1b2441318SGreg Kroah-Hartman // SPDX-License-Identifier: GPL-2.0 2b20a3503SChristoph Lameter /* 314e0f9bcSHugh Dickins * Memory Migration functionality - linux/mm/migrate.c 4b20a3503SChristoph Lameter * 5b20a3503SChristoph Lameter * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter 6b20a3503SChristoph Lameter * 7b20a3503SChristoph Lameter * Page migration was first developed in the context of the memory hotplug 8b20a3503SChristoph Lameter * project. The main authors of the migration code are: 9b20a3503SChristoph Lameter * 10b20a3503SChristoph Lameter * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> 11b20a3503SChristoph Lameter * Hirokazu Takahashi <taka@valinux.co.jp> 12b20a3503SChristoph Lameter * Dave Hansen <haveblue@us.ibm.com> 13cde53535SChristoph Lameter * Christoph Lameter 14b20a3503SChristoph Lameter */ 15b20a3503SChristoph Lameter 16b20a3503SChristoph Lameter #include <linux/migrate.h> 17b95f1b31SPaul Gortmaker #include <linux/export.h> 18b20a3503SChristoph Lameter #include <linux/swap.h> 190697212aSChristoph Lameter #include <linux/swapops.h> 20b20a3503SChristoph Lameter #include <linux/pagemap.h> 21e23ca00bSChristoph Lameter #include <linux/buffer_head.h> 22b20a3503SChristoph Lameter #include <linux/mm_inline.h> 23b488893aSPavel Emelyanov #include <linux/nsproxy.h> 24b20a3503SChristoph Lameter #include <linux/pagevec.h> 25e9995ef9SHugh Dickins #include <linux/ksm.h> 26b20a3503SChristoph Lameter #include <linux/rmap.h> 27b20a3503SChristoph Lameter #include <linux/topology.h> 28b20a3503SChristoph Lameter #include <linux/cpu.h> 29b20a3503SChristoph Lameter #include <linux/cpuset.h> 3004e62a29SChristoph Lameter #include <linux/writeback.h> 31742755a1SChristoph Lameter #include <linux/mempolicy.h> 32742755a1SChristoph Lameter #include <linux/vmalloc.h> 3386c3a764SDavid Quigley #include <linux/security.h> 3442cb14b1SHugh Dickins #include <linux/backing-dev.h> 35bda807d4SMinchan Kim #include <linux/compaction.h> 364f5ca265SAdrian Bunk #include <linux/syscalls.h> 377addf443SDominik Brodowski #include <linux/compat.h> 38290408d4SNaoya Horiguchi #include <linux/hugetlb.h> 398e6ac7faSAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 405a0e3ad6STejun Heo #include <linux/gfp.h> 41df6ad698SJérôme Glisse #include <linux/pfn_t.h> 42a5430ddaSJérôme Glisse #include <linux/memremap.h> 438315ada7SJérôme Glisse #include <linux/userfaultfd_k.h> 44bf6bddf1SRafael Aquini #include <linux/balloon_compaction.h> 4533c3fc71SVladimir Davydov #include <linux/page_idle.h> 46d435edcaSVlastimil Babka #include <linux/page_owner.h> 476e84f315SIngo Molnar #include <linux/sched/mm.h> 48197e7e52SLinus Torvalds #include <linux/ptrace.h> 4934290e2cSRalph Campbell #include <linux/oom.h> 50884a6e5dSDave Hansen #include <linux/memory.h> 51ac16ec83SBaolin Wang #include <linux/random.h> 52c574bbe9SHuang Ying #include <linux/sched/sysctl.h> 53467b171aSAneesh Kumar K.V #include <linux/memory-tiers.h> 54b20a3503SChristoph Lameter 550d1836c3SMichal Nazarewicz #include <asm/tlbflush.h> 560d1836c3SMichal Nazarewicz 577b2a2d4aSMel Gorman #include <trace/events/migrate.h> 587b2a2d4aSMel Gorman 59b20a3503SChristoph Lameter #include "internal.h" 60b20a3503SChristoph Lameter 619e5bcd61SYisheng Xie int isolate_movable_page(struct page *page, isolate_mode_t mode) 62bda807d4SMinchan Kim { 6368f2736aSMatthew Wilcox (Oracle) const struct movable_operations *mops; 64bda807d4SMinchan Kim 65bda807d4SMinchan Kim /* 66bda807d4SMinchan Kim * Avoid burning cycles with pages that are yet under __free_pages(), 67bda807d4SMinchan Kim * or just got freed under us. 68bda807d4SMinchan Kim * 69bda807d4SMinchan Kim * In case we 'win' a race for a movable page being freed under us and 70bda807d4SMinchan Kim * raise its refcount preventing __free_pages() from doing its job 71bda807d4SMinchan Kim * the put_page() at the end of this block will take care of 72bda807d4SMinchan Kim * release this page, thus avoiding a nasty leakage. 73bda807d4SMinchan Kim */ 74bda807d4SMinchan Kim if (unlikely(!get_page_unless_zero(page))) 75bda807d4SMinchan Kim goto out; 76bda807d4SMinchan Kim 778b881763SVlastimil Babka if (unlikely(PageSlab(page))) 788b881763SVlastimil Babka goto out_putpage; 798b881763SVlastimil Babka /* Pairs with smp_wmb() in slab freeing, e.g. SLUB's __free_slab() */ 808b881763SVlastimil Babka smp_rmb(); 81bda807d4SMinchan Kim /* 828b881763SVlastimil Babka * Check movable flag before taking the page lock because 838b881763SVlastimil Babka * we use non-atomic bitops on newly allocated page flags so 848b881763SVlastimil Babka * unconditionally grabbing the lock ruins page's owner side. 85bda807d4SMinchan Kim */ 86bda807d4SMinchan Kim if (unlikely(!__PageMovable(page))) 87bda807d4SMinchan Kim goto out_putpage; 888b881763SVlastimil Babka /* Pairs with smp_wmb() in slab allocation, e.g. SLUB's alloc_slab_page() */ 898b881763SVlastimil Babka smp_rmb(); 908b881763SVlastimil Babka if (unlikely(PageSlab(page))) 918b881763SVlastimil Babka goto out_putpage; 928b881763SVlastimil Babka 93bda807d4SMinchan Kim /* 94bda807d4SMinchan Kim * As movable pages are not isolated from LRU lists, concurrent 95bda807d4SMinchan Kim * compaction threads can race against page migration functions 96bda807d4SMinchan Kim * as well as race against the releasing a page. 97bda807d4SMinchan Kim * 98bda807d4SMinchan Kim * In order to avoid having an already isolated movable page 99bda807d4SMinchan Kim * being (wrongly) re-isolated while it is under migration, 100bda807d4SMinchan Kim * or to avoid attempting to isolate pages being released, 101bda807d4SMinchan Kim * lets be sure we have the page lock 102bda807d4SMinchan Kim * before proceeding with the movable page isolation steps. 103bda807d4SMinchan Kim */ 104bda807d4SMinchan Kim if (unlikely(!trylock_page(page))) 105bda807d4SMinchan Kim goto out_putpage; 106bda807d4SMinchan Kim 107bda807d4SMinchan Kim if (!PageMovable(page) || PageIsolated(page)) 108bda807d4SMinchan Kim goto out_no_isolated; 109bda807d4SMinchan Kim 11068f2736aSMatthew Wilcox (Oracle) mops = page_movable_ops(page); 11168f2736aSMatthew Wilcox (Oracle) VM_BUG_ON_PAGE(!mops, page); 112bda807d4SMinchan Kim 11368f2736aSMatthew Wilcox (Oracle) if (!mops->isolate_page(page, mode)) 114bda807d4SMinchan Kim goto out_no_isolated; 115bda807d4SMinchan Kim 116bda807d4SMinchan Kim /* Driver shouldn't use PG_isolated bit of page->flags */ 117bda807d4SMinchan Kim WARN_ON_ONCE(PageIsolated(page)); 118356ea386Sandrew.yang SetPageIsolated(page); 119bda807d4SMinchan Kim unlock_page(page); 120bda807d4SMinchan Kim 1219e5bcd61SYisheng Xie return 0; 122bda807d4SMinchan Kim 123bda807d4SMinchan Kim out_no_isolated: 124bda807d4SMinchan Kim unlock_page(page); 125bda807d4SMinchan Kim out_putpage: 126bda807d4SMinchan Kim put_page(page); 127bda807d4SMinchan Kim out: 1289e5bcd61SYisheng Xie return -EBUSY; 129bda807d4SMinchan Kim } 130bda807d4SMinchan Kim 131606a6f71SMiaohe Lin static void putback_movable_page(struct page *page) 132bda807d4SMinchan Kim { 13368f2736aSMatthew Wilcox (Oracle) const struct movable_operations *mops = page_movable_ops(page); 134bda807d4SMinchan Kim 13568f2736aSMatthew Wilcox (Oracle) mops->putback_page(page); 136356ea386Sandrew.yang ClearPageIsolated(page); 137bda807d4SMinchan Kim } 138bda807d4SMinchan Kim 139b20a3503SChristoph Lameter /* 1405733c7d1SRafael Aquini * Put previously isolated pages back onto the appropriate lists 1415733c7d1SRafael Aquini * from where they were once taken off for compaction/migration. 1425733c7d1SRafael Aquini * 14359c82b70SJoonsoo Kim * This function shall be used whenever the isolated pageset has been 14459c82b70SJoonsoo Kim * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() 1457ce82f4cSMiaohe Lin * and isolate_hugetlb(). 1465733c7d1SRafael Aquini */ 1475733c7d1SRafael Aquini void putback_movable_pages(struct list_head *l) 1485733c7d1SRafael Aquini { 1495733c7d1SRafael Aquini struct page *page; 1505733c7d1SRafael Aquini struct page *page2; 1515733c7d1SRafael Aquini 1525733c7d1SRafael Aquini list_for_each_entry_safe(page, page2, l, lru) { 15331caf665SNaoya Horiguchi if (unlikely(PageHuge(page))) { 15431caf665SNaoya Horiguchi putback_active_hugepage(page); 15531caf665SNaoya Horiguchi continue; 15631caf665SNaoya Horiguchi } 1575733c7d1SRafael Aquini list_del(&page->lru); 158bda807d4SMinchan Kim /* 159bda807d4SMinchan Kim * We isolated non-lru movable page so here we can use 160bda807d4SMinchan Kim * __PageMovable because LRU page's mapping cannot have 161bda807d4SMinchan Kim * PAGE_MAPPING_MOVABLE. 162bda807d4SMinchan Kim */ 163b1123ea6SMinchan Kim if (unlikely(__PageMovable(page))) { 164bda807d4SMinchan Kim VM_BUG_ON_PAGE(!PageIsolated(page), page); 165bda807d4SMinchan Kim lock_page(page); 166bda807d4SMinchan Kim if (PageMovable(page)) 167bda807d4SMinchan Kim putback_movable_page(page); 168bf6bddf1SRafael Aquini else 169356ea386Sandrew.yang ClearPageIsolated(page); 170bda807d4SMinchan Kim unlock_page(page); 171bda807d4SMinchan Kim put_page(page); 172bda807d4SMinchan Kim } else { 173e8db67ebSNaoya Horiguchi mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + 1746c357848SMatthew Wilcox (Oracle) page_is_file_lru(page), -thp_nr_pages(page)); 175fc280fe8SRabin Vincent putback_lru_page(page); 176b20a3503SChristoph Lameter } 177b20a3503SChristoph Lameter } 178bda807d4SMinchan Kim } 179b20a3503SChristoph Lameter 1800697212aSChristoph Lameter /* 1810697212aSChristoph Lameter * Restore a potential migration pte to a working pte entry 1820697212aSChristoph Lameter */ 1832f031c6fSMatthew Wilcox (Oracle) static bool remove_migration_pte(struct folio *folio, 1842f031c6fSMatthew Wilcox (Oracle) struct vm_area_struct *vma, unsigned long addr, void *old) 1850697212aSChristoph Lameter { 1864eecb8b9SMatthew Wilcox (Oracle) DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION); 1870697212aSChristoph Lameter 1883fe87967SKirill A. Shutemov while (page_vma_mapped_walk(&pvmw)) { 1896c287605SDavid Hildenbrand rmap_t rmap_flags = RMAP_NONE; 1900697212aSChristoph Lameter pte_t pte; 1910697212aSChristoph Lameter swp_entry_t entry; 1924eecb8b9SMatthew Wilcox (Oracle) struct page *new; 1934eecb8b9SMatthew Wilcox (Oracle) unsigned long idx = 0; 1940697212aSChristoph Lameter 1954eecb8b9SMatthew Wilcox (Oracle) /* pgoff is invalid for ksm pages, but they are never large */ 1964eecb8b9SMatthew Wilcox (Oracle) if (folio_test_large(folio) && !folio_test_hugetlb(folio)) 1974eecb8b9SMatthew Wilcox (Oracle) idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff; 1984eecb8b9SMatthew Wilcox (Oracle) new = folio_page(folio, idx); 1990697212aSChristoph Lameter 200616b8371SZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 201616b8371SZi Yan /* PMD-mapped THP migration entry */ 202616b8371SZi Yan if (!pvmw.pte) { 2034eecb8b9SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) || 2044eecb8b9SMatthew Wilcox (Oracle) !folio_test_pmd_mappable(folio), folio); 205616b8371SZi Yan remove_migration_pmd(&pvmw, new); 206616b8371SZi Yan continue; 207616b8371SZi Yan } 208616b8371SZi Yan #endif 209616b8371SZi Yan 2104eecb8b9SMatthew Wilcox (Oracle) folio_get(folio); 2112e346877SPeter Xu pte = mk_pte(new, READ_ONCE(vma->vm_page_prot)); 2123fe87967SKirill A. Shutemov if (pte_swp_soft_dirty(*pvmw.pte)) 213c3d16e16SCyrill Gorcunov pte = pte_mksoft_dirty(pte); 214d3cb8bf6SMel Gorman 2153fe87967SKirill A. Shutemov /* 2163fe87967SKirill A. Shutemov * Recheck VMA as permissions can change since migration started 2173fe87967SKirill A. Shutemov */ 2183fe87967SKirill A. Shutemov entry = pte_to_swp_entry(*pvmw.pte); 2192e346877SPeter Xu if (!is_migration_entry_young(entry)) 2202e346877SPeter Xu pte = pte_mkold(pte); 2212e346877SPeter Xu if (folio_test_dirty(folio) && is_migration_entry_dirty(entry)) 2222e346877SPeter Xu pte = pte_mkdirty(pte); 2234dd845b5SAlistair Popple if (is_writable_migration_entry(entry)) 224d3cb8bf6SMel Gorman pte = maybe_mkwrite(pte, vma); 225f45ec5ffSPeter Xu else if (pte_swp_uffd_wp(*pvmw.pte)) 226f45ec5ffSPeter Xu pte = pte_mkuffd_wp(pte); 227d3cb8bf6SMel Gorman 2286c287605SDavid Hildenbrand if (folio_test_anon(folio) && !is_readable_migration_entry(entry)) 2296c287605SDavid Hildenbrand rmap_flags |= RMAP_EXCLUSIVE; 2306c287605SDavid Hildenbrand 2316128763fSRalph Campbell if (unlikely(is_device_private_page(new))) { 2324dd845b5SAlistair Popple if (pte_write(pte)) 2334dd845b5SAlistair Popple entry = make_writable_device_private_entry( 2344dd845b5SAlistair Popple page_to_pfn(new)); 2354dd845b5SAlistair Popple else 2364dd845b5SAlistair Popple entry = make_readable_device_private_entry( 2374dd845b5SAlistair Popple page_to_pfn(new)); 238a5430ddaSJérôme Glisse pte = swp_entry_to_pte(entry); 2393d321bf8SRalph Campbell if (pte_swp_soft_dirty(*pvmw.pte)) 2403d321bf8SRalph Campbell pte = pte_swp_mksoft_dirty(pte); 241f45ec5ffSPeter Xu if (pte_swp_uffd_wp(*pvmw.pte)) 242ebdf8321SAlistair Popple pte = pte_swp_mkuffd_wp(pte); 243df6ad698SJérôme Glisse } 244a5430ddaSJérôme Glisse 2453ef8fd7fSAndi Kleen #ifdef CONFIG_HUGETLB_PAGE 2464eecb8b9SMatthew Wilcox (Oracle) if (folio_test_hugetlb(folio)) { 24779c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 24879c1c594SChristophe Leroy 249290408d4SNaoya Horiguchi pte = pte_mkhuge(pte); 25079c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 2514eecb8b9SMatthew Wilcox (Oracle) if (folio_test_anon(folio)) 25228c5209dSDavid Hildenbrand hugepage_add_anon_rmap(new, vma, pvmw.address, 2536c287605SDavid Hildenbrand rmap_flags); 254290408d4SNaoya Horiguchi else 255fb3d824dSDavid Hildenbrand page_dup_file_rmap(new, true); 2561eba86c0SPasha Tatashin set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); 257383321abSAneesh Kumar K.V } else 258383321abSAneesh Kumar K.V #endif 259383321abSAneesh Kumar K.V { 2604eecb8b9SMatthew Wilcox (Oracle) if (folio_test_anon(folio)) 261f1e2db12SDavid Hildenbrand page_add_anon_rmap(new, vma, pvmw.address, 2626c287605SDavid Hildenbrand rmap_flags); 26304e62a29SChristoph Lameter else 264cea86fe2SHugh Dickins page_add_file_rmap(new, vma, false); 2651eba86c0SPasha Tatashin set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); 266383321abSAneesh Kumar K.V } 267b7435507SHugh Dickins if (vma->vm_flags & VM_LOCKED) 268*96f97c43SLorenzo Stoakes mlock_drain_local(); 269e125fe40SKirill A. Shutemov 2704cc79b33SAnshuman Khandual trace_remove_migration_pte(pvmw.address, pte_val(pte), 2714cc79b33SAnshuman Khandual compound_order(new)); 2724cc79b33SAnshuman Khandual 27304e62a29SChristoph Lameter /* No need to invalidate - it was non-present before */ 2743fe87967SKirill A. Shutemov update_mmu_cache(vma, pvmw.address, pvmw.pte); 2753fe87967SKirill A. Shutemov } 2763fe87967SKirill A. Shutemov 277e4b82222SMinchan Kim return true; 2780697212aSChristoph Lameter } 2790697212aSChristoph Lameter 2800697212aSChristoph Lameter /* 28104e62a29SChristoph Lameter * Get rid of all migration entries and replace them by 28204e62a29SChristoph Lameter * references to the indicated page. 28304e62a29SChristoph Lameter */ 2844eecb8b9SMatthew Wilcox (Oracle) void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked) 28504e62a29SChristoph Lameter { 286051ac83aSJoonsoo Kim struct rmap_walk_control rwc = { 287051ac83aSJoonsoo Kim .rmap_one = remove_migration_pte, 2884eecb8b9SMatthew Wilcox (Oracle) .arg = src, 289051ac83aSJoonsoo Kim }; 290051ac83aSJoonsoo Kim 291e388466dSKirill A. Shutemov if (locked) 2922f031c6fSMatthew Wilcox (Oracle) rmap_walk_locked(dst, &rwc); 293e388466dSKirill A. Shutemov else 2942f031c6fSMatthew Wilcox (Oracle) rmap_walk(dst, &rwc); 29504e62a29SChristoph Lameter } 29604e62a29SChristoph Lameter 29704e62a29SChristoph Lameter /* 2980697212aSChristoph Lameter * Something used the pte of a page under migration. We need to 2990697212aSChristoph Lameter * get to the page and wait until migration is finished. 3000697212aSChristoph Lameter * When we return from this function the fault will be retried. 3010697212aSChristoph Lameter */ 302e66f17ffSNaoya Horiguchi void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, 30330dad309SNaoya Horiguchi spinlock_t *ptl) 3040697212aSChristoph Lameter { 30530dad309SNaoya Horiguchi pte_t pte; 3060697212aSChristoph Lameter swp_entry_t entry; 3070697212aSChristoph Lameter 30830dad309SNaoya Horiguchi spin_lock(ptl); 3090697212aSChristoph Lameter pte = *ptep; 3100697212aSChristoph Lameter if (!is_swap_pte(pte)) 3110697212aSChristoph Lameter goto out; 3120697212aSChristoph Lameter 3130697212aSChristoph Lameter entry = pte_to_swp_entry(pte); 3140697212aSChristoph Lameter if (!is_migration_entry(entry)) 3150697212aSChristoph Lameter goto out; 3160697212aSChristoph Lameter 317ffa65753SAlistair Popple migration_entry_wait_on_locked(entry, ptep, ptl); 3180697212aSChristoph Lameter return; 3190697212aSChristoph Lameter out: 3200697212aSChristoph Lameter pte_unmap_unlock(ptep, ptl); 3210697212aSChristoph Lameter } 3220697212aSChristoph Lameter 32330dad309SNaoya Horiguchi void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, 32430dad309SNaoya Horiguchi unsigned long address) 32530dad309SNaoya Horiguchi { 32630dad309SNaoya Horiguchi spinlock_t *ptl = pte_lockptr(mm, pmd); 32730dad309SNaoya Horiguchi pte_t *ptep = pte_offset_map(pmd, address); 32830dad309SNaoya Horiguchi __migration_entry_wait(mm, ptep, ptl); 32930dad309SNaoya Horiguchi } 33030dad309SNaoya Horiguchi 331ad1ac596SMiaohe Lin #ifdef CONFIG_HUGETLB_PAGE 332fcd48540SPeter Xu /* 333fcd48540SPeter Xu * The vma read lock must be held upon entry. Holding that lock prevents either 334fcd48540SPeter Xu * the pte or the ptl from being freed. 335fcd48540SPeter Xu * 336fcd48540SPeter Xu * This function will release the vma lock before returning. 337fcd48540SPeter Xu */ 338fcd48540SPeter Xu void __migration_entry_wait_huge(struct vm_area_struct *vma, 339fcd48540SPeter Xu pte_t *ptep, spinlock_t *ptl) 34030dad309SNaoya Horiguchi { 341ad1ac596SMiaohe Lin pte_t pte; 342ad1ac596SMiaohe Lin 343fcd48540SPeter Xu hugetlb_vma_assert_locked(vma); 344ad1ac596SMiaohe Lin spin_lock(ptl); 345ad1ac596SMiaohe Lin pte = huge_ptep_get(ptep); 346ad1ac596SMiaohe Lin 347fcd48540SPeter Xu if (unlikely(!is_hugetlb_entry_migration(pte))) { 348ad1ac596SMiaohe Lin spin_unlock(ptl); 349fcd48540SPeter Xu hugetlb_vma_unlock_read(vma); 350fcd48540SPeter Xu } else { 351fcd48540SPeter Xu /* 352fcd48540SPeter Xu * If migration entry existed, safe to release vma lock 353fcd48540SPeter Xu * here because the pgtable page won't be freed without the 354fcd48540SPeter Xu * pgtable lock released. See comment right above pgtable 355fcd48540SPeter Xu * lock release in migration_entry_wait_on_locked(). 356fcd48540SPeter Xu */ 357fcd48540SPeter Xu hugetlb_vma_unlock_read(vma); 358ad1ac596SMiaohe Lin migration_entry_wait_on_locked(pte_to_swp_entry(pte), NULL, ptl); 35930dad309SNaoya Horiguchi } 360fcd48540SPeter Xu } 36130dad309SNaoya Horiguchi 362ad1ac596SMiaohe Lin void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *pte) 363ad1ac596SMiaohe Lin { 364ad1ac596SMiaohe Lin spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, pte); 365ad1ac596SMiaohe Lin 366fcd48540SPeter Xu __migration_entry_wait_huge(vma, pte, ptl); 367ad1ac596SMiaohe Lin } 368ad1ac596SMiaohe Lin #endif 369ad1ac596SMiaohe Lin 370616b8371SZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 371616b8371SZi Yan void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) 372616b8371SZi Yan { 373616b8371SZi Yan spinlock_t *ptl; 374616b8371SZi Yan 375616b8371SZi Yan ptl = pmd_lock(mm, pmd); 376616b8371SZi Yan if (!is_pmd_migration_entry(*pmd)) 377616b8371SZi Yan goto unlock; 378ffa65753SAlistair Popple migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl); 379616b8371SZi Yan return; 380616b8371SZi Yan unlock: 381616b8371SZi Yan spin_unlock(ptl); 382616b8371SZi Yan } 383616b8371SZi Yan #endif 384616b8371SZi Yan 385108ca835SMatthew Wilcox (Oracle) static int folio_expected_refs(struct address_space *mapping, 386108ca835SMatthew Wilcox (Oracle) struct folio *folio) 3870b3901b3SJan Kara { 388108ca835SMatthew Wilcox (Oracle) int refs = 1; 389108ca835SMatthew Wilcox (Oracle) if (!mapping) 390108ca835SMatthew Wilcox (Oracle) return refs; 3910b3901b3SJan Kara 392108ca835SMatthew Wilcox (Oracle) refs += folio_nr_pages(folio); 393108ca835SMatthew Wilcox (Oracle) if (folio_test_private(folio)) 394108ca835SMatthew Wilcox (Oracle) refs++; 395108ca835SMatthew Wilcox (Oracle) 396108ca835SMatthew Wilcox (Oracle) return refs; 3970b3901b3SJan Kara } 3980b3901b3SJan Kara 399b20a3503SChristoph Lameter /* 400c3fcf8a5SChristoph Lameter * Replace the page in the mapping. 4015b5c7120SChristoph Lameter * 4025b5c7120SChristoph Lameter * The number of remaining references must be: 4035b5c7120SChristoph Lameter * 1 for anonymous pages without a mapping 4045b5c7120SChristoph Lameter * 2 for pages with a mapping 405266cf658SDavid Howells * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. 406b20a3503SChristoph Lameter */ 4073417013eSMatthew Wilcox (Oracle) int folio_migrate_mapping(struct address_space *mapping, 4083417013eSMatthew Wilcox (Oracle) struct folio *newfolio, struct folio *folio, int extra_count) 409b20a3503SChristoph Lameter { 4103417013eSMatthew Wilcox (Oracle) XA_STATE(xas, &mapping->i_pages, folio_index(folio)); 41142cb14b1SHugh Dickins struct zone *oldzone, *newzone; 41242cb14b1SHugh Dickins int dirty; 413108ca835SMatthew Wilcox (Oracle) int expected_count = folio_expected_refs(mapping, folio) + extra_count; 4143417013eSMatthew Wilcox (Oracle) long nr = folio_nr_pages(folio); 4158763cb45SJérôme Glisse 4166c5240aeSChristoph Lameter if (!mapping) { 4170e8c7d0fSChristoph Lameter /* Anonymous page without mapping */ 4183417013eSMatthew Wilcox (Oracle) if (folio_ref_count(folio) != expected_count) 4196c5240aeSChristoph Lameter return -EAGAIN; 420cf4b769aSHugh Dickins 421cf4b769aSHugh Dickins /* No turning back from here */ 4223417013eSMatthew Wilcox (Oracle) newfolio->index = folio->index; 4233417013eSMatthew Wilcox (Oracle) newfolio->mapping = folio->mapping; 4243417013eSMatthew Wilcox (Oracle) if (folio_test_swapbacked(folio)) 4253417013eSMatthew Wilcox (Oracle) __folio_set_swapbacked(newfolio); 426cf4b769aSHugh Dickins 42778bd5209SRafael Aquini return MIGRATEPAGE_SUCCESS; 4286c5240aeSChristoph Lameter } 4296c5240aeSChristoph Lameter 4303417013eSMatthew Wilcox (Oracle) oldzone = folio_zone(folio); 4313417013eSMatthew Wilcox (Oracle) newzone = folio_zone(newfolio); 43242cb14b1SHugh Dickins 43389eb946aSMatthew Wilcox xas_lock_irq(&xas); 4343417013eSMatthew Wilcox (Oracle) if (!folio_ref_freeze(folio, expected_count)) { 43589eb946aSMatthew Wilcox xas_unlock_irq(&xas); 436e286781dSNick Piggin return -EAGAIN; 437e286781dSNick Piggin } 438e286781dSNick Piggin 439b20a3503SChristoph Lameter /* 4403417013eSMatthew Wilcox (Oracle) * Now we know that no one else is looking at the folio: 441cf4b769aSHugh Dickins * no turning back from here. 442b20a3503SChristoph Lameter */ 4433417013eSMatthew Wilcox (Oracle) newfolio->index = folio->index; 4443417013eSMatthew Wilcox (Oracle) newfolio->mapping = folio->mapping; 4453417013eSMatthew Wilcox (Oracle) folio_ref_add(newfolio, nr); /* add cache reference */ 4463417013eSMatthew Wilcox (Oracle) if (folio_test_swapbacked(folio)) { 4473417013eSMatthew Wilcox (Oracle) __folio_set_swapbacked(newfolio); 4483417013eSMatthew Wilcox (Oracle) if (folio_test_swapcache(folio)) { 4493417013eSMatthew Wilcox (Oracle) folio_set_swapcache(newfolio); 4503417013eSMatthew Wilcox (Oracle) newfolio->private = folio_get_private(folio); 451b20a3503SChristoph Lameter } 4526326fec1SNicholas Piggin } else { 4533417013eSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio); 4546326fec1SNicholas Piggin } 455b20a3503SChristoph Lameter 45642cb14b1SHugh Dickins /* Move dirty while page refs frozen and newpage not yet exposed */ 4573417013eSMatthew Wilcox (Oracle) dirty = folio_test_dirty(folio); 45842cb14b1SHugh Dickins if (dirty) { 4593417013eSMatthew Wilcox (Oracle) folio_clear_dirty(folio); 4603417013eSMatthew Wilcox (Oracle) folio_set_dirty(newfolio); 46142cb14b1SHugh Dickins } 46242cb14b1SHugh Dickins 4633417013eSMatthew Wilcox (Oracle) xas_store(&xas, newfolio); 4647cf9c2c7SNick Piggin 4657cf9c2c7SNick Piggin /* 466937a94c9SJacobo Giralt * Drop cache reference from old page by unfreezing 467937a94c9SJacobo Giralt * to one less reference. 4687cf9c2c7SNick Piggin * We know this isn't the last reference. 4697cf9c2c7SNick Piggin */ 4703417013eSMatthew Wilcox (Oracle) folio_ref_unfreeze(folio, expected_count - nr); 4717cf9c2c7SNick Piggin 47289eb946aSMatthew Wilcox xas_unlock(&xas); 47342cb14b1SHugh Dickins /* Leave irq disabled to prevent preemption while updating stats */ 47442cb14b1SHugh Dickins 4750e8c7d0fSChristoph Lameter /* 4760e8c7d0fSChristoph Lameter * If moved to a different zone then also account 4770e8c7d0fSChristoph Lameter * the page for that zone. Other VM counters will be 4780e8c7d0fSChristoph Lameter * taken care of when we establish references to the 4790e8c7d0fSChristoph Lameter * new page and drop references to the old page. 4800e8c7d0fSChristoph Lameter * 4810e8c7d0fSChristoph Lameter * Note that anonymous pages are accounted for 4824b9d0fabSMel Gorman * via NR_FILE_PAGES and NR_ANON_MAPPED if they 4830e8c7d0fSChristoph Lameter * are mapped to swap space. 4840e8c7d0fSChristoph Lameter */ 48542cb14b1SHugh Dickins if (newzone != oldzone) { 4860d1c2072SJohannes Weiner struct lruvec *old_lruvec, *new_lruvec; 4870d1c2072SJohannes Weiner struct mem_cgroup *memcg; 4880d1c2072SJohannes Weiner 4893417013eSMatthew Wilcox (Oracle) memcg = folio_memcg(folio); 4900d1c2072SJohannes Weiner old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat); 4910d1c2072SJohannes Weiner new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat); 4920d1c2072SJohannes Weiner 4935c447d27SShakeel Butt __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr); 4945c447d27SShakeel Butt __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr); 4953417013eSMatthew Wilcox (Oracle) if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) { 4965c447d27SShakeel Butt __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr); 4975c447d27SShakeel Butt __mod_lruvec_state(new_lruvec, NR_SHMEM, nr); 4984b02108aSKOSAKI Motohiro } 499b6038942SShakeel Butt #ifdef CONFIG_SWAP 5003417013eSMatthew Wilcox (Oracle) if (folio_test_swapcache(folio)) { 501b6038942SShakeel Butt __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr); 502b6038942SShakeel Butt __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr); 503b6038942SShakeel Butt } 504b6038942SShakeel Butt #endif 505f56753acSChristoph Hellwig if (dirty && mapping_can_writeback(mapping)) { 5065c447d27SShakeel Butt __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr); 5075c447d27SShakeel Butt __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr); 5085c447d27SShakeel Butt __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr); 5095c447d27SShakeel Butt __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr); 51042cb14b1SHugh Dickins } 51142cb14b1SHugh Dickins } 51242cb14b1SHugh Dickins local_irq_enable(); 513b20a3503SChristoph Lameter 51478bd5209SRafael Aquini return MIGRATEPAGE_SUCCESS; 515b20a3503SChristoph Lameter } 5163417013eSMatthew Wilcox (Oracle) EXPORT_SYMBOL(folio_migrate_mapping); 517b20a3503SChristoph Lameter 518b20a3503SChristoph Lameter /* 519290408d4SNaoya Horiguchi * The expected number of remaining references is the same as that 5203417013eSMatthew Wilcox (Oracle) * of folio_migrate_mapping(). 521290408d4SNaoya Horiguchi */ 522290408d4SNaoya Horiguchi int migrate_huge_page_move_mapping(struct address_space *mapping, 523b890ec2aSMatthew Wilcox (Oracle) struct folio *dst, struct folio *src) 524290408d4SNaoya Horiguchi { 525b890ec2aSMatthew Wilcox (Oracle) XA_STATE(xas, &mapping->i_pages, folio_index(src)); 526290408d4SNaoya Horiguchi int expected_count; 527290408d4SNaoya Horiguchi 52889eb946aSMatthew Wilcox xas_lock_irq(&xas); 529b890ec2aSMatthew Wilcox (Oracle) expected_count = 2 + folio_has_private(src); 530b890ec2aSMatthew Wilcox (Oracle) if (!folio_ref_freeze(src, expected_count)) { 53189eb946aSMatthew Wilcox xas_unlock_irq(&xas); 532290408d4SNaoya Horiguchi return -EAGAIN; 533290408d4SNaoya Horiguchi } 534290408d4SNaoya Horiguchi 535b890ec2aSMatthew Wilcox (Oracle) dst->index = src->index; 536b890ec2aSMatthew Wilcox (Oracle) dst->mapping = src->mapping; 5376a93ca8fSJohannes Weiner 538b890ec2aSMatthew Wilcox (Oracle) folio_get(dst); 539290408d4SNaoya Horiguchi 540b890ec2aSMatthew Wilcox (Oracle) xas_store(&xas, dst); 541290408d4SNaoya Horiguchi 542b890ec2aSMatthew Wilcox (Oracle) folio_ref_unfreeze(src, expected_count - 1); 543290408d4SNaoya Horiguchi 54489eb946aSMatthew Wilcox xas_unlock_irq(&xas); 5456a93ca8fSJohannes Weiner 54678bd5209SRafael Aquini return MIGRATEPAGE_SUCCESS; 547290408d4SNaoya Horiguchi } 548290408d4SNaoya Horiguchi 549290408d4SNaoya Horiguchi /* 55019138349SMatthew Wilcox (Oracle) * Copy the flags and some other ancillary information 551b20a3503SChristoph Lameter */ 55219138349SMatthew Wilcox (Oracle) void folio_migrate_flags(struct folio *newfolio, struct folio *folio) 553b20a3503SChristoph Lameter { 5547851a45cSRik van Riel int cpupid; 5557851a45cSRik van Riel 55619138349SMatthew Wilcox (Oracle) if (folio_test_error(folio)) 55719138349SMatthew Wilcox (Oracle) folio_set_error(newfolio); 55819138349SMatthew Wilcox (Oracle) if (folio_test_referenced(folio)) 55919138349SMatthew Wilcox (Oracle) folio_set_referenced(newfolio); 56019138349SMatthew Wilcox (Oracle) if (folio_test_uptodate(folio)) 56119138349SMatthew Wilcox (Oracle) folio_mark_uptodate(newfolio); 56219138349SMatthew Wilcox (Oracle) if (folio_test_clear_active(folio)) { 56319138349SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio); 56419138349SMatthew Wilcox (Oracle) folio_set_active(newfolio); 56519138349SMatthew Wilcox (Oracle) } else if (folio_test_clear_unevictable(folio)) 56619138349SMatthew Wilcox (Oracle) folio_set_unevictable(newfolio); 56719138349SMatthew Wilcox (Oracle) if (folio_test_workingset(folio)) 56819138349SMatthew Wilcox (Oracle) folio_set_workingset(newfolio); 56919138349SMatthew Wilcox (Oracle) if (folio_test_checked(folio)) 57019138349SMatthew Wilcox (Oracle) folio_set_checked(newfolio); 5716c287605SDavid Hildenbrand /* 5726c287605SDavid Hildenbrand * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via 5736c287605SDavid Hildenbrand * migration entries. We can still have PG_anon_exclusive set on an 5746c287605SDavid Hildenbrand * effectively unmapped and unreferenced first sub-pages of an 5756c287605SDavid Hildenbrand * anonymous THP: we can simply copy it here via PG_mappedtodisk. 5766c287605SDavid Hildenbrand */ 57719138349SMatthew Wilcox (Oracle) if (folio_test_mappedtodisk(folio)) 57819138349SMatthew Wilcox (Oracle) folio_set_mappedtodisk(newfolio); 579b20a3503SChristoph Lameter 5803417013eSMatthew Wilcox (Oracle) /* Move dirty on pages not done by folio_migrate_mapping() */ 58119138349SMatthew Wilcox (Oracle) if (folio_test_dirty(folio)) 58219138349SMatthew Wilcox (Oracle) folio_set_dirty(newfolio); 583b20a3503SChristoph Lameter 58419138349SMatthew Wilcox (Oracle) if (folio_test_young(folio)) 58519138349SMatthew Wilcox (Oracle) folio_set_young(newfolio); 58619138349SMatthew Wilcox (Oracle) if (folio_test_idle(folio)) 58719138349SMatthew Wilcox (Oracle) folio_set_idle(newfolio); 58833c3fc71SVladimir Davydov 5897851a45cSRik van Riel /* 5907851a45cSRik van Riel * Copy NUMA information to the new page, to prevent over-eager 5917851a45cSRik van Riel * future migrations of this same page. 5927851a45cSRik van Riel */ 59319138349SMatthew Wilcox (Oracle) cpupid = page_cpupid_xchg_last(&folio->page, -1); 59433024536SHuang Ying /* 59533024536SHuang Ying * For memory tiering mode, when migrate between slow and fast 59633024536SHuang Ying * memory node, reset cpupid, because that is used to record 59733024536SHuang Ying * page access time in slow memory node. 59833024536SHuang Ying */ 59933024536SHuang Ying if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) { 60033024536SHuang Ying bool f_toptier = node_is_toptier(page_to_nid(&folio->page)); 60133024536SHuang Ying bool t_toptier = node_is_toptier(page_to_nid(&newfolio->page)); 60233024536SHuang Ying 60333024536SHuang Ying if (f_toptier != t_toptier) 60433024536SHuang Ying cpupid = -1; 60533024536SHuang Ying } 60619138349SMatthew Wilcox (Oracle) page_cpupid_xchg_last(&newfolio->page, cpupid); 6077851a45cSRik van Riel 60819138349SMatthew Wilcox (Oracle) folio_migrate_ksm(newfolio, folio); 609c8d6553bSHugh Dickins /* 610c8d6553bSHugh Dickins * Please do not reorder this without considering how mm/ksm.c's 611c8d6553bSHugh Dickins * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). 612c8d6553bSHugh Dickins */ 61319138349SMatthew Wilcox (Oracle) if (folio_test_swapcache(folio)) 61419138349SMatthew Wilcox (Oracle) folio_clear_swapcache(folio); 61519138349SMatthew Wilcox (Oracle) folio_clear_private(folio); 616ad2fa371SMuchun Song 617ad2fa371SMuchun Song /* page->private contains hugetlb specific flags */ 61819138349SMatthew Wilcox (Oracle) if (!folio_test_hugetlb(folio)) 61919138349SMatthew Wilcox (Oracle) folio->private = NULL; 620b20a3503SChristoph Lameter 621b20a3503SChristoph Lameter /* 622b20a3503SChristoph Lameter * If any waiters have accumulated on the new page then 623b20a3503SChristoph Lameter * wake them up. 624b20a3503SChristoph Lameter */ 62519138349SMatthew Wilcox (Oracle) if (folio_test_writeback(newfolio)) 62619138349SMatthew Wilcox (Oracle) folio_end_writeback(newfolio); 627d435edcaSVlastimil Babka 6286aeff241SYang Shi /* 6296aeff241SYang Shi * PG_readahead shares the same bit with PG_reclaim. The above 6306aeff241SYang Shi * end_page_writeback() may clear PG_readahead mistakenly, so set the 6316aeff241SYang Shi * bit after that. 6326aeff241SYang Shi */ 63319138349SMatthew Wilcox (Oracle) if (folio_test_readahead(folio)) 63419138349SMatthew Wilcox (Oracle) folio_set_readahead(newfolio); 6356aeff241SYang Shi 63619138349SMatthew Wilcox (Oracle) folio_copy_owner(newfolio, folio); 63774485cf2SJohannes Weiner 63819138349SMatthew Wilcox (Oracle) if (!folio_test_hugetlb(folio)) 639d21bba2bSMatthew Wilcox (Oracle) mem_cgroup_migrate(folio, newfolio); 640b20a3503SChristoph Lameter } 64119138349SMatthew Wilcox (Oracle) EXPORT_SYMBOL(folio_migrate_flags); 6422916ecc0SJérôme Glisse 643715cbfd6SMatthew Wilcox (Oracle) void folio_migrate_copy(struct folio *newfolio, struct folio *folio) 6442916ecc0SJérôme Glisse { 645715cbfd6SMatthew Wilcox (Oracle) folio_copy(newfolio, folio); 646715cbfd6SMatthew Wilcox (Oracle) folio_migrate_flags(newfolio, folio); 6472916ecc0SJérôme Glisse } 648715cbfd6SMatthew Wilcox (Oracle) EXPORT_SYMBOL(folio_migrate_copy); 649b20a3503SChristoph Lameter 6501d8b85ccSChristoph Lameter /************************************************************ 6511d8b85ccSChristoph Lameter * Migration functions 6521d8b85ccSChristoph Lameter ***********************************************************/ 6531d8b85ccSChristoph Lameter 65416ce101dSAlistair Popple int migrate_folio_extra(struct address_space *mapping, struct folio *dst, 65516ce101dSAlistair Popple struct folio *src, enum migrate_mode mode, int extra_count) 65616ce101dSAlistair Popple { 65716ce101dSAlistair Popple int rc; 65816ce101dSAlistair Popple 65916ce101dSAlistair Popple BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */ 66016ce101dSAlistair Popple 66116ce101dSAlistair Popple rc = folio_migrate_mapping(mapping, dst, src, extra_count); 66216ce101dSAlistair Popple 66316ce101dSAlistair Popple if (rc != MIGRATEPAGE_SUCCESS) 66416ce101dSAlistair Popple return rc; 66516ce101dSAlistair Popple 66616ce101dSAlistair Popple if (mode != MIGRATE_SYNC_NO_COPY) 66716ce101dSAlistair Popple folio_migrate_copy(dst, src); 66816ce101dSAlistair Popple else 66916ce101dSAlistair Popple folio_migrate_flags(dst, src); 67016ce101dSAlistair Popple return MIGRATEPAGE_SUCCESS; 67116ce101dSAlistair Popple } 67216ce101dSAlistair Popple 67354184650SMatthew Wilcox (Oracle) /** 67454184650SMatthew Wilcox (Oracle) * migrate_folio() - Simple folio migration. 67554184650SMatthew Wilcox (Oracle) * @mapping: The address_space containing the folio. 67654184650SMatthew Wilcox (Oracle) * @dst: The folio to migrate the data to. 67754184650SMatthew Wilcox (Oracle) * @src: The folio containing the current data. 67854184650SMatthew Wilcox (Oracle) * @mode: How to migrate the page. 679b20a3503SChristoph Lameter * 68054184650SMatthew Wilcox (Oracle) * Common logic to directly migrate a single LRU folio suitable for 68154184650SMatthew Wilcox (Oracle) * folios that do not use PagePrivate/PagePrivate2. 68254184650SMatthew Wilcox (Oracle) * 68354184650SMatthew Wilcox (Oracle) * Folios are locked upon entry and exit. 684b20a3503SChristoph Lameter */ 68554184650SMatthew Wilcox (Oracle) int migrate_folio(struct address_space *mapping, struct folio *dst, 68654184650SMatthew Wilcox (Oracle) struct folio *src, enum migrate_mode mode) 687b20a3503SChristoph Lameter { 68816ce101dSAlistair Popple return migrate_folio_extra(mapping, dst, src, mode, 0); 689b20a3503SChristoph Lameter } 69054184650SMatthew Wilcox (Oracle) EXPORT_SYMBOL(migrate_folio); 691b20a3503SChristoph Lameter 6929361401eSDavid Howells #ifdef CONFIG_BLOCK 69384ade7c1SJan Kara /* Returns true if all buffers are successfully locked */ 69484ade7c1SJan Kara static bool buffer_migrate_lock_buffers(struct buffer_head *head, 69584ade7c1SJan Kara enum migrate_mode mode) 69684ade7c1SJan Kara { 69784ade7c1SJan Kara struct buffer_head *bh = head; 69884ade7c1SJan Kara 69984ade7c1SJan Kara /* Simple case, sync compaction */ 70084ade7c1SJan Kara if (mode != MIGRATE_ASYNC) { 70184ade7c1SJan Kara do { 70284ade7c1SJan Kara lock_buffer(bh); 70384ade7c1SJan Kara bh = bh->b_this_page; 70484ade7c1SJan Kara 70584ade7c1SJan Kara } while (bh != head); 70684ade7c1SJan Kara 70784ade7c1SJan Kara return true; 70884ade7c1SJan Kara } 70984ade7c1SJan Kara 71084ade7c1SJan Kara /* async case, we cannot block on lock_buffer so use trylock_buffer */ 71184ade7c1SJan Kara do { 71284ade7c1SJan Kara if (!trylock_buffer(bh)) { 71384ade7c1SJan Kara /* 71484ade7c1SJan Kara * We failed to lock the buffer and cannot stall in 71584ade7c1SJan Kara * async migration. Release the taken locks 71684ade7c1SJan Kara */ 71784ade7c1SJan Kara struct buffer_head *failed_bh = bh; 71884ade7c1SJan Kara bh = head; 71984ade7c1SJan Kara while (bh != failed_bh) { 72084ade7c1SJan Kara unlock_buffer(bh); 72184ade7c1SJan Kara bh = bh->b_this_page; 72284ade7c1SJan Kara } 72384ade7c1SJan Kara return false; 72484ade7c1SJan Kara } 72584ade7c1SJan Kara 72684ade7c1SJan Kara bh = bh->b_this_page; 72784ade7c1SJan Kara } while (bh != head); 72884ade7c1SJan Kara return true; 72984ade7c1SJan Kara } 73084ade7c1SJan Kara 73167235182SMatthew Wilcox (Oracle) static int __buffer_migrate_folio(struct address_space *mapping, 73267235182SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode, 73389cb0888SJan Kara bool check_refs) 7341d8b85ccSChristoph Lameter { 7351d8b85ccSChristoph Lameter struct buffer_head *bh, *head; 7361d8b85ccSChristoph Lameter int rc; 737cc4f11e6SJan Kara int expected_count; 7381d8b85ccSChristoph Lameter 73967235182SMatthew Wilcox (Oracle) head = folio_buffers(src); 74067235182SMatthew Wilcox (Oracle) if (!head) 74154184650SMatthew Wilcox (Oracle) return migrate_folio(mapping, dst, src, mode); 7421d8b85ccSChristoph Lameter 743cc4f11e6SJan Kara /* Check whether page does not have extra refs before we do more work */ 744108ca835SMatthew Wilcox (Oracle) expected_count = folio_expected_refs(mapping, src); 74567235182SMatthew Wilcox (Oracle) if (folio_ref_count(src) != expected_count) 746cc4f11e6SJan Kara return -EAGAIN; 747cc4f11e6SJan Kara 748cc4f11e6SJan Kara if (!buffer_migrate_lock_buffers(head, mode)) 749cc4f11e6SJan Kara return -EAGAIN; 7501d8b85ccSChristoph Lameter 75189cb0888SJan Kara if (check_refs) { 75289cb0888SJan Kara bool busy; 75389cb0888SJan Kara bool invalidated = false; 75489cb0888SJan Kara 75589cb0888SJan Kara recheck_buffers: 75689cb0888SJan Kara busy = false; 75789cb0888SJan Kara spin_lock(&mapping->private_lock); 75889cb0888SJan Kara bh = head; 75989cb0888SJan Kara do { 76089cb0888SJan Kara if (atomic_read(&bh->b_count)) { 76189cb0888SJan Kara busy = true; 76289cb0888SJan Kara break; 76389cb0888SJan Kara } 76489cb0888SJan Kara bh = bh->b_this_page; 76589cb0888SJan Kara } while (bh != head); 76689cb0888SJan Kara if (busy) { 76789cb0888SJan Kara if (invalidated) { 76889cb0888SJan Kara rc = -EAGAIN; 76989cb0888SJan Kara goto unlock_buffers; 77089cb0888SJan Kara } 771ebdf4de5SJan Kara spin_unlock(&mapping->private_lock); 77289cb0888SJan Kara invalidate_bh_lrus(); 77389cb0888SJan Kara invalidated = true; 77489cb0888SJan Kara goto recheck_buffers; 77589cb0888SJan Kara } 77689cb0888SJan Kara } 77789cb0888SJan Kara 77867235182SMatthew Wilcox (Oracle) rc = folio_migrate_mapping(mapping, dst, src, 0); 77978bd5209SRafael Aquini if (rc != MIGRATEPAGE_SUCCESS) 780cc4f11e6SJan Kara goto unlock_buffers; 7811d8b85ccSChristoph Lameter 78267235182SMatthew Wilcox (Oracle) folio_attach_private(dst, folio_detach_private(src)); 7831d8b85ccSChristoph Lameter 7841d8b85ccSChristoph Lameter bh = head; 7851d8b85ccSChristoph Lameter do { 78667235182SMatthew Wilcox (Oracle) set_bh_page(bh, &dst->page, bh_offset(bh)); 7871d8b85ccSChristoph Lameter bh = bh->b_this_page; 7881d8b85ccSChristoph Lameter } while (bh != head); 7891d8b85ccSChristoph Lameter 7902916ecc0SJérôme Glisse if (mode != MIGRATE_SYNC_NO_COPY) 79167235182SMatthew Wilcox (Oracle) folio_migrate_copy(dst, src); 7922916ecc0SJérôme Glisse else 79367235182SMatthew Wilcox (Oracle) folio_migrate_flags(dst, src); 7941d8b85ccSChristoph Lameter 795cc4f11e6SJan Kara rc = MIGRATEPAGE_SUCCESS; 796cc4f11e6SJan Kara unlock_buffers: 797ebdf4de5SJan Kara if (check_refs) 798ebdf4de5SJan Kara spin_unlock(&mapping->private_lock); 7991d8b85ccSChristoph Lameter bh = head; 8001d8b85ccSChristoph Lameter do { 8011d8b85ccSChristoph Lameter unlock_buffer(bh); 8021d8b85ccSChristoph Lameter bh = bh->b_this_page; 8031d8b85ccSChristoph Lameter } while (bh != head); 8041d8b85ccSChristoph Lameter 805cc4f11e6SJan Kara return rc; 8061d8b85ccSChristoph Lameter } 80789cb0888SJan Kara 80867235182SMatthew Wilcox (Oracle) /** 80967235182SMatthew Wilcox (Oracle) * buffer_migrate_folio() - Migration function for folios with buffers. 81067235182SMatthew Wilcox (Oracle) * @mapping: The address space containing @src. 81167235182SMatthew Wilcox (Oracle) * @dst: The folio to migrate to. 81267235182SMatthew Wilcox (Oracle) * @src: The folio to migrate from. 81367235182SMatthew Wilcox (Oracle) * @mode: How to migrate the folio. 81467235182SMatthew Wilcox (Oracle) * 81567235182SMatthew Wilcox (Oracle) * This function can only be used if the underlying filesystem guarantees 81667235182SMatthew Wilcox (Oracle) * that no other references to @src exist. For example attached buffer 81767235182SMatthew Wilcox (Oracle) * heads are accessed only under the folio lock. If your filesystem cannot 81867235182SMatthew Wilcox (Oracle) * provide this guarantee, buffer_migrate_folio_norefs() may be more 81967235182SMatthew Wilcox (Oracle) * appropriate. 82067235182SMatthew Wilcox (Oracle) * 82167235182SMatthew Wilcox (Oracle) * Return: 0 on success or a negative errno on failure. 82289cb0888SJan Kara */ 82367235182SMatthew Wilcox (Oracle) int buffer_migrate_folio(struct address_space *mapping, 82467235182SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode) 82589cb0888SJan Kara { 82667235182SMatthew Wilcox (Oracle) return __buffer_migrate_folio(mapping, dst, src, mode, false); 82789cb0888SJan Kara } 82867235182SMatthew Wilcox (Oracle) EXPORT_SYMBOL(buffer_migrate_folio); 82989cb0888SJan Kara 83067235182SMatthew Wilcox (Oracle) /** 83167235182SMatthew Wilcox (Oracle) * buffer_migrate_folio_norefs() - Migration function for folios with buffers. 83267235182SMatthew Wilcox (Oracle) * @mapping: The address space containing @src. 83367235182SMatthew Wilcox (Oracle) * @dst: The folio to migrate to. 83467235182SMatthew Wilcox (Oracle) * @src: The folio to migrate from. 83567235182SMatthew Wilcox (Oracle) * @mode: How to migrate the folio. 83667235182SMatthew Wilcox (Oracle) * 83767235182SMatthew Wilcox (Oracle) * Like buffer_migrate_folio() except that this variant is more careful 83867235182SMatthew Wilcox (Oracle) * and checks that there are also no buffer head references. This function 83967235182SMatthew Wilcox (Oracle) * is the right one for mappings where buffer heads are directly looked 84067235182SMatthew Wilcox (Oracle) * up and referenced (such as block device mappings). 84167235182SMatthew Wilcox (Oracle) * 84267235182SMatthew Wilcox (Oracle) * Return: 0 on success or a negative errno on failure. 84389cb0888SJan Kara */ 84467235182SMatthew Wilcox (Oracle) int buffer_migrate_folio_norefs(struct address_space *mapping, 84567235182SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode) 84689cb0888SJan Kara { 84767235182SMatthew Wilcox (Oracle) return __buffer_migrate_folio(mapping, dst, src, mode, true); 84889cb0888SJan Kara } 849e26355e2SJan Kara EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs); 8509361401eSDavid Howells #endif 8511d8b85ccSChristoph Lameter 8522ec810d5SMatthew Wilcox (Oracle) int filemap_migrate_folio(struct address_space *mapping, 8532ec810d5SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode) 8542ec810d5SMatthew Wilcox (Oracle) { 8552ec810d5SMatthew Wilcox (Oracle) int ret; 8562ec810d5SMatthew Wilcox (Oracle) 8572ec810d5SMatthew Wilcox (Oracle) ret = folio_migrate_mapping(mapping, dst, src, 0); 8582ec810d5SMatthew Wilcox (Oracle) if (ret != MIGRATEPAGE_SUCCESS) 8592ec810d5SMatthew Wilcox (Oracle) return ret; 8602ec810d5SMatthew Wilcox (Oracle) 8612ec810d5SMatthew Wilcox (Oracle) if (folio_get_private(src)) 8622ec810d5SMatthew Wilcox (Oracle) folio_attach_private(dst, folio_detach_private(src)); 8632ec810d5SMatthew Wilcox (Oracle) 8642ec810d5SMatthew Wilcox (Oracle) if (mode != MIGRATE_SYNC_NO_COPY) 8652ec810d5SMatthew Wilcox (Oracle) folio_migrate_copy(dst, src); 8662ec810d5SMatthew Wilcox (Oracle) else 8672ec810d5SMatthew Wilcox (Oracle) folio_migrate_flags(dst, src); 8682ec810d5SMatthew Wilcox (Oracle) return MIGRATEPAGE_SUCCESS; 8692ec810d5SMatthew Wilcox (Oracle) } 8702ec810d5SMatthew Wilcox (Oracle) EXPORT_SYMBOL_GPL(filemap_migrate_folio); 8712ec810d5SMatthew Wilcox (Oracle) 87204e62a29SChristoph Lameter /* 8732be7fa10SMatthew Wilcox (Oracle) * Writeback a folio to clean the dirty state 87404e62a29SChristoph Lameter */ 8752be7fa10SMatthew Wilcox (Oracle) static int writeout(struct address_space *mapping, struct folio *folio) 87604e62a29SChristoph Lameter { 87704e62a29SChristoph Lameter struct writeback_control wbc = { 87804e62a29SChristoph Lameter .sync_mode = WB_SYNC_NONE, 87904e62a29SChristoph Lameter .nr_to_write = 1, 88004e62a29SChristoph Lameter .range_start = 0, 88104e62a29SChristoph Lameter .range_end = LLONG_MAX, 88204e62a29SChristoph Lameter .for_reclaim = 1 88304e62a29SChristoph Lameter }; 88404e62a29SChristoph Lameter int rc; 88504e62a29SChristoph Lameter 88604e62a29SChristoph Lameter if (!mapping->a_ops->writepage) 88704e62a29SChristoph Lameter /* No write method for the address space */ 88804e62a29SChristoph Lameter return -EINVAL; 88904e62a29SChristoph Lameter 8902be7fa10SMatthew Wilcox (Oracle) if (!folio_clear_dirty_for_io(folio)) 89104e62a29SChristoph Lameter /* Someone else already triggered a write */ 89204e62a29SChristoph Lameter return -EAGAIN; 89304e62a29SChristoph Lameter 89404e62a29SChristoph Lameter /* 8952be7fa10SMatthew Wilcox (Oracle) * A dirty folio may imply that the underlying filesystem has 8962be7fa10SMatthew Wilcox (Oracle) * the folio on some queue. So the folio must be clean for 8972be7fa10SMatthew Wilcox (Oracle) * migration. Writeout may mean we lose the lock and the 8982be7fa10SMatthew Wilcox (Oracle) * folio state is no longer what we checked for earlier. 89904e62a29SChristoph Lameter * At this point we know that the migration attempt cannot 90004e62a29SChristoph Lameter * be successful. 90104e62a29SChristoph Lameter */ 9024eecb8b9SMatthew Wilcox (Oracle) remove_migration_ptes(folio, folio, false); 90304e62a29SChristoph Lameter 9042be7fa10SMatthew Wilcox (Oracle) rc = mapping->a_ops->writepage(&folio->page, &wbc); 90504e62a29SChristoph Lameter 90604e62a29SChristoph Lameter if (rc != AOP_WRITEPAGE_ACTIVATE) 90704e62a29SChristoph Lameter /* unlocked. Relock */ 9082be7fa10SMatthew Wilcox (Oracle) folio_lock(folio); 90904e62a29SChristoph Lameter 910bda8550dSHugh Dickins return (rc < 0) ? -EIO : -EAGAIN; 91104e62a29SChristoph Lameter } 91204e62a29SChristoph Lameter 91304e62a29SChristoph Lameter /* 91404e62a29SChristoph Lameter * Default handling if a filesystem does not provide a migration function. 91504e62a29SChristoph Lameter */ 9168faa8ef5SMatthew Wilcox (Oracle) static int fallback_migrate_folio(struct address_space *mapping, 9178faa8ef5SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode) 9188351a6e4SChristoph Lameter { 9198faa8ef5SMatthew Wilcox (Oracle) if (folio_test_dirty(src)) { 9208faa8ef5SMatthew Wilcox (Oracle) /* Only writeback folios in full synchronous migration */ 9212916ecc0SJérôme Glisse switch (mode) { 9222916ecc0SJérôme Glisse case MIGRATE_SYNC: 9232916ecc0SJérôme Glisse case MIGRATE_SYNC_NO_COPY: 9242916ecc0SJérôme Glisse break; 9252916ecc0SJérôme Glisse default: 926b969c4abSMel Gorman return -EBUSY; 9272916ecc0SJérôme Glisse } 9282be7fa10SMatthew Wilcox (Oracle) return writeout(mapping, src); 929b969c4abSMel Gorman } 9308351a6e4SChristoph Lameter 9318351a6e4SChristoph Lameter /* 9328351a6e4SChristoph Lameter * Buffers may be managed in a filesystem specific way. 9338351a6e4SChristoph Lameter * We must have no buffers or drop them. 9348351a6e4SChristoph Lameter */ 9358faa8ef5SMatthew Wilcox (Oracle) if (folio_test_private(src) && 9368faa8ef5SMatthew Wilcox (Oracle) !filemap_release_folio(src, GFP_KERNEL)) 937806031bbSMel Gorman return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; 9388351a6e4SChristoph Lameter 93954184650SMatthew Wilcox (Oracle) return migrate_folio(mapping, dst, src, mode); 9408351a6e4SChristoph Lameter } 9418351a6e4SChristoph Lameter 9421d8b85ccSChristoph Lameter /* 943e24f0b8fSChristoph Lameter * Move a page to a newly allocated page 944e24f0b8fSChristoph Lameter * The page is locked and all ptes have been successfully removed. 945b20a3503SChristoph Lameter * 946e24f0b8fSChristoph Lameter * The new page will have replaced the old page if this function 947e24f0b8fSChristoph Lameter * is successful. 948894bc310SLee Schermerhorn * 949894bc310SLee Schermerhorn * Return value: 950894bc310SLee Schermerhorn * < 0 - error code 95178bd5209SRafael Aquini * MIGRATEPAGE_SUCCESS - success 952b20a3503SChristoph Lameter */ 953e7e3ffebSMatthew Wilcox (Oracle) static int move_to_new_folio(struct folio *dst, struct folio *src, 9545c3f9a67SHugh Dickins enum migrate_mode mode) 955b20a3503SChristoph Lameter { 956bda807d4SMinchan Kim int rc = -EAGAIN; 957e7e3ffebSMatthew Wilcox (Oracle) bool is_lru = !__PageMovable(&src->page); 958b20a3503SChristoph Lameter 959e7e3ffebSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(src), src); 960e7e3ffebSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst); 961b20a3503SChristoph Lameter 962bda807d4SMinchan Kim if (likely(is_lru)) { 96368f2736aSMatthew Wilcox (Oracle) struct address_space *mapping = folio_mapping(src); 96468f2736aSMatthew Wilcox (Oracle) 965b20a3503SChristoph Lameter if (!mapping) 96654184650SMatthew Wilcox (Oracle) rc = migrate_folio(mapping, dst, src, mode); 9675490da4fSMatthew Wilcox (Oracle) else if (mapping->a_ops->migrate_folio) 968b20a3503SChristoph Lameter /* 9695490da4fSMatthew Wilcox (Oracle) * Most folios have a mapping and most filesystems 9705490da4fSMatthew Wilcox (Oracle) * provide a migrate_folio callback. Anonymous folios 971bda807d4SMinchan Kim * are part of swap space which also has its own 9725490da4fSMatthew Wilcox (Oracle) * migrate_folio callback. This is the most common path 973bda807d4SMinchan Kim * for page migration. 974b20a3503SChristoph Lameter */ 9755490da4fSMatthew Wilcox (Oracle) rc = mapping->a_ops->migrate_folio(mapping, dst, src, 9765490da4fSMatthew Wilcox (Oracle) mode); 9778351a6e4SChristoph Lameter else 9788faa8ef5SMatthew Wilcox (Oracle) rc = fallback_migrate_folio(mapping, dst, src, mode); 979bda807d4SMinchan Kim } else { 98068f2736aSMatthew Wilcox (Oracle) const struct movable_operations *mops; 98168f2736aSMatthew Wilcox (Oracle) 982bda807d4SMinchan Kim /* 983bda807d4SMinchan Kim * In case of non-lru page, it could be released after 984bda807d4SMinchan Kim * isolation step. In that case, we shouldn't try migration. 985bda807d4SMinchan Kim */ 986e7e3ffebSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_isolated(src), src); 987e7e3ffebSMatthew Wilcox (Oracle) if (!folio_test_movable(src)) { 988bda807d4SMinchan Kim rc = MIGRATEPAGE_SUCCESS; 989e7e3ffebSMatthew Wilcox (Oracle) folio_clear_isolated(src); 990bda807d4SMinchan Kim goto out; 991bda807d4SMinchan Kim } 992bda807d4SMinchan Kim 99368f2736aSMatthew Wilcox (Oracle) mops = page_movable_ops(&src->page); 99468f2736aSMatthew Wilcox (Oracle) rc = mops->migrate_page(&dst->page, &src->page, mode); 995bda807d4SMinchan Kim WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && 996e7e3ffebSMatthew Wilcox (Oracle) !folio_test_isolated(src)); 997bda807d4SMinchan Kim } 998b20a3503SChristoph Lameter 9995c3f9a67SHugh Dickins /* 1000e7e3ffebSMatthew Wilcox (Oracle) * When successful, old pagecache src->mapping must be cleared before 1001e7e3ffebSMatthew Wilcox (Oracle) * src is freed; but stats require that PageAnon be left as PageAnon. 10025c3f9a67SHugh Dickins */ 10035c3f9a67SHugh Dickins if (rc == MIGRATEPAGE_SUCCESS) { 1004e7e3ffebSMatthew Wilcox (Oracle) if (__PageMovable(&src->page)) { 1005e7e3ffebSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_isolated(src), src); 1006bda807d4SMinchan Kim 1007bda807d4SMinchan Kim /* 1008bda807d4SMinchan Kim * We clear PG_movable under page_lock so any compactor 1009bda807d4SMinchan Kim * cannot try to migrate this page. 1010bda807d4SMinchan Kim */ 1011e7e3ffebSMatthew Wilcox (Oracle) folio_clear_isolated(src); 1012bda807d4SMinchan Kim } 1013bda807d4SMinchan Kim 1014bda807d4SMinchan Kim /* 1015e7e3ffebSMatthew Wilcox (Oracle) * Anonymous and movable src->mapping will be cleared by 1016bda807d4SMinchan Kim * free_pages_prepare so don't reset it here for keeping 1017bda807d4SMinchan Kim * the type to work PageAnon, for example. 1018bda807d4SMinchan Kim */ 1019e7e3ffebSMatthew Wilcox (Oracle) if (!folio_mapping_flags(src)) 1020e7e3ffebSMatthew Wilcox (Oracle) src->mapping = NULL; 1021d2b2c6ddSLars Persson 1022e7e3ffebSMatthew Wilcox (Oracle) if (likely(!folio_is_zone_device(dst))) 1023e7e3ffebSMatthew Wilcox (Oracle) flush_dcache_folio(dst); 10243fe2011fSMel Gorman } 1025bda807d4SMinchan Kim out: 1026e24f0b8fSChristoph Lameter return rc; 1027e24f0b8fSChristoph Lameter } 1028e24f0b8fSChristoph Lameter 1029682a71a1SMatthew Wilcox (Oracle) static int __unmap_and_move(struct folio *src, struct folio *dst, 10309c620e2bSHugh Dickins int force, enum migrate_mode mode) 1031e24f0b8fSChristoph Lameter { 10320dabec93SMinchan Kim int rc = -EAGAIN; 1033213ecb31SBaolin Wang bool page_was_mapped = false; 10343f6c8272SMel Gorman struct anon_vma *anon_vma = NULL; 1035682a71a1SMatthew Wilcox (Oracle) bool is_lru = !__PageMovable(&src->page); 103695a402c3SChristoph Lameter 1037682a71a1SMatthew Wilcox (Oracle) if (!folio_trylock(src)) { 1038a6bc32b8SMel Gorman if (!force || mode == MIGRATE_ASYNC) 10390dabec93SMinchan Kim goto out; 10403e7d3449SMel Gorman 10413e7d3449SMel Gorman /* 10423e7d3449SMel Gorman * It's not safe for direct compaction to call lock_page. 10433e7d3449SMel Gorman * For example, during page readahead pages are added locked 10443e7d3449SMel Gorman * to the LRU. Later, when the IO completes the pages are 10453e7d3449SMel Gorman * marked uptodate and unlocked. However, the queueing 10463e7d3449SMel Gorman * could be merging multiple pages for one bio (e.g. 1047d4388340SMatthew Wilcox (Oracle) * mpage_readahead). If an allocation happens for the 10483e7d3449SMel Gorman * second or third page, the process can end up locking 10493e7d3449SMel Gorman * the same page twice and deadlocking. Rather than 10503e7d3449SMel Gorman * trying to be clever about what pages can be locked, 10513e7d3449SMel Gorman * avoid the use of lock_page for direct compaction 10523e7d3449SMel Gorman * altogether. 10533e7d3449SMel Gorman */ 10543e7d3449SMel Gorman if (current->flags & PF_MEMALLOC) 10550dabec93SMinchan Kim goto out; 10563e7d3449SMel Gorman 1057682a71a1SMatthew Wilcox (Oracle) folio_lock(src); 1058e24f0b8fSChristoph Lameter } 1059e24f0b8fSChristoph Lameter 1060682a71a1SMatthew Wilcox (Oracle) if (folio_test_writeback(src)) { 106111bc82d6SAndrea Arcangeli /* 1062fed5b64aSJianguo Wu * Only in the case of a full synchronous migration is it 1063a6bc32b8SMel Gorman * necessary to wait for PageWriteback. In the async case, 1064a6bc32b8SMel Gorman * the retry loop is too short and in the sync-light case, 1065a6bc32b8SMel Gorman * the overhead of stalling is too much 106611bc82d6SAndrea Arcangeli */ 10672916ecc0SJérôme Glisse switch (mode) { 10682916ecc0SJérôme Glisse case MIGRATE_SYNC: 10692916ecc0SJérôme Glisse case MIGRATE_SYNC_NO_COPY: 10702916ecc0SJérôme Glisse break; 10712916ecc0SJérôme Glisse default: 107211bc82d6SAndrea Arcangeli rc = -EBUSY; 10730a31bc97SJohannes Weiner goto out_unlock; 107411bc82d6SAndrea Arcangeli } 107511bc82d6SAndrea Arcangeli if (!force) 10760a31bc97SJohannes Weiner goto out_unlock; 1077682a71a1SMatthew Wilcox (Oracle) folio_wait_writeback(src); 1078e24f0b8fSChristoph Lameter } 107903f15c86SHugh Dickins 1080e24f0b8fSChristoph Lameter /* 1081682a71a1SMatthew Wilcox (Oracle) * By try_to_migrate(), src->mapcount goes down to 0 here. In this case, 1082682a71a1SMatthew Wilcox (Oracle) * we cannot notice that anon_vma is freed while we migrate a page. 10831ce82b69SHugh Dickins * This get_anon_vma() delays freeing anon_vma pointer until the end 1084dc386d4dSKAMEZAWA Hiroyuki * of migration. File cache pages are no problem because of page_lock() 1085989f89c5SKAMEZAWA Hiroyuki * File Caches may use write_page() or lock_page() in migration, then, 1086989f89c5SKAMEZAWA Hiroyuki * just care Anon page here. 10873fe2011fSMel Gorman * 108829eea9b5SMatthew Wilcox (Oracle) * Only folio_get_anon_vma() understands the subtleties of 108903f15c86SHugh Dickins * getting a hold on an anon_vma from outside one of its mms. 109003f15c86SHugh Dickins * But if we cannot get anon_vma, then we won't need it anyway, 109103f15c86SHugh Dickins * because that implies that the anon page is no longer mapped 109203f15c86SHugh Dickins * (and cannot be remapped so long as we hold the page lock). 10933fe2011fSMel Gorman */ 1094682a71a1SMatthew Wilcox (Oracle) if (folio_test_anon(src) && !folio_test_ksm(src)) 109529eea9b5SMatthew Wilcox (Oracle) anon_vma = folio_get_anon_vma(src); 109662e1c553SShaohua Li 10977db7671fSHugh Dickins /* 10987db7671fSHugh Dickins * Block others from accessing the new page when we get around to 10997db7671fSHugh Dickins * establishing additional references. We are usually the only one 1100682a71a1SMatthew Wilcox (Oracle) * holding a reference to dst at this point. We used to have a BUG 1101682a71a1SMatthew Wilcox (Oracle) * here if folio_trylock(dst) fails, but would like to allow for 1102682a71a1SMatthew Wilcox (Oracle) * cases where there might be a race with the previous use of dst. 11037db7671fSHugh Dickins * This is much like races on refcount of oldpage: just don't BUG(). 11047db7671fSHugh Dickins */ 1105682a71a1SMatthew Wilcox (Oracle) if (unlikely(!folio_trylock(dst))) 11067db7671fSHugh Dickins goto out_unlock; 11077db7671fSHugh Dickins 1108bda807d4SMinchan Kim if (unlikely(!is_lru)) { 1109682a71a1SMatthew Wilcox (Oracle) rc = move_to_new_folio(dst, src, mode); 1110bda807d4SMinchan Kim goto out_unlock_both; 1111bda807d4SMinchan Kim } 1112bda807d4SMinchan Kim 1113dc386d4dSKAMEZAWA Hiroyuki /* 111462e1c553SShaohua Li * Corner case handling: 111562e1c553SShaohua Li * 1. When a new swap-cache page is read into, it is added to the LRU 111662e1c553SShaohua Li * and treated as swapcache but it has no rmap yet. 1117682a71a1SMatthew Wilcox (Oracle) * Calling try_to_unmap() against a src->mapping==NULL page will 111862e1c553SShaohua Li * trigger a BUG. So handle it here. 1119d12b8951SYang Shi * 2. An orphaned page (see truncate_cleanup_page) might have 112062e1c553SShaohua Li * fs-private metadata. The page can be picked up due to memory 112162e1c553SShaohua Li * offlining. Everywhere else except page reclaim, the page is 112262e1c553SShaohua Li * invisible to the vm, so the page can not be migrated. So try to 112362e1c553SShaohua Li * free the metadata, so the page can be freed. 1124dc386d4dSKAMEZAWA Hiroyuki */ 1125682a71a1SMatthew Wilcox (Oracle) if (!src->mapping) { 1126682a71a1SMatthew Wilcox (Oracle) if (folio_test_private(src)) { 1127682a71a1SMatthew Wilcox (Oracle) try_to_free_buffers(src); 11287db7671fSHugh Dickins goto out_unlock_both; 112962e1c553SShaohua Li } 1130682a71a1SMatthew Wilcox (Oracle) } else if (folio_mapped(src)) { 11317db7671fSHugh Dickins /* Establish migration ptes */ 1132682a71a1SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_test_anon(src) && 1133682a71a1SMatthew Wilcox (Oracle) !folio_test_ksm(src) && !anon_vma, src); 1134682a71a1SMatthew Wilcox (Oracle) try_to_migrate(src, 0); 1135213ecb31SBaolin Wang page_was_mapped = true; 11362ebba6b7SHugh Dickins } 1137dc386d4dSKAMEZAWA Hiroyuki 1138682a71a1SMatthew Wilcox (Oracle) if (!folio_mapped(src)) 1139682a71a1SMatthew Wilcox (Oracle) rc = move_to_new_folio(dst, src, mode); 1140e24f0b8fSChristoph Lameter 1141c3096e67SHugh Dickins /* 1142682a71a1SMatthew Wilcox (Oracle) * When successful, push dst to LRU immediately: so that if it 1143c3096e67SHugh Dickins * turns out to be an mlocked page, remove_migration_ptes() will 1144682a71a1SMatthew Wilcox (Oracle) * automatically build up the correct dst->mlock_count for it. 1145c3096e67SHugh Dickins * 1146c3096e67SHugh Dickins * We would like to do something similar for the old page, when 1147c3096e67SHugh Dickins * unsuccessful, and other cases when a page has been temporarily 1148c3096e67SHugh Dickins * isolated from the unevictable LRU: but this case is the easiest. 1149c3096e67SHugh Dickins */ 1150c3096e67SHugh Dickins if (rc == MIGRATEPAGE_SUCCESS) { 1151682a71a1SMatthew Wilcox (Oracle) folio_add_lru(dst); 11525c3f9a67SHugh Dickins if (page_was_mapped) 1153c3096e67SHugh Dickins lru_add_drain(); 1154c3096e67SHugh Dickins } 1155c3096e67SHugh Dickins 11565c3f9a67SHugh Dickins if (page_was_mapped) 1157682a71a1SMatthew Wilcox (Oracle) remove_migration_ptes(src, 1158682a71a1SMatthew Wilcox (Oracle) rc == MIGRATEPAGE_SUCCESS ? dst : src, false); 11593f6c8272SMel Gorman 11607db7671fSHugh Dickins out_unlock_both: 1161682a71a1SMatthew Wilcox (Oracle) folio_unlock(dst); 11627db7671fSHugh Dickins out_unlock: 11633f6c8272SMel Gorman /* Drop an anon_vma reference if we took one */ 116476545066SRik van Riel if (anon_vma) 11659e60109fSPeter Zijlstra put_anon_vma(anon_vma); 1166682a71a1SMatthew Wilcox (Oracle) folio_unlock(src); 11670dabec93SMinchan Kim out: 1168c6c919ebSMinchan Kim /* 1169682a71a1SMatthew Wilcox (Oracle) * If migration is successful, decrease refcount of dst, 1170c6c919ebSMinchan Kim * which will not free the page because new page owner increased 1171c3096e67SHugh Dickins * refcounter. 1172c6c919ebSMinchan Kim */ 1173c3096e67SHugh Dickins if (rc == MIGRATEPAGE_SUCCESS) 1174682a71a1SMatthew Wilcox (Oracle) folio_put(dst); 1175c6c919ebSMinchan Kim 11760dabec93SMinchan Kim return rc; 11770dabec93SMinchan Kim } 117895a402c3SChristoph Lameter 11790dabec93SMinchan Kim /* 118049f51859SHuang Ying * Obtain the lock on folio, remove all ptes and migrate the folio 118149f51859SHuang Ying * to the newly allocated folio in dst. 11820dabec93SMinchan Kim */ 11836ec4476aSLinus Torvalds static int unmap_and_move(new_page_t get_new_page, 1184ef2a5153SGeert Uytterhoeven free_page_t put_new_page, 118549f51859SHuang Ying unsigned long private, struct folio *src, 1186add05cecSNaoya Horiguchi int force, enum migrate_mode mode, 1187dd4ae78aSYang Shi enum migrate_reason reason, 1188dd4ae78aSYang Shi struct list_head *ret) 11890dabec93SMinchan Kim { 119049f51859SHuang Ying struct folio *dst; 11912def7424SHugh Dickins int rc = MIGRATEPAGE_SUCCESS; 119274d4a579SYang Shi struct page *newpage = NULL; 11930dabec93SMinchan Kim 119449f51859SHuang Ying if (!thp_migration_supported() && folio_test_transhuge(src)) 1195d532e2e5SYang Shi return -ENOSYS; 119694723aafSMichal Hocko 119749f51859SHuang Ying if (folio_ref_count(src) == 1) { 119849f51859SHuang Ying /* Folio was freed from under us. So we are done. */ 119949f51859SHuang Ying folio_clear_active(src); 120049f51859SHuang Ying folio_clear_unevictable(src); 1201160088b3SMiaohe Lin /* free_pages_prepare() will clear PG_isolated. */ 12020dabec93SMinchan Kim goto out; 12030dabec93SMinchan Kim } 12040dabec93SMinchan Kim 120549f51859SHuang Ying newpage = get_new_page(&src->page, private); 120674d4a579SYang Shi if (!newpage) 120774d4a579SYang Shi return -ENOMEM; 1208682a71a1SMatthew Wilcox (Oracle) dst = page_folio(newpage); 120974d4a579SYang Shi 12104c74b65fSYang Li dst->private = NULL; 1211682a71a1SMatthew Wilcox (Oracle) rc = __unmap_and_move(src, dst, force, mode); 1212c6c919ebSMinchan Kim if (rc == MIGRATEPAGE_SUCCESS) 121349f51859SHuang Ying set_page_owner_migrate_reason(&dst->page, reason); 1214bf6bddf1SRafael Aquini 12150dabec93SMinchan Kim out: 1216e24f0b8fSChristoph Lameter if (rc != -EAGAIN) { 1217aaa994b3SChristoph Lameter /* 121849f51859SHuang Ying * A folio that has been migrated has all references 121949f51859SHuang Ying * removed and will be freed. A folio that has not been 1220c23a0c99SRalph Campbell * migrated will have kept its references and be restored. 1221aaa994b3SChristoph Lameter */ 122249f51859SHuang Ying list_del(&src->lru); 1223e24f0b8fSChristoph Lameter } 122468711a74SDavid Rientjes 122595a402c3SChristoph Lameter /* 1226c6c919ebSMinchan Kim * If migration is successful, releases reference grabbed during 122749f51859SHuang Ying * isolation. Otherwise, restore the folio to right list unless 1228c6c919ebSMinchan Kim * we want to retry. 122995a402c3SChristoph Lameter */ 1230c6c919ebSMinchan Kim if (rc == MIGRATEPAGE_SUCCESS) { 1231dd4ae78aSYang Shi /* 123249f51859SHuang Ying * Compaction can migrate also non-LRU folios which are 1233dd4ae78aSYang Shi * not accounted to NR_ISOLATED_*. They can be recognized 123449f51859SHuang Ying * as __folio_test_movable 1235dd4ae78aSYang Shi */ 123649f51859SHuang Ying if (likely(!__folio_test_movable(src))) 123749f51859SHuang Ying mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON + 123849f51859SHuang Ying folio_is_file_lru(src), -folio_nr_pages(src)); 1239dd4ae78aSYang Shi 124079f5f8faSOscar Salvador if (reason != MR_MEMORY_FAILURE) 1241c6c919ebSMinchan Kim /* 124249f51859SHuang Ying * We release the folio in page_handle_poison. 1243c6c919ebSMinchan Kim */ 124449f51859SHuang Ying folio_put(src); 1245c6c919ebSMinchan Kim } else { 1246dd4ae78aSYang Shi if (rc != -EAGAIN) 124749f51859SHuang Ying list_add_tail(&src->lru, ret); 1248bda807d4SMinchan Kim 1249cf4b769aSHugh Dickins if (put_new_page) 125049f51859SHuang Ying put_new_page(&dst->page, private); 1251c6c919ebSMinchan Kim else 125249f51859SHuang Ying folio_put(dst); 1253c6c919ebSMinchan Kim } 125468711a74SDavid Rientjes 1255e24f0b8fSChristoph Lameter return rc; 1256e24f0b8fSChristoph Lameter } 1257b20a3503SChristoph Lameter 1258e24f0b8fSChristoph Lameter /* 1259290408d4SNaoya Horiguchi * Counterpart of unmap_and_move_page() for hugepage migration. 1260290408d4SNaoya Horiguchi * 1261290408d4SNaoya Horiguchi * This function doesn't wait the completion of hugepage I/O 1262290408d4SNaoya Horiguchi * because there is no race between I/O and migration for hugepage. 1263290408d4SNaoya Horiguchi * Note that currently hugepage I/O occurs only in direct I/O 1264290408d4SNaoya Horiguchi * where no lock is held and PG_writeback is irrelevant, 1265290408d4SNaoya Horiguchi * and writeback status of all subpages are counted in the reference 1266290408d4SNaoya Horiguchi * count of the head page (i.e. if all subpages of a 2MB hugepage are 1267290408d4SNaoya Horiguchi * under direct I/O, the reference of the head page is 512 and a bit more.) 1268290408d4SNaoya Horiguchi * This means that when we try to migrate hugepage whose subpages are 1269290408d4SNaoya Horiguchi * doing direct I/O, some references remain after try_to_unmap() and 1270290408d4SNaoya Horiguchi * hugepage migration fails without data corruption. 1271290408d4SNaoya Horiguchi * 1272290408d4SNaoya Horiguchi * There is also no race when direct I/O is issued on the page under migration, 1273290408d4SNaoya Horiguchi * because then pte is replaced with migration swap entry and direct I/O code 1274290408d4SNaoya Horiguchi * will wait in the page fault for migration to complete. 1275290408d4SNaoya Horiguchi */ 1276290408d4SNaoya Horiguchi static int unmap_and_move_huge_page(new_page_t get_new_page, 127768711a74SDavid Rientjes free_page_t put_new_page, unsigned long private, 127868711a74SDavid Rientjes struct page *hpage, int force, 1279dd4ae78aSYang Shi enum migrate_mode mode, int reason, 1280dd4ae78aSYang Shi struct list_head *ret) 1281290408d4SNaoya Horiguchi { 12824eecb8b9SMatthew Wilcox (Oracle) struct folio *dst, *src = page_folio(hpage); 12832def7424SHugh Dickins int rc = -EAGAIN; 12842ebba6b7SHugh Dickins int page_was_mapped = 0; 128532665f2bSJoonsoo Kim struct page *new_hpage; 1286290408d4SNaoya Horiguchi struct anon_vma *anon_vma = NULL; 1287c0d0381aSMike Kravetz struct address_space *mapping = NULL; 1288290408d4SNaoya Horiguchi 128983467efbSNaoya Horiguchi /* 12907ed2c31dSAnshuman Khandual * Migratability of hugepages depends on architectures and their size. 129183467efbSNaoya Horiguchi * This check is necessary because some callers of hugepage migration 129283467efbSNaoya Horiguchi * like soft offline and memory hotremove don't walk through page 129383467efbSNaoya Horiguchi * tables or check whether the hugepage is pmd-based or not before 129483467efbSNaoya Horiguchi * kicking migration. 129583467efbSNaoya Horiguchi */ 1296577be05cSHuang Ying if (!hugepage_migration_supported(page_hstate(hpage))) 129783467efbSNaoya Horiguchi return -ENOSYS; 129883467efbSNaoya Horiguchi 1299c33db292SMatthew Wilcox (Oracle) if (folio_ref_count(src) == 1) { 130071a64f61SMuchun Song /* page was freed from under us. So we are done. */ 130171a64f61SMuchun Song putback_active_hugepage(hpage); 130271a64f61SMuchun Song return MIGRATEPAGE_SUCCESS; 130371a64f61SMuchun Song } 130471a64f61SMuchun Song 1305666feb21SMichal Hocko new_hpage = get_new_page(hpage, private); 1306290408d4SNaoya Horiguchi if (!new_hpage) 1307290408d4SNaoya Horiguchi return -ENOMEM; 13084eecb8b9SMatthew Wilcox (Oracle) dst = page_folio(new_hpage); 1309290408d4SNaoya Horiguchi 1310c33db292SMatthew Wilcox (Oracle) if (!folio_trylock(src)) { 13112916ecc0SJérôme Glisse if (!force) 1312290408d4SNaoya Horiguchi goto out; 13132916ecc0SJérôme Glisse switch (mode) { 13142916ecc0SJérôme Glisse case MIGRATE_SYNC: 13152916ecc0SJérôme Glisse case MIGRATE_SYNC_NO_COPY: 13162916ecc0SJérôme Glisse break; 13172916ecc0SJérôme Glisse default: 13182916ecc0SJérôme Glisse goto out; 13192916ecc0SJérôme Glisse } 1320c33db292SMatthew Wilcox (Oracle) folio_lock(src); 1321290408d4SNaoya Horiguchi } 1322290408d4SNaoya Horiguchi 1323cb6acd01SMike Kravetz /* 1324cb6acd01SMike Kravetz * Check for pages which are in the process of being freed. Without 1325c33db292SMatthew Wilcox (Oracle) * folio_mapping() set, hugetlbfs specific move page routine will not 1326cb6acd01SMike Kravetz * be called and we could leak usage counts for subpools. 1327cb6acd01SMike Kravetz */ 1328345c62d1SSidhartha Kumar if (hugetlb_folio_subpool(src) && !folio_mapping(src)) { 1329cb6acd01SMike Kravetz rc = -EBUSY; 1330cb6acd01SMike Kravetz goto out_unlock; 1331cb6acd01SMike Kravetz } 1332cb6acd01SMike Kravetz 1333c33db292SMatthew Wilcox (Oracle) if (folio_test_anon(src)) 133429eea9b5SMatthew Wilcox (Oracle) anon_vma = folio_get_anon_vma(src); 1335290408d4SNaoya Horiguchi 1336c33db292SMatthew Wilcox (Oracle) if (unlikely(!folio_trylock(dst))) 13377db7671fSHugh Dickins goto put_anon; 13387db7671fSHugh Dickins 1339c33db292SMatthew Wilcox (Oracle) if (folio_mapped(src)) { 1340a98a2f0cSAlistair Popple enum ttu_flags ttu = 0; 1341336bf30eSMike Kravetz 1342c33db292SMatthew Wilcox (Oracle) if (!folio_test_anon(src)) { 1343c0d0381aSMike Kravetz /* 1344336bf30eSMike Kravetz * In shared mappings, try_to_unmap could potentially 1345336bf30eSMike Kravetz * call huge_pmd_unshare. Because of this, take 1346336bf30eSMike Kravetz * semaphore in write mode here and set TTU_RMAP_LOCKED 1347336bf30eSMike Kravetz * to let lower levels know we have taken the lock. 1348c0d0381aSMike Kravetz */ 1349c0d0381aSMike Kravetz mapping = hugetlb_page_mapping_lock_write(hpage); 1350c0d0381aSMike Kravetz if (unlikely(!mapping)) 1351c0d0381aSMike Kravetz goto unlock_put_anon; 1352c0d0381aSMike Kravetz 13535202978bSMiaohe Lin ttu = TTU_RMAP_LOCKED; 1354336bf30eSMike Kravetz } 1355336bf30eSMike Kravetz 13564b8554c5SMatthew Wilcox (Oracle) try_to_migrate(src, ttu); 13572ebba6b7SHugh Dickins page_was_mapped = 1; 1358336bf30eSMike Kravetz 13595202978bSMiaohe Lin if (ttu & TTU_RMAP_LOCKED) 1360336bf30eSMike Kravetz i_mmap_unlock_write(mapping); 13612ebba6b7SHugh Dickins } 1362290408d4SNaoya Horiguchi 1363c33db292SMatthew Wilcox (Oracle) if (!folio_mapped(src)) 1364e7e3ffebSMatthew Wilcox (Oracle) rc = move_to_new_folio(dst, src, mode); 1365290408d4SNaoya Horiguchi 1366336bf30eSMike Kravetz if (page_was_mapped) 13674eecb8b9SMatthew Wilcox (Oracle) remove_migration_ptes(src, 13684eecb8b9SMatthew Wilcox (Oracle) rc == MIGRATEPAGE_SUCCESS ? dst : src, false); 1369290408d4SNaoya Horiguchi 1370c0d0381aSMike Kravetz unlock_put_anon: 1371c33db292SMatthew Wilcox (Oracle) folio_unlock(dst); 13727db7671fSHugh Dickins 13737db7671fSHugh Dickins put_anon: 1374fd4a4663SHugh Dickins if (anon_vma) 13759e60109fSPeter Zijlstra put_anon_vma(anon_vma); 13768e6ac7faSAneesh Kumar K.V 13772def7424SHugh Dickins if (rc == MIGRATEPAGE_SUCCESS) { 1378345c62d1SSidhartha Kumar move_hugetlb_state(src, dst, reason); 13792def7424SHugh Dickins put_new_page = NULL; 13802def7424SHugh Dickins } 13818e6ac7faSAneesh Kumar K.V 1382cb6acd01SMike Kravetz out_unlock: 1383c33db292SMatthew Wilcox (Oracle) folio_unlock(src); 138409761333SHillf Danton out: 1385dd4ae78aSYang Shi if (rc == MIGRATEPAGE_SUCCESS) 1386b8ec1ceeSNaoya Horiguchi putback_active_hugepage(hpage); 1387a04840c6SMiaohe Lin else if (rc != -EAGAIN) 1388c33db292SMatthew Wilcox (Oracle) list_move_tail(&src->lru, ret); 138968711a74SDavid Rientjes 139068711a74SDavid Rientjes /* 139168711a74SDavid Rientjes * If migration was not successful and there's a freeing callback, use 139268711a74SDavid Rientjes * it. Otherwise, put_page() will drop the reference grabbed during 139368711a74SDavid Rientjes * isolation. 139468711a74SDavid Rientjes */ 13952def7424SHugh Dickins if (put_new_page) 139668711a74SDavid Rientjes put_new_page(new_hpage, private); 139768711a74SDavid Rientjes else 13983aaa76e1SNaoya Horiguchi putback_active_hugepage(new_hpage); 139968711a74SDavid Rientjes 1400290408d4SNaoya Horiguchi return rc; 1401290408d4SNaoya Horiguchi } 1402290408d4SNaoya Horiguchi 1403eaec4e63SHuang Ying static inline int try_split_folio(struct folio *folio, struct list_head *split_folios) 1404d532e2e5SYang Shi { 14059c62ff00SHuang Ying int rc; 1406d532e2e5SYang Shi 1407eaec4e63SHuang Ying folio_lock(folio); 1408eaec4e63SHuang Ying rc = split_folio_to_list(folio, split_folios); 1409eaec4e63SHuang Ying folio_unlock(folio); 1410e6fa8a79SHuang Ying if (!rc) 1411eaec4e63SHuang Ying list_move_tail(&folio->lru, split_folios); 1412d532e2e5SYang Shi 1413d532e2e5SYang Shi return rc; 1414d532e2e5SYang Shi } 1415d532e2e5SYang Shi 1416290408d4SNaoya Horiguchi /* 1417eaec4e63SHuang Ying * migrate_pages - migrate the folios specified in a list, to the free folios 1418c73e5c9cSSrivatsa S. Bhat * supplied as the target for the page migration 1419e24f0b8fSChristoph Lameter * 1420eaec4e63SHuang Ying * @from: The list of folios to be migrated. 1421eaec4e63SHuang Ying * @get_new_page: The function used to allocate free folios to be used 1422eaec4e63SHuang Ying * as the target of the folio migration. 1423eaec4e63SHuang Ying * @put_new_page: The function used to free target folios if migration 142468711a74SDavid Rientjes * fails, or NULL if no special handling is necessary. 1425c73e5c9cSSrivatsa S. Bhat * @private: Private data to be passed on to get_new_page() 1426c73e5c9cSSrivatsa S. Bhat * @mode: The migration mode that specifies the constraints for 1427eaec4e63SHuang Ying * folio migration, if any. 1428eaec4e63SHuang Ying * @reason: The reason for folio migration. 1429eaec4e63SHuang Ying * @ret_succeeded: Set to the number of folios migrated successfully if 14305ac95884SYang Shi * the caller passes a non-NULL pointer. 1431e24f0b8fSChristoph Lameter * 1432eaec4e63SHuang Ying * The function returns after 10 attempts or if no folios are movable any more 1433eaec4e63SHuang Ying * because the list has become empty or no retryable folios exist any more. 1434eaec4e63SHuang Ying * It is caller's responsibility to call putback_movable_pages() to return folios 1435dd4ae78aSYang Shi * to the LRU or free list only if ret != 0. 1436e24f0b8fSChristoph Lameter * 1437eaec4e63SHuang Ying * Returns the number of {normal folio, large folio, hugetlb} that were not 1438eaec4e63SHuang Ying * migrated, or an error code. The number of large folio splits will be 1439eaec4e63SHuang Ying * considered as the number of non-migrated large folio, no matter how many 1440eaec4e63SHuang Ying * split folios of the large folio are migrated successfully. 1441e24f0b8fSChristoph Lameter */ 14429c620e2bSHugh Dickins int migrate_pages(struct list_head *from, new_page_t get_new_page, 144368711a74SDavid Rientjes free_page_t put_new_page, unsigned long private, 14445ac95884SYang Shi enum migrate_mode mode, int reason, unsigned int *ret_succeeded) 1445e24f0b8fSChristoph Lameter { 1446e24f0b8fSChristoph Lameter int retry = 1; 1447eaec4e63SHuang Ying int large_retry = 1; 14481a5bae25SAnshuman Khandual int thp_retry = 1; 1449e24f0b8fSChristoph Lameter int nr_failed = 0; 1450b5bade97SBaolin Wang int nr_failed_pages = 0; 1451077309bcSHuang Ying int nr_retry_pages = 0; 14525647bc29SMel Gorman int nr_succeeded = 0; 14531a5bae25SAnshuman Khandual int nr_thp_succeeded = 0; 1454eaec4e63SHuang Ying int nr_large_failed = 0; 14551a5bae25SAnshuman Khandual int nr_thp_failed = 0; 14561a5bae25SAnshuman Khandual int nr_thp_split = 0; 1457e24f0b8fSChristoph Lameter int pass = 0; 1458eaec4e63SHuang Ying bool is_large = false; 14591a5bae25SAnshuman Khandual bool is_thp = false; 1460eaec4e63SHuang Ying struct folio *folio, *folio2; 1461eaec4e63SHuang Ying int rc, nr_pages; 1462eaec4e63SHuang Ying LIST_HEAD(ret_folios); 1463eaec4e63SHuang Ying LIST_HEAD(split_folios); 1464b0b515bfSYang Shi bool nosplit = (reason == MR_NUMA_MISPLACED); 1465eaec4e63SHuang Ying bool no_split_folio_counting = false; 14662d1db3b1SChristoph Lameter 14677bc1aec5SLiam Mark trace_mm_migrate_pages_start(mode, reason); 14687bc1aec5SLiam Mark 1469eaec4e63SHuang Ying split_folio_migration: 1470eaec4e63SHuang Ying for (pass = 0; pass < 10 && (retry || large_retry); pass++) { 1471e24f0b8fSChristoph Lameter retry = 0; 1472eaec4e63SHuang Ying large_retry = 0; 14731a5bae25SAnshuman Khandual thp_retry = 0; 1474077309bcSHuang Ying nr_retry_pages = 0; 1475e24f0b8fSChristoph Lameter 1476eaec4e63SHuang Ying list_for_each_entry_safe(folio, folio2, from, lru) { 14771a5bae25SAnshuman Khandual /* 1478eaec4e63SHuang Ying * Large folio statistics is based on the source large 1479eaec4e63SHuang Ying * folio. Capture required information that might get 1480eaec4e63SHuang Ying * lost during migration. 14811a5bae25SAnshuman Khandual */ 1482eaec4e63SHuang Ying is_large = folio_test_large(folio) && !folio_test_hugetlb(folio); 1483eaec4e63SHuang Ying is_thp = is_large && folio_test_pmd_mappable(folio); 1484eaec4e63SHuang Ying nr_pages = folio_nr_pages(folio); 1485e24f0b8fSChristoph Lameter cond_resched(); 1486e24f0b8fSChristoph Lameter 1487eaec4e63SHuang Ying if (folio_test_hugetlb(folio)) 148831caf665SNaoya Horiguchi rc = unmap_and_move_huge_page(get_new_page, 1489eaec4e63SHuang Ying put_new_page, private, 1490eaec4e63SHuang Ying &folio->page, pass > 2, mode, 1491eaec4e63SHuang Ying reason, 1492eaec4e63SHuang Ying &ret_folios); 149331caf665SNaoya Horiguchi else 149468711a74SDavid Rientjes rc = unmap_and_move(get_new_page, put_new_page, 1495eaec4e63SHuang Ying private, folio, pass > 2, mode, 1496eaec4e63SHuang Ying reason, &ret_folios); 1497dd4ae78aSYang Shi /* 1498dd4ae78aSYang Shi * The rules are: 1499eaec4e63SHuang Ying * Success: non hugetlb folio will be freed, hugetlb 1500eaec4e63SHuang Ying * folio will be put back 1501dd4ae78aSYang Shi * -EAGAIN: stay on the from list 1502dd4ae78aSYang Shi * -ENOMEM: stay on the from list 1503577be05cSHuang Ying * -ENOSYS: stay on the from list 1504eaec4e63SHuang Ying * Other errno: put on ret_folios list then splice to 1505dd4ae78aSYang Shi * from list 1506dd4ae78aSYang Shi */ 1507e24f0b8fSChristoph Lameter switch(rc) { 150894723aafSMichal Hocko /* 1509eaec4e63SHuang Ying * Large folio migration might be unsupported or 1510eaec4e63SHuang Ying * the allocation could've failed so we should retry 1511eaec4e63SHuang Ying * on the same folio with the large folio split 1512eaec4e63SHuang Ying * to normal folios. 151394723aafSMichal Hocko * 1514eaec4e63SHuang Ying * Split folios are put in split_folios, and 1515e6fa8a79SHuang Ying * we will migrate them after the rest of the 1516e6fa8a79SHuang Ying * list is processed. 151794723aafSMichal Hocko */ 1518d532e2e5SYang Shi case -ENOSYS: 1519eaec4e63SHuang Ying /* Large folio migration is unsupported */ 1520eaec4e63SHuang Ying if (is_large) { 1521eaec4e63SHuang Ying nr_large_failed++; 1522eaec4e63SHuang Ying nr_thp_failed += is_thp; 1523eaec4e63SHuang Ying if (!try_split_folio(folio, &split_folios)) { 1524eaec4e63SHuang Ying nr_thp_split += is_thp; 1525e6fa8a79SHuang Ying break; 1526d532e2e5SYang Shi } 1527f430893bSMiaohe Lin /* Hugetlb migration is unsupported */ 1528eaec4e63SHuang Ying } else if (!no_split_folio_counting) { 1529f430893bSMiaohe Lin nr_failed++; 1530d532e2e5SYang Shi } 1531d532e2e5SYang Shi 1532eaec4e63SHuang Ying nr_failed_pages += nr_pages; 1533eaec4e63SHuang Ying list_move_tail(&folio->lru, &ret_folios); 1534d532e2e5SYang Shi break; 1535d532e2e5SYang Shi case -ENOMEM: 1536d532e2e5SYang Shi /* 1537d532e2e5SYang Shi * When memory is low, don't bother to try to migrate 1538eaec4e63SHuang Ying * other folios, just exit. 1539d532e2e5SYang Shi */ 1540eaec4e63SHuang Ying if (is_large) { 1541eaec4e63SHuang Ying nr_large_failed++; 1542eaec4e63SHuang Ying nr_thp_failed += is_thp; 1543eaec4e63SHuang Ying /* Large folio NUMA faulting doesn't split to retry. */ 1544fd4a7ac3SBaolin Wang if (!nosplit) { 1545eaec4e63SHuang Ying int ret = try_split_folio(folio, &split_folios); 1546fd4a7ac3SBaolin Wang 1547fd4a7ac3SBaolin Wang if (!ret) { 1548eaec4e63SHuang Ying nr_thp_split += is_thp; 1549e6fa8a79SHuang Ying break; 1550fd4a7ac3SBaolin Wang } else if (reason == MR_LONGTERM_PIN && 1551fd4a7ac3SBaolin Wang ret == -EAGAIN) { 1552fd4a7ac3SBaolin Wang /* 1553eaec4e63SHuang Ying * Try again to split large folio to 1554eaec4e63SHuang Ying * mitigate the failure of longterm pinning. 1555fd4a7ac3SBaolin Wang */ 1556eaec4e63SHuang Ying large_retry++; 1557eaec4e63SHuang Ying thp_retry += is_thp; 1558eaec4e63SHuang Ying nr_retry_pages += nr_pages; 155994723aafSMichal Hocko break; 156094723aafSMichal Hocko } 156194723aafSMichal Hocko } 1562eaec4e63SHuang Ying } else if (!no_split_folio_counting) { 1563f430893bSMiaohe Lin nr_failed++; 15641a5bae25SAnshuman Khandual } 1565b5bade97SBaolin Wang 1566eaec4e63SHuang Ying nr_failed_pages += nr_pages + nr_retry_pages; 156769a041ffSMiaohe Lin /* 1568eaec4e63SHuang Ying * There might be some split folios of fail-to-migrate large 1569eaec4e63SHuang Ying * folios left in split_folios list. Move them back to migration 157069a041ffSMiaohe Lin * list so that they could be put back to the right list by 1571eaec4e63SHuang Ying * the caller otherwise the folio refcnt will be leaked. 157269a041ffSMiaohe Lin */ 1573eaec4e63SHuang Ying list_splice_init(&split_folios, from); 1574fbed53b4SHuang Ying /* nr_failed isn't updated for not used */ 1575eaec4e63SHuang Ying nr_large_failed += large_retry; 157669a041ffSMiaohe Lin nr_thp_failed += thp_retry; 157795a402c3SChristoph Lameter goto out; 1578e24f0b8fSChristoph Lameter case -EAGAIN: 1579eaec4e63SHuang Ying if (is_large) { 1580eaec4e63SHuang Ying large_retry++; 1581eaec4e63SHuang Ying thp_retry += is_thp; 1582eaec4e63SHuang Ying } else if (!no_split_folio_counting) { 1583b20a3503SChristoph Lameter retry++; 1584eaec4e63SHuang Ying } 1585eaec4e63SHuang Ying nr_retry_pages += nr_pages; 1586e24f0b8fSChristoph Lameter break; 158778bd5209SRafael Aquini case MIGRATEPAGE_SUCCESS: 1588eaec4e63SHuang Ying nr_succeeded += nr_pages; 1589eaec4e63SHuang Ying nr_thp_succeeded += is_thp; 15901a5bae25SAnshuman Khandual break; 1591e24f0b8fSChristoph Lameter default: 1592354a3363SNaoya Horiguchi /* 1593d532e2e5SYang Shi * Permanent failure (-EBUSY, etc.): 1594eaec4e63SHuang Ying * unlike -EAGAIN case, the failed folio is 1595eaec4e63SHuang Ying * removed from migration folio list and not 1596354a3363SNaoya Horiguchi * retried in the next outer loop. 1597354a3363SNaoya Horiguchi */ 1598eaec4e63SHuang Ying if (is_large) { 1599eaec4e63SHuang Ying nr_large_failed++; 1600eaec4e63SHuang Ying nr_thp_failed += is_thp; 1601eaec4e63SHuang Ying } else if (!no_split_folio_counting) { 1602b20a3503SChristoph Lameter nr_failed++; 1603eaec4e63SHuang Ying } 1604f430893bSMiaohe Lin 1605eaec4e63SHuang Ying nr_failed_pages += nr_pages; 1606e24f0b8fSChristoph Lameter break; 1607b20a3503SChristoph Lameter } 1608b20a3503SChristoph Lameter } 1609e24f0b8fSChristoph Lameter } 1610b5bade97SBaolin Wang nr_failed += retry; 1611eaec4e63SHuang Ying nr_large_failed += large_retry; 16121a5bae25SAnshuman Khandual nr_thp_failed += thp_retry; 1613077309bcSHuang Ying nr_failed_pages += nr_retry_pages; 1614b5bade97SBaolin Wang /* 1615eaec4e63SHuang Ying * Try to migrate split folios of fail-to-migrate large folios, no 1616eaec4e63SHuang Ying * nr_failed counting in this round, since all split folios of a 1617eaec4e63SHuang Ying * large folio is counted as 1 failure in the first round. 1618b5bade97SBaolin Wang */ 1619eaec4e63SHuang Ying if (!list_empty(&split_folios)) { 1620b5bade97SBaolin Wang /* 1621eaec4e63SHuang Ying * Move non-migrated folios (after 10 retries) to ret_folios 1622b5bade97SBaolin Wang * to avoid migrating them again. 1623b5bade97SBaolin Wang */ 1624eaec4e63SHuang Ying list_splice_init(from, &ret_folios); 1625eaec4e63SHuang Ying list_splice_init(&split_folios, from); 1626eaec4e63SHuang Ying no_split_folio_counting = true; 1627b5bade97SBaolin Wang retry = 1; 1628eaec4e63SHuang Ying goto split_folio_migration; 1629b5bade97SBaolin Wang } 1630b5bade97SBaolin Wang 1631eaec4e63SHuang Ying rc = nr_failed + nr_large_failed; 163295a402c3SChristoph Lameter out: 1633dd4ae78aSYang Shi /* 1634eaec4e63SHuang Ying * Put the permanent failure folio back to migration list, they 1635dd4ae78aSYang Shi * will be put back to the right list by the caller. 1636dd4ae78aSYang Shi */ 1637eaec4e63SHuang Ying list_splice(&ret_folios, from); 1638dd4ae78aSYang Shi 163903e5f82eSBaolin Wang /* 1640eaec4e63SHuang Ying * Return 0 in case all split folios of fail-to-migrate large folios 1641eaec4e63SHuang Ying * are migrated successfully. 164203e5f82eSBaolin Wang */ 164303e5f82eSBaolin Wang if (list_empty(from)) 164403e5f82eSBaolin Wang rc = 0; 164503e5f82eSBaolin Wang 16465647bc29SMel Gorman count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); 1647b5bade97SBaolin Wang count_vm_events(PGMIGRATE_FAIL, nr_failed_pages); 16481a5bae25SAnshuman Khandual count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded); 16491a5bae25SAnshuman Khandual count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed); 16501a5bae25SAnshuman Khandual count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split); 1651b5bade97SBaolin Wang trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded, 16521a5bae25SAnshuman Khandual nr_thp_failed, nr_thp_split, mode, reason); 16537b2a2d4aSMel Gorman 16545ac95884SYang Shi if (ret_succeeded) 16555ac95884SYang Shi *ret_succeeded = nr_succeeded; 16565ac95884SYang Shi 165795a402c3SChristoph Lameter return rc; 1658b20a3503SChristoph Lameter } 1659b20a3503SChristoph Lameter 166019fc7bedSJoonsoo Kim struct page *alloc_migration_target(struct page *page, unsigned long private) 1661b4b38223SJoonsoo Kim { 1662ffe06786SMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 166319fc7bedSJoonsoo Kim struct migration_target_control *mtc; 166419fc7bedSJoonsoo Kim gfp_t gfp_mask; 1665b4b38223SJoonsoo Kim unsigned int order = 0; 1666ffe06786SMatthew Wilcox (Oracle) struct folio *new_folio = NULL; 166719fc7bedSJoonsoo Kim int nid; 166819fc7bedSJoonsoo Kim int zidx; 166919fc7bedSJoonsoo Kim 167019fc7bedSJoonsoo Kim mtc = (struct migration_target_control *)private; 167119fc7bedSJoonsoo Kim gfp_mask = mtc->gfp_mask; 167219fc7bedSJoonsoo Kim nid = mtc->nid; 167319fc7bedSJoonsoo Kim if (nid == NUMA_NO_NODE) 1674ffe06786SMatthew Wilcox (Oracle) nid = folio_nid(folio); 1675b4b38223SJoonsoo Kim 1676ffe06786SMatthew Wilcox (Oracle) if (folio_test_hugetlb(folio)) { 1677e51da3a9SSidhartha Kumar struct hstate *h = folio_hstate(folio); 1678d92bbc27SJoonsoo Kim 167919fc7bedSJoonsoo Kim gfp_mask = htlb_modify_alloc_mask(h, gfp_mask); 168019fc7bedSJoonsoo Kim return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask); 1681d92bbc27SJoonsoo Kim } 1682b4b38223SJoonsoo Kim 1683ffe06786SMatthew Wilcox (Oracle) if (folio_test_large(folio)) { 16849933a0c8SJoonsoo Kim /* 16859933a0c8SJoonsoo Kim * clear __GFP_RECLAIM to make the migration callback 16869933a0c8SJoonsoo Kim * consistent with regular THP allocations. 16879933a0c8SJoonsoo Kim */ 16889933a0c8SJoonsoo Kim gfp_mask &= ~__GFP_RECLAIM; 1689b4b38223SJoonsoo Kim gfp_mask |= GFP_TRANSHUGE; 1690ffe06786SMatthew Wilcox (Oracle) order = folio_order(folio); 1691b4b38223SJoonsoo Kim } 1692ffe06786SMatthew Wilcox (Oracle) zidx = zone_idx(folio_zone(folio)); 169319fc7bedSJoonsoo Kim if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE) 1694b4b38223SJoonsoo Kim gfp_mask |= __GFP_HIGHMEM; 1695b4b38223SJoonsoo Kim 1696ffe06786SMatthew Wilcox (Oracle) new_folio = __folio_alloc(gfp_mask, order, nid, mtc->nmask); 1697b4b38223SJoonsoo Kim 1698ffe06786SMatthew Wilcox (Oracle) return &new_folio->page; 1699b4b38223SJoonsoo Kim } 1700b4b38223SJoonsoo Kim 1701742755a1SChristoph Lameter #ifdef CONFIG_NUMA 1702742755a1SChristoph Lameter 1703a49bd4d7SMichal Hocko static int store_status(int __user *status, int start, int value, int nr) 1704742755a1SChristoph Lameter { 1705a49bd4d7SMichal Hocko while (nr-- > 0) { 1706a49bd4d7SMichal Hocko if (put_user(value, status + start)) 1707a49bd4d7SMichal Hocko return -EFAULT; 1708a49bd4d7SMichal Hocko start++; 1709a49bd4d7SMichal Hocko } 1710742755a1SChristoph Lameter 1711a49bd4d7SMichal Hocko return 0; 1712a49bd4d7SMichal Hocko } 1713742755a1SChristoph Lameter 1714a49bd4d7SMichal Hocko static int do_move_pages_to_node(struct mm_struct *mm, 1715a49bd4d7SMichal Hocko struct list_head *pagelist, int node) 1716a49bd4d7SMichal Hocko { 1717a49bd4d7SMichal Hocko int err; 1718a0976311SJoonsoo Kim struct migration_target_control mtc = { 1719a0976311SJoonsoo Kim .nid = node, 1720a0976311SJoonsoo Kim .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 1721a0976311SJoonsoo Kim }; 1722742755a1SChristoph Lameter 1723a0976311SJoonsoo Kim err = migrate_pages(pagelist, alloc_migration_target, NULL, 17245ac95884SYang Shi (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL); 1725a49bd4d7SMichal Hocko if (err) 1726a49bd4d7SMichal Hocko putback_movable_pages(pagelist); 1727a49bd4d7SMichal Hocko return err; 1728742755a1SChristoph Lameter } 1729742755a1SChristoph Lameter 1730742755a1SChristoph Lameter /* 1731a49bd4d7SMichal Hocko * Resolves the given address to a struct page, isolates it from the LRU and 1732a49bd4d7SMichal Hocko * puts it to the given pagelist. 1733e0153fc2SYang Shi * Returns: 1734e0153fc2SYang Shi * errno - if the page cannot be found/isolated 1735e0153fc2SYang Shi * 0 - when it doesn't have to be migrated because it is already on the 1736e0153fc2SYang Shi * target node 1737e0153fc2SYang Shi * 1 - when it has been queued 1738742755a1SChristoph Lameter */ 1739a49bd4d7SMichal Hocko static int add_page_for_migration(struct mm_struct *mm, unsigned long addr, 1740a49bd4d7SMichal Hocko int node, struct list_head *pagelist, bool migrate_all) 1741742755a1SChristoph Lameter { 1742742755a1SChristoph Lameter struct vm_area_struct *vma; 1743742755a1SChristoph Lameter struct page *page; 1744a49bd4d7SMichal Hocko int err; 1745742755a1SChristoph Lameter 1746d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 1747742755a1SChristoph Lameter err = -EFAULT; 1748cb1c37b1SMiaohe Lin vma = vma_lookup(mm, addr); 1749cb1c37b1SMiaohe Lin if (!vma || !vma_migratable(vma)) 1750a49bd4d7SMichal Hocko goto out; 1751742755a1SChristoph Lameter 1752d899844eSKirill A. Shutemov /* FOLL_DUMP to ignore special (like zero) pages */ 175387d2762eSMiaohe Lin page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 175489f5b7daSLinus Torvalds 175589f5b7daSLinus Torvalds err = PTR_ERR(page); 175689f5b7daSLinus Torvalds if (IS_ERR(page)) 1757a49bd4d7SMichal Hocko goto out; 175889f5b7daSLinus Torvalds 1759742755a1SChristoph Lameter err = -ENOENT; 1760f7091ed6SHaiyue Wang if (!page) 1761a49bd4d7SMichal Hocko goto out; 1762742755a1SChristoph Lameter 1763f7091ed6SHaiyue Wang if (is_zone_device_page(page)) 1764f7091ed6SHaiyue Wang goto out_putpage; 1765f7091ed6SHaiyue Wang 1766a49bd4d7SMichal Hocko err = 0; 1767a49bd4d7SMichal Hocko if (page_to_nid(page) == node) 1768a49bd4d7SMichal Hocko goto out_putpage; 1769742755a1SChristoph Lameter 1770742755a1SChristoph Lameter err = -EACCES; 1771a49bd4d7SMichal Hocko if (page_mapcount(page) > 1 && !migrate_all) 1772a49bd4d7SMichal Hocko goto out_putpage; 1773742755a1SChristoph Lameter 1774e632a938SNaoya Horiguchi if (PageHuge(page)) { 1775e8db67ebSNaoya Horiguchi if (PageHead(page)) { 17767ce82f4cSMiaohe Lin err = isolate_hugetlb(page, pagelist); 17777ce82f4cSMiaohe Lin if (!err) 1778e0153fc2SYang Shi err = 1; 1779e8db67ebSNaoya Horiguchi } 1780a49bd4d7SMichal Hocko } else { 1781a49bd4d7SMichal Hocko struct page *head; 1782e632a938SNaoya Horiguchi 1783e8db67ebSNaoya Horiguchi head = compound_head(page); 1784e8db67ebSNaoya Horiguchi err = isolate_lru_page(head); 1785a49bd4d7SMichal Hocko if (err) 1786a49bd4d7SMichal Hocko goto out_putpage; 1787a49bd4d7SMichal Hocko 1788e0153fc2SYang Shi err = 1; 1789a49bd4d7SMichal Hocko list_add_tail(&head->lru, pagelist); 1790e8db67ebSNaoya Horiguchi mod_node_page_state(page_pgdat(head), 17919de4f22aSHuang Ying NR_ISOLATED_ANON + page_is_file_lru(head), 17926c357848SMatthew Wilcox (Oracle) thp_nr_pages(head)); 17936d9c285aSKOSAKI Motohiro } 1794a49bd4d7SMichal Hocko out_putpage: 1795742755a1SChristoph Lameter /* 1796742755a1SChristoph Lameter * Either remove the duplicate refcount from 1797742755a1SChristoph Lameter * isolate_lru_page() or drop the page ref if it was 1798742755a1SChristoph Lameter * not isolated. 1799742755a1SChristoph Lameter */ 1800742755a1SChristoph Lameter put_page(page); 1801a49bd4d7SMichal Hocko out: 1802d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 1803742755a1SChristoph Lameter return err; 1804742755a1SChristoph Lameter } 1805742755a1SChristoph Lameter 18067ca8783aSWei Yang static int move_pages_and_store_status(struct mm_struct *mm, int node, 18077ca8783aSWei Yang struct list_head *pagelist, int __user *status, 18087ca8783aSWei Yang int start, int i, unsigned long nr_pages) 18097ca8783aSWei Yang { 18107ca8783aSWei Yang int err; 18117ca8783aSWei Yang 18125d7ae891SWei Yang if (list_empty(pagelist)) 18135d7ae891SWei Yang return 0; 18145d7ae891SWei Yang 18157ca8783aSWei Yang err = do_move_pages_to_node(mm, pagelist, node); 18167ca8783aSWei Yang if (err) { 18177ca8783aSWei Yang /* 18187ca8783aSWei Yang * Positive err means the number of failed 18197ca8783aSWei Yang * pages to migrate. Since we are going to 18207ca8783aSWei Yang * abort and return the number of non-migrated 1821ab9dd4f8SLong Li * pages, so need to include the rest of the 18227ca8783aSWei Yang * nr_pages that have not been attempted as 18237ca8783aSWei Yang * well. 18247ca8783aSWei Yang */ 18257ca8783aSWei Yang if (err > 0) 1826a7504ed1SHuang Ying err += nr_pages - i; 18277ca8783aSWei Yang return err; 18287ca8783aSWei Yang } 18297ca8783aSWei Yang return store_status(status, start, node, i - start); 18307ca8783aSWei Yang } 18317ca8783aSWei Yang 1832742755a1SChristoph Lameter /* 18335e9a0f02SBrice Goglin * Migrate an array of page address onto an array of nodes and fill 18345e9a0f02SBrice Goglin * the corresponding array of status. 18355e9a0f02SBrice Goglin */ 18363268c63eSChristoph Lameter static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, 18375e9a0f02SBrice Goglin unsigned long nr_pages, 18385e9a0f02SBrice Goglin const void __user * __user *pages, 18395e9a0f02SBrice Goglin const int __user *nodes, 18405e9a0f02SBrice Goglin int __user *status, int flags) 18415e9a0f02SBrice Goglin { 1842a49bd4d7SMichal Hocko int current_node = NUMA_NO_NODE; 1843a49bd4d7SMichal Hocko LIST_HEAD(pagelist); 1844a49bd4d7SMichal Hocko int start, i; 1845a49bd4d7SMichal Hocko int err = 0, err1; 184635282a2dSBrice Goglin 1847361a2a22SMinchan Kim lru_cache_disable(); 184835282a2dSBrice Goglin 1849a49bd4d7SMichal Hocko for (i = start = 0; i < nr_pages; i++) { 18505e9a0f02SBrice Goglin const void __user *p; 1851a49bd4d7SMichal Hocko unsigned long addr; 18525e9a0f02SBrice Goglin int node; 18535e9a0f02SBrice Goglin 18543140a227SBrice Goglin err = -EFAULT; 1855a49bd4d7SMichal Hocko if (get_user(p, pages + i)) 1856a49bd4d7SMichal Hocko goto out_flush; 1857a49bd4d7SMichal Hocko if (get_user(node, nodes + i)) 1858a49bd4d7SMichal Hocko goto out_flush; 1859057d3389SAndrey Konovalov addr = (unsigned long)untagged_addr(p); 18605e9a0f02SBrice Goglin 18615e9a0f02SBrice Goglin err = -ENODEV; 18626f5a55f1SLinus Torvalds if (node < 0 || node >= MAX_NUMNODES) 1863a49bd4d7SMichal Hocko goto out_flush; 1864389162c2SLai Jiangshan if (!node_state(node, N_MEMORY)) 1865a49bd4d7SMichal Hocko goto out_flush; 18665e9a0f02SBrice Goglin 18675e9a0f02SBrice Goglin err = -EACCES; 18685e9a0f02SBrice Goglin if (!node_isset(node, task_nodes)) 1869a49bd4d7SMichal Hocko goto out_flush; 18705e9a0f02SBrice Goglin 1871a49bd4d7SMichal Hocko if (current_node == NUMA_NO_NODE) { 1872a49bd4d7SMichal Hocko current_node = node; 1873a49bd4d7SMichal Hocko start = i; 1874a49bd4d7SMichal Hocko } else if (node != current_node) { 18757ca8783aSWei Yang err = move_pages_and_store_status(mm, current_node, 18767ca8783aSWei Yang &pagelist, status, start, i, nr_pages); 1877a49bd4d7SMichal Hocko if (err) 1878a49bd4d7SMichal Hocko goto out; 1879a49bd4d7SMichal Hocko start = i; 1880a49bd4d7SMichal Hocko current_node = node; 18815e9a0f02SBrice Goglin } 18825e9a0f02SBrice Goglin 1883a49bd4d7SMichal Hocko /* 1884a49bd4d7SMichal Hocko * Errors in the page lookup or isolation are not fatal and we simply 1885a49bd4d7SMichal Hocko * report them via status 1886a49bd4d7SMichal Hocko */ 1887a49bd4d7SMichal Hocko err = add_page_for_migration(mm, addr, current_node, 1888a49bd4d7SMichal Hocko &pagelist, flags & MPOL_MF_MOVE_ALL); 1889e0153fc2SYang Shi 1890d08221a0SWei Yang if (err > 0) { 1891e0153fc2SYang Shi /* The page is successfully queued for migration */ 1892e0153fc2SYang Shi continue; 1893e0153fc2SYang Shi } 18943140a227SBrice Goglin 1895d08221a0SWei Yang /* 189665462462SJohn Hubbard * The move_pages() man page does not have an -EEXIST choice, so 189765462462SJohn Hubbard * use -EFAULT instead. 189865462462SJohn Hubbard */ 189965462462SJohn Hubbard if (err == -EEXIST) 190065462462SJohn Hubbard err = -EFAULT; 190165462462SJohn Hubbard 190265462462SJohn Hubbard /* 1903d08221a0SWei Yang * If the page is already on the target node (!err), store the 1904d08221a0SWei Yang * node, otherwise, store the err. 1905d08221a0SWei Yang */ 1906d08221a0SWei Yang err = store_status(status, i, err ? : current_node, 1); 1907a49bd4d7SMichal Hocko if (err) 1908a49bd4d7SMichal Hocko goto out_flush; 19093140a227SBrice Goglin 19107ca8783aSWei Yang err = move_pages_and_store_status(mm, current_node, &pagelist, 19117ca8783aSWei Yang status, start, i, nr_pages); 1912a7504ed1SHuang Ying if (err) { 1913a7504ed1SHuang Ying /* We have accounted for page i */ 1914a7504ed1SHuang Ying if (err > 0) 1915a7504ed1SHuang Ying err--; 1916a49bd4d7SMichal Hocko goto out; 1917a7504ed1SHuang Ying } 1918a49bd4d7SMichal Hocko current_node = NUMA_NO_NODE; 19193140a227SBrice Goglin } 1920a49bd4d7SMichal Hocko out_flush: 1921a49bd4d7SMichal Hocko /* Make sure we do not overwrite the existing error */ 19227ca8783aSWei Yang err1 = move_pages_and_store_status(mm, current_node, &pagelist, 19237ca8783aSWei Yang status, start, i, nr_pages); 1924dfe9aa23SWei Yang if (err >= 0) 1925a49bd4d7SMichal Hocko err = err1; 19265e9a0f02SBrice Goglin out: 1927361a2a22SMinchan Kim lru_cache_enable(); 19285e9a0f02SBrice Goglin return err; 19295e9a0f02SBrice Goglin } 19305e9a0f02SBrice Goglin 19315e9a0f02SBrice Goglin /* 19322f007e74SBrice Goglin * Determine the nodes of an array of pages and store it in an array of status. 1933742755a1SChristoph Lameter */ 193480bba129SBrice Goglin static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, 193580bba129SBrice Goglin const void __user **pages, int *status) 1936742755a1SChristoph Lameter { 19372f007e74SBrice Goglin unsigned long i; 1938742755a1SChristoph Lameter 1939d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 19402f007e74SBrice Goglin 19412f007e74SBrice Goglin for (i = 0; i < nr_pages; i++) { 194280bba129SBrice Goglin unsigned long addr = (unsigned long)(*pages); 19432f007e74SBrice Goglin struct vm_area_struct *vma; 19442f007e74SBrice Goglin struct page *page; 1945c095adbcSKOSAKI Motohiro int err = -EFAULT; 19462f007e74SBrice Goglin 1947059b8b48SLiam Howlett vma = vma_lookup(mm, addr); 1948059b8b48SLiam Howlett if (!vma) 1949742755a1SChristoph Lameter goto set_status; 1950742755a1SChristoph Lameter 1951d899844eSKirill A. Shutemov /* FOLL_DUMP to ignore special (like zero) pages */ 195216fd6b31SBaolin Wang page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 195389f5b7daSLinus Torvalds 195489f5b7daSLinus Torvalds err = PTR_ERR(page); 195589f5b7daSLinus Torvalds if (IS_ERR(page)) 195689f5b7daSLinus Torvalds goto set_status; 195789f5b7daSLinus Torvalds 1958f7091ed6SHaiyue Wang err = -ENOENT; 1959f7091ed6SHaiyue Wang if (!page) 1960f7091ed6SHaiyue Wang goto set_status; 1961f7091ed6SHaiyue Wang 1962f7091ed6SHaiyue Wang if (!is_zone_device_page(page)) 19634cd61484SMiaohe Lin err = page_to_nid(page); 1964f7091ed6SHaiyue Wang 19654cd61484SMiaohe Lin put_page(page); 1966742755a1SChristoph Lameter set_status: 196780bba129SBrice Goglin *status = err; 196880bba129SBrice Goglin 196980bba129SBrice Goglin pages++; 197080bba129SBrice Goglin status++; 197180bba129SBrice Goglin } 197280bba129SBrice Goglin 1973d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 197480bba129SBrice Goglin } 197580bba129SBrice Goglin 19765b1b561bSArnd Bergmann static int get_compat_pages_array(const void __user *chunk_pages[], 19775b1b561bSArnd Bergmann const void __user * __user *pages, 19785b1b561bSArnd Bergmann unsigned long chunk_nr) 19795b1b561bSArnd Bergmann { 19805b1b561bSArnd Bergmann compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages; 19815b1b561bSArnd Bergmann compat_uptr_t p; 19825b1b561bSArnd Bergmann int i; 19835b1b561bSArnd Bergmann 19845b1b561bSArnd Bergmann for (i = 0; i < chunk_nr; i++) { 19855b1b561bSArnd Bergmann if (get_user(p, pages32 + i)) 19865b1b561bSArnd Bergmann return -EFAULT; 19875b1b561bSArnd Bergmann chunk_pages[i] = compat_ptr(p); 19885b1b561bSArnd Bergmann } 19895b1b561bSArnd Bergmann 19905b1b561bSArnd Bergmann return 0; 19915b1b561bSArnd Bergmann } 19925b1b561bSArnd Bergmann 199380bba129SBrice Goglin /* 199480bba129SBrice Goglin * Determine the nodes of a user array of pages and store it in 199580bba129SBrice Goglin * a user array of status. 199680bba129SBrice Goglin */ 199780bba129SBrice Goglin static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, 199880bba129SBrice Goglin const void __user * __user *pages, 199980bba129SBrice Goglin int __user *status) 200080bba129SBrice Goglin { 20013eefb826SMiaohe Lin #define DO_PAGES_STAT_CHUNK_NR 16UL 200280bba129SBrice Goglin const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; 200380bba129SBrice Goglin int chunk_status[DO_PAGES_STAT_CHUNK_NR]; 200480bba129SBrice Goglin 200587b8d1adSH. Peter Anvin while (nr_pages) { 20063eefb826SMiaohe Lin unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR); 200787b8d1adSH. Peter Anvin 20085b1b561bSArnd Bergmann if (in_compat_syscall()) { 20095b1b561bSArnd Bergmann if (get_compat_pages_array(chunk_pages, pages, 20105b1b561bSArnd Bergmann chunk_nr)) 201187b8d1adSH. Peter Anvin break; 20125b1b561bSArnd Bergmann } else { 20135b1b561bSArnd Bergmann if (copy_from_user(chunk_pages, pages, 20145b1b561bSArnd Bergmann chunk_nr * sizeof(*chunk_pages))) 20155b1b561bSArnd Bergmann break; 20165b1b561bSArnd Bergmann } 201780bba129SBrice Goglin 201880bba129SBrice Goglin do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); 201980bba129SBrice Goglin 202087b8d1adSH. Peter Anvin if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) 202187b8d1adSH. Peter Anvin break; 2022742755a1SChristoph Lameter 202387b8d1adSH. Peter Anvin pages += chunk_nr; 202487b8d1adSH. Peter Anvin status += chunk_nr; 202587b8d1adSH. Peter Anvin nr_pages -= chunk_nr; 202687b8d1adSH. Peter Anvin } 202787b8d1adSH. Peter Anvin return nr_pages ? -EFAULT : 0; 2028742755a1SChristoph Lameter } 2029742755a1SChristoph Lameter 20304dc200ceSMiaohe Lin static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes) 20314dc200ceSMiaohe Lin { 20324dc200ceSMiaohe Lin struct task_struct *task; 20334dc200ceSMiaohe Lin struct mm_struct *mm; 20344dc200ceSMiaohe Lin 20354dc200ceSMiaohe Lin /* 20364dc200ceSMiaohe Lin * There is no need to check if current process has the right to modify 20374dc200ceSMiaohe Lin * the specified process when they are same. 20384dc200ceSMiaohe Lin */ 20394dc200ceSMiaohe Lin if (!pid) { 20404dc200ceSMiaohe Lin mmget(current->mm); 20414dc200ceSMiaohe Lin *mem_nodes = cpuset_mems_allowed(current); 20424dc200ceSMiaohe Lin return current->mm; 20434dc200ceSMiaohe Lin } 20444dc200ceSMiaohe Lin 20454dc200ceSMiaohe Lin /* Find the mm_struct */ 20464dc200ceSMiaohe Lin rcu_read_lock(); 20474dc200ceSMiaohe Lin task = find_task_by_vpid(pid); 20484dc200ceSMiaohe Lin if (!task) { 20494dc200ceSMiaohe Lin rcu_read_unlock(); 20504dc200ceSMiaohe Lin return ERR_PTR(-ESRCH); 20514dc200ceSMiaohe Lin } 20524dc200ceSMiaohe Lin get_task_struct(task); 20534dc200ceSMiaohe Lin 20544dc200ceSMiaohe Lin /* 20554dc200ceSMiaohe Lin * Check if this process has the right to modify the specified 20564dc200ceSMiaohe Lin * process. Use the regular "ptrace_may_access()" checks. 20574dc200ceSMiaohe Lin */ 20584dc200ceSMiaohe Lin if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { 20594dc200ceSMiaohe Lin rcu_read_unlock(); 20604dc200ceSMiaohe Lin mm = ERR_PTR(-EPERM); 20614dc200ceSMiaohe Lin goto out; 20624dc200ceSMiaohe Lin } 20634dc200ceSMiaohe Lin rcu_read_unlock(); 20644dc200ceSMiaohe Lin 20654dc200ceSMiaohe Lin mm = ERR_PTR(security_task_movememory(task)); 20664dc200ceSMiaohe Lin if (IS_ERR(mm)) 20674dc200ceSMiaohe Lin goto out; 20684dc200ceSMiaohe Lin *mem_nodes = cpuset_mems_allowed(task); 20694dc200ceSMiaohe Lin mm = get_task_mm(task); 20704dc200ceSMiaohe Lin out: 20714dc200ceSMiaohe Lin put_task_struct(task); 20724dc200ceSMiaohe Lin if (!mm) 20734dc200ceSMiaohe Lin mm = ERR_PTR(-EINVAL); 20744dc200ceSMiaohe Lin return mm; 20754dc200ceSMiaohe Lin } 20764dc200ceSMiaohe Lin 2077742755a1SChristoph Lameter /* 2078742755a1SChristoph Lameter * Move a list of pages in the address space of the currently executing 2079742755a1SChristoph Lameter * process. 2080742755a1SChristoph Lameter */ 20817addf443SDominik Brodowski static int kernel_move_pages(pid_t pid, unsigned long nr_pages, 20827addf443SDominik Brodowski const void __user * __user *pages, 20837addf443SDominik Brodowski const int __user *nodes, 20847addf443SDominik Brodowski int __user *status, int flags) 2085742755a1SChristoph Lameter { 2086742755a1SChristoph Lameter struct mm_struct *mm; 20875e9a0f02SBrice Goglin int err; 20883268c63eSChristoph Lameter nodemask_t task_nodes; 2089742755a1SChristoph Lameter 2090742755a1SChristoph Lameter /* Check flags */ 2091742755a1SChristoph Lameter if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) 2092742755a1SChristoph Lameter return -EINVAL; 2093742755a1SChristoph Lameter 2094742755a1SChristoph Lameter if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) 2095742755a1SChristoph Lameter return -EPERM; 2096742755a1SChristoph Lameter 20974dc200ceSMiaohe Lin mm = find_mm_struct(pid, &task_nodes); 20984dc200ceSMiaohe Lin if (IS_ERR(mm)) 20994dc200ceSMiaohe Lin return PTR_ERR(mm); 21006e8b09eaSSasha Levin 21013268c63eSChristoph Lameter if (nodes) 21023268c63eSChristoph Lameter err = do_pages_move(mm, task_nodes, nr_pages, pages, 21033268c63eSChristoph Lameter nodes, status, flags); 21043268c63eSChristoph Lameter else 21055e9a0f02SBrice Goglin err = do_pages_stat(mm, nr_pages, pages, status); 21063268c63eSChristoph Lameter 21073268c63eSChristoph Lameter mmput(mm); 21083268c63eSChristoph Lameter return err; 2109742755a1SChristoph Lameter } 2110742755a1SChristoph Lameter 21117addf443SDominik Brodowski SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, 21127addf443SDominik Brodowski const void __user * __user *, pages, 21137addf443SDominik Brodowski const int __user *, nodes, 21147addf443SDominik Brodowski int __user *, status, int, flags) 21157addf443SDominik Brodowski { 21167addf443SDominik Brodowski return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); 21177addf443SDominik Brodowski } 21187addf443SDominik Brodowski 21197039e1dbSPeter Zijlstra #ifdef CONFIG_NUMA_BALANCING 21207039e1dbSPeter Zijlstra /* 21217039e1dbSPeter Zijlstra * Returns true if this is a safe migration target node for misplaced NUMA 2122bc53008eSWei Yang * pages. Currently it only checks the watermarks which is crude. 21237039e1dbSPeter Zijlstra */ 21247039e1dbSPeter Zijlstra static bool migrate_balanced_pgdat(struct pglist_data *pgdat, 21253abef4e6SMel Gorman unsigned long nr_migrate_pages) 21267039e1dbSPeter Zijlstra { 21277039e1dbSPeter Zijlstra int z; 2128599d0c95SMel Gorman 21297039e1dbSPeter Zijlstra for (z = pgdat->nr_zones - 1; z >= 0; z--) { 21307039e1dbSPeter Zijlstra struct zone *zone = pgdat->node_zones + z; 21317039e1dbSPeter Zijlstra 2132bc53008eSWei Yang if (!managed_zone(zone)) 21337039e1dbSPeter Zijlstra continue; 21347039e1dbSPeter Zijlstra 21357039e1dbSPeter Zijlstra /* Avoid waking kswapd by allocating pages_to_migrate pages. */ 21367039e1dbSPeter Zijlstra if (!zone_watermark_ok(zone, 0, 21377039e1dbSPeter Zijlstra high_wmark_pages(zone) + 21387039e1dbSPeter Zijlstra nr_migrate_pages, 2139bfe9d006SHuang Ying ZONE_MOVABLE, 0)) 21407039e1dbSPeter Zijlstra continue; 21417039e1dbSPeter Zijlstra return true; 21427039e1dbSPeter Zijlstra } 21437039e1dbSPeter Zijlstra return false; 21447039e1dbSPeter Zijlstra } 21457039e1dbSPeter Zijlstra 21467039e1dbSPeter Zijlstra static struct page *alloc_misplaced_dst_page(struct page *page, 2147666feb21SMichal Hocko unsigned long data) 21487039e1dbSPeter Zijlstra { 21497039e1dbSPeter Zijlstra int nid = (int) data; 2150c185e494SMatthew Wilcox (Oracle) int order = compound_order(page); 2151c185e494SMatthew Wilcox (Oracle) gfp_t gfp = __GFP_THISNODE; 2152c185e494SMatthew Wilcox (Oracle) struct folio *new; 21537039e1dbSPeter Zijlstra 2154c185e494SMatthew Wilcox (Oracle) if (order > 0) 2155c185e494SMatthew Wilcox (Oracle) gfp |= GFP_TRANSHUGE_LIGHT; 2156c185e494SMatthew Wilcox (Oracle) else { 2157c185e494SMatthew Wilcox (Oracle) gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY | 2158c185e494SMatthew Wilcox (Oracle) __GFP_NOWARN; 2159c185e494SMatthew Wilcox (Oracle) gfp &= ~__GFP_RECLAIM; 21607039e1dbSPeter Zijlstra } 2161c185e494SMatthew Wilcox (Oracle) new = __folio_alloc_node(gfp, order, nid); 21627039e1dbSPeter Zijlstra 2163c185e494SMatthew Wilcox (Oracle) return &new->page; 2164c5b5a3ddSYang Shi } 2165c5b5a3ddSYang Shi 21661c30e017SMel Gorman static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) 2167b32967ffSMel Gorman { 21682b9b624fSBaolin Wang int nr_pages = thp_nr_pages(page); 2169c574bbe9SHuang Ying int order = compound_order(page); 2170b32967ffSMel Gorman 2171c574bbe9SHuang Ying VM_BUG_ON_PAGE(order && !PageTransHuge(page), page); 21723abef4e6SMel Gorman 2173662aeea7SYang Shi /* Do not migrate THP mapped by multiple processes */ 2174662aeea7SYang Shi if (PageTransHuge(page) && total_mapcount(page) > 1) 2175662aeea7SYang Shi return 0; 2176662aeea7SYang Shi 2177b32967ffSMel Gorman /* Avoid migrating to a node that is nearly full */ 2178c574bbe9SHuang Ying if (!migrate_balanced_pgdat(pgdat, nr_pages)) { 2179c574bbe9SHuang Ying int z; 2180c574bbe9SHuang Ying 2181c574bbe9SHuang Ying if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)) 2182340ef390SHugh Dickins return 0; 2183c574bbe9SHuang Ying for (z = pgdat->nr_zones - 1; z >= 0; z--) { 2184bc53008eSWei Yang if (managed_zone(pgdat->node_zones + z)) 2185c574bbe9SHuang Ying break; 2186c574bbe9SHuang Ying } 2187c574bbe9SHuang Ying wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE); 2188c574bbe9SHuang Ying return 0; 2189c574bbe9SHuang Ying } 2190b32967ffSMel Gorman 2191340ef390SHugh Dickins if (isolate_lru_page(page)) 2192340ef390SHugh Dickins return 0; 2193340ef390SHugh Dickins 2194b75454e1SMiaohe Lin mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_is_file_lru(page), 21952b9b624fSBaolin Wang nr_pages); 2196b32967ffSMel Gorman 2197b32967ffSMel Gorman /* 2198340ef390SHugh Dickins * Isolating the page has taken another reference, so the 2199340ef390SHugh Dickins * caller's reference can be safely dropped without the page 2200340ef390SHugh Dickins * disappearing underneath us during migration. 2201b32967ffSMel Gorman */ 2202b32967ffSMel Gorman put_page(page); 2203340ef390SHugh Dickins return 1; 2204b32967ffSMel Gorman } 2205b32967ffSMel Gorman 2206a8f60772SMel Gorman /* 22077039e1dbSPeter Zijlstra * Attempt to migrate a misplaced page to the specified destination 22087039e1dbSPeter Zijlstra * node. Caller is expected to have an elevated reference count on 22097039e1dbSPeter Zijlstra * the page that will be dropped by this function before returning. 22107039e1dbSPeter Zijlstra */ 22111bc115d8SMel Gorman int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, 22121bc115d8SMel Gorman int node) 22137039e1dbSPeter Zijlstra { 2214a8f60772SMel Gorman pg_data_t *pgdat = NODE_DATA(node); 2215340ef390SHugh Dickins int isolated; 2216b32967ffSMel Gorman int nr_remaining; 2217e39bb6beSHuang Ying unsigned int nr_succeeded; 22187039e1dbSPeter Zijlstra LIST_HEAD(migratepages); 2219b5916c02SAneesh Kumar K.V int nr_pages = thp_nr_pages(page); 2220c5b5a3ddSYang Shi 2221c5b5a3ddSYang Shi /* 22221bc115d8SMel Gorman * Don't migrate file pages that are mapped in multiple processes 22231bc115d8SMel Gorman * with execute permissions as they are probably shared libraries. 22247039e1dbSPeter Zijlstra */ 22257ee820eeSMiaohe Lin if (page_mapcount(page) != 1 && page_is_file_lru(page) && 22267ee820eeSMiaohe Lin (vma->vm_flags & VM_EXEC)) 22277039e1dbSPeter Zijlstra goto out; 22287039e1dbSPeter Zijlstra 2229a8f60772SMel Gorman /* 223009a913a7SMel Gorman * Also do not migrate dirty pages as not all filesystems can move 223109a913a7SMel Gorman * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. 223209a913a7SMel Gorman */ 22339de4f22aSHuang Ying if (page_is_file_lru(page) && PageDirty(page)) 223409a913a7SMel Gorman goto out; 223509a913a7SMel Gorman 2236b32967ffSMel Gorman isolated = numamigrate_isolate_page(pgdat, page); 2237b32967ffSMel Gorman if (!isolated) 22387039e1dbSPeter Zijlstra goto out; 22397039e1dbSPeter Zijlstra 22407039e1dbSPeter Zijlstra list_add(&page->lru, &migratepages); 2241c185e494SMatthew Wilcox (Oracle) nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, 2242c185e494SMatthew Wilcox (Oracle) NULL, node, MIGRATE_ASYNC, 2243c185e494SMatthew Wilcox (Oracle) MR_NUMA_MISPLACED, &nr_succeeded); 22447039e1dbSPeter Zijlstra if (nr_remaining) { 224559c82b70SJoonsoo Kim if (!list_empty(&migratepages)) { 224659c82b70SJoonsoo Kim list_del(&page->lru); 2247c5fc5c3aSYang Shi mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + 2248c5fc5c3aSYang Shi page_is_file_lru(page), -nr_pages); 224959c82b70SJoonsoo Kim putback_lru_page(page); 225059c82b70SJoonsoo Kim } 22517039e1dbSPeter Zijlstra isolated = 0; 2252e39bb6beSHuang Ying } 2253e39bb6beSHuang Ying if (nr_succeeded) { 2254e39bb6beSHuang Ying count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded); 2255e39bb6beSHuang Ying if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node)) 2256e39bb6beSHuang Ying mod_node_page_state(pgdat, PGPROMOTE_SUCCESS, 2257e39bb6beSHuang Ying nr_succeeded); 2258e39bb6beSHuang Ying } 22597039e1dbSPeter Zijlstra BUG_ON(!list_empty(&migratepages)); 22607039e1dbSPeter Zijlstra return isolated; 2261340ef390SHugh Dickins 2262340ef390SHugh Dickins out: 2263340ef390SHugh Dickins put_page(page); 2264340ef390SHugh Dickins return 0; 22657039e1dbSPeter Zijlstra } 2266220018d3SMel Gorman #endif /* CONFIG_NUMA_BALANCING */ 22677d6e2d96SOscar Salvador #endif /* CONFIG_NUMA */ 2268