1f8af4da3SHugh Dickins /* 231dbd01fSIzik Eidus * Memory merging support. 331dbd01fSIzik Eidus * 431dbd01fSIzik Eidus * This code enables dynamic sharing of identical pages found in different 531dbd01fSIzik Eidus * memory areas, even if they are not shared by fork() 631dbd01fSIzik Eidus * 736b2528dSIzik Eidus * Copyright (C) 2008-2009 Red Hat, Inc. 831dbd01fSIzik Eidus * Authors: 931dbd01fSIzik Eidus * Izik Eidus 1031dbd01fSIzik Eidus * Andrea Arcangeli 1131dbd01fSIzik Eidus * Chris Wright 1236b2528dSIzik Eidus * Hugh Dickins 1331dbd01fSIzik Eidus * 1431dbd01fSIzik Eidus * This work is licensed under the terms of the GNU GPL, version 2. 15f8af4da3SHugh Dickins */ 16f8af4da3SHugh Dickins 17f8af4da3SHugh Dickins #include <linux/errno.h> 1831dbd01fSIzik Eidus #include <linux/mm.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> 2831dbd01fSIzik Eidus #include <linux/jhash.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" 4531dbd01fSIzik Eidus 46e850dcf5SHugh Dickins #ifdef CONFIG_NUMA 47e850dcf5SHugh Dickins #define NUMA(x) (x) 48e850dcf5SHugh Dickins #define DO_NUMA(x) do { (x); } while (0) 49e850dcf5SHugh Dickins #else 50e850dcf5SHugh Dickins #define NUMA(x) (0) 51e850dcf5SHugh Dickins #define DO_NUMA(x) do { } while (0) 52e850dcf5SHugh Dickins #endif 53e850dcf5SHugh Dickins 545a2ca3efSMike Rapoport /** 555a2ca3efSMike Rapoport * DOC: Overview 565a2ca3efSMike Rapoport * 5731dbd01fSIzik Eidus * A few notes about the KSM scanning process, 5831dbd01fSIzik Eidus * to make it easier to understand the data structures below: 5931dbd01fSIzik Eidus * 6031dbd01fSIzik Eidus * In order to reduce excessive scanning, KSM sorts the memory pages by their 6131dbd01fSIzik Eidus * contents into a data structure that holds pointers to the pages' locations. 6231dbd01fSIzik Eidus * 6331dbd01fSIzik Eidus * Since the contents of the pages may change at any moment, KSM cannot just 6431dbd01fSIzik Eidus * insert the pages into a normal sorted tree and expect it to find anything. 6531dbd01fSIzik Eidus * Therefore KSM uses two data structures - the stable and the unstable tree. 6631dbd01fSIzik Eidus * 6731dbd01fSIzik Eidus * The stable tree holds pointers to all the merged pages (ksm pages), sorted 6831dbd01fSIzik Eidus * by their contents. Because each such page is write-protected, searching on 6931dbd01fSIzik Eidus * this tree is fully assured to be working (except when pages are unmapped), 7031dbd01fSIzik Eidus * and therefore this tree is called the stable tree. 7131dbd01fSIzik Eidus * 725a2ca3efSMike Rapoport * The stable tree node includes information required for reverse 735a2ca3efSMike Rapoport * mapping from a KSM page to virtual addresses that map this page. 745a2ca3efSMike Rapoport * 755a2ca3efSMike Rapoport * In order to avoid large latencies of the rmap walks on KSM pages, 765a2ca3efSMike Rapoport * KSM maintains two types of nodes in the stable tree: 775a2ca3efSMike Rapoport * 785a2ca3efSMike Rapoport * * the regular nodes that keep the reverse mapping structures in a 795a2ca3efSMike Rapoport * linked list 805a2ca3efSMike Rapoport * * the "chains" that link nodes ("dups") that represent the same 815a2ca3efSMike Rapoport * write protected memory content, but each "dup" corresponds to a 825a2ca3efSMike Rapoport * different KSM page copy of that content 835a2ca3efSMike Rapoport * 845a2ca3efSMike Rapoport * Internally, the regular nodes, "dups" and "chains" are represented 855a2ca3efSMike Rapoport * using the same :c:type:`struct stable_node` structure. 865a2ca3efSMike Rapoport * 8731dbd01fSIzik Eidus * In addition to the stable tree, KSM uses a second data structure called the 8831dbd01fSIzik Eidus * unstable tree: this tree holds pointers to pages which have been found to 8931dbd01fSIzik Eidus * be "unchanged for a period of time". The unstable tree sorts these pages 9031dbd01fSIzik Eidus * by their contents, but since they are not write-protected, KSM cannot rely 9131dbd01fSIzik Eidus * upon the unstable tree to work correctly - the unstable tree is liable to 9231dbd01fSIzik Eidus * be corrupted as its contents are modified, and so it is called unstable. 9331dbd01fSIzik Eidus * 9431dbd01fSIzik Eidus * KSM solves this problem by several techniques: 9531dbd01fSIzik Eidus * 9631dbd01fSIzik Eidus * 1) The unstable tree is flushed every time KSM completes scanning all 9731dbd01fSIzik Eidus * memory areas, and then the tree is rebuilt again from the beginning. 9831dbd01fSIzik Eidus * 2) KSM will only insert into the unstable tree, pages whose hash value 9931dbd01fSIzik Eidus * has not changed since the previous scan of all memory areas. 10031dbd01fSIzik Eidus * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the 10131dbd01fSIzik Eidus * colors of the nodes and not on their contents, assuring that even when 10231dbd01fSIzik Eidus * the tree gets "corrupted" it won't get out of balance, so scanning time 10331dbd01fSIzik Eidus * remains the same (also, searching and inserting nodes in an rbtree uses 10431dbd01fSIzik Eidus * the same algorithm, so we have no overhead when we flush and rebuild). 10531dbd01fSIzik Eidus * 4) KSM never flushes the stable tree, which means that even if it were to 10631dbd01fSIzik Eidus * take 10 attempts to find a page in the unstable tree, once it is found, 10731dbd01fSIzik Eidus * it is secured in the stable tree. (When we scan a new page, we first 10831dbd01fSIzik Eidus * compare it against the stable tree, and then against the unstable tree.) 1098fdb3dbfSHugh Dickins * 1108fdb3dbfSHugh Dickins * If the merge_across_nodes tunable is unset, then KSM maintains multiple 1118fdb3dbfSHugh Dickins * stable trees and multiple unstable trees: one of each for each NUMA node. 11231dbd01fSIzik Eidus */ 11331dbd01fSIzik Eidus 11431dbd01fSIzik Eidus /** 11531dbd01fSIzik Eidus * struct mm_slot - ksm information per mm that is being scanned 11631dbd01fSIzik Eidus * @link: link to the mm_slots hash list 11731dbd01fSIzik Eidus * @mm_list: link into the mm_slots list, rooted in ksm_mm_head 1186514d511SHugh Dickins * @rmap_list: head for this mm_slot's singly-linked list of rmap_items 11931dbd01fSIzik Eidus * @mm: the mm that this information is valid for 12031dbd01fSIzik Eidus */ 12131dbd01fSIzik Eidus struct mm_slot { 12231dbd01fSIzik Eidus struct hlist_node link; 12331dbd01fSIzik Eidus struct list_head mm_list; 1246514d511SHugh Dickins struct rmap_item *rmap_list; 12531dbd01fSIzik Eidus struct mm_struct *mm; 12631dbd01fSIzik Eidus }; 12731dbd01fSIzik Eidus 12831dbd01fSIzik Eidus /** 12931dbd01fSIzik Eidus * struct ksm_scan - cursor for scanning 13031dbd01fSIzik Eidus * @mm_slot: the current mm_slot we are scanning 13131dbd01fSIzik Eidus * @address: the next address inside that to be scanned 1326514d511SHugh Dickins * @rmap_list: link to the next rmap to be scanned in the rmap_list 13331dbd01fSIzik Eidus * @seqnr: count of completed full scans (needed when removing unstable node) 13431dbd01fSIzik Eidus * 13531dbd01fSIzik Eidus * There is only the one ksm_scan instance of this cursor structure. 13631dbd01fSIzik Eidus */ 13731dbd01fSIzik Eidus struct ksm_scan { 13831dbd01fSIzik Eidus struct mm_slot *mm_slot; 13931dbd01fSIzik Eidus unsigned long address; 1406514d511SHugh Dickins struct rmap_item **rmap_list; 14131dbd01fSIzik Eidus unsigned long seqnr; 14231dbd01fSIzik Eidus }; 14331dbd01fSIzik Eidus 14431dbd01fSIzik Eidus /** 1457b6ba2c7SHugh Dickins * struct stable_node - node of the stable rbtree 1467b6ba2c7SHugh Dickins * @node: rb node of this ksm page in the stable tree 1474146d2d6SHugh Dickins * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list 1482c653d0eSAndrea Arcangeli * @hlist_dup: linked into the stable_node->hlist with a stable_node chain 1494146d2d6SHugh Dickins * @list: linked into migrate_nodes, pending placement in the proper node tree 1507b6ba2c7SHugh Dickins * @hlist: hlist head of rmap_items using this ksm page 1514146d2d6SHugh Dickins * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid) 1522c653d0eSAndrea Arcangeli * @chain_prune_time: time of the last full garbage collection 1532c653d0eSAndrea Arcangeli * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN 1544146d2d6SHugh Dickins * @nid: NUMA node id of stable tree in which linked (may not match kpfn) 1557b6ba2c7SHugh Dickins */ 1567b6ba2c7SHugh Dickins struct stable_node { 1574146d2d6SHugh Dickins union { 1584146d2d6SHugh Dickins struct rb_node node; /* when node of stable tree */ 1594146d2d6SHugh Dickins struct { /* when listed for migration */ 1604146d2d6SHugh Dickins struct list_head *head; 1612c653d0eSAndrea Arcangeli struct { 1622c653d0eSAndrea Arcangeli struct hlist_node hlist_dup; 1634146d2d6SHugh Dickins struct list_head list; 1644146d2d6SHugh Dickins }; 1654146d2d6SHugh Dickins }; 1662c653d0eSAndrea Arcangeli }; 1677b6ba2c7SHugh Dickins struct hlist_head hlist; 1682c653d0eSAndrea Arcangeli union { 16962b61f61SHugh Dickins unsigned long kpfn; 1702c653d0eSAndrea Arcangeli unsigned long chain_prune_time; 1712c653d0eSAndrea Arcangeli }; 1722c653d0eSAndrea Arcangeli /* 1732c653d0eSAndrea Arcangeli * STABLE_NODE_CHAIN can be any negative number in 1742c653d0eSAndrea Arcangeli * rmap_hlist_len negative range, but better not -1 to be able 1752c653d0eSAndrea Arcangeli * to reliably detect underflows. 1762c653d0eSAndrea Arcangeli */ 1772c653d0eSAndrea Arcangeli #define STABLE_NODE_CHAIN -1024 1782c653d0eSAndrea Arcangeli int rmap_hlist_len; 1794146d2d6SHugh Dickins #ifdef CONFIG_NUMA 1804146d2d6SHugh Dickins int nid; 1814146d2d6SHugh Dickins #endif 1827b6ba2c7SHugh Dickins }; 1837b6ba2c7SHugh Dickins 1847b6ba2c7SHugh Dickins /** 18531dbd01fSIzik Eidus * struct rmap_item - reverse mapping item for virtual addresses 1866514d511SHugh Dickins * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list 187db114b83SHugh Dickins * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree 188bc56620bSHugh Dickins * @nid: NUMA node id of unstable tree in which linked (may not match page) 18931dbd01fSIzik Eidus * @mm: the memory structure this rmap_item is pointing into 19031dbd01fSIzik Eidus * @address: the virtual address this rmap_item tracks (+ flags in low bits) 19131dbd01fSIzik Eidus * @oldchecksum: previous checksum of the page at that virtual address 1927b6ba2c7SHugh Dickins * @node: rb node of this rmap_item in the unstable tree 1937b6ba2c7SHugh Dickins * @head: pointer to stable_node heading this list in the stable tree 1947b6ba2c7SHugh Dickins * @hlist: link into hlist of rmap_items hanging off that stable_node 19531dbd01fSIzik Eidus */ 19631dbd01fSIzik Eidus struct rmap_item { 1976514d511SHugh Dickins struct rmap_item *rmap_list; 198bc56620bSHugh Dickins union { 199db114b83SHugh Dickins struct anon_vma *anon_vma; /* when stable */ 200bc56620bSHugh Dickins #ifdef CONFIG_NUMA 201bc56620bSHugh Dickins int nid; /* when node of unstable tree */ 202bc56620bSHugh Dickins #endif 203bc56620bSHugh Dickins }; 20431dbd01fSIzik Eidus struct mm_struct *mm; 20531dbd01fSIzik Eidus unsigned long address; /* + low bits used for flags below */ 20631dbd01fSIzik Eidus unsigned int oldchecksum; /* when unstable */ 20731dbd01fSIzik Eidus union { 2087b6ba2c7SHugh Dickins struct rb_node node; /* when node of unstable tree */ 2097b6ba2c7SHugh Dickins struct { /* when listed from stable tree */ 2107b6ba2c7SHugh Dickins struct stable_node *head; 2117b6ba2c7SHugh Dickins struct hlist_node hlist; 2127b6ba2c7SHugh Dickins }; 21331dbd01fSIzik Eidus }; 21431dbd01fSIzik Eidus }; 21531dbd01fSIzik Eidus 21631dbd01fSIzik Eidus #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ 2177b6ba2c7SHugh Dickins #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ 2187b6ba2c7SHugh Dickins #define STABLE_FLAG 0x200 /* is listed from the stable tree */ 21931dbd01fSIzik Eidus 22031dbd01fSIzik Eidus /* The stable and unstable tree heads */ 221ef53d16cSHugh Dickins static struct rb_root one_stable_tree[1] = { RB_ROOT }; 222ef53d16cSHugh Dickins static struct rb_root one_unstable_tree[1] = { RB_ROOT }; 223ef53d16cSHugh Dickins static struct rb_root *root_stable_tree = one_stable_tree; 224ef53d16cSHugh Dickins static struct rb_root *root_unstable_tree = one_unstable_tree; 22531dbd01fSIzik Eidus 2264146d2d6SHugh Dickins /* Recently migrated nodes of stable tree, pending proper placement */ 2274146d2d6SHugh Dickins static LIST_HEAD(migrate_nodes); 2282c653d0eSAndrea Arcangeli #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev) 2294146d2d6SHugh Dickins 2304ca3a69bSSasha Levin #define MM_SLOTS_HASH_BITS 10 2314ca3a69bSSasha Levin static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 23231dbd01fSIzik Eidus 23331dbd01fSIzik Eidus static struct mm_slot ksm_mm_head = { 23431dbd01fSIzik Eidus .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list), 23531dbd01fSIzik Eidus }; 23631dbd01fSIzik Eidus static struct ksm_scan ksm_scan = { 23731dbd01fSIzik Eidus .mm_slot = &ksm_mm_head, 23831dbd01fSIzik Eidus }; 23931dbd01fSIzik Eidus 24031dbd01fSIzik Eidus static struct kmem_cache *rmap_item_cache; 2417b6ba2c7SHugh Dickins static struct kmem_cache *stable_node_cache; 24231dbd01fSIzik Eidus static struct kmem_cache *mm_slot_cache; 24331dbd01fSIzik Eidus 24431dbd01fSIzik Eidus /* The number of nodes in the stable tree */ 245b4028260SHugh Dickins static unsigned long ksm_pages_shared; 24631dbd01fSIzik Eidus 247e178dfdeSHugh Dickins /* The number of page slots additionally sharing those nodes */ 248b4028260SHugh Dickins static unsigned long ksm_pages_sharing; 24931dbd01fSIzik Eidus 250473b0ce4SHugh Dickins /* The number of nodes in the unstable tree */ 251473b0ce4SHugh Dickins static unsigned long ksm_pages_unshared; 252473b0ce4SHugh Dickins 253473b0ce4SHugh Dickins /* The number of rmap_items in use: to calculate pages_volatile */ 254473b0ce4SHugh Dickins static unsigned long ksm_rmap_items; 255473b0ce4SHugh Dickins 2562c653d0eSAndrea Arcangeli /* The number of stable_node chains */ 2572c653d0eSAndrea Arcangeli static unsigned long ksm_stable_node_chains; 2582c653d0eSAndrea Arcangeli 2592c653d0eSAndrea Arcangeli /* The number of stable_node dups linked to the stable_node chains */ 2602c653d0eSAndrea Arcangeli static unsigned long ksm_stable_node_dups; 2612c653d0eSAndrea Arcangeli 2622c653d0eSAndrea Arcangeli /* Delay in pruning stale stable_node_dups in the stable_node_chains */ 2632c653d0eSAndrea Arcangeli static int ksm_stable_node_chains_prune_millisecs = 2000; 2642c653d0eSAndrea Arcangeli 2652c653d0eSAndrea Arcangeli /* Maximum number of page slots sharing a stable node */ 2662c653d0eSAndrea Arcangeli static int ksm_max_page_sharing = 256; 2672c653d0eSAndrea Arcangeli 26831dbd01fSIzik Eidus /* Number of pages ksmd should scan in one batch */ 2692c6854fdSIzik Eidus static unsigned int ksm_thread_pages_to_scan = 100; 27031dbd01fSIzik Eidus 27131dbd01fSIzik Eidus /* Milliseconds ksmd should sleep between batches */ 2722ffd8679SHugh Dickins static unsigned int ksm_thread_sleep_millisecs = 20; 27331dbd01fSIzik Eidus 274e86c59b1SClaudio Imbrenda /* Checksum of an empty (zeroed) page */ 275e86c59b1SClaudio Imbrenda static unsigned int zero_checksum __read_mostly; 276e86c59b1SClaudio Imbrenda 277e86c59b1SClaudio Imbrenda /* Whether to merge empty (zeroed) pages with actual zero pages */ 278e86c59b1SClaudio Imbrenda static bool ksm_use_zero_pages __read_mostly; 279e86c59b1SClaudio Imbrenda 280e850dcf5SHugh Dickins #ifdef CONFIG_NUMA 28190bd6fd3SPetr Holasek /* Zeroed when merging across nodes is not allowed */ 28290bd6fd3SPetr Holasek static unsigned int ksm_merge_across_nodes = 1; 283ef53d16cSHugh Dickins static int ksm_nr_node_ids = 1; 284e850dcf5SHugh Dickins #else 285e850dcf5SHugh Dickins #define ksm_merge_across_nodes 1U 286ef53d16cSHugh Dickins #define ksm_nr_node_ids 1 287e850dcf5SHugh Dickins #endif 28890bd6fd3SPetr Holasek 28931dbd01fSIzik Eidus #define KSM_RUN_STOP 0 29031dbd01fSIzik Eidus #define KSM_RUN_MERGE 1 29131dbd01fSIzik Eidus #define KSM_RUN_UNMERGE 2 292ef4d43a8SHugh Dickins #define KSM_RUN_OFFLINE 4 293ef4d43a8SHugh Dickins static unsigned long ksm_run = KSM_RUN_STOP; 294ef4d43a8SHugh Dickins static void wait_while_offlining(void); 29531dbd01fSIzik Eidus 29631dbd01fSIzik Eidus static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); 29731dbd01fSIzik Eidus static DEFINE_MUTEX(ksm_thread_mutex); 29831dbd01fSIzik Eidus static DEFINE_SPINLOCK(ksm_mmlist_lock); 29931dbd01fSIzik Eidus 30031dbd01fSIzik Eidus #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\ 30131dbd01fSIzik Eidus sizeof(struct __struct), __alignof__(struct __struct),\ 30231dbd01fSIzik Eidus (__flags), NULL) 30331dbd01fSIzik Eidus 30431dbd01fSIzik Eidus static int __init ksm_slab_init(void) 30531dbd01fSIzik Eidus { 30631dbd01fSIzik Eidus rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0); 30731dbd01fSIzik Eidus if (!rmap_item_cache) 30831dbd01fSIzik Eidus goto out; 30931dbd01fSIzik Eidus 3107b6ba2c7SHugh Dickins stable_node_cache = KSM_KMEM_CACHE(stable_node, 0); 3117b6ba2c7SHugh Dickins if (!stable_node_cache) 3127b6ba2c7SHugh Dickins goto out_free1; 3137b6ba2c7SHugh Dickins 31431dbd01fSIzik Eidus mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0); 31531dbd01fSIzik Eidus if (!mm_slot_cache) 3167b6ba2c7SHugh Dickins goto out_free2; 31731dbd01fSIzik Eidus 31831dbd01fSIzik Eidus return 0; 31931dbd01fSIzik Eidus 3207b6ba2c7SHugh Dickins out_free2: 3217b6ba2c7SHugh Dickins kmem_cache_destroy(stable_node_cache); 3227b6ba2c7SHugh Dickins out_free1: 32331dbd01fSIzik Eidus kmem_cache_destroy(rmap_item_cache); 32431dbd01fSIzik Eidus out: 32531dbd01fSIzik Eidus return -ENOMEM; 32631dbd01fSIzik Eidus } 32731dbd01fSIzik Eidus 32831dbd01fSIzik Eidus static void __init ksm_slab_free(void) 32931dbd01fSIzik Eidus { 33031dbd01fSIzik Eidus kmem_cache_destroy(mm_slot_cache); 3317b6ba2c7SHugh Dickins kmem_cache_destroy(stable_node_cache); 33231dbd01fSIzik Eidus kmem_cache_destroy(rmap_item_cache); 33331dbd01fSIzik Eidus mm_slot_cache = NULL; 33431dbd01fSIzik Eidus } 33531dbd01fSIzik Eidus 3362c653d0eSAndrea Arcangeli static __always_inline bool is_stable_node_chain(struct stable_node *chain) 3372c653d0eSAndrea Arcangeli { 3382c653d0eSAndrea Arcangeli return chain->rmap_hlist_len == STABLE_NODE_CHAIN; 3392c653d0eSAndrea Arcangeli } 3402c653d0eSAndrea Arcangeli 3412c653d0eSAndrea Arcangeli static __always_inline bool is_stable_node_dup(struct stable_node *dup) 3422c653d0eSAndrea Arcangeli { 3432c653d0eSAndrea Arcangeli return dup->head == STABLE_NODE_DUP_HEAD; 3442c653d0eSAndrea Arcangeli } 3452c653d0eSAndrea Arcangeli 3462c653d0eSAndrea Arcangeli static inline void stable_node_chain_add_dup(struct stable_node *dup, 3472c653d0eSAndrea Arcangeli struct stable_node *chain) 3482c653d0eSAndrea Arcangeli { 3492c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(dup)); 3502c653d0eSAndrea Arcangeli dup->head = STABLE_NODE_DUP_HEAD; 3512c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_chain(chain)); 3522c653d0eSAndrea Arcangeli hlist_add_head(&dup->hlist_dup, &chain->hlist); 3532c653d0eSAndrea Arcangeli ksm_stable_node_dups++; 3542c653d0eSAndrea Arcangeli } 3552c653d0eSAndrea Arcangeli 3562c653d0eSAndrea Arcangeli static inline void __stable_node_dup_del(struct stable_node *dup) 3572c653d0eSAndrea Arcangeli { 358b4fecc67SAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(dup)); 3592c653d0eSAndrea Arcangeli hlist_del(&dup->hlist_dup); 3602c653d0eSAndrea Arcangeli ksm_stable_node_dups--; 3612c653d0eSAndrea Arcangeli } 3622c653d0eSAndrea Arcangeli 3632c653d0eSAndrea Arcangeli static inline void stable_node_dup_del(struct stable_node *dup) 3642c653d0eSAndrea Arcangeli { 3652c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_chain(dup)); 3662c653d0eSAndrea Arcangeli if (is_stable_node_dup(dup)) 3672c653d0eSAndrea Arcangeli __stable_node_dup_del(dup); 3682c653d0eSAndrea Arcangeli else 3692c653d0eSAndrea Arcangeli rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid)); 3702c653d0eSAndrea Arcangeli #ifdef CONFIG_DEBUG_VM 3712c653d0eSAndrea Arcangeli dup->head = NULL; 3722c653d0eSAndrea Arcangeli #endif 3732c653d0eSAndrea Arcangeli } 3742c653d0eSAndrea Arcangeli 37531dbd01fSIzik Eidus static inline struct rmap_item *alloc_rmap_item(void) 37631dbd01fSIzik Eidus { 377473b0ce4SHugh Dickins struct rmap_item *rmap_item; 378473b0ce4SHugh Dickins 3795b398e41Szhong jiang rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL | 3805b398e41Szhong jiang __GFP_NORETRY | __GFP_NOWARN); 381473b0ce4SHugh Dickins if (rmap_item) 382473b0ce4SHugh Dickins ksm_rmap_items++; 383473b0ce4SHugh Dickins return rmap_item; 38431dbd01fSIzik Eidus } 38531dbd01fSIzik Eidus 38631dbd01fSIzik Eidus static inline void free_rmap_item(struct rmap_item *rmap_item) 38731dbd01fSIzik Eidus { 388473b0ce4SHugh Dickins ksm_rmap_items--; 38931dbd01fSIzik Eidus rmap_item->mm = NULL; /* debug safety */ 39031dbd01fSIzik Eidus kmem_cache_free(rmap_item_cache, rmap_item); 39131dbd01fSIzik Eidus } 39231dbd01fSIzik Eidus 3937b6ba2c7SHugh Dickins static inline struct stable_node *alloc_stable_node(void) 3947b6ba2c7SHugh Dickins { 3956213055fSzhong jiang /* 3966213055fSzhong jiang * The allocation can take too long with GFP_KERNEL when memory is under 3976213055fSzhong jiang * pressure, which may lead to hung task warnings. Adding __GFP_HIGH 3986213055fSzhong jiang * grants access to memory reserves, helping to avoid this problem. 3996213055fSzhong jiang */ 4006213055fSzhong jiang return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH); 4017b6ba2c7SHugh Dickins } 4027b6ba2c7SHugh Dickins 4037b6ba2c7SHugh Dickins static inline void free_stable_node(struct stable_node *stable_node) 4047b6ba2c7SHugh Dickins { 4052c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len && 4062c653d0eSAndrea Arcangeli !is_stable_node_chain(stable_node)); 4077b6ba2c7SHugh Dickins kmem_cache_free(stable_node_cache, stable_node); 4087b6ba2c7SHugh Dickins } 4097b6ba2c7SHugh Dickins 41031dbd01fSIzik Eidus static inline struct mm_slot *alloc_mm_slot(void) 41131dbd01fSIzik Eidus { 41231dbd01fSIzik Eidus if (!mm_slot_cache) /* initialization failed */ 41331dbd01fSIzik Eidus return NULL; 41431dbd01fSIzik Eidus return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); 41531dbd01fSIzik Eidus } 41631dbd01fSIzik Eidus 41731dbd01fSIzik Eidus static inline void free_mm_slot(struct mm_slot *mm_slot) 41831dbd01fSIzik Eidus { 41931dbd01fSIzik Eidus kmem_cache_free(mm_slot_cache, mm_slot); 42031dbd01fSIzik Eidus } 42131dbd01fSIzik Eidus 42231dbd01fSIzik Eidus static struct mm_slot *get_mm_slot(struct mm_struct *mm) 42331dbd01fSIzik Eidus { 4244ca3a69bSSasha Levin struct mm_slot *slot; 42531dbd01fSIzik Eidus 426b67bfe0dSSasha Levin hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm) 4274ca3a69bSSasha Levin if (slot->mm == mm) 4284ca3a69bSSasha Levin return slot; 4294ca3a69bSSasha Levin 43031dbd01fSIzik Eidus return NULL; 43131dbd01fSIzik Eidus } 43231dbd01fSIzik Eidus 43331dbd01fSIzik Eidus static void insert_to_mm_slots_hash(struct mm_struct *mm, 43431dbd01fSIzik Eidus struct mm_slot *mm_slot) 43531dbd01fSIzik Eidus { 43631dbd01fSIzik Eidus mm_slot->mm = mm; 4374ca3a69bSSasha Levin hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm); 43831dbd01fSIzik Eidus } 43931dbd01fSIzik Eidus 44031dbd01fSIzik Eidus /* 441a913e182SHugh Dickins * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's 442a913e182SHugh Dickins * page tables after it has passed through ksm_exit() - which, if necessary, 443a913e182SHugh Dickins * takes mmap_sem briefly to serialize against them. ksm_exit() does not set 444a913e182SHugh Dickins * a special flag: they can just back out as soon as mm_users goes to zero. 445a913e182SHugh Dickins * ksm_test_exit() is used throughout to make this test for exit: in some 446a913e182SHugh Dickins * places for correctness, in some places just to avoid unnecessary work. 447a913e182SHugh Dickins */ 448a913e182SHugh Dickins static inline bool ksm_test_exit(struct mm_struct *mm) 449a913e182SHugh Dickins { 450a913e182SHugh Dickins return atomic_read(&mm->mm_users) == 0; 451a913e182SHugh Dickins } 452a913e182SHugh Dickins 453a913e182SHugh Dickins /* 45431dbd01fSIzik Eidus * We use break_ksm to break COW on a ksm page: it's a stripped down 45531dbd01fSIzik Eidus * 456d4edcf0dSDave Hansen * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1) 45731dbd01fSIzik Eidus * put_page(page); 45831dbd01fSIzik Eidus * 45931dbd01fSIzik Eidus * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma, 46031dbd01fSIzik Eidus * in case the application has unmapped and remapped mm,addr meanwhile. 46131dbd01fSIzik Eidus * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP 46231dbd01fSIzik Eidus * mmap of /dev/mem or /dev/kmem, where we would not want to touch it. 4631b2ee126SDave Hansen * 4641b2ee126SDave Hansen * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context 4651b2ee126SDave Hansen * of the process that owns 'vma'. We also do not want to enforce 4661b2ee126SDave Hansen * protection keys here anyway. 46731dbd01fSIzik Eidus */ 468d952b791SHugh Dickins static int break_ksm(struct vm_area_struct *vma, unsigned long addr) 46931dbd01fSIzik Eidus { 47031dbd01fSIzik Eidus struct page *page; 471d952b791SHugh Dickins int ret = 0; 47231dbd01fSIzik Eidus 47331dbd01fSIzik Eidus do { 47431dbd01fSIzik Eidus cond_resched(); 4751b2ee126SDave Hansen page = follow_page(vma, addr, 4761b2ee126SDave Hansen FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE); 47722eccdd7SDan Carpenter if (IS_ERR_OR_NULL(page)) 47831dbd01fSIzik Eidus break; 47931dbd01fSIzik Eidus if (PageKsm(page)) 480dcddffd4SKirill A. Shutemov ret = handle_mm_fault(vma, addr, 481dcddffd4SKirill A. Shutemov FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE); 48231dbd01fSIzik Eidus else 48331dbd01fSIzik Eidus ret = VM_FAULT_WRITE; 48431dbd01fSIzik Eidus put_page(page); 48533692f27SLinus Torvalds } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM))); 486d952b791SHugh Dickins /* 487d952b791SHugh Dickins * We must loop because handle_mm_fault() may back out if there's 488d952b791SHugh Dickins * any difficulty e.g. if pte accessed bit gets updated concurrently. 489d952b791SHugh Dickins * 490d952b791SHugh Dickins * VM_FAULT_WRITE is what we have been hoping for: it indicates that 491d952b791SHugh Dickins * COW has been broken, even if the vma does not permit VM_WRITE; 492d952b791SHugh Dickins * but note that a concurrent fault might break PageKsm for us. 493d952b791SHugh Dickins * 494d952b791SHugh Dickins * VM_FAULT_SIGBUS could occur if we race with truncation of the 495d952b791SHugh Dickins * backing file, which also invalidates anonymous pages: that's 496d952b791SHugh Dickins * okay, that truncation will have unmapped the PageKsm for us. 497d952b791SHugh Dickins * 498d952b791SHugh Dickins * VM_FAULT_OOM: at the time of writing (late July 2009), setting 499d952b791SHugh Dickins * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the 500d952b791SHugh Dickins * current task has TIF_MEMDIE set, and will be OOM killed on return 501d952b791SHugh Dickins * to user; and ksmd, having no mm, would never be chosen for that. 502d952b791SHugh Dickins * 503d952b791SHugh Dickins * But if the mm is in a limited mem_cgroup, then the fault may fail 504d952b791SHugh Dickins * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and 505d952b791SHugh Dickins * even ksmd can fail in this way - though it's usually breaking ksm 506d952b791SHugh Dickins * just to undo a merge it made a moment before, so unlikely to oom. 507d952b791SHugh Dickins * 508d952b791SHugh Dickins * That's a pity: we might therefore have more kernel pages allocated 509d952b791SHugh Dickins * than we're counting as nodes in the stable tree; but ksm_do_scan 510d952b791SHugh Dickins * will retry to break_cow on each pass, so should recover the page 511d952b791SHugh Dickins * in due course. The important thing is to not let VM_MERGEABLE 512d952b791SHugh Dickins * be cleared while any such pages might remain in the area. 513d952b791SHugh Dickins */ 514d952b791SHugh Dickins return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; 51531dbd01fSIzik Eidus } 51631dbd01fSIzik Eidus 517ef694222SBob Liu static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm, 518ef694222SBob Liu unsigned long addr) 519ef694222SBob Liu { 520ef694222SBob Liu struct vm_area_struct *vma; 521ef694222SBob Liu if (ksm_test_exit(mm)) 522ef694222SBob Liu return NULL; 523ef694222SBob Liu vma = find_vma(mm, addr); 524ef694222SBob Liu if (!vma || vma->vm_start > addr) 525ef694222SBob Liu return NULL; 526ef694222SBob Liu if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) 527ef694222SBob Liu return NULL; 528ef694222SBob Liu return vma; 529ef694222SBob Liu } 530ef694222SBob Liu 5318dd3557aSHugh Dickins static void break_cow(struct rmap_item *rmap_item) 53231dbd01fSIzik Eidus { 5338dd3557aSHugh Dickins struct mm_struct *mm = rmap_item->mm; 5348dd3557aSHugh Dickins unsigned long addr = rmap_item->address; 53531dbd01fSIzik Eidus struct vm_area_struct *vma; 53631dbd01fSIzik Eidus 5374035c07aSHugh Dickins /* 5384035c07aSHugh Dickins * It is not an accident that whenever we want to break COW 5394035c07aSHugh Dickins * to undo, we also need to drop a reference to the anon_vma. 5404035c07aSHugh Dickins */ 5419e60109fSPeter Zijlstra put_anon_vma(rmap_item->anon_vma); 5424035c07aSHugh Dickins 54381464e30SHugh Dickins down_read(&mm->mmap_sem); 544ef694222SBob Liu vma = find_mergeable_vma(mm, addr); 545ef694222SBob Liu if (vma) 54631dbd01fSIzik Eidus break_ksm(vma, addr); 54731dbd01fSIzik Eidus up_read(&mm->mmap_sem); 54831dbd01fSIzik Eidus } 54931dbd01fSIzik Eidus 55031dbd01fSIzik Eidus static struct page *get_mergeable_page(struct rmap_item *rmap_item) 55131dbd01fSIzik Eidus { 55231dbd01fSIzik Eidus struct mm_struct *mm = rmap_item->mm; 55331dbd01fSIzik Eidus unsigned long addr = rmap_item->address; 55431dbd01fSIzik Eidus struct vm_area_struct *vma; 55531dbd01fSIzik Eidus struct page *page; 55631dbd01fSIzik Eidus 55731dbd01fSIzik Eidus down_read(&mm->mmap_sem); 558ef694222SBob Liu vma = find_mergeable_vma(mm, addr); 559ef694222SBob Liu if (!vma) 56031dbd01fSIzik Eidus goto out; 56131dbd01fSIzik Eidus 56231dbd01fSIzik Eidus page = follow_page(vma, addr, FOLL_GET); 56322eccdd7SDan Carpenter if (IS_ERR_OR_NULL(page)) 56431dbd01fSIzik Eidus goto out; 565f765f540SKirill A. Shutemov if (PageAnon(page)) { 56631dbd01fSIzik Eidus flush_anon_page(vma, page, addr); 56731dbd01fSIzik Eidus flush_dcache_page(page); 56831dbd01fSIzik Eidus } else { 56931dbd01fSIzik Eidus put_page(page); 570c8f95ed1SAndrea Arcangeli out: 571c8f95ed1SAndrea Arcangeli page = NULL; 57231dbd01fSIzik Eidus } 57331dbd01fSIzik Eidus up_read(&mm->mmap_sem); 57431dbd01fSIzik Eidus return page; 57531dbd01fSIzik Eidus } 57631dbd01fSIzik Eidus 57790bd6fd3SPetr Holasek /* 57890bd6fd3SPetr Holasek * This helper is used for getting right index into array of tree roots. 57990bd6fd3SPetr Holasek * When merge_across_nodes knob is set to 1, there are only two rb-trees for 58090bd6fd3SPetr Holasek * stable and unstable pages from all nodes with roots in index 0. Otherwise, 58190bd6fd3SPetr Holasek * every node has its own stable and unstable tree. 58290bd6fd3SPetr Holasek */ 58390bd6fd3SPetr Holasek static inline int get_kpfn_nid(unsigned long kpfn) 58490bd6fd3SPetr Holasek { 585d8fc16a8SHugh Dickins return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn)); 58690bd6fd3SPetr Holasek } 58790bd6fd3SPetr Holasek 5882c653d0eSAndrea Arcangeli static struct stable_node *alloc_stable_node_chain(struct stable_node *dup, 5892c653d0eSAndrea Arcangeli struct rb_root *root) 5902c653d0eSAndrea Arcangeli { 5912c653d0eSAndrea Arcangeli struct stable_node *chain = alloc_stable_node(); 5922c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_chain(dup)); 5932c653d0eSAndrea Arcangeli if (likely(chain)) { 5942c653d0eSAndrea Arcangeli INIT_HLIST_HEAD(&chain->hlist); 5952c653d0eSAndrea Arcangeli chain->chain_prune_time = jiffies; 5962c653d0eSAndrea Arcangeli chain->rmap_hlist_len = STABLE_NODE_CHAIN; 5972c653d0eSAndrea Arcangeli #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA) 5982c653d0eSAndrea Arcangeli chain->nid = -1; /* debug */ 5992c653d0eSAndrea Arcangeli #endif 6002c653d0eSAndrea Arcangeli ksm_stable_node_chains++; 6012c653d0eSAndrea Arcangeli 6022c653d0eSAndrea Arcangeli /* 6032c653d0eSAndrea Arcangeli * Put the stable node chain in the first dimension of 6042c653d0eSAndrea Arcangeli * the stable tree and at the same time remove the old 6052c653d0eSAndrea Arcangeli * stable node. 6062c653d0eSAndrea Arcangeli */ 6072c653d0eSAndrea Arcangeli rb_replace_node(&dup->node, &chain->node, root); 6082c653d0eSAndrea Arcangeli 6092c653d0eSAndrea Arcangeli /* 6102c653d0eSAndrea Arcangeli * Move the old stable node to the second dimension 6112c653d0eSAndrea Arcangeli * queued in the hlist_dup. The invariant is that all 6122c653d0eSAndrea Arcangeli * dup stable_nodes in the chain->hlist point to pages 6132c653d0eSAndrea Arcangeli * that are wrprotected and have the exact same 6142c653d0eSAndrea Arcangeli * content. 6152c653d0eSAndrea Arcangeli */ 6162c653d0eSAndrea Arcangeli stable_node_chain_add_dup(dup, chain); 6172c653d0eSAndrea Arcangeli } 6182c653d0eSAndrea Arcangeli return chain; 6192c653d0eSAndrea Arcangeli } 6202c653d0eSAndrea Arcangeli 6212c653d0eSAndrea Arcangeli static inline void free_stable_node_chain(struct stable_node *chain, 6222c653d0eSAndrea Arcangeli struct rb_root *root) 6232c653d0eSAndrea Arcangeli { 6242c653d0eSAndrea Arcangeli rb_erase(&chain->node, root); 6252c653d0eSAndrea Arcangeli free_stable_node(chain); 6262c653d0eSAndrea Arcangeli ksm_stable_node_chains--; 6272c653d0eSAndrea Arcangeli } 6282c653d0eSAndrea Arcangeli 6294035c07aSHugh Dickins static void remove_node_from_stable_tree(struct stable_node *stable_node) 6304035c07aSHugh Dickins { 6314035c07aSHugh Dickins struct rmap_item *rmap_item; 6324035c07aSHugh Dickins 6332c653d0eSAndrea Arcangeli /* check it's not STABLE_NODE_CHAIN or negative */ 6342c653d0eSAndrea Arcangeli BUG_ON(stable_node->rmap_hlist_len < 0); 6352c653d0eSAndrea Arcangeli 636b67bfe0dSSasha Levin hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { 6374035c07aSHugh Dickins if (rmap_item->hlist.next) 6384035c07aSHugh Dickins ksm_pages_sharing--; 6394035c07aSHugh Dickins else 6404035c07aSHugh Dickins ksm_pages_shared--; 6412c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len <= 0); 6422c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len--; 6439e60109fSPeter Zijlstra put_anon_vma(rmap_item->anon_vma); 6444035c07aSHugh Dickins rmap_item->address &= PAGE_MASK; 6454035c07aSHugh Dickins cond_resched(); 6464035c07aSHugh Dickins } 6474035c07aSHugh Dickins 6482c653d0eSAndrea Arcangeli /* 6492c653d0eSAndrea Arcangeli * We need the second aligned pointer of the migrate_nodes 6502c653d0eSAndrea Arcangeli * list_head to stay clear from the rb_parent_color union 6512c653d0eSAndrea Arcangeli * (aligned and different than any node) and also different 6522c653d0eSAndrea Arcangeli * from &migrate_nodes. This will verify that future list.h changes 6532c653d0eSAndrea Arcangeli * don't break STABLE_NODE_DUP_HEAD. 6542c653d0eSAndrea Arcangeli */ 6552c653d0eSAndrea Arcangeli #if GCC_VERSION >= 40903 /* only recent gcc can handle it */ 6562c653d0eSAndrea Arcangeli BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes); 6572c653d0eSAndrea Arcangeli BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1); 6582c653d0eSAndrea Arcangeli #endif 6592c653d0eSAndrea Arcangeli 6604146d2d6SHugh Dickins if (stable_node->head == &migrate_nodes) 6614146d2d6SHugh Dickins list_del(&stable_node->list); 6624146d2d6SHugh Dickins else 6632c653d0eSAndrea Arcangeli stable_node_dup_del(stable_node); 6644035c07aSHugh Dickins free_stable_node(stable_node); 6654035c07aSHugh Dickins } 6664035c07aSHugh Dickins 6674035c07aSHugh Dickins /* 6684035c07aSHugh Dickins * get_ksm_page: checks if the page indicated by the stable node 6694035c07aSHugh Dickins * is still its ksm page, despite having held no reference to it. 6704035c07aSHugh Dickins * In which case we can trust the content of the page, and it 6714035c07aSHugh Dickins * returns the gotten page; but if the page has now been zapped, 6724035c07aSHugh Dickins * remove the stale node from the stable tree and return NULL. 673c8d6553bSHugh Dickins * But beware, the stable node's page might be being migrated. 6744035c07aSHugh Dickins * 6754035c07aSHugh Dickins * You would expect the stable_node to hold a reference to the ksm page. 6764035c07aSHugh Dickins * But if it increments the page's count, swapping out has to wait for 6774035c07aSHugh Dickins * ksmd to come around again before it can free the page, which may take 6784035c07aSHugh Dickins * seconds or even minutes: much too unresponsive. So instead we use a 6794035c07aSHugh Dickins * "keyhole reference": access to the ksm page from the stable node peeps 6804035c07aSHugh Dickins * out through its keyhole to see if that page still holds the right key, 6814035c07aSHugh Dickins * pointing back to this stable node. This relies on freeing a PageAnon 6824035c07aSHugh Dickins * page to reset its page->mapping to NULL, and relies on no other use of 6834035c07aSHugh Dickins * a page to put something that might look like our key in page->mapping. 6844035c07aSHugh Dickins * is on its way to being freed; but it is an anomaly to bear in mind. 6854035c07aSHugh Dickins */ 6868fdb3dbfSHugh Dickins static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it) 6874035c07aSHugh Dickins { 6884035c07aSHugh Dickins struct page *page; 6894035c07aSHugh Dickins void *expected_mapping; 690c8d6553bSHugh Dickins unsigned long kpfn; 6914035c07aSHugh Dickins 692bda807d4SMinchan Kim expected_mapping = (void *)((unsigned long)stable_node | 693bda807d4SMinchan Kim PAGE_MAPPING_KSM); 694c8d6553bSHugh Dickins again: 69508df4774SPaul E. McKenney kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */ 696c8d6553bSHugh Dickins page = pfn_to_page(kpfn); 6974db0c3c2SJason Low if (READ_ONCE(page->mapping) != expected_mapping) 6984035c07aSHugh Dickins goto stale; 699c8d6553bSHugh Dickins 700c8d6553bSHugh Dickins /* 701c8d6553bSHugh Dickins * We cannot do anything with the page while its refcount is 0. 702c8d6553bSHugh Dickins * Usually 0 means free, or tail of a higher-order page: in which 703c8d6553bSHugh Dickins * case this node is no longer referenced, and should be freed; 704c8d6553bSHugh Dickins * however, it might mean that the page is under page_freeze_refs(). 705c8d6553bSHugh Dickins * The __remove_mapping() case is easy, again the node is now stale; 706c8d6553bSHugh Dickins * but if page is swapcache in migrate_page_move_mapping(), it might 707c8d6553bSHugh Dickins * still be our page, in which case it's essential to keep the node. 708c8d6553bSHugh Dickins */ 709c8d6553bSHugh Dickins while (!get_page_unless_zero(page)) { 710c8d6553bSHugh Dickins /* 711c8d6553bSHugh Dickins * Another check for page->mapping != expected_mapping would 712c8d6553bSHugh Dickins * work here too. We have chosen the !PageSwapCache test to 713c8d6553bSHugh Dickins * optimize the common case, when the page is or is about to 714c8d6553bSHugh Dickins * be freed: PageSwapCache is cleared (under spin_lock_irq) 715c8d6553bSHugh Dickins * in the freeze_refs section of __remove_mapping(); but Anon 716c8d6553bSHugh Dickins * page->mapping reset to NULL later, in free_pages_prepare(). 717c8d6553bSHugh Dickins */ 718c8d6553bSHugh Dickins if (!PageSwapCache(page)) 7194035c07aSHugh Dickins goto stale; 720c8d6553bSHugh Dickins cpu_relax(); 721c8d6553bSHugh Dickins } 722c8d6553bSHugh Dickins 7234db0c3c2SJason Low if (READ_ONCE(page->mapping) != expected_mapping) { 7244035c07aSHugh Dickins put_page(page); 7254035c07aSHugh Dickins goto stale; 7264035c07aSHugh Dickins } 727c8d6553bSHugh Dickins 7288fdb3dbfSHugh Dickins if (lock_it) { 7298aafa6a4SHugh Dickins lock_page(page); 7304db0c3c2SJason Low if (READ_ONCE(page->mapping) != expected_mapping) { 7318aafa6a4SHugh Dickins unlock_page(page); 7328aafa6a4SHugh Dickins put_page(page); 7338aafa6a4SHugh Dickins goto stale; 7348aafa6a4SHugh Dickins } 7358aafa6a4SHugh Dickins } 7364035c07aSHugh Dickins return page; 737c8d6553bSHugh Dickins 7384035c07aSHugh Dickins stale: 739c8d6553bSHugh Dickins /* 740c8d6553bSHugh Dickins * We come here from above when page->mapping or !PageSwapCache 741c8d6553bSHugh Dickins * suggests that the node is stale; but it might be under migration. 742c8d6553bSHugh Dickins * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(), 743c8d6553bSHugh Dickins * before checking whether node->kpfn has been changed. 744c8d6553bSHugh Dickins */ 745c8d6553bSHugh Dickins smp_rmb(); 7464db0c3c2SJason Low if (READ_ONCE(stable_node->kpfn) != kpfn) 747c8d6553bSHugh Dickins goto again; 7484035c07aSHugh Dickins remove_node_from_stable_tree(stable_node); 7494035c07aSHugh Dickins return NULL; 7504035c07aSHugh Dickins } 7514035c07aSHugh Dickins 75231dbd01fSIzik Eidus /* 75331dbd01fSIzik Eidus * Removing rmap_item from stable or unstable tree. 75431dbd01fSIzik Eidus * This function will clean the information from the stable/unstable tree. 75531dbd01fSIzik Eidus */ 75631dbd01fSIzik Eidus static void remove_rmap_item_from_tree(struct rmap_item *rmap_item) 75731dbd01fSIzik Eidus { 7587b6ba2c7SHugh Dickins if (rmap_item->address & STABLE_FLAG) { 7597b6ba2c7SHugh Dickins struct stable_node *stable_node; 7605ad64688SHugh Dickins struct page *page; 76131dbd01fSIzik Eidus 7627b6ba2c7SHugh Dickins stable_node = rmap_item->head; 7638aafa6a4SHugh Dickins page = get_ksm_page(stable_node, true); 7644035c07aSHugh Dickins if (!page) 7654035c07aSHugh Dickins goto out; 7665ad64688SHugh Dickins 7677b6ba2c7SHugh Dickins hlist_del(&rmap_item->hlist); 7685ad64688SHugh Dickins unlock_page(page); 7695ad64688SHugh Dickins put_page(page); 77008beca44SHugh Dickins 77198666f8aSAndrea Arcangeli if (!hlist_empty(&stable_node->hlist)) 7724035c07aSHugh Dickins ksm_pages_sharing--; 7734035c07aSHugh Dickins else 774b4028260SHugh Dickins ksm_pages_shared--; 7752c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len <= 0); 7762c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len--; 77731dbd01fSIzik Eidus 7789e60109fSPeter Zijlstra put_anon_vma(rmap_item->anon_vma); 77993d17715SHugh Dickins rmap_item->address &= PAGE_MASK; 78031dbd01fSIzik Eidus 7817b6ba2c7SHugh Dickins } else if (rmap_item->address & UNSTABLE_FLAG) { 78231dbd01fSIzik Eidus unsigned char age; 78331dbd01fSIzik Eidus /* 7849ba69294SHugh Dickins * Usually ksmd can and must skip the rb_erase, because 78531dbd01fSIzik Eidus * root_unstable_tree was already reset to RB_ROOT. 7869ba69294SHugh Dickins * But be careful when an mm is exiting: do the rb_erase 7879ba69294SHugh Dickins * if this rmap_item was inserted by this scan, rather 7889ba69294SHugh Dickins * than left over from before. 78931dbd01fSIzik Eidus */ 79031dbd01fSIzik Eidus age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); 791cd551f97SHugh Dickins BUG_ON(age > 1); 79231dbd01fSIzik Eidus if (!age) 79390bd6fd3SPetr Holasek rb_erase(&rmap_item->node, 794ef53d16cSHugh Dickins root_unstable_tree + NUMA(rmap_item->nid)); 79593d17715SHugh Dickins ksm_pages_unshared--; 79631dbd01fSIzik Eidus rmap_item->address &= PAGE_MASK; 79793d17715SHugh Dickins } 7984035c07aSHugh Dickins out: 79931dbd01fSIzik Eidus cond_resched(); /* we're called from many long loops */ 80031dbd01fSIzik Eidus } 80131dbd01fSIzik Eidus 80231dbd01fSIzik Eidus static void remove_trailing_rmap_items(struct mm_slot *mm_slot, 8036514d511SHugh Dickins struct rmap_item **rmap_list) 80431dbd01fSIzik Eidus { 8056514d511SHugh Dickins while (*rmap_list) { 8066514d511SHugh Dickins struct rmap_item *rmap_item = *rmap_list; 8076514d511SHugh Dickins *rmap_list = rmap_item->rmap_list; 80831dbd01fSIzik Eidus remove_rmap_item_from_tree(rmap_item); 80931dbd01fSIzik Eidus free_rmap_item(rmap_item); 81031dbd01fSIzik Eidus } 81131dbd01fSIzik Eidus } 81231dbd01fSIzik Eidus 81331dbd01fSIzik Eidus /* 814e850dcf5SHugh Dickins * Though it's very tempting to unmerge rmap_items from stable tree rather 81531dbd01fSIzik Eidus * than check every pte of a given vma, the locking doesn't quite work for 81631dbd01fSIzik Eidus * that - an rmap_item is assigned to the stable tree after inserting ksm 81731dbd01fSIzik Eidus * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing 81831dbd01fSIzik Eidus * rmap_items from parent to child at fork time (so as not to waste time 81931dbd01fSIzik Eidus * if exit comes before the next scan reaches it). 82081464e30SHugh Dickins * 82181464e30SHugh Dickins * Similarly, although we'd like to remove rmap_items (so updating counts 82281464e30SHugh Dickins * and freeing memory) when unmerging an area, it's easier to leave that 82381464e30SHugh Dickins * to the next pass of ksmd - consider, for example, how ksmd might be 82481464e30SHugh Dickins * in cmp_and_merge_page on one of the rmap_items we would be removing. 82531dbd01fSIzik Eidus */ 826d952b791SHugh Dickins static int unmerge_ksm_pages(struct vm_area_struct *vma, 82731dbd01fSIzik Eidus unsigned long start, unsigned long end) 82831dbd01fSIzik Eidus { 82931dbd01fSIzik Eidus unsigned long addr; 830d952b791SHugh Dickins int err = 0; 83131dbd01fSIzik Eidus 832d952b791SHugh Dickins for (addr = start; addr < end && !err; addr += PAGE_SIZE) { 8339ba69294SHugh Dickins if (ksm_test_exit(vma->vm_mm)) 8349ba69294SHugh Dickins break; 835d952b791SHugh Dickins if (signal_pending(current)) 836d952b791SHugh Dickins err = -ERESTARTSYS; 837d952b791SHugh Dickins else 838d952b791SHugh Dickins err = break_ksm(vma, addr); 839d952b791SHugh Dickins } 840d952b791SHugh Dickins return err; 84131dbd01fSIzik Eidus } 84231dbd01fSIzik Eidus 843*88484826SMike Rapoport static inline struct stable_node *page_stable_node(struct page *page) 844*88484826SMike Rapoport { 845*88484826SMike Rapoport return PageKsm(page) ? page_rmapping(page) : NULL; 846*88484826SMike Rapoport } 847*88484826SMike Rapoport 848*88484826SMike Rapoport static inline void set_page_stable_node(struct page *page, 849*88484826SMike Rapoport struct stable_node *stable_node) 850*88484826SMike Rapoport { 851*88484826SMike Rapoport page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM); 852*88484826SMike Rapoport } 853*88484826SMike Rapoport 8542ffd8679SHugh Dickins #ifdef CONFIG_SYSFS 8552ffd8679SHugh Dickins /* 8562ffd8679SHugh Dickins * Only called through the sysfs control interface: 8572ffd8679SHugh Dickins */ 858cbf86cfeSHugh Dickins static int remove_stable_node(struct stable_node *stable_node) 859cbf86cfeSHugh Dickins { 860cbf86cfeSHugh Dickins struct page *page; 861cbf86cfeSHugh Dickins int err; 862cbf86cfeSHugh Dickins 863cbf86cfeSHugh Dickins page = get_ksm_page(stable_node, true); 864cbf86cfeSHugh Dickins if (!page) { 865cbf86cfeSHugh Dickins /* 866cbf86cfeSHugh Dickins * get_ksm_page did remove_node_from_stable_tree itself. 867cbf86cfeSHugh Dickins */ 868cbf86cfeSHugh Dickins return 0; 869cbf86cfeSHugh Dickins } 870cbf86cfeSHugh Dickins 8718fdb3dbfSHugh Dickins if (WARN_ON_ONCE(page_mapped(page))) { 872cbf86cfeSHugh Dickins /* 8738fdb3dbfSHugh Dickins * This should not happen: but if it does, just refuse to let 8748fdb3dbfSHugh Dickins * merge_across_nodes be switched - there is no need to panic. 8758fdb3dbfSHugh Dickins */ 8768fdb3dbfSHugh Dickins err = -EBUSY; 8778fdb3dbfSHugh Dickins } else { 8788fdb3dbfSHugh Dickins /* 8798fdb3dbfSHugh Dickins * The stable node did not yet appear stale to get_ksm_page(), 8808fdb3dbfSHugh Dickins * since that allows for an unmapped ksm page to be recognized 8818fdb3dbfSHugh Dickins * right up until it is freed; but the node is safe to remove. 882cbf86cfeSHugh Dickins * This page might be in a pagevec waiting to be freed, 883cbf86cfeSHugh Dickins * or it might be PageSwapCache (perhaps under writeback), 884cbf86cfeSHugh Dickins * or it might have been removed from swapcache a moment ago. 885cbf86cfeSHugh Dickins */ 886cbf86cfeSHugh Dickins set_page_stable_node(page, NULL); 887cbf86cfeSHugh Dickins remove_node_from_stable_tree(stable_node); 888cbf86cfeSHugh Dickins err = 0; 889cbf86cfeSHugh Dickins } 890cbf86cfeSHugh Dickins 891cbf86cfeSHugh Dickins unlock_page(page); 892cbf86cfeSHugh Dickins put_page(page); 893cbf86cfeSHugh Dickins return err; 894cbf86cfeSHugh Dickins } 895cbf86cfeSHugh Dickins 8962c653d0eSAndrea Arcangeli static int remove_stable_node_chain(struct stable_node *stable_node, 8972c653d0eSAndrea Arcangeli struct rb_root *root) 8982c653d0eSAndrea Arcangeli { 8992c653d0eSAndrea Arcangeli struct stable_node *dup; 9002c653d0eSAndrea Arcangeli struct hlist_node *hlist_safe; 9012c653d0eSAndrea Arcangeli 9022c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 9032c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node)); 9042c653d0eSAndrea Arcangeli if (remove_stable_node(stable_node)) 9052c653d0eSAndrea Arcangeli return true; 9062c653d0eSAndrea Arcangeli else 9072c653d0eSAndrea Arcangeli return false; 9082c653d0eSAndrea Arcangeli } 9092c653d0eSAndrea Arcangeli 9102c653d0eSAndrea Arcangeli hlist_for_each_entry_safe(dup, hlist_safe, 9112c653d0eSAndrea Arcangeli &stable_node->hlist, hlist_dup) { 9122c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(dup)); 9132c653d0eSAndrea Arcangeli if (remove_stable_node(dup)) 9142c653d0eSAndrea Arcangeli return true; 9152c653d0eSAndrea Arcangeli } 9162c653d0eSAndrea Arcangeli BUG_ON(!hlist_empty(&stable_node->hlist)); 9172c653d0eSAndrea Arcangeli free_stable_node_chain(stable_node, root); 9182c653d0eSAndrea Arcangeli return false; 9192c653d0eSAndrea Arcangeli } 9202c653d0eSAndrea Arcangeli 921cbf86cfeSHugh Dickins static int remove_all_stable_nodes(void) 922cbf86cfeSHugh Dickins { 92303640418SGeliang Tang struct stable_node *stable_node, *next; 924cbf86cfeSHugh Dickins int nid; 925cbf86cfeSHugh Dickins int err = 0; 926cbf86cfeSHugh Dickins 927ef53d16cSHugh Dickins for (nid = 0; nid < ksm_nr_node_ids; nid++) { 928cbf86cfeSHugh Dickins while (root_stable_tree[nid].rb_node) { 929cbf86cfeSHugh Dickins stable_node = rb_entry(root_stable_tree[nid].rb_node, 930cbf86cfeSHugh Dickins struct stable_node, node); 9312c653d0eSAndrea Arcangeli if (remove_stable_node_chain(stable_node, 9322c653d0eSAndrea Arcangeli root_stable_tree + nid)) { 933cbf86cfeSHugh Dickins err = -EBUSY; 934cbf86cfeSHugh Dickins break; /* proceed to next nid */ 935cbf86cfeSHugh Dickins } 936cbf86cfeSHugh Dickins cond_resched(); 937cbf86cfeSHugh Dickins } 938cbf86cfeSHugh Dickins } 93903640418SGeliang Tang list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { 9404146d2d6SHugh Dickins if (remove_stable_node(stable_node)) 9414146d2d6SHugh Dickins err = -EBUSY; 9424146d2d6SHugh Dickins cond_resched(); 9434146d2d6SHugh Dickins } 944cbf86cfeSHugh Dickins return err; 945cbf86cfeSHugh Dickins } 946cbf86cfeSHugh Dickins 947d952b791SHugh Dickins static int unmerge_and_remove_all_rmap_items(void) 94831dbd01fSIzik Eidus { 94931dbd01fSIzik Eidus struct mm_slot *mm_slot; 95031dbd01fSIzik Eidus struct mm_struct *mm; 95131dbd01fSIzik Eidus struct vm_area_struct *vma; 952d952b791SHugh Dickins int err = 0; 95331dbd01fSIzik Eidus 954d952b791SHugh Dickins spin_lock(&ksm_mmlist_lock); 9559ba69294SHugh Dickins ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next, 956d952b791SHugh Dickins struct mm_slot, mm_list); 957d952b791SHugh Dickins spin_unlock(&ksm_mmlist_lock); 958d952b791SHugh Dickins 9599ba69294SHugh Dickins for (mm_slot = ksm_scan.mm_slot; 9609ba69294SHugh Dickins mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) { 96131dbd01fSIzik Eidus mm = mm_slot->mm; 96231dbd01fSIzik Eidus down_read(&mm->mmap_sem); 96331dbd01fSIzik Eidus for (vma = mm->mmap; vma; vma = vma->vm_next) { 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 9746514d511SHugh Dickins remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list); 9757496fea9SZhou Chengming up_read(&mm->mmap_sem); 97631dbd01fSIzik Eidus 97731dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 9789ba69294SHugh Dickins ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next, 979d952b791SHugh Dickins struct mm_slot, mm_list); 9809ba69294SHugh Dickins if (ksm_test_exit(mm)) { 9814ca3a69bSSasha Levin hash_del(&mm_slot->link); 9829ba69294SHugh Dickins list_del(&mm_slot->mm_list); 98331dbd01fSIzik Eidus spin_unlock(&ksm_mmlist_lock); 9849ba69294SHugh Dickins 9859ba69294SHugh Dickins free_mm_slot(mm_slot); 9869ba69294SHugh Dickins clear_bit(MMF_VM_MERGEABLE, &mm->flags); 9879ba69294SHugh Dickins mmdrop(mm); 9887496fea9SZhou Chengming } else 9899ba69294SHugh Dickins spin_unlock(&ksm_mmlist_lock); 99031dbd01fSIzik Eidus } 99131dbd01fSIzik Eidus 992cbf86cfeSHugh Dickins /* Clean up stable nodes, but don't worry if some are still busy */ 993cbf86cfeSHugh Dickins remove_all_stable_nodes(); 994d952b791SHugh Dickins ksm_scan.seqnr = 0; 9959ba69294SHugh Dickins return 0; 9969ba69294SHugh Dickins 9979ba69294SHugh Dickins error: 9989ba69294SHugh Dickins up_read(&mm->mmap_sem); 999d952b791SHugh Dickins spin_lock(&ksm_mmlist_lock); 1000d952b791SHugh Dickins ksm_scan.mm_slot = &ksm_mm_head; 1001d952b791SHugh Dickins spin_unlock(&ksm_mmlist_lock); 1002d952b791SHugh Dickins return err; 1003d952b791SHugh Dickins } 10042ffd8679SHugh Dickins #endif /* CONFIG_SYSFS */ 1005d952b791SHugh Dickins 100631dbd01fSIzik Eidus static u32 calc_checksum(struct page *page) 100731dbd01fSIzik Eidus { 100831dbd01fSIzik Eidus u32 checksum; 10099b04c5feSCong Wang void *addr = kmap_atomic(page); 101031dbd01fSIzik Eidus checksum = jhash2(addr, PAGE_SIZE / 4, 17); 10119b04c5feSCong Wang kunmap_atomic(addr); 101231dbd01fSIzik Eidus return checksum; 101331dbd01fSIzik Eidus } 101431dbd01fSIzik Eidus 101531dbd01fSIzik Eidus static int memcmp_pages(struct page *page1, struct page *page2) 101631dbd01fSIzik Eidus { 101731dbd01fSIzik Eidus char *addr1, *addr2; 101831dbd01fSIzik Eidus int ret; 101931dbd01fSIzik Eidus 10209b04c5feSCong Wang addr1 = kmap_atomic(page1); 10219b04c5feSCong Wang addr2 = kmap_atomic(page2); 102231dbd01fSIzik Eidus ret = memcmp(addr1, addr2, PAGE_SIZE); 10239b04c5feSCong Wang kunmap_atomic(addr2); 10249b04c5feSCong Wang kunmap_atomic(addr1); 102531dbd01fSIzik Eidus return ret; 102631dbd01fSIzik Eidus } 102731dbd01fSIzik Eidus 102831dbd01fSIzik Eidus static inline int pages_identical(struct page *page1, struct page *page2) 102931dbd01fSIzik Eidus { 103031dbd01fSIzik Eidus return !memcmp_pages(page1, page2); 103131dbd01fSIzik Eidus } 103231dbd01fSIzik Eidus 103331dbd01fSIzik Eidus static int write_protect_page(struct vm_area_struct *vma, struct page *page, 103431dbd01fSIzik Eidus pte_t *orig_pte) 103531dbd01fSIzik Eidus { 103631dbd01fSIzik Eidus struct mm_struct *mm = vma->vm_mm; 103736eaff33SKirill A. Shutemov struct page_vma_mapped_walk pvmw = { 103836eaff33SKirill A. Shutemov .page = page, 103936eaff33SKirill A. Shutemov .vma = vma, 104036eaff33SKirill A. Shutemov }; 104131dbd01fSIzik Eidus int swapped; 104231dbd01fSIzik Eidus int err = -EFAULT; 10436bdb913fSHaggai Eran unsigned long mmun_start; /* For mmu_notifiers */ 10446bdb913fSHaggai Eran unsigned long mmun_end; /* For mmu_notifiers */ 104531dbd01fSIzik Eidus 104636eaff33SKirill A. Shutemov pvmw.address = page_address_in_vma(page, vma); 104736eaff33SKirill A. Shutemov if (pvmw.address == -EFAULT) 104831dbd01fSIzik Eidus goto out; 104931dbd01fSIzik Eidus 105029ad768cSAndrea Arcangeli BUG_ON(PageTransCompound(page)); 10516bdb913fSHaggai Eran 105236eaff33SKirill A. Shutemov mmun_start = pvmw.address; 105336eaff33SKirill A. Shutemov mmun_end = pvmw.address + PAGE_SIZE; 10546bdb913fSHaggai Eran mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 10556bdb913fSHaggai Eran 105636eaff33SKirill A. Shutemov if (!page_vma_mapped_walk(&pvmw)) 10576bdb913fSHaggai Eran goto out_mn; 105836eaff33SKirill A. Shutemov if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?")) 105936eaff33SKirill A. Shutemov goto out_unlock; 106031dbd01fSIzik Eidus 1061595cd8f2SAneesh Kumar K.V if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) || 1062b3a81d08SMinchan Kim (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) || 1063b3a81d08SMinchan Kim mm_tlb_flush_pending(mm)) { 106431dbd01fSIzik Eidus pte_t entry; 106531dbd01fSIzik Eidus 106631dbd01fSIzik Eidus swapped = PageSwapCache(page); 106736eaff33SKirill A. Shutemov flush_cache_page(vma, pvmw.address, page_to_pfn(page)); 106831dbd01fSIzik Eidus /* 106925985edcSLucas De Marchi * Ok this is tricky, when get_user_pages_fast() run it doesn't 107031dbd01fSIzik Eidus * take any lock, therefore the check that we are going to make 107131dbd01fSIzik Eidus * with the pagecount against the mapcount is racey and 107231dbd01fSIzik Eidus * O_DIRECT can happen right after the check. 107331dbd01fSIzik Eidus * So we clear the pte and flush the tlb before the check 107431dbd01fSIzik Eidus * this assure us that no O_DIRECT can happen after the check 107531dbd01fSIzik Eidus * or in the middle of the check. 10760f10851eSJérôme Glisse * 10770f10851eSJérôme Glisse * No need to notify as we are downgrading page table to read 10780f10851eSJérôme Glisse * only not changing it to point to a new page. 10790f10851eSJérôme Glisse * 1080ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 108131dbd01fSIzik Eidus */ 10820f10851eSJérôme Glisse entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte); 108331dbd01fSIzik Eidus /* 108431dbd01fSIzik Eidus * Check that no O_DIRECT or similar I/O is in progress on the 108531dbd01fSIzik Eidus * page 108631dbd01fSIzik Eidus */ 108731e855eaSHugh Dickins if (page_mapcount(page) + 1 + swapped != page_count(page)) { 108836eaff33SKirill A. Shutemov set_pte_at(mm, pvmw.address, pvmw.pte, entry); 108931dbd01fSIzik Eidus goto out_unlock; 109031dbd01fSIzik Eidus } 10914e31635cSHugh Dickins if (pte_dirty(entry)) 10924e31635cSHugh Dickins set_page_dirty(page); 1093595cd8f2SAneesh Kumar K.V 1094595cd8f2SAneesh Kumar K.V if (pte_protnone(entry)) 1095595cd8f2SAneesh Kumar K.V entry = pte_mkclean(pte_clear_savedwrite(entry)); 1096595cd8f2SAneesh Kumar K.V else 10974e31635cSHugh Dickins entry = pte_mkclean(pte_wrprotect(entry)); 109836eaff33SKirill A. Shutemov set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry); 109931dbd01fSIzik Eidus } 110036eaff33SKirill A. Shutemov *orig_pte = *pvmw.pte; 110131dbd01fSIzik Eidus err = 0; 110231dbd01fSIzik Eidus 110331dbd01fSIzik Eidus out_unlock: 110436eaff33SKirill A. Shutemov page_vma_mapped_walk_done(&pvmw); 11056bdb913fSHaggai Eran out_mn: 11066bdb913fSHaggai Eran mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 110731dbd01fSIzik Eidus out: 110831dbd01fSIzik Eidus return err; 110931dbd01fSIzik Eidus } 111031dbd01fSIzik Eidus 111131dbd01fSIzik Eidus /** 111231dbd01fSIzik Eidus * replace_page - replace page in vma by new ksm page 11138dd3557aSHugh Dickins * @vma: vma that holds the pte pointing to page 11148dd3557aSHugh Dickins * @page: the page we are replacing by kpage 11158dd3557aSHugh Dickins * @kpage: the ksm page we replace page by 111631dbd01fSIzik Eidus * @orig_pte: the original value of the pte 111731dbd01fSIzik Eidus * 111831dbd01fSIzik Eidus * Returns 0 on success, -EFAULT on failure. 111931dbd01fSIzik Eidus */ 11208dd3557aSHugh Dickins static int replace_page(struct vm_area_struct *vma, struct page *page, 11218dd3557aSHugh Dickins struct page *kpage, pte_t orig_pte) 112231dbd01fSIzik Eidus { 112331dbd01fSIzik Eidus struct mm_struct *mm = vma->vm_mm; 112431dbd01fSIzik Eidus pmd_t *pmd; 112531dbd01fSIzik Eidus pte_t *ptep; 1126e86c59b1SClaudio Imbrenda pte_t newpte; 112731dbd01fSIzik Eidus spinlock_t *ptl; 112831dbd01fSIzik Eidus unsigned long addr; 112931dbd01fSIzik Eidus int err = -EFAULT; 11306bdb913fSHaggai Eran unsigned long mmun_start; /* For mmu_notifiers */ 11316bdb913fSHaggai Eran unsigned long mmun_end; /* For mmu_notifiers */ 113231dbd01fSIzik Eidus 11338dd3557aSHugh Dickins addr = page_address_in_vma(page, vma); 113431dbd01fSIzik Eidus if (addr == -EFAULT) 113531dbd01fSIzik Eidus goto out; 113631dbd01fSIzik Eidus 11376219049aSBob Liu pmd = mm_find_pmd(mm, addr); 11386219049aSBob Liu if (!pmd) 113931dbd01fSIzik Eidus goto out; 114031dbd01fSIzik Eidus 11416bdb913fSHaggai Eran mmun_start = addr; 11426bdb913fSHaggai Eran mmun_end = addr + PAGE_SIZE; 11436bdb913fSHaggai Eran mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 11446bdb913fSHaggai Eran 114531dbd01fSIzik Eidus ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); 114631dbd01fSIzik Eidus if (!pte_same(*ptep, orig_pte)) { 114731dbd01fSIzik Eidus pte_unmap_unlock(ptep, ptl); 11486bdb913fSHaggai Eran goto out_mn; 114931dbd01fSIzik Eidus } 115031dbd01fSIzik Eidus 1151e86c59b1SClaudio Imbrenda /* 1152e86c59b1SClaudio Imbrenda * No need to check ksm_use_zero_pages here: we can only have a 1153e86c59b1SClaudio Imbrenda * zero_page here if ksm_use_zero_pages was enabled alreaady. 1154e86c59b1SClaudio Imbrenda */ 1155e86c59b1SClaudio Imbrenda if (!is_zero_pfn(page_to_pfn(kpage))) { 11568dd3557aSHugh Dickins get_page(kpage); 1157d281ee61SKirill A. Shutemov page_add_anon_rmap(kpage, vma, addr, false); 1158e86c59b1SClaudio Imbrenda newpte = mk_pte(kpage, vma->vm_page_prot); 1159e86c59b1SClaudio Imbrenda } else { 1160e86c59b1SClaudio Imbrenda newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage), 1161e86c59b1SClaudio Imbrenda vma->vm_page_prot)); 1162a38c015fSClaudio Imbrenda /* 1163a38c015fSClaudio Imbrenda * We're replacing an anonymous page with a zero page, which is 1164a38c015fSClaudio Imbrenda * not anonymous. We need to do proper accounting otherwise we 1165a38c015fSClaudio Imbrenda * will get wrong values in /proc, and a BUG message in dmesg 1166a38c015fSClaudio Imbrenda * when tearing down the mm. 1167a38c015fSClaudio Imbrenda */ 1168a38c015fSClaudio Imbrenda dec_mm_counter(mm, MM_ANONPAGES); 1169e86c59b1SClaudio Imbrenda } 117031dbd01fSIzik Eidus 117131dbd01fSIzik Eidus flush_cache_page(vma, addr, pte_pfn(*ptep)); 11720f10851eSJérôme Glisse /* 11730f10851eSJérôme Glisse * No need to notify as we are replacing a read only page with another 11740f10851eSJérôme Glisse * read only page with the same content. 11750f10851eSJérôme Glisse * 1176ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 11770f10851eSJérôme Glisse */ 11780f10851eSJérôme Glisse ptep_clear_flush(vma, addr, ptep); 1179e86c59b1SClaudio Imbrenda set_pte_at_notify(mm, addr, ptep, newpte); 118031dbd01fSIzik Eidus 1181d281ee61SKirill A. Shutemov page_remove_rmap(page, false); 1182ae52a2adSHugh Dickins if (!page_mapped(page)) 1183ae52a2adSHugh Dickins try_to_free_swap(page); 11848dd3557aSHugh Dickins put_page(page); 118531dbd01fSIzik Eidus 118631dbd01fSIzik Eidus pte_unmap_unlock(ptep, ptl); 118731dbd01fSIzik Eidus err = 0; 11886bdb913fSHaggai Eran out_mn: 11896bdb913fSHaggai Eran mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 119031dbd01fSIzik Eidus out: 119131dbd01fSIzik Eidus return err; 119231dbd01fSIzik Eidus } 119331dbd01fSIzik Eidus 119431dbd01fSIzik Eidus /* 119531dbd01fSIzik Eidus * try_to_merge_one_page - take two pages and merge them into one 11968dd3557aSHugh Dickins * @vma: the vma that holds the pte pointing to page 11978dd3557aSHugh Dickins * @page: the PageAnon page that we want to replace with kpage 119880e14822SHugh Dickins * @kpage: the PageKsm page that we want to map instead of page, 119980e14822SHugh Dickins * or NULL the first time when we want to use page as kpage. 120031dbd01fSIzik Eidus * 120131dbd01fSIzik Eidus * This function returns 0 if the pages were merged, -EFAULT otherwise. 120231dbd01fSIzik Eidus */ 120331dbd01fSIzik Eidus static int try_to_merge_one_page(struct vm_area_struct *vma, 12048dd3557aSHugh Dickins struct page *page, struct page *kpage) 120531dbd01fSIzik Eidus { 120631dbd01fSIzik Eidus pte_t orig_pte = __pte(0); 120731dbd01fSIzik Eidus int err = -EFAULT; 120831dbd01fSIzik Eidus 1209db114b83SHugh Dickins if (page == kpage) /* ksm page forked */ 1210db114b83SHugh Dickins return 0; 1211db114b83SHugh Dickins 12128dd3557aSHugh Dickins if (!PageAnon(page)) 121331dbd01fSIzik Eidus goto out; 121431dbd01fSIzik Eidus 121531dbd01fSIzik Eidus /* 121631dbd01fSIzik Eidus * We need the page lock to read a stable PageSwapCache in 121731dbd01fSIzik Eidus * write_protect_page(). We use trylock_page() instead of 121831dbd01fSIzik Eidus * lock_page() because we don't want to wait here - we 121931dbd01fSIzik Eidus * prefer to continue scanning and merging different pages, 122031dbd01fSIzik Eidus * then come back to this page when it is unlocked. 122131dbd01fSIzik Eidus */ 12228dd3557aSHugh Dickins if (!trylock_page(page)) 122331e855eaSHugh Dickins goto out; 1224f765f540SKirill A. Shutemov 1225f765f540SKirill A. Shutemov if (PageTransCompound(page)) { 1226a7306c34SAndrea Arcangeli if (split_huge_page(page)) 1227f765f540SKirill A. Shutemov goto out_unlock; 1228f765f540SKirill A. Shutemov } 1229f765f540SKirill A. Shutemov 123031dbd01fSIzik Eidus /* 123131dbd01fSIzik Eidus * If this anonymous page is mapped only here, its pte may need 123231dbd01fSIzik Eidus * to be write-protected. If it's mapped elsewhere, all of its 123331dbd01fSIzik Eidus * ptes are necessarily already write-protected. But in either 123431dbd01fSIzik Eidus * case, we need to lock and check page_count is not raised. 123531dbd01fSIzik Eidus */ 123680e14822SHugh Dickins if (write_protect_page(vma, page, &orig_pte) == 0) { 123780e14822SHugh Dickins if (!kpage) { 123880e14822SHugh Dickins /* 123980e14822SHugh Dickins * While we hold page lock, upgrade page from 124080e14822SHugh Dickins * PageAnon+anon_vma to PageKsm+NULL stable_node: 124180e14822SHugh Dickins * stable_tree_insert() will update stable_node. 124280e14822SHugh Dickins */ 124380e14822SHugh Dickins set_page_stable_node(page, NULL); 124480e14822SHugh Dickins mark_page_accessed(page); 1245337ed7ebSMinchan Kim /* 1246337ed7ebSMinchan Kim * Page reclaim just frees a clean page with no dirty 1247337ed7ebSMinchan Kim * ptes: make sure that the ksm page would be swapped. 1248337ed7ebSMinchan Kim */ 1249337ed7ebSMinchan Kim if (!PageDirty(page)) 1250337ed7ebSMinchan Kim SetPageDirty(page); 125180e14822SHugh Dickins err = 0; 125280e14822SHugh Dickins } else if (pages_identical(page, kpage)) 12538dd3557aSHugh Dickins err = replace_page(vma, page, kpage, orig_pte); 125480e14822SHugh Dickins } 125531dbd01fSIzik Eidus 125680e14822SHugh Dickins if ((vma->vm_flags & VM_LOCKED) && kpage && !err) { 125773848b46SHugh Dickins munlock_vma_page(page); 12585ad64688SHugh Dickins if (!PageMlocked(kpage)) { 12595ad64688SHugh Dickins unlock_page(page); 12605ad64688SHugh Dickins lock_page(kpage); 12615ad64688SHugh Dickins mlock_vma_page(kpage); 12625ad64688SHugh Dickins page = kpage; /* for final unlock */ 12635ad64688SHugh Dickins } 12645ad64688SHugh Dickins } 126573848b46SHugh Dickins 1266f765f540SKirill A. Shutemov out_unlock: 12678dd3557aSHugh Dickins unlock_page(page); 126831dbd01fSIzik Eidus out: 126931dbd01fSIzik Eidus return err; 127031dbd01fSIzik Eidus } 127131dbd01fSIzik Eidus 127231dbd01fSIzik Eidus /* 127381464e30SHugh Dickins * try_to_merge_with_ksm_page - like try_to_merge_two_pages, 127481464e30SHugh Dickins * but no new kernel page is allocated: kpage must already be a ksm page. 12758dd3557aSHugh Dickins * 12768dd3557aSHugh Dickins * This function returns 0 if the pages were merged, -EFAULT otherwise. 127781464e30SHugh Dickins */ 12788dd3557aSHugh Dickins static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item, 12798dd3557aSHugh Dickins struct page *page, struct page *kpage) 128081464e30SHugh Dickins { 12818dd3557aSHugh Dickins struct mm_struct *mm = rmap_item->mm; 128281464e30SHugh Dickins struct vm_area_struct *vma; 128381464e30SHugh Dickins int err = -EFAULT; 128481464e30SHugh Dickins 12858dd3557aSHugh Dickins down_read(&mm->mmap_sem); 128685c6e8ddSAndrea Arcangeli vma = find_mergeable_vma(mm, rmap_item->address); 128785c6e8ddSAndrea Arcangeli if (!vma) 12889ba69294SHugh Dickins goto out; 12899ba69294SHugh Dickins 12908dd3557aSHugh Dickins err = try_to_merge_one_page(vma, page, kpage); 1291db114b83SHugh Dickins if (err) 1292db114b83SHugh Dickins goto out; 1293db114b83SHugh Dickins 1294bc56620bSHugh Dickins /* Unstable nid is in union with stable anon_vma: remove first */ 1295bc56620bSHugh Dickins remove_rmap_item_from_tree(rmap_item); 1296bc56620bSHugh Dickins 1297db114b83SHugh Dickins /* Must get reference to anon_vma while still holding mmap_sem */ 12989e60109fSPeter Zijlstra rmap_item->anon_vma = vma->anon_vma; 12999e60109fSPeter Zijlstra get_anon_vma(vma->anon_vma); 130081464e30SHugh Dickins out: 13018dd3557aSHugh Dickins up_read(&mm->mmap_sem); 130281464e30SHugh Dickins return err; 130381464e30SHugh Dickins } 130481464e30SHugh Dickins 130581464e30SHugh Dickins /* 130631dbd01fSIzik Eidus * try_to_merge_two_pages - take two identical pages and prepare them 130731dbd01fSIzik Eidus * to be merged into one page. 130831dbd01fSIzik Eidus * 13098dd3557aSHugh Dickins * This function returns the kpage if we successfully merged two identical 13108dd3557aSHugh Dickins * pages into one ksm page, NULL otherwise. 131131dbd01fSIzik Eidus * 131280e14822SHugh Dickins * Note that this function upgrades page to ksm page: if one of the pages 131331dbd01fSIzik Eidus * is already a ksm page, try_to_merge_with_ksm_page should be used. 131431dbd01fSIzik Eidus */ 13158dd3557aSHugh Dickins static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item, 13168dd3557aSHugh Dickins struct page *page, 13178dd3557aSHugh Dickins struct rmap_item *tree_rmap_item, 13188dd3557aSHugh Dickins struct page *tree_page) 131931dbd01fSIzik Eidus { 132080e14822SHugh Dickins int err; 132131dbd01fSIzik Eidus 132280e14822SHugh Dickins err = try_to_merge_with_ksm_page(rmap_item, page, NULL); 132331dbd01fSIzik Eidus if (!err) { 13248dd3557aSHugh Dickins err = try_to_merge_with_ksm_page(tree_rmap_item, 132580e14822SHugh Dickins tree_page, page); 132631dbd01fSIzik Eidus /* 132781464e30SHugh Dickins * If that fails, we have a ksm page with only one pte 132881464e30SHugh Dickins * pointing to it: so break it. 132931dbd01fSIzik Eidus */ 13304035c07aSHugh Dickins if (err) 13318dd3557aSHugh Dickins break_cow(rmap_item); 133231dbd01fSIzik Eidus } 133380e14822SHugh Dickins return err ? NULL : page; 133431dbd01fSIzik Eidus } 133531dbd01fSIzik Eidus 13362c653d0eSAndrea Arcangeli static __always_inline 13372c653d0eSAndrea Arcangeli bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset) 13382c653d0eSAndrea Arcangeli { 13392c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len < 0); 13402c653d0eSAndrea Arcangeli /* 13412c653d0eSAndrea Arcangeli * Check that at least one mapping still exists, otherwise 13422c653d0eSAndrea Arcangeli * there's no much point to merge and share with this 13432c653d0eSAndrea Arcangeli * stable_node, as the underlying tree_page of the other 13442c653d0eSAndrea Arcangeli * sharer is going to be freed soon. 13452c653d0eSAndrea Arcangeli */ 13462c653d0eSAndrea Arcangeli return stable_node->rmap_hlist_len && 13472c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len + offset < ksm_max_page_sharing; 13482c653d0eSAndrea Arcangeli } 13492c653d0eSAndrea Arcangeli 13502c653d0eSAndrea Arcangeli static __always_inline 13512c653d0eSAndrea Arcangeli bool is_page_sharing_candidate(struct stable_node *stable_node) 13522c653d0eSAndrea Arcangeli { 13532c653d0eSAndrea Arcangeli return __is_page_sharing_candidate(stable_node, 0); 13542c653d0eSAndrea Arcangeli } 13552c653d0eSAndrea Arcangeli 1356c01f0b54SColin Ian King static struct page *stable_node_dup(struct stable_node **_stable_node_dup, 13578dc5ffcdSAndrea Arcangeli struct stable_node **_stable_node, 13582c653d0eSAndrea Arcangeli struct rb_root *root, 13592c653d0eSAndrea Arcangeli bool prune_stale_stable_nodes) 13602c653d0eSAndrea Arcangeli { 1361b4fecc67SAndrea Arcangeli struct stable_node *dup, *found = NULL, *stable_node = *_stable_node; 13622c653d0eSAndrea Arcangeli struct hlist_node *hlist_safe; 13638dc5ffcdSAndrea Arcangeli struct page *_tree_page, *tree_page = NULL; 13642c653d0eSAndrea Arcangeli int nr = 0; 13652c653d0eSAndrea Arcangeli int found_rmap_hlist_len; 13662c653d0eSAndrea Arcangeli 13672c653d0eSAndrea Arcangeli if (!prune_stale_stable_nodes || 13682c653d0eSAndrea Arcangeli time_before(jiffies, stable_node->chain_prune_time + 13692c653d0eSAndrea Arcangeli msecs_to_jiffies( 13702c653d0eSAndrea Arcangeli ksm_stable_node_chains_prune_millisecs))) 13712c653d0eSAndrea Arcangeli prune_stale_stable_nodes = false; 13722c653d0eSAndrea Arcangeli else 13732c653d0eSAndrea Arcangeli stable_node->chain_prune_time = jiffies; 13742c653d0eSAndrea Arcangeli 13752c653d0eSAndrea Arcangeli hlist_for_each_entry_safe(dup, hlist_safe, 13762c653d0eSAndrea Arcangeli &stable_node->hlist, hlist_dup) { 13772c653d0eSAndrea Arcangeli cond_resched(); 13782c653d0eSAndrea Arcangeli /* 13792c653d0eSAndrea Arcangeli * We must walk all stable_node_dup to prune the stale 13802c653d0eSAndrea Arcangeli * stable nodes during lookup. 13812c653d0eSAndrea Arcangeli * 13822c653d0eSAndrea Arcangeli * get_ksm_page can drop the nodes from the 13832c653d0eSAndrea Arcangeli * stable_node->hlist if they point to freed pages 13842c653d0eSAndrea Arcangeli * (that's why we do a _safe walk). The "dup" 13852c653d0eSAndrea Arcangeli * stable_node parameter itself will be freed from 13862c653d0eSAndrea Arcangeli * under us if it returns NULL. 13872c653d0eSAndrea Arcangeli */ 13882c653d0eSAndrea Arcangeli _tree_page = get_ksm_page(dup, false); 13892c653d0eSAndrea Arcangeli if (!_tree_page) 13902c653d0eSAndrea Arcangeli continue; 13912c653d0eSAndrea Arcangeli nr += 1; 13922c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(dup)) { 13932c653d0eSAndrea Arcangeli if (!found || 13942c653d0eSAndrea Arcangeli dup->rmap_hlist_len > found_rmap_hlist_len) { 13952c653d0eSAndrea Arcangeli if (found) 13968dc5ffcdSAndrea Arcangeli put_page(tree_page); 13972c653d0eSAndrea Arcangeli found = dup; 13982c653d0eSAndrea Arcangeli found_rmap_hlist_len = found->rmap_hlist_len; 13998dc5ffcdSAndrea Arcangeli tree_page = _tree_page; 14002c653d0eSAndrea Arcangeli 14018dc5ffcdSAndrea Arcangeli /* skip put_page for found dup */ 14022c653d0eSAndrea Arcangeli if (!prune_stale_stable_nodes) 14032c653d0eSAndrea Arcangeli break; 14042c653d0eSAndrea Arcangeli continue; 14052c653d0eSAndrea Arcangeli } 14062c653d0eSAndrea Arcangeli } 14072c653d0eSAndrea Arcangeli put_page(_tree_page); 14082c653d0eSAndrea Arcangeli } 14092c653d0eSAndrea Arcangeli 141080b18dfaSAndrea Arcangeli if (found) { 14112c653d0eSAndrea Arcangeli /* 141280b18dfaSAndrea Arcangeli * nr is counting all dups in the chain only if 141380b18dfaSAndrea Arcangeli * prune_stale_stable_nodes is true, otherwise we may 141480b18dfaSAndrea Arcangeli * break the loop at nr == 1 even if there are 141580b18dfaSAndrea Arcangeli * multiple entries. 14162c653d0eSAndrea Arcangeli */ 141780b18dfaSAndrea Arcangeli if (prune_stale_stable_nodes && nr == 1) { 14182c653d0eSAndrea Arcangeli /* 14192c653d0eSAndrea Arcangeli * If there's not just one entry it would 14202c653d0eSAndrea Arcangeli * corrupt memory, better BUG_ON. In KSM 14212c653d0eSAndrea Arcangeli * context with no lock held it's not even 14222c653d0eSAndrea Arcangeli * fatal. 14232c653d0eSAndrea Arcangeli */ 14242c653d0eSAndrea Arcangeli BUG_ON(stable_node->hlist.first->next); 14252c653d0eSAndrea Arcangeli 14262c653d0eSAndrea Arcangeli /* 14272c653d0eSAndrea Arcangeli * There's just one entry and it is below the 14282c653d0eSAndrea Arcangeli * deduplication limit so drop the chain. 14292c653d0eSAndrea Arcangeli */ 14302c653d0eSAndrea Arcangeli rb_replace_node(&stable_node->node, &found->node, 14312c653d0eSAndrea Arcangeli root); 14322c653d0eSAndrea Arcangeli free_stable_node(stable_node); 14332c653d0eSAndrea Arcangeli ksm_stable_node_chains--; 14342c653d0eSAndrea Arcangeli ksm_stable_node_dups--; 1435b4fecc67SAndrea Arcangeli /* 14360ba1d0f7SAndrea Arcangeli * NOTE: the caller depends on the stable_node 14370ba1d0f7SAndrea Arcangeli * to be equal to stable_node_dup if the chain 14380ba1d0f7SAndrea Arcangeli * was collapsed. 1439b4fecc67SAndrea Arcangeli */ 14400ba1d0f7SAndrea Arcangeli *_stable_node = found; 14410ba1d0f7SAndrea Arcangeli /* 14420ba1d0f7SAndrea Arcangeli * Just for robustneess as stable_node is 14430ba1d0f7SAndrea Arcangeli * otherwise left as a stable pointer, the 14440ba1d0f7SAndrea Arcangeli * compiler shall optimize it away at build 14450ba1d0f7SAndrea Arcangeli * time. 14460ba1d0f7SAndrea Arcangeli */ 14470ba1d0f7SAndrea Arcangeli stable_node = NULL; 144880b18dfaSAndrea Arcangeli } else if (stable_node->hlist.first != &found->hlist_dup && 144980b18dfaSAndrea Arcangeli __is_page_sharing_candidate(found, 1)) { 14502c653d0eSAndrea Arcangeli /* 145180b18dfaSAndrea Arcangeli * If the found stable_node dup can accept one 145280b18dfaSAndrea Arcangeli * more future merge (in addition to the one 145380b18dfaSAndrea Arcangeli * that is underway) and is not at the head of 145480b18dfaSAndrea Arcangeli * the chain, put it there so next search will 145580b18dfaSAndrea Arcangeli * be quicker in the !prune_stale_stable_nodes 145680b18dfaSAndrea Arcangeli * case. 145780b18dfaSAndrea Arcangeli * 145880b18dfaSAndrea Arcangeli * NOTE: it would be inaccurate to use nr > 1 145980b18dfaSAndrea Arcangeli * instead of checking the hlist.first pointer 146080b18dfaSAndrea Arcangeli * directly, because in the 146180b18dfaSAndrea Arcangeli * prune_stale_stable_nodes case "nr" isn't 146280b18dfaSAndrea Arcangeli * the position of the found dup in the chain, 146380b18dfaSAndrea Arcangeli * but the total number of dups in the chain. 14642c653d0eSAndrea Arcangeli */ 14652c653d0eSAndrea Arcangeli hlist_del(&found->hlist_dup); 14662c653d0eSAndrea Arcangeli hlist_add_head(&found->hlist_dup, 14672c653d0eSAndrea Arcangeli &stable_node->hlist); 14682c653d0eSAndrea Arcangeli } 14692c653d0eSAndrea Arcangeli } 14702c653d0eSAndrea Arcangeli 14718dc5ffcdSAndrea Arcangeli *_stable_node_dup = found; 14728dc5ffcdSAndrea Arcangeli return tree_page; 14732c653d0eSAndrea Arcangeli } 14742c653d0eSAndrea Arcangeli 14752c653d0eSAndrea Arcangeli static struct stable_node *stable_node_dup_any(struct stable_node *stable_node, 14762c653d0eSAndrea Arcangeli struct rb_root *root) 14772c653d0eSAndrea Arcangeli { 14782c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) 14792c653d0eSAndrea Arcangeli return stable_node; 14802c653d0eSAndrea Arcangeli if (hlist_empty(&stable_node->hlist)) { 14812c653d0eSAndrea Arcangeli free_stable_node_chain(stable_node, root); 14822c653d0eSAndrea Arcangeli return NULL; 14832c653d0eSAndrea Arcangeli } 14842c653d0eSAndrea Arcangeli return hlist_entry(stable_node->hlist.first, 14852c653d0eSAndrea Arcangeli typeof(*stable_node), hlist_dup); 14862c653d0eSAndrea Arcangeli } 14872c653d0eSAndrea Arcangeli 14888dc5ffcdSAndrea Arcangeli /* 14898dc5ffcdSAndrea Arcangeli * Like for get_ksm_page, this function can free the *_stable_node and 14908dc5ffcdSAndrea Arcangeli * *_stable_node_dup if the returned tree_page is NULL. 14918dc5ffcdSAndrea Arcangeli * 14928dc5ffcdSAndrea Arcangeli * It can also free and overwrite *_stable_node with the found 14938dc5ffcdSAndrea Arcangeli * stable_node_dup if the chain is collapsed (in which case 14948dc5ffcdSAndrea Arcangeli * *_stable_node will be equal to *_stable_node_dup like if the chain 14958dc5ffcdSAndrea Arcangeli * never existed). It's up to the caller to verify tree_page is not 14968dc5ffcdSAndrea Arcangeli * NULL before dereferencing *_stable_node or *_stable_node_dup. 14978dc5ffcdSAndrea Arcangeli * 14988dc5ffcdSAndrea Arcangeli * *_stable_node_dup is really a second output parameter of this 14998dc5ffcdSAndrea Arcangeli * function and will be overwritten in all cases, the caller doesn't 15008dc5ffcdSAndrea Arcangeli * need to initialize it. 15018dc5ffcdSAndrea Arcangeli */ 15028dc5ffcdSAndrea Arcangeli static struct page *__stable_node_chain(struct stable_node **_stable_node_dup, 15038dc5ffcdSAndrea Arcangeli struct stable_node **_stable_node, 15042c653d0eSAndrea Arcangeli struct rb_root *root, 15052c653d0eSAndrea Arcangeli bool prune_stale_stable_nodes) 15062c653d0eSAndrea Arcangeli { 1507b4fecc67SAndrea Arcangeli struct stable_node *stable_node = *_stable_node; 15082c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 15092c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(stable_node)) { 15108dc5ffcdSAndrea Arcangeli *_stable_node_dup = stable_node; 15118dc5ffcdSAndrea Arcangeli return get_ksm_page(stable_node, false); 15122c653d0eSAndrea Arcangeli } 15138dc5ffcdSAndrea Arcangeli /* 15148dc5ffcdSAndrea Arcangeli * _stable_node_dup set to NULL means the stable_node 15158dc5ffcdSAndrea Arcangeli * reached the ksm_max_page_sharing limit. 15168dc5ffcdSAndrea Arcangeli */ 15178dc5ffcdSAndrea Arcangeli *_stable_node_dup = NULL; 15182c653d0eSAndrea Arcangeli return NULL; 15192c653d0eSAndrea Arcangeli } 15208dc5ffcdSAndrea Arcangeli return stable_node_dup(_stable_node_dup, _stable_node, root, 15212c653d0eSAndrea Arcangeli prune_stale_stable_nodes); 15222c653d0eSAndrea Arcangeli } 15232c653d0eSAndrea Arcangeli 15248dc5ffcdSAndrea Arcangeli static __always_inline struct page *chain_prune(struct stable_node **s_n_d, 15258dc5ffcdSAndrea Arcangeli struct stable_node **s_n, 15262c653d0eSAndrea Arcangeli struct rb_root *root) 15272c653d0eSAndrea Arcangeli { 15288dc5ffcdSAndrea Arcangeli return __stable_node_chain(s_n_d, s_n, root, true); 15292c653d0eSAndrea Arcangeli } 15302c653d0eSAndrea Arcangeli 15318dc5ffcdSAndrea Arcangeli static __always_inline struct page *chain(struct stable_node **s_n_d, 15328dc5ffcdSAndrea Arcangeli struct stable_node *s_n, 15332c653d0eSAndrea Arcangeli struct rb_root *root) 15342c653d0eSAndrea Arcangeli { 15358dc5ffcdSAndrea Arcangeli struct stable_node *old_stable_node = s_n; 15368dc5ffcdSAndrea Arcangeli struct page *tree_page; 15378dc5ffcdSAndrea Arcangeli 15388dc5ffcdSAndrea Arcangeli tree_page = __stable_node_chain(s_n_d, &s_n, root, false); 15398dc5ffcdSAndrea Arcangeli /* not pruning dups so s_n cannot have changed */ 15408dc5ffcdSAndrea Arcangeli VM_BUG_ON(s_n != old_stable_node); 15418dc5ffcdSAndrea Arcangeli return tree_page; 15422c653d0eSAndrea Arcangeli } 15432c653d0eSAndrea Arcangeli 154431dbd01fSIzik Eidus /* 15458dd3557aSHugh Dickins * stable_tree_search - search for page inside the stable tree 154631dbd01fSIzik Eidus * 154731dbd01fSIzik Eidus * This function checks if there is a page inside the stable tree 154831dbd01fSIzik Eidus * with identical content to the page that we are scanning right now. 154931dbd01fSIzik Eidus * 15507b6ba2c7SHugh Dickins * This function returns the stable tree node of identical content if found, 155131dbd01fSIzik Eidus * NULL otherwise. 155231dbd01fSIzik Eidus */ 155362b61f61SHugh Dickins static struct page *stable_tree_search(struct page *page) 155431dbd01fSIzik Eidus { 155590bd6fd3SPetr Holasek int nid; 1556ef53d16cSHugh Dickins struct rb_root *root; 15574146d2d6SHugh Dickins struct rb_node **new; 15584146d2d6SHugh Dickins struct rb_node *parent; 15592c653d0eSAndrea Arcangeli struct stable_node *stable_node, *stable_node_dup, *stable_node_any; 15604146d2d6SHugh Dickins struct stable_node *page_node; 156131dbd01fSIzik Eidus 15624146d2d6SHugh Dickins page_node = page_stable_node(page); 15634146d2d6SHugh Dickins if (page_node && page_node->head != &migrate_nodes) { 15644146d2d6SHugh Dickins /* ksm page forked */ 156508beca44SHugh Dickins get_page(page); 156662b61f61SHugh Dickins return page; 156708beca44SHugh Dickins } 156808beca44SHugh Dickins 156990bd6fd3SPetr Holasek nid = get_kpfn_nid(page_to_pfn(page)); 1570ef53d16cSHugh Dickins root = root_stable_tree + nid; 15714146d2d6SHugh Dickins again: 1572ef53d16cSHugh Dickins new = &root->rb_node; 15734146d2d6SHugh Dickins parent = NULL; 157490bd6fd3SPetr Holasek 15754146d2d6SHugh Dickins while (*new) { 15764035c07aSHugh Dickins struct page *tree_page; 157731dbd01fSIzik Eidus int ret; 157831dbd01fSIzik Eidus 157931dbd01fSIzik Eidus cond_resched(); 15804146d2d6SHugh Dickins stable_node = rb_entry(*new, struct stable_node, node); 15812c653d0eSAndrea Arcangeli stable_node_any = NULL; 15828dc5ffcdSAndrea Arcangeli tree_page = chain_prune(&stable_node_dup, &stable_node, root); 1583b4fecc67SAndrea Arcangeli /* 1584b4fecc67SAndrea Arcangeli * NOTE: stable_node may have been freed by 1585b4fecc67SAndrea Arcangeli * chain_prune() if the returned stable_node_dup is 1586b4fecc67SAndrea Arcangeli * not NULL. stable_node_dup may have been inserted in 1587b4fecc67SAndrea Arcangeli * the rbtree instead as a regular stable_node (in 1588b4fecc67SAndrea Arcangeli * order to collapse the stable_node chain if a single 15890ba1d0f7SAndrea Arcangeli * stable_node dup was found in it). In such case the 15900ba1d0f7SAndrea Arcangeli * stable_node is overwritten by the calleee to point 15910ba1d0f7SAndrea Arcangeli * to the stable_node_dup that was collapsed in the 15920ba1d0f7SAndrea Arcangeli * stable rbtree and stable_node will be equal to 15930ba1d0f7SAndrea Arcangeli * stable_node_dup like if the chain never existed. 1594b4fecc67SAndrea Arcangeli */ 15952c653d0eSAndrea Arcangeli if (!stable_node_dup) { 15962c653d0eSAndrea Arcangeli /* 15972c653d0eSAndrea Arcangeli * Either all stable_node dups were full in 15982c653d0eSAndrea Arcangeli * this stable_node chain, or this chain was 15992c653d0eSAndrea Arcangeli * empty and should be rb_erased. 16002c653d0eSAndrea Arcangeli */ 16012c653d0eSAndrea Arcangeli stable_node_any = stable_node_dup_any(stable_node, 16022c653d0eSAndrea Arcangeli root); 16032c653d0eSAndrea Arcangeli if (!stable_node_any) { 16042c653d0eSAndrea Arcangeli /* rb_erase just run */ 16052c653d0eSAndrea Arcangeli goto again; 16062c653d0eSAndrea Arcangeli } 16072c653d0eSAndrea Arcangeli /* 16082c653d0eSAndrea Arcangeli * Take any of the stable_node dups page of 16092c653d0eSAndrea Arcangeli * this stable_node chain to let the tree walk 16102c653d0eSAndrea Arcangeli * continue. All KSM pages belonging to the 16112c653d0eSAndrea Arcangeli * stable_node dups in a stable_node chain 16122c653d0eSAndrea Arcangeli * have the same content and they're 16132c653d0eSAndrea Arcangeli * wrprotected at all times. Any will work 16142c653d0eSAndrea Arcangeli * fine to continue the walk. 16152c653d0eSAndrea Arcangeli */ 16162c653d0eSAndrea Arcangeli tree_page = get_ksm_page(stable_node_any, false); 16172c653d0eSAndrea Arcangeli } 16182c653d0eSAndrea Arcangeli VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); 1619f2e5ff85SAndrea Arcangeli if (!tree_page) { 1620f2e5ff85SAndrea Arcangeli /* 1621f2e5ff85SAndrea Arcangeli * If we walked over a stale stable_node, 1622f2e5ff85SAndrea Arcangeli * get_ksm_page() will call rb_erase() and it 1623f2e5ff85SAndrea Arcangeli * may rebalance the tree from under us. So 1624f2e5ff85SAndrea Arcangeli * restart the search from scratch. Returning 1625f2e5ff85SAndrea Arcangeli * NULL would be safe too, but we'd generate 1626f2e5ff85SAndrea Arcangeli * false negative insertions just because some 1627f2e5ff85SAndrea Arcangeli * stable_node was stale. 1628f2e5ff85SAndrea Arcangeli */ 1629f2e5ff85SAndrea Arcangeli goto again; 1630f2e5ff85SAndrea Arcangeli } 163131dbd01fSIzik Eidus 16324035c07aSHugh Dickins ret = memcmp_pages(page, tree_page); 1633c8d6553bSHugh Dickins put_page(tree_page); 163431dbd01fSIzik Eidus 16354146d2d6SHugh Dickins parent = *new; 1636c8d6553bSHugh Dickins if (ret < 0) 16374146d2d6SHugh Dickins new = &parent->rb_left; 1638c8d6553bSHugh Dickins else if (ret > 0) 16394146d2d6SHugh Dickins new = &parent->rb_right; 1640c8d6553bSHugh Dickins else { 16412c653d0eSAndrea Arcangeli if (page_node) { 16422c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 16432c653d0eSAndrea Arcangeli /* 16442c653d0eSAndrea Arcangeli * Test if the migrated page should be merged 16452c653d0eSAndrea Arcangeli * into a stable node dup. If the mapcount is 16462c653d0eSAndrea Arcangeli * 1 we can migrate it with another KSM page 16472c653d0eSAndrea Arcangeli * without adding it to the chain. 16482c653d0eSAndrea Arcangeli */ 16492c653d0eSAndrea Arcangeli if (page_mapcount(page) > 1) 16502c653d0eSAndrea Arcangeli goto chain_append; 16512c653d0eSAndrea Arcangeli } 16522c653d0eSAndrea Arcangeli 16532c653d0eSAndrea Arcangeli if (!stable_node_dup) { 16542c653d0eSAndrea Arcangeli /* 16552c653d0eSAndrea Arcangeli * If the stable_node is a chain and 16562c653d0eSAndrea Arcangeli * we got a payload match in memcmp 16572c653d0eSAndrea Arcangeli * but we cannot merge the scanned 16582c653d0eSAndrea Arcangeli * page in any of the existing 16592c653d0eSAndrea Arcangeli * stable_node dups because they're 16602c653d0eSAndrea Arcangeli * all full, we need to wait the 16612c653d0eSAndrea Arcangeli * scanned page to find itself a match 16622c653d0eSAndrea Arcangeli * in the unstable tree to create a 16632c653d0eSAndrea Arcangeli * brand new KSM page to add later to 16642c653d0eSAndrea Arcangeli * the dups of this stable_node. 16652c653d0eSAndrea Arcangeli */ 16662c653d0eSAndrea Arcangeli return NULL; 16672c653d0eSAndrea Arcangeli } 16682c653d0eSAndrea Arcangeli 1669c8d6553bSHugh Dickins /* 1670c8d6553bSHugh Dickins * Lock and unlock the stable_node's page (which 1671c8d6553bSHugh Dickins * might already have been migrated) so that page 1672c8d6553bSHugh Dickins * migration is sure to notice its raised count. 1673c8d6553bSHugh Dickins * It would be more elegant to return stable_node 1674c8d6553bSHugh Dickins * than kpage, but that involves more changes. 1675c8d6553bSHugh Dickins */ 16762c653d0eSAndrea Arcangeli tree_page = get_ksm_page(stable_node_dup, true); 16772c653d0eSAndrea Arcangeli if (unlikely(!tree_page)) 16782c653d0eSAndrea Arcangeli /* 16792c653d0eSAndrea Arcangeli * The tree may have been rebalanced, 16802c653d0eSAndrea Arcangeli * so re-evaluate parent and new. 16812c653d0eSAndrea Arcangeli */ 16822c653d0eSAndrea Arcangeli goto again; 1683c8d6553bSHugh Dickins unlock_page(tree_page); 16842c653d0eSAndrea Arcangeli 16852c653d0eSAndrea Arcangeli if (get_kpfn_nid(stable_node_dup->kpfn) != 16862c653d0eSAndrea Arcangeli NUMA(stable_node_dup->nid)) { 16874146d2d6SHugh Dickins put_page(tree_page); 16884146d2d6SHugh Dickins goto replace; 16894146d2d6SHugh Dickins } 169062b61f61SHugh Dickins return tree_page; 169131dbd01fSIzik Eidus } 1692c8d6553bSHugh Dickins } 169331dbd01fSIzik Eidus 16944146d2d6SHugh Dickins if (!page_node) 169531dbd01fSIzik Eidus return NULL; 16964146d2d6SHugh Dickins 16974146d2d6SHugh Dickins list_del(&page_node->list); 16984146d2d6SHugh Dickins DO_NUMA(page_node->nid = nid); 16994146d2d6SHugh Dickins rb_link_node(&page_node->node, parent, new); 1700ef53d16cSHugh Dickins rb_insert_color(&page_node->node, root); 17012c653d0eSAndrea Arcangeli out: 17022c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(page_node)) { 17034146d2d6SHugh Dickins get_page(page); 17044146d2d6SHugh Dickins return page; 17052c653d0eSAndrea Arcangeli } else 17062c653d0eSAndrea Arcangeli return NULL; 17074146d2d6SHugh Dickins 17084146d2d6SHugh Dickins replace: 1709b4fecc67SAndrea Arcangeli /* 1710b4fecc67SAndrea Arcangeli * If stable_node was a chain and chain_prune collapsed it, 17110ba1d0f7SAndrea Arcangeli * stable_node has been updated to be the new regular 17120ba1d0f7SAndrea Arcangeli * stable_node. A collapse of the chain is indistinguishable 17130ba1d0f7SAndrea Arcangeli * from the case there was no chain in the stable 17140ba1d0f7SAndrea Arcangeli * rbtree. Otherwise stable_node is the chain and 17150ba1d0f7SAndrea Arcangeli * stable_node_dup is the dup to replace. 1716b4fecc67SAndrea Arcangeli */ 17170ba1d0f7SAndrea Arcangeli if (stable_node_dup == stable_node) { 1718b4fecc67SAndrea Arcangeli VM_BUG_ON(is_stable_node_chain(stable_node_dup)); 1719b4fecc67SAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node_dup)); 17202c653d0eSAndrea Arcangeli /* there is no chain */ 17214146d2d6SHugh Dickins if (page_node) { 17222c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 17234146d2d6SHugh Dickins list_del(&page_node->list); 17244146d2d6SHugh Dickins DO_NUMA(page_node->nid = nid); 1725b4fecc67SAndrea Arcangeli rb_replace_node(&stable_node_dup->node, 1726b4fecc67SAndrea Arcangeli &page_node->node, 17272c653d0eSAndrea Arcangeli root); 17282c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(page_node)) 17294146d2d6SHugh Dickins get_page(page); 17302c653d0eSAndrea Arcangeli else 17312c653d0eSAndrea Arcangeli page = NULL; 17324146d2d6SHugh Dickins } else { 1733b4fecc67SAndrea Arcangeli rb_erase(&stable_node_dup->node, root); 17344146d2d6SHugh Dickins page = NULL; 17354146d2d6SHugh Dickins } 17362c653d0eSAndrea Arcangeli } else { 17372c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_chain(stable_node)); 17382c653d0eSAndrea Arcangeli __stable_node_dup_del(stable_node_dup); 17392c653d0eSAndrea Arcangeli if (page_node) { 17402c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 17412c653d0eSAndrea Arcangeli list_del(&page_node->list); 17422c653d0eSAndrea Arcangeli DO_NUMA(page_node->nid = nid); 17432c653d0eSAndrea Arcangeli stable_node_chain_add_dup(page_node, stable_node); 17442c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(page_node)) 17452c653d0eSAndrea Arcangeli get_page(page); 17462c653d0eSAndrea Arcangeli else 17472c653d0eSAndrea Arcangeli page = NULL; 17482c653d0eSAndrea Arcangeli } else { 17492c653d0eSAndrea Arcangeli page = NULL; 17502c653d0eSAndrea Arcangeli } 17512c653d0eSAndrea Arcangeli } 17522c653d0eSAndrea Arcangeli stable_node_dup->head = &migrate_nodes; 17532c653d0eSAndrea Arcangeli list_add(&stable_node_dup->list, stable_node_dup->head); 17544146d2d6SHugh Dickins return page; 17552c653d0eSAndrea Arcangeli 17562c653d0eSAndrea Arcangeli chain_append: 17572c653d0eSAndrea Arcangeli /* stable_node_dup could be null if it reached the limit */ 17582c653d0eSAndrea Arcangeli if (!stable_node_dup) 17592c653d0eSAndrea Arcangeli stable_node_dup = stable_node_any; 1760b4fecc67SAndrea Arcangeli /* 1761b4fecc67SAndrea Arcangeli * If stable_node was a chain and chain_prune collapsed it, 17620ba1d0f7SAndrea Arcangeli * stable_node has been updated to be the new regular 17630ba1d0f7SAndrea Arcangeli * stable_node. A collapse of the chain is indistinguishable 17640ba1d0f7SAndrea Arcangeli * from the case there was no chain in the stable 17650ba1d0f7SAndrea Arcangeli * rbtree. Otherwise stable_node is the chain and 17660ba1d0f7SAndrea Arcangeli * stable_node_dup is the dup to replace. 1767b4fecc67SAndrea Arcangeli */ 17680ba1d0f7SAndrea Arcangeli if (stable_node_dup == stable_node) { 1769b4fecc67SAndrea Arcangeli VM_BUG_ON(is_stable_node_chain(stable_node_dup)); 1770b4fecc67SAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node_dup)); 17712c653d0eSAndrea Arcangeli /* chain is missing so create it */ 17722c653d0eSAndrea Arcangeli stable_node = alloc_stable_node_chain(stable_node_dup, 17732c653d0eSAndrea Arcangeli root); 17742c653d0eSAndrea Arcangeli if (!stable_node) 17752c653d0eSAndrea Arcangeli return NULL; 17762c653d0eSAndrea Arcangeli } 17772c653d0eSAndrea Arcangeli /* 17782c653d0eSAndrea Arcangeli * Add this stable_node dup that was 17792c653d0eSAndrea Arcangeli * migrated to the stable_node chain 17802c653d0eSAndrea Arcangeli * of the current nid for this page 17812c653d0eSAndrea Arcangeli * content. 17822c653d0eSAndrea Arcangeli */ 1783b4fecc67SAndrea Arcangeli VM_BUG_ON(!is_stable_node_chain(stable_node)); 1784b4fecc67SAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(stable_node_dup)); 17852c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 17862c653d0eSAndrea Arcangeli list_del(&page_node->list); 17872c653d0eSAndrea Arcangeli DO_NUMA(page_node->nid = nid); 17882c653d0eSAndrea Arcangeli stable_node_chain_add_dup(page_node, stable_node); 17892c653d0eSAndrea Arcangeli goto out; 179031dbd01fSIzik Eidus } 179131dbd01fSIzik Eidus 179231dbd01fSIzik Eidus /* 1793e850dcf5SHugh Dickins * stable_tree_insert - insert stable tree node pointing to new ksm page 179431dbd01fSIzik Eidus * into the stable tree. 179531dbd01fSIzik Eidus * 17967b6ba2c7SHugh Dickins * This function returns the stable tree node just allocated on success, 17977b6ba2c7SHugh Dickins * NULL otherwise. 179831dbd01fSIzik Eidus */ 17997b6ba2c7SHugh Dickins static struct stable_node *stable_tree_insert(struct page *kpage) 180031dbd01fSIzik Eidus { 180190bd6fd3SPetr Holasek int nid; 180290bd6fd3SPetr Holasek unsigned long kpfn; 1803ef53d16cSHugh Dickins struct rb_root *root; 180490bd6fd3SPetr Holasek struct rb_node **new; 1805f2e5ff85SAndrea Arcangeli struct rb_node *parent; 18062c653d0eSAndrea Arcangeli struct stable_node *stable_node, *stable_node_dup, *stable_node_any; 18072c653d0eSAndrea Arcangeli bool need_chain = false; 180831dbd01fSIzik Eidus 180990bd6fd3SPetr Holasek kpfn = page_to_pfn(kpage); 181090bd6fd3SPetr Holasek nid = get_kpfn_nid(kpfn); 1811ef53d16cSHugh Dickins root = root_stable_tree + nid; 1812f2e5ff85SAndrea Arcangeli again: 1813f2e5ff85SAndrea Arcangeli parent = NULL; 1814ef53d16cSHugh Dickins new = &root->rb_node; 181590bd6fd3SPetr Holasek 181631dbd01fSIzik Eidus while (*new) { 18174035c07aSHugh Dickins struct page *tree_page; 181831dbd01fSIzik Eidus int ret; 181931dbd01fSIzik Eidus 182031dbd01fSIzik Eidus cond_resched(); 182108beca44SHugh Dickins stable_node = rb_entry(*new, struct stable_node, node); 18222c653d0eSAndrea Arcangeli stable_node_any = NULL; 18238dc5ffcdSAndrea Arcangeli tree_page = chain(&stable_node_dup, stable_node, root); 18242c653d0eSAndrea Arcangeli if (!stable_node_dup) { 18252c653d0eSAndrea Arcangeli /* 18262c653d0eSAndrea Arcangeli * Either all stable_node dups were full in 18272c653d0eSAndrea Arcangeli * this stable_node chain, or this chain was 18282c653d0eSAndrea Arcangeli * empty and should be rb_erased. 18292c653d0eSAndrea Arcangeli */ 18302c653d0eSAndrea Arcangeli stable_node_any = stable_node_dup_any(stable_node, 18312c653d0eSAndrea Arcangeli root); 18322c653d0eSAndrea Arcangeli if (!stable_node_any) { 18332c653d0eSAndrea Arcangeli /* rb_erase just run */ 18342c653d0eSAndrea Arcangeli goto again; 18352c653d0eSAndrea Arcangeli } 18362c653d0eSAndrea Arcangeli /* 18372c653d0eSAndrea Arcangeli * Take any of the stable_node dups page of 18382c653d0eSAndrea Arcangeli * this stable_node chain to let the tree walk 18392c653d0eSAndrea Arcangeli * continue. All KSM pages belonging to the 18402c653d0eSAndrea Arcangeli * stable_node dups in a stable_node chain 18412c653d0eSAndrea Arcangeli * have the same content and they're 18422c653d0eSAndrea Arcangeli * wrprotected at all times. Any will work 18432c653d0eSAndrea Arcangeli * fine to continue the walk. 18442c653d0eSAndrea Arcangeli */ 18452c653d0eSAndrea Arcangeli tree_page = get_ksm_page(stable_node_any, false); 18462c653d0eSAndrea Arcangeli } 18472c653d0eSAndrea Arcangeli VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); 1848f2e5ff85SAndrea Arcangeli if (!tree_page) { 1849f2e5ff85SAndrea Arcangeli /* 1850f2e5ff85SAndrea Arcangeli * If we walked over a stale stable_node, 1851f2e5ff85SAndrea Arcangeli * get_ksm_page() will call rb_erase() and it 1852f2e5ff85SAndrea Arcangeli * may rebalance the tree from under us. So 1853f2e5ff85SAndrea Arcangeli * restart the search from scratch. Returning 1854f2e5ff85SAndrea Arcangeli * NULL would be safe too, but we'd generate 1855f2e5ff85SAndrea Arcangeli * false negative insertions just because some 1856f2e5ff85SAndrea Arcangeli * stable_node was stale. 1857f2e5ff85SAndrea Arcangeli */ 1858f2e5ff85SAndrea Arcangeli goto again; 1859f2e5ff85SAndrea Arcangeli } 186031dbd01fSIzik Eidus 18614035c07aSHugh Dickins ret = memcmp_pages(kpage, tree_page); 18624035c07aSHugh Dickins put_page(tree_page); 186331dbd01fSIzik Eidus 186431dbd01fSIzik Eidus parent = *new; 186531dbd01fSIzik Eidus if (ret < 0) 186631dbd01fSIzik Eidus new = &parent->rb_left; 186731dbd01fSIzik Eidus else if (ret > 0) 186831dbd01fSIzik Eidus new = &parent->rb_right; 186931dbd01fSIzik Eidus else { 18702c653d0eSAndrea Arcangeli need_chain = true; 18712c653d0eSAndrea Arcangeli break; 187231dbd01fSIzik Eidus } 187331dbd01fSIzik Eidus } 187431dbd01fSIzik Eidus 18752c653d0eSAndrea Arcangeli stable_node_dup = alloc_stable_node(); 18762c653d0eSAndrea Arcangeli if (!stable_node_dup) 18777b6ba2c7SHugh Dickins return NULL; 187831dbd01fSIzik Eidus 18792c653d0eSAndrea Arcangeli INIT_HLIST_HEAD(&stable_node_dup->hlist); 18802c653d0eSAndrea Arcangeli stable_node_dup->kpfn = kpfn; 18812c653d0eSAndrea Arcangeli set_page_stable_node(kpage, stable_node_dup); 18822c653d0eSAndrea Arcangeli stable_node_dup->rmap_hlist_len = 0; 18832c653d0eSAndrea Arcangeli DO_NUMA(stable_node_dup->nid = nid); 18842c653d0eSAndrea Arcangeli if (!need_chain) { 18852c653d0eSAndrea Arcangeli rb_link_node(&stable_node_dup->node, parent, new); 18862c653d0eSAndrea Arcangeli rb_insert_color(&stable_node_dup->node, root); 18872c653d0eSAndrea Arcangeli } else { 18882c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 18892c653d0eSAndrea Arcangeli struct stable_node *orig = stable_node; 18902c653d0eSAndrea Arcangeli /* chain is missing so create it */ 18912c653d0eSAndrea Arcangeli stable_node = alloc_stable_node_chain(orig, root); 18922c653d0eSAndrea Arcangeli if (!stable_node) { 18932c653d0eSAndrea Arcangeli free_stable_node(stable_node_dup); 18942c653d0eSAndrea Arcangeli return NULL; 18952c653d0eSAndrea Arcangeli } 18962c653d0eSAndrea Arcangeli } 18972c653d0eSAndrea Arcangeli stable_node_chain_add_dup(stable_node_dup, stable_node); 18982c653d0eSAndrea Arcangeli } 189908beca44SHugh Dickins 19002c653d0eSAndrea Arcangeli return stable_node_dup; 190131dbd01fSIzik Eidus } 190231dbd01fSIzik Eidus 190331dbd01fSIzik Eidus /* 19048dd3557aSHugh Dickins * unstable_tree_search_insert - search for identical page, 19058dd3557aSHugh Dickins * else insert rmap_item into the unstable tree. 190631dbd01fSIzik Eidus * 190731dbd01fSIzik Eidus * This function searches for a page in the unstable tree identical to the 190831dbd01fSIzik Eidus * page currently being scanned; and if no identical page is found in the 190931dbd01fSIzik Eidus * tree, we insert rmap_item as a new object into the unstable tree. 191031dbd01fSIzik Eidus * 191131dbd01fSIzik Eidus * This function returns pointer to rmap_item found to be identical 191231dbd01fSIzik Eidus * to the currently scanned page, NULL otherwise. 191331dbd01fSIzik Eidus * 191431dbd01fSIzik Eidus * This function does both searching and inserting, because they share 191531dbd01fSIzik Eidus * the same walking algorithm in an rbtree. 191631dbd01fSIzik Eidus */ 19178dd3557aSHugh Dickins static 19188dd3557aSHugh Dickins struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item, 19198dd3557aSHugh Dickins struct page *page, 19208dd3557aSHugh Dickins struct page **tree_pagep) 192131dbd01fSIzik Eidus { 192290bd6fd3SPetr Holasek struct rb_node **new; 192390bd6fd3SPetr Holasek struct rb_root *root; 192431dbd01fSIzik Eidus struct rb_node *parent = NULL; 192590bd6fd3SPetr Holasek int nid; 192690bd6fd3SPetr Holasek 192790bd6fd3SPetr Holasek nid = get_kpfn_nid(page_to_pfn(page)); 1928ef53d16cSHugh Dickins root = root_unstable_tree + nid; 192990bd6fd3SPetr Holasek new = &root->rb_node; 193031dbd01fSIzik Eidus 193131dbd01fSIzik Eidus while (*new) { 193231dbd01fSIzik Eidus struct rmap_item *tree_rmap_item; 19338dd3557aSHugh Dickins struct page *tree_page; 193431dbd01fSIzik Eidus int ret; 193531dbd01fSIzik Eidus 1936d178f27fSHugh Dickins cond_resched(); 193731dbd01fSIzik Eidus tree_rmap_item = rb_entry(*new, struct rmap_item, node); 19388dd3557aSHugh Dickins tree_page = get_mergeable_page(tree_rmap_item); 1939c8f95ed1SAndrea Arcangeli if (!tree_page) 194031dbd01fSIzik Eidus return NULL; 194131dbd01fSIzik Eidus 194231dbd01fSIzik Eidus /* 19438dd3557aSHugh Dickins * Don't substitute a ksm page for a forked page. 194431dbd01fSIzik Eidus */ 19458dd3557aSHugh Dickins if (page == tree_page) { 19468dd3557aSHugh Dickins put_page(tree_page); 194731dbd01fSIzik Eidus return NULL; 194831dbd01fSIzik Eidus } 194931dbd01fSIzik Eidus 19508dd3557aSHugh Dickins ret = memcmp_pages(page, tree_page); 195131dbd01fSIzik Eidus 195231dbd01fSIzik Eidus parent = *new; 195331dbd01fSIzik Eidus if (ret < 0) { 19548dd3557aSHugh Dickins put_page(tree_page); 195531dbd01fSIzik Eidus new = &parent->rb_left; 195631dbd01fSIzik Eidus } else if (ret > 0) { 19578dd3557aSHugh Dickins put_page(tree_page); 195831dbd01fSIzik Eidus new = &parent->rb_right; 1959b599cbdfSHugh Dickins } else if (!ksm_merge_across_nodes && 1960b599cbdfSHugh Dickins page_to_nid(tree_page) != nid) { 1961b599cbdfSHugh Dickins /* 1962b599cbdfSHugh Dickins * If tree_page has been migrated to another NUMA node, 1963b599cbdfSHugh Dickins * it will be flushed out and put in the right unstable 1964b599cbdfSHugh Dickins * tree next time: only merge with it when across_nodes. 1965b599cbdfSHugh Dickins */ 1966b599cbdfSHugh Dickins put_page(tree_page); 1967b599cbdfSHugh Dickins return NULL; 196831dbd01fSIzik Eidus } else { 19698dd3557aSHugh Dickins *tree_pagep = tree_page; 197031dbd01fSIzik Eidus return tree_rmap_item; 197131dbd01fSIzik Eidus } 197231dbd01fSIzik Eidus } 197331dbd01fSIzik Eidus 19747b6ba2c7SHugh Dickins rmap_item->address |= UNSTABLE_FLAG; 197531dbd01fSIzik Eidus rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); 1976e850dcf5SHugh Dickins DO_NUMA(rmap_item->nid = nid); 197731dbd01fSIzik Eidus rb_link_node(&rmap_item->node, parent, new); 197890bd6fd3SPetr Holasek rb_insert_color(&rmap_item->node, root); 197931dbd01fSIzik Eidus 1980473b0ce4SHugh Dickins ksm_pages_unshared++; 198131dbd01fSIzik Eidus return NULL; 198231dbd01fSIzik Eidus } 198331dbd01fSIzik Eidus 198431dbd01fSIzik Eidus /* 198531dbd01fSIzik Eidus * stable_tree_append - add another rmap_item to the linked list of 198631dbd01fSIzik Eidus * rmap_items hanging off a given node of the stable tree, all sharing 198731dbd01fSIzik Eidus * the same ksm page. 198831dbd01fSIzik Eidus */ 198931dbd01fSIzik Eidus static void stable_tree_append(struct rmap_item *rmap_item, 19902c653d0eSAndrea Arcangeli struct stable_node *stable_node, 19912c653d0eSAndrea Arcangeli bool max_page_sharing_bypass) 199231dbd01fSIzik Eidus { 19932c653d0eSAndrea Arcangeli /* 19942c653d0eSAndrea Arcangeli * rmap won't find this mapping if we don't insert the 19952c653d0eSAndrea Arcangeli * rmap_item in the right stable_node 19962c653d0eSAndrea Arcangeli * duplicate. page_migration could break later if rmap breaks, 19972c653d0eSAndrea Arcangeli * so we can as well crash here. We really need to check for 19982c653d0eSAndrea Arcangeli * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check 19992c653d0eSAndrea Arcangeli * for other negative values as an undeflow if detected here 20002c653d0eSAndrea Arcangeli * for the first time (and not when decreasing rmap_hlist_len) 20012c653d0eSAndrea Arcangeli * would be sign of memory corruption in the stable_node. 20022c653d0eSAndrea Arcangeli */ 20032c653d0eSAndrea Arcangeli BUG_ON(stable_node->rmap_hlist_len < 0); 20042c653d0eSAndrea Arcangeli 20052c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len++; 20062c653d0eSAndrea Arcangeli if (!max_page_sharing_bypass) 20072c653d0eSAndrea Arcangeli /* possibly non fatal but unexpected overflow, only warn */ 20082c653d0eSAndrea Arcangeli WARN_ON_ONCE(stable_node->rmap_hlist_len > 20092c653d0eSAndrea Arcangeli ksm_max_page_sharing); 20102c653d0eSAndrea Arcangeli 20117b6ba2c7SHugh Dickins rmap_item->head = stable_node; 201231dbd01fSIzik Eidus rmap_item->address |= STABLE_FLAG; 20137b6ba2c7SHugh Dickins hlist_add_head(&rmap_item->hlist, &stable_node->hlist); 2014e178dfdeSHugh Dickins 20157b6ba2c7SHugh Dickins if (rmap_item->hlist.next) 2016e178dfdeSHugh Dickins ksm_pages_sharing++; 20177b6ba2c7SHugh Dickins else 20187b6ba2c7SHugh Dickins ksm_pages_shared++; 201931dbd01fSIzik Eidus } 202031dbd01fSIzik Eidus 202131dbd01fSIzik Eidus /* 202281464e30SHugh Dickins * cmp_and_merge_page - first see if page can be merged into the stable tree; 202381464e30SHugh Dickins * if not, compare checksum to previous and if it's the same, see if page can 202481464e30SHugh Dickins * be inserted into the unstable tree, or merged with a page already there and 202581464e30SHugh Dickins * both transferred to the stable tree. 202631dbd01fSIzik Eidus * 202731dbd01fSIzik Eidus * @page: the page that we are searching identical page to. 202831dbd01fSIzik Eidus * @rmap_item: the reverse mapping into the virtual address of this page 202931dbd01fSIzik Eidus */ 203031dbd01fSIzik Eidus static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item) 203131dbd01fSIzik Eidus { 20324b22927fSKirill Tkhai struct mm_struct *mm = rmap_item->mm; 203331dbd01fSIzik Eidus struct rmap_item *tree_rmap_item; 20348dd3557aSHugh Dickins struct page *tree_page = NULL; 20357b6ba2c7SHugh Dickins struct stable_node *stable_node; 20368dd3557aSHugh Dickins struct page *kpage; 203731dbd01fSIzik Eidus unsigned int checksum; 203831dbd01fSIzik Eidus int err; 20392c653d0eSAndrea Arcangeli bool max_page_sharing_bypass = false; 204031dbd01fSIzik Eidus 20414146d2d6SHugh Dickins stable_node = page_stable_node(page); 20424146d2d6SHugh Dickins if (stable_node) { 20434146d2d6SHugh Dickins if (stable_node->head != &migrate_nodes && 20442c653d0eSAndrea Arcangeli get_kpfn_nid(READ_ONCE(stable_node->kpfn)) != 20452c653d0eSAndrea Arcangeli NUMA(stable_node->nid)) { 20462c653d0eSAndrea Arcangeli stable_node_dup_del(stable_node); 20474146d2d6SHugh Dickins stable_node->head = &migrate_nodes; 20484146d2d6SHugh Dickins list_add(&stable_node->list, stable_node->head); 20494146d2d6SHugh Dickins } 20504146d2d6SHugh Dickins if (stable_node->head != &migrate_nodes && 20514146d2d6SHugh Dickins rmap_item->head == stable_node) 20524146d2d6SHugh Dickins return; 20532c653d0eSAndrea Arcangeli /* 20542c653d0eSAndrea Arcangeli * If it's a KSM fork, allow it to go over the sharing limit 20552c653d0eSAndrea Arcangeli * without warnings. 20562c653d0eSAndrea Arcangeli */ 20572c653d0eSAndrea Arcangeli if (!is_page_sharing_candidate(stable_node)) 20582c653d0eSAndrea Arcangeli max_page_sharing_bypass = true; 20594146d2d6SHugh Dickins } 206031dbd01fSIzik Eidus 206131dbd01fSIzik Eidus /* We first start with searching the page inside the stable tree */ 206262b61f61SHugh Dickins kpage = stable_tree_search(page); 20634146d2d6SHugh Dickins if (kpage == page && rmap_item->head == stable_node) { 20644146d2d6SHugh Dickins put_page(kpage); 20654146d2d6SHugh Dickins return; 20664146d2d6SHugh Dickins } 20674146d2d6SHugh Dickins 20684146d2d6SHugh Dickins remove_rmap_item_from_tree(rmap_item); 20694146d2d6SHugh Dickins 207062b61f61SHugh Dickins if (kpage) { 207108beca44SHugh Dickins err = try_to_merge_with_ksm_page(rmap_item, page, kpage); 207231dbd01fSIzik Eidus if (!err) { 207331dbd01fSIzik Eidus /* 207431dbd01fSIzik Eidus * The page was successfully merged: 207531dbd01fSIzik Eidus * add its rmap_item to the stable tree. 207631dbd01fSIzik Eidus */ 20775ad64688SHugh Dickins lock_page(kpage); 20782c653d0eSAndrea Arcangeli stable_tree_append(rmap_item, page_stable_node(kpage), 20792c653d0eSAndrea Arcangeli max_page_sharing_bypass); 20805ad64688SHugh Dickins unlock_page(kpage); 208131dbd01fSIzik Eidus } 20828dd3557aSHugh Dickins put_page(kpage); 208331dbd01fSIzik Eidus return; 208431dbd01fSIzik Eidus } 208531dbd01fSIzik Eidus 208631dbd01fSIzik Eidus /* 20874035c07aSHugh Dickins * If the hash value of the page has changed from the last time 20884035c07aSHugh Dickins * we calculated it, this page is changing frequently: therefore we 20894035c07aSHugh Dickins * don't want to insert it in the unstable tree, and we don't want 20904035c07aSHugh Dickins * to waste our time searching for something identical to it there. 209131dbd01fSIzik Eidus */ 209231dbd01fSIzik Eidus checksum = calc_checksum(page); 209331dbd01fSIzik Eidus if (rmap_item->oldchecksum != checksum) { 209431dbd01fSIzik Eidus rmap_item->oldchecksum = checksum; 209531dbd01fSIzik Eidus return; 209631dbd01fSIzik Eidus } 209731dbd01fSIzik Eidus 2098e86c59b1SClaudio Imbrenda /* 2099e86c59b1SClaudio Imbrenda * Same checksum as an empty page. We attempt to merge it with the 2100e86c59b1SClaudio Imbrenda * appropriate zero page if the user enabled this via sysfs. 2101e86c59b1SClaudio Imbrenda */ 2102e86c59b1SClaudio Imbrenda if (ksm_use_zero_pages && (checksum == zero_checksum)) { 2103e86c59b1SClaudio Imbrenda struct vm_area_struct *vma; 2104e86c59b1SClaudio Imbrenda 21054b22927fSKirill Tkhai down_read(&mm->mmap_sem); 21064b22927fSKirill Tkhai vma = find_mergeable_vma(mm, rmap_item->address); 2107e86c59b1SClaudio Imbrenda err = try_to_merge_one_page(vma, page, 2108e86c59b1SClaudio Imbrenda ZERO_PAGE(rmap_item->address)); 21094b22927fSKirill Tkhai up_read(&mm->mmap_sem); 2110e86c59b1SClaudio Imbrenda /* 2111e86c59b1SClaudio Imbrenda * In case of failure, the page was not really empty, so we 2112e86c59b1SClaudio Imbrenda * need to continue. Otherwise we're done. 2113e86c59b1SClaudio Imbrenda */ 2114e86c59b1SClaudio Imbrenda if (!err) 2115e86c59b1SClaudio Imbrenda return; 2116e86c59b1SClaudio Imbrenda } 21178dd3557aSHugh Dickins tree_rmap_item = 21188dd3557aSHugh Dickins unstable_tree_search_insert(rmap_item, page, &tree_page); 211931dbd01fSIzik Eidus if (tree_rmap_item) { 212077da2ba0SClaudio Imbrenda bool split; 212177da2ba0SClaudio Imbrenda 21228dd3557aSHugh Dickins kpage = try_to_merge_two_pages(rmap_item, page, 21238dd3557aSHugh Dickins tree_rmap_item, tree_page); 212477da2ba0SClaudio Imbrenda /* 212577da2ba0SClaudio Imbrenda * If both pages we tried to merge belong to the same compound 212677da2ba0SClaudio Imbrenda * page, then we actually ended up increasing the reference 212777da2ba0SClaudio Imbrenda * count of the same compound page twice, and split_huge_page 212877da2ba0SClaudio Imbrenda * failed. 212977da2ba0SClaudio Imbrenda * Here we set a flag if that happened, and we use it later to 213077da2ba0SClaudio Imbrenda * try split_huge_page again. Since we call put_page right 213177da2ba0SClaudio Imbrenda * afterwards, the reference count will be correct and 213277da2ba0SClaudio Imbrenda * split_huge_page should succeed. 213377da2ba0SClaudio Imbrenda */ 213477da2ba0SClaudio Imbrenda split = PageTransCompound(page) 213577da2ba0SClaudio Imbrenda && compound_head(page) == compound_head(tree_page); 21368dd3557aSHugh Dickins put_page(tree_page); 21378dd3557aSHugh Dickins if (kpage) { 2138bc56620bSHugh Dickins /* 2139bc56620bSHugh Dickins * The pages were successfully merged: insert new 2140bc56620bSHugh Dickins * node in the stable tree and add both rmap_items. 2141bc56620bSHugh Dickins */ 21425ad64688SHugh Dickins lock_page(kpage); 21437b6ba2c7SHugh Dickins stable_node = stable_tree_insert(kpage); 21447b6ba2c7SHugh Dickins if (stable_node) { 21452c653d0eSAndrea Arcangeli stable_tree_append(tree_rmap_item, stable_node, 21462c653d0eSAndrea Arcangeli false); 21472c653d0eSAndrea Arcangeli stable_tree_append(rmap_item, stable_node, 21482c653d0eSAndrea Arcangeli false); 21497b6ba2c7SHugh Dickins } 21505ad64688SHugh Dickins unlock_page(kpage); 21517b6ba2c7SHugh Dickins 215231dbd01fSIzik Eidus /* 215331dbd01fSIzik Eidus * If we fail to insert the page into the stable tree, 215431dbd01fSIzik Eidus * we will have 2 virtual addresses that are pointing 215531dbd01fSIzik Eidus * to a ksm page left outside the stable tree, 215631dbd01fSIzik Eidus * in which case we need to break_cow on both. 215731dbd01fSIzik Eidus */ 21587b6ba2c7SHugh Dickins if (!stable_node) { 21598dd3557aSHugh Dickins break_cow(tree_rmap_item); 21608dd3557aSHugh Dickins break_cow(rmap_item); 216131dbd01fSIzik Eidus } 216277da2ba0SClaudio Imbrenda } else if (split) { 216377da2ba0SClaudio Imbrenda /* 216477da2ba0SClaudio Imbrenda * We are here if we tried to merge two pages and 216577da2ba0SClaudio Imbrenda * failed because they both belonged to the same 216677da2ba0SClaudio Imbrenda * compound page. We will split the page now, but no 216777da2ba0SClaudio Imbrenda * merging will take place. 216877da2ba0SClaudio Imbrenda * We do not want to add the cost of a full lock; if 216977da2ba0SClaudio Imbrenda * the page is locked, it is better to skip it and 217077da2ba0SClaudio Imbrenda * perhaps try again later. 217177da2ba0SClaudio Imbrenda */ 217277da2ba0SClaudio Imbrenda if (!trylock_page(page)) 217377da2ba0SClaudio Imbrenda return; 217477da2ba0SClaudio Imbrenda split_huge_page(page); 217577da2ba0SClaudio Imbrenda unlock_page(page); 217631dbd01fSIzik Eidus } 217731dbd01fSIzik Eidus } 217831dbd01fSIzik Eidus } 217931dbd01fSIzik Eidus 218031dbd01fSIzik Eidus static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot, 21816514d511SHugh Dickins struct rmap_item **rmap_list, 218231dbd01fSIzik Eidus unsigned long addr) 218331dbd01fSIzik Eidus { 218431dbd01fSIzik Eidus struct rmap_item *rmap_item; 218531dbd01fSIzik Eidus 21866514d511SHugh Dickins while (*rmap_list) { 21876514d511SHugh Dickins rmap_item = *rmap_list; 218893d17715SHugh Dickins if ((rmap_item->address & PAGE_MASK) == addr) 218931dbd01fSIzik Eidus return rmap_item; 219031dbd01fSIzik Eidus if (rmap_item->address > addr) 219131dbd01fSIzik Eidus break; 21926514d511SHugh Dickins *rmap_list = rmap_item->rmap_list; 219331dbd01fSIzik Eidus remove_rmap_item_from_tree(rmap_item); 219431dbd01fSIzik Eidus free_rmap_item(rmap_item); 219531dbd01fSIzik Eidus } 219631dbd01fSIzik Eidus 219731dbd01fSIzik Eidus rmap_item = alloc_rmap_item(); 219831dbd01fSIzik Eidus if (rmap_item) { 219931dbd01fSIzik Eidus /* It has already been zeroed */ 220031dbd01fSIzik Eidus rmap_item->mm = mm_slot->mm; 220131dbd01fSIzik Eidus rmap_item->address = addr; 22026514d511SHugh Dickins rmap_item->rmap_list = *rmap_list; 22036514d511SHugh Dickins *rmap_list = rmap_item; 220431dbd01fSIzik Eidus } 220531dbd01fSIzik Eidus return rmap_item; 220631dbd01fSIzik Eidus } 220731dbd01fSIzik Eidus 220831dbd01fSIzik Eidus static struct rmap_item *scan_get_next_rmap_item(struct page **page) 220931dbd01fSIzik Eidus { 221031dbd01fSIzik Eidus struct mm_struct *mm; 221131dbd01fSIzik Eidus struct mm_slot *slot; 221231dbd01fSIzik Eidus struct vm_area_struct *vma; 221331dbd01fSIzik Eidus struct rmap_item *rmap_item; 221490bd6fd3SPetr Holasek int nid; 221531dbd01fSIzik Eidus 221631dbd01fSIzik Eidus if (list_empty(&ksm_mm_head.mm_list)) 221731dbd01fSIzik Eidus return NULL; 221831dbd01fSIzik Eidus 221931dbd01fSIzik Eidus slot = ksm_scan.mm_slot; 222031dbd01fSIzik Eidus if (slot == &ksm_mm_head) { 22212919bfd0SHugh Dickins /* 22222919bfd0SHugh Dickins * A number of pages can hang around indefinitely on per-cpu 22232919bfd0SHugh Dickins * pagevecs, raised page count preventing write_protect_page 22242919bfd0SHugh Dickins * from merging them. Though it doesn't really matter much, 22252919bfd0SHugh Dickins * it is puzzling to see some stuck in pages_volatile until 22262919bfd0SHugh Dickins * other activity jostles them out, and they also prevented 22272919bfd0SHugh Dickins * LTP's KSM test from succeeding deterministically; so drain 22282919bfd0SHugh Dickins * them here (here rather than on entry to ksm_do_scan(), 22292919bfd0SHugh Dickins * so we don't IPI too often when pages_to_scan is set low). 22302919bfd0SHugh Dickins */ 22312919bfd0SHugh Dickins lru_add_drain_all(); 22322919bfd0SHugh Dickins 22334146d2d6SHugh Dickins /* 22344146d2d6SHugh Dickins * Whereas stale stable_nodes on the stable_tree itself 22354146d2d6SHugh Dickins * get pruned in the regular course of stable_tree_search(), 22364146d2d6SHugh Dickins * those moved out to the migrate_nodes list can accumulate: 22374146d2d6SHugh Dickins * so prune them once before each full scan. 22384146d2d6SHugh Dickins */ 22394146d2d6SHugh Dickins if (!ksm_merge_across_nodes) { 224003640418SGeliang Tang struct stable_node *stable_node, *next; 22414146d2d6SHugh Dickins struct page *page; 22424146d2d6SHugh Dickins 224303640418SGeliang Tang list_for_each_entry_safe(stable_node, next, 224403640418SGeliang Tang &migrate_nodes, list) { 22454146d2d6SHugh Dickins page = get_ksm_page(stable_node, false); 22464146d2d6SHugh Dickins if (page) 22474146d2d6SHugh Dickins put_page(page); 22484146d2d6SHugh Dickins cond_resched(); 22494146d2d6SHugh Dickins } 22504146d2d6SHugh Dickins } 22514146d2d6SHugh Dickins 2252ef53d16cSHugh Dickins for (nid = 0; nid < ksm_nr_node_ids; nid++) 225390bd6fd3SPetr Holasek root_unstable_tree[nid] = RB_ROOT; 225431dbd01fSIzik Eidus 225531dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 225631dbd01fSIzik Eidus slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list); 225731dbd01fSIzik Eidus ksm_scan.mm_slot = slot; 225831dbd01fSIzik Eidus spin_unlock(&ksm_mmlist_lock); 22592b472611SHugh Dickins /* 22602b472611SHugh Dickins * Although we tested list_empty() above, a racing __ksm_exit 22612b472611SHugh Dickins * of the last mm on the list may have removed it since then. 22622b472611SHugh Dickins */ 22632b472611SHugh Dickins if (slot == &ksm_mm_head) 22642b472611SHugh Dickins return NULL; 226531dbd01fSIzik Eidus next_mm: 226631dbd01fSIzik Eidus ksm_scan.address = 0; 22676514d511SHugh Dickins ksm_scan.rmap_list = &slot->rmap_list; 226831dbd01fSIzik Eidus } 226931dbd01fSIzik Eidus 227031dbd01fSIzik Eidus mm = slot->mm; 227131dbd01fSIzik Eidus down_read(&mm->mmap_sem); 22729ba69294SHugh Dickins if (ksm_test_exit(mm)) 22739ba69294SHugh Dickins vma = NULL; 22749ba69294SHugh Dickins else 22759ba69294SHugh Dickins vma = find_vma(mm, ksm_scan.address); 22769ba69294SHugh Dickins 22779ba69294SHugh Dickins for (; vma; vma = vma->vm_next) { 227831dbd01fSIzik Eidus if (!(vma->vm_flags & VM_MERGEABLE)) 227931dbd01fSIzik Eidus continue; 228031dbd01fSIzik Eidus if (ksm_scan.address < vma->vm_start) 228131dbd01fSIzik Eidus ksm_scan.address = vma->vm_start; 228231dbd01fSIzik Eidus if (!vma->anon_vma) 228331dbd01fSIzik Eidus ksm_scan.address = vma->vm_end; 228431dbd01fSIzik Eidus 228531dbd01fSIzik Eidus while (ksm_scan.address < vma->vm_end) { 22869ba69294SHugh Dickins if (ksm_test_exit(mm)) 22879ba69294SHugh Dickins break; 228831dbd01fSIzik Eidus *page = follow_page(vma, ksm_scan.address, FOLL_GET); 228921ae5b01SAndrea Arcangeli if (IS_ERR_OR_NULL(*page)) { 229021ae5b01SAndrea Arcangeli ksm_scan.address += PAGE_SIZE; 229121ae5b01SAndrea Arcangeli cond_resched(); 229221ae5b01SAndrea Arcangeli continue; 229321ae5b01SAndrea Arcangeli } 2294f765f540SKirill A. Shutemov if (PageAnon(*page)) { 229531dbd01fSIzik Eidus flush_anon_page(vma, *page, ksm_scan.address); 229631dbd01fSIzik Eidus flush_dcache_page(*page); 229731dbd01fSIzik Eidus rmap_item = get_next_rmap_item(slot, 22986514d511SHugh Dickins ksm_scan.rmap_list, ksm_scan.address); 229931dbd01fSIzik Eidus if (rmap_item) { 23006514d511SHugh Dickins ksm_scan.rmap_list = 23016514d511SHugh Dickins &rmap_item->rmap_list; 230231dbd01fSIzik Eidus ksm_scan.address += PAGE_SIZE; 230331dbd01fSIzik Eidus } else 230431dbd01fSIzik Eidus put_page(*page); 230531dbd01fSIzik Eidus up_read(&mm->mmap_sem); 230631dbd01fSIzik Eidus return rmap_item; 230731dbd01fSIzik Eidus } 230831dbd01fSIzik Eidus put_page(*page); 230931dbd01fSIzik Eidus ksm_scan.address += PAGE_SIZE; 231031dbd01fSIzik Eidus cond_resched(); 231131dbd01fSIzik Eidus } 231231dbd01fSIzik Eidus } 231331dbd01fSIzik Eidus 23149ba69294SHugh Dickins if (ksm_test_exit(mm)) { 23159ba69294SHugh Dickins ksm_scan.address = 0; 23166514d511SHugh Dickins ksm_scan.rmap_list = &slot->rmap_list; 23179ba69294SHugh Dickins } 231831dbd01fSIzik Eidus /* 231931dbd01fSIzik Eidus * Nuke all the rmap_items that are above this current rmap: 232031dbd01fSIzik Eidus * because there were no VM_MERGEABLE vmas with such addresses. 232131dbd01fSIzik Eidus */ 23226514d511SHugh Dickins remove_trailing_rmap_items(slot, ksm_scan.rmap_list); 232331dbd01fSIzik Eidus 232431dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 2325cd551f97SHugh Dickins ksm_scan.mm_slot = list_entry(slot->mm_list.next, 2326cd551f97SHugh Dickins struct mm_slot, mm_list); 2327cd551f97SHugh Dickins if (ksm_scan.address == 0) { 2328cd551f97SHugh Dickins /* 2329cd551f97SHugh Dickins * We've completed a full scan of all vmas, holding mmap_sem 2330cd551f97SHugh Dickins * throughout, and found no VM_MERGEABLE: so do the same as 2331cd551f97SHugh Dickins * __ksm_exit does to remove this mm from all our lists now. 23329ba69294SHugh Dickins * This applies either when cleaning up after __ksm_exit 23339ba69294SHugh Dickins * (but beware: we can reach here even before __ksm_exit), 23349ba69294SHugh Dickins * or when all VM_MERGEABLE areas have been unmapped (and 23359ba69294SHugh Dickins * mmap_sem then protects against race with MADV_MERGEABLE). 2336cd551f97SHugh Dickins */ 23374ca3a69bSSasha Levin hash_del(&slot->link); 2338cd551f97SHugh Dickins list_del(&slot->mm_list); 23399ba69294SHugh Dickins spin_unlock(&ksm_mmlist_lock); 23409ba69294SHugh Dickins 2341cd551f97SHugh Dickins free_mm_slot(slot); 2342cd551f97SHugh Dickins clear_bit(MMF_VM_MERGEABLE, &mm->flags); 23439ba69294SHugh Dickins up_read(&mm->mmap_sem); 23449ba69294SHugh Dickins mmdrop(mm); 23459ba69294SHugh Dickins } else { 2346cd551f97SHugh Dickins up_read(&mm->mmap_sem); 23477496fea9SZhou Chengming /* 23487496fea9SZhou Chengming * up_read(&mm->mmap_sem) first because after 23497496fea9SZhou Chengming * spin_unlock(&ksm_mmlist_lock) run, the "mm" may 23507496fea9SZhou Chengming * already have been freed under us by __ksm_exit() 23517496fea9SZhou Chengming * because the "mm_slot" is still hashed and 23527496fea9SZhou Chengming * ksm_scan.mm_slot doesn't point to it anymore. 23537496fea9SZhou Chengming */ 23547496fea9SZhou Chengming spin_unlock(&ksm_mmlist_lock); 23559ba69294SHugh Dickins } 235631dbd01fSIzik Eidus 235731dbd01fSIzik Eidus /* Repeat until we've completed scanning the whole list */ 2358cd551f97SHugh Dickins slot = ksm_scan.mm_slot; 235931dbd01fSIzik Eidus if (slot != &ksm_mm_head) 236031dbd01fSIzik Eidus goto next_mm; 236131dbd01fSIzik Eidus 236231dbd01fSIzik Eidus ksm_scan.seqnr++; 236331dbd01fSIzik Eidus return NULL; 236431dbd01fSIzik Eidus } 236531dbd01fSIzik Eidus 236631dbd01fSIzik Eidus /** 236731dbd01fSIzik Eidus * ksm_do_scan - the ksm scanner main worker function. 2368b7701a5fSMike Rapoport * @scan_npages: number of pages we want to scan before we return. 236931dbd01fSIzik Eidus */ 237031dbd01fSIzik Eidus static void ksm_do_scan(unsigned int scan_npages) 237131dbd01fSIzik Eidus { 237231dbd01fSIzik Eidus struct rmap_item *rmap_item; 237322eccdd7SDan Carpenter struct page *uninitialized_var(page); 237431dbd01fSIzik Eidus 2375878aee7dSAndrea Arcangeli while (scan_npages-- && likely(!freezing(current))) { 237631dbd01fSIzik Eidus cond_resched(); 237731dbd01fSIzik Eidus rmap_item = scan_get_next_rmap_item(&page); 237831dbd01fSIzik Eidus if (!rmap_item) 237931dbd01fSIzik Eidus return; 238031dbd01fSIzik Eidus cmp_and_merge_page(page, rmap_item); 238131dbd01fSIzik Eidus put_page(page); 238231dbd01fSIzik Eidus } 238331dbd01fSIzik Eidus } 238431dbd01fSIzik Eidus 23856e158384SHugh Dickins static int ksmd_should_run(void) 23866e158384SHugh Dickins { 23876e158384SHugh Dickins return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list); 23886e158384SHugh Dickins } 23896e158384SHugh Dickins 239031dbd01fSIzik Eidus static int ksm_scan_thread(void *nothing) 239131dbd01fSIzik Eidus { 2392878aee7dSAndrea Arcangeli set_freezable(); 2393339aa624SIzik Eidus set_user_nice(current, 5); 239431dbd01fSIzik Eidus 239531dbd01fSIzik Eidus while (!kthread_should_stop()) { 239631dbd01fSIzik Eidus mutex_lock(&ksm_thread_mutex); 2397ef4d43a8SHugh Dickins wait_while_offlining(); 23986e158384SHugh Dickins if (ksmd_should_run()) 239931dbd01fSIzik Eidus ksm_do_scan(ksm_thread_pages_to_scan); 240031dbd01fSIzik Eidus mutex_unlock(&ksm_thread_mutex); 24016e158384SHugh Dickins 2402878aee7dSAndrea Arcangeli try_to_freeze(); 2403878aee7dSAndrea Arcangeli 24046e158384SHugh Dickins if (ksmd_should_run()) { 240531dbd01fSIzik Eidus schedule_timeout_interruptible( 240631dbd01fSIzik Eidus msecs_to_jiffies(ksm_thread_sleep_millisecs)); 240731dbd01fSIzik Eidus } else { 2408878aee7dSAndrea Arcangeli wait_event_freezable(ksm_thread_wait, 24096e158384SHugh Dickins ksmd_should_run() || kthread_should_stop()); 241031dbd01fSIzik Eidus } 241131dbd01fSIzik Eidus } 241231dbd01fSIzik Eidus return 0; 241331dbd01fSIzik Eidus } 241431dbd01fSIzik Eidus 2415f8af4da3SHugh Dickins int ksm_madvise(struct vm_area_struct *vma, unsigned long start, 2416f8af4da3SHugh Dickins unsigned long end, int advice, unsigned long *vm_flags) 2417f8af4da3SHugh Dickins { 2418f8af4da3SHugh Dickins struct mm_struct *mm = vma->vm_mm; 2419d952b791SHugh Dickins int err; 2420f8af4da3SHugh Dickins 2421f8af4da3SHugh Dickins switch (advice) { 2422f8af4da3SHugh Dickins case MADV_MERGEABLE: 2423f8af4da3SHugh Dickins /* 2424f8af4da3SHugh Dickins * Be somewhat over-protective for now! 2425f8af4da3SHugh Dickins */ 2426f8af4da3SHugh Dickins if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE | 2427f8af4da3SHugh Dickins VM_PFNMAP | VM_IO | VM_DONTEXPAND | 24280661a336SKirill A. Shutemov VM_HUGETLB | VM_MIXEDMAP)) 2429f8af4da3SHugh Dickins return 0; /* just ignore the advice */ 2430f8af4da3SHugh Dickins 2431cc2383ecSKonstantin Khlebnikov #ifdef VM_SAO 2432cc2383ecSKonstantin Khlebnikov if (*vm_flags & VM_SAO) 2433cc2383ecSKonstantin Khlebnikov return 0; 2434cc2383ecSKonstantin Khlebnikov #endif 243574a04967SKhalid Aziz #ifdef VM_SPARC_ADI 243674a04967SKhalid Aziz if (*vm_flags & VM_SPARC_ADI) 243774a04967SKhalid Aziz return 0; 243874a04967SKhalid Aziz #endif 2439cc2383ecSKonstantin Khlebnikov 2440d952b791SHugh Dickins if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { 2441d952b791SHugh Dickins err = __ksm_enter(mm); 2442d952b791SHugh Dickins if (err) 2443d952b791SHugh Dickins return err; 2444d952b791SHugh Dickins } 2445f8af4da3SHugh Dickins 2446f8af4da3SHugh Dickins *vm_flags |= VM_MERGEABLE; 2447f8af4da3SHugh Dickins break; 2448f8af4da3SHugh Dickins 2449f8af4da3SHugh Dickins case MADV_UNMERGEABLE: 2450f8af4da3SHugh Dickins if (!(*vm_flags & VM_MERGEABLE)) 2451f8af4da3SHugh Dickins return 0; /* just ignore the advice */ 2452f8af4da3SHugh Dickins 2453d952b791SHugh Dickins if (vma->anon_vma) { 2454d952b791SHugh Dickins err = unmerge_ksm_pages(vma, start, end); 2455d952b791SHugh Dickins if (err) 2456d952b791SHugh Dickins return err; 2457d952b791SHugh Dickins } 2458f8af4da3SHugh Dickins 2459f8af4da3SHugh Dickins *vm_flags &= ~VM_MERGEABLE; 2460f8af4da3SHugh Dickins break; 2461f8af4da3SHugh Dickins } 2462f8af4da3SHugh Dickins 2463f8af4da3SHugh Dickins return 0; 2464f8af4da3SHugh Dickins } 2465f8af4da3SHugh Dickins 2466f8af4da3SHugh Dickins int __ksm_enter(struct mm_struct *mm) 2467f8af4da3SHugh Dickins { 24686e158384SHugh Dickins struct mm_slot *mm_slot; 24696e158384SHugh Dickins int needs_wakeup; 24706e158384SHugh Dickins 24716e158384SHugh Dickins mm_slot = alloc_mm_slot(); 247231dbd01fSIzik Eidus if (!mm_slot) 247331dbd01fSIzik Eidus return -ENOMEM; 247431dbd01fSIzik Eidus 24756e158384SHugh Dickins /* Check ksm_run too? Would need tighter locking */ 24766e158384SHugh Dickins needs_wakeup = list_empty(&ksm_mm_head.mm_list); 24776e158384SHugh Dickins 247831dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 247931dbd01fSIzik Eidus insert_to_mm_slots_hash(mm, mm_slot); 248031dbd01fSIzik Eidus /* 2481cbf86cfeSHugh Dickins * When KSM_RUN_MERGE (or KSM_RUN_STOP), 2482cbf86cfeSHugh Dickins * insert just behind the scanning cursor, to let the area settle 248331dbd01fSIzik Eidus * down a little; when fork is followed by immediate exec, we don't 248431dbd01fSIzik Eidus * want ksmd to waste time setting up and tearing down an rmap_list. 2485cbf86cfeSHugh Dickins * 2486cbf86cfeSHugh Dickins * But when KSM_RUN_UNMERGE, it's important to insert ahead of its 2487cbf86cfeSHugh Dickins * scanning cursor, otherwise KSM pages in newly forked mms will be 2488cbf86cfeSHugh Dickins * missed: then we might as well insert at the end of the list. 248931dbd01fSIzik Eidus */ 2490cbf86cfeSHugh Dickins if (ksm_run & KSM_RUN_UNMERGE) 2491cbf86cfeSHugh Dickins list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list); 2492cbf86cfeSHugh Dickins else 249331dbd01fSIzik Eidus list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list); 249431dbd01fSIzik Eidus spin_unlock(&ksm_mmlist_lock); 249531dbd01fSIzik Eidus 2496f8af4da3SHugh Dickins set_bit(MMF_VM_MERGEABLE, &mm->flags); 2497f1f10076SVegard Nossum mmgrab(mm); 24986e158384SHugh Dickins 24996e158384SHugh Dickins if (needs_wakeup) 25006e158384SHugh Dickins wake_up_interruptible(&ksm_thread_wait); 25016e158384SHugh Dickins 2502f8af4da3SHugh Dickins return 0; 2503f8af4da3SHugh Dickins } 2504f8af4da3SHugh Dickins 25051c2fb7a4SAndrea Arcangeli void __ksm_exit(struct mm_struct *mm) 2506f8af4da3SHugh Dickins { 2507cd551f97SHugh Dickins struct mm_slot *mm_slot; 25089ba69294SHugh Dickins int easy_to_free = 0; 2509cd551f97SHugh Dickins 251031dbd01fSIzik Eidus /* 25119ba69294SHugh Dickins * This process is exiting: if it's straightforward (as is the 25129ba69294SHugh Dickins * case when ksmd was never running), free mm_slot immediately. 25139ba69294SHugh Dickins * But if it's at the cursor or has rmap_items linked to it, use 25149ba69294SHugh Dickins * mmap_sem to synchronize with any break_cows before pagetables 25159ba69294SHugh Dickins * are freed, and leave the mm_slot on the list for ksmd to free. 25169ba69294SHugh Dickins * Beware: ksm may already have noticed it exiting and freed the slot. 251731dbd01fSIzik Eidus */ 25189ba69294SHugh Dickins 2519cd551f97SHugh Dickins spin_lock(&ksm_mmlist_lock); 2520cd551f97SHugh Dickins mm_slot = get_mm_slot(mm); 25219ba69294SHugh Dickins if (mm_slot && ksm_scan.mm_slot != mm_slot) { 25226514d511SHugh Dickins if (!mm_slot->rmap_list) { 25234ca3a69bSSasha Levin hash_del(&mm_slot->link); 2524cd551f97SHugh Dickins list_del(&mm_slot->mm_list); 25259ba69294SHugh Dickins easy_to_free = 1; 25269ba69294SHugh Dickins } else { 25279ba69294SHugh Dickins list_move(&mm_slot->mm_list, 25289ba69294SHugh Dickins &ksm_scan.mm_slot->mm_list); 25299ba69294SHugh Dickins } 25309ba69294SHugh Dickins } 2531cd551f97SHugh Dickins spin_unlock(&ksm_mmlist_lock); 2532cd551f97SHugh Dickins 25339ba69294SHugh Dickins if (easy_to_free) { 2534cd551f97SHugh Dickins free_mm_slot(mm_slot); 2535cd551f97SHugh Dickins clear_bit(MMF_VM_MERGEABLE, &mm->flags); 25369ba69294SHugh Dickins mmdrop(mm); 25379ba69294SHugh Dickins } else if (mm_slot) { 25389ba69294SHugh Dickins down_write(&mm->mmap_sem); 25399ba69294SHugh Dickins up_write(&mm->mmap_sem); 25409ba69294SHugh Dickins } 2541f8af4da3SHugh Dickins } 254231dbd01fSIzik Eidus 2543cbf86cfeSHugh Dickins struct page *ksm_might_need_to_copy(struct page *page, 25445ad64688SHugh Dickins struct vm_area_struct *vma, unsigned long address) 25455ad64688SHugh Dickins { 2546cbf86cfeSHugh Dickins struct anon_vma *anon_vma = page_anon_vma(page); 25475ad64688SHugh Dickins struct page *new_page; 25485ad64688SHugh Dickins 2549cbf86cfeSHugh Dickins if (PageKsm(page)) { 2550cbf86cfeSHugh Dickins if (page_stable_node(page) && 2551cbf86cfeSHugh Dickins !(ksm_run & KSM_RUN_UNMERGE)) 2552cbf86cfeSHugh Dickins return page; /* no need to copy it */ 2553cbf86cfeSHugh Dickins } else if (!anon_vma) { 2554cbf86cfeSHugh Dickins return page; /* no need to copy it */ 2555cbf86cfeSHugh Dickins } else if (anon_vma->root == vma->anon_vma->root && 2556cbf86cfeSHugh Dickins page->index == linear_page_index(vma, address)) { 2557cbf86cfeSHugh Dickins return page; /* still no need to copy it */ 2558cbf86cfeSHugh Dickins } 2559cbf86cfeSHugh Dickins if (!PageUptodate(page)) 2560cbf86cfeSHugh Dickins return page; /* let do_swap_page report the error */ 2561cbf86cfeSHugh Dickins 25625ad64688SHugh Dickins new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); 25635ad64688SHugh Dickins if (new_page) { 25645ad64688SHugh Dickins copy_user_highpage(new_page, page, address, vma); 25655ad64688SHugh Dickins 25665ad64688SHugh Dickins SetPageDirty(new_page); 25675ad64688SHugh Dickins __SetPageUptodate(new_page); 256848c935adSKirill A. Shutemov __SetPageLocked(new_page); 25695ad64688SHugh Dickins } 25705ad64688SHugh Dickins 25715ad64688SHugh Dickins return new_page; 25725ad64688SHugh Dickins } 25735ad64688SHugh Dickins 25741df631aeSMinchan Kim void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc) 2575e9995ef9SHugh Dickins { 2576e9995ef9SHugh Dickins struct stable_node *stable_node; 2577e9995ef9SHugh Dickins struct rmap_item *rmap_item; 2578e9995ef9SHugh Dickins int search_new_forks = 0; 2579e9995ef9SHugh Dickins 2580309381feSSasha Levin VM_BUG_ON_PAGE(!PageKsm(page), page); 25819f32624bSJoonsoo Kim 25829f32624bSJoonsoo Kim /* 25839f32624bSJoonsoo Kim * Rely on the page lock to protect against concurrent modifications 25849f32624bSJoonsoo Kim * to that page's node of the stable tree. 25859f32624bSJoonsoo Kim */ 2586309381feSSasha Levin VM_BUG_ON_PAGE(!PageLocked(page), page); 2587e9995ef9SHugh Dickins 2588e9995ef9SHugh Dickins stable_node = page_stable_node(page); 2589e9995ef9SHugh Dickins if (!stable_node) 25901df631aeSMinchan Kim return; 2591e9995ef9SHugh Dickins again: 2592b67bfe0dSSasha Levin hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { 2593e9995ef9SHugh Dickins struct anon_vma *anon_vma = rmap_item->anon_vma; 25945beb4930SRik van Riel struct anon_vma_chain *vmac; 2595e9995ef9SHugh Dickins struct vm_area_struct *vma; 2596e9995ef9SHugh Dickins 2597ad12695fSAndrea Arcangeli cond_resched(); 2598b6b19f25SHugh Dickins anon_vma_lock_read(anon_vma); 2599bf181b9fSMichel Lespinasse anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, 2600bf181b9fSMichel Lespinasse 0, ULONG_MAX) { 2601ad12695fSAndrea Arcangeli cond_resched(); 26025beb4930SRik van Riel vma = vmac->vma; 2603e9995ef9SHugh Dickins if (rmap_item->address < vma->vm_start || 2604e9995ef9SHugh Dickins rmap_item->address >= vma->vm_end) 2605e9995ef9SHugh Dickins continue; 2606e9995ef9SHugh Dickins /* 2607e9995ef9SHugh Dickins * Initially we examine only the vma which covers this 2608e9995ef9SHugh Dickins * rmap_item; but later, if there is still work to do, 2609e9995ef9SHugh Dickins * we examine covering vmas in other mms: in case they 2610e9995ef9SHugh Dickins * were forked from the original since ksmd passed. 2611e9995ef9SHugh Dickins */ 2612e9995ef9SHugh Dickins if ((rmap_item->mm == vma->vm_mm) == search_new_forks) 2613e9995ef9SHugh Dickins continue; 2614e9995ef9SHugh Dickins 26150dd1c7bbSJoonsoo Kim if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) 26160dd1c7bbSJoonsoo Kim continue; 26170dd1c7bbSJoonsoo Kim 2618e4b82222SMinchan Kim if (!rwc->rmap_one(page, vma, 26191df631aeSMinchan Kim rmap_item->address, rwc->arg)) { 2620b6b19f25SHugh Dickins anon_vma_unlock_read(anon_vma); 26211df631aeSMinchan Kim return; 2622e9995ef9SHugh Dickins } 26230dd1c7bbSJoonsoo Kim if (rwc->done && rwc->done(page)) { 26240dd1c7bbSJoonsoo Kim anon_vma_unlock_read(anon_vma); 26251df631aeSMinchan Kim return; 26260dd1c7bbSJoonsoo Kim } 2627e9995ef9SHugh Dickins } 2628b6b19f25SHugh Dickins anon_vma_unlock_read(anon_vma); 2629e9995ef9SHugh Dickins } 2630e9995ef9SHugh Dickins if (!search_new_forks++) 2631e9995ef9SHugh Dickins goto again; 2632e9995ef9SHugh Dickins } 2633e9995ef9SHugh Dickins 263452629506SJoonsoo Kim #ifdef CONFIG_MIGRATION 2635e9995ef9SHugh Dickins void ksm_migrate_page(struct page *newpage, struct page *oldpage) 2636e9995ef9SHugh Dickins { 2637e9995ef9SHugh Dickins struct stable_node *stable_node; 2638e9995ef9SHugh Dickins 2639309381feSSasha Levin VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage); 2640309381feSSasha Levin VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); 2641309381feSSasha Levin VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage); 2642e9995ef9SHugh Dickins 2643e9995ef9SHugh Dickins stable_node = page_stable_node(newpage); 2644e9995ef9SHugh Dickins if (stable_node) { 2645309381feSSasha Levin VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage); 264662b61f61SHugh Dickins stable_node->kpfn = page_to_pfn(newpage); 2647c8d6553bSHugh Dickins /* 2648c8d6553bSHugh Dickins * newpage->mapping was set in advance; now we need smp_wmb() 2649c8d6553bSHugh Dickins * to make sure that the new stable_node->kpfn is visible 2650c8d6553bSHugh Dickins * to get_ksm_page() before it can see that oldpage->mapping 2651c8d6553bSHugh Dickins * has gone stale (or that PageSwapCache has been cleared). 2652c8d6553bSHugh Dickins */ 2653c8d6553bSHugh Dickins smp_wmb(); 2654c8d6553bSHugh Dickins set_page_stable_node(oldpage, NULL); 2655e9995ef9SHugh Dickins } 2656e9995ef9SHugh Dickins } 2657e9995ef9SHugh Dickins #endif /* CONFIG_MIGRATION */ 2658e9995ef9SHugh Dickins 265962b61f61SHugh Dickins #ifdef CONFIG_MEMORY_HOTREMOVE 2660ef4d43a8SHugh Dickins static void wait_while_offlining(void) 2661ef4d43a8SHugh Dickins { 2662ef4d43a8SHugh Dickins while (ksm_run & KSM_RUN_OFFLINE) { 2663ef4d43a8SHugh Dickins mutex_unlock(&ksm_thread_mutex); 2664ef4d43a8SHugh Dickins wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE), 266574316201SNeilBrown TASK_UNINTERRUPTIBLE); 2666ef4d43a8SHugh Dickins mutex_lock(&ksm_thread_mutex); 2667ef4d43a8SHugh Dickins } 2668ef4d43a8SHugh Dickins } 2669ef4d43a8SHugh Dickins 26702c653d0eSAndrea Arcangeli static bool stable_node_dup_remove_range(struct stable_node *stable_node, 26712c653d0eSAndrea Arcangeli unsigned long start_pfn, 26722c653d0eSAndrea Arcangeli unsigned long end_pfn) 26732c653d0eSAndrea Arcangeli { 26742c653d0eSAndrea Arcangeli if (stable_node->kpfn >= start_pfn && 26752c653d0eSAndrea Arcangeli stable_node->kpfn < end_pfn) { 26762c653d0eSAndrea Arcangeli /* 26772c653d0eSAndrea Arcangeli * Don't get_ksm_page, page has already gone: 26782c653d0eSAndrea Arcangeli * which is why we keep kpfn instead of page* 26792c653d0eSAndrea Arcangeli */ 26802c653d0eSAndrea Arcangeli remove_node_from_stable_tree(stable_node); 26812c653d0eSAndrea Arcangeli return true; 26822c653d0eSAndrea Arcangeli } 26832c653d0eSAndrea Arcangeli return false; 26842c653d0eSAndrea Arcangeli } 26852c653d0eSAndrea Arcangeli 26862c653d0eSAndrea Arcangeli static bool stable_node_chain_remove_range(struct stable_node *stable_node, 26872c653d0eSAndrea Arcangeli unsigned long start_pfn, 26882c653d0eSAndrea Arcangeli unsigned long end_pfn, 26892c653d0eSAndrea Arcangeli struct rb_root *root) 26902c653d0eSAndrea Arcangeli { 26912c653d0eSAndrea Arcangeli struct stable_node *dup; 26922c653d0eSAndrea Arcangeli struct hlist_node *hlist_safe; 26932c653d0eSAndrea Arcangeli 26942c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 26952c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node)); 26962c653d0eSAndrea Arcangeli return stable_node_dup_remove_range(stable_node, start_pfn, 26972c653d0eSAndrea Arcangeli end_pfn); 26982c653d0eSAndrea Arcangeli } 26992c653d0eSAndrea Arcangeli 27002c653d0eSAndrea Arcangeli hlist_for_each_entry_safe(dup, hlist_safe, 27012c653d0eSAndrea Arcangeli &stable_node->hlist, hlist_dup) { 27022c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(dup)); 27032c653d0eSAndrea Arcangeli stable_node_dup_remove_range(dup, start_pfn, end_pfn); 27042c653d0eSAndrea Arcangeli } 27052c653d0eSAndrea Arcangeli if (hlist_empty(&stable_node->hlist)) { 27062c653d0eSAndrea Arcangeli free_stable_node_chain(stable_node, root); 27072c653d0eSAndrea Arcangeli return true; /* notify caller that tree was rebalanced */ 27082c653d0eSAndrea Arcangeli } else 27092c653d0eSAndrea Arcangeli return false; 27102c653d0eSAndrea Arcangeli } 27112c653d0eSAndrea Arcangeli 2712ee0ea59cSHugh Dickins static void ksm_check_stable_tree(unsigned long start_pfn, 271362b61f61SHugh Dickins unsigned long end_pfn) 271462b61f61SHugh Dickins { 271503640418SGeliang Tang struct stable_node *stable_node, *next; 271662b61f61SHugh Dickins struct rb_node *node; 271790bd6fd3SPetr Holasek int nid; 271862b61f61SHugh Dickins 2719ef53d16cSHugh Dickins for (nid = 0; nid < ksm_nr_node_ids; nid++) { 2720ef53d16cSHugh Dickins node = rb_first(root_stable_tree + nid); 2721ee0ea59cSHugh Dickins while (node) { 272262b61f61SHugh Dickins stable_node = rb_entry(node, struct stable_node, node); 27232c653d0eSAndrea Arcangeli if (stable_node_chain_remove_range(stable_node, 27242c653d0eSAndrea Arcangeli start_pfn, end_pfn, 27252c653d0eSAndrea Arcangeli root_stable_tree + 27262c653d0eSAndrea Arcangeli nid)) 2727ef53d16cSHugh Dickins node = rb_first(root_stable_tree + nid); 27282c653d0eSAndrea Arcangeli else 2729ee0ea59cSHugh Dickins node = rb_next(node); 2730ee0ea59cSHugh Dickins cond_resched(); 273162b61f61SHugh Dickins } 2732ee0ea59cSHugh Dickins } 273303640418SGeliang Tang list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { 27344146d2d6SHugh Dickins if (stable_node->kpfn >= start_pfn && 27354146d2d6SHugh Dickins stable_node->kpfn < end_pfn) 27364146d2d6SHugh Dickins remove_node_from_stable_tree(stable_node); 27374146d2d6SHugh Dickins cond_resched(); 27384146d2d6SHugh Dickins } 273962b61f61SHugh Dickins } 274062b61f61SHugh Dickins 274162b61f61SHugh Dickins static int ksm_memory_callback(struct notifier_block *self, 274262b61f61SHugh Dickins unsigned long action, void *arg) 274362b61f61SHugh Dickins { 274462b61f61SHugh Dickins struct memory_notify *mn = arg; 274562b61f61SHugh Dickins 274662b61f61SHugh Dickins switch (action) { 274762b61f61SHugh Dickins case MEM_GOING_OFFLINE: 274862b61f61SHugh Dickins /* 2749ef4d43a8SHugh Dickins * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items() 2750ef4d43a8SHugh Dickins * and remove_all_stable_nodes() while memory is going offline: 2751ef4d43a8SHugh Dickins * it is unsafe for them to touch the stable tree at this time. 2752ef4d43a8SHugh Dickins * But unmerge_ksm_pages(), rmap lookups and other entry points 2753ef4d43a8SHugh Dickins * which do not need the ksm_thread_mutex are all safe. 275462b61f61SHugh Dickins */ 2755ef4d43a8SHugh Dickins mutex_lock(&ksm_thread_mutex); 2756ef4d43a8SHugh Dickins ksm_run |= KSM_RUN_OFFLINE; 2757ef4d43a8SHugh Dickins mutex_unlock(&ksm_thread_mutex); 275862b61f61SHugh Dickins break; 275962b61f61SHugh Dickins 276062b61f61SHugh Dickins case MEM_OFFLINE: 276162b61f61SHugh Dickins /* 276262b61f61SHugh Dickins * Most of the work is done by page migration; but there might 276362b61f61SHugh Dickins * be a few stable_nodes left over, still pointing to struct 2764ee0ea59cSHugh Dickins * pages which have been offlined: prune those from the tree, 2765ee0ea59cSHugh Dickins * otherwise get_ksm_page() might later try to access a 2766ee0ea59cSHugh Dickins * non-existent struct page. 276762b61f61SHugh Dickins */ 2768ee0ea59cSHugh Dickins ksm_check_stable_tree(mn->start_pfn, 2769ee0ea59cSHugh Dickins mn->start_pfn + mn->nr_pages); 277062b61f61SHugh Dickins /* fallthrough */ 277162b61f61SHugh Dickins 277262b61f61SHugh Dickins case MEM_CANCEL_OFFLINE: 2773ef4d43a8SHugh Dickins mutex_lock(&ksm_thread_mutex); 2774ef4d43a8SHugh Dickins ksm_run &= ~KSM_RUN_OFFLINE; 277562b61f61SHugh Dickins mutex_unlock(&ksm_thread_mutex); 2776ef4d43a8SHugh Dickins 2777ef4d43a8SHugh Dickins smp_mb(); /* wake_up_bit advises this */ 2778ef4d43a8SHugh Dickins wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE)); 277962b61f61SHugh Dickins break; 278062b61f61SHugh Dickins } 278162b61f61SHugh Dickins return NOTIFY_OK; 278262b61f61SHugh Dickins } 2783ef4d43a8SHugh Dickins #else 2784ef4d43a8SHugh Dickins static void wait_while_offlining(void) 2785ef4d43a8SHugh Dickins { 2786ef4d43a8SHugh Dickins } 278762b61f61SHugh Dickins #endif /* CONFIG_MEMORY_HOTREMOVE */ 278862b61f61SHugh Dickins 27892ffd8679SHugh Dickins #ifdef CONFIG_SYSFS 27902ffd8679SHugh Dickins /* 27912ffd8679SHugh Dickins * This all compiles without CONFIG_SYSFS, but is a waste of space. 27922ffd8679SHugh Dickins */ 27932ffd8679SHugh Dickins 279431dbd01fSIzik Eidus #define KSM_ATTR_RO(_name) \ 279531dbd01fSIzik Eidus static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 279631dbd01fSIzik Eidus #define KSM_ATTR(_name) \ 279731dbd01fSIzik Eidus static struct kobj_attribute _name##_attr = \ 279831dbd01fSIzik Eidus __ATTR(_name, 0644, _name##_show, _name##_store) 279931dbd01fSIzik Eidus 280031dbd01fSIzik Eidus static ssize_t sleep_millisecs_show(struct kobject *kobj, 280131dbd01fSIzik Eidus struct kobj_attribute *attr, char *buf) 280231dbd01fSIzik Eidus { 280331dbd01fSIzik Eidus return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs); 280431dbd01fSIzik Eidus } 280531dbd01fSIzik Eidus 280631dbd01fSIzik Eidus static ssize_t sleep_millisecs_store(struct kobject *kobj, 280731dbd01fSIzik Eidus struct kobj_attribute *attr, 280831dbd01fSIzik Eidus const char *buf, size_t count) 280931dbd01fSIzik Eidus { 281031dbd01fSIzik Eidus unsigned long msecs; 281131dbd01fSIzik Eidus int err; 281231dbd01fSIzik Eidus 28133dbb95f7SJingoo Han err = kstrtoul(buf, 10, &msecs); 281431dbd01fSIzik Eidus if (err || msecs > UINT_MAX) 281531dbd01fSIzik Eidus return -EINVAL; 281631dbd01fSIzik Eidus 281731dbd01fSIzik Eidus ksm_thread_sleep_millisecs = msecs; 281831dbd01fSIzik Eidus 281931dbd01fSIzik Eidus return count; 282031dbd01fSIzik Eidus } 282131dbd01fSIzik Eidus KSM_ATTR(sleep_millisecs); 282231dbd01fSIzik Eidus 282331dbd01fSIzik Eidus static ssize_t pages_to_scan_show(struct kobject *kobj, 282431dbd01fSIzik Eidus struct kobj_attribute *attr, char *buf) 282531dbd01fSIzik Eidus { 282631dbd01fSIzik Eidus return sprintf(buf, "%u\n", ksm_thread_pages_to_scan); 282731dbd01fSIzik Eidus } 282831dbd01fSIzik Eidus 282931dbd01fSIzik Eidus static ssize_t pages_to_scan_store(struct kobject *kobj, 283031dbd01fSIzik Eidus struct kobj_attribute *attr, 283131dbd01fSIzik Eidus const char *buf, size_t count) 283231dbd01fSIzik Eidus { 283331dbd01fSIzik Eidus int err; 283431dbd01fSIzik Eidus unsigned long nr_pages; 283531dbd01fSIzik Eidus 28363dbb95f7SJingoo Han err = kstrtoul(buf, 10, &nr_pages); 283731dbd01fSIzik Eidus if (err || nr_pages > UINT_MAX) 283831dbd01fSIzik Eidus return -EINVAL; 283931dbd01fSIzik Eidus 284031dbd01fSIzik Eidus ksm_thread_pages_to_scan = nr_pages; 284131dbd01fSIzik Eidus 284231dbd01fSIzik Eidus return count; 284331dbd01fSIzik Eidus } 284431dbd01fSIzik Eidus KSM_ATTR(pages_to_scan); 284531dbd01fSIzik Eidus 284631dbd01fSIzik Eidus static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, 284731dbd01fSIzik Eidus char *buf) 284831dbd01fSIzik Eidus { 2849ef4d43a8SHugh Dickins return sprintf(buf, "%lu\n", ksm_run); 285031dbd01fSIzik Eidus } 285131dbd01fSIzik Eidus 285231dbd01fSIzik Eidus static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, 285331dbd01fSIzik Eidus const char *buf, size_t count) 285431dbd01fSIzik Eidus { 285531dbd01fSIzik Eidus int err; 285631dbd01fSIzik Eidus unsigned long flags; 285731dbd01fSIzik Eidus 28583dbb95f7SJingoo Han err = kstrtoul(buf, 10, &flags); 285931dbd01fSIzik Eidus if (err || flags > UINT_MAX) 286031dbd01fSIzik Eidus return -EINVAL; 286131dbd01fSIzik Eidus if (flags > KSM_RUN_UNMERGE) 286231dbd01fSIzik Eidus return -EINVAL; 286331dbd01fSIzik Eidus 286431dbd01fSIzik Eidus /* 286531dbd01fSIzik Eidus * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. 286631dbd01fSIzik Eidus * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, 2867d0f209f6SHugh Dickins * breaking COW to free the pages_shared (but leaves mm_slots 2868d0f209f6SHugh Dickins * on the list for when ksmd may be set running again). 286931dbd01fSIzik Eidus */ 287031dbd01fSIzik Eidus 287131dbd01fSIzik Eidus mutex_lock(&ksm_thread_mutex); 2872ef4d43a8SHugh Dickins wait_while_offlining(); 287331dbd01fSIzik Eidus if (ksm_run != flags) { 287431dbd01fSIzik Eidus ksm_run = flags; 2875d952b791SHugh Dickins if (flags & KSM_RUN_UNMERGE) { 2876e1e12d2fSDavid Rientjes set_current_oom_origin(); 2877d952b791SHugh Dickins err = unmerge_and_remove_all_rmap_items(); 2878e1e12d2fSDavid Rientjes clear_current_oom_origin(); 2879d952b791SHugh Dickins if (err) { 2880d952b791SHugh Dickins ksm_run = KSM_RUN_STOP; 2881d952b791SHugh Dickins count = err; 2882d952b791SHugh Dickins } 2883d952b791SHugh Dickins } 288431dbd01fSIzik Eidus } 288531dbd01fSIzik Eidus mutex_unlock(&ksm_thread_mutex); 288631dbd01fSIzik Eidus 288731dbd01fSIzik Eidus if (flags & KSM_RUN_MERGE) 288831dbd01fSIzik Eidus wake_up_interruptible(&ksm_thread_wait); 288931dbd01fSIzik Eidus 289031dbd01fSIzik Eidus return count; 289131dbd01fSIzik Eidus } 289231dbd01fSIzik Eidus KSM_ATTR(run); 289331dbd01fSIzik Eidus 289490bd6fd3SPetr Holasek #ifdef CONFIG_NUMA 289590bd6fd3SPetr Holasek static ssize_t merge_across_nodes_show(struct kobject *kobj, 289690bd6fd3SPetr Holasek struct kobj_attribute *attr, char *buf) 289790bd6fd3SPetr Holasek { 289890bd6fd3SPetr Holasek return sprintf(buf, "%u\n", ksm_merge_across_nodes); 289990bd6fd3SPetr Holasek } 290090bd6fd3SPetr Holasek 290190bd6fd3SPetr Holasek static ssize_t merge_across_nodes_store(struct kobject *kobj, 290290bd6fd3SPetr Holasek struct kobj_attribute *attr, 290390bd6fd3SPetr Holasek const char *buf, size_t count) 290490bd6fd3SPetr Holasek { 290590bd6fd3SPetr Holasek int err; 290690bd6fd3SPetr Holasek unsigned long knob; 290790bd6fd3SPetr Holasek 290890bd6fd3SPetr Holasek err = kstrtoul(buf, 10, &knob); 290990bd6fd3SPetr Holasek if (err) 291090bd6fd3SPetr Holasek return err; 291190bd6fd3SPetr Holasek if (knob > 1) 291290bd6fd3SPetr Holasek return -EINVAL; 291390bd6fd3SPetr Holasek 291490bd6fd3SPetr Holasek mutex_lock(&ksm_thread_mutex); 2915ef4d43a8SHugh Dickins wait_while_offlining(); 291690bd6fd3SPetr Holasek if (ksm_merge_across_nodes != knob) { 2917cbf86cfeSHugh Dickins if (ksm_pages_shared || remove_all_stable_nodes()) 291890bd6fd3SPetr Holasek err = -EBUSY; 2919ef53d16cSHugh Dickins else if (root_stable_tree == one_stable_tree) { 2920ef53d16cSHugh Dickins struct rb_root *buf; 2921ef53d16cSHugh Dickins /* 2922ef53d16cSHugh Dickins * This is the first time that we switch away from the 2923ef53d16cSHugh Dickins * default of merging across nodes: must now allocate 2924ef53d16cSHugh Dickins * a buffer to hold as many roots as may be needed. 2925ef53d16cSHugh Dickins * Allocate stable and unstable together: 2926ef53d16cSHugh Dickins * MAXSMP NODES_SHIFT 10 will use 16kB. 2927ef53d16cSHugh Dickins */ 2928bafe1e14SJoe Perches buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf), 2929bafe1e14SJoe Perches GFP_KERNEL); 2930ef53d16cSHugh Dickins /* Let us assume that RB_ROOT is NULL is zero */ 2931ef53d16cSHugh Dickins if (!buf) 2932ef53d16cSHugh Dickins err = -ENOMEM; 2933ef53d16cSHugh Dickins else { 2934ef53d16cSHugh Dickins root_stable_tree = buf; 2935ef53d16cSHugh Dickins root_unstable_tree = buf + nr_node_ids; 2936ef53d16cSHugh Dickins /* Stable tree is empty but not the unstable */ 2937ef53d16cSHugh Dickins root_unstable_tree[0] = one_unstable_tree[0]; 2938ef53d16cSHugh Dickins } 2939ef53d16cSHugh Dickins } 2940ef53d16cSHugh Dickins if (!err) { 294190bd6fd3SPetr Holasek ksm_merge_across_nodes = knob; 2942ef53d16cSHugh Dickins ksm_nr_node_ids = knob ? 1 : nr_node_ids; 2943ef53d16cSHugh Dickins } 294490bd6fd3SPetr Holasek } 294590bd6fd3SPetr Holasek mutex_unlock(&ksm_thread_mutex); 294690bd6fd3SPetr Holasek 294790bd6fd3SPetr Holasek return err ? err : count; 294890bd6fd3SPetr Holasek } 294990bd6fd3SPetr Holasek KSM_ATTR(merge_across_nodes); 295090bd6fd3SPetr Holasek #endif 295190bd6fd3SPetr Holasek 2952e86c59b1SClaudio Imbrenda static ssize_t use_zero_pages_show(struct kobject *kobj, 2953e86c59b1SClaudio Imbrenda struct kobj_attribute *attr, char *buf) 2954e86c59b1SClaudio Imbrenda { 2955e86c59b1SClaudio Imbrenda return sprintf(buf, "%u\n", ksm_use_zero_pages); 2956e86c59b1SClaudio Imbrenda } 2957e86c59b1SClaudio Imbrenda static ssize_t use_zero_pages_store(struct kobject *kobj, 2958e86c59b1SClaudio Imbrenda struct kobj_attribute *attr, 2959e86c59b1SClaudio Imbrenda const char *buf, size_t count) 2960e86c59b1SClaudio Imbrenda { 2961e86c59b1SClaudio Imbrenda int err; 2962e86c59b1SClaudio Imbrenda bool value; 2963e86c59b1SClaudio Imbrenda 2964e86c59b1SClaudio Imbrenda err = kstrtobool(buf, &value); 2965e86c59b1SClaudio Imbrenda if (err) 2966e86c59b1SClaudio Imbrenda return -EINVAL; 2967e86c59b1SClaudio Imbrenda 2968e86c59b1SClaudio Imbrenda ksm_use_zero_pages = value; 2969e86c59b1SClaudio Imbrenda 2970e86c59b1SClaudio Imbrenda return count; 2971e86c59b1SClaudio Imbrenda } 2972e86c59b1SClaudio Imbrenda KSM_ATTR(use_zero_pages); 2973e86c59b1SClaudio Imbrenda 29742c653d0eSAndrea Arcangeli static ssize_t max_page_sharing_show(struct kobject *kobj, 29752c653d0eSAndrea Arcangeli struct kobj_attribute *attr, char *buf) 29762c653d0eSAndrea Arcangeli { 29772c653d0eSAndrea Arcangeli return sprintf(buf, "%u\n", ksm_max_page_sharing); 29782c653d0eSAndrea Arcangeli } 29792c653d0eSAndrea Arcangeli 29802c653d0eSAndrea Arcangeli static ssize_t max_page_sharing_store(struct kobject *kobj, 29812c653d0eSAndrea Arcangeli struct kobj_attribute *attr, 29822c653d0eSAndrea Arcangeli const char *buf, size_t count) 29832c653d0eSAndrea Arcangeli { 29842c653d0eSAndrea Arcangeli int err; 29852c653d0eSAndrea Arcangeli int knob; 29862c653d0eSAndrea Arcangeli 29872c653d0eSAndrea Arcangeli err = kstrtoint(buf, 10, &knob); 29882c653d0eSAndrea Arcangeli if (err) 29892c653d0eSAndrea Arcangeli return err; 29902c653d0eSAndrea Arcangeli /* 29912c653d0eSAndrea Arcangeli * When a KSM page is created it is shared by 2 mappings. This 29922c653d0eSAndrea Arcangeli * being a signed comparison, it implicitly verifies it's not 29932c653d0eSAndrea Arcangeli * negative. 29942c653d0eSAndrea Arcangeli */ 29952c653d0eSAndrea Arcangeli if (knob < 2) 29962c653d0eSAndrea Arcangeli return -EINVAL; 29972c653d0eSAndrea Arcangeli 29982c653d0eSAndrea Arcangeli if (READ_ONCE(ksm_max_page_sharing) == knob) 29992c653d0eSAndrea Arcangeli return count; 30002c653d0eSAndrea Arcangeli 30012c653d0eSAndrea Arcangeli mutex_lock(&ksm_thread_mutex); 30022c653d0eSAndrea Arcangeli wait_while_offlining(); 30032c653d0eSAndrea Arcangeli if (ksm_max_page_sharing != knob) { 30042c653d0eSAndrea Arcangeli if (ksm_pages_shared || remove_all_stable_nodes()) 30052c653d0eSAndrea Arcangeli err = -EBUSY; 30062c653d0eSAndrea Arcangeli else 30072c653d0eSAndrea Arcangeli ksm_max_page_sharing = knob; 30082c653d0eSAndrea Arcangeli } 30092c653d0eSAndrea Arcangeli mutex_unlock(&ksm_thread_mutex); 30102c653d0eSAndrea Arcangeli 30112c653d0eSAndrea Arcangeli return err ? err : count; 30122c653d0eSAndrea Arcangeli } 30132c653d0eSAndrea Arcangeli KSM_ATTR(max_page_sharing); 30142c653d0eSAndrea Arcangeli 3015b4028260SHugh Dickins static ssize_t pages_shared_show(struct kobject *kobj, 3016b4028260SHugh Dickins struct kobj_attribute *attr, char *buf) 3017b4028260SHugh Dickins { 3018b4028260SHugh Dickins return sprintf(buf, "%lu\n", ksm_pages_shared); 3019b4028260SHugh Dickins } 3020b4028260SHugh Dickins KSM_ATTR_RO(pages_shared); 3021b4028260SHugh Dickins 3022b4028260SHugh Dickins static ssize_t pages_sharing_show(struct kobject *kobj, 3023b4028260SHugh Dickins struct kobj_attribute *attr, char *buf) 3024b4028260SHugh Dickins { 3025e178dfdeSHugh Dickins return sprintf(buf, "%lu\n", ksm_pages_sharing); 3026b4028260SHugh Dickins } 3027b4028260SHugh Dickins KSM_ATTR_RO(pages_sharing); 3028b4028260SHugh Dickins 3029473b0ce4SHugh Dickins static ssize_t pages_unshared_show(struct kobject *kobj, 3030473b0ce4SHugh Dickins struct kobj_attribute *attr, char *buf) 3031473b0ce4SHugh Dickins { 3032473b0ce4SHugh Dickins return sprintf(buf, "%lu\n", ksm_pages_unshared); 3033473b0ce4SHugh Dickins } 3034473b0ce4SHugh Dickins KSM_ATTR_RO(pages_unshared); 3035473b0ce4SHugh Dickins 3036473b0ce4SHugh Dickins static ssize_t pages_volatile_show(struct kobject *kobj, 3037473b0ce4SHugh Dickins struct kobj_attribute *attr, char *buf) 3038473b0ce4SHugh Dickins { 3039473b0ce4SHugh Dickins long ksm_pages_volatile; 3040473b0ce4SHugh Dickins 3041473b0ce4SHugh Dickins ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared 3042473b0ce4SHugh Dickins - ksm_pages_sharing - ksm_pages_unshared; 3043473b0ce4SHugh Dickins /* 3044473b0ce4SHugh Dickins * It was not worth any locking to calculate that statistic, 3045473b0ce4SHugh Dickins * but it might therefore sometimes be negative: conceal that. 3046473b0ce4SHugh Dickins */ 3047473b0ce4SHugh Dickins if (ksm_pages_volatile < 0) 3048473b0ce4SHugh Dickins ksm_pages_volatile = 0; 3049473b0ce4SHugh Dickins return sprintf(buf, "%ld\n", ksm_pages_volatile); 3050473b0ce4SHugh Dickins } 3051473b0ce4SHugh Dickins KSM_ATTR_RO(pages_volatile); 3052473b0ce4SHugh Dickins 30532c653d0eSAndrea Arcangeli static ssize_t stable_node_dups_show(struct kobject *kobj, 30542c653d0eSAndrea Arcangeli struct kobj_attribute *attr, char *buf) 30552c653d0eSAndrea Arcangeli { 30562c653d0eSAndrea Arcangeli return sprintf(buf, "%lu\n", ksm_stable_node_dups); 30572c653d0eSAndrea Arcangeli } 30582c653d0eSAndrea Arcangeli KSM_ATTR_RO(stable_node_dups); 30592c653d0eSAndrea Arcangeli 30602c653d0eSAndrea Arcangeli static ssize_t stable_node_chains_show(struct kobject *kobj, 30612c653d0eSAndrea Arcangeli struct kobj_attribute *attr, char *buf) 30622c653d0eSAndrea Arcangeli { 30632c653d0eSAndrea Arcangeli return sprintf(buf, "%lu\n", ksm_stable_node_chains); 30642c653d0eSAndrea Arcangeli } 30652c653d0eSAndrea Arcangeli KSM_ATTR_RO(stable_node_chains); 30662c653d0eSAndrea Arcangeli 30672c653d0eSAndrea Arcangeli static ssize_t 30682c653d0eSAndrea Arcangeli stable_node_chains_prune_millisecs_show(struct kobject *kobj, 30692c653d0eSAndrea Arcangeli struct kobj_attribute *attr, 30702c653d0eSAndrea Arcangeli char *buf) 30712c653d0eSAndrea Arcangeli { 30722c653d0eSAndrea Arcangeli return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs); 30732c653d0eSAndrea Arcangeli } 30742c653d0eSAndrea Arcangeli 30752c653d0eSAndrea Arcangeli static ssize_t 30762c653d0eSAndrea Arcangeli stable_node_chains_prune_millisecs_store(struct kobject *kobj, 30772c653d0eSAndrea Arcangeli struct kobj_attribute *attr, 30782c653d0eSAndrea Arcangeli const char *buf, size_t count) 30792c653d0eSAndrea Arcangeli { 30802c653d0eSAndrea Arcangeli unsigned long msecs; 30812c653d0eSAndrea Arcangeli int err; 30822c653d0eSAndrea Arcangeli 30832c653d0eSAndrea Arcangeli err = kstrtoul(buf, 10, &msecs); 30842c653d0eSAndrea Arcangeli if (err || msecs > UINT_MAX) 30852c653d0eSAndrea Arcangeli return -EINVAL; 30862c653d0eSAndrea Arcangeli 30872c653d0eSAndrea Arcangeli ksm_stable_node_chains_prune_millisecs = msecs; 30882c653d0eSAndrea Arcangeli 30892c653d0eSAndrea Arcangeli return count; 30902c653d0eSAndrea Arcangeli } 30912c653d0eSAndrea Arcangeli KSM_ATTR(stable_node_chains_prune_millisecs); 30922c653d0eSAndrea Arcangeli 3093473b0ce4SHugh Dickins static ssize_t full_scans_show(struct kobject *kobj, 3094473b0ce4SHugh Dickins struct kobj_attribute *attr, char *buf) 3095473b0ce4SHugh Dickins { 3096473b0ce4SHugh Dickins return sprintf(buf, "%lu\n", ksm_scan.seqnr); 3097473b0ce4SHugh Dickins } 3098473b0ce4SHugh Dickins KSM_ATTR_RO(full_scans); 3099473b0ce4SHugh Dickins 310031dbd01fSIzik Eidus static struct attribute *ksm_attrs[] = { 310131dbd01fSIzik Eidus &sleep_millisecs_attr.attr, 310231dbd01fSIzik Eidus &pages_to_scan_attr.attr, 310331dbd01fSIzik Eidus &run_attr.attr, 3104b4028260SHugh Dickins &pages_shared_attr.attr, 3105b4028260SHugh Dickins &pages_sharing_attr.attr, 3106473b0ce4SHugh Dickins &pages_unshared_attr.attr, 3107473b0ce4SHugh Dickins &pages_volatile_attr.attr, 3108473b0ce4SHugh Dickins &full_scans_attr.attr, 310990bd6fd3SPetr Holasek #ifdef CONFIG_NUMA 311090bd6fd3SPetr Holasek &merge_across_nodes_attr.attr, 311190bd6fd3SPetr Holasek #endif 31122c653d0eSAndrea Arcangeli &max_page_sharing_attr.attr, 31132c653d0eSAndrea Arcangeli &stable_node_chains_attr.attr, 31142c653d0eSAndrea Arcangeli &stable_node_dups_attr.attr, 31152c653d0eSAndrea Arcangeli &stable_node_chains_prune_millisecs_attr.attr, 3116e86c59b1SClaudio Imbrenda &use_zero_pages_attr.attr, 311731dbd01fSIzik Eidus NULL, 311831dbd01fSIzik Eidus }; 311931dbd01fSIzik Eidus 3120f907c26aSArvind Yadav static const struct attribute_group ksm_attr_group = { 312131dbd01fSIzik Eidus .attrs = ksm_attrs, 312231dbd01fSIzik Eidus .name = "ksm", 312331dbd01fSIzik Eidus }; 31242ffd8679SHugh Dickins #endif /* CONFIG_SYSFS */ 312531dbd01fSIzik Eidus 312631dbd01fSIzik Eidus static int __init ksm_init(void) 312731dbd01fSIzik Eidus { 312831dbd01fSIzik Eidus struct task_struct *ksm_thread; 312931dbd01fSIzik Eidus int err; 313031dbd01fSIzik Eidus 3131e86c59b1SClaudio Imbrenda /* The correct value depends on page size and endianness */ 3132e86c59b1SClaudio Imbrenda zero_checksum = calc_checksum(ZERO_PAGE(0)); 3133e86c59b1SClaudio Imbrenda /* Default to false for backwards compatibility */ 3134e86c59b1SClaudio Imbrenda ksm_use_zero_pages = false; 3135e86c59b1SClaudio Imbrenda 313631dbd01fSIzik Eidus err = ksm_slab_init(); 313731dbd01fSIzik Eidus if (err) 313831dbd01fSIzik Eidus goto out; 313931dbd01fSIzik Eidus 314031dbd01fSIzik Eidus ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); 314131dbd01fSIzik Eidus if (IS_ERR(ksm_thread)) { 314225acde31SPaul McQuade pr_err("ksm: creating kthread failed\n"); 314331dbd01fSIzik Eidus err = PTR_ERR(ksm_thread); 3144d9f8984cSLai Jiangshan goto out_free; 314531dbd01fSIzik Eidus } 314631dbd01fSIzik Eidus 31472ffd8679SHugh Dickins #ifdef CONFIG_SYSFS 314831dbd01fSIzik Eidus err = sysfs_create_group(mm_kobj, &ksm_attr_group); 314931dbd01fSIzik Eidus if (err) { 315025acde31SPaul McQuade pr_err("ksm: register sysfs failed\n"); 31512ffd8679SHugh Dickins kthread_stop(ksm_thread); 3152d9f8984cSLai Jiangshan goto out_free; 315331dbd01fSIzik Eidus } 3154c73602adSHugh Dickins #else 3155c73602adSHugh Dickins ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ 3156c73602adSHugh Dickins 31572ffd8679SHugh Dickins #endif /* CONFIG_SYSFS */ 315831dbd01fSIzik Eidus 315962b61f61SHugh Dickins #ifdef CONFIG_MEMORY_HOTREMOVE 3160ef4d43a8SHugh Dickins /* There is no significance to this priority 100 */ 316162b61f61SHugh Dickins hotplug_memory_notifier(ksm_memory_callback, 100); 316262b61f61SHugh Dickins #endif 316331dbd01fSIzik Eidus return 0; 316431dbd01fSIzik Eidus 3165d9f8984cSLai Jiangshan out_free: 316631dbd01fSIzik Eidus ksm_slab_free(); 316731dbd01fSIzik Eidus out: 316831dbd01fSIzik Eidus return err; 316931dbd01fSIzik Eidus } 3170a64fb3cdSPaul Gortmaker subsys_initcall(ksm_init); 3171