d1SYang Shi page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); 860cbf86cfeSHugh Dickins if (!page) { 861cbf86cfeSHugh Dickins /* 862cbf86cfeSHugh Dickins * get_ksm_page did remove_node_from_stable_tree itself. 863cbf86cfeSHugh Dickins */ 864cbf86cfeSHugh Dickins return 0; 865cbf86cfeSHugh Dickins } 866cbf86cfeSHugh Dickins 867cbf86cfeSHugh Dickins /* 8689a63236fSAndrey Ryabinin * Page could be still mapped if this races with __mmput() running in 8699a63236fSAndrey Ryabinin * between ksm_exit() and exit_mmap(). Just refuse to let 8709a63236fSAndrey Ryabinin * merge_across_nodes/max_page_sharing be switched. 8718fdb3dbfSHugh Dickins */ 8728fdb3dbfSHugh Dickins err = -EBUSY; 8739a63236fSAndrey Ryabinin if (!page_mapped(page)) { 8748fdb3dbfSHugh Dickins /* 8758fdb3dbfSHugh Dickins * The stable node did not yet appear stale to get_ksm_page(), 8768fdb3dbfSHugh Dickins * since that allows for an unmapped ksm page to be recognized 8778fdb3dbfSHugh Dickins * right up until it is freed; but the node is safe to remove. 878cbf86cfeSHugh Dickins * This page might be in a pagevec waiting to be freed, 879cbf86cfeSHugh Dickins * or it might be PageSwapCache (perhaps under writeback), 880cbf86cfeSHugh Dickins * or it might have been removed from swapcache a moment ago. 881cbf86cfeSHugh Dickins */ 882cbf86cfeSHugh Dickins set_page_stable_node(page, NULL); 883cbf86cfeSHugh Dickins remove_node_from_stable_tree(stable_node); 884cbf86cfeSHugh Dickins err = 0; 885cbf86cfeSHugh Dickins } 886cbf86cfeSHugh Dickins 887cbf86cfeSHugh Dickins unlock_page(page); 888cbf86cfeSHugh Dickins put_page(page); 889cbf86cfeSHugh Dickins return err; 890cbf86cfeSHugh Dickins } 891cbf86cfeSHugh Dickins 89221fbd591SQi Zheng static int remove_stable_node_chain(struct ksm_stable_node *stable_node, 8932c653d0eSAndrea Arcangeli struct rb_root *root) 8942c653d0eSAndrea Arcangeli { 89521fbd591SQi Zheng struct ksm_stable_node *dup; 8962c653d0eSAndrea Arcangeli struct hlist_node *hlist_safe; 8972c653d0eSAndrea Arcangeli 8982c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 8992c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node)); 9002c653d0eSAndrea Arcangeli if (remove_stable_node(stable_node)) 9012c653d0eSAndrea Arcangeli return true; 9022c653d0eSAndrea Arcangeli else 9032c653d0eSAndrea Arcangeli return false; 9042c653d0eSAndrea Arcangeli } 9052c653d0eSAndrea Arcangeli 9062c653d0eSAndrea Arcangeli hlist_for_each_entry_safe(dup, hlist_safe, 9072c653d0eSAndrea Arcangeli &stable_node->hlist, hlist_dup) { 9082c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(dup)); 9092c653d0eSAndrea Arcangeli if (remove_stable_node(dup)) 9102c653d0eSAndrea Arcangeli return true; 9112c653d0eSAndrea Arcangeli } 9122c653d0eSAndrea Arcangeli BUG_ON(!hlist_empty(&stable_node->hlist)); 9132c653d0eSAndrea Arcangeli free_stable_node_chain(stable_node, root); 9142c653d0eSAndrea Arcangeli return false; 9152c653d0eSAndrea Arcangeli } 9162c653d0eSAndrea Arcangeli 917cbf86cfeSHugh Dickins static int remove_all_stable_nodes(void) 918cbf86cfeSHugh Dickins { 91921fbd591SQi Zheng struct ksm_stable_node *stable_node, *next; 920cbf86cfeSHugh Dickins int nid; 921cbf86cfeSHugh Dickins int err = 0; 922cbf86cfeSHugh Dickins 923ef53d16cSHugh Dickins for (nid = 0; nid < ksm_nr_node_ids; nid++) { 924cbf86cfeSHugh Dickins while (root_stable_tree[nid].rb_node) { 925cbf86cfeSHugh Dickins stable_node = rb_entry(root_stable_tree[nid].rb_node, 92621fbd591SQi Zheng struct ksm_stable_node, node); 9272c653d0eSAndrea Arcangeli if (remove_stable_node_chain(stable_node, 9282c653d0eSAndrea Arcangeli root_stable_tree + nid)) { 929cbf86cfeSHugh Dickins err = -EBUSY; 930cbf86cfeSHugh Dickins break; /* proceed to next nid */ 931cbf86cfeSHugh Dickins } 932cbf86cfeSHugh Dickins cond_resched(); 933cbf86cfeSHugh Dickins } 934cbf86cfeSHugh Dickins } 93503640418SGeliang Tang list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { 9364146d2d6SHugh Dickins if (remove_stable_node(stable_node)) 9374146d2d6SHugh Dickins err = -EBUSY; 9384146d2d6SHugh Dickins cond_resched(); 9394146d2d6SHugh Dickins } 940cbf86cfeSHugh Dickins return err; 941cbf86cfeSHugh Dickins } 942cbf86cfeSHugh Dickins 943d952b791SHugh Dickins static int unmerge_and_remove_all_rmap_items(void) 94431dbd01fSIzik Eidus { 94521fbd591SQi Zheng struct ksm_mm_slot *mm_slot; 94658730ab6SQi Zheng struct mm_slot *slot; 94731dbd01fSIzik Eidus struct mm_struct *mm; 94831dbd01fSIzik Eidus struct vm_area_struct *vma; 949d952b791SHugh Dickins int err = 0; 95031dbd01fSIzik Eidus 951d952b791SHugh Dickins spin_lock(&ksm_mmlist_lock); 95258730ab6SQi Zheng slot = list_entry(ksm_mm_head.slot.mm_node.next, 95358730ab6SQi Zheng struct mm_slot, mm_node); 95458730ab6SQi Zheng ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); 955d952b791SHugh Dickins spin_unlock(&ksm_mmlist_lock); 956d952b791SHugh Dickins 957a5f18ba0SMatthew Wilcox (Oracle) for (mm_slot = ksm_scan.mm_slot; mm_slot != &ksm_mm_head; 958a5f18ba0SMatthew Wilcox (Oracle) mm_slot = ksm_scan.mm_slot) { 95958730ab6SQi Zheng VMA_ITERATOR(vmi, mm_slot->slot.mm, 0); 960a5f18ba0SMatthew Wilcox (Oracle) 96158730ab6SQi Zheng mm = mm_slot->slot.mm; 962d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 963a5f18ba0SMatthew Wilcox (Oracle) for_each_vma(vmi, vma) { 9649ba69294SHugh Dickins if (ksm_test_exit(mm)) 9659ba69294SHugh Dickins break; 96631dbd01fSIzik Eidus if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) 96731dbd01fSIzik Eidus continue; 968d952b791SHugh Dickins err = unmerge_ksm_pages(vma, 969d952b791SHugh Dickins vma->vm_start, vma->vm_end); 9709ba69294SHugh Dickins if (err) 9719ba69294SHugh Dickins goto error; 972d952b791SHugh Dickins } 9739ba69294SHugh Dickins 974420be4edSChengyang Fan remove_trailing_rmap_items(&mm_slot->rmap_list); 975d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 97631dbd01fSIzik Eidus 97731dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 97858730ab6SQi Zheng slot = list_entry(mm_slot->slot.mm_node.next, 97958730ab6SQi Zheng struct mm_slot, mm_node); 98058730ab6SQi Zheng ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); 9819ba69294SHugh Dickins if (ksm_test_exit(mm)) { 98258730ab6SQi Zheng hash_del(&mm_slot->slot.hash); 98358730ab6SQi Zheng list_del(&mm_slot->slot.mm_node); 98431dbd01fSIzik Eidus spin_unlock(&ksm_mmlist_lock); 9859ba69294SHugh Dickins 98658730ab6SQi Zheng mm_slot_free(mm_slot_cache, mm_slot); 9879ba69294SHugh Dickins clear_bit(MMF_VM_MERGEABLE, &mm->flags); 9889ba69294SHugh Dickins mmdrop(mm); 9897496fea9SZhou Chengming } else 9909ba69294SHugh Dickins spin_unlock(&ksm_mmlist_lock); 99131dbd01fSIzik Eidus } 99231dbd01fSIzik Eidus 993cbf86cfeSHugh Dickins /* Clean up stable nodes, but don't worry if some are still busy */ 994cbf86cfeSHugh Dickins remove_all_stable_nodes(); 995d952b791SHugh Dickins ksm_scan.seqnr = 0; 9969ba69294SHugh Dickins return 0; 9979ba69294SHugh Dickins 9989ba69294SHugh Dickins error: 999d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1000d952b791SHugh Dickins spin_lock(&ksm_mmlist_lock); 1001d952b791SHugh Dickins ksm_scan.mm_slot = &ksm_mm_head; 1002d952b791SHugh Dickins spin_unlock(&ksm_mmlist_lock); 1003d952b791SHugh Dickins return err; 1004d952b791SHugh Dickins } 10052ffd8679SHugh Dickins #endif /* CONFIG_SYSFS */ 1006d952b791SHugh Dickins 100731dbd01fSIzik Eidus static u32 calc_checksum(struct page *page) 100831dbd01fSIzik Eidus { 100931dbd01fSIzik Eidus u32 checksum; 10109b04c5feSCong Wang void *addr = kmap_atomic(page); 101159e1a2f4STimofey Titovets checksum = xxhash(addr, PAGE_SIZE, 0); 10129b04c5feSCong Wang kunmap_atomic(addr); 101331dbd01fSIzik Eidus return checksum; 101431dbd01fSIzik Eidus } 101531dbd01fSIzik Eidus 101631dbd01fSIzik Eidus static int write_protect_page(struct vm_area_struct *vma, struct page *page, 101731dbd01fSIzik Eidus pte_t *orig_pte) 101831dbd01fSIzik Eidus { 101931dbd01fSIzik Eidus struct mm_struct *mm = vma->vm_mm; 1020eed05e54SMatthew Wilcox (Oracle) DEFINE_PAGE_VMA_WALK(pvmw, page, vma, 0, 0); 102131dbd01fSIzik Eidus int swapped; 102231dbd01fSIzik Eidus int err = -EFAULT; 1023ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 10246c287605SDavid Hildenbrand bool anon_exclusive; 102531dbd01fSIzik Eidus 102636eaff33SKirill A. Shutemov pvmw.address = page_address_in_vma(page, vma); 102736eaff33SKirill A. Shutemov if (pvmw.address == -EFAULT) 102831dbd01fSIzik Eidus goto out; 102931dbd01fSIzik Eidus 103029ad768cSAndrea Arcangeli BUG_ON(PageTransCompound(page)); 10316bdb913fSHaggai Eran 10327269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 10336f4f13e8SJérôme Glisse pvmw.address, 1034ac46d4f3SJérôme Glisse pvmw.address + PAGE_SIZE); 1035ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 10366bdb913fSHaggai Eran 103736eaff33SKirill A. Shutemov if (!page_vma_mapped_walk(&pvmw)) 10386bdb913fSHaggai Eran goto out_mn; 103936eaff33SKirill A. Shutemov if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?")) 104036eaff33SKirill A. Shutemov goto out_unlock; 104131dbd01fSIzik Eidus 10426c287605SDavid Hildenbrand anon_exclusive = PageAnonExclusive(page); 1043595cd8f2SAneesh Kumar K.V if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) || 1044b3a81d08SMinchan Kim (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) || 10456c287605SDavid Hildenbrand anon_exclusive || mm_tlb_flush_pending(mm)) { 104631dbd01fSIzik Eidus pte_t entry; 104731dbd01fSIzik Eidus 104831dbd01fSIzik Eidus swapped = PageSwapCache(page); 104936eaff33SKirill A. Shutemov flush_cache_page(vma, pvmw.address, page_to_pfn(page)); 105031dbd01fSIzik Eidus /* 105125985edcSLucas De Marchi * Ok this is tricky, when get_user_pages_fast() run it doesn't 105231dbd01fSIzik Eidus * take any lock, therefore the check that we are going to make 1053f0953a1bSIngo Molnar * with the pagecount against the mapcount is racy and 105431dbd01fSIzik Eidus * O_DIRECT can happen right after the check. 105531dbd01fSIzik Eidus * So we clear the pte and flush the tlb before the check 105631dbd01fSIzik Eidus * this assure us that no O_DIRECT can happen after the check 105731dbd01fSIzik Eidus * or in the middle of the check. 10580f10851eSJérôme Glisse * 10590f10851eSJérôme Glisse * No need to notify as we are downgrading page table to read 10600f10851eSJérôme Glisse * only not changing it to point to a new page. 10610f10851eSJérôme Glisse * 1062ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 106331dbd01fSIzik Eidus */ 10640f10851eSJérôme Glisse entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte); 106531dbd01fSIzik Eidus /* 106631dbd01fSIzik Eidus * Check that no O_DIRECT or similar I/O is in progress on the 106731dbd01fSIzik Eidus * page 106831dbd01fSIzik Eidus */ 106931e855eaSHugh Dickins if (page_mapcount(page) + 1 + swapped != page_count(page)) { 107036eaff33SKirill A. Shutemov set_pte_at(mm, pvmw.address, pvmw.pte, entry); 107131dbd01fSIzik Eidus goto out_unlock; 107231dbd01fSIzik Eidus } 10736c287605SDavid Hildenbrand 1074088b8aa5SDavid Hildenbrand /* See page_try_share_anon_rmap(): clear PTE first. */ 10756c287605SDavid Hildenbrand if (anon_exclusive && page_try_share_anon_rmap(page)) { 10766c287605SDavid Hildenbrand set_pte_at(mm, pvmw.address, pvmw.pte, entry); 10776c287605SDavid Hildenbrand goto out_unlock; 10786c287605SDavid Hildenbrand } 10796c287605SDavid Hildenbrand 10804e31635cSHugh Dickins if (pte_dirty(entry)) 10814e31635cSHugh Dickins set_page_dirty(page); 1082595cd8f2SAneesh Kumar K.V 1083595cd8f2SAneesh Kumar K.V if (pte_protnone(entry)) 1084595cd8f2SAneesh Kumar K.V entry = pte_mkclean(pte_clear_savedwrite(entry)); 1085595cd8f2SAneesh Kumar K.V else 10864e31635cSHugh Dickins entry = pte_mkclean(pte_wrprotect(entry)); 108736eaff33SKirill A. Shutemov set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry); 108831dbd01fSIzik Eidus } 108936eaff33SKirill A. Shutemov *orig_pte = *pvmw.pte; 109031dbd01fSIzik Eidus err = 0; 109131dbd01fSIzik Eidus 109231dbd01fSIzik Eidus out_unlock: 109336eaff33SKirill A. Shutemov page_vma_mapped_walk_done(&pvmw); 10946bdb913fSHaggai Eran out_mn: 1095ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 109631dbd01fSIzik Eidus out: 109731dbd01fSIzik Eidus return err; 109831dbd01fSIzik Eidus } 109931dbd01fSIzik Eidus 110031dbd01fSIzik Eidus /** 110131dbd01fSIzik Eidus * replace_page - replace page in vma by new ksm page 11028dd3557aSHugh Dickins * @vma: vma that holds the pte pointing to page 11038dd3557aSHugh Dickins * @page: the page we are replacing by kpage 11048dd3557aSHugh Dickins * @kpage: the ksm page we replace page by 110531dbd01fSIzik Eidus * @orig_pte: the original value of the pte 110631dbd01fSIzik Eidus * 110731dbd01fSIzik Eidus * Returns 0 on success, -EFAULT on failure. 110831dbd01fSIzik Eidus */ 11098dd3557aSHugh Dickins static int replace_page(struct vm_area_struct *vma, struct page *page, 11108dd3557aSHugh Dickins struct page *kpage, pte_t orig_pte) 111131dbd01fSIzik Eidus { 111231dbd01fSIzik Eidus struct mm_struct *mm = vma->vm_mm; 1113*b4e6f66eSMatthew Wilcox (Oracle) struct folio *folio; 111431dbd01fSIzik Eidus pmd_t *pmd; 111550722804SZach O'Keefe pmd_t pmde; 111631dbd01fSIzik Eidus pte_t *ptep; 1117e86c59b1SClaudio Imbrenda pte_t newpte; 111831dbd01fSIzik Eidus spinlock_t *ptl; 111931dbd01fSIzik Eidus unsigned long addr; 112031dbd01fSIzik Eidus int err = -EFAULT; 1121ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 112231dbd01fSIzik Eidus 11238dd3557aSHugh Dickins addr = page_address_in_vma(page, vma); 112431dbd01fSIzik Eidus if (addr == -EFAULT) 112531dbd01fSIzik Eidus goto out; 112631dbd01fSIzik Eidus 11276219049aSBob Liu pmd = mm_find_pmd(mm, addr); 11286219049aSBob Liu if (!pmd) 112931dbd01fSIzik Eidus goto out; 113050722804SZach O'Keefe /* 113150722804SZach O'Keefe * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at() 113250722804SZach O'Keefe * without holding anon_vma lock for write. So when looking for a 113350722804SZach O'Keefe * genuine pmde (in which to find pte), test present and !THP together. 113450722804SZach O'Keefe */ 113550722804SZach O'Keefe pmde = *pmd; 113650722804SZach O'Keefe barrier(); 113750722804SZach O'Keefe if (!pmd_present(pmde) || pmd_trans_huge(pmde)) 113850722804SZach O'Keefe goto out; 113931dbd01fSIzik Eidus 11407269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr, 11416f4f13e8SJérôme Glisse addr + PAGE_SIZE); 1142ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 11436bdb913fSHaggai Eran 114431dbd01fSIzik Eidus ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); 114531dbd01fSIzik Eidus if (!pte_same(*ptep, orig_pte)) { 114631dbd01fSIzik Eidus pte_unmap_unlock(ptep, ptl); 11476bdb913fSHaggai Eran goto out_mn; 114831dbd01fSIzik Eidus } 11496c287605SDavid Hildenbrand VM_BUG_ON_PAGE(PageAnonExclusive(page), page); 11506c287605SDavid Hildenbrand VM_BUG_ON_PAGE(PageAnon(kpage) && PageAnonExclusive(kpage), kpage); 115131dbd01fSIzik Eidus 1152e86c59b1SClaudio Imbrenda /* 1153e86c59b1SClaudio Imbrenda * No need to check ksm_use_zero_pages here: we can only have a 1154457aef94SEthon Paul * zero_page here if ksm_use_zero_pages was enabled already. 1155e86c59b1SClaudio Imbrenda */ 1156e86c59b1SClaudio Imbrenda if (!is_zero_pfn(page_to_pfn(kpage))) { 11578dd3557aSHugh Dickins get_page(kpage); 1158f1e2db12SDavid Hildenbrand page_add_anon_rmap(kpage, vma, addr, RMAP_NONE); 1159e86c59b1SClaudio Imbrenda newpte = mk_pte(kpage, vma->vm_page_prot); 1160e86c59b1SClaudio Imbrenda } else { 1161e86c59b1SClaudio Imbrenda newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage), 1162e86c59b1SClaudio Imbrenda vma->vm_page_prot)); 1163a38c015fSClaudio Imbrenda /* 1164a38c015fSClaudio Imbrenda * We're replacing an anonymous page with a zero page, which is 1165a38c015fSClaudio Imbrenda * not anonymous. We need to do proper accounting otherwise we 1166a38c015fSClaudio Imbrenda * will get wrong values in /proc, and a BUG message in dmesg 1167a38c015fSClaudio Imbrenda * when tearing down the mm. 1168a38c015fSClaudio Imbrenda */ 1169a38c015fSClaudio Imbrenda dec_mm_counter(mm, MM_ANONPAGES); 1170e86c59b1SClaudio Imbrenda } 117131dbd01fSIzik Eidus 117231dbd01fSIzik Eidus flush_cache_page(vma, addr, pte_pfn(*ptep)); 11730f10851eSJérôme Glisse /* 11740f10851eSJérôme Glisse * No need to notify as we are replacing a read only page with another 11750f10851eSJérôme Glisse * read only page with the same content. 11760f10851eSJérôme Glisse * 1177ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 11780f10851eSJérôme Glisse */ 11790f10851eSJérôme Glisse ptep_clear_flush(vma, addr, ptep); 1180e86c59b1SClaudio Imbrenda set_pte_at_notify(mm, addr, ptep, newpte); 118131dbd01fSIzik Eidus 1182*b4e6f66eSMatthew Wilcox (Oracle) folio = page_folio(page); 1183cea86fe2SHugh Dickins page_remove_rmap(page, vma, false); 1184*b4e6f66eSMatthew Wilcox (Oracle) if (!folio_mapped(folio)) 1185*b4e6f66eSMatthew Wilcox (Oracle) folio_free_swap(folio); 1186*b4e6f66eSMatthew Wilcox (Oracle) folio_put(folio); 118731dbd01fSIzik Eidus 118831dbd01fSIzik Eidus pte_unmap_unlock(ptep, ptl); 118931dbd01fSIzik Eidus err = 0; 11906bdb913fSHaggai Eran out_mn: 1191ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 119231dbd01fSIzik Eidus out: 119331dbd01fSIzik Eidus return err; 119431dbd01fSIzik Eidus } 119531dbd01fSIzik Eidus 119631dbd01fSIzik Eidus /* 119731dbd01fSIzik Eidus * try_to_merge_one_page - take two pages and merge them into one 11988dd3557aSHugh Dickins * @vma: the vma that holds the pte pointing to page 11998dd3557aSHugh Dickins * @page: the PageAnon page that we want to replace with kpage 120080e14822SHugh Dickins * @kpage: the PageKsm page that we want to map instead of page, 120180e14822SHugh Dickins * or NULL the first time when we want to use page as kpage. 120231dbd01fSIzik Eidus * 120331dbd01fSIzik Eidus * This function returns 0 if the pages were merged, -EFAULT otherwise. 120431dbd01fSIzik Eidus */ 120531dbd01fSIzik Eidus static int try_to_merge_one_page(struct vm_area_struct *vma, 12068dd3557aSHugh Dickins struct page *page, struct page *kpage) 120731dbd01fSIzik Eidus { 120831dbd01fSIzik Eidus pte_t orig_pte = __pte(0); 120931dbd01fSIzik Eidus int err = -EFAULT; 121031dbd01fSIzik Eidus 1211db114b83SHugh Dickins if (page == kpage) /* ksm page forked */ 1212db114b83SHugh Dickins return 0; 1213db114b83SHugh Dickins 12148dd3557aSHugh Dickins if (!PageAnon(page)) 121531dbd01fSIzik Eidus goto out; 121631dbd01fSIzik Eidus 121731dbd01fSIzik Eidus /* 121831dbd01fSIzik Eidus * We need the page lock to read a stable PageSwapCache in 121931dbd01fSIzik Eidus * write_protect_page(). We use trylock_page() instead of 122031dbd01fSIzik Eidus * lock_page() because we don't want to wait here - we 122131dbd01fSIzik Eidus * prefer to continue scanning and merging different pages, 122231dbd01fSIzik Eidus * then come back to this page when it is unlocked. 122331dbd01fSIzik Eidus */ 12248dd3557aSHugh Dickins if (!trylock_page(page)) 122531e855eaSHugh Dickins goto out; 1226f765f540SKirill A. Shutemov 1227f765f540SKirill A. Shutemov if (PageTransCompound(page)) { 1228a7306c34SAndrea Arcangeli if (split_huge_page(page)) 1229f765f540SKirill A. Shutemov goto out_unlock; 1230f765f540SKirill A. Shutemov } 1231f765f540SKirill A. Shutemov 123231dbd01fSIzik Eidus /* 123331dbd01fSIzik Eidus * If this anonymous page is mapped only here, its pte may need 123431dbd01fSIzik Eidus * to be write-protected. If it's mapped elsewhere, all of its 123531dbd01fSIzik Eidus * ptes are necessarily already write-protected. But in either 123631dbd01fSIzik Eidus * case, we need to lock and check page_count is not raised. 123731dbd01fSIzik Eidus */ 123880e14822SHugh Dickins if (write_protect_page(vma, page, &orig_pte) == 0) { 123980e14822SHugh Dickins if (!kpage) { 124080e14822SHugh Dickins /* 124180e14822SHugh Dickins * While we hold page lock, upgrade page from 124280e14822SHugh Dickins * PageAnon+anon_vma to PageKsm+NULL stable_node: 124380e14822SHugh Dickins * stable_tree_insert() will update stable_node. 124480e14822SHugh Dickins */ 124580e14822SHugh Dickins set_page_stable_node(page, NULL); 124680e14822SHugh Dickins mark_page_accessed(page); 1247337ed7ebSMinchan Kim /* 1248337ed7ebSMinchan Kim * Page reclaim just frees a clean page with no dirty 1249337ed7ebSMinchan Kim * ptes: make sure that the ksm page would be swapped. 1250337ed7ebSMinchan Kim */ 1251337ed7ebSMinchan Kim if (!PageDirty(page)) 1252337ed7ebSMinchan Kim SetPageDirty(page); 125380e14822SHugh Dickins err = 0; 125480e14822SHugh Dickins } else if (pages_identical(page, kpage)) 12558dd3557aSHugh Dickins err = replace_page(vma, page, kpage, orig_pte); 125680e14822SHugh Dickins } 125731dbd01fSIzik Eidus 1258f765f540SKirill A. Shutemov out_unlock: 12598dd3557aSHugh Dickins unlock_page(page); 126031dbd01fSIzik Eidus out: 126131dbd01fSIzik Eidus return err; 126231dbd01fSIzik Eidus } 126331dbd01fSIzik Eidus 126431dbd01fSIzik Eidus /* 126581464e30SHugh Dickins * try_to_merge_with_ksm_page - like try_to_merge_two_pages, 126681464e30SHugh Dickins * but no new kernel page is allocated: kpage must already be a ksm page. 12678dd3557aSHugh Dickins * 12688dd3557aSHugh Dickins * This function returns 0 if the pages were merged, -EFAULT otherwise. 126981464e30SHugh Dickins */ 127021fbd591SQi Zheng static int try_to_merge_with_ksm_page(struct ksm_rmap_item *rmap_item, 12718dd3557aSHugh Dickins struct page *page, struct page *kpage) 127281464e30SHugh Dickins { 12738dd3557aSHugh Dickins struct mm_struct *mm = rmap_item->mm; 127481464e30SHugh Dickins struct vm_area_struct *vma; 127581464e30SHugh Dickins int err = -EFAULT; 127681464e30SHugh Dickins 1277d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 127885c6e8ddSAndrea Arcangeli vma = find_mergeable_vma(mm, rmap_item->address); 127985c6e8ddSAndrea Arcangeli if (!vma) 12809ba69294SHugh Dickins goto out; 12819ba69294SHugh Dickins 12828dd3557aSHugh Dickins err = try_to_merge_one_page(vma, page, kpage); 1283db114b83SHugh Dickins if (err) 1284db114b83SHugh Dickins goto out; 1285db114b83SHugh Dickins 1286bc56620bSHugh Dickins /* Unstable nid is in union with stable anon_vma: remove first */ 1287bc56620bSHugh Dickins remove_rmap_item_from_tree(rmap_item); 1288bc56620bSHugh Dickins 1289c1e8d7c6SMichel Lespinasse /* Must get reference to anon_vma while still holding mmap_lock */ 12909e60109fSPeter Zijlstra rmap_item->anon_vma = vma->anon_vma; 12919e60109fSPeter Zijlstra get_anon_vma(vma->anon_vma); 129281464e30SHugh Dickins out: 1293d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 129481464e30SHugh Dickins return err; 129581464e30SHugh Dickins } 129681464e30SHugh Dickins 129781464e30SHugh Dickins /* 129831dbd01fSIzik Eidus * try_to_merge_two_pages - take two identical pages and prepare them 129931dbd01fSIzik Eidus * to be merged into one page. 130031dbd01fSIzik Eidus * 13018dd3557aSHugh Dickins * This function returns the kpage if we successfully merged two identical 13028dd3557aSHugh Dickins * pages into one ksm page, NULL otherwise. 130331dbd01fSIzik Eidus * 130480e14822SHugh Dickins * Note that this function upgrades page to ksm page: if one of the pages 130531dbd01fSIzik Eidus * is already a ksm page, try_to_merge_with_ksm_page should be used. 130631dbd01fSIzik Eidus */ 130721fbd591SQi Zheng static struct page *try_to_merge_two_pages(struct ksm_rmap_item *rmap_item, 13088dd3557aSHugh Dickins struct page *page, 130921fbd591SQi Zheng struct ksm_rmap_item *tree_rmap_item, 13108dd3557aSHugh Dickins struct page *tree_page) 131131dbd01fSIzik Eidus { 131280e14822SHugh Dickins int err; 131331dbd01fSIzik Eidus 131480e14822SHugh Dickins err = try_to_merge_with_ksm_page(rmap_item, page, NULL); 131531dbd01fSIzik Eidus if (!err) { 13168dd3557aSHugh Dickins err = try_to_merge_with_ksm_page(tree_rmap_item, 131780e14822SHugh Dickins tree_page, page); 131831dbd01fSIzik Eidus /* 131981464e30SHugh Dickins * If that fails, we have a ksm page with only one pte 132081464e30SHugh Dickins * pointing to it: so break it. 132131dbd01fSIzik Eidus */ 13224035c07aSHugh Dickins if (err) 13238dd3557aSHugh Dickins break_cow(rmap_item); 132431dbd01fSIzik Eidus } 132580e14822SHugh Dickins return err ? NULL : page; 132631dbd01fSIzik Eidus } 132731dbd01fSIzik Eidus 13282c653d0eSAndrea Arcangeli static __always_inline 132921fbd591SQi Zheng bool __is_page_sharing_candidate(struct ksm_stable_node *stable_node, int offset) 13302c653d0eSAndrea Arcangeli { 13312c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len < 0); 13322c653d0eSAndrea Arcangeli /* 13332c653d0eSAndrea Arcangeli * Check that at least one mapping still exists, otherwise 13342c653d0eSAndrea Arcangeli * there's no much point to merge and share with this 13352c653d0eSAndrea Arcangeli * stable_node, as the underlying tree_page of the other 13362c653d0eSAndrea Arcangeli * sharer is going to be freed soon. 13372c653d0eSAndrea Arcangeli */ 13382c653d0eSAndrea Arcangeli return stable_node->rmap_hlist_len && 13392c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len + offset < ksm_max_page_sharing; 13402c653d0eSAndrea Arcangeli } 13412c653d0eSAndrea Arcangeli 13422c653d0eSAndrea Arcangeli static __always_inline 134321fbd591SQi Zheng bool is_page_sharing_candidate(struct ksm_stable_node *stable_node) 13442c653d0eSAndrea Arcangeli { 13452c653d0eSAndrea Arcangeli return __is_page_sharing_candidate(stable_node, 0); 13462c653d0eSAndrea Arcangeli } 13472c653d0eSAndrea Arcangeli 134821fbd591SQi Zheng static struct page *stable_node_dup(struct ksm_stable_node **_stable_node_dup, 134921fbd591SQi Zheng struct ksm_stable_node **_stable_node, 13502c653d0eSAndrea Arcangeli struct rb_root *root, 13512c653d0eSAndrea Arcangeli bool prune_stale_stable_nodes) 13522c653d0eSAndrea Arcangeli { 135321fbd591SQi Zheng struct ksm_stable_node *dup, *found = NULL, *stable_node = *_stable_node; 13542c653d0eSAndrea Arcangeli struct hlist_node *hlist_safe; 13558dc5ffcdSAndrea Arcangeli struct page *_tree_page, *tree_page = NULL; 13562c653d0eSAndrea Arcangeli int nr = 0; 13572c653d0eSAndrea Arcangeli int found_rmap_hlist_len; 13582c653d0eSAndrea Arcangeli 13592c653d0eSAndrea Arcangeli if (!prune_stale_stable_nodes || 13602c653d0eSAndrea Arcangeli time_before(jiffies, stable_node->chain_prune_time + 13612c653d0eSAndrea Arcangeli msecs_to_jiffies( 13622c653d0eSAndrea Arcangeli ksm_stable_node_chains_prune_millisecs))) 13632c653d0eSAndrea Arcangeli prune_stale_stable_nodes = false; 13642c653d0eSAndrea Arcangeli else 13652c653d0eSAndrea Arcangeli stable_node->chain_prune_time = jiffies; 13662c653d0eSAndrea Arcangeli 13672c653d0eSAndrea Arcangeli hlist_for_each_entry_safe(dup, hlist_safe, 13682c653d0eSAndrea Arcangeli &stable_node->hlist, hlist_dup) { 13692c653d0eSAndrea Arcangeli cond_resched(); 13702c653d0eSAndrea Arcangeli /* 13712c653d0eSAndrea Arcangeli * We must walk all stable_node_dup to prune the stale 13722c653d0eSAndrea Arcangeli * stable nodes during lookup. 13732c653d0eSAndrea Arcangeli * 13742c653d0eSAndrea Arcangeli * get_ksm_page can drop the nodes from the 13752c653d0eSAndrea Arcangeli * stable_node->hlist if they point to freed pages 13762c653d0eSAndrea Arcangeli * (that's why we do a _safe walk). The "dup" 13772c653d0eSAndrea Arcangeli * stable_node parameter itself will be freed from 13782c653d0eSAndrea Arcangeli * under us if it returns NULL. 13792c653d0eSAndrea Arcangeli */ 13802cee57d1SYang Shi _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK); 13812c653d0eSAndrea Arcangeli if (!_tree_page) 13822c653d0eSAndrea Arcangeli continue; 13832c653d0eSAndrea Arcangeli nr += 1; 13842c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(dup)) { 13852c653d0eSAndrea Arcangeli if (!found || 13862c653d0eSAndrea Arcangeli dup->rmap_hlist_len > found_rmap_hlist_len) { 13872c653d0eSAndrea Arcangeli if (found) 13888dc5ffcdSAndrea Arcangeli put_page(tree_page); 13892c653d0eSAndrea Arcangeli found = dup; 13902c653d0eSAndrea Arcangeli found_rmap_hlist_len = found->rmap_hlist_len; 13918dc5ffcdSAndrea Arcangeli tree_page = _tree_page; 13922c653d0eSAndrea Arcangeli 13938dc5ffcdSAndrea Arcangeli /* skip put_page for found dup */ 13942c653d0eSAndrea Arcangeli if (!prune_stale_stable_nodes) 13952c653d0eSAndrea Arcangeli break; 13962c653d0eSAndrea Arcangeli continue; 13972c653d0eSAndrea Arcangeli } 13982c653d0eSAndrea Arcangeli } 13992c653d0eSAndrea Arcangeli put_page(_tree_page); 14002c653d0eSAndrea Arcangeli } 14012c653d0eSAndrea Arcangeli 140280b18dfaSAndrea Arcangeli if (found) { 14032c653d0eSAndrea Arcangeli /* 140480b18dfaSAndrea Arcangeli * nr is counting all dups in the chain only if 140580b18dfaSAndrea Arcangeli * prune_stale_stable_nodes is true, otherwise we may 140680b18dfaSAndrea Arcangeli * break the loop at nr == 1 even if there are 140780b18dfaSAndrea Arcangeli * multiple entries. 14082c653d0eSAndrea Arcangeli */ 140980b18dfaSAndrea Arcangeli if (prune_stale_stable_nodes && nr == 1) { 14102c653d0eSAndrea Arcangeli /* 14112c653d0eSAndrea Arcangeli * If there's not just one entry it would 14122c653d0eSAndrea Arcangeli * corrupt memory, better BUG_ON. In KSM 14132c653d0eSAndrea Arcangeli * context with no lock held it's not even 14142c653d0eSAndrea Arcangeli * fatal. 14152c653d0eSAndrea Arcangeli */ 14162c653d0eSAndrea Arcangeli BUG_ON(stable_node->hlist.first->next); 14172c653d0eSAndrea Arcangeli 14182c653d0eSAndrea Arcangeli /* 14192c653d0eSAndrea Arcangeli * There's just one entry and it is below the 14202c653d0eSAndrea Arcangeli * deduplication limit so drop the chain. 14212c653d0eSAndrea Arcangeli */ 14222c653d0eSAndrea Arcangeli rb_replace_node(&stable_node->node, &found->node, 14232c653d0eSAndrea Arcangeli root); 14242c653d0eSAndrea Arcangeli free_stable_node(stable_node); 14252c653d0eSAndrea Arcangeli ksm_stable_node_chains--; 14262c653d0eSAndrea Arcangeli ksm_stable_node_dups--; 1427b4fecc67SAndrea Arcangeli /* 14280ba1d0f7SAndrea Arcangeli * NOTE: the caller depends on the stable_node 14290ba1d0f7SAndrea Arcangeli * to be equal to stable_node_dup if the chain 14300ba1d0f7SAndrea Arcangeli * was collapsed. 1431b4fecc67SAndrea Arcangeli */ 14320ba1d0f7SAndrea Arcangeli *_stable_node = found; 14330ba1d0f7SAndrea Arcangeli /* 1434f0953a1bSIngo Molnar * Just for robustness, as stable_node is 14350ba1d0f7SAndrea Arcangeli * otherwise left as a stable pointer, the 14360ba1d0f7SAndrea Arcangeli * compiler shall optimize it away at build 14370ba1d0f7SAndrea Arcangeli * time. 14380ba1d0f7SAndrea Arcangeli */ 14390ba1d0f7SAndrea Arcangeli stable_node = NULL; 144080b18dfaSAndrea Arcangeli } else if (stable_node->hlist.first != &found->hlist_dup && 144180b18dfaSAndrea Arcangeli __is_page_sharing_candidate(found, 1)) { 14422c653d0eSAndrea Arcangeli /* 144380b18dfaSAndrea Arcangeli * If the found stable_node dup can accept one 144480b18dfaSAndrea Arcangeli * more future merge (in addition to the one 144580b18dfaSAndrea Arcangeli * that is underway) and is not at the head of 144680b18dfaSAndrea Arcangeli * the chain, put it there so next search will 144780b18dfaSAndrea Arcangeli * be quicker in the !prune_stale_stable_nodes 144880b18dfaSAndrea Arcangeli * case. 144980b18dfaSAndrea Arcangeli * 145080b18dfaSAndrea Arcangeli * NOTE: it would be inaccurate to use nr > 1 145180b18dfaSAndrea Arcangeli * instead of checking the hlist.first pointer 145280b18dfaSAndrea Arcangeli * directly, because in the 145380b18dfaSAndrea Arcangeli * prune_stale_stable_nodes case "nr" isn't 145480b18dfaSAndrea Arcangeli * the position of the found dup in the chain, 145580b18dfaSAndrea Arcangeli * but the total number of dups in the chain. 14562c653d0eSAndrea Arcangeli */ 14572c653d0eSAndrea Arcangeli hlist_del(&found->hlist_dup); 14582c653d0eSAndrea Arcangeli hlist_add_head(&found->hlist_dup, 14592c653d0eSAndrea Arcangeli &stable_node->hlist); 14602c653d0eSAndrea Arcangeli } 14612c653d0eSAndrea Arcangeli } 14622c653d0eSAndrea Arcangeli 14638dc5ffcdSAndrea Arcangeli *_stable_node_dup = found; 14648dc5ffcdSAndrea Arcangeli return tree_page; 14652c653d0eSAndrea Arcangeli } 14662c653d0eSAndrea Arcangeli 146721fbd591SQi Zheng static struct ksm_stable_node *stable_node_dup_any(struct ksm_stable_node *stable_node, 14682c653d0eSAndrea Arcangeli struct rb_root *root) 14692c653d0eSAndrea Arcangeli { 14702c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) 14712c653d0eSAndrea Arcangeli return stable_node; 14722c653d0eSAndrea Arcangeli if (hlist_empty(&stable_node->hlist)) { 14732c653d0eSAndrea Arcangeli free_stable_node_chain(stable_node, root); 14742c653d0eSAndrea Arcangeli return NULL; 14752c653d0eSAndrea Arcangeli } 14762c653d0eSAndrea Arcangeli return hlist_entry(stable_node->hlist.first, 14772c653d0eSAndrea Arcangeli typeof(*stable_node), hlist_dup); 14782c653d0eSAndrea Arcangeli } 14792c653d0eSAndrea Arcangeli 14808dc5ffcdSAndrea Arcangeli /* 14818dc5ffcdSAndrea Arcangeli * Like for get_ksm_page, this function can free the *_stable_node and 14828dc5ffcdSAndrea Arcangeli * *_stable_node_dup if the returned tree_page is NULL. 14838dc5ffcdSAndrea Arcangeli * 14848dc5ffcdSAndrea Arcangeli * It can also free and overwrite *_stable_node with the found 14858dc5ffcdSAndrea Arcangeli * stable_node_dup if the chain is collapsed (in which case 14868dc5ffcdSAndrea Arcangeli * *_stable_node will be equal to *_stable_node_dup like if the chain 14878dc5ffcdSAndrea Arcangeli * never existed). It's up to the caller to verify tree_page is not 14888dc5ffcdSAndrea Arcangeli * NULL before dereferencing *_stable_node or *_stable_node_dup. 14898dc5ffcdSAndrea Arcangeli * 14908dc5ffcdSAndrea Arcangeli * *_stable_node_dup is really a second output parameter of this 14918dc5ffcdSAndrea Arcangeli * function and will be overwritten in all cases, the caller doesn't 14928dc5ffcdSAndrea Arcangeli * need to initialize it. 14938dc5ffcdSAndrea Arcangeli */ 149421fbd591SQi Zheng static struct page *__stable_node_chain(struct ksm_stable_node **_stable_node_dup, 149521fbd591SQi Zheng struct ksm_stable_node **_stable_node, 14962c653d0eSAndrea Arcangeli struct rb_root *root, 14972c653d0eSAndrea Arcangeli bool prune_stale_stable_nodes) 14982c653d0eSAndrea Arcangeli { 149921fbd591SQi Zheng struct ksm_stable_node *stable_node = *_stable_node; 15002c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 15012c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(stable_node)) { 15028dc5ffcdSAndrea Arcangeli *_stable_node_dup = stable_node; 15032cee57d1SYang Shi return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK); 15042c653d0eSAndrea Arcangeli } 15058dc5ffcdSAndrea Arcangeli /* 15068dc5ffcdSAndrea Arcangeli * _stable_node_dup set to NULL means the stable_node 15078dc5ffcdSAndrea Arcangeli * reached the ksm_max_page_sharing limit. 15088dc5ffcdSAndrea Arcangeli */ 15098dc5ffcdSAndrea Arcangeli *_stable_node_dup = NULL; 15102c653d0eSAndrea Arcangeli return NULL; 15112c653d0eSAndrea Arcangeli } 15128dc5ffcdSAndrea Arcangeli return stable_node_dup(_stable_node_dup, _stable_node, root, 15132c653d0eSAndrea Arcangeli prune_stale_stable_nodes); 15142c653d0eSAndrea Arcangeli } 15152c653d0eSAndrea Arcangeli 151621fbd591SQi Zheng static __always_inline struct page *chain_prune(struct ksm_stable_node **s_n_d, 151721fbd591SQi Zheng struct ksm_stable_node **s_n, 15182c653d0eSAndrea Arcangeli struct rb_root *root) 15192c653d0eSAndrea Arcangeli { 15208dc5ffcdSAndrea Arcangeli return __stable_node_chain(s_n_d, s_n, root, true); 15212c653d0eSAndrea Arcangeli } 15222c653d0eSAndrea Arcangeli 152321fbd591SQi Zheng static __always_inline struct page *chain(struct ksm_stable_node **s_n_d, 152421fbd591SQi Zheng struct ksm_stable_node *s_n, 15252c653d0eSAndrea Arcangeli struct rb_root *root) 15262c653d0eSAndrea Arcangeli { 152721fbd591SQi Zheng struct ksm_stable_node *old_stable_node = s_n; 15288dc5ffcdSAndrea Arcangeli struct page *tree_page; 15298dc5ffcdSAndrea Arcangeli 15308dc5ffcdSAndrea Arcangeli tree_page = __stable_node_chain(s_n_d, &s_n, root, false); 15318dc5ffcdSAndrea Arcangeli /* not pruning dups so s_n cannot have changed */ 15328dc5ffcdSAndrea Arcangeli VM_BUG_ON(s_n != old_stable_node); 15338dc5ffcdSAndrea Arcangeli return tree_page; 15342c653d0eSAndrea Arcangeli } 15352c653d0eSAndrea Arcangeli 153631dbd01fSIzik Eidus /* 15378dd3557aSHugh Dickins * stable_tree_search - search for page inside the stable tree 153831dbd01fSIzik Eidus * 153931dbd01fSIzik Eidus * This function checks if there is a page inside the stable tree 154031dbd01fSIzik Eidus * with identical content to the page that we are scanning right now. 154131dbd01fSIzik Eidus * 15427b6ba2c7SHugh Dickins * This function returns the stable tree node of identical content if found, 154331dbd01fSIzik Eidus * NULL otherwise. 154431dbd01fSIzik Eidus */ 154562b61f61SHugh Dickins static struct page *stable_tree_search(struct page *page) 154631dbd01fSIzik Eidus { 154790bd6fd3SPetr Holasek int nid; 1548ef53d16cSHugh Dickins struct rb_root *root; 15494146d2d6SHugh Dickins struct rb_node **new; 15504146d2d6SHugh Dickins struct rb_node *parent; 155121fbd591SQi Zheng struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; 155221fbd591SQi Zheng struct ksm_stable_node *page_node; 155331dbd01fSIzik Eidus 15544146d2d6SHugh Dickins page_node = page_stable_node(page); 15554146d2d6SHugh Dickins if (page_node && page_node->head != &migrate_nodes) { 15564146d2d6SHugh Dickins /* ksm page forked */ 155708beca44SHugh Dickins get_page(page); 155862b61f61SHugh Dickins return page; 155908beca44SHugh Dickins } 156008beca44SHugh Dickins 156190bd6fd3SPetr Holasek nid = get_kpfn_nid(page_to_pfn(page)); 1562ef53d16cSHugh Dickins root = root_stable_tree + nid; 15634146d2d6SHugh Dickins again: 1564ef53d16cSHugh Dickins new = &root->rb_node; 15654146d2d6SHugh Dickins parent = NULL; 156690bd6fd3SPetr Holasek 15674146d2d6SHugh Dickins while (*new) { 15684035c07aSHugh Dickins struct page *tree_page; 156931dbd01fSIzik Eidus int ret; 157031dbd01fSIzik Eidus 157131dbd01fSIzik Eidus cond_resched(); 157221fbd591SQi Zheng stable_node = rb_entry(*new, struct ksm_stable_node, node); 15732c653d0eSAndrea Arcangeli stable_node_any = NULL; 15748dc5ffcdSAndrea Arcangeli tree_page = chain_prune(&stable_node_dup, &stable_node, root); 1575b4fecc67SAndrea Arcangeli /* 1576b4fecc67SAndrea Arcangeli * NOTE: stable_node may have been freed by 1577b4fecc67SAndrea Arcangeli * chain_prune() if the returned stable_node_dup is 1578b4fecc67SAndrea Arcangeli * not NULL. stable_node_dup may have been inserted in 1579b4fecc67SAndrea Arcangeli * the rbtree instead as a regular stable_node (in 1580b4fecc67SAndrea Arcangeli * order to collapse the stable_node chain if a single 15810ba1d0f7SAndrea Arcangeli * stable_node dup was found in it). In such case the 15823413b2c8SJulia Lawall * stable_node is overwritten by the callee to point 15830ba1d0f7SAndrea Arcangeli * to the stable_node_dup that was collapsed in the 15840ba1d0f7SAndrea Arcangeli * stable rbtree and stable_node will be equal to 15850ba1d0f7SAndrea Arcangeli * stable_node_dup like if the chain never existed. 1586b4fecc67SAndrea Arcangeli */ 15872c653d0eSAndrea Arcangeli if (!stable_node_dup) { 15882c653d0eSAndrea Arcangeli /* 15892c653d0eSAndrea Arcangeli * Either all stable_node dups were full in 15902c653d0eSAndrea Arcangeli * this stable_node chain, or this chain was 15912c653d0eSAndrea Arcangeli * empty and should be rb_erased. 15922c653d0eSAndrea Arcangeli */ 15932c653d0eSAndrea Arcangeli stable_node_any = stable_node_dup_any(stable_node, 15942c653d0eSAndrea Arcangeli root); 15952c653d0eSAndrea Arcangeli if (!stable_node_any) { 15962c653d0eSAndrea Arcangeli /* rb_erase just run */ 15972c653d0eSAndrea Arcangeli goto again; 15982c653d0eSAndrea Arcangeli } 15992c653d0eSAndrea Arcangeli /* 16002c653d0eSAndrea Arcangeli * Take any of the stable_node dups page of 16012c653d0eSAndrea Arcangeli * this stable_node chain to let the tree walk 16022c653d0eSAndrea Arcangeli * continue. All KSM pages belonging to the 16032c653d0eSAndrea Arcangeli * stable_node dups in a stable_node chain 16042c653d0eSAndrea Arcangeli * have the same content and they're 1605457aef94SEthon Paul * write protected at all times. Any will work 16062c653d0eSAndrea Arcangeli * fine to continue the walk. 16072c653d0eSAndrea Arcangeli */ 16082cee57d1SYang Shi tree_page = get_ksm_page(stable_node_any, 16092cee57d1SYang Shi GET_KSM_PAGE_NOLOCK); 16102c653d0eSAndrea Arcangeli } 16112c653d0eSAndrea Arcangeli VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); 1612f2e5ff85SAndrea Arcangeli if (!tree_page) { 1613f2e5ff85SAndrea Arcangeli /* 1614f2e5ff85SAndrea Arcangeli * If we walked over a stale stable_node, 1615f2e5ff85SAndrea Arcangeli * get_ksm_page() will call rb_erase() and it 1616f2e5ff85SAndrea Arcangeli * may rebalance the tree from under us. So 1617f2e5ff85SAndrea Arcangeli * restart the search from scratch. Returning 1618f2e5ff85SAndrea Arcangeli * NULL would be safe too, but we'd generate 1619f2e5ff85SAndrea Arcangeli * false negative insertions just because some 1620f2e5ff85SAndrea Arcangeli * stable_node was stale. 1621f2e5ff85SAndrea Arcangeli */ 1622f2e5ff85SAndrea Arcangeli goto again; 1623f2e5ff85SAndrea Arcangeli } 162431dbd01fSIzik Eidus 16254035c07aSHugh Dickins ret = memcmp_pages(page, tree_page); 1626c8d6553bSHugh Dickins put_page(tree_page); 162731dbd01fSIzik Eidus 16284146d2d6SHugh Dickins parent = *new; 1629c8d6553bSHugh Dickins if (ret < 0) 16304146d2d6SHugh Dickins new = &parent->rb_left; 1631c8d6553bSHugh Dickins else if (ret > 0) 16324146d2d6SHugh Dickins new = &parent->rb_right; 1633c8d6553bSHugh Dickins else { 16342c653d0eSAndrea Arcangeli if (page_node) { 16352c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 16362c653d0eSAndrea Arcangeli /* 16372c653d0eSAndrea Arcangeli * Test if the migrated page should be merged 16382c653d0eSAndrea Arcangeli * into a stable node dup. If the mapcount is 16392c653d0eSAndrea Arcangeli * 1 we can migrate it with another KSM page 16402c653d0eSAndrea Arcangeli * without adding it to the chain. 16412c653d0eSAndrea Arcangeli */ 16422c653d0eSAndrea Arcangeli if (page_mapcount(page) > 1) 16432c653d0eSAndrea Arcangeli goto chain_append; 16442c653d0eSAndrea Arcangeli } 16452c653d0eSAndrea Arcangeli 16462c653d0eSAndrea Arcangeli if (!stable_node_dup) { 16472c653d0eSAndrea Arcangeli /* 16482c653d0eSAndrea Arcangeli * If the stable_node is a chain and 16492c653d0eSAndrea Arcangeli * we got a payload match in memcmp 16502c653d0eSAndrea Arcangeli * but we cannot merge the scanned 16512c653d0eSAndrea Arcangeli * page in any of the existing 16522c653d0eSAndrea Arcangeli * stable_node dups because they're 16532c653d0eSAndrea Arcangeli * all full, we need to wait the 16542c653d0eSAndrea Arcangeli * scanned page to find itself a match 16552c653d0eSAndrea Arcangeli * in the unstable tree to create a 16562c653d0eSAndrea Arcangeli * brand new KSM page to add later to 16572c653d0eSAndrea Arcangeli * the dups of this stable_node. 16582c653d0eSAndrea Arcangeli */ 16592c653d0eSAndrea Arcangeli return NULL; 16602c653d0eSAndrea Arcangeli } 16612c653d0eSAndrea Arcangeli 1662c8d6553bSHugh Dickins /* 1663c8d6553bSHugh Dickins * Lock and unlock the stable_node's page (which 1664c8d6553bSHugh Dickins * might already have been migrated) so that page 1665c8d6553bSHugh Dickins * migration is sure to notice its raised count. 1666c8d6553bSHugh Dickins * It would be more elegant to return stable_node 1667c8d6553bSHugh Dickins * than kpage, but that involves more changes. 1668c8d6553bSHugh Dickins */ 16692cee57d1SYang Shi tree_page = get_ksm_page(stable_node_dup, 16702cee57d1SYang Shi GET_KSM_PAGE_TRYLOCK); 16712cee57d1SYang Shi 16722cee57d1SYang Shi if (PTR_ERR(tree_page) == -EBUSY) 16732cee57d1SYang Shi return ERR_PTR(-EBUSY); 16742cee57d1SYang Shi 16752c653d0eSAndrea Arcangeli if (unlikely(!tree_page)) 16762c653d0eSAndrea Arcangeli /* 16772c653d0eSAndrea Arcangeli * The tree may have been rebalanced, 16782c653d0eSAndrea Arcangeli * so re-evaluate parent and new. 16792c653d0eSAndrea Arcangeli */ 16802c653d0eSAndrea Arcangeli goto again; 1681c8d6553bSHugh Dickins unlock_page(tree_page); 16822c653d0eSAndrea Arcangeli 16832c653d0eSAndrea Arcangeli if (get_kpfn_nid(stable_node_dup->kpfn) != 16842c653d0eSAndrea Arcangeli NUMA(stable_node_dup->nid)) { 16854146d2d6SHugh Dickins put_page(tree_page); 16864146d2d6SHugh Dickins goto replace; 16874146d2d6SHugh Dickins } 168862b61f61SHugh Dickins return tree_page; 168931dbd01fSIzik Eidus } 1690c8d6553bSHugh Dickins } 169131dbd01fSIzik Eidus 16924146d2d6SHugh Dickins if (!page_node) 169331dbd01fSIzik Eidus return NULL; 16944146d2d6SHugh Dickins 16954146d2d6SHugh Dickins list_del(&page_node->list); 16964146d2d6SHugh Dickins DO_NUMA(page_node->nid = nid); 16974146d2d6SHugh Dickins rb_link_node(&page_node->node, parent, new); 1698ef53d16cSHugh Dickins rb_insert_color(&page_node->node, root); 16992c653d0eSAndrea Arcangeli out: 17002c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(page_node)) { 17014146d2d6SHugh Dickins get_page(page); 17024146d2d6SHugh Dickins return page; 17032c653d0eSAndrea Arcangeli } else 17042c653d0eSAndrea Arcangeli return NULL; 17054146d2d6SHugh Dickins 17064146d2d6SHugh Dickins replace: 1707b4fecc67SAndrea Arcangeli /* 1708b4fecc67SAndrea Arcangeli * If stable_node was a chain and chain_prune collapsed it, 17090ba1d0f7SAndrea Arcangeli * stable_node has been updated to be the new regular 17100ba1d0f7SAndrea Arcangeli * stable_node. A collapse of the chain is indistinguishable 17110ba1d0f7SAndrea Arcangeli * from the case there was no chain in the stable 17120ba1d0f7SAndrea Arcangeli * rbtree. Otherwise stable_node is the chain and 17130ba1d0f7SAndrea Arcangeli * stable_node_dup is the dup to replace. 1714b4fecc67SAndrea Arcangeli */ 17150ba1d0f7SAndrea Arcangeli if (stable_node_dup == stable_node) { 1716b4fecc67SAndrea Arcangeli VM_BUG_ON(is_stable_node_chain(stable_node_dup)); 1717b4fecc67SAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node_dup)); 17182c653d0eSAndrea Arcangeli /* there is no chain */ 17194146d2d6SHugh Dickins if (page_node) { 17202c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 17214146d2d6SHugh Dickins list_del(&page_node->list); 17224146d2d6SHugh Dickins DO_NUMA(page_node->nid = nid); 1723b4fecc67SAndrea Arcangeli rb_replace_node(&stable_node_dup->node, 1724b4fecc67SAndrea Arcangeli &page_node->node, 17252c653d0eSAndrea Arcangeli root); 17262c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(page_node)) 17274146d2d6SHugh Dickins get_page(page); 17282c653d0eSAndrea Arcangeli else 17292c653d0eSAndrea Arcangeli page = NULL; 17304146d2d6SHugh Dickins } else { 1731b4fecc67SAndrea Arcangeli rb_erase(&stable_node_dup->node, root); 17324146d2d6SHugh Dickins page = NULL; 17334146d2d6SHugh Dickins } 17342c653d0eSAndrea Arcangeli } else { 17352c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_chain(stable_node)); 17362c653d0eSAndrea Arcangeli __stable_node_dup_del(stable_node_dup); 17372c653d0eSAndrea Arcangeli if (page_node) { 17382c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 17392c653d0eSAndrea Arcangeli list_del(&page_node->list); 17402c653d0eSAndrea Arcangeli DO_NUMA(page_node->nid = nid); 17412c653d0eSAndrea Arcangeli stable_node_chain_add_dup(page_node, stable_node); 17422c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(page_node)) 17432c653d0eSAndrea Arcangeli get_page(page); 17442c653d0eSAndrea Arcangeli else 17452c653d0eSAndrea Arcangeli page = NULL; 17462c653d0eSAndrea Arcangeli } else { 17472c653d0eSAndrea Arcangeli page = NULL; 17482c653d0eSAndrea Arcangeli } 17492c653d0eSAndrea Arcangeli } 17502c653d0eSAndrea Arcangeli stable_node_dup->head = &migrate_nodes; 17512c653d0eSAndrea Arcangeli list_add(&stable_node_dup->list, stable_node_dup->head); 17524146d2d6SHugh Dickins return page; 17532c653d0eSAndrea Arcangeli 17542c653d0eSAndrea Arcangeli chain_append: 17552c653d0eSAndrea Arcangeli /* stable_node_dup could be null if it reached the limit */ 17562c653d0eSAndrea Arcangeli if (!stable_node_dup) 17572c653d0eSAndrea Arcangeli stable_node_dup = stable_node_any; 1758b4fecc67SAndrea Arcangeli /* 1759b4fecc67SAndrea Arcangeli * If stable_node was a chain and chain_prune collapsed it, 17600ba1d0f7SAndrea Arcangeli * stable_node has been updated to be the new regular 17610ba1d0f7SAndrea Arcangeli * stable_node. A collapse of the chain is indistinguishable 17620ba1d0f7SAndrea Arcangeli * from the case there was no chain in the stable 17630ba1d0f7SAndrea Arcangeli * rbtree. Otherwise stable_node is the chain and 17640ba1d0f7SAndrea Arcangeli * stable_node_dup is the dup to replace. 1765b4fecc67SAndrea Arcangeli */ 17660ba1d0f7SAndrea Arcangeli if (stable_node_dup == stable_node) { 1767b4fecc67SAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node_dup)); 17682c653d0eSAndrea Arcangeli /* chain is missing so create it */ 17692c653d0eSAndrea Arcangeli stable_node = alloc_stable_node_chain(stable_node_dup, 17702c653d0eSAndrea Arcangeli root); 17712c653d0eSAndrea Arcangeli if (!stable_node) 17722c653d0eSAndrea Arcangeli return NULL; 17732c653d0eSAndrea Arcangeli } 17742c653d0eSAndrea Arcangeli /* 17752c653d0eSAndrea Arcangeli * Add this stable_node dup that was 17762c653d0eSAndrea Arcangeli * migrated to the stable_node chain 17772c653d0eSAndrea Arcangeli * of the current nid for this page 17782c653d0eSAndrea Arcangeli * content. 17792c653d0eSAndrea Arcangeli */ 1780b4fecc67SAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(stable_node_dup)); 17812c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 17822c653d0eSAndrea Arcangeli list_del(&page_node->list); 17832c653d0eSAndrea Arcangeli DO_NUMA(page_node->nid = nid); 17842c653d0eSAndrea Arcangeli stable_node_chain_add_dup(page_node, stable_node); 17852c653d0eSAndrea Arcangeli goto out; 178631dbd01fSIzik Eidus } 178731dbd01fSIzik Eidus 178831dbd01fSIzik Eidus /* 1789e850dcf5SHugh Dickins * stable_tree_insert - insert stable tree node pointing to new ksm page 179031dbd01fSIzik Eidus * into the stable tree. 179131dbd01fSIzik Eidus * 17927b6ba2c7SHugh Dickins * This function returns the stable tree node just allocated on success, 17937b6ba2c7SHugh Dickins * NULL otherwise. 179431dbd01fSIzik Eidus */ 179521fbd591SQi Zheng static struct ksm_stable_node *stable_tree_insert(struct page *kpage) 179631dbd01fSIzik Eidus { 179790bd6fd3SPetr Holasek int nid; 179890bd6fd3SPetr Holasek unsigned long kpfn; 1799ef53d16cSHugh Dickins struct rb_root *root; 180090bd6fd3SPetr Holasek struct rb_node **new; 1801f2e5ff85SAndrea Arcangeli struct rb_node *parent; 180221fbd591SQi Zheng struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; 18032c653d0eSAndrea Arcangeli bool need_chain = false; 180431dbd01fSIzik Eidus 180590bd6fd3SPetr Holasek kpfn = page_to_pfn(kpage); 180690bd6fd3SPetr Holasek nid = get_kpfn_nid(kpfn); 1807ef53d16cSHugh Dickins root = root_stable_tree + nid; 1808f2e5ff85SAndrea Arcangeli again: 1809f2e5ff85SAndrea Arcangeli parent = NULL; 1810ef53d16cSHugh Dickins new = &root->rb_node; 181190bd6fd3SPetr Holasek 181231dbd01fSIzik Eidus while (*new) { 18134035c07aSHugh Dickins struct page *tree_page; 181431dbd01fSIzik Eidus int ret; 181531dbd01fSIzik Eidus 181631dbd01fSIzik Eidus cond_resched(); 181721fbd591SQi Zheng stable_node = rb_entry(*new, struct ksm_stable_node, node); 18182c653d0eSAndrea Arcangeli stable_node_any = NULL; 18198dc5ffcdSAndrea Arcangeli tree_page = chain(&stable_node_dup, stable_node, root); 18202c653d0eSAndrea Arcangeli if (!stable_node_dup) { 18212c653d0eSAndrea Arcangeli /* 18222c653d0eSAndrea Arcangeli * Either all stable_node dups were full in 18232c653d0eSAndrea Arcangeli * this stable_node chain, or this chain was 18242c653d0eSAndrea Arcangeli * empty and should be rb_erased. 18252c653d0eSAndrea Arcangeli */ 18262c653d0eSAndrea Arcangeli stable_node_any = stable_node_dup_any(stable_node, 18272c653d0eSAndrea Arcangeli root); 18282c653d0eSAndrea Arcangeli if (!stable_node_any) { 18292c653d0eSAndrea Arcangeli /* rb_erase just run */ 18302c653d0eSAndrea Arcangeli goto again; 18312c653d0eSAndrea Arcangeli } 18322c653d0eSAndrea Arcangeli /* 18332c653d0eSAndrea Arcangeli * Take any of the stable_node dups page of 18342c653d0eSAndrea Arcangeli * this stable_node chain to let the tree walk 18352c653d0eSAndrea Arcangeli * continue. All KSM pages belonging to the 18362c653d0eSAndrea Arcangeli * stable_node dups in a stable_node chain 18372c653d0eSAndrea Arcangeli * have the same content and they're 1838457aef94SEthon Paul * write protected at all times. Any will work 18392c653d0eSAndrea Arcangeli * fine to continue the walk. 18402c653d0eSAndrea Arcangeli */ 18412cee57d1SYang Shi tree_page = get_ksm_page(stable_node_any, 18422cee57d1SYang Shi GET_KSM_PAGE_NOLOCK); 18432c653d0eSAndrea Arcangeli } 18442c653d0eSAndrea Arcangeli VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); 1845f2e5ff85SAndrea Arcangeli if (!tree_page) { 1846f2e5ff85SAndrea Arcangeli /* 1847f2e5ff85SAndrea Arcangeli * If we walked over a stale stable_node, 1848f2e5ff85SAndrea Arcangeli * get_ksm_page() will call rb_erase() and it 1849f2e5ff85SAndrea Arcangeli * may rebalance the tree from under us. So 1850f2e5ff85SAndrea Arcangeli * restart the search from scratch. Returning 1851f2e5ff85SAndrea Arcangeli * NULL would be safe too, but we'd generate 1852f2e5ff85SAndrea Arcangeli * false negative insertions just because some 1853f2e5ff85SAndrea Arcangeli * stable_node was stale. 1854f2e5ff85SAndrea Arcangeli */ 1855f2e5ff85SAndrea Arcangeli goto again; 1856f2e5ff85SAndrea Arcangeli } 185731dbd01fSIzik Eidus 18584035c07aSHugh Dickins ret = memcmp_pages(kpage, tree_page); 18594035c07aSHugh Dickins put_page(tree_page); 186031dbd01fSIzik Eidus 186131dbd01fSIzik Eidus parent = *new; 186231dbd01fSIzik Eidus if (ret < 0) 186331dbd01fSIzik Eidus new = &parent->rb_left; 186431dbd01fSIzik Eidus else if (ret > 0) 186531dbd01fSIzik Eidus new = &parent->rb_right; 186631dbd01fSIzik Eidus else { 18672c653d0eSAndrea Arcangeli need_chain = true; 18682c653d0eSAndrea Arcangeli break; 186931dbd01fSIzik Eidus } 187031dbd01fSIzik Eidus } 187131dbd01fSIzik Eidus 18722c653d0eSAndrea Arcangeli stable_node_dup = alloc_stable_node(); 18732c653d0eSAndrea Arcangeli if (!stable_node_dup) 18747b6ba2c7SHugh Dickins return NULL; 187531dbd01fSIzik Eidus 18762c653d0eSAndrea Arcangeli INIT_HLIST_HEAD(&stable_node_dup->hlist); 18772c653d0eSAndrea Arcangeli stable_node_dup->kpfn = kpfn; 18782c653d0eSAndrea Arcangeli set_page_stable_node(kpage, stable_node_dup); 18792c653d0eSAndrea Arcangeli stable_node_dup->rmap_hlist_len = 0; 18802c653d0eSAndrea Arcangeli DO_NUMA(stable_node_dup->nid = nid); 18812c653d0eSAndrea Arcangeli if (!need_chain) { 18822c653d0eSAndrea Arcangeli rb_link_node(&stable_node_dup->node, parent, new); 18832c653d0eSAndrea Arcangeli rb_insert_color(&stable_node_dup->node, root); 18842c653d0eSAndrea Arcangeli } else { 18852c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 188621fbd591SQi Zheng struct ksm_stable_node *orig = stable_node; 18872c653d0eSAndrea Arcangeli /* chain is missing so create it */ 18882c653d0eSAndrea Arcangeli stable_node = alloc_stable_node_chain(orig, root); 18892c653d0eSAndrea Arcangeli if (!stable_node) { 18902c653d0eSAndrea Arcangeli free_stable_node(stable_node_dup); 18912c653d0eSAndrea Arcangeli return NULL; 18922c653d0eSAndrea Arcangeli } 18932c653d0eSAndrea Arcangeli } 18942c653d0eSAndrea Arcangeli stable_node_chain_add_dup(stable_node_dup, stable_node); 18952c653d0eSAndrea Arcangeli } 189608beca44SHugh Dickins 18972c653d0eSAndrea Arcangeli return stable_node_dup; 189831dbd01fSIzik Eidus } 189931dbd01fSIzik Eidus 190031dbd01fSIzik Eidus /* 19018dd3557aSHugh Dickins * unstable_tree_search_insert - search for identical page, 19028dd3557aSHugh Dickins * else insert rmap_item into the unstable tree. 190331dbd01fSIzik Eidus * 190431dbd01fSIzik Eidus * This function searches for a page in the unstable tree identical to the 190531dbd01fSIzik Eidus * page currently being scanned; and if no identical page is found in the 190631dbd01fSIzik Eidus * tree, we insert rmap_item as a new object into the unstable tree. 190731dbd01fSIzik Eidus * 190831dbd01fSIzik Eidus * This function returns pointer to rmap_item found to be identical 190931dbd01fSIzik Eidus * to the currently scanned page, NULL otherwise. 191031dbd01fSIzik Eidus * 191131dbd01fSIzik Eidus * This function does both searching and inserting, because they share 191231dbd01fSIzik Eidus * the same walking algorithm in an rbtree. 191331dbd01fSIzik Eidus */ 19148dd3557aSHugh Dickins static 191521fbd591SQi Zheng struct ksm_rmap_item *unstable_tree_search_insert(struct ksm_rmap_item *rmap_item, 19168dd3557aSHugh Dickins struct page *page, 19178dd3557aSHugh Dickins struct page **tree_pagep) 191831dbd01fSIzik Eidus { 191990bd6fd3SPetr Holasek struct rb_node **new; 192090bd6fd3SPetr Holasek struct rb_root *root; 192131dbd01fSIzik Eidus struct rb_node *parent = NULL; 192290bd6fd3SPetr Holasek int nid; 192390bd6fd3SPetr Holasek 192490bd6fd3SPetr Holasek nid = get_kpfn_nid(page_to_pfn(page)); 1925ef53d16cSHugh Dickins root = root_unstable_tree + nid; 192690bd6fd3SPetr Holasek new = &root->rb_node; 192731dbd01fSIzik Eidus 192831dbd01fSIzik Eidus while (*new) { 192921fbd591SQi Zheng struct ksm_rmap_item *tree_rmap_item; 19308dd3557aSHugh Dickins struct page *tree_page; 193131dbd01fSIzik Eidus int ret; 193231dbd01fSIzik Eidus 1933d178f27fSHugh Dickins cond_resched(); 193421fbd591SQi Zheng tree_rmap_item = rb_entry(*new, struct ksm_rmap_item, node); 19358dd3557aSHugh Dickins tree_page = get_mergeable_page(tree_rmap_item); 1936c8f95ed1SAndrea Arcangeli if (!tree_page) 193731dbd01fSIzik Eidus return NULL; 193831dbd01fSIzik Eidus 193931dbd01fSIzik Eidus /* 19408dd3557aSHugh Dickins * Don't substitute a ksm page for a forked page. 194131dbd01fSIzik Eidus */ 19428dd3557aSHugh Dickins if (page == tree_page) { 19438dd3557aSHugh Dickins put_page(tree_page); 194431dbd01fSIzik Eidus return NULL; 194531dbd01fSIzik Eidus } 194631dbd01fSIzik Eidus 19478dd3557aSHugh Dickins ret = memcmp_pages(page, tree_page); 194831dbd01fSIzik Eidus 194931dbd01fSIzik Eidus parent = *new; 195031dbd01fSIzik Eidus if (ret < 0) { 19518dd3557aSHugh Dickins put_page(tree_page); 195231dbd01fSIzik Eidus new = &parent->rb_left; 195331dbd01fSIzik Eidus } else if (ret > 0) { 19548dd3557aSHugh Dickins put_page(tree_page); 195531dbd01fSIzik Eidus new = &parent->rb_right; 1956b599cbdfSHugh Dickins } else if (!ksm_merge_across_nodes && 1957b599cbdfSHugh Dickins page_to_nid(tree_page) != nid) { 1958b599cbdfSHugh Dickins /* 1959b599cbdfSHugh Dickins * If tree_page has been migrated to another NUMA node, 1960b599cbdfSHugh Dickins * it will be flushed out and put in the right unstable 1961b599cbdfSHugh Dickins * tree next time: only merge with it when across_nodes. 1962b599cbdfSHugh Dickins */ 1963b599cbdfSHugh Dickins put_page(tree_page); 1964b599cbdfSHugh Dickins return NULL; 196531dbd01fSIzik Eidus } else { 19668dd3557aSHugh Dickins *tree_pagep = tree_page; 196731dbd01fSIzik Eidus return tree_rmap_item; 196831dbd01fSIzik Eidus } 196931dbd01fSIzik Eidus } 197031dbd01fSIzik Eidus 19717b6ba2c7SHugh Dickins rmap_item->address |= UNSTABLE_FLAG; 197231dbd01fSIzik Eidus rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); 1973e850dcf5SHugh Dickins DO_NUMA(rmap_item->nid = nid); 197431dbd01fSIzik Eidus rb_link_node(&rmap_item->node, parent, new); 197590bd6fd3SPetr Holasek rb_insert_color(&rmap_item->node, root); 197631dbd01fSIzik Eidus 1977473b0ce4SHugh Dickins ksm_pages_unshared++; 197831dbd01fSIzik Eidus return NULL; 197931dbd01fSIzik Eidus } 198031dbd01fSIzik Eidus 198131dbd01fSIzik Eidus /* 198231dbd01fSIzik Eidus * stable_tree_append - add another rmap_item to the linked list of 198331dbd01fSIzik Eidus * rmap_items hanging off a given node of the stable tree, all sharing 198431dbd01fSIzik Eidus * the same ksm page. 198531dbd01fSIzik Eidus */ 198621fbd591SQi Zheng static void stable_tree_append(struct ksm_rmap_item *rmap_item, 198721fbd591SQi Zheng struct ksm_stable_node *stable_node, 19882c653d0eSAndrea Arcangeli bool max_page_sharing_bypass) 198931dbd01fSIzik Eidus { 19902c653d0eSAndrea Arcangeli /* 19912c653d0eSAndrea Arcangeli * rmap won't find this mapping if we don't insert the 19922c653d0eSAndrea Arcangeli * rmap_item in the right stable_node 19932c653d0eSAndrea Arcangeli * duplicate. page_migration could break later if rmap breaks, 19942c653d0eSAndrea Arcangeli * so we can as well crash here. We really need to check for 19952c653d0eSAndrea Arcangeli * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check 1996457aef94SEthon Paul * for other negative values as an underflow if detected here 19972c653d0eSAndrea Arcangeli * for the first time (and not when decreasing rmap_hlist_len) 19982c653d0eSAndrea Arcangeli * would be sign of memory corruption in the stable_node. 19992c653d0eSAndrea Arcangeli */ 20002c653d0eSAndrea Arcangeli BUG_ON(stable_node->rmap_hlist_len < 0); 20012c653d0eSAndrea Arcangeli 20022c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len++; 20032c653d0eSAndrea Arcangeli if (!max_page_sharing_bypass) 20042c653d0eSAndrea Arcangeli /* possibly non fatal but unexpected overflow, only warn */ 20052c653d0eSAndrea Arcangeli WARN_ON_ONCE(stable_node->rmap_hlist_len > 20062c653d0eSAndrea Arcangeli ksm_max_page_sharing); 20072c653d0eSAndrea Arcangeli 20087b6ba2c7SHugh Dickins rmap_item->head = stable_node; 200931dbd01fSIzik Eidus rmap_item->address |= STABLE_FLAG; 20107b6ba2c7SHugh Dickins hlist_add_head(&rmap_item->hlist, &stable_node->hlist); 2011e178dfdeSHugh Dickins 20127b6ba2c7SHugh Dickins if (rmap_item->hlist.next) 2013e178dfdeSHugh Dickins ksm_pages_sharing++; 20147b6ba2c7SHugh Dickins else 20157b6ba2c7SHugh Dickins ksm_pages_shared++; 201676093853Sxu xin 201776093853Sxu xin rmap_item->mm->ksm_merging_pages++; 201831dbd01fSIzik Eidus } 201931dbd01fSIzik Eidus 202031dbd01fSIzik Eidus /* 202181464e30SHugh Dickins * cmp_and_merge_page - first see if page can be merged into the stable tree; 202281464e30SHugh Dickins * if not, compare checksum to previous and if it's the same, see if page can 202381464e30SHugh Dickins * be inserted into the unstable tree, or merged with a page already there and 202481464e30SHugh Dickins * both transferred to the stable tree. 202531dbd01fSIzik Eidus * 202631dbd01fSIzik Eidus * @page: the page that we are searching identical page to. 202731dbd01fSIzik Eidus * @rmap_item: the reverse mapping into the virtual address of this page 202831dbd01fSIzik Eidus */ 202921fbd591SQi Zheng static void cmp_and_merge_page(struct page *page, struct ksm_rmap_item *rmap_item) 203031dbd01fSIzik Eidus { 20314b22927fSKirill Tkhai struct mm_struct *mm = rmap_item->mm; 203221fbd591SQi Zheng struct ksm_rmap_item *tree_rmap_item; 20338dd3557aSHugh Dickins struct page *tree_page = NULL; 203421fbd591SQi Zheng struct ksm_stable_node *stable_node; 20358dd3557aSHugh Dickins struct page *kpage; 203631dbd01fSIzik Eidus unsigned int checksum; 203731dbd01fSIzik Eidus int err; 20382c653d0eSAndrea Arcangeli bool max_page_sharing_bypass = false; 203931dbd01fSIzik Eidus 20404146d2d6SHugh Dickins stable_node = page_stable_node(page); 20414146d2d6SHugh Dickins if (stable_node) { 20424146d2d6SHugh Dickins if (stable_node->head != &migrate_nodes && 20432c653d0eSAndrea Arcangeli get_kpfn_nid(READ_ONCE(stable_node->kpfn)) != 20442c653d0eSAndrea Arcangeli NUMA(stable_node->nid)) { 20452c653d0eSAndrea Arcangeli stable_node_dup_del(stable_node); 20464146d2d6SHugh Dickins stable_node->head = &migrate_nodes; 20474146d2d6SHugh Dickins list_add(&stable_node->list, stable_node->head); 20484146d2d6SHugh Dickins } 20494146d2d6SHugh Dickins if (stable_node->head != &migrate_nodes && 20504146d2d6SHugh Dickins rmap_item->head == stable_node) 20514146d2d6SHugh Dickins return; 20522c653d0eSAndrea Arcangeli /* 20532c653d0eSAndrea Arcangeli * If it's a KSM fork, allow it to go over the sharing limit 20542c653d0eSAndrea Arcangeli * without warnings. 20552c653d0eSAndrea Arcangeli */ 20562c653d0eSAndrea Arcangeli if (!is_page_sharing_candidate(stable_node)) 20572c653d0eSAndrea Arcangeli max_page_sharing_bypass = true; 20584146d2d6SHugh Dickins } 205931dbd01fSIzik Eidus 206031dbd01fSIzik Eidus /* We first start with searching the page inside the stable tree */ 206162b61f61SHugh Dickins kpage = stable_tree_search(page); 20624146d2d6SHugh Dickins if (kpage == page && rmap_item->head == stable_node) { 20634146d2d6SHugh Dickins put_page(kpage); 20644146d2d6SHugh Dickins return; 20654146d2d6SHugh Dickins } 20664146d2d6SHugh Dickins 20674146d2d6SHugh Dickins remove_rmap_item_from_tree(rmap_item); 20684146d2d6SHugh Dickins 206962b61f61SHugh Dickins if (kpage) { 20702cee57d1SYang Shi if (PTR_ERR(kpage) == -EBUSY) 20712cee57d1SYang Shi return; 20722cee57d1SYang Shi 207308beca44SHugh Dickins err = try_to_merge_with_ksm_page(rmap_item, page, kpage); 207431dbd01fSIzik Eidus if (!err) { 207531dbd01fSIzik Eidus /* 207631dbd01fSIzik Eidus * The page was successfully merged: 207731dbd01fSIzik Eidus * add its rmap_item to the stable tree. 207831dbd01fSIzik Eidus */ 20795ad64688SHugh Dickins lock_page(kpage); 20802c653d0eSAndrea Arcangeli stable_tree_append(rmap_item, page_stable_node(kpage), 20812c653d0eSAndrea Arcangeli max_page_sharing_bypass); 20825ad64688SHugh Dickins unlock_page(kpage); 208331dbd01fSIzik Eidus } 20848dd3557aSHugh Dickins put_page(kpage); 208531dbd01fSIzik Eidus return; 208631dbd01fSIzik Eidus } 208731dbd01fSIzik Eidus 208831dbd01fSIzik Eidus /* 20894035c07aSHugh Dickins * If the hash value of the page has changed from the last time 20904035c07aSHugh Dickins * we calculated it, this page is changing frequently: therefore we 20914035c07aSHugh Dickins * don't want to insert it in the unstable tree, and we don't want 20924035c07aSHugh Dickins * to waste our time searching for something identical to it there. 209331dbd01fSIzik Eidus */ 209431dbd01fSIzik Eidus checksum = calc_checksum(page); 209531dbd01fSIzik Eidus if (rmap_item->oldchecksum != checksum) { 209631dbd01fSIzik Eidus rmap_item->oldchecksum = checksum; 209731dbd01fSIzik Eidus return; 209831dbd01fSIzik Eidus } 209931dbd01fSIzik Eidus 2100e86c59b1SClaudio Imbrenda /* 2101e86c59b1SClaudio Imbrenda * Same checksum as an empty page. We attempt to merge it with the 2102e86c59b1SClaudio Imbrenda * appropriate zero page if the user enabled this via sysfs. 2103e86c59b1SClaudio Imbrenda */ 2104e86c59b1SClaudio Imbrenda if (ksm_use_zero_pages && (checksum == zero_checksum)) { 2105e86c59b1SClaudio Imbrenda struct vm_area_struct *vma; 2106e86c59b1SClaudio Imbrenda 2107d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 21084b22927fSKirill Tkhai vma = find_mergeable_vma(mm, rmap_item->address); 210956df70a6SMuchun Song if (vma) { 2110e86c59b1SClaudio Imbrenda err = try_to_merge_one_page(vma, page, 2111e86c59b1SClaudio Imbrenda ZERO_PAGE(rmap_item->address)); 211256df70a6SMuchun Song } else { 211356df70a6SMuchun Song /* 211456df70a6SMuchun Song * If the vma is out of date, we do not need to 211556df70a6SMuchun Song * continue. 211656df70a6SMuchun Song */ 211756df70a6SMuchun Song err = 0; 211856df70a6SMuchun Song } 2119d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 2120e86c59b1SClaudio Imbrenda /* 2121e86c59b1SClaudio Imbrenda * In case of failure, the page was not really empty, so we 2122e86c59b1SClaudio Imbrenda * need to continue. Otherwise we're done. 2123e86c59b1SClaudio Imbrenda */ 2124e86c59b1SClaudio Imbrenda if (!err) 2125e86c59b1SClaudio Imbrenda return; 2126e86c59b1SClaudio Imbrenda } 21278dd3557aSHugh Dickins tree_rmap_item = 21288dd3557aSHugh Dickins unstable_tree_search_insert(rmap_item, page, &tree_page); 212931dbd01fSIzik Eidus if (tree_rmap_item) { 213077da2ba0SClaudio Imbrenda bool split; 213177da2ba0SClaudio Imbrenda 21328dd3557aSHugh Dickins kpage = try_to_merge_two_pages(rmap_item, page, 21338dd3557aSHugh Dickins tree_rmap_item, tree_page); 213477da2ba0SClaudio Imbrenda /* 213577da2ba0SClaudio Imbrenda * If both pages we tried to merge belong to the same compound 213677da2ba0SClaudio Imbrenda * page, then we actually ended up increasing the reference 213777da2ba0SClaudio Imbrenda * count of the same compound page twice, and split_huge_page 213877da2ba0SClaudio Imbrenda * failed. 213977da2ba0SClaudio Imbrenda * Here we set a flag if that happened, and we use it later to 214077da2ba0SClaudio Imbrenda * try split_huge_page again. Since we call put_page right 214177da2ba0SClaudio Imbrenda * afterwards, the reference count will be correct and 214277da2ba0SClaudio Imbrenda * split_huge_page should succeed. 214377da2ba0SClaudio Imbrenda */ 214477da2ba0SClaudio Imbrenda split = PageTransCompound(page) 214577da2ba0SClaudio Imbrenda && compound_head(page) == compound_head(tree_page); 21468dd3557aSHugh Dickins put_page(tree_page); 21478dd3557aSHugh Dickins if (kpage) { 2148bc56620bSHugh Dickins /* 2149bc56620bSHugh Dickins * The pages were successfully merged: insert new 2150bc56620bSHugh Dickins * node in the stable tree and add both rmap_items. 2151bc56620bSHugh Dickins */ 21525ad64688SHugh Dickins lock_page(kpage); 21537b6ba2c7SHugh Dickins stable_node = stable_tree_insert(kpage); 21547b6ba2c7SHugh Dickins if (stable_node) { 21552c653d0eSAndrea Arcangeli stable_tree_append(tree_rmap_item, stable_node, 21562c653d0eSAndrea Arcangeli false); 21572c653d0eSAndrea Arcangeli stable_tree_append(rmap_item, stable_node, 21582c653d0eSAndrea Arcangeli false); 21597b6ba2c7SHugh Dickins } 21605ad64688SHugh Dickins unlock_page(kpage); 21617b6ba2c7SHugh Dickins 216231dbd01fSIzik Eidus /* 216331dbd01fSIzik Eidus * If we fail to insert the page into the stable tree, 216431dbd01fSIzik Eidus * we will have 2 virtual addresses that are pointing 216531dbd01fSIzik Eidus * to a ksm page left outside the stable tree, 216631dbd01fSIzik Eidus * in which case we need to break_cow on both. 216731dbd01fSIzik Eidus */ 21687b6ba2c7SHugh Dickins if (!stable_node) { 21698dd3557aSHugh Dickins break_cow(tree_rmap_item); 21708dd3557aSHugh Dickins break_cow(rmap_item); 217131dbd01fSIzik Eidus } 217277da2ba0SClaudio Imbrenda } else if (split) { 217377da2ba0SClaudio Imbrenda /* 217477da2ba0SClaudio Imbrenda * We are here if we tried to merge two pages and 217577da2ba0SClaudio Imbrenda * failed because they both belonged to the same 217677da2ba0SClaudio Imbrenda * compound page. We will split the page now, but no 217777da2ba0SClaudio Imbrenda * merging will take place. 217877da2ba0SClaudio Imbrenda * We do not want to add the cost of a full lock; if 217977da2ba0SClaudio Imbrenda * the page is locked, it is better to skip it and 218077da2ba0SClaudio Imbrenda * perhaps try again later. 218177da2ba0SClaudio Imbrenda */ 218277da2ba0SClaudio Imbrenda if (!trylock_page(page)) 218377da2ba0SClaudio Imbrenda return; 218477da2ba0SClaudio Imbrenda split_huge_page(page); 218577da2ba0SClaudio Imbrenda unlock_page(page); 218631dbd01fSIzik Eidus } 218731dbd01fSIzik Eidus } 218831dbd01fSIzik Eidus } 218931dbd01fSIzik Eidus 219021fbd591SQi Zheng static struct ksm_rmap_item *get_next_rmap_item(struct ksm_mm_slot *mm_slot, 219121fbd591SQi Zheng struct ksm_rmap_item **rmap_list, 219231dbd01fSIzik Eidus unsigned long addr) 219331dbd01fSIzik Eidus { 219421fbd591SQi Zheng struct ksm_rmap_item *rmap_item; 219531dbd01fSIzik Eidus 21966514d511SHugh Dickins while (*rmap_list) { 21976514d511SHugh Dickins rmap_item = *rmap_list; 219893d17715SHugh Dickins if ((rmap_item->address & PAGE_MASK) == addr) 219931dbd01fSIzik Eidus return rmap_item; 220031dbd01fSIzik Eidus if (rmap_item->address > addr) 220131dbd01fSIzik Eidus break; 22026514d511SHugh Dickins *rmap_list = rmap_item->rmap_list; 220331dbd01fSIzik Eidus remove_rmap_item_from_tree(rmap_item); 220431dbd01fSIzik Eidus free_rmap_item(rmap_item); 220531dbd01fSIzik Eidus } 220631dbd01fSIzik Eidus 220731dbd01fSIzik Eidus rmap_item = alloc_rmap_item(); 220831dbd01fSIzik Eidus if (rmap_item) { 220931dbd01fSIzik Eidus /* It has already been zeroed */ 221058730ab6SQi Zheng rmap_item->mm = mm_slot->slot.mm; 2211cb4df4caSxu xin rmap_item->mm->ksm_rmap_items++; 221231dbd01fSIzik Eidus rmap_item->address = addr; 22136514d511SHugh Dickins rmap_item->rmap_list = *rmap_list; 22146514d511SHugh Dickins *rmap_list = rmap_item; 221531dbd01fSIzik Eidus } 221631dbd01fSIzik Eidus return rmap_item; 221731dbd01fSIzik Eidus } 221831dbd01fSIzik Eidus 221921fbd591SQi Zheng static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page) 222031dbd01fSIzik Eidus { 222131dbd01fSIzik Eidus struct mm_struct *mm; 222258730ab6SQi Zheng struct ksm_mm_slot *mm_slot; 222358730ab6SQi Zheng struct mm_slot *slot; 222431dbd01fSIzik Eidus struct vm_area_struct *vma; 222521fbd591SQi Zheng struct ksm_rmap_item *rmap_item; 2226a5f18ba0SMatthew Wilcox (Oracle) struct vma_iterator vmi; 222790bd6fd3SPetr Holasek int nid; 222831dbd01fSIzik Eidus 222958730ab6SQi Zheng if (list_empty(&ksm_mm_head.slot.mm_node)) 223031dbd01fSIzik Eidus return NULL; 223131dbd01fSIzik Eidus 223258730ab6SQi Zheng mm_slot = ksm_scan.mm_slot; 223358730ab6SQi Zheng if (mm_slot == &ksm_mm_head) { 22342919bfd0SHugh Dickins /* 22352919bfd0SHugh Dickins * A number of pages can hang around indefinitely on per-cpu 22362919bfd0SHugh Dickins * pagevecs, raised page count preventing write_protect_page 22372919bfd0SHugh Dickins * from merging them. Though it doesn't really matter much, 22382919bfd0SHugh Dickins * it is puzzling to see some stuck in pages_volatile until 22392919bfd0SHugh Dickins * other activity jostles them out, and they also prevented 22402919bfd0SHugh Dickins * LTP's KSM test from succeeding deterministically; so drain 22412919bfd0SHugh Dickins * them here (here rather than on entry to ksm_do_scan(), 22422919bfd0SHugh Dickins * so we don't IPI too often when pages_to_scan is set low). 22432919bfd0SHugh Dickins */ 22442919bfd0SHugh Dickins lru_add_drain_all(); 22452919bfd0SHugh Dickins 22464146d2d6SHugh Dickins /* 22474146d2d6SHugh Dickins * Whereas stale stable_nodes on the stable_tree itself 22484146d2d6SHugh Dickins * get pruned in the regular course of stable_tree_search(), 22494146d2d6SHugh Dickins * those moved out to the migrate_nodes list can accumulate: 22504146d2d6SHugh Dickins * so prune them once before each full scan. 22514146d2d6SHugh Dickins */ 22524146d2d6SHugh Dickins if (!ksm_merge_across_nodes) { 225321fbd591SQi Zheng struct ksm_stable_node *stable_node, *next; 22544146d2d6SHugh Dickins struct page *page; 22554146d2d6SHugh Dickins 225603640418SGeliang Tang list_for_each_entry_safe(stable_node, next, 225703640418SGeliang Tang &migrate_nodes, list) { 22582cee57d1SYang Shi page = get_ksm_page(stable_node, 22592cee57d1SYang Shi GET_KSM_PAGE_NOLOCK); 22604146d2d6SHugh Dickins if (page) 22614146d2d6SHugh Dickins put_page(page); 22624146d2d6SHugh Dickins cond_resched(); 22634146d2d6SHugh Dickins } 22644146d2d6SHugh Dickins } 22654146d2d6SHugh Dickins 2266ef53d16cSHugh Dickins for (nid = 0; nid < ksm_nr_node_ids; nid++) 226790bd6fd3SPetr Holasek root_unstable_tree[nid] = RB_ROOT; 226831dbd01fSIzik Eidus 226931dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 227058730ab6SQi Zheng slot = list_entry(mm_slot->slot.mm_node.next, 227158730ab6SQi Zheng struct mm_slot, mm_node); 227258730ab6SQi Zheng mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); 227358730ab6SQi Zheng ksm_scan.mm_slot = mm_slot; 227431dbd01fSIzik Eidus spin_unlock(&ksm_mmlist_lock); 22752b472611SHugh Dickins /* 22762b472611SHugh Dickins * Although we tested list_empty() above, a racing __ksm_exit 22772b472611SHugh Dickins * of the last mm on the list may have removed it since then. 22782b472611SHugh Dickins */ 227958730ab6SQi Zheng if (mm_slot == &ksm_mm_head) 22802b472611SHugh Dickins return NULL; 228131dbd01fSIzik Eidus next_mm: 228231dbd01fSIzik Eidus ksm_scan.address = 0; 228358730ab6SQi Zheng ksm_scan.rmap_list = &mm_slot->rmap_list; 228431dbd01fSIzik Eidus } 228531dbd01fSIzik Eidus 228658730ab6SQi Zheng slot = &mm_slot->slot; 228731dbd01fSIzik Eidus mm = slot->mm; 2288a5f18ba0SMatthew Wilcox (Oracle) vma_iter_init(&vmi, mm, ksm_scan.address); 2289a5f18ba0SMatthew Wilcox (Oracle) 2290d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 22919ba69294SHugh Dickins if (ksm_test_exit(mm)) 2292a5f18ba0SMatthew Wilcox (Oracle) goto no_vmas; 22939ba69294SHugh Dickins 2294a5f18ba0SMatthew Wilcox (Oracle) for_each_vma(vmi, vma) { 229531dbd01fSIzik Eidus if (!(vma->vm_flags & VM_MERGEABLE)) 229631dbd01fSIzik Eidus continue; 229731dbd01fSIzik Eidus if (ksm_scan.address < vma->vm_start) 229831dbd01fSIzik Eidus ksm_scan.address = vma->vm_start; 229931dbd01fSIzik Eidus if (!vma->anon_vma) 230031dbd01fSIzik Eidus ksm_scan.address = vma->vm_end; 230131dbd01fSIzik Eidus 230231dbd01fSIzik Eidus while (ksm_scan.address < vma->vm_end) { 23039ba69294SHugh Dickins if (ksm_test_exit(mm)) 23049ba69294SHugh Dickins break; 230531dbd01fSIzik Eidus *page = follow_page(vma, ksm_scan.address, FOLL_GET); 2306f7091ed6SHaiyue Wang if (IS_ERR_OR_NULL(*page)) { 230721ae5b01SAndrea Arcangeli ksm_scan.address += PAGE_SIZE; 230821ae5b01SAndrea Arcangeli cond_resched(); 230921ae5b01SAndrea Arcangeli continue; 231021ae5b01SAndrea Arcangeli } 2311f7091ed6SHaiyue Wang if (is_zone_device_page(*page)) 2312f7091ed6SHaiyue Wang goto next_page; 2313f765f540SKirill A. Shutemov if (PageAnon(*page)) { 231431dbd01fSIzik Eidus flush_anon_page(vma, *page, ksm_scan.address); 231531dbd01fSIzik Eidus flush_dcache_page(*page); 231658730ab6SQi Zheng rmap_item = get_next_rmap_item(mm_slot, 23176514d511SHugh Dickins ksm_scan.rmap_list, ksm_scan.address); 231831dbd01fSIzik Eidus if (rmap_item) { 23196514d511SHugh Dickins ksm_scan.rmap_list = 23206514d511SHugh Dickins &rmap_item->rmap_list; 232131dbd01fSIzik Eidus ksm_scan.address += PAGE_SIZE; 232231dbd01fSIzik Eidus } else 232331dbd01fSIzik Eidus put_page(*page); 2324d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 232531dbd01fSIzik Eidus return rmap_item; 232631dbd01fSIzik Eidus } 2327f7091ed6SHaiyue Wang next_page: 232831dbd01fSIzik Eidus put_page(*page); 232931dbd01fSIzik Eidus ksm_scan.address += PAGE_SIZE; 233031dbd01fSIzik Eidus cond_resched(); 233131dbd01fSIzik Eidus } 233231dbd01fSIzik Eidus } 233331dbd01fSIzik Eidus 23349ba69294SHugh Dickins if (ksm_test_exit(mm)) { 2335a5f18ba0SMatthew Wilcox (Oracle) no_vmas: 23369ba69294SHugh Dickins ksm_scan.address = 0; 233758730ab6SQi Zheng ksm_scan.rmap_list = &mm_slot->rmap_list; 23389ba69294SHugh Dickins } 233931dbd01fSIzik Eidus /* 234031dbd01fSIzik Eidus * Nuke all the rmap_items that are above this current rmap: 234131dbd01fSIzik Eidus * because there were no VM_MERGEABLE vmas with such addresses. 234231dbd01fSIzik Eidus */ 2343420be4edSChengyang Fan remove_trailing_rmap_items(ksm_scan.rmap_list); 234431dbd01fSIzik Eidus 234531dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 234658730ab6SQi Zheng slot = list_entry(mm_slot->slot.mm_node.next, 234758730ab6SQi Zheng struct mm_slot, mm_node); 234858730ab6SQi Zheng ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); 2349cd551f97SHugh Dickins if (ksm_scan.address == 0) { 2350cd551f97SHugh Dickins /* 2351c1e8d7c6SMichel Lespinasse * We've completed a full scan of all vmas, holding mmap_lock 2352cd551f97SHugh Dickins * throughout, and found no VM_MERGEABLE: so do the same as 2353cd551f97SHugh Dickins * __ksm_exit does to remove this mm from all our lists now. 23549ba69294SHugh Dickins * This applies either when cleaning up after __ksm_exit 23559ba69294SHugh Dickins * (but beware: we can reach here even before __ksm_exit), 23569ba69294SHugh Dickins * or when all VM_MERGEABLE areas have been unmapped (and 2357c1e8d7c6SMichel Lespinasse * mmap_lock then protects against race with MADV_MERGEABLE). 2358cd551f97SHugh Dickins */ 235958730ab6SQi Zheng hash_del(&mm_slot->slot.hash); 236058730ab6SQi Zheng list_del(&mm_slot->slot.mm_node); 23619ba69294SHugh Dickins spin_unlock(&ksm_mmlist_lock); 23629ba69294SHugh Dickins 236358730ab6SQi Zheng mm_slot_free(mm_slot_cache, mm_slot); 2364cd551f97SHugh Dickins clear_bit(MMF_VM_MERGEABLE, &mm->flags); 2365d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 23669ba69294SHugh Dickins mmdrop(mm); 23679ba69294SHugh Dickins } else { 2368d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 23697496fea9SZhou Chengming /* 23703e4e28c5SMichel Lespinasse * mmap_read_unlock(mm) first because after 23717496fea9SZhou Chengming * spin_unlock(&ksm_mmlist_lock) run, the "mm" may 23727496fea9SZhou Chengming * already have been freed under us by __ksm_exit() 23737496fea9SZhou Chengming * because the "mm_slot" is still hashed and 23747496fea9SZhou Chengming * ksm_scan.mm_slot doesn't point to it anymore. 23757496fea9SZhou Chengming */ 23767496fea9SZhou Chengming spin_unlock(&ksm_mmlist_lock); 23779ba69294SHugh Dickins } 237831dbd01fSIzik Eidus 237931dbd01fSIzik Eidus /* Repeat until we've completed scanning the whole list */ 238058730ab6SQi Zheng mm_slot = ksm_scan.mm_slot; 238158730ab6SQi Zheng if (mm_slot != &ksm_mm_head) 238231dbd01fSIzik Eidus goto next_mm; 238331dbd01fSIzik Eidus 238431dbd01fSIzik Eidus ksm_scan.seqnr++; 238531dbd01fSIzik Eidus return NULL; 238631dbd01fSIzik Eidus } 238731dbd01fSIzik Eidus 238831dbd01fSIzik Eidus /** 238931dbd01fSIzik Eidus * ksm_do_scan - the ksm scanner main worker function. 2390b7701a5fSMike Rapoport * @scan_npages: number of pages we want to scan before we return. 239131dbd01fSIzik Eidus */ 239231dbd01fSIzik Eidus static void ksm_do_scan(unsigned int scan_npages) 239331dbd01fSIzik Eidus { 239421fbd591SQi Zheng struct ksm_rmap_item *rmap_item; 23953f649ab7SKees Cook struct page *page; 239631dbd01fSIzik Eidus 2397878aee7dSAndrea Arcangeli while (scan_npages-- && likely(!freezing(current))) { 239831dbd01fSIzik Eidus cond_resched(); 239931dbd01fSIzik Eidus rmap_item = scan_get_next_rmap_item(&page); 240031dbd01fSIzik Eidus if (!rmap_item) 240131dbd01fSIzik Eidus return; 240231dbd01fSIzik Eidus cmp_and_merge_page(page, rmap_item); 240331dbd01fSIzik Eidus put_page(page); 240431dbd01fSIzik Eidus } 240531dbd01fSIzik Eidus } 240631dbd01fSIzik Eidus 24076e158384SHugh Dickins static int ksmd_should_run(void) 24086e158384SHugh Dickins { 240958730ab6SQi Zheng return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.slot.mm_node); 24106e158384SHugh Dickins } 24116e158384SHugh Dickins 241231dbd01fSIzik Eidus static int ksm_scan_thread(void *nothing) 241331dbd01fSIzik Eidus { 2414fcf9a0efSKirill Tkhai unsigned int sleep_ms; 2415fcf9a0efSKirill Tkhai 2416878aee7dSAndrea Arcangeli set_freezable(); 2417339aa624SIzik Eidus set_user_nice(current, 5); 241831dbd01fSIzik Eidus 241931dbd01fSIzik Eidus while (!kthread_should_stop()) { 242031dbd01fSIzik Eidus mutex_lock(&ksm_thread_mutex); 2421ef4d43a8SHugh Dickins wait_while_offlining(); 24226e158384SHugh Dickins if (ksmd_should_run()) 242331dbd01fSIzik Eidus ksm_do_scan(ksm_thread_pages_to_scan); 242431dbd01fSIzik Eidus mutex_unlock(&ksm_thread_mutex); 24256e158384SHugh Dickins 2426878aee7dSAndrea Arcangeli try_to_freeze(); 2427878aee7dSAndrea Arcangeli 24286e158384SHugh Dickins if (ksmd_should_run()) { 2429fcf9a0efSKirill Tkhai sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs); 2430fcf9a0efSKirill Tkhai wait_event_interruptible_timeout(ksm_iter_wait, 2431fcf9a0efSKirill Tkhai sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs), 2432fcf9a0efSKirill Tkhai msecs_to_jiffies(sleep_ms)); 243331dbd01fSIzik Eidus } else { 2434878aee7dSAndrea Arcangeli wait_event_freezable(ksm_thread_wait, 24356e158384SHugh Dickins ksmd_should_run() || kthread_should_stop()); 243631dbd01fSIzik Eidus } 243731dbd01fSIzik Eidus } 243831dbd01fSIzik Eidus return 0; 243931dbd01fSIzik Eidus } 244031dbd01fSIzik Eidus 2441f8af4da3SHugh Dickins int ksm_madvise(struct vm_area_struct *vma, unsigned long start, 2442f8af4da3SHugh Dickins unsigned long end, int advice, unsigned long *vm_flags) 2443f8af4da3SHugh Dickins { 2444f8af4da3SHugh Dickins struct mm_struct *mm = vma->vm_mm; 2445d952b791SHugh Dickins int err; 2446f8af4da3SHugh Dickins 2447f8af4da3SHugh Dickins switch (advice) { 2448f8af4da3SHugh Dickins case MADV_MERGEABLE: 2449f8af4da3SHugh Dickins /* 2450f8af4da3SHugh Dickins * Be somewhat over-protective for now! 2451f8af4da3SHugh Dickins */ 2452f8af4da3SHugh Dickins if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE | 2453f8af4da3SHugh Dickins VM_PFNMAP | VM_IO | VM_DONTEXPAND | 24540661a336SKirill A. Shutemov VM_HUGETLB | VM_MIXEDMAP)) 2455f8af4da3SHugh Dickins return 0; /* just ignore the advice */ 2456f8af4da3SHugh Dickins 2457e1fb4a08SDave Jiang if (vma_is_dax(vma)) 2458e1fb4a08SDave Jiang return 0; 2459e1fb4a08SDave Jiang 246012564485SShawn Anastasio #ifdef VM_SAO 246112564485SShawn Anastasio if (*vm_flags & VM_SAO) 246212564485SShawn Anastasio return 0; 246312564485SShawn Anastasio #endif 246474a04967SKhalid Aziz #ifdef VM_SPARC_ADI 246574a04967SKhalid Aziz if (*vm_flags & VM_SPARC_ADI) 246674a04967SKhalid Aziz return 0; 246774a04967SKhalid Aziz #endif 2468cc2383ecSKonstantin Khlebnikov 2469d952b791SHugh Dickins if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { 2470d952b791SHugh Dickins err = __ksm_enter(mm); 2471d952b791SHugh Dickins if (err) 2472d952b791SHugh Dickins return err; 2473d952b791SHugh Dickins } 2474f8af4da3SHugh Dickins 2475f8af4da3SHugh Dickins *vm_flags |= VM_MERGEABLE; 2476f8af4da3SHugh Dickins break; 2477f8af4da3SHugh Dickins 2478f8af4da3SHugh Dickins case MADV_UNMERGEABLE: 2479f8af4da3SHugh Dickins if (!(*vm_flags & VM_MERGEABLE)) 2480f8af4da3SHugh Dickins return 0; /* just ignore the advice */ 2481f8af4da3SHugh Dickins 2482d952b791SHugh Dickins if (vma->anon_vma) { 2483d952b791SHugh Dickins err = unmerge_ksm_pages(vma, start, end); 2484d952b791SHugh Dickins if (err) 2485d952b791SHugh Dickins return err; 2486d952b791SHugh Dickins } 2487f8af4da3SHugh Dickins 2488f8af4da3SHugh Dickins *vm_flags &= ~VM_MERGEABLE; 2489f8af4da3SHugh Dickins break; 2490f8af4da3SHugh Dickins } 2491f8af4da3SHugh Dickins 2492f8af4da3SHugh Dickins return 0; 2493f8af4da3SHugh Dickins } 249433cf1707SBharata B Rao EXPORT_SYMBOL_GPL(ksm_madvise); 2495f8af4da3SHugh Dickins 2496f8af4da3SHugh Dickins int __ksm_enter(struct mm_struct *mm) 2497f8af4da3SHugh Dickins { 249821fbd591SQi Zheng struct ksm_mm_slot *mm_slot; 249958730ab6SQi Zheng struct mm_slot *slot; 25006e158384SHugh Dickins int needs_wakeup; 25016e158384SHugh Dickins 250258730ab6SQi Zheng mm_slot = mm_slot_alloc(mm_slot_cache); 250331dbd01fSIzik Eidus if (!mm_slot) 250431dbd01fSIzik Eidus return -ENOMEM; 250531dbd01fSIzik Eidus 250658730ab6SQi Zheng slot = &mm_slot->slot; 250758730ab6SQi Zheng 25086e158384SHugh Dickins /* Check ksm_run too? Would need tighter locking */ 250958730ab6SQi Zheng needs_wakeup = list_empty(&ksm_mm_head.slot.mm_node); 25106e158384SHugh Dickins 251131dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 251258730ab6SQi Zheng mm_slot_insert(mm_slots_hash, mm, slot); 251331dbd01fSIzik Eidus /* 2514cbf86cfeSHugh Dickins * When KSM_RUN_MERGE (or KSM_RUN_STOP), 2515cbf86cfeSHugh Dickins * insert just behind the scanning cursor, to let the area settle 251631dbd01fSIzik Eidus * down a little; when fork is followed by immediate exec, we don't 251731dbd01fSIzik Eidus * want ksmd to waste time setting up and tearing down an rmap_list. 2518cbf86cfeSHugh Dickins * 2519cbf86cfeSHugh Dickins * But when KSM_RUN_UNMERGE, it's important to insert ahead of its 2520cbf86cfeSHugh Dickins * scanning cursor, otherwise KSM pages in newly forked mms will be 2521cbf86cfeSHugh Dickins * missed: then we might as well insert at the end of the list. 252231dbd01fSIzik Eidus */ 2523cbf86cfeSHugh Dickins if (ksm_run & KSM_RUN_UNMERGE) 252458730ab6SQi Zheng list_add_tail(&slot->mm_node, &ksm_mm_head.slot.mm_node); 2525cbf86cfeSHugh Dickins else 252658730ab6SQi Zheng list_add_tail(&slot->mm_node, &ksm_scan.mm_slot->slot.mm_node); 252731dbd01fSIzik Eidus spin_unlock(&ksm_mmlist_lock); 252831dbd01fSIzik Eidus 2529f8af4da3SHugh Dickins set_bit(MMF_VM_MERGEABLE, &mm->flags); 2530f1f10076SVegard Nossum mmgrab(mm); 25316e158384SHugh Dickins 25326e158384SHugh Dickins if (needs_wakeup) 25336e158384SHugh Dickins wake_up_interruptible(&ksm_thread_wait); 25346e158384SHugh Dickins 2535f8af4da3SHugh Dickins return 0; 2536f8af4da3SHugh Dickins } 2537f8af4da3SHugh Dickins 25381c2fb7a4SAndrea Arcangeli void __ksm_exit(struct mm_struct *mm) 2539f8af4da3SHugh Dickins { 254021fbd591SQi Zheng struct ksm_mm_slot *mm_slot; 254158730ab6SQi Zheng struct mm_slot *slot; 25429ba69294SHugh Dickins int easy_to_free = 0; 2543cd551f97SHugh Dickins 254431dbd01fSIzik Eidus /* 25459ba69294SHugh Dickins * This process is exiting: if it's straightforward (as is the 25469ba69294SHugh Dickins * case when ksmd was never running), free mm_slot immediately. 25479ba69294SHugh Dickins * But if it's at the cursor or has rmap_items linked to it, use 2548c1e8d7c6SMichel Lespinasse * mmap_lock to synchronize with any break_cows before pagetables 25499ba69294SHugh Dickins * are freed, and leave the mm_slot on the list for ksmd to free. 25509ba69294SHugh Dickins * Beware: ksm may already have noticed it exiting and freed the slot. 255131dbd01fSIzik Eidus */ 25529ba69294SHugh Dickins 2553cd551f97SHugh Dickins spin_lock(&ksm_mmlist_lock); 255458730ab6SQi Zheng slot = mm_slot_lookup(mm_slots_hash, mm); 255558730ab6SQi Zheng mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); 25569ba69294SHugh Dickins if (mm_slot && ksm_scan.mm_slot != mm_slot) { 25576514d511SHugh Dickins if (!mm_slot->rmap_list) { 255858730ab6SQi Zheng hash_del(&slot->hash); 255958730ab6SQi Zheng list_del(&slot->mm_node); 25609ba69294SHugh Dickins easy_to_free = 1; 25619ba69294SHugh Dickins } else { 256258730ab6SQi Zheng list_move(&slot->mm_node, 256358730ab6SQi Zheng &ksm_scan.mm_slot->slot.mm_node); 25649ba69294SHugh Dickins } 25659ba69294SHugh Dickins } 2566cd551f97SHugh Dickins spin_unlock(&ksm_mmlist_lock); 2567cd551f97SHugh Dickins 25689ba69294SHugh Dickins if (easy_to_free) { 256958730ab6SQi Zheng mm_slot_free(mm_slot_cache, mm_slot); 2570cd551f97SHugh Dickins clear_bit(MMF_VM_MERGEABLE, &mm->flags); 25719ba69294SHugh Dickins mmdrop(mm); 25729ba69294SHugh Dickins } else if (mm_slot) { 2573d8ed45c5SMichel Lespinasse mmap_write_lock(mm); 2574d8ed45c5SMichel Lespinasse mmap_write_unlock(mm); 25759ba69294SHugh Dickins } 2576f8af4da3SHugh Dickins } 257731dbd01fSIzik Eidus 2578cbf86cfeSHugh Dickins struct page *ksm_might_need_to_copy(struct page *page, 25795ad64688SHugh Dickins struct vm_area_struct *vma, unsigned long address) 25805ad64688SHugh Dickins { 2581e05b3453SMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 2582e05b3453SMatthew Wilcox (Oracle) struct anon_vma *anon_vma = folio_anon_vma(folio); 25835ad64688SHugh Dickins struct page *new_page; 25845ad64688SHugh Dickins 2585cbf86cfeSHugh Dickins if (PageKsm(page)) { 2586cbf86cfeSHugh Dickins if (page_stable_node(page) && 2587cbf86cfeSHugh Dickins !(ksm_run & KSM_RUN_UNMERGE)) 2588cbf86cfeSHugh Dickins return page; /* no need to copy it */ 2589cbf86cfeSHugh Dickins } else if (!anon_vma) { 2590cbf86cfeSHugh Dickins return page; /* no need to copy it */ 2591e1c63e11SNanyong Sun } else if (page->index == linear_page_index(vma, address) && 2592e1c63e11SNanyong Sun anon_vma->root == vma->anon_vma->root) { 2593cbf86cfeSHugh Dickins return page; /* still no need to copy it */ 2594cbf86cfeSHugh Dickins } 2595cbf86cfeSHugh Dickins if (!PageUptodate(page)) 2596cbf86cfeSHugh Dickins return page; /* let do_swap_page report the error */ 2597cbf86cfeSHugh Dickins 25985ad64688SHugh Dickins new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); 25998f425e4eSMatthew Wilcox (Oracle) if (new_page && 26008f425e4eSMatthew Wilcox (Oracle) mem_cgroup_charge(page_folio(new_page), vma->vm_mm, GFP_KERNEL)) { 260162fdb163SHugh Dickins put_page(new_page); 260262fdb163SHugh Dickins new_page = NULL; 260362fdb163SHugh Dickins } 26045ad64688SHugh Dickins if (new_page) { 26055ad64688SHugh Dickins copy_user_highpage(new_page, page, address, vma); 26065ad64688SHugh Dickins 26075ad64688SHugh Dickins SetPageDirty(new_page); 26085ad64688SHugh Dickins __SetPageUptodate(new_page); 260948c935adSKirill A. Shutemov __SetPageLocked(new_page); 26104d45c3afSYang Yang #ifdef CONFIG_SWAP 26114d45c3afSYang Yang count_vm_event(KSM_SWPIN_COPY); 26124d45c3afSYang Yang #endif 26135ad64688SHugh Dickins } 26145ad64688SHugh Dickins 26155ad64688SHugh Dickins return new_page; 26165ad64688SHugh Dickins } 26175ad64688SHugh Dickins 26186d4675e6SMinchan Kim void rmap_walk_ksm(struct folio *folio, struct rmap_walk_control *rwc) 2619e9995ef9SHugh Dickins { 262021fbd591SQi Zheng struct ksm_stable_node *stable_node; 262121fbd591SQi Zheng struct ksm_rmap_item *rmap_item; 2622e9995ef9SHugh Dickins int search_new_forks = 0; 2623e9995ef9SHugh Dickins 26242f031c6fSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_ksm(folio), folio); 26259f32624bSJoonsoo Kim 26269f32624bSJoonsoo Kim /* 26279f32624bSJoonsoo Kim * Rely on the page lock to protect against concurrent modifications 26289f32624bSJoonsoo Kim * to that page's node of the stable tree. 26299f32624bSJoonsoo Kim */ 26302f031c6fSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 2631e9995ef9SHugh Dickins 26322f031c6fSMatthew Wilcox (Oracle) stable_node = folio_stable_node(folio); 2633e9995ef9SHugh Dickins if (!stable_node) 26341df631aeSMinchan Kim return; 2635e9995ef9SHugh Dickins again: 2636b67bfe0dSSasha Levin hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { 2637e9995ef9SHugh Dickins struct anon_vma *anon_vma = rmap_item->anon_vma; 26385beb4930SRik van Riel struct anon_vma_chain *vmac; 2639e9995ef9SHugh Dickins struct vm_area_struct *vma; 2640e9995ef9SHugh Dickins 2641ad12695fSAndrea Arcangeli cond_resched(); 26426d4675e6SMinchan Kim if (!anon_vma_trylock_read(anon_vma)) { 26436d4675e6SMinchan Kim if (rwc->try_lock) { 26446d4675e6SMinchan Kim rwc->contended = true; 26456d4675e6SMinchan Kim return; 26466d4675e6SMinchan Kim } 2647b6b19f25SHugh Dickins anon_vma_lock_read(anon_vma); 26486d4675e6SMinchan Kim } 2649bf181b9fSMichel Lespinasse anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, 2650bf181b9fSMichel Lespinasse 0, ULONG_MAX) { 26511105a2fcSJia He unsigned long addr; 26521105a2fcSJia He 2653ad12695fSAndrea Arcangeli cond_resched(); 26545beb4930SRik van Riel vma = vmac->vma; 26551105a2fcSJia He 26561105a2fcSJia He /* Ignore the stable/unstable/sqnr flags */ 2657cd7fae26SMiaohe Lin addr = rmap_item->address & PAGE_MASK; 26581105a2fcSJia He 26591105a2fcSJia He if (addr < vma->vm_start || addr >= vma->vm_end) 2660e9995ef9SHugh Dickins continue; 2661e9995ef9SHugh Dickins /* 2662e9995ef9SHugh Dickins * Initially we examine only the vma which covers this 2663e9995ef9SHugh Dickins * rmap_item; but later, if there is still work to do, 2664e9995ef9SHugh Dickins * we examine covering vmas in other mms: in case they 2665e9995ef9SHugh Dickins * were forked from the original since ksmd passed. 2666e9995ef9SHugh Dickins */ 2667e9995ef9SHugh Dickins if ((rmap_item->mm == vma->vm_mm) == search_new_forks) 2668e9995ef9SHugh Dickins continue; 2669e9995ef9SHugh Dickins 26700dd1c7bbSJoonsoo Kim if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) 26710dd1c7bbSJoonsoo Kim continue; 26720dd1c7bbSJoonsoo Kim 26732f031c6fSMatthew Wilcox (Oracle) if (!rwc->rmap_one(folio, vma, addr, rwc->arg)) { 2674b6b19f25SHugh Dickins anon_vma_unlock_read(anon_vma); 26751df631aeSMinchan Kim return; 2676e9995ef9SHugh Dickins } 26772f031c6fSMatthew Wilcox (Oracle) if (rwc->done && rwc->done(folio)) { 26780dd1c7bbSJoonsoo Kim anon_vma_unlock_read(anon_vma); 26791df631aeSMinchan Kim return; 26800dd1c7bbSJoonsoo Kim } 2681e9995ef9SHugh Dickins } 2682b6b19f25SHugh Dickins anon_vma_unlock_read(anon_vma); 2683e9995ef9SHugh Dickins } 2684e9995ef9SHugh Dickins if (!search_new_forks++) 2685e9995ef9SHugh Dickins goto again; 2686e9995ef9SHugh Dickins } 2687e9995ef9SHugh Dickins 268852629506SJoonsoo Kim #ifdef CONFIG_MIGRATION 268919138349SMatthew Wilcox (Oracle) void folio_migrate_ksm(struct folio *newfolio, struct folio *folio) 2690e9995ef9SHugh Dickins { 269121fbd591SQi Zheng struct ksm_stable_node *stable_node; 2692e9995ef9SHugh Dickins 269319138349SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 269419138349SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(newfolio), newfolio); 269519138349SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(newfolio->mapping != folio->mapping, newfolio); 2696e9995ef9SHugh Dickins 269719138349SMatthew Wilcox (Oracle) stable_node = folio_stable_node(folio); 2698e9995ef9SHugh Dickins if (stable_node) { 269919138349SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(stable_node->kpfn != folio_pfn(folio), folio); 270019138349SMatthew Wilcox (Oracle) stable_node->kpfn = folio_pfn(newfolio); 2701c8d6553bSHugh Dickins /* 270219138349SMatthew Wilcox (Oracle) * newfolio->mapping was set in advance; now we need smp_wmb() 2703c8d6553bSHugh Dickins * to make sure that the new stable_node->kpfn is visible 270419138349SMatthew Wilcox (Oracle) * to get_ksm_page() before it can see that folio->mapping 270519138349SMatthew Wilcox (Oracle) * has gone stale (or that folio_test_swapcache has been cleared). 2706c8d6553bSHugh Dickins */ 2707c8d6553bSHugh Dickins smp_wmb(); 270819138349SMatthew Wilcox (Oracle) set_page_stable_node(&folio->page, NULL); 2709e9995ef9SHugh Dickins } 2710e9995ef9SHugh Dickins } 2711e9995ef9SHugh Dickins #endif /* CONFIG_MIGRATION */ 2712e9995ef9SHugh Dickins 271362b61f61SHugh Dickins #ifdef CONFIG_MEMORY_HOTREMOVE 2714ef4d43a8SHugh Dickins static void wait_while_offlining(void) 2715ef4d43a8SHugh Dickins { 2716ef4d43a8SHugh Dickins while (ksm_run & KSM_RUN_OFFLINE) { 2717ef4d43a8SHugh Dickins mutex_unlock(&ksm_thread_mutex); 2718ef4d43a8SHugh Dickins wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE), 271974316201SNeilBrown TASK_UNINTERRUPTIBLE); 2720ef4d43a8SHugh Dickins mutex_lock(&ksm_thread_mutex); 2721ef4d43a8SHugh Dickins } 2722ef4d43a8SHugh Dickins } 2723ef4d43a8SHugh Dickins 272421fbd591SQi Zheng static bool stable_node_dup_remove_range(struct ksm_stable_node *stable_node, 27252c653d0eSAndrea Arcangeli unsigned long start_pfn, 27262c653d0eSAndrea Arcangeli unsigned long end_pfn) 27272c653d0eSAndrea Arcangeli { 27282c653d0eSAndrea Arcangeli if (stable_node->kpfn >= start_pfn && 27292c653d0eSAndrea Arcangeli stable_node->kpfn < end_pfn) { 27302c653d0eSAndrea Arcangeli /* 27312c653d0eSAndrea Arcangeli * Don't get_ksm_page, page has already gone: 27322c653d0eSAndrea Arcangeli * which is why we keep kpfn instead of page* 27332c653d0eSAndrea Arcangeli */ 27342c653d0eSAndrea Arcangeli remove_node_from_stable_tree(stable_node); 27352c653d0eSAndrea Arcangeli return true; 27362c653d0eSAndrea Arcangeli } 27372c653d0eSAndrea Arcangeli return false; 27382c653d0eSAndrea Arcangeli } 27392c653d0eSAndrea Arcangeli 274021fbd591SQi Zheng static bool stable_node_chain_remove_range(struct ksm_stable_node *stable_node, 27412c653d0eSAndrea Arcangeli unsigned long start_pfn, 27422c653d0eSAndrea Arcangeli unsigned long end_pfn, 27432c653d0eSAndrea Arcangeli struct rb_root *root) 27442c653d0eSAndrea Arcangeli { 274521fbd591SQi Zheng struct ksm_stable_node *dup; 27462c653d0eSAndrea Arcangeli struct hlist_node *hlist_safe; 27472c653d0eSAndrea Arcangeli 27482c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 27492c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node)); 27502c653d0eSAndrea Arcangeli return stable_node_dup_remove_range(stable_node, start_pfn, 27512c653d0eSAndrea Arcangeli end_pfn); 27522c653d0eSAndrea Arcangeli } 27532c653d0eSAndrea Arcangeli 27542c653d0eSAndrea Arcangeli hlist_for_each_entry_safe(dup, hlist_safe, 27552c653d0eSAndrea Arcangeli &stable_node->hlist, hlist_dup) { 27562c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(dup)); 27572c653d0eSAndrea Arcangeli stable_node_dup_remove_range(dup, start_pfn, end_pfn); 27582c653d0eSAndrea Arcangeli } 27592c653d0eSAndrea Arcangeli if (hlist_empty(&stable_node->hlist)) { 27602c653d0eSAndrea Arcangeli free_stable_node_chain(stable_node, root); 27612c653d0eSAndrea Arcangeli return true; /* notify caller that tree was rebalanced */ 27622c653d0eSAndrea Arcangeli } else 27632c653d0eSAndrea Arcangeli return false; 27642c653d0eSAndrea Arcangeli } 27652c653d0eSAndrea Arcangeli 2766ee0ea59cSHugh Dickins static void ksm_check_stable_tree(unsigned long start_pfn, 276762b61f61SHugh Dickins unsigned long end_pfn) 276862b61f61SHugh Dickins { 276921fbd591SQi Zheng struct ksm_stable_node *stable_node, *next; 277062b61f61SHugh Dickins struct rb_node *node; 277190bd6fd3SPetr Holasek int nid; 277262b61f61SHugh Dickins 2773ef53d16cSHugh Dickins for (nid = 0; nid < ksm_nr_node_ids; nid++) { 2774ef53d16cSHugh Dickins node = rb_first(root_stable_tree + nid); 2775ee0ea59cSHugh Dickins while (node) { 277621fbd591SQi Zheng stable_node = rb_entry(node, struct ksm_stable_node, node); 27772c653d0eSAndrea Arcangeli if (stable_node_chain_remove_range(stable_node, 27782c653d0eSAndrea Arcangeli start_pfn, end_pfn, 27792c653d0eSAndrea Arcangeli root_stable_tree + 27802c653d0eSAndrea Arcangeli nid)) 2781ef53d16cSHugh Dickins node = rb_first(root_stable_tree + nid); 27822c653d0eSAndrea Arcangeli else 2783ee0ea59cSHugh Dickins node = rb_next(node); 2784ee0ea59cSHugh Dickins cond_resched(); 278562b61f61SHugh Dickins } 2786ee0ea59cSHugh Dickins } 278703640418SGeliang Tang list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { 27884146d2d6SHugh Dickins if (stable_node->kpfn >= start_pfn && 27894146d2d6SHugh Dickins stable_node->kpfn < end_pfn) 27904146d2d6SHugh Dickins remove_node_from_stable_tree(stable_node); 27914146d2d6SHugh Dickins cond_resched(); 27924146d2d6SHugh Dickins } 279362b61f61SHugh Dickins } 279462b61f61SHugh Dickins 279562b61f61SHugh Dickins static int ksm_memory_callback(struct notifier_block *self, 279662b61f61SHugh Dickins unsigned long action, void *arg) 279762b61f61SHugh Dickins { 279862b61f61SHugh Dickins struct memory_notify *mn = arg; 279962b61f61SHugh Dickins 280062b61f61SHugh Dickins switch (action) { 280162b61f61SHugh Dickins case MEM_GOING_OFFLINE: 280262b61f61SHugh Dickins /* 2803ef4d43a8SHugh Dickins * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items() 2804ef4d43a8SHugh Dickins * and remove_all_stable_nodes() while memory is going offline: 2805ef4d43a8SHugh Dickins * it is unsafe for them to touch the stable tree at this time. 2806ef4d43a8SHugh Dickins * But unmerge_ksm_pages(), rmap lookups and other entry points 2807ef4d43a8SHugh Dickins * which do not need the ksm_thread_mutex are all safe. 280862b61f61SHugh Dickins */ 2809ef4d43a8SHugh Dickins mutex_lock(&ksm_thread_mutex); 2810ef4d43a8SHugh Dickins ksm_run |= KSM_RUN_OFFLINE; 2811ef4d43a8SHugh Dickins mutex_unlock(&ksm_thread_mutex); 281262b61f61SHugh Dickins break; 281362b61f61SHugh Dickins 281462b61f61SHugh Dickins case MEM_OFFLINE: 281562b61f61SHugh Dickins /* 281662b61f61SHugh Dickins * Most of the work is done by page migration; but there might 281762b61f61SHugh Dickins * be a few stable_nodes left over, still pointing to struct 2818ee0ea59cSHugh Dickins * pages which have been offlined: prune those from the tree, 2819ee0ea59cSHugh Dickins * otherwise get_ksm_page() might later try to access a 2820ee0ea59cSHugh Dickins * non-existent struct page. 282162b61f61SHugh Dickins */ 2822ee0ea59cSHugh Dickins ksm_check_stable_tree(mn->start_pfn, 2823ee0ea59cSHugh Dickins mn->start_pfn + mn->nr_pages); 2824e4a9bc58SJoe Perches fallthrough; 282562b61f61SHugh Dickins case MEM_CANCEL_OFFLINE: 2826ef4d43a8SHugh Dickins mutex_lock(&ksm_thread_mutex); 2827ef4d43a8SHugh Dickins ksm_run &= ~KSM_RUN_OFFLINE; 282862b61f61SHugh Dickins mutex_unlock(&ksm_thread_mutex); 2829ef4d43a8SHugh Dickins 2830ef4d43a8SHugh Dickins smp_mb(); /* wake_up_bit advises this */ 2831ef4d43a8SHugh Dickins wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE)); 283262b61f61SHugh Dickins break; 283362b61f61SHugh Dickins } 283462b61f61SHugh Dickins return NOTIFY_OK; 283562b61f61SHugh Dickins } 2836ef4d43a8SHugh Dickins #else 2837ef4d43a8SHugh Dickins static void wait_while_offlining(void) 2838ef4d43a8SHugh Dickins { 2839ef4d43a8SHugh Dickins } 284062b61f61SHugh Dickins #endif /* CONFIG_MEMORY_HOTREMOVE */ 284162b61f61SHugh Dickins 28422ffd8679SHugh Dickins #ifdef CONFIG_SYSFS 28432ffd8679SHugh Dickins /* 28442ffd8679SHugh Dickins * This all compiles without CONFIG_SYSFS, but is a waste of space. 28452ffd8679SHugh Dickins */ 28462ffd8679SHugh Dickins 284731dbd01fSIzik Eidus #define KSM_ATTR_RO(_name) \ 284831dbd01fSIzik Eidus static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 284931dbd01fSIzik Eidus #define KSM_ATTR(_name) \ 28501bad2e5cSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 285131dbd01fSIzik Eidus 285231dbd01fSIzik Eidus static ssize_t sleep_millisecs_show(struct kobject *kobj, 285331dbd01fSIzik Eidus struct kobj_attribute *attr, char *buf) 285431dbd01fSIzik Eidus { 2855ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs); 285631dbd01fSIzik Eidus } 285731dbd01fSIzik Eidus 285831dbd01fSIzik Eidus static ssize_t sleep_millisecs_store(struct kobject *kobj, 285931dbd01fSIzik Eidus struct kobj_attribute *attr, 286031dbd01fSIzik Eidus const char *buf, size_t count) 286131dbd01fSIzik Eidus { 2862dfefd226SAlexey Dobriyan unsigned int msecs; 286331dbd01fSIzik Eidus int err; 286431dbd01fSIzik Eidus 2865dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &msecs); 2866dfefd226SAlexey Dobriyan if (err) 286731dbd01fSIzik Eidus return -EINVAL; 286831dbd01fSIzik Eidus 286931dbd01fSIzik Eidus ksm_thread_sleep_millisecs = msecs; 2870fcf9a0efSKirill Tkhai wake_up_interruptible(&ksm_iter_wait); 287131dbd01fSIzik Eidus 287231dbd01fSIzik Eidus return count; 287331dbd01fSIzik Eidus } 287431dbd01fSIzik Eidus KSM_ATTR(sleep_millisecs); 287531dbd01fSIzik Eidus 287631dbd01fSIzik Eidus static ssize_t pages_to_scan_show(struct kobject *kobj, 287731dbd01fSIzik Eidus struct kobj_attribute *attr, char *buf) 287831dbd01fSIzik Eidus { 2879ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan); 288031dbd01fSIzik Eidus } 288131dbd01fSIzik Eidus 288231dbd01fSIzik Eidus static ssize_t pages_to_scan_store(struct kobject *kobj, 288331dbd01fSIzik Eidus struct kobj_attribute *attr, 288431dbd01fSIzik Eidus const char *buf, size_t count) 288531dbd01fSIzik Eidus { 2886dfefd226SAlexey Dobriyan unsigned int nr_pages; 288731dbd01fSIzik Eidus int err; 288831dbd01fSIzik Eidus 2889dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &nr_pages); 2890dfefd226SAlexey Dobriyan if (err) 289131dbd01fSIzik Eidus return -EINVAL; 289231dbd01fSIzik Eidus 289331dbd01fSIzik Eidus ksm_thread_pages_to_scan = nr_pages; 289431dbd01fSIzik Eidus 289531dbd01fSIzik Eidus return count; 289631dbd01fSIzik Eidus } 289731dbd01fSIzik Eidus KSM_ATTR(pages_to_scan); 289831dbd01fSIzik Eidus 289931dbd01fSIzik Eidus static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, 290031dbd01fSIzik Eidus char *buf) 290131dbd01fSIzik Eidus { 2902ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", ksm_run); 290331dbd01fSIzik Eidus } 290431dbd01fSIzik Eidus 290531dbd01fSIzik Eidus static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, 290631dbd01fSIzik Eidus const char *buf, size_t count) 290731dbd01fSIzik Eidus { 2908dfefd226SAlexey Dobriyan unsigned int flags; 290931dbd01fSIzik Eidus int err; 291031dbd01fSIzik Eidus 2911dfefd226SAlexey Dobriyan err = kstrtouint(buf, 10, &flags); 2912dfefd226SAlexey Dobriyan if (err) 291331dbd01fSIzik Eidus return -EINVAL; 291431dbd01fSIzik Eidus if (flags > KSM_RUN_UNMERGE) 291531dbd01fSIzik Eidus return -EINVAL; 291631dbd01fSIzik Eidus 291731dbd01fSIzik Eidus /* 291831dbd01fSIzik Eidus * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. 291931dbd01fSIzik Eidus * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, 2920d0f209f6SHugh Dickins * breaking COW to free the pages_shared (but leaves mm_slots 2921d0f209f6SHugh Dickins * on the list for when ksmd may be set running again). 292231dbd01fSIzik Eidus */ 292331dbd01fSIzik Eidus 292431dbd01fSIzik Eidus mutex_lock(&ksm_thread_mutex); 2925ef4d43a8SHugh Dickins wait_while_offlining(); 292631dbd01fSIzik Eidus if (ksm_run != flags) { 292731dbd01fSIzik Eidus ksm_run = flags; 2928d952b791SHugh Dickins if (flags & KSM_RUN_UNMERGE) { 2929e1e12d2fSDavid Rientjes set_current_oom_origin(); 2930d952b791SHugh Dickins err = unmerge_and_remove_all_rmap_items(); 2931e1e12d2fSDavid Rientjes clear_current_oom_origin(); 2932d952b791SHugh Dickins if (err) { 2933d952b791SHugh Dickins ksm_run = KSM_RUN_STOP; 2934d952b791SHugh Dickins count = err; 2935d952b791SHugh Dickins } 2936d952b791SHugh Dickins } 293731dbd01fSIzik Eidus } 293831dbd01fSIzik Eidus mutex_unlock(&ksm_thread_mutex); 293931dbd01fSIzik Eidus 294031dbd01fSIzik Eidus if (flags & KSM_RUN_MERGE) 294131dbd01fSIzik Eidus wake_up_interruptible(&ksm_thread_wait); 294231dbd01fSIzik Eidus 294331dbd01fSIzik Eidus return count; 294431dbd01fSIzik Eidus } 294531dbd01fSIzik Eidus KSM_ATTR(run); 294631dbd01fSIzik Eidus 294790bd6fd3SPetr Holasek #ifdef CONFIG_NUMA 294890bd6fd3SPetr Holasek static ssize_t merge_across_nodes_show(struct kobject *kobj, 294990bd6fd3SPetr Holasek struct kobj_attribute *attr, char *buf) 295090bd6fd3SPetr Holasek { 2951ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes); 295290bd6fd3SPetr Holasek } 295390bd6fd3SPetr Holasek 295490bd6fd3SPetr Holasek static ssize_t merge_across_nodes_store(struct kobject *kobj, 295590bd6fd3SPetr Holasek struct kobj_attribute *attr, 295690bd6fd3SPetr Holasek const char *buf, size_t count) 295790bd6fd3SPetr Holasek { 295890bd6fd3SPetr Holasek int err; 295990bd6fd3SPetr Holasek unsigned long knob; 296090bd6fd3SPetr Holasek 296190bd6fd3SPetr Holasek err = kstrtoul(buf, 10, &knob); 296290bd6fd3SPetr Holasek if (err) 296390bd6fd3SPetr Holasek return err; 296490bd6fd3SPetr Holasek if (knob > 1) 296590bd6fd3SPetr Holasek return -EINVAL; 296690bd6fd3SPetr Holasek 296790bd6fd3SPetr Holasek mutex_lock(&ksm_thread_mutex); 2968ef4d43a8SHugh Dickins wait_while_offlining(); 296990bd6fd3SPetr Holasek if (ksm_merge_across_nodes != knob) { 2970cbf86cfeSHugh Dickins if (ksm_pages_shared || remove_all_stable_nodes()) 297190bd6fd3SPetr Holasek err = -EBUSY; 2972ef53d16cSHugh Dickins else if (root_stable_tree == one_stable_tree) { 2973ef53d16cSHugh Dickins struct rb_root *buf; 2974ef53d16cSHugh Dickins /* 2975ef53d16cSHugh Dickins * This is the first time that we switch away from the 2976ef53d16cSHugh Dickins * default of merging across nodes: must now allocate 2977ef53d16cSHugh Dickins * a buffer to hold as many roots as may be needed. 2978ef53d16cSHugh Dickins * Allocate stable and unstable together: 2979ef53d16cSHugh Dickins * MAXSMP NODES_SHIFT 10 will use 16kB. 2980ef53d16cSHugh Dickins */ 2981bafe1e14SJoe Perches buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf), 2982bafe1e14SJoe Perches GFP_KERNEL); 2983ef53d16cSHugh Dickins /* Let us assume that RB_ROOT is NULL is zero */ 2984ef53d16cSHugh Dickins if (!buf) 2985ef53d16cSHugh Dickins err = -ENOMEM; 2986ef53d16cSHugh Dickins else { 2987ef53d16cSHugh Dickins root_stable_tree = buf; 2988ef53d16cSHugh Dickins root_unstable_tree = buf + nr_node_ids; 2989ef53d16cSHugh Dickins /* Stable tree is empty but not the unstable */ 2990ef53d16cSHugh Dickins root_unstable_tree[0] = one_unstable_tree[0]; 2991ef53d16cSHugh Dickins } 2992ef53d16cSHugh Dickins } 2993ef53d16cSHugh Dickins if (!err) { 299490bd6fd3SPetr Holasek ksm_merge_across_nodes = knob; 2995ef53d16cSHugh Dickins ksm_nr_node_ids = knob ? 1 : nr_node_ids; 2996ef53d16cSHugh Dickins } 299790bd6fd3SPetr Holasek } 299890bd6fd3SPetr Holasek mutex_unlock(&ksm_thread_mutex); 299990bd6fd3SPetr Holasek 300090bd6fd3SPetr Holasek return err ? err : count; 300190bd6fd3SPetr Holasek } 300290bd6fd3SPetr Holasek KSM_ATTR(merge_across_nodes); 300390bd6fd3SPetr Holasek #endif 300490bd6fd3SPetr Holasek 3005e86c59b1SClaudio Imbrenda static ssize_t use_zero_pages_show(struct kobject *kobj, 3006e86c59b1SClaudio Imbrenda struct kobj_attribute *attr, char *buf) 3007e86c59b1SClaudio Imbrenda { 3008ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", ksm_use_zero_pages); 3009e86c59b1SClaudio Imbrenda } 3010e86c59b1SClaudio Imbrenda static ssize_t use_zero_pages_store(struct kobject *kobj, 3011e86c59b1SClaudio Imbrenda struct kobj_attribute *attr, 3012e86c59b1SClaudio Imbrenda const char *buf, size_t count) 3013e86c59b1SClaudio Imbrenda { 3014e86c59b1SClaudio Imbrenda int err; 3015e86c59b1SClaudio Imbrenda bool value; 3016e86c59b1SClaudio Imbrenda 3017e86c59b1SClaudio Imbrenda err = kstrtobool(buf, &value); 3018e86c59b1SClaudio Imbrenda if (err) 3019e86c59b1SClaudio Imbrenda return -EINVAL; 3020e86c59b1SClaudio Imbrenda 3021e86c59b1SClaudio Imbrenda ksm_use_zero_pages = value; 3022e86c59b1SClaudio Imbrenda 3023e86c59b1SClaudio Imbrenda return count; 3024e86c59b1SClaudio Imbrenda } 3025e86c59b1SClaudio Imbrenda KSM_ATTR(use_zero_pages); 3026e86c59b1SClaudio Imbrenda 30272c653d0eSAndrea Arcangeli static ssize_t max_page_sharing_show(struct kobject *kobj, 30282c653d0eSAndrea Arcangeli struct kobj_attribute *attr, char *buf) 30292c653d0eSAndrea Arcangeli { 3030ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", ksm_max_page_sharing); 30312c653d0eSAndrea Arcangeli } 30322c653d0eSAndrea Arcangeli 30332c653d0eSAndrea Arcangeli static ssize_t max_page_sharing_store(struct kobject *kobj, 30342c653d0eSAndrea Arcangeli struct kobj_attribute *attr, 30352c653d0eSAndrea Arcangeli const char *buf, size_t count) 30362c653d0eSAndrea Arcangeli { 30372c653d0eSAndrea Arcangeli int err; 30382c653d0eSAndrea Arcangeli int knob; 30392c653d0eSAndrea Arcangeli 30402c653d0eSAndrea Arcangeli err = kstrtoint(buf, 10, &knob); 30412c653d0eSAndrea Arcangeli if (err) 30422c653d0eSAndrea Arcangeli return err; 30432c653d0eSAndrea Arcangeli /* 30442c653d0eSAndrea Arcangeli * When a KSM page is created it is shared by 2 mappings. This 30452c653d0eSAndrea Arcangeli * being a signed comparison, it implicitly verifies it's not 30462c653d0eSAndrea Arcangeli * negative. 30472c653d0eSAndrea Arcangeli */ 30482c653d0eSAndrea Arcangeli if (knob < 2) 30492c653d0eSAndrea Arcangeli return -EINVAL; 30502c653d0eSAndrea Arcangeli 30512c653d0eSAndrea Arcangeli if (READ_ONCE(ksm_max_page_sharing) == knob) 30522c653d0eSAndrea Arcangeli return count; 30532c653d0eSAndrea Arcangeli 30542c653d0eSAndrea Arcangeli mutex_lock(&ksm_thread_mutex); 30552c653d0eSAndrea Arcangeli wait_while_offlining(); 30562c653d0eSAndrea Arcangeli if (ksm_max_page_sharing != knob) { 30572c653d0eSAndrea Arcangeli if (ksm_pages_shared || remove_all_stable_nodes()) 30582c653d0eSAndrea Arcangeli err = -EBUSY; 30592c653d0eSAndrea Arcangeli else 30602c653d0eSAndrea Arcangeli ksm_max_page_sharing = knob; 30612c653d0eSAndrea Arcangeli } 30622c653d0eSAndrea Arcangeli mutex_unlock(&ksm_thread_mutex); 30632c653d0eSAndrea Arcangeli 30642c653d0eSAndrea Arcangeli return err ? err : count; 30652c653d0eSAndrea Arcangeli } 30662c653d0eSAndrea Arcangeli KSM_ATTR(max_page_sharing); 30672c653d0eSAndrea Arcangeli 3068b4028260SHugh Dickins static ssize_t pages_shared_show(struct kobject *kobj, 3069b4028260SHugh Dickins struct kobj_attribute *attr, char *buf) 3070b4028260SHugh Dickins { 3071ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", ksm_pages_shared); 3072b4028260SHugh Dickins } 3073b4028260SHugh Dickins KSM_ATTR_RO(pages_shared); 3074b4028260SHugh Dickins 3075b4028260SHugh Dickins static ssize_t pages_sharing_show(struct kobject *kobj, 3076b4028260SHugh Dickins struct kobj_attribute *attr, char *buf) 3077b4028260SHugh Dickins { 3078ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", ksm_pages_sharing); 3079b4028260SHugh Dickins } 3080b4028260SHugh Dickins KSM_ATTR_RO(pages_sharing); 3081b4028260SHugh Dickins 3082473b0ce4SHugh Dickins static ssize_t pages_unshared_show(struct kobject *kobj, 3083473b0ce4SHugh Dickins struct kobj_attribute *attr, char *buf) 3084473b0ce4SHugh Dickins { 3085ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", ksm_pages_unshared); 3086473b0ce4SHugh Dickins } 3087473b0ce4SHugh Dickins KSM_ATTR_RO(pages_unshared); 3088473b0ce4SHugh Dickins 3089473b0ce4SHugh Dickins static ssize_t pages_volatile_show(struct kobject *kobj, 3090473b0ce4SHugh Dickins struct kobj_attribute *attr, char *buf) 3091473b0ce4SHugh Dickins { 3092473b0ce4SHugh Dickins long ksm_pages_volatile; 3093473b0ce4SHugh Dickins 3094473b0ce4SHugh Dickins ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared 3095473b0ce4SHugh Dickins - ksm_pages_sharing - ksm_pages_unshared; 3096473b0ce4SHugh Dickins /* 3097473b0ce4SHugh Dickins * It was not worth any locking to calculate that statistic, 3098473b0ce4SHugh Dickins * but it might therefore sometimes be negative: conceal that. 3099473b0ce4SHugh Dickins */ 3100473b0ce4SHugh Dickins if (ksm_pages_volatile < 0) 3101473b0ce4SHugh Dickins ksm_pages_volatile = 0; 3102ae7a927dSJoe Perches return sysfs_emit(buf, "%ld\n", ksm_pages_volatile); 3103473b0ce4SHugh Dickins } 3104473b0ce4SHugh Dickins KSM_ATTR_RO(pages_volatile); 3105473b0ce4SHugh Dickins 31062c653d0eSAndrea Arcangeli static ssize_t stable_node_dups_show(struct kobject *kobj, 31072c653d0eSAndrea Arcangeli struct kobj_attribute *attr, char *buf) 31082c653d0eSAndrea Arcangeli { 3109ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups); 31102c653d0eSAndrea Arcangeli } 31112c653d0eSAndrea Arcangeli KSM_ATTR_RO(stable_node_dups); 31122c653d0eSAndrea Arcangeli 31132c653d0eSAndrea Arcangeli static ssize_t stable_node_chains_show(struct kobject *kobj, 31142c653d0eSAndrea Arcangeli struct kobj_attribute *attr, char *buf) 31152c653d0eSAndrea Arcangeli { 3116ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains); 31172c653d0eSAndrea Arcangeli } 31182c653d0eSAndrea Arcangeli KSM_ATTR_RO(stable_node_chains); 31192c653d0eSAndrea Arcangeli 31202c653d0eSAndrea Arcangeli static ssize_t 31212c653d0eSAndrea Arcangeli stable_node_chains_prune_millisecs_show(struct kobject *kobj, 31222c653d0eSAndrea Arcangeli struct kobj_attribute *attr, 31232c653d0eSAndrea Arcangeli char *buf) 31242c653d0eSAndrea Arcangeli { 3125ae7a927dSJoe Perches return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs); 31262c653d0eSAndrea Arcangeli } 31272c653d0eSAndrea Arcangeli 31282c653d0eSAndrea Arcangeli static ssize_t 31292c653d0eSAndrea Arcangeli stable_node_chains_prune_millisecs_store(struct kobject *kobj, 31302c653d0eSAndrea Arcangeli struct kobj_attribute *attr, 31312c653d0eSAndrea Arcangeli const char *buf, size_t count) 31322c653d0eSAndrea Arcangeli { 3133584ff0dfSZhansaya Bagdauletkyzy unsigned int msecs; 31342c653d0eSAndrea Arcangeli int err; 31352c653d0eSAndrea Arcangeli 3136584ff0dfSZhansaya Bagdauletkyzy err = kstrtouint(buf, 10, &msecs); 3137584ff0dfSZhansaya Bagdauletkyzy if (err) 31382c653d0eSAndrea Arcangeli return -EINVAL; 31392c653d0eSAndrea Arcangeli 31402c653d0eSAndrea Arcangeli ksm_stable_node_chains_prune_millisecs = msecs; 31412c653d0eSAndrea Arcangeli 31422c653d0eSAndrea Arcangeli return count; 31432c653d0eSAndrea Arcangeli } 31442c653d0eSAndrea Arcangeli KSM_ATTR(stable_node_chains_prune_millisecs); 31452c653d0eSAndrea Arcangeli 3146473b0ce4SHugh Dickins static ssize_t full_scans_show(struct kobject *kobj, 3147473b0ce4SHugh Dickins struct kobj_attribute *attr, char *buf) 3148473b0ce4SHugh Dickins { 3149ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr); 3150473b0ce4SHugh Dickins } 3151473b0ce4SHugh Dickins KSM_ATTR_RO(full_scans); 3152473b0ce4SHugh Dickins 315331dbd01fSIzik Eidus static struct attribute *ksm_attrs[] = { 315431dbd01fSIzik Eidus &sleep_millisecs_attr.attr, 315531dbd01fSIzik Eidus &pages_to_scan_attr.attr, 315631dbd01fSIzik Eidus &run_attr.attr, 3157b4028260SHugh Dickins &pages_shared_attr.attr, 3158b4028260SHugh Dickins &pages_sharing_attr.attr, 3159473b0ce4SHugh Dickins &pages_unshared_attr.attr, 3160473b0ce4SHugh Dickins &pages_volatile_attr.attr, 3161473b0ce4SHugh Dickins &full_scans_attr.attr, 316290bd6fd3SPetr Holasek #ifdef CONFIG_NUMA 316390bd6fd3SPetr Holasek &merge_across_nodes_attr.attr, 316490bd6fd3SPetr Holasek #endif 31652c653d0eSAndrea Arcangeli &max_page_sharing_attr.attr, 31662c653d0eSAndrea Arcangeli &stable_node_chains_attr.attr, 31672c653d0eSAndrea Arcangeli &stable_node_dups_attr.attr, 31682c653d0eSAndrea Arcangeli &stable_node_chains_prune_millisecs_attr.attr, 3169e86c59b1SClaudio Imbrenda &use_zero_pages_attr.attr, 317031dbd01fSIzik Eidus NULL, 317131dbd01fSIzik Eidus }; 317231dbd01fSIzik Eidus 3173f907c26aSArvind Yadav static const struct attribute_group ksm_attr_group = { 317431dbd01fSIzik Eidus .attrs = ksm_attrs, 317531dbd01fSIzik Eidus .name = "ksm", 317631dbd01fSIzik Eidus }; 31772ffd8679SHugh Dickins #endif /* CONFIG_SYSFS */ 317831dbd01fSIzik Eidus 317931dbd01fSIzik Eidus static int __init ksm_init(void) 318031dbd01fSIzik Eidus { 318131dbd01fSIzik Eidus struct task_struct *ksm_thread; 318231dbd01fSIzik Eidus int err; 318331dbd01fSIzik Eidus 3184e86c59b1SClaudio Imbrenda /* The correct value depends on page size and endianness */ 3185e86c59b1SClaudio Imbrenda zero_checksum = calc_checksum(ZERO_PAGE(0)); 3186e86c59b1SClaudio Imbrenda /* Default to false for backwards compatibility */ 3187e86c59b1SClaudio Imbrenda ksm_use_zero_pages = false; 3188e86c59b1SClaudio Imbrenda 318931dbd01fSIzik Eidus err = ksm_slab_init(); 319031dbd01fSIzik Eidus if (err) 319131dbd01fSIzik Eidus goto out; 319231dbd01fSIzik Eidus 319331dbd01fSIzik Eidus ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); 319431dbd01fSIzik Eidus if (IS_ERR(ksm_thread)) { 319525acde31SPaul McQuade pr_err("ksm: creating kthread failed\n"); 319631dbd01fSIzik Eidus err = PTR_ERR(ksm_thread); 3197d9f8984cSLai Jiangshan goto out_free; 319831dbd01fSIzik Eidus } 319931dbd01fSIzik Eidus 32002ffd8679SHugh Dickins #ifdef CONFIG_SYSFS 320131dbd01fSIzik Eidus err = sysfs_create_group(mm_kobj, &ksm_attr_group); 320231dbd01fSIzik Eidus if (err) { 320325acde31SPaul McQuade pr_err("ksm: register sysfs failed\n"); 32042ffd8679SHugh Dickins kthread_stop(ksm_thread); 3205d9f8984cSLai Jiangshan goto out_free; 320631dbd01fSIzik Eidus } 3207c73602adSHugh Dickins #else 3208c73602adSHugh Dickins ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ 3209c73602adSHugh Dickins 32102ffd8679SHugh Dickins #endif /* CONFIG_SYSFS */ 321131dbd01fSIzik Eidus 321262b61f61SHugh Dickins #ifdef CONFIG_MEMORY_HOTREMOVE 3213ef4d43a8SHugh Dickins /* There is no significance to this priority 100 */ 321462b61f61SHugh Dickins hotplug_memory_notifier(ksm_memory_callback, 100); 321562b61f61SHugh Dickins #endif 321631dbd01fSIzik Eidus return 0; 321731dbd01fSIzik Eidus 3218d9f8984cSLai Jiangshan out_free: 321931dbd01fSIzik Eidus ksm_slab_free(); 322031dbd01fSIzik Eidus out: 322131dbd01fSIzik Eidus return err; 322231dbd01fSIzik Eidus } 3223a64fb3cdSPaul Gortmaker subsys_initcall(ksm_init); 3224