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 */ 2191105a2fcSJia He #define KSM_FLAG_MASK (SEQNR_MASK|UNSTABLE_FLAG|STABLE_FLAG) 2201105a2fcSJia He /* to mask all the flags */ 22131dbd01fSIzik Eidus 22231dbd01fSIzik Eidus /* The stable and unstable tree heads */ 223ef53d16cSHugh Dickins static struct rb_root one_stable_tree[1] = { RB_ROOT }; 224ef53d16cSHugh Dickins static struct rb_root one_unstable_tree[1] = { RB_ROOT }; 225ef53d16cSHugh Dickins static struct rb_root *root_stable_tree = one_stable_tree; 226ef53d16cSHugh Dickins static struct rb_root *root_unstable_tree = one_unstable_tree; 22731dbd01fSIzik Eidus 2284146d2d6SHugh Dickins /* Recently migrated nodes of stable tree, pending proper placement */ 2294146d2d6SHugh Dickins static LIST_HEAD(migrate_nodes); 2302c653d0eSAndrea Arcangeli #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev) 2314146d2d6SHugh Dickins 2324ca3a69bSSasha Levin #define MM_SLOTS_HASH_BITS 10 2334ca3a69bSSasha Levin static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); 23431dbd01fSIzik Eidus 23531dbd01fSIzik Eidus static struct mm_slot ksm_mm_head = { 23631dbd01fSIzik Eidus .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list), 23731dbd01fSIzik Eidus }; 23831dbd01fSIzik Eidus static struct ksm_scan ksm_scan = { 23931dbd01fSIzik Eidus .mm_slot = &ksm_mm_head, 24031dbd01fSIzik Eidus }; 24131dbd01fSIzik Eidus 24231dbd01fSIzik Eidus static struct kmem_cache *rmap_item_cache; 2437b6ba2c7SHugh Dickins static struct kmem_cache *stable_node_cache; 24431dbd01fSIzik Eidus static struct kmem_cache *mm_slot_cache; 24531dbd01fSIzik Eidus 24631dbd01fSIzik Eidus /* The number of nodes in the stable tree */ 247b4028260SHugh Dickins static unsigned long ksm_pages_shared; 24831dbd01fSIzik Eidus 249e178dfdeSHugh Dickins /* The number of page slots additionally sharing those nodes */ 250b4028260SHugh Dickins static unsigned long ksm_pages_sharing; 25131dbd01fSIzik Eidus 252473b0ce4SHugh Dickins /* The number of nodes in the unstable tree */ 253473b0ce4SHugh Dickins static unsigned long ksm_pages_unshared; 254473b0ce4SHugh Dickins 255473b0ce4SHugh Dickins /* The number of rmap_items in use: to calculate pages_volatile */ 256473b0ce4SHugh Dickins static unsigned long ksm_rmap_items; 257473b0ce4SHugh Dickins 2582c653d0eSAndrea Arcangeli /* The number of stable_node chains */ 2592c653d0eSAndrea Arcangeli static unsigned long ksm_stable_node_chains; 2602c653d0eSAndrea Arcangeli 2612c653d0eSAndrea Arcangeli /* The number of stable_node dups linked to the stable_node chains */ 2622c653d0eSAndrea Arcangeli static unsigned long ksm_stable_node_dups; 2632c653d0eSAndrea Arcangeli 2642c653d0eSAndrea Arcangeli /* Delay in pruning stale stable_node_dups in the stable_node_chains */ 2652c653d0eSAndrea Arcangeli static int ksm_stable_node_chains_prune_millisecs = 2000; 2662c653d0eSAndrea Arcangeli 2672c653d0eSAndrea Arcangeli /* Maximum number of page slots sharing a stable node */ 2682c653d0eSAndrea Arcangeli static int ksm_max_page_sharing = 256; 2692c653d0eSAndrea Arcangeli 27031dbd01fSIzik Eidus /* Number of pages ksmd should scan in one batch */ 2712c6854fdSIzik Eidus static unsigned int ksm_thread_pages_to_scan = 100; 27231dbd01fSIzik Eidus 27331dbd01fSIzik Eidus /* Milliseconds ksmd should sleep between batches */ 2742ffd8679SHugh Dickins static unsigned int ksm_thread_sleep_millisecs = 20; 27531dbd01fSIzik Eidus 276e86c59b1SClaudio Imbrenda /* Checksum of an empty (zeroed) page */ 277e86c59b1SClaudio Imbrenda static unsigned int zero_checksum __read_mostly; 278e86c59b1SClaudio Imbrenda 279e86c59b1SClaudio Imbrenda /* Whether to merge empty (zeroed) pages with actual zero pages */ 280e86c59b1SClaudio Imbrenda static bool ksm_use_zero_pages __read_mostly; 281e86c59b1SClaudio Imbrenda 282e850dcf5SHugh Dickins #ifdef CONFIG_NUMA 28390bd6fd3SPetr Holasek /* Zeroed when merging across nodes is not allowed */ 28490bd6fd3SPetr Holasek static unsigned int ksm_merge_across_nodes = 1; 285ef53d16cSHugh Dickins static int ksm_nr_node_ids = 1; 286e850dcf5SHugh Dickins #else 287e850dcf5SHugh Dickins #define ksm_merge_across_nodes 1U 288ef53d16cSHugh Dickins #define ksm_nr_node_ids 1 289e850dcf5SHugh Dickins #endif 29090bd6fd3SPetr Holasek 29131dbd01fSIzik Eidus #define KSM_RUN_STOP 0 29231dbd01fSIzik Eidus #define KSM_RUN_MERGE 1 29331dbd01fSIzik Eidus #define KSM_RUN_UNMERGE 2 294ef4d43a8SHugh Dickins #define KSM_RUN_OFFLINE 4 295ef4d43a8SHugh Dickins static unsigned long ksm_run = KSM_RUN_STOP; 296ef4d43a8SHugh Dickins static void wait_while_offlining(void); 29731dbd01fSIzik Eidus 29831dbd01fSIzik Eidus static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); 29931dbd01fSIzik Eidus static DEFINE_MUTEX(ksm_thread_mutex); 30031dbd01fSIzik Eidus static DEFINE_SPINLOCK(ksm_mmlist_lock); 30131dbd01fSIzik Eidus 30231dbd01fSIzik Eidus #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\ 30331dbd01fSIzik Eidus sizeof(struct __struct), __alignof__(struct __struct),\ 30431dbd01fSIzik Eidus (__flags), NULL) 30531dbd01fSIzik Eidus 30631dbd01fSIzik Eidus static int __init ksm_slab_init(void) 30731dbd01fSIzik Eidus { 30831dbd01fSIzik Eidus rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0); 30931dbd01fSIzik Eidus if (!rmap_item_cache) 31031dbd01fSIzik Eidus goto out; 31131dbd01fSIzik Eidus 3127b6ba2c7SHugh Dickins stable_node_cache = KSM_KMEM_CACHE(stable_node, 0); 3137b6ba2c7SHugh Dickins if (!stable_node_cache) 3147b6ba2c7SHugh Dickins goto out_free1; 3157b6ba2c7SHugh Dickins 31631dbd01fSIzik Eidus mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0); 31731dbd01fSIzik Eidus if (!mm_slot_cache) 3187b6ba2c7SHugh Dickins goto out_free2; 31931dbd01fSIzik Eidus 32031dbd01fSIzik Eidus return 0; 32131dbd01fSIzik Eidus 3227b6ba2c7SHugh Dickins out_free2: 3237b6ba2c7SHugh Dickins kmem_cache_destroy(stable_node_cache); 3247b6ba2c7SHugh Dickins out_free1: 32531dbd01fSIzik Eidus kmem_cache_destroy(rmap_item_cache); 32631dbd01fSIzik Eidus out: 32731dbd01fSIzik Eidus return -ENOMEM; 32831dbd01fSIzik Eidus } 32931dbd01fSIzik Eidus 33031dbd01fSIzik Eidus static void __init ksm_slab_free(void) 33131dbd01fSIzik Eidus { 33231dbd01fSIzik Eidus kmem_cache_destroy(mm_slot_cache); 3337b6ba2c7SHugh Dickins kmem_cache_destroy(stable_node_cache); 33431dbd01fSIzik Eidus kmem_cache_destroy(rmap_item_cache); 33531dbd01fSIzik Eidus mm_slot_cache = NULL; 33631dbd01fSIzik Eidus } 33731dbd01fSIzik Eidus 3382c653d0eSAndrea Arcangeli static __always_inline bool is_stable_node_chain(struct stable_node *chain) 3392c653d0eSAndrea Arcangeli { 3402c653d0eSAndrea Arcangeli return chain->rmap_hlist_len == STABLE_NODE_CHAIN; 3412c653d0eSAndrea Arcangeli } 3422c653d0eSAndrea Arcangeli 3432c653d0eSAndrea Arcangeli static __always_inline bool is_stable_node_dup(struct stable_node *dup) 3442c653d0eSAndrea Arcangeli { 3452c653d0eSAndrea Arcangeli return dup->head == STABLE_NODE_DUP_HEAD; 3462c653d0eSAndrea Arcangeli } 3472c653d0eSAndrea Arcangeli 3482c653d0eSAndrea Arcangeli static inline void stable_node_chain_add_dup(struct stable_node *dup, 3492c653d0eSAndrea Arcangeli struct stable_node *chain) 3502c653d0eSAndrea Arcangeli { 3512c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(dup)); 3522c653d0eSAndrea Arcangeli dup->head = STABLE_NODE_DUP_HEAD; 3532c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_chain(chain)); 3542c653d0eSAndrea Arcangeli hlist_add_head(&dup->hlist_dup, &chain->hlist); 3552c653d0eSAndrea Arcangeli ksm_stable_node_dups++; 3562c653d0eSAndrea Arcangeli } 3572c653d0eSAndrea Arcangeli 3582c653d0eSAndrea Arcangeli static inline void __stable_node_dup_del(struct stable_node *dup) 3592c653d0eSAndrea Arcangeli { 360b4fecc67SAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(dup)); 3612c653d0eSAndrea Arcangeli hlist_del(&dup->hlist_dup); 3622c653d0eSAndrea Arcangeli ksm_stable_node_dups--; 3632c653d0eSAndrea Arcangeli } 3642c653d0eSAndrea Arcangeli 3652c653d0eSAndrea Arcangeli static inline void stable_node_dup_del(struct stable_node *dup) 3662c653d0eSAndrea Arcangeli { 3672c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_chain(dup)); 3682c653d0eSAndrea Arcangeli if (is_stable_node_dup(dup)) 3692c653d0eSAndrea Arcangeli __stable_node_dup_del(dup); 3702c653d0eSAndrea Arcangeli else 3712c653d0eSAndrea Arcangeli rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid)); 3722c653d0eSAndrea Arcangeli #ifdef CONFIG_DEBUG_VM 3732c653d0eSAndrea Arcangeli dup->head = NULL; 3742c653d0eSAndrea Arcangeli #endif 3752c653d0eSAndrea Arcangeli } 3762c653d0eSAndrea Arcangeli 37731dbd01fSIzik Eidus static inline struct rmap_item *alloc_rmap_item(void) 37831dbd01fSIzik Eidus { 379473b0ce4SHugh Dickins struct rmap_item *rmap_item; 380473b0ce4SHugh Dickins 3815b398e41Szhong jiang rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL | 3825b398e41Szhong jiang __GFP_NORETRY | __GFP_NOWARN); 383473b0ce4SHugh Dickins if (rmap_item) 384473b0ce4SHugh Dickins ksm_rmap_items++; 385473b0ce4SHugh Dickins return rmap_item; 38631dbd01fSIzik Eidus } 38731dbd01fSIzik Eidus 38831dbd01fSIzik Eidus static inline void free_rmap_item(struct rmap_item *rmap_item) 38931dbd01fSIzik Eidus { 390473b0ce4SHugh Dickins ksm_rmap_items--; 39131dbd01fSIzik Eidus rmap_item->mm = NULL; /* debug safety */ 39231dbd01fSIzik Eidus kmem_cache_free(rmap_item_cache, rmap_item); 39331dbd01fSIzik Eidus } 39431dbd01fSIzik Eidus 3957b6ba2c7SHugh Dickins static inline struct stable_node *alloc_stable_node(void) 3967b6ba2c7SHugh Dickins { 3976213055fSzhong jiang /* 3986213055fSzhong jiang * The allocation can take too long with GFP_KERNEL when memory is under 3996213055fSzhong jiang * pressure, which may lead to hung task warnings. Adding __GFP_HIGH 4006213055fSzhong jiang * grants access to memory reserves, helping to avoid this problem. 4016213055fSzhong jiang */ 4026213055fSzhong jiang return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH); 4037b6ba2c7SHugh Dickins } 4047b6ba2c7SHugh Dickins 4057b6ba2c7SHugh Dickins static inline void free_stable_node(struct stable_node *stable_node) 4067b6ba2c7SHugh Dickins { 4072c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len && 4082c653d0eSAndrea Arcangeli !is_stable_node_chain(stable_node)); 4097b6ba2c7SHugh Dickins kmem_cache_free(stable_node_cache, stable_node); 4107b6ba2c7SHugh Dickins } 4117b6ba2c7SHugh Dickins 41231dbd01fSIzik Eidus static inline struct mm_slot *alloc_mm_slot(void) 41331dbd01fSIzik Eidus { 41431dbd01fSIzik Eidus if (!mm_slot_cache) /* initialization failed */ 41531dbd01fSIzik Eidus return NULL; 41631dbd01fSIzik Eidus return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); 41731dbd01fSIzik Eidus } 41831dbd01fSIzik Eidus 41931dbd01fSIzik Eidus static inline void free_mm_slot(struct mm_slot *mm_slot) 42031dbd01fSIzik Eidus { 42131dbd01fSIzik Eidus kmem_cache_free(mm_slot_cache, mm_slot); 42231dbd01fSIzik Eidus } 42331dbd01fSIzik Eidus 42431dbd01fSIzik Eidus static struct mm_slot *get_mm_slot(struct mm_struct *mm) 42531dbd01fSIzik Eidus { 4264ca3a69bSSasha Levin struct mm_slot *slot; 42731dbd01fSIzik Eidus 428b67bfe0dSSasha Levin hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm) 4294ca3a69bSSasha Levin if (slot->mm == mm) 4304ca3a69bSSasha Levin return slot; 4314ca3a69bSSasha Levin 43231dbd01fSIzik Eidus return NULL; 43331dbd01fSIzik Eidus } 43431dbd01fSIzik Eidus 43531dbd01fSIzik Eidus static void insert_to_mm_slots_hash(struct mm_struct *mm, 43631dbd01fSIzik Eidus struct mm_slot *mm_slot) 43731dbd01fSIzik Eidus { 43831dbd01fSIzik Eidus mm_slot->mm = mm; 4394ca3a69bSSasha Levin hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm); 44031dbd01fSIzik Eidus } 44131dbd01fSIzik Eidus 44231dbd01fSIzik Eidus /* 443a913e182SHugh Dickins * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's 444a913e182SHugh Dickins * page tables after it has passed through ksm_exit() - which, if necessary, 445a913e182SHugh Dickins * takes mmap_sem briefly to serialize against them. ksm_exit() does not set 446a913e182SHugh Dickins * a special flag: they can just back out as soon as mm_users goes to zero. 447a913e182SHugh Dickins * ksm_test_exit() is used throughout to make this test for exit: in some 448a913e182SHugh Dickins * places for correctness, in some places just to avoid unnecessary work. 449a913e182SHugh Dickins */ 450a913e182SHugh Dickins static inline bool ksm_test_exit(struct mm_struct *mm) 451a913e182SHugh Dickins { 452a913e182SHugh Dickins return atomic_read(&mm->mm_users) == 0; 453a913e182SHugh Dickins } 454a913e182SHugh Dickins 455a913e182SHugh Dickins /* 45631dbd01fSIzik Eidus * We use break_ksm to break COW on a ksm page: it's a stripped down 45731dbd01fSIzik Eidus * 458d4edcf0dSDave Hansen * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1) 45931dbd01fSIzik Eidus * put_page(page); 46031dbd01fSIzik Eidus * 46131dbd01fSIzik Eidus * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma, 46231dbd01fSIzik Eidus * in case the application has unmapped and remapped mm,addr meanwhile. 46331dbd01fSIzik Eidus * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP 46431dbd01fSIzik Eidus * mmap of /dev/mem or /dev/kmem, where we would not want to touch it. 4651b2ee126SDave Hansen * 4661b2ee126SDave Hansen * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context 4671b2ee126SDave Hansen * of the process that owns 'vma'. We also do not want to enforce 4681b2ee126SDave Hansen * protection keys here anyway. 46931dbd01fSIzik Eidus */ 470d952b791SHugh Dickins static int break_ksm(struct vm_area_struct *vma, unsigned long addr) 47131dbd01fSIzik Eidus { 47231dbd01fSIzik Eidus struct page *page; 473d952b791SHugh Dickins int ret = 0; 47431dbd01fSIzik Eidus 47531dbd01fSIzik Eidus do { 47631dbd01fSIzik Eidus cond_resched(); 4771b2ee126SDave Hansen page = follow_page(vma, addr, 4781b2ee126SDave Hansen FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE); 47922eccdd7SDan Carpenter if (IS_ERR_OR_NULL(page)) 48031dbd01fSIzik Eidus break; 48131dbd01fSIzik Eidus if (PageKsm(page)) 482dcddffd4SKirill A. Shutemov ret = handle_mm_fault(vma, addr, 483dcddffd4SKirill A. Shutemov FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE); 48431dbd01fSIzik Eidus else 48531dbd01fSIzik Eidus ret = VM_FAULT_WRITE; 48631dbd01fSIzik Eidus put_page(page); 48733692f27SLinus Torvalds } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM))); 488d952b791SHugh Dickins /* 489d952b791SHugh Dickins * We must loop because handle_mm_fault() may back out if there's 490d952b791SHugh Dickins * any difficulty e.g. if pte accessed bit gets updated concurrently. 491d952b791SHugh Dickins * 492d952b791SHugh Dickins * VM_FAULT_WRITE is what we have been hoping for: it indicates that 493d952b791SHugh Dickins * COW has been broken, even if the vma does not permit VM_WRITE; 494d952b791SHugh Dickins * but note that a concurrent fault might break PageKsm for us. 495d952b791SHugh Dickins * 496d952b791SHugh Dickins * VM_FAULT_SIGBUS could occur if we race with truncation of the 497d952b791SHugh Dickins * backing file, which also invalidates anonymous pages: that's 498d952b791SHugh Dickins * okay, that truncation will have unmapped the PageKsm for us. 499d952b791SHugh Dickins * 500d952b791SHugh Dickins * VM_FAULT_OOM: at the time of writing (late July 2009), setting 501d952b791SHugh Dickins * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the 502d952b791SHugh Dickins * current task has TIF_MEMDIE set, and will be OOM killed on return 503d952b791SHugh Dickins * to user; and ksmd, having no mm, would never be chosen for that. 504d952b791SHugh Dickins * 505d952b791SHugh Dickins * But if the mm is in a limited mem_cgroup, then the fault may fail 506d952b791SHugh Dickins * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and 507d952b791SHugh Dickins * even ksmd can fail in this way - though it's usually breaking ksm 508d952b791SHugh Dickins * just to undo a merge it made a moment before, so unlikely to oom. 509d952b791SHugh Dickins * 510d952b791SHugh Dickins * That's a pity: we might therefore have more kernel pages allocated 511d952b791SHugh Dickins * than we're counting as nodes in the stable tree; but ksm_do_scan 512d952b791SHugh Dickins * will retry to break_cow on each pass, so should recover the page 513d952b791SHugh Dickins * in due course. The important thing is to not let VM_MERGEABLE 514d952b791SHugh Dickins * be cleared while any such pages might remain in the area. 515d952b791SHugh Dickins */ 516d952b791SHugh Dickins return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; 51731dbd01fSIzik Eidus } 51831dbd01fSIzik Eidus 519ef694222SBob Liu static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm, 520ef694222SBob Liu unsigned long addr) 521ef694222SBob Liu { 522ef694222SBob Liu struct vm_area_struct *vma; 523ef694222SBob Liu if (ksm_test_exit(mm)) 524ef694222SBob Liu return NULL; 525ef694222SBob Liu vma = find_vma(mm, addr); 526ef694222SBob Liu if (!vma || vma->vm_start > addr) 527ef694222SBob Liu return NULL; 528ef694222SBob Liu if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) 529ef694222SBob Liu return NULL; 530ef694222SBob Liu return vma; 531ef694222SBob Liu } 532ef694222SBob Liu 5338dd3557aSHugh Dickins static void break_cow(struct rmap_item *rmap_item) 53431dbd01fSIzik Eidus { 5358dd3557aSHugh Dickins struct mm_struct *mm = rmap_item->mm; 5368dd3557aSHugh Dickins unsigned long addr = rmap_item->address; 53731dbd01fSIzik Eidus struct vm_area_struct *vma; 53831dbd01fSIzik Eidus 5394035c07aSHugh Dickins /* 5404035c07aSHugh Dickins * It is not an accident that whenever we want to break COW 5414035c07aSHugh Dickins * to undo, we also need to drop a reference to the anon_vma. 5424035c07aSHugh Dickins */ 5439e60109fSPeter Zijlstra put_anon_vma(rmap_item->anon_vma); 5444035c07aSHugh Dickins 54581464e30SHugh Dickins down_read(&mm->mmap_sem); 546ef694222SBob Liu vma = find_mergeable_vma(mm, addr); 547ef694222SBob Liu if (vma) 54831dbd01fSIzik Eidus break_ksm(vma, addr); 54931dbd01fSIzik Eidus up_read(&mm->mmap_sem); 55031dbd01fSIzik Eidus } 55131dbd01fSIzik Eidus 55231dbd01fSIzik Eidus static struct page *get_mergeable_page(struct rmap_item *rmap_item) 55331dbd01fSIzik Eidus { 55431dbd01fSIzik Eidus struct mm_struct *mm = rmap_item->mm; 55531dbd01fSIzik Eidus unsigned long addr = rmap_item->address; 55631dbd01fSIzik Eidus struct vm_area_struct *vma; 55731dbd01fSIzik Eidus struct page *page; 55831dbd01fSIzik Eidus 55931dbd01fSIzik Eidus down_read(&mm->mmap_sem); 560ef694222SBob Liu vma = find_mergeable_vma(mm, addr); 561ef694222SBob Liu if (!vma) 56231dbd01fSIzik Eidus goto out; 56331dbd01fSIzik Eidus 56431dbd01fSIzik Eidus page = follow_page(vma, addr, FOLL_GET); 56522eccdd7SDan Carpenter if (IS_ERR_OR_NULL(page)) 56631dbd01fSIzik Eidus goto out; 567f765f540SKirill A. Shutemov if (PageAnon(page)) { 56831dbd01fSIzik Eidus flush_anon_page(vma, page, addr); 56931dbd01fSIzik Eidus flush_dcache_page(page); 57031dbd01fSIzik Eidus } else { 57131dbd01fSIzik Eidus put_page(page); 572c8f95ed1SAndrea Arcangeli out: 573c8f95ed1SAndrea Arcangeli page = NULL; 57431dbd01fSIzik Eidus } 57531dbd01fSIzik Eidus up_read(&mm->mmap_sem); 57631dbd01fSIzik Eidus return page; 57731dbd01fSIzik Eidus } 57831dbd01fSIzik Eidus 57990bd6fd3SPetr Holasek /* 58090bd6fd3SPetr Holasek * This helper is used for getting right index into array of tree roots. 58190bd6fd3SPetr Holasek * When merge_across_nodes knob is set to 1, there are only two rb-trees for 58290bd6fd3SPetr Holasek * stable and unstable pages from all nodes with roots in index 0. Otherwise, 58390bd6fd3SPetr Holasek * every node has its own stable and unstable tree. 58490bd6fd3SPetr Holasek */ 58590bd6fd3SPetr Holasek static inline int get_kpfn_nid(unsigned long kpfn) 58690bd6fd3SPetr Holasek { 587d8fc16a8SHugh Dickins return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn)); 58890bd6fd3SPetr Holasek } 58990bd6fd3SPetr Holasek 5902c653d0eSAndrea Arcangeli static struct stable_node *alloc_stable_node_chain(struct stable_node *dup, 5912c653d0eSAndrea Arcangeli struct rb_root *root) 5922c653d0eSAndrea Arcangeli { 5932c653d0eSAndrea Arcangeli struct stable_node *chain = alloc_stable_node(); 5942c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_chain(dup)); 5952c653d0eSAndrea Arcangeli if (likely(chain)) { 5962c653d0eSAndrea Arcangeli INIT_HLIST_HEAD(&chain->hlist); 5972c653d0eSAndrea Arcangeli chain->chain_prune_time = jiffies; 5982c653d0eSAndrea Arcangeli chain->rmap_hlist_len = STABLE_NODE_CHAIN; 5992c653d0eSAndrea Arcangeli #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA) 6002c653d0eSAndrea Arcangeli chain->nid = -1; /* debug */ 6012c653d0eSAndrea Arcangeli #endif 6022c653d0eSAndrea Arcangeli ksm_stable_node_chains++; 6032c653d0eSAndrea Arcangeli 6042c653d0eSAndrea Arcangeli /* 6052c653d0eSAndrea Arcangeli * Put the stable node chain in the first dimension of 6062c653d0eSAndrea Arcangeli * the stable tree and at the same time remove the old 6072c653d0eSAndrea Arcangeli * stable node. 6082c653d0eSAndrea Arcangeli */ 6092c653d0eSAndrea Arcangeli rb_replace_node(&dup->node, &chain->node, root); 6102c653d0eSAndrea Arcangeli 6112c653d0eSAndrea Arcangeli /* 6122c653d0eSAndrea Arcangeli * Move the old stable node to the second dimension 6132c653d0eSAndrea Arcangeli * queued in the hlist_dup. The invariant is that all 6142c653d0eSAndrea Arcangeli * dup stable_nodes in the chain->hlist point to pages 6152c653d0eSAndrea Arcangeli * that are wrprotected and have the exact same 6162c653d0eSAndrea Arcangeli * content. 6172c653d0eSAndrea Arcangeli */ 6182c653d0eSAndrea Arcangeli stable_node_chain_add_dup(dup, chain); 6192c653d0eSAndrea Arcangeli } 6202c653d0eSAndrea Arcangeli return chain; 6212c653d0eSAndrea Arcangeli } 6222c653d0eSAndrea Arcangeli 6232c653d0eSAndrea Arcangeli static inline void free_stable_node_chain(struct stable_node *chain, 6242c653d0eSAndrea Arcangeli struct rb_root *root) 6252c653d0eSAndrea Arcangeli { 6262c653d0eSAndrea Arcangeli rb_erase(&chain->node, root); 6272c653d0eSAndrea Arcangeli free_stable_node(chain); 6282c653d0eSAndrea Arcangeli ksm_stable_node_chains--; 6292c653d0eSAndrea Arcangeli } 6302c653d0eSAndrea Arcangeli 6314035c07aSHugh Dickins static void remove_node_from_stable_tree(struct stable_node *stable_node) 6324035c07aSHugh Dickins { 6334035c07aSHugh Dickins struct rmap_item *rmap_item; 6344035c07aSHugh Dickins 6352c653d0eSAndrea Arcangeli /* check it's not STABLE_NODE_CHAIN or negative */ 6362c653d0eSAndrea Arcangeli BUG_ON(stable_node->rmap_hlist_len < 0); 6372c653d0eSAndrea Arcangeli 638b67bfe0dSSasha Levin hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { 6394035c07aSHugh Dickins if (rmap_item->hlist.next) 6404035c07aSHugh Dickins ksm_pages_sharing--; 6414035c07aSHugh Dickins else 6424035c07aSHugh Dickins ksm_pages_shared--; 6432c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len <= 0); 6442c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len--; 6459e60109fSPeter Zijlstra put_anon_vma(rmap_item->anon_vma); 6464035c07aSHugh Dickins rmap_item->address &= PAGE_MASK; 6474035c07aSHugh Dickins cond_resched(); 6484035c07aSHugh Dickins } 6494035c07aSHugh Dickins 6502c653d0eSAndrea Arcangeli /* 6512c653d0eSAndrea Arcangeli * We need the second aligned pointer of the migrate_nodes 6522c653d0eSAndrea Arcangeli * list_head to stay clear from the rb_parent_color union 6532c653d0eSAndrea Arcangeli * (aligned and different than any node) and also different 6542c653d0eSAndrea Arcangeli * from &migrate_nodes. This will verify that future list.h changes 6552c653d0eSAndrea Arcangeli * don't break STABLE_NODE_DUP_HEAD. 6562c653d0eSAndrea Arcangeli */ 6572c653d0eSAndrea Arcangeli #if GCC_VERSION >= 40903 /* only recent gcc can handle it */ 6582c653d0eSAndrea Arcangeli BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes); 6592c653d0eSAndrea Arcangeli BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1); 6602c653d0eSAndrea Arcangeli #endif 6612c653d0eSAndrea Arcangeli 6624146d2d6SHugh Dickins if (stable_node->head == &migrate_nodes) 6634146d2d6SHugh Dickins list_del(&stable_node->list); 6644146d2d6SHugh Dickins else 6652c653d0eSAndrea Arcangeli stable_node_dup_del(stable_node); 6664035c07aSHugh Dickins free_stable_node(stable_node); 6674035c07aSHugh Dickins } 6684035c07aSHugh Dickins 6694035c07aSHugh Dickins /* 6704035c07aSHugh Dickins * get_ksm_page: checks if the page indicated by the stable node 6714035c07aSHugh Dickins * is still its ksm page, despite having held no reference to it. 6724035c07aSHugh Dickins * In which case we can trust the content of the page, and it 6734035c07aSHugh Dickins * returns the gotten page; but if the page has now been zapped, 6744035c07aSHugh Dickins * remove the stale node from the stable tree and return NULL. 675c8d6553bSHugh Dickins * But beware, the stable node's page might be being migrated. 6764035c07aSHugh Dickins * 6774035c07aSHugh Dickins * You would expect the stable_node to hold a reference to the ksm page. 6784035c07aSHugh Dickins * But if it increments the page's count, swapping out has to wait for 6794035c07aSHugh Dickins * ksmd to come around again before it can free the page, which may take 6804035c07aSHugh Dickins * seconds or even minutes: much too unresponsive. So instead we use a 6814035c07aSHugh Dickins * "keyhole reference": access to the ksm page from the stable node peeps 6824035c07aSHugh Dickins * out through its keyhole to see if that page still holds the right key, 6834035c07aSHugh Dickins * pointing back to this stable node. This relies on freeing a PageAnon 6844035c07aSHugh Dickins * page to reset its page->mapping to NULL, and relies on no other use of 6854035c07aSHugh Dickins * a page to put something that might look like our key in page->mapping. 6864035c07aSHugh Dickins * is on its way to being freed; but it is an anomaly to bear in mind. 6874035c07aSHugh Dickins */ 6888fdb3dbfSHugh Dickins static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it) 6894035c07aSHugh Dickins { 6904035c07aSHugh Dickins struct page *page; 6914035c07aSHugh Dickins void *expected_mapping; 692c8d6553bSHugh Dickins unsigned long kpfn; 6934035c07aSHugh Dickins 694bda807d4SMinchan Kim expected_mapping = (void *)((unsigned long)stable_node | 695bda807d4SMinchan Kim PAGE_MAPPING_KSM); 696c8d6553bSHugh Dickins again: 69708df4774SPaul E. McKenney kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */ 698c8d6553bSHugh Dickins page = pfn_to_page(kpfn); 6994db0c3c2SJason Low if (READ_ONCE(page->mapping) != expected_mapping) 7004035c07aSHugh Dickins goto stale; 701c8d6553bSHugh Dickins 702c8d6553bSHugh Dickins /* 703c8d6553bSHugh Dickins * We cannot do anything with the page while its refcount is 0. 704c8d6553bSHugh Dickins * Usually 0 means free, or tail of a higher-order page: in which 705c8d6553bSHugh Dickins * case this node is no longer referenced, and should be freed; 706c8d6553bSHugh Dickins * however, it might mean that the page is under page_freeze_refs(). 707c8d6553bSHugh Dickins * The __remove_mapping() case is easy, again the node is now stale; 708c8d6553bSHugh Dickins * but if page is swapcache in migrate_page_move_mapping(), it might 709c8d6553bSHugh Dickins * still be our page, in which case it's essential to keep the node. 710c8d6553bSHugh Dickins */ 711c8d6553bSHugh Dickins while (!get_page_unless_zero(page)) { 712c8d6553bSHugh Dickins /* 713c8d6553bSHugh Dickins * Another check for page->mapping != expected_mapping would 714c8d6553bSHugh Dickins * work here too. We have chosen the !PageSwapCache test to 715c8d6553bSHugh Dickins * optimize the common case, when the page is or is about to 716c8d6553bSHugh Dickins * be freed: PageSwapCache is cleared (under spin_lock_irq) 717c8d6553bSHugh Dickins * in the freeze_refs section of __remove_mapping(); but Anon 718c8d6553bSHugh Dickins * page->mapping reset to NULL later, in free_pages_prepare(). 719c8d6553bSHugh Dickins */ 720c8d6553bSHugh Dickins if (!PageSwapCache(page)) 7214035c07aSHugh Dickins goto stale; 722c8d6553bSHugh Dickins cpu_relax(); 723c8d6553bSHugh Dickins } 724c8d6553bSHugh Dickins 7254db0c3c2SJason Low if (READ_ONCE(page->mapping) != expected_mapping) { 7264035c07aSHugh Dickins put_page(page); 7274035c07aSHugh Dickins goto stale; 7284035c07aSHugh Dickins } 729c8d6553bSHugh Dickins 7308fdb3dbfSHugh Dickins if (lock_it) { 7318aafa6a4SHugh Dickins lock_page(page); 7324db0c3c2SJason Low if (READ_ONCE(page->mapping) != expected_mapping) { 7338aafa6a4SHugh Dickins unlock_page(page); 7348aafa6a4SHugh Dickins put_page(page); 7358aafa6a4SHugh Dickins goto stale; 7368aafa6a4SHugh Dickins } 7378aafa6a4SHugh Dickins } 7384035c07aSHugh Dickins return page; 739c8d6553bSHugh Dickins 7404035c07aSHugh Dickins stale: 741c8d6553bSHugh Dickins /* 742c8d6553bSHugh Dickins * We come here from above when page->mapping or !PageSwapCache 743c8d6553bSHugh Dickins * suggests that the node is stale; but it might be under migration. 744c8d6553bSHugh Dickins * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(), 745c8d6553bSHugh Dickins * before checking whether node->kpfn has been changed. 746c8d6553bSHugh Dickins */ 747c8d6553bSHugh Dickins smp_rmb(); 7484db0c3c2SJason Low if (READ_ONCE(stable_node->kpfn) != kpfn) 749c8d6553bSHugh Dickins goto again; 7504035c07aSHugh Dickins remove_node_from_stable_tree(stable_node); 7514035c07aSHugh Dickins return NULL; 7524035c07aSHugh Dickins } 7534035c07aSHugh Dickins 75431dbd01fSIzik Eidus /* 75531dbd01fSIzik Eidus * Removing rmap_item from stable or unstable tree. 75631dbd01fSIzik Eidus * This function will clean the information from the stable/unstable tree. 75731dbd01fSIzik Eidus */ 75831dbd01fSIzik Eidus static void remove_rmap_item_from_tree(struct rmap_item *rmap_item) 75931dbd01fSIzik Eidus { 7607b6ba2c7SHugh Dickins if (rmap_item->address & STABLE_FLAG) { 7617b6ba2c7SHugh Dickins struct stable_node *stable_node; 7625ad64688SHugh Dickins struct page *page; 76331dbd01fSIzik Eidus 7647b6ba2c7SHugh Dickins stable_node = rmap_item->head; 7658aafa6a4SHugh Dickins page = get_ksm_page(stable_node, true); 7664035c07aSHugh Dickins if (!page) 7674035c07aSHugh Dickins goto out; 7685ad64688SHugh Dickins 7697b6ba2c7SHugh Dickins hlist_del(&rmap_item->hlist); 7705ad64688SHugh Dickins unlock_page(page); 7715ad64688SHugh Dickins put_page(page); 77208beca44SHugh Dickins 77398666f8aSAndrea Arcangeli if (!hlist_empty(&stable_node->hlist)) 7744035c07aSHugh Dickins ksm_pages_sharing--; 7754035c07aSHugh Dickins else 776b4028260SHugh Dickins ksm_pages_shared--; 7772c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len <= 0); 7782c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len--; 77931dbd01fSIzik Eidus 7809e60109fSPeter Zijlstra put_anon_vma(rmap_item->anon_vma); 78193d17715SHugh Dickins rmap_item->address &= PAGE_MASK; 78231dbd01fSIzik Eidus 7837b6ba2c7SHugh Dickins } else if (rmap_item->address & UNSTABLE_FLAG) { 78431dbd01fSIzik Eidus unsigned char age; 78531dbd01fSIzik Eidus /* 7869ba69294SHugh Dickins * Usually ksmd can and must skip the rb_erase, because 78731dbd01fSIzik Eidus * root_unstable_tree was already reset to RB_ROOT. 7889ba69294SHugh Dickins * But be careful when an mm is exiting: do the rb_erase 7899ba69294SHugh Dickins * if this rmap_item was inserted by this scan, rather 7909ba69294SHugh Dickins * than left over from before. 79131dbd01fSIzik Eidus */ 79231dbd01fSIzik Eidus age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); 793cd551f97SHugh Dickins BUG_ON(age > 1); 79431dbd01fSIzik Eidus if (!age) 79590bd6fd3SPetr Holasek rb_erase(&rmap_item->node, 796ef53d16cSHugh Dickins root_unstable_tree + NUMA(rmap_item->nid)); 79793d17715SHugh Dickins ksm_pages_unshared--; 79831dbd01fSIzik Eidus rmap_item->address &= PAGE_MASK; 79993d17715SHugh Dickins } 8004035c07aSHugh Dickins out: 80131dbd01fSIzik Eidus cond_resched(); /* we're called from many long loops */ 80231dbd01fSIzik Eidus } 80331dbd01fSIzik Eidus 80431dbd01fSIzik Eidus static void remove_trailing_rmap_items(struct mm_slot *mm_slot, 8056514d511SHugh Dickins struct rmap_item **rmap_list) 80631dbd01fSIzik Eidus { 8076514d511SHugh Dickins while (*rmap_list) { 8086514d511SHugh Dickins struct rmap_item *rmap_item = *rmap_list; 8096514d511SHugh Dickins *rmap_list = rmap_item->rmap_list; 81031dbd01fSIzik Eidus remove_rmap_item_from_tree(rmap_item); 81131dbd01fSIzik Eidus free_rmap_item(rmap_item); 81231dbd01fSIzik Eidus } 81331dbd01fSIzik Eidus } 81431dbd01fSIzik Eidus 81531dbd01fSIzik Eidus /* 816e850dcf5SHugh Dickins * Though it's very tempting to unmerge rmap_items from stable tree rather 81731dbd01fSIzik Eidus * than check every pte of a given vma, the locking doesn't quite work for 81831dbd01fSIzik Eidus * that - an rmap_item is assigned to the stable tree after inserting ksm 81931dbd01fSIzik Eidus * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing 82031dbd01fSIzik Eidus * rmap_items from parent to child at fork time (so as not to waste time 82131dbd01fSIzik Eidus * if exit comes before the next scan reaches it). 82281464e30SHugh Dickins * 82381464e30SHugh Dickins * Similarly, although we'd like to remove rmap_items (so updating counts 82481464e30SHugh Dickins * and freeing memory) when unmerging an area, it's easier to leave that 82581464e30SHugh Dickins * to the next pass of ksmd - consider, for example, how ksmd might be 82681464e30SHugh Dickins * in cmp_and_merge_page on one of the rmap_items we would be removing. 82731dbd01fSIzik Eidus */ 828d952b791SHugh Dickins static int unmerge_ksm_pages(struct vm_area_struct *vma, 82931dbd01fSIzik Eidus unsigned long start, unsigned long end) 83031dbd01fSIzik Eidus { 83131dbd01fSIzik Eidus unsigned long addr; 832d952b791SHugh Dickins int err = 0; 83331dbd01fSIzik Eidus 834d952b791SHugh Dickins for (addr = start; addr < end && !err; addr += PAGE_SIZE) { 8359ba69294SHugh Dickins if (ksm_test_exit(vma->vm_mm)) 8369ba69294SHugh Dickins break; 837d952b791SHugh Dickins if (signal_pending(current)) 838d952b791SHugh Dickins err = -ERESTARTSYS; 839d952b791SHugh Dickins else 840d952b791SHugh Dickins err = break_ksm(vma, addr); 841d952b791SHugh Dickins } 842d952b791SHugh Dickins return err; 84331dbd01fSIzik Eidus } 84431dbd01fSIzik Eidus 84588484826SMike Rapoport static inline struct stable_node *page_stable_node(struct page *page) 84688484826SMike Rapoport { 84788484826SMike Rapoport return PageKsm(page) ? page_rmapping(page) : NULL; 84888484826SMike Rapoport } 84988484826SMike Rapoport 85088484826SMike Rapoport static inline void set_page_stable_node(struct page *page, 85188484826SMike Rapoport struct stable_node *stable_node) 85288484826SMike Rapoport { 85388484826SMike Rapoport page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM); 85488484826SMike Rapoport } 85588484826SMike Rapoport 8562ffd8679SHugh Dickins #ifdef CONFIG_SYSFS 8572ffd8679SHugh Dickins /* 8582ffd8679SHugh Dickins * Only called through the sysfs control interface: 8592ffd8679SHugh Dickins */ 860cbf86cfeSHugh Dickins static int remove_stable_node(struct stable_node *stable_node) 861cbf86cfeSHugh Dickins { 862cbf86cfeSHugh Dickins struct page *page; 863cbf86cfeSHugh Dickins int err; 864cbf86cfeSHugh Dickins 865cbf86cfeSHugh Dickins page = get_ksm_page(stable_node, true); 866cbf86cfeSHugh Dickins if (!page) { 867cbf86cfeSHugh Dickins /* 868cbf86cfeSHugh Dickins * get_ksm_page did remove_node_from_stable_tree itself. 869cbf86cfeSHugh Dickins */ 870cbf86cfeSHugh Dickins return 0; 871cbf86cfeSHugh Dickins } 872cbf86cfeSHugh Dickins 8738fdb3dbfSHugh Dickins if (WARN_ON_ONCE(page_mapped(page))) { 874cbf86cfeSHugh Dickins /* 8758fdb3dbfSHugh Dickins * This should not happen: but if it does, just refuse to let 8768fdb3dbfSHugh Dickins * merge_across_nodes be switched - there is no need to panic. 8778fdb3dbfSHugh Dickins */ 8788fdb3dbfSHugh Dickins err = -EBUSY; 8798fdb3dbfSHugh Dickins } else { 8808fdb3dbfSHugh Dickins /* 8818fdb3dbfSHugh Dickins * The stable node did not yet appear stale to get_ksm_page(), 8828fdb3dbfSHugh Dickins * since that allows for an unmapped ksm page to be recognized 8838fdb3dbfSHugh Dickins * right up until it is freed; but the node is safe to remove. 884cbf86cfeSHugh Dickins * This page might be in a pagevec waiting to be freed, 885cbf86cfeSHugh Dickins * or it might be PageSwapCache (perhaps under writeback), 886cbf86cfeSHugh Dickins * or it might have been removed from swapcache a moment ago. 887cbf86cfeSHugh Dickins */ 888cbf86cfeSHugh Dickins set_page_stable_node(page, NULL); 889cbf86cfeSHugh Dickins remove_node_from_stable_tree(stable_node); 890cbf86cfeSHugh Dickins err = 0; 891cbf86cfeSHugh Dickins } 892cbf86cfeSHugh Dickins 893cbf86cfeSHugh Dickins unlock_page(page); 894cbf86cfeSHugh Dickins put_page(page); 895cbf86cfeSHugh Dickins return err; 896cbf86cfeSHugh Dickins } 897cbf86cfeSHugh Dickins 8982c653d0eSAndrea Arcangeli static int remove_stable_node_chain(struct stable_node *stable_node, 8992c653d0eSAndrea Arcangeli struct rb_root *root) 9002c653d0eSAndrea Arcangeli { 9012c653d0eSAndrea Arcangeli struct stable_node *dup; 9022c653d0eSAndrea Arcangeli struct hlist_node *hlist_safe; 9032c653d0eSAndrea Arcangeli 9042c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 9052c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node)); 9062c653d0eSAndrea Arcangeli if (remove_stable_node(stable_node)) 9072c653d0eSAndrea Arcangeli return true; 9082c653d0eSAndrea Arcangeli else 9092c653d0eSAndrea Arcangeli return false; 9102c653d0eSAndrea Arcangeli } 9112c653d0eSAndrea Arcangeli 9122c653d0eSAndrea Arcangeli hlist_for_each_entry_safe(dup, hlist_safe, 9132c653d0eSAndrea Arcangeli &stable_node->hlist, hlist_dup) { 9142c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(dup)); 9152c653d0eSAndrea Arcangeli if (remove_stable_node(dup)) 9162c653d0eSAndrea Arcangeli return true; 9172c653d0eSAndrea Arcangeli } 9182c653d0eSAndrea Arcangeli BUG_ON(!hlist_empty(&stable_node->hlist)); 9192c653d0eSAndrea Arcangeli free_stable_node_chain(stable_node, root); 9202c653d0eSAndrea Arcangeli return false; 9212c653d0eSAndrea Arcangeli } 9222c653d0eSAndrea Arcangeli 923cbf86cfeSHugh Dickins static int remove_all_stable_nodes(void) 924cbf86cfeSHugh Dickins { 92503640418SGeliang Tang struct stable_node *stable_node, *next; 926cbf86cfeSHugh Dickins int nid; 927cbf86cfeSHugh Dickins int err = 0; 928cbf86cfeSHugh Dickins 929ef53d16cSHugh Dickins for (nid = 0; nid < ksm_nr_node_ids; nid++) { 930cbf86cfeSHugh Dickins while (root_stable_tree[nid].rb_node) { 931cbf86cfeSHugh Dickins stable_node = rb_entry(root_stable_tree[nid].rb_node, 932cbf86cfeSHugh Dickins struct stable_node, node); 9332c653d0eSAndrea Arcangeli if (remove_stable_node_chain(stable_node, 9342c653d0eSAndrea Arcangeli root_stable_tree + nid)) { 935cbf86cfeSHugh Dickins err = -EBUSY; 936cbf86cfeSHugh Dickins break; /* proceed to next nid */ 937cbf86cfeSHugh Dickins } 938cbf86cfeSHugh Dickins cond_resched(); 939cbf86cfeSHugh Dickins } 940cbf86cfeSHugh Dickins } 94103640418SGeliang Tang list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { 9424146d2d6SHugh Dickins if (remove_stable_node(stable_node)) 9434146d2d6SHugh Dickins err = -EBUSY; 9444146d2d6SHugh Dickins cond_resched(); 9454146d2d6SHugh Dickins } 946cbf86cfeSHugh Dickins return err; 947cbf86cfeSHugh Dickins } 948cbf86cfeSHugh Dickins 949d952b791SHugh Dickins static int unmerge_and_remove_all_rmap_items(void) 95031dbd01fSIzik Eidus { 95131dbd01fSIzik Eidus struct mm_slot *mm_slot; 95231dbd01fSIzik Eidus struct mm_struct *mm; 95331dbd01fSIzik Eidus struct vm_area_struct *vma; 954d952b791SHugh Dickins int err = 0; 95531dbd01fSIzik Eidus 956d952b791SHugh Dickins spin_lock(&ksm_mmlist_lock); 9579ba69294SHugh Dickins ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next, 958d952b791SHugh Dickins struct mm_slot, mm_list); 959d952b791SHugh Dickins spin_unlock(&ksm_mmlist_lock); 960d952b791SHugh Dickins 9619ba69294SHugh Dickins for (mm_slot = ksm_scan.mm_slot; 9629ba69294SHugh Dickins mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) { 96331dbd01fSIzik Eidus mm = mm_slot->mm; 96431dbd01fSIzik Eidus down_read(&mm->mmap_sem); 96531dbd01fSIzik Eidus for (vma = mm->mmap; vma; vma = vma->vm_next) { 9669ba69294SHugh Dickins if (ksm_test_exit(mm)) 9679ba69294SHugh Dickins break; 96831dbd01fSIzik Eidus if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) 96931dbd01fSIzik Eidus continue; 970d952b791SHugh Dickins err = unmerge_ksm_pages(vma, 971d952b791SHugh Dickins vma->vm_start, vma->vm_end); 9729ba69294SHugh Dickins if (err) 9739ba69294SHugh Dickins goto error; 974d952b791SHugh Dickins } 9759ba69294SHugh Dickins 9766514d511SHugh Dickins remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list); 9777496fea9SZhou Chengming up_read(&mm->mmap_sem); 97831dbd01fSIzik Eidus 97931dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 9809ba69294SHugh Dickins ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next, 981d952b791SHugh Dickins struct mm_slot, mm_list); 9829ba69294SHugh Dickins if (ksm_test_exit(mm)) { 9834ca3a69bSSasha Levin hash_del(&mm_slot->link); 9849ba69294SHugh Dickins list_del(&mm_slot->mm_list); 98531dbd01fSIzik Eidus spin_unlock(&ksm_mmlist_lock); 9869ba69294SHugh Dickins 9879ba69294SHugh Dickins free_mm_slot(mm_slot); 9889ba69294SHugh Dickins clear_bit(MMF_VM_MERGEABLE, &mm->flags); 9899ba69294SHugh Dickins mmdrop(mm); 9907496fea9SZhou Chengming } else 9919ba69294SHugh Dickins spin_unlock(&ksm_mmlist_lock); 99231dbd01fSIzik Eidus } 99331dbd01fSIzik Eidus 994cbf86cfeSHugh Dickins /* Clean up stable nodes, but don't worry if some are still busy */ 995cbf86cfeSHugh Dickins remove_all_stable_nodes(); 996d952b791SHugh Dickins ksm_scan.seqnr = 0; 9979ba69294SHugh Dickins return 0; 9989ba69294SHugh Dickins 9999ba69294SHugh Dickins error: 10009ba69294SHugh Dickins up_read(&mm->mmap_sem); 1001d952b791SHugh Dickins spin_lock(&ksm_mmlist_lock); 1002d952b791SHugh Dickins ksm_scan.mm_slot = &ksm_mm_head; 1003d952b791SHugh Dickins spin_unlock(&ksm_mmlist_lock); 1004d952b791SHugh Dickins return err; 1005d952b791SHugh Dickins } 10062ffd8679SHugh Dickins #endif /* CONFIG_SYSFS */ 1007d952b791SHugh Dickins 100831dbd01fSIzik Eidus static u32 calc_checksum(struct page *page) 100931dbd01fSIzik Eidus { 101031dbd01fSIzik Eidus u32 checksum; 10119b04c5feSCong Wang void *addr = kmap_atomic(page); 101231dbd01fSIzik Eidus checksum = jhash2(addr, PAGE_SIZE / 4, 17); 10139b04c5feSCong Wang kunmap_atomic(addr); 101431dbd01fSIzik Eidus return checksum; 101531dbd01fSIzik Eidus } 101631dbd01fSIzik Eidus 101731dbd01fSIzik Eidus static int memcmp_pages(struct page *page1, struct page *page2) 101831dbd01fSIzik Eidus { 101931dbd01fSIzik Eidus char *addr1, *addr2; 102031dbd01fSIzik Eidus int ret; 102131dbd01fSIzik Eidus 10229b04c5feSCong Wang addr1 = kmap_atomic(page1); 10239b04c5feSCong Wang addr2 = kmap_atomic(page2); 102431dbd01fSIzik Eidus ret = memcmp(addr1, addr2, PAGE_SIZE); 10259b04c5feSCong Wang kunmap_atomic(addr2); 10269b04c5feSCong Wang kunmap_atomic(addr1); 102731dbd01fSIzik Eidus return ret; 102831dbd01fSIzik Eidus } 102931dbd01fSIzik Eidus 103031dbd01fSIzik Eidus static inline int pages_identical(struct page *page1, struct page *page2) 103131dbd01fSIzik Eidus { 103231dbd01fSIzik Eidus return !memcmp_pages(page1, page2); 103331dbd01fSIzik Eidus } 103431dbd01fSIzik Eidus 103531dbd01fSIzik Eidus static int write_protect_page(struct vm_area_struct *vma, struct page *page, 103631dbd01fSIzik Eidus pte_t *orig_pte) 103731dbd01fSIzik Eidus { 103831dbd01fSIzik Eidus struct mm_struct *mm = vma->vm_mm; 103936eaff33SKirill A. Shutemov struct page_vma_mapped_walk pvmw = { 104036eaff33SKirill A. Shutemov .page = page, 104136eaff33SKirill A. Shutemov .vma = vma, 104236eaff33SKirill A. Shutemov }; 104331dbd01fSIzik Eidus int swapped; 104431dbd01fSIzik Eidus int err = -EFAULT; 10456bdb913fSHaggai Eran unsigned long mmun_start; /* For mmu_notifiers */ 10466bdb913fSHaggai Eran unsigned long mmun_end; /* For mmu_notifiers */ 104731dbd01fSIzik Eidus 104836eaff33SKirill A. Shutemov pvmw.address = page_address_in_vma(page, vma); 104936eaff33SKirill A. Shutemov if (pvmw.address == -EFAULT) 105031dbd01fSIzik Eidus goto out; 105131dbd01fSIzik Eidus 105229ad768cSAndrea Arcangeli BUG_ON(PageTransCompound(page)); 10536bdb913fSHaggai Eran 105436eaff33SKirill A. Shutemov mmun_start = pvmw.address; 105536eaff33SKirill A. Shutemov mmun_end = pvmw.address + PAGE_SIZE; 10566bdb913fSHaggai Eran mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 10576bdb913fSHaggai Eran 105836eaff33SKirill A. Shutemov if (!page_vma_mapped_walk(&pvmw)) 10596bdb913fSHaggai Eran goto out_mn; 106036eaff33SKirill A. Shutemov if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?")) 106136eaff33SKirill A. Shutemov goto out_unlock; 106231dbd01fSIzik Eidus 1063595cd8f2SAneesh Kumar K.V if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) || 1064b3a81d08SMinchan Kim (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) || 1065b3a81d08SMinchan Kim mm_tlb_flush_pending(mm)) { 106631dbd01fSIzik Eidus pte_t entry; 106731dbd01fSIzik Eidus 106831dbd01fSIzik Eidus swapped = PageSwapCache(page); 106936eaff33SKirill A. Shutemov flush_cache_page(vma, pvmw.address, page_to_pfn(page)); 107031dbd01fSIzik Eidus /* 107125985edcSLucas De Marchi * Ok this is tricky, when get_user_pages_fast() run it doesn't 107231dbd01fSIzik Eidus * take any lock, therefore the check that we are going to make 107331dbd01fSIzik Eidus * with the pagecount against the mapcount is racey and 107431dbd01fSIzik Eidus * O_DIRECT can happen right after the check. 107531dbd01fSIzik Eidus * So we clear the pte and flush the tlb before the check 107631dbd01fSIzik Eidus * this assure us that no O_DIRECT can happen after the check 107731dbd01fSIzik Eidus * or in the middle of the check. 10780f10851eSJérôme Glisse * 10790f10851eSJérôme Glisse * No need to notify as we are downgrading page table to read 10800f10851eSJérôme Glisse * only not changing it to point to a new page. 10810f10851eSJérôme Glisse * 1082ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 108331dbd01fSIzik Eidus */ 10840f10851eSJérôme Glisse entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte); 108531dbd01fSIzik Eidus /* 108631dbd01fSIzik Eidus * Check that no O_DIRECT or similar I/O is in progress on the 108731dbd01fSIzik Eidus * page 108831dbd01fSIzik Eidus */ 108931e855eaSHugh Dickins if (page_mapcount(page) + 1 + swapped != page_count(page)) { 109036eaff33SKirill A. Shutemov set_pte_at(mm, pvmw.address, pvmw.pte, entry); 109131dbd01fSIzik Eidus goto out_unlock; 109231dbd01fSIzik Eidus } 10934e31635cSHugh Dickins if (pte_dirty(entry)) 10944e31635cSHugh Dickins set_page_dirty(page); 1095595cd8f2SAneesh Kumar K.V 1096595cd8f2SAneesh Kumar K.V if (pte_protnone(entry)) 1097595cd8f2SAneesh Kumar K.V entry = pte_mkclean(pte_clear_savedwrite(entry)); 1098595cd8f2SAneesh Kumar K.V else 10994e31635cSHugh Dickins entry = pte_mkclean(pte_wrprotect(entry)); 110036eaff33SKirill A. Shutemov set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry); 110131dbd01fSIzik Eidus } 110236eaff33SKirill A. Shutemov *orig_pte = *pvmw.pte; 110331dbd01fSIzik Eidus err = 0; 110431dbd01fSIzik Eidus 110531dbd01fSIzik Eidus out_unlock: 110636eaff33SKirill A. Shutemov page_vma_mapped_walk_done(&pvmw); 11076bdb913fSHaggai Eran out_mn: 11086bdb913fSHaggai Eran mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 110931dbd01fSIzik Eidus out: 111031dbd01fSIzik Eidus return err; 111131dbd01fSIzik Eidus } 111231dbd01fSIzik Eidus 111331dbd01fSIzik Eidus /** 111431dbd01fSIzik Eidus * replace_page - replace page in vma by new ksm page 11158dd3557aSHugh Dickins * @vma: vma that holds the pte pointing to page 11168dd3557aSHugh Dickins * @page: the page we are replacing by kpage 11178dd3557aSHugh Dickins * @kpage: the ksm page we replace page by 111831dbd01fSIzik Eidus * @orig_pte: the original value of the pte 111931dbd01fSIzik Eidus * 112031dbd01fSIzik Eidus * Returns 0 on success, -EFAULT on failure. 112131dbd01fSIzik Eidus */ 11228dd3557aSHugh Dickins static int replace_page(struct vm_area_struct *vma, struct page *page, 11238dd3557aSHugh Dickins struct page *kpage, pte_t orig_pte) 112431dbd01fSIzik Eidus { 112531dbd01fSIzik Eidus struct mm_struct *mm = vma->vm_mm; 112631dbd01fSIzik Eidus pmd_t *pmd; 112731dbd01fSIzik Eidus pte_t *ptep; 1128e86c59b1SClaudio Imbrenda pte_t newpte; 112931dbd01fSIzik Eidus spinlock_t *ptl; 113031dbd01fSIzik Eidus unsigned long addr; 113131dbd01fSIzik Eidus int err = -EFAULT; 11326bdb913fSHaggai Eran unsigned long mmun_start; /* For mmu_notifiers */ 11336bdb913fSHaggai Eran unsigned long mmun_end; /* For mmu_notifiers */ 113431dbd01fSIzik Eidus 11358dd3557aSHugh Dickins addr = page_address_in_vma(page, vma); 113631dbd01fSIzik Eidus if (addr == -EFAULT) 113731dbd01fSIzik Eidus goto out; 113831dbd01fSIzik Eidus 11396219049aSBob Liu pmd = mm_find_pmd(mm, addr); 11406219049aSBob Liu if (!pmd) 114131dbd01fSIzik Eidus goto out; 114231dbd01fSIzik Eidus 11436bdb913fSHaggai Eran mmun_start = addr; 11446bdb913fSHaggai Eran mmun_end = addr + PAGE_SIZE; 11456bdb913fSHaggai Eran mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 11466bdb913fSHaggai Eran 114731dbd01fSIzik Eidus ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); 114831dbd01fSIzik Eidus if (!pte_same(*ptep, orig_pte)) { 114931dbd01fSIzik Eidus pte_unmap_unlock(ptep, ptl); 11506bdb913fSHaggai Eran goto out_mn; 115131dbd01fSIzik Eidus } 115231dbd01fSIzik Eidus 1153e86c59b1SClaudio Imbrenda /* 1154e86c59b1SClaudio Imbrenda * No need to check ksm_use_zero_pages here: we can only have a 1155e86c59b1SClaudio Imbrenda * zero_page here if ksm_use_zero_pages was enabled alreaady. 1156e86c59b1SClaudio Imbrenda */ 1157e86c59b1SClaudio Imbrenda if (!is_zero_pfn(page_to_pfn(kpage))) { 11588dd3557aSHugh Dickins get_page(kpage); 1159d281ee61SKirill A. Shutemov page_add_anon_rmap(kpage, vma, addr, false); 1160e86c59b1SClaudio Imbrenda newpte = mk_pte(kpage, vma->vm_page_prot); 1161e86c59b1SClaudio Imbrenda } else { 1162e86c59b1SClaudio Imbrenda newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage), 1163e86c59b1SClaudio Imbrenda vma->vm_page_prot)); 1164a38c015fSClaudio Imbrenda /* 1165a38c015fSClaudio Imbrenda * We're replacing an anonymous page with a zero page, which is 1166a38c015fSClaudio Imbrenda * not anonymous. We need to do proper accounting otherwise we 1167a38c015fSClaudio Imbrenda * will get wrong values in /proc, and a BUG message in dmesg 1168a38c015fSClaudio Imbrenda * when tearing down the mm. 1169a38c015fSClaudio Imbrenda */ 1170a38c015fSClaudio Imbrenda dec_mm_counter(mm, MM_ANONPAGES); 1171e86c59b1SClaudio Imbrenda } 117231dbd01fSIzik Eidus 117331dbd01fSIzik Eidus flush_cache_page(vma, addr, pte_pfn(*ptep)); 11740f10851eSJérôme Glisse /* 11750f10851eSJérôme Glisse * No need to notify as we are replacing a read only page with another 11760f10851eSJérôme Glisse * read only page with the same content. 11770f10851eSJérôme Glisse * 1178ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 11790f10851eSJérôme Glisse */ 11800f10851eSJérôme Glisse ptep_clear_flush(vma, addr, ptep); 1181e86c59b1SClaudio Imbrenda set_pte_at_notify(mm, addr, ptep, newpte); 118231dbd01fSIzik Eidus 1183d281ee61SKirill A. Shutemov page_remove_rmap(page, false); 1184ae52a2adSHugh Dickins if (!page_mapped(page)) 1185ae52a2adSHugh Dickins try_to_free_swap(page); 11868dd3557aSHugh Dickins put_page(page); 118731dbd01fSIzik Eidus 118831dbd01fSIzik Eidus pte_unmap_unlock(ptep, ptl); 118931dbd01fSIzik Eidus err = 0; 11906bdb913fSHaggai Eran out_mn: 11916bdb913fSHaggai Eran mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 119231dbd01fSIzik Eidus out: 119331dbd01fSIzik Eidus return err; 119431dbd01fSIzik Eidus } 119531dbd01fSIzik Eidus 119631dbd01fSIzik Eidus /* 119731dbd01fSIzik Eidus * try_to_merge_one_page - take two pages and merge them into one 11988dd3557aSHugh Dickins * @vma: the vma that holds the pte pointing to page 11998dd3557aSHugh Dickins * @page: the PageAnon page that we want to replace with kpage 120080e14822SHugh Dickins * @kpage: the PageKsm page that we want to map instead of page, 120180e14822SHugh Dickins * or NULL the first time when we want to use page as kpage. 120231dbd01fSIzik Eidus * 120331dbd01fSIzik Eidus * This function returns 0 if the pages were merged, -EFAULT otherwise. 120431dbd01fSIzik Eidus */ 120531dbd01fSIzik Eidus static int try_to_merge_one_page(struct vm_area_struct *vma, 12068dd3557aSHugh Dickins struct page *page, struct page *kpage) 120731dbd01fSIzik Eidus { 120831dbd01fSIzik Eidus pte_t orig_pte = __pte(0); 120931dbd01fSIzik Eidus int err = -EFAULT; 121031dbd01fSIzik Eidus 1211db114b83SHugh Dickins if (page == kpage) /* ksm page forked */ 1212db114b83SHugh Dickins return 0; 1213db114b83SHugh Dickins 12148dd3557aSHugh Dickins if (!PageAnon(page)) 121531dbd01fSIzik Eidus goto out; 121631dbd01fSIzik Eidus 121731dbd01fSIzik Eidus /* 121831dbd01fSIzik Eidus * We need the page lock to read a stable PageSwapCache in 121931dbd01fSIzik Eidus * write_protect_page(). We use trylock_page() instead of 122031dbd01fSIzik Eidus * lock_page() because we don't want to wait here - we 122131dbd01fSIzik Eidus * prefer to continue scanning and merging different pages, 122231dbd01fSIzik Eidus * then come back to this page when it is unlocked. 122331dbd01fSIzik Eidus */ 12248dd3557aSHugh Dickins if (!trylock_page(page)) 122531e855eaSHugh Dickins goto out; 1226f765f540SKirill A. Shutemov 1227f765f540SKirill A. Shutemov if (PageTransCompound(page)) { 1228a7306c34SAndrea Arcangeli if (split_huge_page(page)) 1229f765f540SKirill A. Shutemov goto out_unlock; 1230f765f540SKirill A. Shutemov } 1231f765f540SKirill A. Shutemov 123231dbd01fSIzik Eidus /* 123331dbd01fSIzik Eidus * If this anonymous page is mapped only here, its pte may need 123431dbd01fSIzik Eidus * to be write-protected. If it's mapped elsewhere, all of its 123531dbd01fSIzik Eidus * ptes are necessarily already write-protected. But in either 123631dbd01fSIzik Eidus * case, we need to lock and check page_count is not raised. 123731dbd01fSIzik Eidus */ 123880e14822SHugh Dickins if (write_protect_page(vma, page, &orig_pte) == 0) { 123980e14822SHugh Dickins if (!kpage) { 124080e14822SHugh Dickins /* 124180e14822SHugh Dickins * While we hold page lock, upgrade page from 124280e14822SHugh Dickins * PageAnon+anon_vma to PageKsm+NULL stable_node: 124380e14822SHugh Dickins * stable_tree_insert() will update stable_node. 124480e14822SHugh Dickins */ 124580e14822SHugh Dickins set_page_stable_node(page, NULL); 124680e14822SHugh Dickins mark_page_accessed(page); 1247337ed7ebSMinchan Kim /* 1248337ed7ebSMinchan Kim * Page reclaim just frees a clean page with no dirty 1249337ed7ebSMinchan Kim * ptes: make sure that the ksm page would be swapped. 1250337ed7ebSMinchan Kim */ 1251337ed7ebSMinchan Kim if (!PageDirty(page)) 1252337ed7ebSMinchan Kim SetPageDirty(page); 125380e14822SHugh Dickins err = 0; 125480e14822SHugh Dickins } else if (pages_identical(page, kpage)) 12558dd3557aSHugh Dickins err = replace_page(vma, page, kpage, orig_pte); 125680e14822SHugh Dickins } 125731dbd01fSIzik Eidus 125880e14822SHugh Dickins if ((vma->vm_flags & VM_LOCKED) && kpage && !err) { 125973848b46SHugh Dickins munlock_vma_page(page); 12605ad64688SHugh Dickins if (!PageMlocked(kpage)) { 12615ad64688SHugh Dickins unlock_page(page); 12625ad64688SHugh Dickins lock_page(kpage); 12635ad64688SHugh Dickins mlock_vma_page(kpage); 12645ad64688SHugh Dickins page = kpage; /* for final unlock */ 12655ad64688SHugh Dickins } 12665ad64688SHugh Dickins } 126773848b46SHugh Dickins 1268f765f540SKirill A. Shutemov out_unlock: 12698dd3557aSHugh Dickins unlock_page(page); 127031dbd01fSIzik Eidus out: 127131dbd01fSIzik Eidus return err; 127231dbd01fSIzik Eidus } 127331dbd01fSIzik Eidus 127431dbd01fSIzik Eidus /* 127581464e30SHugh Dickins * try_to_merge_with_ksm_page - like try_to_merge_two_pages, 127681464e30SHugh Dickins * but no new kernel page is allocated: kpage must already be a ksm page. 12778dd3557aSHugh Dickins * 12788dd3557aSHugh Dickins * This function returns 0 if the pages were merged, -EFAULT otherwise. 127981464e30SHugh Dickins */ 12808dd3557aSHugh Dickins static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item, 12818dd3557aSHugh Dickins struct page *page, struct page *kpage) 128281464e30SHugh Dickins { 12838dd3557aSHugh Dickins struct mm_struct *mm = rmap_item->mm; 128481464e30SHugh Dickins struct vm_area_struct *vma; 128581464e30SHugh Dickins int err = -EFAULT; 128681464e30SHugh Dickins 12878dd3557aSHugh Dickins down_read(&mm->mmap_sem); 128885c6e8ddSAndrea Arcangeli vma = find_mergeable_vma(mm, rmap_item->address); 128985c6e8ddSAndrea Arcangeli if (!vma) 12909ba69294SHugh Dickins goto out; 12919ba69294SHugh Dickins 12928dd3557aSHugh Dickins err = try_to_merge_one_page(vma, page, kpage); 1293db114b83SHugh Dickins if (err) 1294db114b83SHugh Dickins goto out; 1295db114b83SHugh Dickins 1296bc56620bSHugh Dickins /* Unstable nid is in union with stable anon_vma: remove first */ 1297bc56620bSHugh Dickins remove_rmap_item_from_tree(rmap_item); 1298bc56620bSHugh Dickins 1299db114b83SHugh Dickins /* Must get reference to anon_vma while still holding mmap_sem */ 13009e60109fSPeter Zijlstra rmap_item->anon_vma = vma->anon_vma; 13019e60109fSPeter Zijlstra get_anon_vma(vma->anon_vma); 130281464e30SHugh Dickins out: 13038dd3557aSHugh Dickins up_read(&mm->mmap_sem); 130481464e30SHugh Dickins return err; 130581464e30SHugh Dickins } 130681464e30SHugh Dickins 130781464e30SHugh Dickins /* 130831dbd01fSIzik Eidus * try_to_merge_two_pages - take two identical pages and prepare them 130931dbd01fSIzik Eidus * to be merged into one page. 131031dbd01fSIzik Eidus * 13118dd3557aSHugh Dickins * This function returns the kpage if we successfully merged two identical 13128dd3557aSHugh Dickins * pages into one ksm page, NULL otherwise. 131331dbd01fSIzik Eidus * 131480e14822SHugh Dickins * Note that this function upgrades page to ksm page: if one of the pages 131531dbd01fSIzik Eidus * is already a ksm page, try_to_merge_with_ksm_page should be used. 131631dbd01fSIzik Eidus */ 13178dd3557aSHugh Dickins static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item, 13188dd3557aSHugh Dickins struct page *page, 13198dd3557aSHugh Dickins struct rmap_item *tree_rmap_item, 13208dd3557aSHugh Dickins struct page *tree_page) 132131dbd01fSIzik Eidus { 132280e14822SHugh Dickins int err; 132331dbd01fSIzik Eidus 132480e14822SHugh Dickins err = try_to_merge_with_ksm_page(rmap_item, page, NULL); 132531dbd01fSIzik Eidus if (!err) { 13268dd3557aSHugh Dickins err = try_to_merge_with_ksm_page(tree_rmap_item, 132780e14822SHugh Dickins tree_page, page); 132831dbd01fSIzik Eidus /* 132981464e30SHugh Dickins * If that fails, we have a ksm page with only one pte 133081464e30SHugh Dickins * pointing to it: so break it. 133131dbd01fSIzik Eidus */ 13324035c07aSHugh Dickins if (err) 13338dd3557aSHugh Dickins break_cow(rmap_item); 133431dbd01fSIzik Eidus } 133580e14822SHugh Dickins return err ? NULL : page; 133631dbd01fSIzik Eidus } 133731dbd01fSIzik Eidus 13382c653d0eSAndrea Arcangeli static __always_inline 13392c653d0eSAndrea Arcangeli bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset) 13402c653d0eSAndrea Arcangeli { 13412c653d0eSAndrea Arcangeli VM_BUG_ON(stable_node->rmap_hlist_len < 0); 13422c653d0eSAndrea Arcangeli /* 13432c653d0eSAndrea Arcangeli * Check that at least one mapping still exists, otherwise 13442c653d0eSAndrea Arcangeli * there's no much point to merge and share with this 13452c653d0eSAndrea Arcangeli * stable_node, as the underlying tree_page of the other 13462c653d0eSAndrea Arcangeli * sharer is going to be freed soon. 13472c653d0eSAndrea Arcangeli */ 13482c653d0eSAndrea Arcangeli return stable_node->rmap_hlist_len && 13492c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len + offset < ksm_max_page_sharing; 13502c653d0eSAndrea Arcangeli } 13512c653d0eSAndrea Arcangeli 13522c653d0eSAndrea Arcangeli static __always_inline 13532c653d0eSAndrea Arcangeli bool is_page_sharing_candidate(struct stable_node *stable_node) 13542c653d0eSAndrea Arcangeli { 13552c653d0eSAndrea Arcangeli return __is_page_sharing_candidate(stable_node, 0); 13562c653d0eSAndrea Arcangeli } 13572c653d0eSAndrea Arcangeli 1358c01f0b54SColin Ian King static struct page *stable_node_dup(struct stable_node **_stable_node_dup, 13598dc5ffcdSAndrea Arcangeli struct stable_node **_stable_node, 13602c653d0eSAndrea Arcangeli struct rb_root *root, 13612c653d0eSAndrea Arcangeli bool prune_stale_stable_nodes) 13622c653d0eSAndrea Arcangeli { 1363b4fecc67SAndrea Arcangeli struct stable_node *dup, *found = NULL, *stable_node = *_stable_node; 13642c653d0eSAndrea Arcangeli struct hlist_node *hlist_safe; 13658dc5ffcdSAndrea Arcangeli struct page *_tree_page, *tree_page = NULL; 13662c653d0eSAndrea Arcangeli int nr = 0; 13672c653d0eSAndrea Arcangeli int found_rmap_hlist_len; 13682c653d0eSAndrea Arcangeli 13692c653d0eSAndrea Arcangeli if (!prune_stale_stable_nodes || 13702c653d0eSAndrea Arcangeli time_before(jiffies, stable_node->chain_prune_time + 13712c653d0eSAndrea Arcangeli msecs_to_jiffies( 13722c653d0eSAndrea Arcangeli ksm_stable_node_chains_prune_millisecs))) 13732c653d0eSAndrea Arcangeli prune_stale_stable_nodes = false; 13742c653d0eSAndrea Arcangeli else 13752c653d0eSAndrea Arcangeli stable_node->chain_prune_time = jiffies; 13762c653d0eSAndrea Arcangeli 13772c653d0eSAndrea Arcangeli hlist_for_each_entry_safe(dup, hlist_safe, 13782c653d0eSAndrea Arcangeli &stable_node->hlist, hlist_dup) { 13792c653d0eSAndrea Arcangeli cond_resched(); 13802c653d0eSAndrea Arcangeli /* 13812c653d0eSAndrea Arcangeli * We must walk all stable_node_dup to prune the stale 13822c653d0eSAndrea Arcangeli * stable nodes during lookup. 13832c653d0eSAndrea Arcangeli * 13842c653d0eSAndrea Arcangeli * get_ksm_page can drop the nodes from the 13852c653d0eSAndrea Arcangeli * stable_node->hlist if they point to freed pages 13862c653d0eSAndrea Arcangeli * (that's why we do a _safe walk). The "dup" 13872c653d0eSAndrea Arcangeli * stable_node parameter itself will be freed from 13882c653d0eSAndrea Arcangeli * under us if it returns NULL. 13892c653d0eSAndrea Arcangeli */ 13902c653d0eSAndrea Arcangeli _tree_page = get_ksm_page(dup, false); 13912c653d0eSAndrea Arcangeli if (!_tree_page) 13922c653d0eSAndrea Arcangeli continue; 13932c653d0eSAndrea Arcangeli nr += 1; 13942c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(dup)) { 13952c653d0eSAndrea Arcangeli if (!found || 13962c653d0eSAndrea Arcangeli dup->rmap_hlist_len > found_rmap_hlist_len) { 13972c653d0eSAndrea Arcangeli if (found) 13988dc5ffcdSAndrea Arcangeli put_page(tree_page); 13992c653d0eSAndrea Arcangeli found = dup; 14002c653d0eSAndrea Arcangeli found_rmap_hlist_len = found->rmap_hlist_len; 14018dc5ffcdSAndrea Arcangeli tree_page = _tree_page; 14022c653d0eSAndrea Arcangeli 14038dc5ffcdSAndrea Arcangeli /* skip put_page for found dup */ 14042c653d0eSAndrea Arcangeli if (!prune_stale_stable_nodes) 14052c653d0eSAndrea Arcangeli break; 14062c653d0eSAndrea Arcangeli continue; 14072c653d0eSAndrea Arcangeli } 14082c653d0eSAndrea Arcangeli } 14092c653d0eSAndrea Arcangeli put_page(_tree_page); 14102c653d0eSAndrea Arcangeli } 14112c653d0eSAndrea Arcangeli 141280b18dfaSAndrea Arcangeli if (found) { 14132c653d0eSAndrea Arcangeli /* 141480b18dfaSAndrea Arcangeli * nr is counting all dups in the chain only if 141580b18dfaSAndrea Arcangeli * prune_stale_stable_nodes is true, otherwise we may 141680b18dfaSAndrea Arcangeli * break the loop at nr == 1 even if there are 141780b18dfaSAndrea Arcangeli * multiple entries. 14182c653d0eSAndrea Arcangeli */ 141980b18dfaSAndrea Arcangeli if (prune_stale_stable_nodes && nr == 1) { 14202c653d0eSAndrea Arcangeli /* 14212c653d0eSAndrea Arcangeli * If there's not just one entry it would 14222c653d0eSAndrea Arcangeli * corrupt memory, better BUG_ON. In KSM 14232c653d0eSAndrea Arcangeli * context with no lock held it's not even 14242c653d0eSAndrea Arcangeli * fatal. 14252c653d0eSAndrea Arcangeli */ 14262c653d0eSAndrea Arcangeli BUG_ON(stable_node->hlist.first->next); 14272c653d0eSAndrea Arcangeli 14282c653d0eSAndrea Arcangeli /* 14292c653d0eSAndrea Arcangeli * There's just one entry and it is below the 14302c653d0eSAndrea Arcangeli * deduplication limit so drop the chain. 14312c653d0eSAndrea Arcangeli */ 14322c653d0eSAndrea Arcangeli rb_replace_node(&stable_node->node, &found->node, 14332c653d0eSAndrea Arcangeli root); 14342c653d0eSAndrea Arcangeli free_stable_node(stable_node); 14352c653d0eSAndrea Arcangeli ksm_stable_node_chains--; 14362c653d0eSAndrea Arcangeli ksm_stable_node_dups--; 1437b4fecc67SAndrea Arcangeli /* 14380ba1d0f7SAndrea Arcangeli * NOTE: the caller depends on the stable_node 14390ba1d0f7SAndrea Arcangeli * to be equal to stable_node_dup if the chain 14400ba1d0f7SAndrea Arcangeli * was collapsed. 1441b4fecc67SAndrea Arcangeli */ 14420ba1d0f7SAndrea Arcangeli *_stable_node = found; 14430ba1d0f7SAndrea Arcangeli /* 14440ba1d0f7SAndrea Arcangeli * Just for robustneess as stable_node is 14450ba1d0f7SAndrea Arcangeli * otherwise left as a stable pointer, the 14460ba1d0f7SAndrea Arcangeli * compiler shall optimize it away at build 14470ba1d0f7SAndrea Arcangeli * time. 14480ba1d0f7SAndrea Arcangeli */ 14490ba1d0f7SAndrea Arcangeli stable_node = NULL; 145080b18dfaSAndrea Arcangeli } else if (stable_node->hlist.first != &found->hlist_dup && 145180b18dfaSAndrea Arcangeli __is_page_sharing_candidate(found, 1)) { 14522c653d0eSAndrea Arcangeli /* 145380b18dfaSAndrea Arcangeli * If the found stable_node dup can accept one 145480b18dfaSAndrea Arcangeli * more future merge (in addition to the one 145580b18dfaSAndrea Arcangeli * that is underway) and is not at the head of 145680b18dfaSAndrea Arcangeli * the chain, put it there so next search will 145780b18dfaSAndrea Arcangeli * be quicker in the !prune_stale_stable_nodes 145880b18dfaSAndrea Arcangeli * case. 145980b18dfaSAndrea Arcangeli * 146080b18dfaSAndrea Arcangeli * NOTE: it would be inaccurate to use nr > 1 146180b18dfaSAndrea Arcangeli * instead of checking the hlist.first pointer 146280b18dfaSAndrea Arcangeli * directly, because in the 146380b18dfaSAndrea Arcangeli * prune_stale_stable_nodes case "nr" isn't 146480b18dfaSAndrea Arcangeli * the position of the found dup in the chain, 146580b18dfaSAndrea Arcangeli * but the total number of dups in the chain. 14662c653d0eSAndrea Arcangeli */ 14672c653d0eSAndrea Arcangeli hlist_del(&found->hlist_dup); 14682c653d0eSAndrea Arcangeli hlist_add_head(&found->hlist_dup, 14692c653d0eSAndrea Arcangeli &stable_node->hlist); 14702c653d0eSAndrea Arcangeli } 14712c653d0eSAndrea Arcangeli } 14722c653d0eSAndrea Arcangeli 14738dc5ffcdSAndrea Arcangeli *_stable_node_dup = found; 14748dc5ffcdSAndrea Arcangeli return tree_page; 14752c653d0eSAndrea Arcangeli } 14762c653d0eSAndrea Arcangeli 14772c653d0eSAndrea Arcangeli static struct stable_node *stable_node_dup_any(struct stable_node *stable_node, 14782c653d0eSAndrea Arcangeli struct rb_root *root) 14792c653d0eSAndrea Arcangeli { 14802c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) 14812c653d0eSAndrea Arcangeli return stable_node; 14822c653d0eSAndrea Arcangeli if (hlist_empty(&stable_node->hlist)) { 14832c653d0eSAndrea Arcangeli free_stable_node_chain(stable_node, root); 14842c653d0eSAndrea Arcangeli return NULL; 14852c653d0eSAndrea Arcangeli } 14862c653d0eSAndrea Arcangeli return hlist_entry(stable_node->hlist.first, 14872c653d0eSAndrea Arcangeli typeof(*stable_node), hlist_dup); 14882c653d0eSAndrea Arcangeli } 14892c653d0eSAndrea Arcangeli 14908dc5ffcdSAndrea Arcangeli /* 14918dc5ffcdSAndrea Arcangeli * Like for get_ksm_page, this function can free the *_stable_node and 14928dc5ffcdSAndrea Arcangeli * *_stable_node_dup if the returned tree_page is NULL. 14938dc5ffcdSAndrea Arcangeli * 14948dc5ffcdSAndrea Arcangeli * It can also free and overwrite *_stable_node with the found 14958dc5ffcdSAndrea Arcangeli * stable_node_dup if the chain is collapsed (in which case 14968dc5ffcdSAndrea Arcangeli * *_stable_node will be equal to *_stable_node_dup like if the chain 14978dc5ffcdSAndrea Arcangeli * never existed). It's up to the caller to verify tree_page is not 14988dc5ffcdSAndrea Arcangeli * NULL before dereferencing *_stable_node or *_stable_node_dup. 14998dc5ffcdSAndrea Arcangeli * 15008dc5ffcdSAndrea Arcangeli * *_stable_node_dup is really a second output parameter of this 15018dc5ffcdSAndrea Arcangeli * function and will be overwritten in all cases, the caller doesn't 15028dc5ffcdSAndrea Arcangeli * need to initialize it. 15038dc5ffcdSAndrea Arcangeli */ 15048dc5ffcdSAndrea Arcangeli static struct page *__stable_node_chain(struct stable_node **_stable_node_dup, 15058dc5ffcdSAndrea Arcangeli struct stable_node **_stable_node, 15062c653d0eSAndrea Arcangeli struct rb_root *root, 15072c653d0eSAndrea Arcangeli bool prune_stale_stable_nodes) 15082c653d0eSAndrea Arcangeli { 1509b4fecc67SAndrea Arcangeli struct stable_node *stable_node = *_stable_node; 15102c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 15112c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(stable_node)) { 15128dc5ffcdSAndrea Arcangeli *_stable_node_dup = stable_node; 15138dc5ffcdSAndrea Arcangeli return get_ksm_page(stable_node, false); 15142c653d0eSAndrea Arcangeli } 15158dc5ffcdSAndrea Arcangeli /* 15168dc5ffcdSAndrea Arcangeli * _stable_node_dup set to NULL means the stable_node 15178dc5ffcdSAndrea Arcangeli * reached the ksm_max_page_sharing limit. 15188dc5ffcdSAndrea Arcangeli */ 15198dc5ffcdSAndrea Arcangeli *_stable_node_dup = NULL; 15202c653d0eSAndrea Arcangeli return NULL; 15212c653d0eSAndrea Arcangeli } 15228dc5ffcdSAndrea Arcangeli return stable_node_dup(_stable_node_dup, _stable_node, root, 15232c653d0eSAndrea Arcangeli prune_stale_stable_nodes); 15242c653d0eSAndrea Arcangeli } 15252c653d0eSAndrea Arcangeli 15268dc5ffcdSAndrea Arcangeli static __always_inline struct page *chain_prune(struct stable_node **s_n_d, 15278dc5ffcdSAndrea Arcangeli struct stable_node **s_n, 15282c653d0eSAndrea Arcangeli struct rb_root *root) 15292c653d0eSAndrea Arcangeli { 15308dc5ffcdSAndrea Arcangeli return __stable_node_chain(s_n_d, s_n, root, true); 15312c653d0eSAndrea Arcangeli } 15322c653d0eSAndrea Arcangeli 15338dc5ffcdSAndrea Arcangeli static __always_inline struct page *chain(struct stable_node **s_n_d, 15348dc5ffcdSAndrea Arcangeli struct stable_node *s_n, 15352c653d0eSAndrea Arcangeli struct rb_root *root) 15362c653d0eSAndrea Arcangeli { 15378dc5ffcdSAndrea Arcangeli struct stable_node *old_stable_node = s_n; 15388dc5ffcdSAndrea Arcangeli struct page *tree_page; 15398dc5ffcdSAndrea Arcangeli 15408dc5ffcdSAndrea Arcangeli tree_page = __stable_node_chain(s_n_d, &s_n, root, false); 15418dc5ffcdSAndrea Arcangeli /* not pruning dups so s_n cannot have changed */ 15428dc5ffcdSAndrea Arcangeli VM_BUG_ON(s_n != old_stable_node); 15438dc5ffcdSAndrea Arcangeli return tree_page; 15442c653d0eSAndrea Arcangeli } 15452c653d0eSAndrea Arcangeli 154631dbd01fSIzik Eidus /* 15478dd3557aSHugh Dickins * stable_tree_search - search for page inside the stable tree 154831dbd01fSIzik Eidus * 154931dbd01fSIzik Eidus * This function checks if there is a page inside the stable tree 155031dbd01fSIzik Eidus * with identical content to the page that we are scanning right now. 155131dbd01fSIzik Eidus * 15527b6ba2c7SHugh Dickins * This function returns the stable tree node of identical content if found, 155331dbd01fSIzik Eidus * NULL otherwise. 155431dbd01fSIzik Eidus */ 155562b61f61SHugh Dickins static struct page *stable_tree_search(struct page *page) 155631dbd01fSIzik Eidus { 155790bd6fd3SPetr Holasek int nid; 1558ef53d16cSHugh Dickins struct rb_root *root; 15594146d2d6SHugh Dickins struct rb_node **new; 15604146d2d6SHugh Dickins struct rb_node *parent; 15612c653d0eSAndrea Arcangeli struct stable_node *stable_node, *stable_node_dup, *stable_node_any; 15624146d2d6SHugh Dickins struct stable_node *page_node; 156331dbd01fSIzik Eidus 15644146d2d6SHugh Dickins page_node = page_stable_node(page); 15654146d2d6SHugh Dickins if (page_node && page_node->head != &migrate_nodes) { 15664146d2d6SHugh Dickins /* ksm page forked */ 156708beca44SHugh Dickins get_page(page); 156862b61f61SHugh Dickins return page; 156908beca44SHugh Dickins } 157008beca44SHugh Dickins 157190bd6fd3SPetr Holasek nid = get_kpfn_nid(page_to_pfn(page)); 1572ef53d16cSHugh Dickins root = root_stable_tree + nid; 15734146d2d6SHugh Dickins again: 1574ef53d16cSHugh Dickins new = &root->rb_node; 15754146d2d6SHugh Dickins parent = NULL; 157690bd6fd3SPetr Holasek 15774146d2d6SHugh Dickins while (*new) { 15784035c07aSHugh Dickins struct page *tree_page; 157931dbd01fSIzik Eidus int ret; 158031dbd01fSIzik Eidus 158131dbd01fSIzik Eidus cond_resched(); 15824146d2d6SHugh Dickins stable_node = rb_entry(*new, struct stable_node, node); 15832c653d0eSAndrea Arcangeli stable_node_any = NULL; 15848dc5ffcdSAndrea Arcangeli tree_page = chain_prune(&stable_node_dup, &stable_node, root); 1585b4fecc67SAndrea Arcangeli /* 1586b4fecc67SAndrea Arcangeli * NOTE: stable_node may have been freed by 1587b4fecc67SAndrea Arcangeli * chain_prune() if the returned stable_node_dup is 1588b4fecc67SAndrea Arcangeli * not NULL. stable_node_dup may have been inserted in 1589b4fecc67SAndrea Arcangeli * the rbtree instead as a regular stable_node (in 1590b4fecc67SAndrea Arcangeli * order to collapse the stable_node chain if a single 15910ba1d0f7SAndrea Arcangeli * stable_node dup was found in it). In such case the 15920ba1d0f7SAndrea Arcangeli * stable_node is overwritten by the calleee to point 15930ba1d0f7SAndrea Arcangeli * to the stable_node_dup that was collapsed in the 15940ba1d0f7SAndrea Arcangeli * stable rbtree and stable_node will be equal to 15950ba1d0f7SAndrea Arcangeli * stable_node_dup like if the chain never existed. 1596b4fecc67SAndrea Arcangeli */ 15972c653d0eSAndrea Arcangeli if (!stable_node_dup) { 15982c653d0eSAndrea Arcangeli /* 15992c653d0eSAndrea Arcangeli * Either all stable_node dups were full in 16002c653d0eSAndrea Arcangeli * this stable_node chain, or this chain was 16012c653d0eSAndrea Arcangeli * empty and should be rb_erased. 16022c653d0eSAndrea Arcangeli */ 16032c653d0eSAndrea Arcangeli stable_node_any = stable_node_dup_any(stable_node, 16042c653d0eSAndrea Arcangeli root); 16052c653d0eSAndrea Arcangeli if (!stable_node_any) { 16062c653d0eSAndrea Arcangeli /* rb_erase just run */ 16072c653d0eSAndrea Arcangeli goto again; 16082c653d0eSAndrea Arcangeli } 16092c653d0eSAndrea Arcangeli /* 16102c653d0eSAndrea Arcangeli * Take any of the stable_node dups page of 16112c653d0eSAndrea Arcangeli * this stable_node chain to let the tree walk 16122c653d0eSAndrea Arcangeli * continue. All KSM pages belonging to the 16132c653d0eSAndrea Arcangeli * stable_node dups in a stable_node chain 16142c653d0eSAndrea Arcangeli * have the same content and they're 16152c653d0eSAndrea Arcangeli * wrprotected at all times. Any will work 16162c653d0eSAndrea Arcangeli * fine to continue the walk. 16172c653d0eSAndrea Arcangeli */ 16182c653d0eSAndrea Arcangeli tree_page = get_ksm_page(stable_node_any, false); 16192c653d0eSAndrea Arcangeli } 16202c653d0eSAndrea Arcangeli VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); 1621f2e5ff85SAndrea Arcangeli if (!tree_page) { 1622f2e5ff85SAndrea Arcangeli /* 1623f2e5ff85SAndrea Arcangeli * If we walked over a stale stable_node, 1624f2e5ff85SAndrea Arcangeli * get_ksm_page() will call rb_erase() and it 1625f2e5ff85SAndrea Arcangeli * may rebalance the tree from under us. So 1626f2e5ff85SAndrea Arcangeli * restart the search from scratch. Returning 1627f2e5ff85SAndrea Arcangeli * NULL would be safe too, but we'd generate 1628f2e5ff85SAndrea Arcangeli * false negative insertions just because some 1629f2e5ff85SAndrea Arcangeli * stable_node was stale. 1630f2e5ff85SAndrea Arcangeli */ 1631f2e5ff85SAndrea Arcangeli goto again; 1632f2e5ff85SAndrea Arcangeli } 163331dbd01fSIzik Eidus 16344035c07aSHugh Dickins ret = memcmp_pages(page, tree_page); 1635c8d6553bSHugh Dickins put_page(tree_page); 163631dbd01fSIzik Eidus 16374146d2d6SHugh Dickins parent = *new; 1638c8d6553bSHugh Dickins if (ret < 0) 16394146d2d6SHugh Dickins new = &parent->rb_left; 1640c8d6553bSHugh Dickins else if (ret > 0) 16414146d2d6SHugh Dickins new = &parent->rb_right; 1642c8d6553bSHugh Dickins else { 16432c653d0eSAndrea Arcangeli if (page_node) { 16442c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 16452c653d0eSAndrea Arcangeli /* 16462c653d0eSAndrea Arcangeli * Test if the migrated page should be merged 16472c653d0eSAndrea Arcangeli * into a stable node dup. If the mapcount is 16482c653d0eSAndrea Arcangeli * 1 we can migrate it with another KSM page 16492c653d0eSAndrea Arcangeli * without adding it to the chain. 16502c653d0eSAndrea Arcangeli */ 16512c653d0eSAndrea Arcangeli if (page_mapcount(page) > 1) 16522c653d0eSAndrea Arcangeli goto chain_append; 16532c653d0eSAndrea Arcangeli } 16542c653d0eSAndrea Arcangeli 16552c653d0eSAndrea Arcangeli if (!stable_node_dup) { 16562c653d0eSAndrea Arcangeli /* 16572c653d0eSAndrea Arcangeli * If the stable_node is a chain and 16582c653d0eSAndrea Arcangeli * we got a payload match in memcmp 16592c653d0eSAndrea Arcangeli * but we cannot merge the scanned 16602c653d0eSAndrea Arcangeli * page in any of the existing 16612c653d0eSAndrea Arcangeli * stable_node dups because they're 16622c653d0eSAndrea Arcangeli * all full, we need to wait the 16632c653d0eSAndrea Arcangeli * scanned page to find itself a match 16642c653d0eSAndrea Arcangeli * in the unstable tree to create a 16652c653d0eSAndrea Arcangeli * brand new KSM page to add later to 16662c653d0eSAndrea Arcangeli * the dups of this stable_node. 16672c653d0eSAndrea Arcangeli */ 16682c653d0eSAndrea Arcangeli return NULL; 16692c653d0eSAndrea Arcangeli } 16702c653d0eSAndrea Arcangeli 1671c8d6553bSHugh Dickins /* 1672c8d6553bSHugh Dickins * Lock and unlock the stable_node's page (which 1673c8d6553bSHugh Dickins * might already have been migrated) so that page 1674c8d6553bSHugh Dickins * migration is sure to notice its raised count. 1675c8d6553bSHugh Dickins * It would be more elegant to return stable_node 1676c8d6553bSHugh Dickins * than kpage, but that involves more changes. 1677c8d6553bSHugh Dickins */ 16782c653d0eSAndrea Arcangeli tree_page = get_ksm_page(stable_node_dup, true); 16792c653d0eSAndrea Arcangeli if (unlikely(!tree_page)) 16802c653d0eSAndrea Arcangeli /* 16812c653d0eSAndrea Arcangeli * The tree may have been rebalanced, 16822c653d0eSAndrea Arcangeli * so re-evaluate parent and new. 16832c653d0eSAndrea Arcangeli */ 16842c653d0eSAndrea Arcangeli goto again; 1685c8d6553bSHugh Dickins unlock_page(tree_page); 16862c653d0eSAndrea Arcangeli 16872c653d0eSAndrea Arcangeli if (get_kpfn_nid(stable_node_dup->kpfn) != 16882c653d0eSAndrea Arcangeli NUMA(stable_node_dup->nid)) { 16894146d2d6SHugh Dickins put_page(tree_page); 16904146d2d6SHugh Dickins goto replace; 16914146d2d6SHugh Dickins } 169262b61f61SHugh Dickins return tree_page; 169331dbd01fSIzik Eidus } 1694c8d6553bSHugh Dickins } 169531dbd01fSIzik Eidus 16964146d2d6SHugh Dickins if (!page_node) 169731dbd01fSIzik Eidus return NULL; 16984146d2d6SHugh Dickins 16994146d2d6SHugh Dickins list_del(&page_node->list); 17004146d2d6SHugh Dickins DO_NUMA(page_node->nid = nid); 17014146d2d6SHugh Dickins rb_link_node(&page_node->node, parent, new); 1702ef53d16cSHugh Dickins rb_insert_color(&page_node->node, root); 17032c653d0eSAndrea Arcangeli out: 17042c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(page_node)) { 17054146d2d6SHugh Dickins get_page(page); 17064146d2d6SHugh Dickins return page; 17072c653d0eSAndrea Arcangeli } else 17082c653d0eSAndrea Arcangeli return NULL; 17094146d2d6SHugh Dickins 17104146d2d6SHugh Dickins replace: 1711b4fecc67SAndrea Arcangeli /* 1712b4fecc67SAndrea Arcangeli * If stable_node was a chain and chain_prune collapsed it, 17130ba1d0f7SAndrea Arcangeli * stable_node has been updated to be the new regular 17140ba1d0f7SAndrea Arcangeli * stable_node. A collapse of the chain is indistinguishable 17150ba1d0f7SAndrea Arcangeli * from the case there was no chain in the stable 17160ba1d0f7SAndrea Arcangeli * rbtree. Otherwise stable_node is the chain and 17170ba1d0f7SAndrea Arcangeli * stable_node_dup is the dup to replace. 1718b4fecc67SAndrea Arcangeli */ 17190ba1d0f7SAndrea Arcangeli if (stable_node_dup == stable_node) { 1720b4fecc67SAndrea Arcangeli VM_BUG_ON(is_stable_node_chain(stable_node_dup)); 1721b4fecc67SAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node_dup)); 17222c653d0eSAndrea Arcangeli /* there is no chain */ 17234146d2d6SHugh Dickins if (page_node) { 17242c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 17254146d2d6SHugh Dickins list_del(&page_node->list); 17264146d2d6SHugh Dickins DO_NUMA(page_node->nid = nid); 1727b4fecc67SAndrea Arcangeli rb_replace_node(&stable_node_dup->node, 1728b4fecc67SAndrea Arcangeli &page_node->node, 17292c653d0eSAndrea Arcangeli root); 17302c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(page_node)) 17314146d2d6SHugh Dickins get_page(page); 17322c653d0eSAndrea Arcangeli else 17332c653d0eSAndrea Arcangeli page = NULL; 17344146d2d6SHugh Dickins } else { 1735b4fecc67SAndrea Arcangeli rb_erase(&stable_node_dup->node, root); 17364146d2d6SHugh Dickins page = NULL; 17374146d2d6SHugh Dickins } 17382c653d0eSAndrea Arcangeli } else { 17392c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_chain(stable_node)); 17402c653d0eSAndrea Arcangeli __stable_node_dup_del(stable_node_dup); 17412c653d0eSAndrea Arcangeli if (page_node) { 17422c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 17432c653d0eSAndrea Arcangeli list_del(&page_node->list); 17442c653d0eSAndrea Arcangeli DO_NUMA(page_node->nid = nid); 17452c653d0eSAndrea Arcangeli stable_node_chain_add_dup(page_node, stable_node); 17462c653d0eSAndrea Arcangeli if (is_page_sharing_candidate(page_node)) 17472c653d0eSAndrea Arcangeli get_page(page); 17482c653d0eSAndrea Arcangeli else 17492c653d0eSAndrea Arcangeli page = NULL; 17502c653d0eSAndrea Arcangeli } else { 17512c653d0eSAndrea Arcangeli page = NULL; 17522c653d0eSAndrea Arcangeli } 17532c653d0eSAndrea Arcangeli } 17542c653d0eSAndrea Arcangeli stable_node_dup->head = &migrate_nodes; 17552c653d0eSAndrea Arcangeli list_add(&stable_node_dup->list, stable_node_dup->head); 17564146d2d6SHugh Dickins return page; 17572c653d0eSAndrea Arcangeli 17582c653d0eSAndrea Arcangeli chain_append: 17592c653d0eSAndrea Arcangeli /* stable_node_dup could be null if it reached the limit */ 17602c653d0eSAndrea Arcangeli if (!stable_node_dup) 17612c653d0eSAndrea Arcangeli stable_node_dup = stable_node_any; 1762b4fecc67SAndrea Arcangeli /* 1763b4fecc67SAndrea Arcangeli * If stable_node was a chain and chain_prune collapsed it, 17640ba1d0f7SAndrea Arcangeli * stable_node has been updated to be the new regular 17650ba1d0f7SAndrea Arcangeli * stable_node. A collapse of the chain is indistinguishable 17660ba1d0f7SAndrea Arcangeli * from the case there was no chain in the stable 17670ba1d0f7SAndrea Arcangeli * rbtree. Otherwise stable_node is the chain and 17680ba1d0f7SAndrea Arcangeli * stable_node_dup is the dup to replace. 1769b4fecc67SAndrea Arcangeli */ 17700ba1d0f7SAndrea Arcangeli if (stable_node_dup == stable_node) { 1771b4fecc67SAndrea Arcangeli VM_BUG_ON(is_stable_node_chain(stable_node_dup)); 1772b4fecc67SAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node_dup)); 17732c653d0eSAndrea Arcangeli /* chain is missing so create it */ 17742c653d0eSAndrea Arcangeli stable_node = alloc_stable_node_chain(stable_node_dup, 17752c653d0eSAndrea Arcangeli root); 17762c653d0eSAndrea Arcangeli if (!stable_node) 17772c653d0eSAndrea Arcangeli return NULL; 17782c653d0eSAndrea Arcangeli } 17792c653d0eSAndrea Arcangeli /* 17802c653d0eSAndrea Arcangeli * Add this stable_node dup that was 17812c653d0eSAndrea Arcangeli * migrated to the stable_node chain 17822c653d0eSAndrea Arcangeli * of the current nid for this page 17832c653d0eSAndrea Arcangeli * content. 17842c653d0eSAndrea Arcangeli */ 1785b4fecc67SAndrea Arcangeli VM_BUG_ON(!is_stable_node_chain(stable_node)); 1786b4fecc67SAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(stable_node_dup)); 17872c653d0eSAndrea Arcangeli VM_BUG_ON(page_node->head != &migrate_nodes); 17882c653d0eSAndrea Arcangeli list_del(&page_node->list); 17892c653d0eSAndrea Arcangeli DO_NUMA(page_node->nid = nid); 17902c653d0eSAndrea Arcangeli stable_node_chain_add_dup(page_node, stable_node); 17912c653d0eSAndrea Arcangeli goto out; 179231dbd01fSIzik Eidus } 179331dbd01fSIzik Eidus 179431dbd01fSIzik Eidus /* 1795e850dcf5SHugh Dickins * stable_tree_insert - insert stable tree node pointing to new ksm page 179631dbd01fSIzik Eidus * into the stable tree. 179731dbd01fSIzik Eidus * 17987b6ba2c7SHugh Dickins * This function returns the stable tree node just allocated on success, 17997b6ba2c7SHugh Dickins * NULL otherwise. 180031dbd01fSIzik Eidus */ 18017b6ba2c7SHugh Dickins static struct stable_node *stable_tree_insert(struct page *kpage) 180231dbd01fSIzik Eidus { 180390bd6fd3SPetr Holasek int nid; 180490bd6fd3SPetr Holasek unsigned long kpfn; 1805ef53d16cSHugh Dickins struct rb_root *root; 180690bd6fd3SPetr Holasek struct rb_node **new; 1807f2e5ff85SAndrea Arcangeli struct rb_node *parent; 18082c653d0eSAndrea Arcangeli struct stable_node *stable_node, *stable_node_dup, *stable_node_any; 18092c653d0eSAndrea Arcangeli bool need_chain = false; 181031dbd01fSIzik Eidus 181190bd6fd3SPetr Holasek kpfn = page_to_pfn(kpage); 181290bd6fd3SPetr Holasek nid = get_kpfn_nid(kpfn); 1813ef53d16cSHugh Dickins root = root_stable_tree + nid; 1814f2e5ff85SAndrea Arcangeli again: 1815f2e5ff85SAndrea Arcangeli parent = NULL; 1816ef53d16cSHugh Dickins new = &root->rb_node; 181790bd6fd3SPetr Holasek 181831dbd01fSIzik Eidus while (*new) { 18194035c07aSHugh Dickins struct page *tree_page; 182031dbd01fSIzik Eidus int ret; 182131dbd01fSIzik Eidus 182231dbd01fSIzik Eidus cond_resched(); 182308beca44SHugh Dickins stable_node = rb_entry(*new, struct stable_node, node); 18242c653d0eSAndrea Arcangeli stable_node_any = NULL; 18258dc5ffcdSAndrea Arcangeli tree_page = chain(&stable_node_dup, stable_node, root); 18262c653d0eSAndrea Arcangeli if (!stable_node_dup) { 18272c653d0eSAndrea Arcangeli /* 18282c653d0eSAndrea Arcangeli * Either all stable_node dups were full in 18292c653d0eSAndrea Arcangeli * this stable_node chain, or this chain was 18302c653d0eSAndrea Arcangeli * empty and should be rb_erased. 18312c653d0eSAndrea Arcangeli */ 18322c653d0eSAndrea Arcangeli stable_node_any = stable_node_dup_any(stable_node, 18332c653d0eSAndrea Arcangeli root); 18342c653d0eSAndrea Arcangeli if (!stable_node_any) { 18352c653d0eSAndrea Arcangeli /* rb_erase just run */ 18362c653d0eSAndrea Arcangeli goto again; 18372c653d0eSAndrea Arcangeli } 18382c653d0eSAndrea Arcangeli /* 18392c653d0eSAndrea Arcangeli * Take any of the stable_node dups page of 18402c653d0eSAndrea Arcangeli * this stable_node chain to let the tree walk 18412c653d0eSAndrea Arcangeli * continue. All KSM pages belonging to the 18422c653d0eSAndrea Arcangeli * stable_node dups in a stable_node chain 18432c653d0eSAndrea Arcangeli * have the same content and they're 18442c653d0eSAndrea Arcangeli * wrprotected at all times. Any will work 18452c653d0eSAndrea Arcangeli * fine to continue the walk. 18462c653d0eSAndrea Arcangeli */ 18472c653d0eSAndrea Arcangeli tree_page = get_ksm_page(stable_node_any, false); 18482c653d0eSAndrea Arcangeli } 18492c653d0eSAndrea Arcangeli VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); 1850f2e5ff85SAndrea Arcangeli if (!tree_page) { 1851f2e5ff85SAndrea Arcangeli /* 1852f2e5ff85SAndrea Arcangeli * If we walked over a stale stable_node, 1853f2e5ff85SAndrea Arcangeli * get_ksm_page() will call rb_erase() and it 1854f2e5ff85SAndrea Arcangeli * may rebalance the tree from under us. So 1855f2e5ff85SAndrea Arcangeli * restart the search from scratch. Returning 1856f2e5ff85SAndrea Arcangeli * NULL would be safe too, but we'd generate 1857f2e5ff85SAndrea Arcangeli * false negative insertions just because some 1858f2e5ff85SAndrea Arcangeli * stable_node was stale. 1859f2e5ff85SAndrea Arcangeli */ 1860f2e5ff85SAndrea Arcangeli goto again; 1861f2e5ff85SAndrea Arcangeli } 186231dbd01fSIzik Eidus 18634035c07aSHugh Dickins ret = memcmp_pages(kpage, tree_page); 18644035c07aSHugh Dickins put_page(tree_page); 186531dbd01fSIzik Eidus 186631dbd01fSIzik Eidus parent = *new; 186731dbd01fSIzik Eidus if (ret < 0) 186831dbd01fSIzik Eidus new = &parent->rb_left; 186931dbd01fSIzik Eidus else if (ret > 0) 187031dbd01fSIzik Eidus new = &parent->rb_right; 187131dbd01fSIzik Eidus else { 18722c653d0eSAndrea Arcangeli need_chain = true; 18732c653d0eSAndrea Arcangeli break; 187431dbd01fSIzik Eidus } 187531dbd01fSIzik Eidus } 187631dbd01fSIzik Eidus 18772c653d0eSAndrea Arcangeli stable_node_dup = alloc_stable_node(); 18782c653d0eSAndrea Arcangeli if (!stable_node_dup) 18797b6ba2c7SHugh Dickins return NULL; 188031dbd01fSIzik Eidus 18812c653d0eSAndrea Arcangeli INIT_HLIST_HEAD(&stable_node_dup->hlist); 18822c653d0eSAndrea Arcangeli stable_node_dup->kpfn = kpfn; 18832c653d0eSAndrea Arcangeli set_page_stable_node(kpage, stable_node_dup); 18842c653d0eSAndrea Arcangeli stable_node_dup->rmap_hlist_len = 0; 18852c653d0eSAndrea Arcangeli DO_NUMA(stable_node_dup->nid = nid); 18862c653d0eSAndrea Arcangeli if (!need_chain) { 18872c653d0eSAndrea Arcangeli rb_link_node(&stable_node_dup->node, parent, new); 18882c653d0eSAndrea Arcangeli rb_insert_color(&stable_node_dup->node, root); 18892c653d0eSAndrea Arcangeli } else { 18902c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 18912c653d0eSAndrea Arcangeli struct stable_node *orig = stable_node; 18922c653d0eSAndrea Arcangeli /* chain is missing so create it */ 18932c653d0eSAndrea Arcangeli stable_node = alloc_stable_node_chain(orig, root); 18942c653d0eSAndrea Arcangeli if (!stable_node) { 18952c653d0eSAndrea Arcangeli free_stable_node(stable_node_dup); 18962c653d0eSAndrea Arcangeli return NULL; 18972c653d0eSAndrea Arcangeli } 18982c653d0eSAndrea Arcangeli } 18992c653d0eSAndrea Arcangeli stable_node_chain_add_dup(stable_node_dup, stable_node); 19002c653d0eSAndrea Arcangeli } 190108beca44SHugh Dickins 19022c653d0eSAndrea Arcangeli return stable_node_dup; 190331dbd01fSIzik Eidus } 190431dbd01fSIzik Eidus 190531dbd01fSIzik Eidus /* 19068dd3557aSHugh Dickins * unstable_tree_search_insert - search for identical page, 19078dd3557aSHugh Dickins * else insert rmap_item into the unstable tree. 190831dbd01fSIzik Eidus * 190931dbd01fSIzik Eidus * This function searches for a page in the unstable tree identical to the 191031dbd01fSIzik Eidus * page currently being scanned; and if no identical page is found in the 191131dbd01fSIzik Eidus * tree, we insert rmap_item as a new object into the unstable tree. 191231dbd01fSIzik Eidus * 191331dbd01fSIzik Eidus * This function returns pointer to rmap_item found to be identical 191431dbd01fSIzik Eidus * to the currently scanned page, NULL otherwise. 191531dbd01fSIzik Eidus * 191631dbd01fSIzik Eidus * This function does both searching and inserting, because they share 191731dbd01fSIzik Eidus * the same walking algorithm in an rbtree. 191831dbd01fSIzik Eidus */ 19198dd3557aSHugh Dickins static 19208dd3557aSHugh Dickins struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item, 19218dd3557aSHugh Dickins struct page *page, 19228dd3557aSHugh Dickins struct page **tree_pagep) 192331dbd01fSIzik Eidus { 192490bd6fd3SPetr Holasek struct rb_node **new; 192590bd6fd3SPetr Holasek struct rb_root *root; 192631dbd01fSIzik Eidus struct rb_node *parent = NULL; 192790bd6fd3SPetr Holasek int nid; 192890bd6fd3SPetr Holasek 192990bd6fd3SPetr Holasek nid = get_kpfn_nid(page_to_pfn(page)); 1930ef53d16cSHugh Dickins root = root_unstable_tree + nid; 193190bd6fd3SPetr Holasek new = &root->rb_node; 193231dbd01fSIzik Eidus 193331dbd01fSIzik Eidus while (*new) { 193431dbd01fSIzik Eidus struct rmap_item *tree_rmap_item; 19358dd3557aSHugh Dickins struct page *tree_page; 193631dbd01fSIzik Eidus int ret; 193731dbd01fSIzik Eidus 1938d178f27fSHugh Dickins cond_resched(); 193931dbd01fSIzik Eidus tree_rmap_item = rb_entry(*new, struct rmap_item, node); 19408dd3557aSHugh Dickins tree_page = get_mergeable_page(tree_rmap_item); 1941c8f95ed1SAndrea Arcangeli if (!tree_page) 194231dbd01fSIzik Eidus return NULL; 194331dbd01fSIzik Eidus 194431dbd01fSIzik Eidus /* 19458dd3557aSHugh Dickins * Don't substitute a ksm page for a forked page. 194631dbd01fSIzik Eidus */ 19478dd3557aSHugh Dickins if (page == tree_page) { 19488dd3557aSHugh Dickins put_page(tree_page); 194931dbd01fSIzik Eidus return NULL; 195031dbd01fSIzik Eidus } 195131dbd01fSIzik Eidus 19528dd3557aSHugh Dickins ret = memcmp_pages(page, tree_page); 195331dbd01fSIzik Eidus 195431dbd01fSIzik Eidus parent = *new; 195531dbd01fSIzik Eidus if (ret < 0) { 19568dd3557aSHugh Dickins put_page(tree_page); 195731dbd01fSIzik Eidus new = &parent->rb_left; 195831dbd01fSIzik Eidus } else if (ret > 0) { 19598dd3557aSHugh Dickins put_page(tree_page); 196031dbd01fSIzik Eidus new = &parent->rb_right; 1961b599cbdfSHugh Dickins } else if (!ksm_merge_across_nodes && 1962b599cbdfSHugh Dickins page_to_nid(tree_page) != nid) { 1963b599cbdfSHugh Dickins /* 1964b599cbdfSHugh Dickins * If tree_page has been migrated to another NUMA node, 1965b599cbdfSHugh Dickins * it will be flushed out and put in the right unstable 1966b599cbdfSHugh Dickins * tree next time: only merge with it when across_nodes. 1967b599cbdfSHugh Dickins */ 1968b599cbdfSHugh Dickins put_page(tree_page); 1969b599cbdfSHugh Dickins return NULL; 197031dbd01fSIzik Eidus } else { 19718dd3557aSHugh Dickins *tree_pagep = tree_page; 197231dbd01fSIzik Eidus return tree_rmap_item; 197331dbd01fSIzik Eidus } 197431dbd01fSIzik Eidus } 197531dbd01fSIzik Eidus 19767b6ba2c7SHugh Dickins rmap_item->address |= UNSTABLE_FLAG; 197731dbd01fSIzik Eidus rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); 1978e850dcf5SHugh Dickins DO_NUMA(rmap_item->nid = nid); 197931dbd01fSIzik Eidus rb_link_node(&rmap_item->node, parent, new); 198090bd6fd3SPetr Holasek rb_insert_color(&rmap_item->node, root); 198131dbd01fSIzik Eidus 1982473b0ce4SHugh Dickins ksm_pages_unshared++; 198331dbd01fSIzik Eidus return NULL; 198431dbd01fSIzik Eidus } 198531dbd01fSIzik Eidus 198631dbd01fSIzik Eidus /* 198731dbd01fSIzik Eidus * stable_tree_append - add another rmap_item to the linked list of 198831dbd01fSIzik Eidus * rmap_items hanging off a given node of the stable tree, all sharing 198931dbd01fSIzik Eidus * the same ksm page. 199031dbd01fSIzik Eidus */ 199131dbd01fSIzik Eidus static void stable_tree_append(struct rmap_item *rmap_item, 19922c653d0eSAndrea Arcangeli struct stable_node *stable_node, 19932c653d0eSAndrea Arcangeli bool max_page_sharing_bypass) 199431dbd01fSIzik Eidus { 19952c653d0eSAndrea Arcangeli /* 19962c653d0eSAndrea Arcangeli * rmap won't find this mapping if we don't insert the 19972c653d0eSAndrea Arcangeli * rmap_item in the right stable_node 19982c653d0eSAndrea Arcangeli * duplicate. page_migration could break later if rmap breaks, 19992c653d0eSAndrea Arcangeli * so we can as well crash here. We really need to check for 20002c653d0eSAndrea Arcangeli * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check 20012c653d0eSAndrea Arcangeli * for other negative values as an undeflow if detected here 20022c653d0eSAndrea Arcangeli * for the first time (and not when decreasing rmap_hlist_len) 20032c653d0eSAndrea Arcangeli * would be sign of memory corruption in the stable_node. 20042c653d0eSAndrea Arcangeli */ 20052c653d0eSAndrea Arcangeli BUG_ON(stable_node->rmap_hlist_len < 0); 20062c653d0eSAndrea Arcangeli 20072c653d0eSAndrea Arcangeli stable_node->rmap_hlist_len++; 20082c653d0eSAndrea Arcangeli if (!max_page_sharing_bypass) 20092c653d0eSAndrea Arcangeli /* possibly non fatal but unexpected overflow, only warn */ 20102c653d0eSAndrea Arcangeli WARN_ON_ONCE(stable_node->rmap_hlist_len > 20112c653d0eSAndrea Arcangeli ksm_max_page_sharing); 20122c653d0eSAndrea Arcangeli 20137b6ba2c7SHugh Dickins rmap_item->head = stable_node; 201431dbd01fSIzik Eidus rmap_item->address |= STABLE_FLAG; 20157b6ba2c7SHugh Dickins hlist_add_head(&rmap_item->hlist, &stable_node->hlist); 2016e178dfdeSHugh Dickins 20177b6ba2c7SHugh Dickins if (rmap_item->hlist.next) 2018e178dfdeSHugh Dickins ksm_pages_sharing++; 20197b6ba2c7SHugh Dickins else 20207b6ba2c7SHugh Dickins ksm_pages_shared++; 202131dbd01fSIzik Eidus } 202231dbd01fSIzik Eidus 202331dbd01fSIzik Eidus /* 202481464e30SHugh Dickins * cmp_and_merge_page - first see if page can be merged into the stable tree; 202581464e30SHugh Dickins * if not, compare checksum to previous and if it's the same, see if page can 202681464e30SHugh Dickins * be inserted into the unstable tree, or merged with a page already there and 202781464e30SHugh Dickins * both transferred to the stable tree. 202831dbd01fSIzik Eidus * 202931dbd01fSIzik Eidus * @page: the page that we are searching identical page to. 203031dbd01fSIzik Eidus * @rmap_item: the reverse mapping into the virtual address of this page 203131dbd01fSIzik Eidus */ 203231dbd01fSIzik Eidus static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item) 203331dbd01fSIzik Eidus { 20344b22927fSKirill Tkhai struct mm_struct *mm = rmap_item->mm; 203531dbd01fSIzik Eidus struct rmap_item *tree_rmap_item; 20368dd3557aSHugh Dickins struct page *tree_page = NULL; 20377b6ba2c7SHugh Dickins struct stable_node *stable_node; 20388dd3557aSHugh Dickins struct page *kpage; 203931dbd01fSIzik Eidus unsigned int checksum; 204031dbd01fSIzik Eidus int err; 20412c653d0eSAndrea Arcangeli bool max_page_sharing_bypass = false; 204231dbd01fSIzik Eidus 20434146d2d6SHugh Dickins stable_node = page_stable_node(page); 20444146d2d6SHugh Dickins if (stable_node) { 20454146d2d6SHugh Dickins if (stable_node->head != &migrate_nodes && 20462c653d0eSAndrea Arcangeli get_kpfn_nid(READ_ONCE(stable_node->kpfn)) != 20472c653d0eSAndrea Arcangeli NUMA(stable_node->nid)) { 20482c653d0eSAndrea Arcangeli stable_node_dup_del(stable_node); 20494146d2d6SHugh Dickins stable_node->head = &migrate_nodes; 20504146d2d6SHugh Dickins list_add(&stable_node->list, stable_node->head); 20514146d2d6SHugh Dickins } 20524146d2d6SHugh Dickins if (stable_node->head != &migrate_nodes && 20534146d2d6SHugh Dickins rmap_item->head == stable_node) 20544146d2d6SHugh Dickins return; 20552c653d0eSAndrea Arcangeli /* 20562c653d0eSAndrea Arcangeli * If it's a KSM fork, allow it to go over the sharing limit 20572c653d0eSAndrea Arcangeli * without warnings. 20582c653d0eSAndrea Arcangeli */ 20592c653d0eSAndrea Arcangeli if (!is_page_sharing_candidate(stable_node)) 20602c653d0eSAndrea Arcangeli max_page_sharing_bypass = true; 20614146d2d6SHugh Dickins } 206231dbd01fSIzik Eidus 206331dbd01fSIzik Eidus /* We first start with searching the page inside the stable tree */ 206462b61f61SHugh Dickins kpage = stable_tree_search(page); 20654146d2d6SHugh Dickins if (kpage == page && rmap_item->head == stable_node) { 20664146d2d6SHugh Dickins put_page(kpage); 20674146d2d6SHugh Dickins return; 20684146d2d6SHugh Dickins } 20694146d2d6SHugh Dickins 20704146d2d6SHugh Dickins remove_rmap_item_from_tree(rmap_item); 20714146d2d6SHugh Dickins 207262b61f61SHugh Dickins if (kpage) { 207308beca44SHugh Dickins err = try_to_merge_with_ksm_page(rmap_item, page, kpage); 207431dbd01fSIzik Eidus if (!err) { 207531dbd01fSIzik Eidus /* 207631dbd01fSIzik Eidus * The page was successfully merged: 207731dbd01fSIzik Eidus * add its rmap_item to the stable tree. 207831dbd01fSIzik Eidus */ 20795ad64688SHugh Dickins lock_page(kpage); 20802c653d0eSAndrea Arcangeli stable_tree_append(rmap_item, page_stable_node(kpage), 20812c653d0eSAndrea Arcangeli max_page_sharing_bypass); 20825ad64688SHugh Dickins unlock_page(kpage); 208331dbd01fSIzik Eidus } 20848dd3557aSHugh Dickins put_page(kpage); 208531dbd01fSIzik Eidus return; 208631dbd01fSIzik Eidus } 208731dbd01fSIzik Eidus 208831dbd01fSIzik Eidus /* 20894035c07aSHugh Dickins * If the hash value of the page has changed from the last time 20904035c07aSHugh Dickins * we calculated it, this page is changing frequently: therefore we 20914035c07aSHugh Dickins * don't want to insert it in the unstable tree, and we don't want 20924035c07aSHugh Dickins * to waste our time searching for something identical to it there. 209331dbd01fSIzik Eidus */ 209431dbd01fSIzik Eidus checksum = calc_checksum(page); 209531dbd01fSIzik Eidus if (rmap_item->oldchecksum != checksum) { 209631dbd01fSIzik Eidus rmap_item->oldchecksum = checksum; 209731dbd01fSIzik Eidus return; 209831dbd01fSIzik Eidus } 209931dbd01fSIzik Eidus 2100e86c59b1SClaudio Imbrenda /* 2101e86c59b1SClaudio Imbrenda * Same checksum as an empty page. We attempt to merge it with the 2102e86c59b1SClaudio Imbrenda * appropriate zero page if the user enabled this via sysfs. 2103e86c59b1SClaudio Imbrenda */ 2104e86c59b1SClaudio Imbrenda if (ksm_use_zero_pages && (checksum == zero_checksum)) { 2105e86c59b1SClaudio Imbrenda struct vm_area_struct *vma; 2106e86c59b1SClaudio Imbrenda 21074b22927fSKirill Tkhai down_read(&mm->mmap_sem); 21084b22927fSKirill Tkhai vma = find_mergeable_vma(mm, rmap_item->address); 2109e86c59b1SClaudio Imbrenda err = try_to_merge_one_page(vma, page, 2110e86c59b1SClaudio Imbrenda ZERO_PAGE(rmap_item->address)); 21114b22927fSKirill Tkhai up_read(&mm->mmap_sem); 2112e86c59b1SClaudio Imbrenda /* 2113e86c59b1SClaudio Imbrenda * In case of failure, the page was not really empty, so we 2114e86c59b1SClaudio Imbrenda * need to continue. Otherwise we're done. 2115e86c59b1SClaudio Imbrenda */ 2116e86c59b1SClaudio Imbrenda if (!err) 2117e86c59b1SClaudio Imbrenda return; 2118e86c59b1SClaudio Imbrenda } 21198dd3557aSHugh Dickins tree_rmap_item = 21208dd3557aSHugh Dickins unstable_tree_search_insert(rmap_item, page, &tree_page); 212131dbd01fSIzik Eidus if (tree_rmap_item) { 212277da2ba0SClaudio Imbrenda bool split; 212377da2ba0SClaudio Imbrenda 21248dd3557aSHugh Dickins kpage = try_to_merge_two_pages(rmap_item, page, 21258dd3557aSHugh Dickins tree_rmap_item, tree_page); 212677da2ba0SClaudio Imbrenda /* 212777da2ba0SClaudio Imbrenda * If both pages we tried to merge belong to the same compound 212877da2ba0SClaudio Imbrenda * page, then we actually ended up increasing the reference 212977da2ba0SClaudio Imbrenda * count of the same compound page twice, and split_huge_page 213077da2ba0SClaudio Imbrenda * failed. 213177da2ba0SClaudio Imbrenda * Here we set a flag if that happened, and we use it later to 213277da2ba0SClaudio Imbrenda * try split_huge_page again. Since we call put_page right 213377da2ba0SClaudio Imbrenda * afterwards, the reference count will be correct and 213477da2ba0SClaudio Imbrenda * split_huge_page should succeed. 213577da2ba0SClaudio Imbrenda */ 213677da2ba0SClaudio Imbrenda split = PageTransCompound(page) 213777da2ba0SClaudio Imbrenda && compound_head(page) == compound_head(tree_page); 21388dd3557aSHugh Dickins put_page(tree_page); 21398dd3557aSHugh Dickins if (kpage) { 2140bc56620bSHugh Dickins /* 2141bc56620bSHugh Dickins * The pages were successfully merged: insert new 2142bc56620bSHugh Dickins * node in the stable tree and add both rmap_items. 2143bc56620bSHugh Dickins */ 21445ad64688SHugh Dickins lock_page(kpage); 21457b6ba2c7SHugh Dickins stable_node = stable_tree_insert(kpage); 21467b6ba2c7SHugh Dickins if (stable_node) { 21472c653d0eSAndrea Arcangeli stable_tree_append(tree_rmap_item, stable_node, 21482c653d0eSAndrea Arcangeli false); 21492c653d0eSAndrea Arcangeli stable_tree_append(rmap_item, stable_node, 21502c653d0eSAndrea Arcangeli false); 21517b6ba2c7SHugh Dickins } 21525ad64688SHugh Dickins unlock_page(kpage); 21537b6ba2c7SHugh Dickins 215431dbd01fSIzik Eidus /* 215531dbd01fSIzik Eidus * If we fail to insert the page into the stable tree, 215631dbd01fSIzik Eidus * we will have 2 virtual addresses that are pointing 215731dbd01fSIzik Eidus * to a ksm page left outside the stable tree, 215831dbd01fSIzik Eidus * in which case we need to break_cow on both. 215931dbd01fSIzik Eidus */ 21607b6ba2c7SHugh Dickins if (!stable_node) { 21618dd3557aSHugh Dickins break_cow(tree_rmap_item); 21628dd3557aSHugh Dickins break_cow(rmap_item); 216331dbd01fSIzik Eidus } 216477da2ba0SClaudio Imbrenda } else if (split) { 216577da2ba0SClaudio Imbrenda /* 216677da2ba0SClaudio Imbrenda * We are here if we tried to merge two pages and 216777da2ba0SClaudio Imbrenda * failed because they both belonged to the same 216877da2ba0SClaudio Imbrenda * compound page. We will split the page now, but no 216977da2ba0SClaudio Imbrenda * merging will take place. 217077da2ba0SClaudio Imbrenda * We do not want to add the cost of a full lock; if 217177da2ba0SClaudio Imbrenda * the page is locked, it is better to skip it and 217277da2ba0SClaudio Imbrenda * perhaps try again later. 217377da2ba0SClaudio Imbrenda */ 217477da2ba0SClaudio Imbrenda if (!trylock_page(page)) 217577da2ba0SClaudio Imbrenda return; 217677da2ba0SClaudio Imbrenda split_huge_page(page); 217777da2ba0SClaudio Imbrenda unlock_page(page); 217831dbd01fSIzik Eidus } 217931dbd01fSIzik Eidus } 218031dbd01fSIzik Eidus } 218131dbd01fSIzik Eidus 218231dbd01fSIzik Eidus static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot, 21836514d511SHugh Dickins struct rmap_item **rmap_list, 218431dbd01fSIzik Eidus unsigned long addr) 218531dbd01fSIzik Eidus { 218631dbd01fSIzik Eidus struct rmap_item *rmap_item; 218731dbd01fSIzik Eidus 21886514d511SHugh Dickins while (*rmap_list) { 21896514d511SHugh Dickins rmap_item = *rmap_list; 219093d17715SHugh Dickins if ((rmap_item->address & PAGE_MASK) == addr) 219131dbd01fSIzik Eidus return rmap_item; 219231dbd01fSIzik Eidus if (rmap_item->address > addr) 219331dbd01fSIzik Eidus break; 21946514d511SHugh Dickins *rmap_list = rmap_item->rmap_list; 219531dbd01fSIzik Eidus remove_rmap_item_from_tree(rmap_item); 219631dbd01fSIzik Eidus free_rmap_item(rmap_item); 219731dbd01fSIzik Eidus } 219831dbd01fSIzik Eidus 219931dbd01fSIzik Eidus rmap_item = alloc_rmap_item(); 220031dbd01fSIzik Eidus if (rmap_item) { 220131dbd01fSIzik Eidus /* It has already been zeroed */ 220231dbd01fSIzik Eidus rmap_item->mm = mm_slot->mm; 220331dbd01fSIzik Eidus rmap_item->address = addr; 22046514d511SHugh Dickins rmap_item->rmap_list = *rmap_list; 22056514d511SHugh Dickins *rmap_list = rmap_item; 220631dbd01fSIzik Eidus } 220731dbd01fSIzik Eidus return rmap_item; 220831dbd01fSIzik Eidus } 220931dbd01fSIzik Eidus 221031dbd01fSIzik Eidus static struct rmap_item *scan_get_next_rmap_item(struct page **page) 221131dbd01fSIzik Eidus { 221231dbd01fSIzik Eidus struct mm_struct *mm; 221331dbd01fSIzik Eidus struct mm_slot *slot; 221431dbd01fSIzik Eidus struct vm_area_struct *vma; 221531dbd01fSIzik Eidus struct rmap_item *rmap_item; 221690bd6fd3SPetr Holasek int nid; 221731dbd01fSIzik Eidus 221831dbd01fSIzik Eidus if (list_empty(&ksm_mm_head.mm_list)) 221931dbd01fSIzik Eidus return NULL; 222031dbd01fSIzik Eidus 222131dbd01fSIzik Eidus slot = ksm_scan.mm_slot; 222231dbd01fSIzik Eidus if (slot == &ksm_mm_head) { 22232919bfd0SHugh Dickins /* 22242919bfd0SHugh Dickins * A number of pages can hang around indefinitely on per-cpu 22252919bfd0SHugh Dickins * pagevecs, raised page count preventing write_protect_page 22262919bfd0SHugh Dickins * from merging them. Though it doesn't really matter much, 22272919bfd0SHugh Dickins * it is puzzling to see some stuck in pages_volatile until 22282919bfd0SHugh Dickins * other activity jostles them out, and they also prevented 22292919bfd0SHugh Dickins * LTP's KSM test from succeeding deterministically; so drain 22302919bfd0SHugh Dickins * them here (here rather than on entry to ksm_do_scan(), 22312919bfd0SHugh Dickins * so we don't IPI too often when pages_to_scan is set low). 22322919bfd0SHugh Dickins */ 22332919bfd0SHugh Dickins lru_add_drain_all(); 22342919bfd0SHugh Dickins 22354146d2d6SHugh Dickins /* 22364146d2d6SHugh Dickins * Whereas stale stable_nodes on the stable_tree itself 22374146d2d6SHugh Dickins * get pruned in the regular course of stable_tree_search(), 22384146d2d6SHugh Dickins * those moved out to the migrate_nodes list can accumulate: 22394146d2d6SHugh Dickins * so prune them once before each full scan. 22404146d2d6SHugh Dickins */ 22414146d2d6SHugh Dickins if (!ksm_merge_across_nodes) { 224203640418SGeliang Tang struct stable_node *stable_node, *next; 22434146d2d6SHugh Dickins struct page *page; 22444146d2d6SHugh Dickins 224503640418SGeliang Tang list_for_each_entry_safe(stable_node, next, 224603640418SGeliang Tang &migrate_nodes, list) { 22474146d2d6SHugh Dickins page = get_ksm_page(stable_node, false); 22484146d2d6SHugh Dickins if (page) 22494146d2d6SHugh Dickins put_page(page); 22504146d2d6SHugh Dickins cond_resched(); 22514146d2d6SHugh Dickins } 22524146d2d6SHugh Dickins } 22534146d2d6SHugh Dickins 2254ef53d16cSHugh Dickins for (nid = 0; nid < ksm_nr_node_ids; nid++) 225590bd6fd3SPetr Holasek root_unstable_tree[nid] = RB_ROOT; 225631dbd01fSIzik Eidus 225731dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 225831dbd01fSIzik Eidus slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list); 225931dbd01fSIzik Eidus ksm_scan.mm_slot = slot; 226031dbd01fSIzik Eidus spin_unlock(&ksm_mmlist_lock); 22612b472611SHugh Dickins /* 22622b472611SHugh Dickins * Although we tested list_empty() above, a racing __ksm_exit 22632b472611SHugh Dickins * of the last mm on the list may have removed it since then. 22642b472611SHugh Dickins */ 22652b472611SHugh Dickins if (slot == &ksm_mm_head) 22662b472611SHugh Dickins return NULL; 226731dbd01fSIzik Eidus next_mm: 226831dbd01fSIzik Eidus ksm_scan.address = 0; 22696514d511SHugh Dickins ksm_scan.rmap_list = &slot->rmap_list; 227031dbd01fSIzik Eidus } 227131dbd01fSIzik Eidus 227231dbd01fSIzik Eidus mm = slot->mm; 227331dbd01fSIzik Eidus down_read(&mm->mmap_sem); 22749ba69294SHugh Dickins if (ksm_test_exit(mm)) 22759ba69294SHugh Dickins vma = NULL; 22769ba69294SHugh Dickins else 22779ba69294SHugh Dickins vma = find_vma(mm, ksm_scan.address); 22789ba69294SHugh Dickins 22799ba69294SHugh Dickins for (; vma; vma = vma->vm_next) { 228031dbd01fSIzik Eidus if (!(vma->vm_flags & VM_MERGEABLE)) 228131dbd01fSIzik Eidus continue; 228231dbd01fSIzik Eidus if (ksm_scan.address < vma->vm_start) 228331dbd01fSIzik Eidus ksm_scan.address = vma->vm_start; 228431dbd01fSIzik Eidus if (!vma->anon_vma) 228531dbd01fSIzik Eidus ksm_scan.address = vma->vm_end; 228631dbd01fSIzik Eidus 228731dbd01fSIzik Eidus while (ksm_scan.address < vma->vm_end) { 22889ba69294SHugh Dickins if (ksm_test_exit(mm)) 22899ba69294SHugh Dickins break; 229031dbd01fSIzik Eidus *page = follow_page(vma, ksm_scan.address, FOLL_GET); 229121ae5b01SAndrea Arcangeli if (IS_ERR_OR_NULL(*page)) { 229221ae5b01SAndrea Arcangeli ksm_scan.address += PAGE_SIZE; 229321ae5b01SAndrea Arcangeli cond_resched(); 229421ae5b01SAndrea Arcangeli continue; 229521ae5b01SAndrea Arcangeli } 2296f765f540SKirill A. Shutemov if (PageAnon(*page)) { 229731dbd01fSIzik Eidus flush_anon_page(vma, *page, ksm_scan.address); 229831dbd01fSIzik Eidus flush_dcache_page(*page); 229931dbd01fSIzik Eidus rmap_item = get_next_rmap_item(slot, 23006514d511SHugh Dickins ksm_scan.rmap_list, ksm_scan.address); 230131dbd01fSIzik Eidus if (rmap_item) { 23026514d511SHugh Dickins ksm_scan.rmap_list = 23036514d511SHugh Dickins &rmap_item->rmap_list; 230431dbd01fSIzik Eidus ksm_scan.address += PAGE_SIZE; 230531dbd01fSIzik Eidus } else 230631dbd01fSIzik Eidus put_page(*page); 230731dbd01fSIzik Eidus up_read(&mm->mmap_sem); 230831dbd01fSIzik Eidus return rmap_item; 230931dbd01fSIzik Eidus } 231031dbd01fSIzik Eidus put_page(*page); 231131dbd01fSIzik Eidus ksm_scan.address += PAGE_SIZE; 231231dbd01fSIzik Eidus cond_resched(); 231331dbd01fSIzik Eidus } 231431dbd01fSIzik Eidus } 231531dbd01fSIzik Eidus 23169ba69294SHugh Dickins if (ksm_test_exit(mm)) { 23179ba69294SHugh Dickins ksm_scan.address = 0; 23186514d511SHugh Dickins ksm_scan.rmap_list = &slot->rmap_list; 23199ba69294SHugh Dickins } 232031dbd01fSIzik Eidus /* 232131dbd01fSIzik Eidus * Nuke all the rmap_items that are above this current rmap: 232231dbd01fSIzik Eidus * because there were no VM_MERGEABLE vmas with such addresses. 232331dbd01fSIzik Eidus */ 23246514d511SHugh Dickins remove_trailing_rmap_items(slot, ksm_scan.rmap_list); 232531dbd01fSIzik Eidus 232631dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 2327cd551f97SHugh Dickins ksm_scan.mm_slot = list_entry(slot->mm_list.next, 2328cd551f97SHugh Dickins struct mm_slot, mm_list); 2329cd551f97SHugh Dickins if (ksm_scan.address == 0) { 2330cd551f97SHugh Dickins /* 2331cd551f97SHugh Dickins * We've completed a full scan of all vmas, holding mmap_sem 2332cd551f97SHugh Dickins * throughout, and found no VM_MERGEABLE: so do the same as 2333cd551f97SHugh Dickins * __ksm_exit does to remove this mm from all our lists now. 23349ba69294SHugh Dickins * This applies either when cleaning up after __ksm_exit 23359ba69294SHugh Dickins * (but beware: we can reach here even before __ksm_exit), 23369ba69294SHugh Dickins * or when all VM_MERGEABLE areas have been unmapped (and 23379ba69294SHugh Dickins * mmap_sem then protects against race with MADV_MERGEABLE). 2338cd551f97SHugh Dickins */ 23394ca3a69bSSasha Levin hash_del(&slot->link); 2340cd551f97SHugh Dickins list_del(&slot->mm_list); 23419ba69294SHugh Dickins spin_unlock(&ksm_mmlist_lock); 23429ba69294SHugh Dickins 2343cd551f97SHugh Dickins free_mm_slot(slot); 2344cd551f97SHugh Dickins clear_bit(MMF_VM_MERGEABLE, &mm->flags); 23459ba69294SHugh Dickins up_read(&mm->mmap_sem); 23469ba69294SHugh Dickins mmdrop(mm); 23479ba69294SHugh Dickins } else { 2348cd551f97SHugh Dickins up_read(&mm->mmap_sem); 23497496fea9SZhou Chengming /* 23507496fea9SZhou Chengming * up_read(&mm->mmap_sem) first because after 23517496fea9SZhou Chengming * spin_unlock(&ksm_mmlist_lock) run, the "mm" may 23527496fea9SZhou Chengming * already have been freed under us by __ksm_exit() 23537496fea9SZhou Chengming * because the "mm_slot" is still hashed and 23547496fea9SZhou Chengming * ksm_scan.mm_slot doesn't point to it anymore. 23557496fea9SZhou Chengming */ 23567496fea9SZhou Chengming spin_unlock(&ksm_mmlist_lock); 23579ba69294SHugh Dickins } 235831dbd01fSIzik Eidus 235931dbd01fSIzik Eidus /* Repeat until we've completed scanning the whole list */ 2360cd551f97SHugh Dickins slot = ksm_scan.mm_slot; 236131dbd01fSIzik Eidus if (slot != &ksm_mm_head) 236231dbd01fSIzik Eidus goto next_mm; 236331dbd01fSIzik Eidus 236431dbd01fSIzik Eidus ksm_scan.seqnr++; 236531dbd01fSIzik Eidus return NULL; 236631dbd01fSIzik Eidus } 236731dbd01fSIzik Eidus 236831dbd01fSIzik Eidus /** 236931dbd01fSIzik Eidus * ksm_do_scan - the ksm scanner main worker function. 2370b7701a5fSMike Rapoport * @scan_npages: number of pages we want to scan before we return. 237131dbd01fSIzik Eidus */ 237231dbd01fSIzik Eidus static void ksm_do_scan(unsigned int scan_npages) 237331dbd01fSIzik Eidus { 237431dbd01fSIzik Eidus struct rmap_item *rmap_item; 237522eccdd7SDan Carpenter struct page *uninitialized_var(page); 237631dbd01fSIzik Eidus 2377878aee7dSAndrea Arcangeli while (scan_npages-- && likely(!freezing(current))) { 237831dbd01fSIzik Eidus cond_resched(); 237931dbd01fSIzik Eidus rmap_item = scan_get_next_rmap_item(&page); 238031dbd01fSIzik Eidus if (!rmap_item) 238131dbd01fSIzik Eidus return; 238231dbd01fSIzik Eidus cmp_and_merge_page(page, rmap_item); 238331dbd01fSIzik Eidus put_page(page); 238431dbd01fSIzik Eidus } 238531dbd01fSIzik Eidus } 238631dbd01fSIzik Eidus 23876e158384SHugh Dickins static int ksmd_should_run(void) 23886e158384SHugh Dickins { 23896e158384SHugh Dickins return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list); 23906e158384SHugh Dickins } 23916e158384SHugh Dickins 239231dbd01fSIzik Eidus static int ksm_scan_thread(void *nothing) 239331dbd01fSIzik Eidus { 2394878aee7dSAndrea Arcangeli set_freezable(); 2395339aa624SIzik Eidus set_user_nice(current, 5); 239631dbd01fSIzik Eidus 239731dbd01fSIzik Eidus while (!kthread_should_stop()) { 239831dbd01fSIzik Eidus mutex_lock(&ksm_thread_mutex); 2399ef4d43a8SHugh Dickins wait_while_offlining(); 24006e158384SHugh Dickins if (ksmd_should_run()) 240131dbd01fSIzik Eidus ksm_do_scan(ksm_thread_pages_to_scan); 240231dbd01fSIzik Eidus mutex_unlock(&ksm_thread_mutex); 24036e158384SHugh Dickins 2404878aee7dSAndrea Arcangeli try_to_freeze(); 2405878aee7dSAndrea Arcangeli 24066e158384SHugh Dickins if (ksmd_should_run()) { 240731dbd01fSIzik Eidus schedule_timeout_interruptible( 240831dbd01fSIzik Eidus msecs_to_jiffies(ksm_thread_sleep_millisecs)); 240931dbd01fSIzik Eidus } else { 2410878aee7dSAndrea Arcangeli wait_event_freezable(ksm_thread_wait, 24116e158384SHugh Dickins ksmd_should_run() || kthread_should_stop()); 241231dbd01fSIzik Eidus } 241331dbd01fSIzik Eidus } 241431dbd01fSIzik Eidus return 0; 241531dbd01fSIzik Eidus } 241631dbd01fSIzik Eidus 2417f8af4da3SHugh Dickins int ksm_madvise(struct vm_area_struct *vma, unsigned long start, 2418f8af4da3SHugh Dickins unsigned long end, int advice, unsigned long *vm_flags) 2419f8af4da3SHugh Dickins { 2420f8af4da3SHugh Dickins struct mm_struct *mm = vma->vm_mm; 2421d952b791SHugh Dickins int err; 2422f8af4da3SHugh Dickins 2423f8af4da3SHugh Dickins switch (advice) { 2424f8af4da3SHugh Dickins case MADV_MERGEABLE: 2425f8af4da3SHugh Dickins /* 2426f8af4da3SHugh Dickins * Be somewhat over-protective for now! 2427f8af4da3SHugh Dickins */ 2428f8af4da3SHugh Dickins if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE | 2429f8af4da3SHugh Dickins VM_PFNMAP | VM_IO | VM_DONTEXPAND | 24300661a336SKirill A. Shutemov VM_HUGETLB | VM_MIXEDMAP)) 2431f8af4da3SHugh Dickins return 0; /* just ignore the advice */ 2432f8af4da3SHugh Dickins 2433*e1fb4a08SDave Jiang if (vma_is_dax(vma)) 2434*e1fb4a08SDave Jiang return 0; 2435*e1fb4a08SDave Jiang 2436cc2383ecSKonstantin Khlebnikov #ifdef VM_SAO 2437cc2383ecSKonstantin Khlebnikov if (*vm_flags & VM_SAO) 2438cc2383ecSKonstantin Khlebnikov return 0; 2439cc2383ecSKonstantin Khlebnikov #endif 244074a04967SKhalid Aziz #ifdef VM_SPARC_ADI 244174a04967SKhalid Aziz if (*vm_flags & VM_SPARC_ADI) 244274a04967SKhalid Aziz return 0; 244374a04967SKhalid Aziz #endif 2444cc2383ecSKonstantin Khlebnikov 2445d952b791SHugh Dickins if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { 2446d952b791SHugh Dickins err = __ksm_enter(mm); 2447d952b791SHugh Dickins if (err) 2448d952b791SHugh Dickins return err; 2449d952b791SHugh Dickins } 2450f8af4da3SHugh Dickins 2451f8af4da3SHugh Dickins *vm_flags |= VM_MERGEABLE; 2452f8af4da3SHugh Dickins break; 2453f8af4da3SHugh Dickins 2454f8af4da3SHugh Dickins case MADV_UNMERGEABLE: 2455f8af4da3SHugh Dickins if (!(*vm_flags & VM_MERGEABLE)) 2456f8af4da3SHugh Dickins return 0; /* just ignore the advice */ 2457f8af4da3SHugh Dickins 2458d952b791SHugh Dickins if (vma->anon_vma) { 2459d952b791SHugh Dickins err = unmerge_ksm_pages(vma, start, end); 2460d952b791SHugh Dickins if (err) 2461d952b791SHugh Dickins return err; 2462d952b791SHugh Dickins } 2463f8af4da3SHugh Dickins 2464f8af4da3SHugh Dickins *vm_flags &= ~VM_MERGEABLE; 2465f8af4da3SHugh Dickins break; 2466f8af4da3SHugh Dickins } 2467f8af4da3SHugh Dickins 2468f8af4da3SHugh Dickins return 0; 2469f8af4da3SHugh Dickins } 2470f8af4da3SHugh Dickins 2471f8af4da3SHugh Dickins int __ksm_enter(struct mm_struct *mm) 2472f8af4da3SHugh Dickins { 24736e158384SHugh Dickins struct mm_slot *mm_slot; 24746e158384SHugh Dickins int needs_wakeup; 24756e158384SHugh Dickins 24766e158384SHugh Dickins mm_slot = alloc_mm_slot(); 247731dbd01fSIzik Eidus if (!mm_slot) 247831dbd01fSIzik Eidus return -ENOMEM; 247931dbd01fSIzik Eidus 24806e158384SHugh Dickins /* Check ksm_run too? Would need tighter locking */ 24816e158384SHugh Dickins needs_wakeup = list_empty(&ksm_mm_head.mm_list); 24826e158384SHugh Dickins 248331dbd01fSIzik Eidus spin_lock(&ksm_mmlist_lock); 248431dbd01fSIzik Eidus insert_to_mm_slots_hash(mm, mm_slot); 248531dbd01fSIzik Eidus /* 2486cbf86cfeSHugh Dickins * When KSM_RUN_MERGE (or KSM_RUN_STOP), 2487cbf86cfeSHugh Dickins * insert just behind the scanning cursor, to let the area settle 248831dbd01fSIzik Eidus * down a little; when fork is followed by immediate exec, we don't 248931dbd01fSIzik Eidus * want ksmd to waste time setting up and tearing down an rmap_list. 2490cbf86cfeSHugh Dickins * 2491cbf86cfeSHugh Dickins * But when KSM_RUN_UNMERGE, it's important to insert ahead of its 2492cbf86cfeSHugh Dickins * scanning cursor, otherwise KSM pages in newly forked mms will be 2493cbf86cfeSHugh Dickins * missed: then we might as well insert at the end of the list. 249431dbd01fSIzik Eidus */ 2495cbf86cfeSHugh Dickins if (ksm_run & KSM_RUN_UNMERGE) 2496cbf86cfeSHugh Dickins list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list); 2497cbf86cfeSHugh Dickins else 249831dbd01fSIzik Eidus list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list); 249931dbd01fSIzik Eidus spin_unlock(&ksm_mmlist_lock); 250031dbd01fSIzik Eidus 2501f8af4da3SHugh Dickins set_bit(MMF_VM_MERGEABLE, &mm->flags); 2502f1f10076SVegard Nossum mmgrab(mm); 25036e158384SHugh Dickins 25046e158384SHugh Dickins if (needs_wakeup) 25056e158384SHugh Dickins wake_up_interruptible(&ksm_thread_wait); 25066e158384SHugh Dickins 2507f8af4da3SHugh Dickins return 0; 2508f8af4da3SHugh Dickins } 2509f8af4da3SHugh Dickins 25101c2fb7a4SAndrea Arcangeli void __ksm_exit(struct mm_struct *mm) 2511f8af4da3SHugh Dickins { 2512cd551f97SHugh Dickins struct mm_slot *mm_slot; 25139ba69294SHugh Dickins int easy_to_free = 0; 2514cd551f97SHugh Dickins 251531dbd01fSIzik Eidus /* 25169ba69294SHugh Dickins * This process is exiting: if it's straightforward (as is the 25179ba69294SHugh Dickins * case when ksmd was never running), free mm_slot immediately. 25189ba69294SHugh Dickins * But if it's at the cursor or has rmap_items linked to it, use 25199ba69294SHugh Dickins * mmap_sem to synchronize with any break_cows before pagetables 25209ba69294SHugh Dickins * are freed, and leave the mm_slot on the list for ksmd to free. 25219ba69294SHugh Dickins * Beware: ksm may already have noticed it exiting and freed the slot. 252231dbd01fSIzik Eidus */ 25239ba69294SHugh Dickins 2524cd551f97SHugh Dickins spin_lock(&ksm_mmlist_lock); 2525cd551f97SHugh Dickins mm_slot = get_mm_slot(mm); 25269ba69294SHugh Dickins if (mm_slot && ksm_scan.mm_slot != mm_slot) { 25276514d511SHugh Dickins if (!mm_slot->rmap_list) { 25284ca3a69bSSasha Levin hash_del(&mm_slot->link); 2529cd551f97SHugh Dickins list_del(&mm_slot->mm_list); 25309ba69294SHugh Dickins easy_to_free = 1; 25319ba69294SHugh Dickins } else { 25329ba69294SHugh Dickins list_move(&mm_slot->mm_list, 25339ba69294SHugh Dickins &ksm_scan.mm_slot->mm_list); 25349ba69294SHugh Dickins } 25359ba69294SHugh Dickins } 2536cd551f97SHugh Dickins spin_unlock(&ksm_mmlist_lock); 2537cd551f97SHugh Dickins 25389ba69294SHugh Dickins if (easy_to_free) { 2539cd551f97SHugh Dickins free_mm_slot(mm_slot); 2540cd551f97SHugh Dickins clear_bit(MMF_VM_MERGEABLE, &mm->flags); 25419ba69294SHugh Dickins mmdrop(mm); 25429ba69294SHugh Dickins } else if (mm_slot) { 25439ba69294SHugh Dickins down_write(&mm->mmap_sem); 25449ba69294SHugh Dickins up_write(&mm->mmap_sem); 25459ba69294SHugh Dickins } 2546f8af4da3SHugh Dickins } 254731dbd01fSIzik Eidus 2548cbf86cfeSHugh Dickins struct page *ksm_might_need_to_copy(struct page *page, 25495ad64688SHugh Dickins struct vm_area_struct *vma, unsigned long address) 25505ad64688SHugh Dickins { 2551cbf86cfeSHugh Dickins struct anon_vma *anon_vma = page_anon_vma(page); 25525ad64688SHugh Dickins struct page *new_page; 25535ad64688SHugh Dickins 2554cbf86cfeSHugh Dickins if (PageKsm(page)) { 2555cbf86cfeSHugh Dickins if (page_stable_node(page) && 2556cbf86cfeSHugh Dickins !(ksm_run & KSM_RUN_UNMERGE)) 2557cbf86cfeSHugh Dickins return page; /* no need to copy it */ 2558cbf86cfeSHugh Dickins } else if (!anon_vma) { 2559cbf86cfeSHugh Dickins return page; /* no need to copy it */ 2560cbf86cfeSHugh Dickins } else if (anon_vma->root == vma->anon_vma->root && 2561cbf86cfeSHugh Dickins page->index == linear_page_index(vma, address)) { 2562cbf86cfeSHugh Dickins return page; /* still no need to copy it */ 2563cbf86cfeSHugh Dickins } 2564cbf86cfeSHugh Dickins if (!PageUptodate(page)) 2565cbf86cfeSHugh Dickins return page; /* let do_swap_page report the error */ 2566cbf86cfeSHugh Dickins 25675ad64688SHugh Dickins new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); 25685ad64688SHugh Dickins if (new_page) { 25695ad64688SHugh Dickins copy_user_highpage(new_page, page, address, vma); 25705ad64688SHugh Dickins 25715ad64688SHugh Dickins SetPageDirty(new_page); 25725ad64688SHugh Dickins __SetPageUptodate(new_page); 257348c935adSKirill A. Shutemov __SetPageLocked(new_page); 25745ad64688SHugh Dickins } 25755ad64688SHugh Dickins 25765ad64688SHugh Dickins return new_page; 25775ad64688SHugh Dickins } 25785ad64688SHugh Dickins 25791df631aeSMinchan Kim void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc) 2580e9995ef9SHugh Dickins { 2581e9995ef9SHugh Dickins struct stable_node *stable_node; 2582e9995ef9SHugh Dickins struct rmap_item *rmap_item; 2583e9995ef9SHugh Dickins int search_new_forks = 0; 2584e9995ef9SHugh Dickins 2585309381feSSasha Levin VM_BUG_ON_PAGE(!PageKsm(page), page); 25869f32624bSJoonsoo Kim 25879f32624bSJoonsoo Kim /* 25889f32624bSJoonsoo Kim * Rely on the page lock to protect against concurrent modifications 25899f32624bSJoonsoo Kim * to that page's node of the stable tree. 25909f32624bSJoonsoo Kim */ 2591309381feSSasha Levin VM_BUG_ON_PAGE(!PageLocked(page), page); 2592e9995ef9SHugh Dickins 2593e9995ef9SHugh Dickins stable_node = page_stable_node(page); 2594e9995ef9SHugh Dickins if (!stable_node) 25951df631aeSMinchan Kim return; 2596e9995ef9SHugh Dickins again: 2597b67bfe0dSSasha Levin hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { 2598e9995ef9SHugh Dickins struct anon_vma *anon_vma = rmap_item->anon_vma; 25995beb4930SRik van Riel struct anon_vma_chain *vmac; 2600e9995ef9SHugh Dickins struct vm_area_struct *vma; 2601e9995ef9SHugh Dickins 2602ad12695fSAndrea Arcangeli cond_resched(); 2603b6b19f25SHugh Dickins anon_vma_lock_read(anon_vma); 2604bf181b9fSMichel Lespinasse anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, 2605bf181b9fSMichel Lespinasse 0, ULONG_MAX) { 26061105a2fcSJia He unsigned long addr; 26071105a2fcSJia He 2608ad12695fSAndrea Arcangeli cond_resched(); 26095beb4930SRik van Riel vma = vmac->vma; 26101105a2fcSJia He 26111105a2fcSJia He /* Ignore the stable/unstable/sqnr flags */ 26121105a2fcSJia He addr = rmap_item->address & ~KSM_FLAG_MASK; 26131105a2fcSJia He 26141105a2fcSJia He if (addr < vma->vm_start || addr >= vma->vm_end) 2615e9995ef9SHugh Dickins continue; 2616e9995ef9SHugh Dickins /* 2617e9995ef9SHugh Dickins * Initially we examine only the vma which covers this 2618e9995ef9SHugh Dickins * rmap_item; but later, if there is still work to do, 2619e9995ef9SHugh Dickins * we examine covering vmas in other mms: in case they 2620e9995ef9SHugh Dickins * were forked from the original since ksmd passed. 2621e9995ef9SHugh Dickins */ 2622e9995ef9SHugh Dickins if ((rmap_item->mm == vma->vm_mm) == search_new_forks) 2623e9995ef9SHugh Dickins continue; 2624e9995ef9SHugh Dickins 26250dd1c7bbSJoonsoo Kim if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) 26260dd1c7bbSJoonsoo Kim continue; 26270dd1c7bbSJoonsoo Kim 26281105a2fcSJia He if (!rwc->rmap_one(page, vma, addr, rwc->arg)) { 2629b6b19f25SHugh Dickins anon_vma_unlock_read(anon_vma); 26301df631aeSMinchan Kim return; 2631e9995ef9SHugh Dickins } 26320dd1c7bbSJoonsoo Kim if (rwc->done && rwc->done(page)) { 26330dd1c7bbSJoonsoo Kim anon_vma_unlock_read(anon_vma); 26341df631aeSMinchan Kim return; 26350dd1c7bbSJoonsoo Kim } 2636e9995ef9SHugh Dickins } 2637b6b19f25SHugh Dickins anon_vma_unlock_read(anon_vma); 2638e9995ef9SHugh Dickins } 2639e9995ef9SHugh Dickins if (!search_new_forks++) 2640e9995ef9SHugh Dickins goto again; 2641e9995ef9SHugh Dickins } 2642e9995ef9SHugh Dickins 264352629506SJoonsoo Kim #ifdef CONFIG_MIGRATION 2644e9995ef9SHugh Dickins void ksm_migrate_page(struct page *newpage, struct page *oldpage) 2645e9995ef9SHugh Dickins { 2646e9995ef9SHugh Dickins struct stable_node *stable_node; 2647e9995ef9SHugh Dickins 2648309381feSSasha Levin VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage); 2649309381feSSasha Levin VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); 2650309381feSSasha Levin VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage); 2651e9995ef9SHugh Dickins 2652e9995ef9SHugh Dickins stable_node = page_stable_node(newpage); 2653e9995ef9SHugh Dickins if (stable_node) { 2654309381feSSasha Levin VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage); 265562b61f61SHugh Dickins stable_node->kpfn = page_to_pfn(newpage); 2656c8d6553bSHugh Dickins /* 2657c8d6553bSHugh Dickins * newpage->mapping was set in advance; now we need smp_wmb() 2658c8d6553bSHugh Dickins * to make sure that the new stable_node->kpfn is visible 2659c8d6553bSHugh Dickins * to get_ksm_page() before it can see that oldpage->mapping 2660c8d6553bSHugh Dickins * has gone stale (or that PageSwapCache has been cleared). 2661c8d6553bSHugh Dickins */ 2662c8d6553bSHugh Dickins smp_wmb(); 2663c8d6553bSHugh Dickins set_page_stable_node(oldpage, NULL); 2664e9995ef9SHugh Dickins } 2665e9995ef9SHugh Dickins } 2666e9995ef9SHugh Dickins #endif /* CONFIG_MIGRATION */ 2667e9995ef9SHugh Dickins 266862b61f61SHugh Dickins #ifdef CONFIG_MEMORY_HOTREMOVE 2669ef4d43a8SHugh Dickins static void wait_while_offlining(void) 2670ef4d43a8SHugh Dickins { 2671ef4d43a8SHugh Dickins while (ksm_run & KSM_RUN_OFFLINE) { 2672ef4d43a8SHugh Dickins mutex_unlock(&ksm_thread_mutex); 2673ef4d43a8SHugh Dickins wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE), 267474316201SNeilBrown TASK_UNINTERRUPTIBLE); 2675ef4d43a8SHugh Dickins mutex_lock(&ksm_thread_mutex); 2676ef4d43a8SHugh Dickins } 2677ef4d43a8SHugh Dickins } 2678ef4d43a8SHugh Dickins 26792c653d0eSAndrea Arcangeli static bool stable_node_dup_remove_range(struct stable_node *stable_node, 26802c653d0eSAndrea Arcangeli unsigned long start_pfn, 26812c653d0eSAndrea Arcangeli unsigned long end_pfn) 26822c653d0eSAndrea Arcangeli { 26832c653d0eSAndrea Arcangeli if (stable_node->kpfn >= start_pfn && 26842c653d0eSAndrea Arcangeli stable_node->kpfn < end_pfn) { 26852c653d0eSAndrea Arcangeli /* 26862c653d0eSAndrea Arcangeli * Don't get_ksm_page, page has already gone: 26872c653d0eSAndrea Arcangeli * which is why we keep kpfn instead of page* 26882c653d0eSAndrea Arcangeli */ 26892c653d0eSAndrea Arcangeli remove_node_from_stable_tree(stable_node); 26902c653d0eSAndrea Arcangeli return true; 26912c653d0eSAndrea Arcangeli } 26922c653d0eSAndrea Arcangeli return false; 26932c653d0eSAndrea Arcangeli } 26942c653d0eSAndrea Arcangeli 26952c653d0eSAndrea Arcangeli static bool stable_node_chain_remove_range(struct stable_node *stable_node, 26962c653d0eSAndrea Arcangeli unsigned long start_pfn, 26972c653d0eSAndrea Arcangeli unsigned long end_pfn, 26982c653d0eSAndrea Arcangeli struct rb_root *root) 26992c653d0eSAndrea Arcangeli { 27002c653d0eSAndrea Arcangeli struct stable_node *dup; 27012c653d0eSAndrea Arcangeli struct hlist_node *hlist_safe; 27022c653d0eSAndrea Arcangeli 27032c653d0eSAndrea Arcangeli if (!is_stable_node_chain(stable_node)) { 27042c653d0eSAndrea Arcangeli VM_BUG_ON(is_stable_node_dup(stable_node)); 27052c653d0eSAndrea Arcangeli return stable_node_dup_remove_range(stable_node, start_pfn, 27062c653d0eSAndrea Arcangeli end_pfn); 27072c653d0eSAndrea Arcangeli } 27082c653d0eSAndrea Arcangeli 27092c653d0eSAndrea Arcangeli hlist_for_each_entry_safe(dup, hlist_safe, 27102c653d0eSAndrea Arcangeli &stable_node->hlist, hlist_dup) { 27112c653d0eSAndrea Arcangeli VM_BUG_ON(!is_stable_node_dup(dup)); 27122c653d0eSAndrea Arcangeli stable_node_dup_remove_range(dup, start_pfn, end_pfn); 27132c653d0eSAndrea Arcangeli } 27142c653d0eSAndrea Arcangeli if (hlist_empty(&stable_node->hlist)) { 27152c653d0eSAndrea Arcangeli free_stable_node_chain(stable_node, root); 27162c653d0eSAndrea Arcangeli return true; /* notify caller that tree was rebalanced */ 27172c653d0eSAndrea Arcangeli } else 27182c653d0eSAndrea Arcangeli return false; 27192c653d0eSAndrea Arcangeli } 27202c653d0eSAndrea Arcangeli 2721ee0ea59cSHugh Dickins static void ksm_check_stable_tree(unsigned long start_pfn, 272262b61f61SHugh Dickins unsigned long end_pfn) 272362b61f61SHugh Dickins { 272403640418SGeliang Tang struct stable_node *stable_node, *next; 272562b61f61SHugh Dickins struct rb_node *node; 272690bd6fd3SPetr Holasek int nid; 272762b61f61SHugh Dickins 2728ef53d16cSHugh Dickins for (nid = 0; nid < ksm_nr_node_ids; nid++) { 2729ef53d16cSHugh Dickins node = rb_first(root_stable_tree + nid); 2730ee0ea59cSHugh Dickins while (node) { 273162b61f61SHugh Dickins stable_node = rb_entry(node, struct stable_node, node); 27322c653d0eSAndrea Arcangeli if (stable_node_chain_remove_range(stable_node, 27332c653d0eSAndrea Arcangeli start_pfn, end_pfn, 27342c653d0eSAndrea Arcangeli root_stable_tree + 27352c653d0eSAndrea Arcangeli nid)) 2736ef53d16cSHugh Dickins node = rb_first(root_stable_tree + nid); 27372c653d0eSAndrea Arcangeli else 2738ee0ea59cSHugh Dickins node = rb_next(node); 2739ee0ea59cSHugh Dickins cond_resched(); 274062b61f61SHugh Dickins } 2741ee0ea59cSHugh Dickins } 274203640418SGeliang Tang list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { 27434146d2d6SHugh Dickins if (stable_node->kpfn >= start_pfn && 27444146d2d6SHugh Dickins stable_node->kpfn < end_pfn) 27454146d2d6SHugh Dickins remove_node_from_stable_tree(stable_node); 27464146d2d6SHugh Dickins cond_resched(); 27474146d2d6SHugh Dickins } 274862b61f61SHugh Dickins } 274962b61f61SHugh Dickins 275062b61f61SHugh Dickins static int ksm_memory_callback(struct notifier_block *self, 275162b61f61SHugh Dickins unsigned long action, void *arg) 275262b61f61SHugh Dickins { 275362b61f61SHugh Dickins struct memory_notify *mn = arg; 275462b61f61SHugh Dickins 275562b61f61SHugh Dickins switch (action) { 275662b61f61SHugh Dickins case MEM_GOING_OFFLINE: 275762b61f61SHugh Dickins /* 2758ef4d43a8SHugh Dickins * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items() 2759ef4d43a8SHugh Dickins * and remove_all_stable_nodes() while memory is going offline: 2760ef4d43a8SHugh Dickins * it is unsafe for them to touch the stable tree at this time. 2761ef4d43a8SHugh Dickins * But unmerge_ksm_pages(), rmap lookups and other entry points 2762ef4d43a8SHugh Dickins * which do not need the ksm_thread_mutex are all safe. 276362b61f61SHugh Dickins */ 2764ef4d43a8SHugh Dickins mutex_lock(&ksm_thread_mutex); 2765ef4d43a8SHugh Dickins ksm_run |= KSM_RUN_OFFLINE; 2766ef4d43a8SHugh Dickins mutex_unlock(&ksm_thread_mutex); 276762b61f61SHugh Dickins break; 276862b61f61SHugh Dickins 276962b61f61SHugh Dickins case MEM_OFFLINE: 277062b61f61SHugh Dickins /* 277162b61f61SHugh Dickins * Most of the work is done by page migration; but there might 277262b61f61SHugh Dickins * be a few stable_nodes left over, still pointing to struct 2773ee0ea59cSHugh Dickins * pages which have been offlined: prune those from the tree, 2774ee0ea59cSHugh Dickins * otherwise get_ksm_page() might later try to access a 2775ee0ea59cSHugh Dickins * non-existent struct page. 277662b61f61SHugh Dickins */ 2777ee0ea59cSHugh Dickins ksm_check_stable_tree(mn->start_pfn, 2778ee0ea59cSHugh Dickins mn->start_pfn + mn->nr_pages); 277962b61f61SHugh Dickins /* fallthrough */ 278062b61f61SHugh Dickins 278162b61f61SHugh Dickins case MEM_CANCEL_OFFLINE: 2782ef4d43a8SHugh Dickins mutex_lock(&ksm_thread_mutex); 2783ef4d43a8SHugh Dickins ksm_run &= ~KSM_RUN_OFFLINE; 278462b61f61SHugh Dickins mutex_unlock(&ksm_thread_mutex); 2785ef4d43a8SHugh Dickins 2786ef4d43a8SHugh Dickins smp_mb(); /* wake_up_bit advises this */ 2787ef4d43a8SHugh Dickins wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE)); 278862b61f61SHugh Dickins break; 278962b61f61SHugh Dickins } 279062b61f61SHugh Dickins return NOTIFY_OK; 279162b61f61SHugh Dickins } 2792ef4d43a8SHugh Dickins #else 2793ef4d43a8SHugh Dickins static void wait_while_offlining(void) 2794ef4d43a8SHugh Dickins { 2795ef4d43a8SHugh Dickins } 279662b61f61SHugh Dickins #endif /* CONFIG_MEMORY_HOTREMOVE */ 279762b61f61SHugh Dickins 27982ffd8679SHugh Dickins #ifdef CONFIG_SYSFS 27992ffd8679SHugh Dickins /* 28002ffd8679SHugh Dickins * This all compiles without CONFIG_SYSFS, but is a waste of space. 28012ffd8679SHugh Dickins */ 28022ffd8679SHugh Dickins 280331dbd01fSIzik Eidus #define KSM_ATTR_RO(_name) \ 280431dbd01fSIzik Eidus static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 280531dbd01fSIzik Eidus #define KSM_ATTR(_name) \ 280631dbd01fSIzik Eidus static struct kobj_attribute _name##_attr = \ 280731dbd01fSIzik Eidus __ATTR(_name, 0644, _name##_show, _name##_store) 280831dbd01fSIzik Eidus 280931dbd01fSIzik Eidus static ssize_t sleep_millisecs_show(struct kobject *kobj, 281031dbd01fSIzik Eidus struct kobj_attribute *attr, char *buf) 281131dbd01fSIzik Eidus { 281231dbd01fSIzik Eidus return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs); 281331dbd01fSIzik Eidus } 281431dbd01fSIzik Eidus 281531dbd01fSIzik Eidus static ssize_t sleep_millisecs_store(struct kobject *kobj, 281631dbd01fSIzik Eidus struct kobj_attribute *attr, 281731dbd01fSIzik Eidus const char *buf, size_t count) 281831dbd01fSIzik Eidus { 281931dbd01fSIzik Eidus unsigned long msecs; 282031dbd01fSIzik Eidus int err; 282131dbd01fSIzik Eidus 28223dbb95f7SJingoo Han err = kstrtoul(buf, 10, &msecs); 282331dbd01fSIzik Eidus if (err || msecs > UINT_MAX) 282431dbd01fSIzik Eidus return -EINVAL; 282531dbd01fSIzik Eidus 282631dbd01fSIzik Eidus ksm_thread_sleep_millisecs = msecs; 282731dbd01fSIzik Eidus 282831dbd01fSIzik Eidus return count; 282931dbd01fSIzik Eidus } 283031dbd01fSIzik Eidus KSM_ATTR(sleep_millisecs); 283131dbd01fSIzik Eidus 283231dbd01fSIzik Eidus static ssize_t pages_to_scan_show(struct kobject *kobj, 283331dbd01fSIzik Eidus struct kobj_attribute *attr, char *buf) 283431dbd01fSIzik Eidus { 283531dbd01fSIzik Eidus return sprintf(buf, "%u\n", ksm_thread_pages_to_scan); 283631dbd01fSIzik Eidus } 283731dbd01fSIzik Eidus 283831dbd01fSIzik Eidus static ssize_t pages_to_scan_store(struct kobject *kobj, 283931dbd01fSIzik Eidus struct kobj_attribute *attr, 284031dbd01fSIzik Eidus const char *buf, size_t count) 284131dbd01fSIzik Eidus { 284231dbd01fSIzik Eidus int err; 284331dbd01fSIzik Eidus unsigned long nr_pages; 284431dbd01fSIzik Eidus 28453dbb95f7SJingoo Han err = kstrtoul(buf, 10, &nr_pages); 284631dbd01fSIzik Eidus if (err || nr_pages > UINT_MAX) 284731dbd01fSIzik Eidus return -EINVAL; 284831dbd01fSIzik Eidus 284931dbd01fSIzik Eidus ksm_thread_pages_to_scan = nr_pages; 285031dbd01fSIzik Eidus 285131dbd01fSIzik Eidus return count; 285231dbd01fSIzik Eidus } 285331dbd01fSIzik Eidus KSM_ATTR(pages_to_scan); 285431dbd01fSIzik Eidus 285531dbd01fSIzik Eidus static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, 285631dbd01fSIzik Eidus char *buf) 285731dbd01fSIzik Eidus { 2858ef4d43a8SHugh Dickins return sprintf(buf, "%lu\n", ksm_run); 285931dbd01fSIzik Eidus } 286031dbd01fSIzik Eidus 286131dbd01fSIzik Eidus static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, 286231dbd01fSIzik Eidus const char *buf, size_t count) 286331dbd01fSIzik Eidus { 286431dbd01fSIzik Eidus int err; 286531dbd01fSIzik Eidus unsigned long flags; 286631dbd01fSIzik Eidus 28673dbb95f7SJingoo Han err = kstrtoul(buf, 10, &flags); 286831dbd01fSIzik Eidus if (err || flags > UINT_MAX) 286931dbd01fSIzik Eidus return -EINVAL; 287031dbd01fSIzik Eidus if (flags > KSM_RUN_UNMERGE) 287131dbd01fSIzik Eidus return -EINVAL; 287231dbd01fSIzik Eidus 287331dbd01fSIzik Eidus /* 287431dbd01fSIzik Eidus * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. 287531dbd01fSIzik Eidus * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, 2876d0f209f6SHugh Dickins * breaking COW to free the pages_shared (but leaves mm_slots 2877d0f209f6SHugh Dickins * on the list for when ksmd may be set running again). 287831dbd01fSIzik Eidus */ 287931dbd01fSIzik Eidus 288031dbd01fSIzik Eidus mutex_lock(&ksm_thread_mutex); 2881ef4d43a8SHugh Dickins wait_while_offlining(); 288231dbd01fSIzik Eidus if (ksm_run != flags) { 288331dbd01fSIzik Eidus ksm_run = flags; 2884d952b791SHugh Dickins if (flags & KSM_RUN_UNMERGE) { 2885e1e12d2fSDavid Rientjes set_current_oom_origin(); 2886d952b791SHugh Dickins err = unmerge_and_remove_all_rmap_items(); 2887e1e12d2fSDavid Rientjes clear_current_oom_origin(); 2888d952b791SHugh Dickins if (err) { 2889d952b791SHugh Dickins ksm_run = KSM_RUN_STOP; 2890d952b791SHugh Dickins count = err; 2891d952b791SHugh Dickins } 2892d952b791SHugh Dickins } 289331dbd01fSIzik Eidus } 289431dbd01fSIzik Eidus mutex_unlock(&ksm_thread_mutex); 289531dbd01fSIzik Eidus 289631dbd01fSIzik Eidus if (flags & KSM_RUN_MERGE) 289731dbd01fSIzik Eidus wake_up_interruptible(&ksm_thread_wait); 289831dbd01fSIzik Eidus 289931dbd01fSIzik Eidus return count; 290031dbd01fSIzik Eidus } 290131dbd01fSIzik Eidus KSM_ATTR(run); 290231dbd01fSIzik Eidus 290390bd6fd3SPetr Holasek #ifdef CONFIG_NUMA 290490bd6fd3SPetr Holasek static ssize_t merge_across_nodes_show(struct kobject *kobj, 290590bd6fd3SPetr Holasek struct kobj_attribute *attr, char *buf) 290690bd6fd3SPetr Holasek { 290790bd6fd3SPetr Holasek return sprintf(buf, "%u\n", ksm_merge_across_nodes); 290890bd6fd3SPetr Holasek } 290990bd6fd3SPetr Holasek 291090bd6fd3SPetr Holasek static ssize_t merge_across_nodes_store(struct kobject *kobj, 291190bd6fd3SPetr Holasek struct kobj_attribute *attr, 291290bd6fd3SPetr Holasek const char *buf, size_t count) 291390bd6fd3SPetr Holasek { 291490bd6fd3SPetr Holasek int err; 291590bd6fd3SPetr Holasek unsigned long knob; 291690bd6fd3SPetr Holasek 291790bd6fd3SPetr Holasek err = kstrtoul(buf, 10, &knob); 291890bd6fd3SPetr Holasek if (err) 291990bd6fd3SPetr Holasek return err; 292090bd6fd3SPetr Holasek if (knob > 1) 292190bd6fd3SPetr Holasek return -EINVAL; 292290bd6fd3SPetr Holasek 292390bd6fd3SPetr Holasek mutex_lock(&ksm_thread_mutex); 2924ef4d43a8SHugh Dickins wait_while_offlining(); 292590bd6fd3SPetr Holasek if (ksm_merge_across_nodes != knob) { 2926cbf86cfeSHugh Dickins if (ksm_pages_shared || remove_all_stable_nodes()) 292790bd6fd3SPetr Holasek err = -EBUSY; 2928ef53d16cSHugh Dickins else if (root_stable_tree == one_stable_tree) { 2929ef53d16cSHugh Dickins struct rb_root *buf; 2930ef53d16cSHugh Dickins /* 2931ef53d16cSHugh Dickins * This is the first time that we switch away from the 2932ef53d16cSHugh Dickins * default of merging across nodes: must now allocate 2933ef53d16cSHugh Dickins * a buffer to hold as many roots as may be needed. 2934ef53d16cSHugh Dickins * Allocate stable and unstable together: 2935ef53d16cSHugh Dickins * MAXSMP NODES_SHIFT 10 will use 16kB. 2936ef53d16cSHugh Dickins */ 2937bafe1e14SJoe Perches buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf), 2938bafe1e14SJoe Perches GFP_KERNEL); 2939ef53d16cSHugh Dickins /* Let us assume that RB_ROOT is NULL is zero */ 2940ef53d16cSHugh Dickins if (!buf) 2941ef53d16cSHugh Dickins err = -ENOMEM; 2942ef53d16cSHugh Dickins else { 2943ef53d16cSHugh Dickins root_stable_tree = buf; 2944ef53d16cSHugh Dickins root_unstable_tree = buf + nr_node_ids; 2945ef53d16cSHugh Dickins /* Stable tree is empty but not the unstable */ 2946ef53d16cSHugh Dickins root_unstable_tree[0] = one_unstable_tree[0]; 2947ef53d16cSHugh Dickins } 2948ef53d16cSHugh Dickins } 2949ef53d16cSHugh Dickins if (!err) { 295090bd6fd3SPetr Holasek ksm_merge_across_nodes = knob; 2951ef53d16cSHugh Dickins ksm_nr_node_ids = knob ? 1 : nr_node_ids; 2952ef53d16cSHugh Dickins } 295390bd6fd3SPetr Holasek } 295490bd6fd3SPetr Holasek mutex_unlock(&ksm_thread_mutex); 295590bd6fd3SPetr Holasek 295690bd6fd3SPetr Holasek return err ? err : count; 295790bd6fd3SPetr Holasek } 295890bd6fd3SPetr Holasek KSM_ATTR(merge_across_nodes); 295990bd6fd3SPetr Holasek #endif 296090bd6fd3SPetr Holasek 2961e86c59b1SClaudio Imbrenda static ssize_t use_zero_pages_show(struct kobject *kobj, 2962e86c59b1SClaudio Imbrenda struct kobj_attribute *attr, char *buf) 2963e86c59b1SClaudio Imbrenda { 2964e86c59b1SClaudio Imbrenda return sprintf(buf, "%u\n", ksm_use_zero_pages); 2965e86c59b1SClaudio Imbrenda } 2966e86c59b1SClaudio Imbrenda static ssize_t use_zero_pages_store(struct kobject *kobj, 2967e86c59b1SClaudio Imbrenda struct kobj_attribute *attr, 2968e86c59b1SClaudio Imbrenda const char *buf, size_t count) 2969e86c59b1SClaudio Imbrenda { 2970e86c59b1SClaudio Imbrenda int err; 2971e86c59b1SClaudio Imbrenda bool value; 2972e86c59b1SClaudio Imbrenda 2973e86c59b1SClaudio Imbrenda err = kstrtobool(buf, &value); 2974e86c59b1SClaudio Imbrenda if (err) 2975e86c59b1SClaudio Imbrenda return -EINVAL; 2976e86c59b1SClaudio Imbrenda 2977e86c59b1SClaudio Imbrenda ksm_use_zero_pages = value; 2978e86c59b1SClaudio Imbrenda 2979e86c59b1SClaudio Imbrenda return count; 2980e86c59b1SClaudio Imbrenda } 2981e86c59b1SClaudio Imbrenda KSM_ATTR(use_zero_pages); 2982e86c59b1SClaudio Imbrenda 29832c653d0eSAndrea Arcangeli static ssize_t max_page_sharing_show(struct kobject *kobj, 29842c653d0eSAndrea Arcangeli struct kobj_attribute *attr, char *buf) 29852c653d0eSAndrea Arcangeli { 29862c653d0eSAndrea Arcangeli return sprintf(buf, "%u\n", ksm_max_page_sharing); 29872c653d0eSAndrea Arcangeli } 29882c653d0eSAndrea Arcangeli 29892c653d0eSAndrea Arcangeli static ssize_t max_page_sharing_store(struct kobject *kobj, 29902c653d0eSAndrea Arcangeli struct kobj_attribute *attr, 29912c653d0eSAndrea Arcangeli const char *buf, size_t count) 29922c653d0eSAndrea Arcangeli { 29932c653d0eSAndrea Arcangeli int err; 29942c653d0eSAndrea Arcangeli int knob; 29952c653d0eSAndrea Arcangeli 29962c653d0eSAndrea Arcangeli err = kstrtoint(buf, 10, &knob); 29972c653d0eSAndrea Arcangeli if (err) 29982c653d0eSAndrea Arcangeli return err; 29992c653d0eSAndrea Arcangeli /* 30002c653d0eSAndrea Arcangeli * When a KSM page is created it is shared by 2 mappings. This 30012c653d0eSAndrea Arcangeli * being a signed comparison, it implicitly verifies it's not 30022c653d0eSAndrea Arcangeli * negative. 30032c653d0eSAndrea Arcangeli */ 30042c653d0eSAndrea Arcangeli if (knob < 2) 30052c653d0eSAndrea Arcangeli return -EINVAL; 30062c653d0eSAndrea Arcangeli 30072c653d0eSAndrea Arcangeli if (READ_ONCE(ksm_max_page_sharing) == knob) 30082c653d0eSAndrea Arcangeli return count; 30092c653d0eSAndrea Arcangeli 30102c653d0eSAndrea Arcangeli mutex_lock(&ksm_thread_mutex); 30112c653d0eSAndrea Arcangeli wait_while_offlining(); 30122c653d0eSAndrea Arcangeli if (ksm_max_page_sharing != knob) { 30132c653d0eSAndrea Arcangeli if (ksm_pages_shared || remove_all_stable_nodes()) 30142c653d0eSAndrea Arcangeli err = -EBUSY; 30152c653d0eSAndrea Arcangeli else 30162c653d0eSAndrea Arcangeli ksm_max_page_sharing = knob; 30172c653d0eSAndrea Arcangeli } 30182c653d0eSAndrea Arcangeli mutex_unlock(&ksm_thread_mutex); 30192c653d0eSAndrea Arcangeli 30202c653d0eSAndrea Arcangeli return err ? err : count; 30212c653d0eSAndrea Arcangeli } 30222c653d0eSAndrea Arcangeli KSM_ATTR(max_page_sharing); 30232c653d0eSAndrea Arcangeli 3024b4028260SHugh Dickins static ssize_t pages_shared_show(struct kobject *kobj, 3025b4028260SHugh Dickins struct kobj_attribute *attr, char *buf) 3026b4028260SHugh Dickins { 3027b4028260SHugh Dickins return sprintf(buf, "%lu\n", ksm_pages_shared); 3028b4028260SHugh Dickins } 3029b4028260SHugh Dickins KSM_ATTR_RO(pages_shared); 3030b4028260SHugh Dickins 3031b4028260SHugh Dickins static ssize_t pages_sharing_show(struct kobject *kobj, 3032b4028260SHugh Dickins struct kobj_attribute *attr, char *buf) 3033b4028260SHugh Dickins { 3034e178dfdeSHugh Dickins return sprintf(buf, "%lu\n", ksm_pages_sharing); 3035b4028260SHugh Dickins } 3036b4028260SHugh Dickins KSM_ATTR_RO(pages_sharing); 3037b4028260SHugh Dickins 3038473b0ce4SHugh Dickins static ssize_t pages_unshared_show(struct kobject *kobj, 3039473b0ce4SHugh Dickins struct kobj_attribute *attr, char *buf) 3040473b0ce4SHugh Dickins { 3041473b0ce4SHugh Dickins return sprintf(buf, "%lu\n", ksm_pages_unshared); 3042473b0ce4SHugh Dickins } 3043473b0ce4SHugh Dickins KSM_ATTR_RO(pages_unshared); 3044473b0ce4SHugh Dickins 3045473b0ce4SHugh Dickins static ssize_t pages_volatile_show(struct kobject *kobj, 3046473b0ce4SHugh Dickins struct kobj_attribute *attr, char *buf) 3047473b0ce4SHugh Dickins { 3048473b0ce4SHugh Dickins long ksm_pages_volatile; 3049473b0ce4SHugh Dickins 3050473b0ce4SHugh Dickins ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared 3051473b0ce4SHugh Dickins - ksm_pages_sharing - ksm_pages_unshared; 3052473b0ce4SHugh Dickins /* 3053473b0ce4SHugh Dickins * It was not worth any locking to calculate that statistic, 3054473b0ce4SHugh Dickins * but it might therefore sometimes be negative: conceal that. 3055473b0ce4SHugh Dickins */ 3056473b0ce4SHugh Dickins if (ksm_pages_volatile < 0) 3057473b0ce4SHugh Dickins ksm_pages_volatile = 0; 3058473b0ce4SHugh Dickins return sprintf(buf, "%ld\n", ksm_pages_volatile); 3059473b0ce4SHugh Dickins } 3060473b0ce4SHugh Dickins KSM_ATTR_RO(pages_volatile); 3061473b0ce4SHugh Dickins 30622c653d0eSAndrea Arcangeli static ssize_t stable_node_dups_show(struct kobject *kobj, 30632c653d0eSAndrea Arcangeli struct kobj_attribute *attr, char *buf) 30642c653d0eSAndrea Arcangeli { 30652c653d0eSAndrea Arcangeli return sprintf(buf, "%lu\n", ksm_stable_node_dups); 30662c653d0eSAndrea Arcangeli } 30672c653d0eSAndrea Arcangeli KSM_ATTR_RO(stable_node_dups); 30682c653d0eSAndrea Arcangeli 30692c653d0eSAndrea Arcangeli static ssize_t stable_node_chains_show(struct kobject *kobj, 30702c653d0eSAndrea Arcangeli struct kobj_attribute *attr, char *buf) 30712c653d0eSAndrea Arcangeli { 30722c653d0eSAndrea Arcangeli return sprintf(buf, "%lu\n", ksm_stable_node_chains); 30732c653d0eSAndrea Arcangeli } 30742c653d0eSAndrea Arcangeli KSM_ATTR_RO(stable_node_chains); 30752c653d0eSAndrea Arcangeli 30762c653d0eSAndrea Arcangeli static ssize_t 30772c653d0eSAndrea Arcangeli stable_node_chains_prune_millisecs_show(struct kobject *kobj, 30782c653d0eSAndrea Arcangeli struct kobj_attribute *attr, 30792c653d0eSAndrea Arcangeli char *buf) 30802c653d0eSAndrea Arcangeli { 30812c653d0eSAndrea Arcangeli return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs); 30822c653d0eSAndrea Arcangeli } 30832c653d0eSAndrea Arcangeli 30842c653d0eSAndrea Arcangeli static ssize_t 30852c653d0eSAndrea Arcangeli stable_node_chains_prune_millisecs_store(struct kobject *kobj, 30862c653d0eSAndrea Arcangeli struct kobj_attribute *attr, 30872c653d0eSAndrea Arcangeli const char *buf, size_t count) 30882c653d0eSAndrea Arcangeli { 30892c653d0eSAndrea Arcangeli unsigned long msecs; 30902c653d0eSAndrea Arcangeli int err; 30912c653d0eSAndrea Arcangeli 30922c653d0eSAndrea Arcangeli err = kstrtoul(buf, 10, &msecs); 30932c653d0eSAndrea Arcangeli if (err || msecs > UINT_MAX) 30942c653d0eSAndrea Arcangeli return -EINVAL; 30952c653d0eSAndrea Arcangeli 30962c653d0eSAndrea Arcangeli ksm_stable_node_chains_prune_millisecs = msecs; 30972c653d0eSAndrea Arcangeli 30982c653d0eSAndrea Arcangeli return count; 30992c653d0eSAndrea Arcangeli } 31002c653d0eSAndrea Arcangeli KSM_ATTR(stable_node_chains_prune_millisecs); 31012c653d0eSAndrea Arcangeli 3102473b0ce4SHugh Dickins static ssize_t full_scans_show(struct kobject *kobj, 3103473b0ce4SHugh Dickins struct kobj_attribute *attr, char *buf) 3104473b0ce4SHugh Dickins { 3105473b0ce4SHugh Dickins return sprintf(buf, "%lu\n", ksm_scan.seqnr); 3106473b0ce4SHugh Dickins } 3107473b0ce4SHugh Dickins KSM_ATTR_RO(full_scans); 3108473b0ce4SHugh Dickins 310931dbd01fSIzik Eidus static struct attribute *ksm_attrs[] = { 311031dbd01fSIzik Eidus &sleep_millisecs_attr.attr, 311131dbd01fSIzik Eidus &pages_to_scan_attr.attr, 311231dbd01fSIzik Eidus &run_attr.attr, 3113b4028260SHugh Dickins &pages_shared_attr.attr, 3114b4028260SHugh Dickins &pages_sharing_attr.attr, 3115473b0ce4SHugh Dickins &pages_unshared_attr.attr, 3116473b0ce4SHugh Dickins &pages_volatile_attr.attr, 3117473b0ce4SHugh Dickins &full_scans_attr.attr, 311890bd6fd3SPetr Holasek #ifdef CONFIG_NUMA 311990bd6fd3SPetr Holasek &merge_across_nodes_attr.attr, 312090bd6fd3SPetr Holasek #endif 31212c653d0eSAndrea Arcangeli &max_page_sharing_attr.attr, 31222c653d0eSAndrea Arcangeli &stable_node_chains_attr.attr, 31232c653d0eSAndrea Arcangeli &stable_node_dups_attr.attr, 31242c653d0eSAndrea Arcangeli &stable_node_chains_prune_millisecs_attr.attr, 3125e86c59b1SClaudio Imbrenda &use_zero_pages_attr.attr, 312631dbd01fSIzik Eidus NULL, 312731dbd01fSIzik Eidus }; 312831dbd01fSIzik Eidus 3129f907c26aSArvind Yadav static const struct attribute_group ksm_attr_group = { 313031dbd01fSIzik Eidus .attrs = ksm_attrs, 313131dbd01fSIzik Eidus .name = "ksm", 313231dbd01fSIzik Eidus }; 31332ffd8679SHugh Dickins #endif /* CONFIG_SYSFS */ 313431dbd01fSIzik Eidus 313531dbd01fSIzik Eidus static int __init ksm_init(void) 313631dbd01fSIzik Eidus { 313731dbd01fSIzik Eidus struct task_struct *ksm_thread; 313831dbd01fSIzik Eidus int err; 313931dbd01fSIzik Eidus 3140e86c59b1SClaudio Imbrenda /* The correct value depends on page size and endianness */ 3141e86c59b1SClaudio Imbrenda zero_checksum = calc_checksum(ZERO_PAGE(0)); 3142e86c59b1SClaudio Imbrenda /* Default to false for backwards compatibility */ 3143e86c59b1SClaudio Imbrenda ksm_use_zero_pages = false; 3144e86c59b1SClaudio Imbrenda 314531dbd01fSIzik Eidus err = ksm_slab_init(); 314631dbd01fSIzik Eidus if (err) 314731dbd01fSIzik Eidus goto out; 314831dbd01fSIzik Eidus 314931dbd01fSIzik Eidus ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); 315031dbd01fSIzik Eidus if (IS_ERR(ksm_thread)) { 315125acde31SPaul McQuade pr_err("ksm: creating kthread failed\n"); 315231dbd01fSIzik Eidus err = PTR_ERR(ksm_thread); 3153d9f8984cSLai Jiangshan goto out_free; 315431dbd01fSIzik Eidus } 315531dbd01fSIzik Eidus 31562ffd8679SHugh Dickins #ifdef CONFIG_SYSFS 315731dbd01fSIzik Eidus err = sysfs_create_group(mm_kobj, &ksm_attr_group); 315831dbd01fSIzik Eidus if (err) { 315925acde31SPaul McQuade pr_err("ksm: register sysfs failed\n"); 31602ffd8679SHugh Dickins kthread_stop(ksm_thread); 3161d9f8984cSLai Jiangshan goto out_free; 316231dbd01fSIzik Eidus } 3163c73602adSHugh Dickins #else 3164c73602adSHugh Dickins ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ 3165c73602adSHugh Dickins 31662ffd8679SHugh Dickins #endif /* CONFIG_SYSFS */ 316731dbd01fSIzik Eidus 316862b61f61SHugh Dickins #ifdef CONFIG_MEMORY_HOTREMOVE 3169ef4d43a8SHugh Dickins /* There is no significance to this priority 100 */ 317062b61f61SHugh Dickins hotplug_memory_notifier(ksm_memory_callback, 100); 317162b61f61SHugh Dickins #endif 317231dbd01fSIzik Eidus return 0; 317331dbd01fSIzik Eidus 3174d9f8984cSLai Jiangshan out_free: 317531dbd01fSIzik Eidus ksm_slab_free(); 317631dbd01fSIzik Eidus out: 317731dbd01fSIzik Eidus return err; 317831dbd01fSIzik Eidus } 3179a64fb3cdSPaul Gortmaker subsys_initcall(ksm_init); 3180