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> 24e9995ef9SHugh Dickins #include <linux/ksm.h> 25b20a3503SChristoph Lameter #include <linux/rmap.h> 26b20a3503SChristoph Lameter #include <linux/topology.h> 27b20a3503SChristoph Lameter #include <linux/cpu.h> 28b20a3503SChristoph Lameter #include <linux/cpuset.h> 2904e62a29SChristoph Lameter #include <linux/writeback.h> 30742755a1SChristoph Lameter #include <linux/mempolicy.h> 31742755a1SChristoph Lameter #include <linux/vmalloc.h> 3286c3a764SDavid Quigley #include <linux/security.h> 3342cb14b1SHugh Dickins #include <linux/backing-dev.h> 34bda807d4SMinchan Kim #include <linux/compaction.h> 354f5ca265SAdrian Bunk #include <linux/syscalls.h> 367addf443SDominik Brodowski #include <linux/compat.h> 37290408d4SNaoya Horiguchi #include <linux/hugetlb.h> 388e6ac7faSAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 395a0e3ad6STejun Heo #include <linux/gfp.h> 40df6ad698SJérôme Glisse #include <linux/pfn_t.h> 41a5430ddaSJérôme Glisse #include <linux/memremap.h> 428315ada7SJérôme Glisse #include <linux/userfaultfd_k.h> 43bf6bddf1SRafael Aquini #include <linux/balloon_compaction.h> 4433c3fc71SVladimir Davydov #include <linux/page_idle.h> 45d435edcaSVlastimil Babka #include <linux/page_owner.h> 466e84f315SIngo Molnar #include <linux/sched/mm.h> 47197e7e52SLinus Torvalds #include <linux/ptrace.h> 4834290e2cSRalph Campbell #include <linux/oom.h> 49884a6e5dSDave Hansen #include <linux/memory.h> 50ac16ec83SBaolin Wang #include <linux/random.h> 51c574bbe9SHuang Ying #include <linux/sched/sysctl.h> 52467b171aSAneesh Kumar K.V #include <linux/memory-tiers.h> 53b20a3503SChristoph Lameter 540d1836c3SMichal Nazarewicz #include <asm/tlbflush.h> 550d1836c3SMichal Nazarewicz 567b2a2d4aSMel Gorman #include <trace/events/migrate.h> 577b2a2d4aSMel Gorman 58b20a3503SChristoph Lameter #include "internal.h" 59b20a3503SChristoph Lameter 60cd775580SBaolin Wang bool isolate_movable_page(struct page *page, isolate_mode_t mode) 61bda807d4SMinchan Kim { 6219979497SVishal Moola (Oracle) struct folio *folio = folio_get_nontail_page(page); 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 */ 7419979497SVishal Moola (Oracle) if (!folio) 75bda807d4SMinchan Kim goto out; 76bda807d4SMinchan Kim 7719979497SVishal Moola (Oracle) if (unlikely(folio_test_slab(folio))) 7819979497SVishal Moola (Oracle) goto out_putfolio; 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 */ 8619979497SVishal Moola (Oracle) if (unlikely(!__folio_test_movable(folio))) 8719979497SVishal Moola (Oracle) goto out_putfolio; 888b881763SVlastimil Babka /* Pairs with smp_wmb() in slab allocation, e.g. SLUB's alloc_slab_page() */ 898b881763SVlastimil Babka smp_rmb(); 9019979497SVishal Moola (Oracle) if (unlikely(folio_test_slab(folio))) 9119979497SVishal Moola (Oracle) goto out_putfolio; 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 */ 10419979497SVishal Moola (Oracle) if (unlikely(!folio_trylock(folio))) 10519979497SVishal Moola (Oracle) goto out_putfolio; 106bda807d4SMinchan Kim 10719979497SVishal Moola (Oracle) if (!folio_test_movable(folio) || folio_test_isolated(folio)) 108bda807d4SMinchan Kim goto out_no_isolated; 109bda807d4SMinchan Kim 11019979497SVishal Moola (Oracle) mops = folio_movable_ops(folio); 11119979497SVishal Moola (Oracle) VM_BUG_ON_FOLIO(!mops, folio); 112bda807d4SMinchan Kim 11319979497SVishal Moola (Oracle) if (!mops->isolate_page(&folio->page, mode)) 114bda807d4SMinchan Kim goto out_no_isolated; 115bda807d4SMinchan Kim 116bda807d4SMinchan Kim /* Driver shouldn't use PG_isolated bit of page->flags */ 11719979497SVishal Moola (Oracle) WARN_ON_ONCE(folio_test_isolated(folio)); 11819979497SVishal Moola (Oracle) folio_set_isolated(folio); 11919979497SVishal Moola (Oracle) folio_unlock(folio); 120bda807d4SMinchan Kim 121cd775580SBaolin Wang return true; 122bda807d4SMinchan Kim 123bda807d4SMinchan Kim out_no_isolated: 12419979497SVishal Moola (Oracle) folio_unlock(folio); 12519979497SVishal Moola (Oracle) out_putfolio: 12619979497SVishal Moola (Oracle) folio_put(folio); 127bda807d4SMinchan Kim out: 128cd775580SBaolin Wang return false; 129bda807d4SMinchan Kim } 130bda807d4SMinchan Kim 131280d724aSVishal Moola (Oracle) static void putback_movable_folio(struct folio *folio) 132bda807d4SMinchan Kim { 133280d724aSVishal Moola (Oracle) const struct movable_operations *mops = folio_movable_ops(folio); 134bda807d4SMinchan Kim 135280d724aSVishal Moola (Oracle) mops->putback_page(&folio->page); 136280d724aSVishal Moola (Oracle) folio_clear_isolated(folio); 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 { 149280d724aSVishal Moola (Oracle) struct folio *folio; 150280d724aSVishal Moola (Oracle) struct folio *folio2; 1515733c7d1SRafael Aquini 152280d724aSVishal Moola (Oracle) list_for_each_entry_safe(folio, folio2, l, lru) { 153280d724aSVishal Moola (Oracle) if (unlikely(folio_test_hugetlb(folio))) { 154280d724aSVishal Moola (Oracle) folio_putback_active_hugetlb(folio); 15531caf665SNaoya Horiguchi continue; 15631caf665SNaoya Horiguchi } 157280d724aSVishal Moola (Oracle) list_del(&folio->lru); 158bda807d4SMinchan Kim /* 159280d724aSVishal Moola (Oracle) * We isolated non-lru movable folio so here we can use 160280d724aSVishal Moola (Oracle) * __PageMovable because LRU folio's mapping cannot have 161bda807d4SMinchan Kim * PAGE_MAPPING_MOVABLE. 162bda807d4SMinchan Kim */ 163280d724aSVishal Moola (Oracle) if (unlikely(__folio_test_movable(folio))) { 164280d724aSVishal Moola (Oracle) VM_BUG_ON_FOLIO(!folio_test_isolated(folio), folio); 165280d724aSVishal Moola (Oracle) folio_lock(folio); 166280d724aSVishal Moola (Oracle) if (folio_test_movable(folio)) 167280d724aSVishal Moola (Oracle) putback_movable_folio(folio); 168bf6bddf1SRafael Aquini else 169280d724aSVishal Moola (Oracle) folio_clear_isolated(folio); 170280d724aSVishal Moola (Oracle) folio_unlock(folio); 171280d724aSVishal Moola (Oracle) folio_put(folio); 172bda807d4SMinchan Kim } else { 173280d724aSVishal Moola (Oracle) node_stat_mod_folio(folio, NR_ISOLATED_ANON + 174280d724aSVishal Moola (Oracle) folio_is_file_lru(folio), -folio_nr_pages(folio)); 175280d724aSVishal Moola (Oracle) folio_putback_lru(folio); 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; 190c33c7948SRyan Roberts pte_t old_pte; 1910697212aSChristoph Lameter pte_t pte; 1920697212aSChristoph Lameter swp_entry_t entry; 1934eecb8b9SMatthew Wilcox (Oracle) struct page *new; 1944eecb8b9SMatthew Wilcox (Oracle) unsigned long idx = 0; 1950697212aSChristoph Lameter 1964eecb8b9SMatthew Wilcox (Oracle) /* pgoff is invalid for ksm pages, but they are never large */ 1974eecb8b9SMatthew Wilcox (Oracle) if (folio_test_large(folio) && !folio_test_hugetlb(folio)) 1984eecb8b9SMatthew Wilcox (Oracle) idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff; 1994eecb8b9SMatthew Wilcox (Oracle) new = folio_page(folio, idx); 2000697212aSChristoph Lameter 201616b8371SZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 202616b8371SZi Yan /* PMD-mapped THP migration entry */ 203616b8371SZi Yan if (!pvmw.pte) { 2044eecb8b9SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) || 2054eecb8b9SMatthew Wilcox (Oracle) !folio_test_pmd_mappable(folio), folio); 206616b8371SZi Yan remove_migration_pmd(&pvmw, new); 207616b8371SZi Yan continue; 208616b8371SZi Yan } 209616b8371SZi Yan #endif 210616b8371SZi Yan 2114eecb8b9SMatthew Wilcox (Oracle) folio_get(folio); 2122e346877SPeter Xu pte = mk_pte(new, READ_ONCE(vma->vm_page_prot)); 213c33c7948SRyan Roberts old_pte = ptep_get(pvmw.pte); 214c33c7948SRyan Roberts if (pte_swp_soft_dirty(old_pte)) 215c3d16e16SCyrill Gorcunov pte = pte_mksoft_dirty(pte); 216d3cb8bf6SMel Gorman 217c33c7948SRyan Roberts entry = pte_to_swp_entry(old_pte); 2182e346877SPeter Xu if (!is_migration_entry_young(entry)) 2192e346877SPeter Xu pte = pte_mkold(pte); 2202e346877SPeter Xu if (folio_test_dirty(folio) && is_migration_entry_dirty(entry)) 2212e346877SPeter Xu pte = pte_mkdirty(pte); 2224dd845b5SAlistair Popple if (is_writable_migration_entry(entry)) 223161e393cSRick Edgecombe pte = pte_mkwrite(pte, vma); 224c33c7948SRyan Roberts else if (pte_swp_uffd_wp(old_pte)) 225f45ec5ffSPeter Xu pte = pte_mkuffd_wp(pte); 226d3cb8bf6SMel Gorman 2276c287605SDavid Hildenbrand if (folio_test_anon(folio) && !is_readable_migration_entry(entry)) 2286c287605SDavid Hildenbrand rmap_flags |= RMAP_EXCLUSIVE; 2296c287605SDavid Hildenbrand 2306128763fSRalph Campbell if (unlikely(is_device_private_page(new))) { 2314dd845b5SAlistair Popple if (pte_write(pte)) 2324dd845b5SAlistair Popple entry = make_writable_device_private_entry( 2334dd845b5SAlistair Popple page_to_pfn(new)); 2344dd845b5SAlistair Popple else 2354dd845b5SAlistair Popple entry = make_readable_device_private_entry( 2364dd845b5SAlistair Popple page_to_pfn(new)); 237a5430ddaSJérôme Glisse pte = swp_entry_to_pte(entry); 238c33c7948SRyan Roberts if (pte_swp_soft_dirty(old_pte)) 2393d321bf8SRalph Campbell pte = pte_swp_mksoft_dirty(pte); 240c33c7948SRyan Roberts if (pte_swp_uffd_wp(old_pte)) 241ebdf8321SAlistair Popple pte = pte_swp_mkuffd_wp(pte); 242df6ad698SJérôme Glisse } 243a5430ddaSJérôme Glisse 2443ef8fd7fSAndi Kleen #ifdef CONFIG_HUGETLB_PAGE 2454eecb8b9SMatthew Wilcox (Oracle) if (folio_test_hugetlb(folio)) { 246935d4f0cSRyan Roberts struct hstate *h = hstate_vma(vma); 247935d4f0cSRyan Roberts unsigned int shift = huge_page_shift(h); 248935d4f0cSRyan Roberts unsigned long psize = huge_page_size(h); 24979c1c594SChristophe Leroy 25079c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 2514eecb8b9SMatthew Wilcox (Oracle) if (folio_test_anon(folio)) 25209c55050SDavid Hildenbrand hugepage_add_anon_rmap(folio, vma, pvmw.address, 2536c287605SDavid Hildenbrand rmap_flags); 254290408d4SNaoya Horiguchi else 255fb3d824dSDavid Hildenbrand page_dup_file_rmap(new, true); 256935d4f0cSRyan Roberts set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte, 257935d4f0cSRyan Roberts psize); 258383321abSAneesh Kumar K.V } else 259383321abSAneesh Kumar K.V #endif 260383321abSAneesh Kumar K.V { 2614eecb8b9SMatthew Wilcox (Oracle) if (folio_test_anon(folio)) 262f1e2db12SDavid Hildenbrand page_add_anon_rmap(new, vma, pvmw.address, 2636c287605SDavid Hildenbrand rmap_flags); 26404e62a29SChristoph Lameter else 265cea86fe2SHugh Dickins page_add_file_rmap(new, vma, false); 2661eba86c0SPasha Tatashin set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); 267383321abSAneesh Kumar K.V } 268b7435507SHugh Dickins if (vma->vm_flags & VM_LOCKED) 26996f97c43SLorenzo Stoakes mlock_drain_local(); 270e125fe40SKirill A. Shutemov 2714cc79b33SAnshuman Khandual trace_remove_migration_pte(pvmw.address, pte_val(pte), 2724cc79b33SAnshuman Khandual compound_order(new)); 2734cc79b33SAnshuman Khandual 27404e62a29SChristoph Lameter /* No need to invalidate - it was non-present before */ 2753fe87967SKirill A. Shutemov update_mmu_cache(vma, pvmw.address, pvmw.pte); 2763fe87967SKirill A. Shutemov } 2773fe87967SKirill A. Shutemov 278e4b82222SMinchan Kim return true; 2790697212aSChristoph Lameter } 2800697212aSChristoph Lameter 2810697212aSChristoph Lameter /* 28204e62a29SChristoph Lameter * Get rid of all migration entries and replace them by 28304e62a29SChristoph Lameter * references to the indicated page. 28404e62a29SChristoph Lameter */ 2854eecb8b9SMatthew Wilcox (Oracle) void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked) 28604e62a29SChristoph Lameter { 287051ac83aSJoonsoo Kim struct rmap_walk_control rwc = { 288051ac83aSJoonsoo Kim .rmap_one = remove_migration_pte, 2894eecb8b9SMatthew Wilcox (Oracle) .arg = src, 290051ac83aSJoonsoo Kim }; 291051ac83aSJoonsoo Kim 292e388466dSKirill A. Shutemov if (locked) 2932f031c6fSMatthew Wilcox (Oracle) rmap_walk_locked(dst, &rwc); 294e388466dSKirill A. Shutemov else 2952f031c6fSMatthew Wilcox (Oracle) rmap_walk(dst, &rwc); 29604e62a29SChristoph Lameter } 29704e62a29SChristoph Lameter 29804e62a29SChristoph Lameter /* 2990697212aSChristoph Lameter * Something used the pte of a page under migration. We need to 3000697212aSChristoph Lameter * get to the page and wait until migration is finished. 3010697212aSChristoph Lameter * When we return from this function the fault will be retried. 3020697212aSChristoph Lameter */ 3030cb8fd4dSHugh Dickins void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, 3040cb8fd4dSHugh Dickins unsigned long address) 3050697212aSChristoph Lameter { 3060cb8fd4dSHugh Dickins spinlock_t *ptl; 3070cb8fd4dSHugh Dickins pte_t *ptep; 30830dad309SNaoya Horiguchi pte_t pte; 3090697212aSChristoph Lameter swp_entry_t entry; 3100697212aSChristoph Lameter 3110cb8fd4dSHugh Dickins ptep = pte_offset_map_lock(mm, pmd, address, &ptl); 31204dee9e8SHugh Dickins if (!ptep) 31304dee9e8SHugh Dickins return; 31404dee9e8SHugh Dickins 315c33c7948SRyan Roberts pte = ptep_get(ptep); 3160cb8fd4dSHugh Dickins pte_unmap(ptep); 3170cb8fd4dSHugh Dickins 3180697212aSChristoph Lameter if (!is_swap_pte(pte)) 3190697212aSChristoph Lameter goto out; 3200697212aSChristoph Lameter 3210697212aSChristoph Lameter entry = pte_to_swp_entry(pte); 3220697212aSChristoph Lameter if (!is_migration_entry(entry)) 3230697212aSChristoph Lameter goto out; 3240697212aSChristoph Lameter 3250cb8fd4dSHugh Dickins migration_entry_wait_on_locked(entry, ptl); 3260697212aSChristoph Lameter return; 3270697212aSChristoph Lameter out: 3280cb8fd4dSHugh Dickins spin_unlock(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 */ 3380cb8fd4dSHugh Dickins void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *ptep) 33930dad309SNaoya Horiguchi { 3400cb8fd4dSHugh Dickins spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, ptep); 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); 3580cb8fd4dSHugh Dickins migration_entry_wait_on_locked(pte_to_swp_entry(pte), ptl); 35930dad309SNaoya Horiguchi } 360fcd48540SPeter Xu } 361ad1ac596SMiaohe Lin #endif 362ad1ac596SMiaohe Lin 363616b8371SZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 364616b8371SZi Yan void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) 365616b8371SZi Yan { 366616b8371SZi Yan spinlock_t *ptl; 367616b8371SZi Yan 368616b8371SZi Yan ptl = pmd_lock(mm, pmd); 369616b8371SZi Yan if (!is_pmd_migration_entry(*pmd)) 370616b8371SZi Yan goto unlock; 3710cb8fd4dSHugh Dickins migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), ptl); 372616b8371SZi Yan return; 373616b8371SZi Yan unlock: 374616b8371SZi Yan spin_unlock(ptl); 375616b8371SZi Yan } 376616b8371SZi Yan #endif 377616b8371SZi Yan 378108ca835SMatthew Wilcox (Oracle) static int folio_expected_refs(struct address_space *mapping, 379108ca835SMatthew Wilcox (Oracle) struct folio *folio) 3800b3901b3SJan Kara { 381108ca835SMatthew Wilcox (Oracle) int refs = 1; 382108ca835SMatthew Wilcox (Oracle) if (!mapping) 383108ca835SMatthew Wilcox (Oracle) return refs; 3840b3901b3SJan Kara 385108ca835SMatthew Wilcox (Oracle) refs += folio_nr_pages(folio); 386108ca835SMatthew Wilcox (Oracle) if (folio_test_private(folio)) 387108ca835SMatthew Wilcox (Oracle) refs++; 388108ca835SMatthew Wilcox (Oracle) 389108ca835SMatthew Wilcox (Oracle) return refs; 3900b3901b3SJan Kara } 3910b3901b3SJan Kara 392b20a3503SChristoph Lameter /* 393c3fcf8a5SChristoph Lameter * Replace the page in the mapping. 3945b5c7120SChristoph Lameter * 3955b5c7120SChristoph Lameter * The number of remaining references must be: 3965b5c7120SChristoph Lameter * 1 for anonymous pages without a mapping 3975b5c7120SChristoph Lameter * 2 for pages with a mapping 398266cf658SDavid Howells * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. 399b20a3503SChristoph Lameter */ 4003417013eSMatthew Wilcox (Oracle) int folio_migrate_mapping(struct address_space *mapping, 4013417013eSMatthew Wilcox (Oracle) struct folio *newfolio, struct folio *folio, int extra_count) 402b20a3503SChristoph Lameter { 4033417013eSMatthew Wilcox (Oracle) XA_STATE(xas, &mapping->i_pages, folio_index(folio)); 40442cb14b1SHugh Dickins struct zone *oldzone, *newzone; 40542cb14b1SHugh Dickins int dirty; 406108ca835SMatthew Wilcox (Oracle) int expected_count = folio_expected_refs(mapping, folio) + extra_count; 4073417013eSMatthew Wilcox (Oracle) long nr = folio_nr_pages(folio); 4088763cb45SJérôme Glisse 4096c5240aeSChristoph Lameter if (!mapping) { 4100e8c7d0fSChristoph Lameter /* Anonymous page without mapping */ 4113417013eSMatthew Wilcox (Oracle) if (folio_ref_count(folio) != expected_count) 4126c5240aeSChristoph Lameter return -EAGAIN; 413cf4b769aSHugh Dickins 414cf4b769aSHugh Dickins /* No turning back from here */ 4153417013eSMatthew Wilcox (Oracle) newfolio->index = folio->index; 4163417013eSMatthew Wilcox (Oracle) newfolio->mapping = folio->mapping; 4173417013eSMatthew Wilcox (Oracle) if (folio_test_swapbacked(folio)) 4183417013eSMatthew Wilcox (Oracle) __folio_set_swapbacked(newfolio); 419cf4b769aSHugh Dickins 42078bd5209SRafael Aquini return MIGRATEPAGE_SUCCESS; 4216c5240aeSChristoph Lameter } 4226c5240aeSChristoph Lameter 4233417013eSMatthew Wilcox (Oracle) oldzone = folio_zone(folio); 4243417013eSMatthew Wilcox (Oracle) newzone = folio_zone(newfolio); 42542cb14b1SHugh Dickins 42689eb946aSMatthew Wilcox xas_lock_irq(&xas); 4273417013eSMatthew Wilcox (Oracle) if (!folio_ref_freeze(folio, expected_count)) { 42889eb946aSMatthew Wilcox xas_unlock_irq(&xas); 429e286781dSNick Piggin return -EAGAIN; 430e286781dSNick Piggin } 431e286781dSNick Piggin 432b20a3503SChristoph Lameter /* 4333417013eSMatthew Wilcox (Oracle) * Now we know that no one else is looking at the folio: 434cf4b769aSHugh Dickins * no turning back from here. 435b20a3503SChristoph Lameter */ 4363417013eSMatthew Wilcox (Oracle) newfolio->index = folio->index; 4373417013eSMatthew Wilcox (Oracle) newfolio->mapping = folio->mapping; 4383417013eSMatthew Wilcox (Oracle) folio_ref_add(newfolio, nr); /* add cache reference */ 4393417013eSMatthew Wilcox (Oracle) if (folio_test_swapbacked(folio)) { 4403417013eSMatthew Wilcox (Oracle) __folio_set_swapbacked(newfolio); 4413417013eSMatthew Wilcox (Oracle) if (folio_test_swapcache(folio)) { 4423417013eSMatthew Wilcox (Oracle) folio_set_swapcache(newfolio); 4433417013eSMatthew Wilcox (Oracle) newfolio->private = folio_get_private(folio); 444b20a3503SChristoph Lameter } 4456326fec1SNicholas Piggin } else { 4463417013eSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio); 4476326fec1SNicholas Piggin } 448b20a3503SChristoph Lameter 44942cb14b1SHugh Dickins /* Move dirty while page refs frozen and newpage not yet exposed */ 4503417013eSMatthew Wilcox (Oracle) dirty = folio_test_dirty(folio); 45142cb14b1SHugh Dickins if (dirty) { 4523417013eSMatthew Wilcox (Oracle) folio_clear_dirty(folio); 4533417013eSMatthew Wilcox (Oracle) folio_set_dirty(newfolio); 45442cb14b1SHugh Dickins } 45542cb14b1SHugh Dickins 4563417013eSMatthew Wilcox (Oracle) xas_store(&xas, newfolio); 4577cf9c2c7SNick Piggin 4587cf9c2c7SNick Piggin /* 459937a94c9SJacobo Giralt * Drop cache reference from old page by unfreezing 460937a94c9SJacobo Giralt * to one less reference. 4617cf9c2c7SNick Piggin * We know this isn't the last reference. 4627cf9c2c7SNick Piggin */ 4633417013eSMatthew Wilcox (Oracle) folio_ref_unfreeze(folio, expected_count - nr); 4647cf9c2c7SNick Piggin 46589eb946aSMatthew Wilcox xas_unlock(&xas); 46642cb14b1SHugh Dickins /* Leave irq disabled to prevent preemption while updating stats */ 46742cb14b1SHugh Dickins 4680e8c7d0fSChristoph Lameter /* 4690e8c7d0fSChristoph Lameter * If moved to a different zone then also account 4700e8c7d0fSChristoph Lameter * the page for that zone. Other VM counters will be 4710e8c7d0fSChristoph Lameter * taken care of when we establish references to the 4720e8c7d0fSChristoph Lameter * new page and drop references to the old page. 4730e8c7d0fSChristoph Lameter * 4740e8c7d0fSChristoph Lameter * Note that anonymous pages are accounted for 4754b9d0fabSMel Gorman * via NR_FILE_PAGES and NR_ANON_MAPPED if they 4760e8c7d0fSChristoph Lameter * are mapped to swap space. 4770e8c7d0fSChristoph Lameter */ 47842cb14b1SHugh Dickins if (newzone != oldzone) { 4790d1c2072SJohannes Weiner struct lruvec *old_lruvec, *new_lruvec; 4800d1c2072SJohannes Weiner struct mem_cgroup *memcg; 4810d1c2072SJohannes Weiner 4823417013eSMatthew Wilcox (Oracle) memcg = folio_memcg(folio); 4830d1c2072SJohannes Weiner old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat); 4840d1c2072SJohannes Weiner new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat); 4850d1c2072SJohannes Weiner 4865c447d27SShakeel Butt __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr); 4875c447d27SShakeel Butt __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr); 4883417013eSMatthew Wilcox (Oracle) if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) { 4895c447d27SShakeel Butt __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr); 4905c447d27SShakeel Butt __mod_lruvec_state(new_lruvec, NR_SHMEM, nr); 4910b52c420SJan Glauber 4920b52c420SJan Glauber if (folio_test_pmd_mappable(folio)) { 4930b52c420SJan Glauber __mod_lruvec_state(old_lruvec, NR_SHMEM_THPS, -nr); 4940b52c420SJan Glauber __mod_lruvec_state(new_lruvec, NR_SHMEM_THPS, nr); 4950b52c420SJan Glauber } 4964b02108aSKOSAKI Motohiro } 497b6038942SShakeel Butt #ifdef CONFIG_SWAP 4983417013eSMatthew Wilcox (Oracle) if (folio_test_swapcache(folio)) { 499b6038942SShakeel Butt __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr); 500b6038942SShakeel Butt __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr); 501b6038942SShakeel Butt } 502b6038942SShakeel Butt #endif 503f56753acSChristoph Hellwig if (dirty && mapping_can_writeback(mapping)) { 5045c447d27SShakeel Butt __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr); 5055c447d27SShakeel Butt __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr); 5065c447d27SShakeel Butt __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr); 5075c447d27SShakeel Butt __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr); 50842cb14b1SHugh Dickins } 50942cb14b1SHugh Dickins } 51042cb14b1SHugh Dickins local_irq_enable(); 511b20a3503SChristoph Lameter 51278bd5209SRafael Aquini return MIGRATEPAGE_SUCCESS; 513b20a3503SChristoph Lameter } 5143417013eSMatthew Wilcox (Oracle) EXPORT_SYMBOL(folio_migrate_mapping); 515b20a3503SChristoph Lameter 516b20a3503SChristoph Lameter /* 517290408d4SNaoya Horiguchi * The expected number of remaining references is the same as that 5183417013eSMatthew Wilcox (Oracle) * of folio_migrate_mapping(). 519290408d4SNaoya Horiguchi */ 520290408d4SNaoya Horiguchi int migrate_huge_page_move_mapping(struct address_space *mapping, 521b890ec2aSMatthew Wilcox (Oracle) struct folio *dst, struct folio *src) 522290408d4SNaoya Horiguchi { 523b890ec2aSMatthew Wilcox (Oracle) XA_STATE(xas, &mapping->i_pages, folio_index(src)); 524290408d4SNaoya Horiguchi int expected_count; 525290408d4SNaoya Horiguchi 52689eb946aSMatthew Wilcox xas_lock_irq(&xas); 527b890ec2aSMatthew Wilcox (Oracle) expected_count = 2 + folio_has_private(src); 528b890ec2aSMatthew Wilcox (Oracle) if (!folio_ref_freeze(src, expected_count)) { 52989eb946aSMatthew Wilcox xas_unlock_irq(&xas); 530290408d4SNaoya Horiguchi return -EAGAIN; 531290408d4SNaoya Horiguchi } 532290408d4SNaoya Horiguchi 533b890ec2aSMatthew Wilcox (Oracle) dst->index = src->index; 534b890ec2aSMatthew Wilcox (Oracle) dst->mapping = src->mapping; 5356a93ca8fSJohannes Weiner 536b890ec2aSMatthew Wilcox (Oracle) folio_get(dst); 537290408d4SNaoya Horiguchi 538b890ec2aSMatthew Wilcox (Oracle) xas_store(&xas, dst); 539290408d4SNaoya Horiguchi 540b890ec2aSMatthew Wilcox (Oracle) folio_ref_unfreeze(src, expected_count - 1); 541290408d4SNaoya Horiguchi 54289eb946aSMatthew Wilcox xas_unlock_irq(&xas); 5436a93ca8fSJohannes Weiner 54478bd5209SRafael Aquini return MIGRATEPAGE_SUCCESS; 545290408d4SNaoya Horiguchi } 546290408d4SNaoya Horiguchi 547290408d4SNaoya Horiguchi /* 54819138349SMatthew Wilcox (Oracle) * Copy the flags and some other ancillary information 549b20a3503SChristoph Lameter */ 55019138349SMatthew Wilcox (Oracle) void folio_migrate_flags(struct folio *newfolio, struct folio *folio) 551b20a3503SChristoph Lameter { 5527851a45cSRik van Riel int cpupid; 5537851a45cSRik van Riel 55419138349SMatthew Wilcox (Oracle) if (folio_test_error(folio)) 55519138349SMatthew Wilcox (Oracle) folio_set_error(newfolio); 55619138349SMatthew Wilcox (Oracle) if (folio_test_referenced(folio)) 55719138349SMatthew Wilcox (Oracle) folio_set_referenced(newfolio); 55819138349SMatthew Wilcox (Oracle) if (folio_test_uptodate(folio)) 55919138349SMatthew Wilcox (Oracle) folio_mark_uptodate(newfolio); 56019138349SMatthew Wilcox (Oracle) if (folio_test_clear_active(folio)) { 56119138349SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio); 56219138349SMatthew Wilcox (Oracle) folio_set_active(newfolio); 56319138349SMatthew Wilcox (Oracle) } else if (folio_test_clear_unevictable(folio)) 56419138349SMatthew Wilcox (Oracle) folio_set_unevictable(newfolio); 56519138349SMatthew Wilcox (Oracle) if (folio_test_workingset(folio)) 56619138349SMatthew Wilcox (Oracle) folio_set_workingset(newfolio); 56719138349SMatthew Wilcox (Oracle) if (folio_test_checked(folio)) 56819138349SMatthew Wilcox (Oracle) folio_set_checked(newfolio); 5696c287605SDavid Hildenbrand /* 5706c287605SDavid Hildenbrand * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via 5716c287605SDavid Hildenbrand * migration entries. We can still have PG_anon_exclusive set on an 5726c287605SDavid Hildenbrand * effectively unmapped and unreferenced first sub-pages of an 5736c287605SDavid Hildenbrand * anonymous THP: we can simply copy it here via PG_mappedtodisk. 5746c287605SDavid Hildenbrand */ 57519138349SMatthew Wilcox (Oracle) if (folio_test_mappedtodisk(folio)) 57619138349SMatthew Wilcox (Oracle) folio_set_mappedtodisk(newfolio); 577b20a3503SChristoph Lameter 5783417013eSMatthew Wilcox (Oracle) /* Move dirty on pages not done by folio_migrate_mapping() */ 57919138349SMatthew Wilcox (Oracle) if (folio_test_dirty(folio)) 58019138349SMatthew Wilcox (Oracle) folio_set_dirty(newfolio); 581b20a3503SChristoph Lameter 58219138349SMatthew Wilcox (Oracle) if (folio_test_young(folio)) 58319138349SMatthew Wilcox (Oracle) folio_set_young(newfolio); 58419138349SMatthew Wilcox (Oracle) if (folio_test_idle(folio)) 58519138349SMatthew Wilcox (Oracle) folio_set_idle(newfolio); 58633c3fc71SVladimir Davydov 5877851a45cSRik van Riel /* 5887851a45cSRik van Riel * Copy NUMA information to the new page, to prevent over-eager 5897851a45cSRik van Riel * future migrations of this same page. 5907851a45cSRik van Riel */ 59119138349SMatthew Wilcox (Oracle) cpupid = page_cpupid_xchg_last(&folio->page, -1); 59233024536SHuang Ying /* 59333024536SHuang Ying * For memory tiering mode, when migrate between slow and fast 59433024536SHuang Ying * memory node, reset cpupid, because that is used to record 59533024536SHuang Ying * page access time in slow memory node. 59633024536SHuang Ying */ 59733024536SHuang Ying if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) { 59833024536SHuang Ying bool f_toptier = node_is_toptier(page_to_nid(&folio->page)); 59933024536SHuang Ying bool t_toptier = node_is_toptier(page_to_nid(&newfolio->page)); 60033024536SHuang Ying 60133024536SHuang Ying if (f_toptier != t_toptier) 60233024536SHuang Ying cpupid = -1; 60333024536SHuang Ying } 60419138349SMatthew Wilcox (Oracle) page_cpupid_xchg_last(&newfolio->page, cpupid); 6057851a45cSRik van Riel 60619138349SMatthew Wilcox (Oracle) folio_migrate_ksm(newfolio, folio); 607c8d6553bSHugh Dickins /* 608c8d6553bSHugh Dickins * Please do not reorder this without considering how mm/ksm.c's 609c8d6553bSHugh Dickins * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). 610c8d6553bSHugh Dickins */ 61119138349SMatthew Wilcox (Oracle) if (folio_test_swapcache(folio)) 61219138349SMatthew Wilcox (Oracle) folio_clear_swapcache(folio); 61319138349SMatthew Wilcox (Oracle) folio_clear_private(folio); 614ad2fa371SMuchun Song 615ad2fa371SMuchun Song /* page->private contains hugetlb specific flags */ 61619138349SMatthew Wilcox (Oracle) if (!folio_test_hugetlb(folio)) 61719138349SMatthew Wilcox (Oracle) folio->private = NULL; 618b20a3503SChristoph Lameter 619b20a3503SChristoph Lameter /* 620b20a3503SChristoph Lameter * If any waiters have accumulated on the new page then 621b20a3503SChristoph Lameter * wake them up. 622b20a3503SChristoph Lameter */ 62319138349SMatthew Wilcox (Oracle) if (folio_test_writeback(newfolio)) 62419138349SMatthew Wilcox (Oracle) folio_end_writeback(newfolio); 625d435edcaSVlastimil Babka 6266aeff241SYang Shi /* 6276aeff241SYang Shi * PG_readahead shares the same bit with PG_reclaim. The above 6286aeff241SYang Shi * end_page_writeback() may clear PG_readahead mistakenly, so set the 6296aeff241SYang Shi * bit after that. 6306aeff241SYang Shi */ 63119138349SMatthew Wilcox (Oracle) if (folio_test_readahead(folio)) 63219138349SMatthew Wilcox (Oracle) folio_set_readahead(newfolio); 6336aeff241SYang Shi 63419138349SMatthew Wilcox (Oracle) folio_copy_owner(newfolio, folio); 63574485cf2SJohannes Weiner 63619138349SMatthew Wilcox (Oracle) if (!folio_test_hugetlb(folio)) 637d21bba2bSMatthew Wilcox (Oracle) mem_cgroup_migrate(folio, newfolio); 638b20a3503SChristoph Lameter } 63919138349SMatthew Wilcox (Oracle) EXPORT_SYMBOL(folio_migrate_flags); 6402916ecc0SJérôme Glisse 641715cbfd6SMatthew Wilcox (Oracle) void folio_migrate_copy(struct folio *newfolio, struct folio *folio) 6422916ecc0SJérôme Glisse { 643715cbfd6SMatthew Wilcox (Oracle) folio_copy(newfolio, folio); 644715cbfd6SMatthew Wilcox (Oracle) folio_migrate_flags(newfolio, folio); 6452916ecc0SJérôme Glisse } 646715cbfd6SMatthew Wilcox (Oracle) EXPORT_SYMBOL(folio_migrate_copy); 647b20a3503SChristoph Lameter 6481d8b85ccSChristoph Lameter /************************************************************ 6491d8b85ccSChristoph Lameter * Migration functions 6501d8b85ccSChristoph Lameter ***********************************************************/ 6511d8b85ccSChristoph Lameter 65216ce101dSAlistair Popple int migrate_folio_extra(struct address_space *mapping, struct folio *dst, 65316ce101dSAlistair Popple struct folio *src, enum migrate_mode mode, int extra_count) 65416ce101dSAlistair Popple { 65516ce101dSAlistair Popple int rc; 65616ce101dSAlistair Popple 65716ce101dSAlistair Popple BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */ 65816ce101dSAlistair Popple 65916ce101dSAlistair Popple rc = folio_migrate_mapping(mapping, dst, src, extra_count); 66016ce101dSAlistair Popple 66116ce101dSAlistair Popple if (rc != MIGRATEPAGE_SUCCESS) 66216ce101dSAlistair Popple return rc; 66316ce101dSAlistair Popple 66416ce101dSAlistair Popple if (mode != MIGRATE_SYNC_NO_COPY) 66516ce101dSAlistair Popple folio_migrate_copy(dst, src); 66616ce101dSAlistair Popple else 66716ce101dSAlistair Popple folio_migrate_flags(dst, src); 66816ce101dSAlistair Popple return MIGRATEPAGE_SUCCESS; 66916ce101dSAlistair Popple } 67016ce101dSAlistair Popple 67154184650SMatthew Wilcox (Oracle) /** 67254184650SMatthew Wilcox (Oracle) * migrate_folio() - Simple folio migration. 67354184650SMatthew Wilcox (Oracle) * @mapping: The address_space containing the folio. 67454184650SMatthew Wilcox (Oracle) * @dst: The folio to migrate the data to. 67554184650SMatthew Wilcox (Oracle) * @src: The folio containing the current data. 67654184650SMatthew Wilcox (Oracle) * @mode: How to migrate the page. 677b20a3503SChristoph Lameter * 67854184650SMatthew Wilcox (Oracle) * Common logic to directly migrate a single LRU folio suitable for 67954184650SMatthew Wilcox (Oracle) * folios that do not use PagePrivate/PagePrivate2. 68054184650SMatthew Wilcox (Oracle) * 68154184650SMatthew Wilcox (Oracle) * Folios are locked upon entry and exit. 682b20a3503SChristoph Lameter */ 68354184650SMatthew Wilcox (Oracle) int migrate_folio(struct address_space *mapping, struct folio *dst, 68454184650SMatthew Wilcox (Oracle) struct folio *src, enum migrate_mode mode) 685b20a3503SChristoph Lameter { 68616ce101dSAlistair Popple return migrate_folio_extra(mapping, dst, src, mode, 0); 687b20a3503SChristoph Lameter } 68854184650SMatthew Wilcox (Oracle) EXPORT_SYMBOL(migrate_folio); 689b20a3503SChristoph Lameter 690925c86a1SChristoph Hellwig #ifdef CONFIG_BUFFER_HEAD 69184ade7c1SJan Kara /* Returns true if all buffers are successfully locked */ 69284ade7c1SJan Kara static bool buffer_migrate_lock_buffers(struct buffer_head *head, 69384ade7c1SJan Kara enum migrate_mode mode) 69484ade7c1SJan Kara { 69584ade7c1SJan Kara struct buffer_head *bh = head; 6964bb6dc79SDouglas Anderson struct buffer_head *failed_bh; 69784ade7c1SJan Kara 69884ade7c1SJan Kara do { 6994bb6dc79SDouglas Anderson if (!trylock_buffer(bh)) { 7004bb6dc79SDouglas Anderson if (mode == MIGRATE_ASYNC) 7014bb6dc79SDouglas Anderson goto unlock; 7024bb6dc79SDouglas Anderson if (mode == MIGRATE_SYNC_LIGHT && !buffer_uptodate(bh)) 7034bb6dc79SDouglas Anderson goto unlock; 70484ade7c1SJan Kara lock_buffer(bh); 7054bb6dc79SDouglas Anderson } 70684ade7c1SJan Kara 7074bb6dc79SDouglas Anderson bh = bh->b_this_page; 70884ade7c1SJan Kara } while (bh != head); 70984ade7c1SJan Kara 71084ade7c1SJan Kara return true; 71184ade7c1SJan Kara 7124bb6dc79SDouglas Anderson unlock: 7134bb6dc79SDouglas Anderson /* We failed to lock the buffer and cannot stall. */ 7144bb6dc79SDouglas Anderson failed_bh = bh; 71584ade7c1SJan Kara bh = head; 71684ade7c1SJan Kara while (bh != failed_bh) { 71784ade7c1SJan Kara unlock_buffer(bh); 71884ade7c1SJan Kara bh = bh->b_this_page; 71984ade7c1SJan Kara } 72084ade7c1SJan Kara 7214bb6dc79SDouglas Anderson return false; 72284ade7c1SJan Kara } 72384ade7c1SJan Kara 72467235182SMatthew Wilcox (Oracle) static int __buffer_migrate_folio(struct address_space *mapping, 72567235182SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode, 72689cb0888SJan Kara bool check_refs) 7271d8b85ccSChristoph Lameter { 7281d8b85ccSChristoph Lameter struct buffer_head *bh, *head; 7291d8b85ccSChristoph Lameter int rc; 730cc4f11e6SJan Kara int expected_count; 7311d8b85ccSChristoph Lameter 73267235182SMatthew Wilcox (Oracle) head = folio_buffers(src); 73367235182SMatthew Wilcox (Oracle) if (!head) 73454184650SMatthew Wilcox (Oracle) return migrate_folio(mapping, dst, src, mode); 7351d8b85ccSChristoph Lameter 736cc4f11e6SJan Kara /* Check whether page does not have extra refs before we do more work */ 737108ca835SMatthew Wilcox (Oracle) expected_count = folio_expected_refs(mapping, src); 73867235182SMatthew Wilcox (Oracle) if (folio_ref_count(src) != expected_count) 739cc4f11e6SJan Kara return -EAGAIN; 740cc4f11e6SJan Kara 741cc4f11e6SJan Kara if (!buffer_migrate_lock_buffers(head, mode)) 742cc4f11e6SJan Kara return -EAGAIN; 7431d8b85ccSChristoph Lameter 74489cb0888SJan Kara if (check_refs) { 74589cb0888SJan Kara bool busy; 74689cb0888SJan Kara bool invalidated = false; 74789cb0888SJan Kara 74889cb0888SJan Kara recheck_buffers: 74989cb0888SJan Kara busy = false; 75089cb0888SJan Kara spin_lock(&mapping->private_lock); 75189cb0888SJan Kara bh = head; 75289cb0888SJan Kara do { 75389cb0888SJan Kara if (atomic_read(&bh->b_count)) { 75489cb0888SJan Kara busy = true; 75589cb0888SJan Kara break; 75689cb0888SJan Kara } 75789cb0888SJan Kara bh = bh->b_this_page; 75889cb0888SJan Kara } while (bh != head); 75989cb0888SJan Kara if (busy) { 76089cb0888SJan Kara if (invalidated) { 76189cb0888SJan Kara rc = -EAGAIN; 76289cb0888SJan Kara goto unlock_buffers; 76389cb0888SJan Kara } 764ebdf4de5SJan Kara spin_unlock(&mapping->private_lock); 76589cb0888SJan Kara invalidate_bh_lrus(); 76689cb0888SJan Kara invalidated = true; 76789cb0888SJan Kara goto recheck_buffers; 76889cb0888SJan Kara } 76989cb0888SJan Kara } 77089cb0888SJan Kara 77167235182SMatthew Wilcox (Oracle) rc = folio_migrate_mapping(mapping, dst, src, 0); 77278bd5209SRafael Aquini if (rc != MIGRATEPAGE_SUCCESS) 773cc4f11e6SJan Kara goto unlock_buffers; 7741d8b85ccSChristoph Lameter 77567235182SMatthew Wilcox (Oracle) folio_attach_private(dst, folio_detach_private(src)); 7761d8b85ccSChristoph Lameter 7771d8b85ccSChristoph Lameter bh = head; 7781d8b85ccSChristoph Lameter do { 779d5db4f9dSMatthew Wilcox (Oracle) folio_set_bh(bh, dst, bh_offset(bh)); 7801d8b85ccSChristoph Lameter bh = bh->b_this_page; 7811d8b85ccSChristoph Lameter } while (bh != head); 7821d8b85ccSChristoph Lameter 7832916ecc0SJérôme Glisse if (mode != MIGRATE_SYNC_NO_COPY) 78467235182SMatthew Wilcox (Oracle) folio_migrate_copy(dst, src); 7852916ecc0SJérôme Glisse else 78667235182SMatthew Wilcox (Oracle) folio_migrate_flags(dst, src); 7871d8b85ccSChristoph Lameter 788cc4f11e6SJan Kara rc = MIGRATEPAGE_SUCCESS; 789cc4f11e6SJan Kara unlock_buffers: 790ebdf4de5SJan Kara if (check_refs) 791ebdf4de5SJan Kara spin_unlock(&mapping->private_lock); 7921d8b85ccSChristoph Lameter bh = head; 7931d8b85ccSChristoph Lameter do { 7941d8b85ccSChristoph Lameter unlock_buffer(bh); 7951d8b85ccSChristoph Lameter bh = bh->b_this_page; 7961d8b85ccSChristoph Lameter } while (bh != head); 7971d8b85ccSChristoph Lameter 798cc4f11e6SJan Kara return rc; 7991d8b85ccSChristoph Lameter } 80089cb0888SJan Kara 80167235182SMatthew Wilcox (Oracle) /** 80267235182SMatthew Wilcox (Oracle) * buffer_migrate_folio() - Migration function for folios with buffers. 80367235182SMatthew Wilcox (Oracle) * @mapping: The address space containing @src. 80467235182SMatthew Wilcox (Oracle) * @dst: The folio to migrate to. 80567235182SMatthew Wilcox (Oracle) * @src: The folio to migrate from. 80667235182SMatthew Wilcox (Oracle) * @mode: How to migrate the folio. 80767235182SMatthew Wilcox (Oracle) * 80867235182SMatthew Wilcox (Oracle) * This function can only be used if the underlying filesystem guarantees 80967235182SMatthew Wilcox (Oracle) * that no other references to @src exist. For example attached buffer 81067235182SMatthew Wilcox (Oracle) * heads are accessed only under the folio lock. If your filesystem cannot 81167235182SMatthew Wilcox (Oracle) * provide this guarantee, buffer_migrate_folio_norefs() may be more 81267235182SMatthew Wilcox (Oracle) * appropriate. 81367235182SMatthew Wilcox (Oracle) * 81467235182SMatthew Wilcox (Oracle) * Return: 0 on success or a negative errno on failure. 81589cb0888SJan Kara */ 81667235182SMatthew Wilcox (Oracle) int buffer_migrate_folio(struct address_space *mapping, 81767235182SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode) 81889cb0888SJan Kara { 81967235182SMatthew Wilcox (Oracle) return __buffer_migrate_folio(mapping, dst, src, mode, false); 82089cb0888SJan Kara } 82167235182SMatthew Wilcox (Oracle) EXPORT_SYMBOL(buffer_migrate_folio); 82289cb0888SJan Kara 82367235182SMatthew Wilcox (Oracle) /** 82467235182SMatthew Wilcox (Oracle) * buffer_migrate_folio_norefs() - Migration function for folios with buffers. 82567235182SMatthew Wilcox (Oracle) * @mapping: The address space containing @src. 82667235182SMatthew Wilcox (Oracle) * @dst: The folio to migrate to. 82767235182SMatthew Wilcox (Oracle) * @src: The folio to migrate from. 82867235182SMatthew Wilcox (Oracle) * @mode: How to migrate the folio. 82967235182SMatthew Wilcox (Oracle) * 83067235182SMatthew Wilcox (Oracle) * Like buffer_migrate_folio() except that this variant is more careful 83167235182SMatthew Wilcox (Oracle) * and checks that there are also no buffer head references. This function 83267235182SMatthew Wilcox (Oracle) * is the right one for mappings where buffer heads are directly looked 83367235182SMatthew Wilcox (Oracle) * up and referenced (such as block device mappings). 83467235182SMatthew Wilcox (Oracle) * 83567235182SMatthew Wilcox (Oracle) * Return: 0 on success or a negative errno on failure. 83689cb0888SJan Kara */ 83767235182SMatthew Wilcox (Oracle) int buffer_migrate_folio_norefs(struct address_space *mapping, 83867235182SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode) 83989cb0888SJan Kara { 84067235182SMatthew Wilcox (Oracle) return __buffer_migrate_folio(mapping, dst, src, mode, true); 84189cb0888SJan Kara } 842e26355e2SJan Kara EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs); 843925c86a1SChristoph Hellwig #endif /* CONFIG_BUFFER_HEAD */ 8441d8b85ccSChristoph Lameter 8452ec810d5SMatthew Wilcox (Oracle) int filemap_migrate_folio(struct address_space *mapping, 8462ec810d5SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode) 8472ec810d5SMatthew Wilcox (Oracle) { 8482ec810d5SMatthew Wilcox (Oracle) int ret; 8492ec810d5SMatthew Wilcox (Oracle) 8502ec810d5SMatthew Wilcox (Oracle) ret = folio_migrate_mapping(mapping, dst, src, 0); 8512ec810d5SMatthew Wilcox (Oracle) if (ret != MIGRATEPAGE_SUCCESS) 8522ec810d5SMatthew Wilcox (Oracle) return ret; 8532ec810d5SMatthew Wilcox (Oracle) 8542ec810d5SMatthew Wilcox (Oracle) if (folio_get_private(src)) 8552ec810d5SMatthew Wilcox (Oracle) folio_attach_private(dst, folio_detach_private(src)); 8562ec810d5SMatthew Wilcox (Oracle) 8572ec810d5SMatthew Wilcox (Oracle) if (mode != MIGRATE_SYNC_NO_COPY) 8582ec810d5SMatthew Wilcox (Oracle) folio_migrate_copy(dst, src); 8592ec810d5SMatthew Wilcox (Oracle) else 8602ec810d5SMatthew Wilcox (Oracle) folio_migrate_flags(dst, src); 8612ec810d5SMatthew Wilcox (Oracle) return MIGRATEPAGE_SUCCESS; 8622ec810d5SMatthew Wilcox (Oracle) } 8632ec810d5SMatthew Wilcox (Oracle) EXPORT_SYMBOL_GPL(filemap_migrate_folio); 8642ec810d5SMatthew Wilcox (Oracle) 86504e62a29SChristoph Lameter /* 8662be7fa10SMatthew Wilcox (Oracle) * Writeback a folio to clean the dirty state 86704e62a29SChristoph Lameter */ 8682be7fa10SMatthew Wilcox (Oracle) static int writeout(struct address_space *mapping, struct folio *folio) 86904e62a29SChristoph Lameter { 87004e62a29SChristoph Lameter struct writeback_control wbc = { 87104e62a29SChristoph Lameter .sync_mode = WB_SYNC_NONE, 87204e62a29SChristoph Lameter .nr_to_write = 1, 87304e62a29SChristoph Lameter .range_start = 0, 87404e62a29SChristoph Lameter .range_end = LLONG_MAX, 87504e62a29SChristoph Lameter .for_reclaim = 1 87604e62a29SChristoph Lameter }; 87704e62a29SChristoph Lameter int rc; 87804e62a29SChristoph Lameter 87904e62a29SChristoph Lameter if (!mapping->a_ops->writepage) 88004e62a29SChristoph Lameter /* No write method for the address space */ 88104e62a29SChristoph Lameter return -EINVAL; 88204e62a29SChristoph Lameter 8832be7fa10SMatthew Wilcox (Oracle) if (!folio_clear_dirty_for_io(folio)) 88404e62a29SChristoph Lameter /* Someone else already triggered a write */ 88504e62a29SChristoph Lameter return -EAGAIN; 88604e62a29SChristoph Lameter 88704e62a29SChristoph Lameter /* 8882be7fa10SMatthew Wilcox (Oracle) * A dirty folio may imply that the underlying filesystem has 8892be7fa10SMatthew Wilcox (Oracle) * the folio on some queue. So the folio must be clean for 8902be7fa10SMatthew Wilcox (Oracle) * migration. Writeout may mean we lose the lock and the 8912be7fa10SMatthew Wilcox (Oracle) * folio state is no longer what we checked for earlier. 89204e62a29SChristoph Lameter * At this point we know that the migration attempt cannot 89304e62a29SChristoph Lameter * be successful. 89404e62a29SChristoph Lameter */ 8954eecb8b9SMatthew Wilcox (Oracle) remove_migration_ptes(folio, folio, false); 89604e62a29SChristoph Lameter 8972be7fa10SMatthew Wilcox (Oracle) rc = mapping->a_ops->writepage(&folio->page, &wbc); 89804e62a29SChristoph Lameter 89904e62a29SChristoph Lameter if (rc != AOP_WRITEPAGE_ACTIVATE) 90004e62a29SChristoph Lameter /* unlocked. Relock */ 9012be7fa10SMatthew Wilcox (Oracle) folio_lock(folio); 90204e62a29SChristoph Lameter 903bda8550dSHugh Dickins return (rc < 0) ? -EIO : -EAGAIN; 90404e62a29SChristoph Lameter } 90504e62a29SChristoph Lameter 90604e62a29SChristoph Lameter /* 90704e62a29SChristoph Lameter * Default handling if a filesystem does not provide a migration function. 90804e62a29SChristoph Lameter */ 9098faa8ef5SMatthew Wilcox (Oracle) static int fallback_migrate_folio(struct address_space *mapping, 9108faa8ef5SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode) 9118351a6e4SChristoph Lameter { 9128faa8ef5SMatthew Wilcox (Oracle) if (folio_test_dirty(src)) { 9138faa8ef5SMatthew Wilcox (Oracle) /* Only writeback folios in full synchronous migration */ 9142916ecc0SJérôme Glisse switch (mode) { 9152916ecc0SJérôme Glisse case MIGRATE_SYNC: 9162916ecc0SJérôme Glisse case MIGRATE_SYNC_NO_COPY: 9172916ecc0SJérôme Glisse break; 9182916ecc0SJérôme Glisse default: 919b969c4abSMel Gorman return -EBUSY; 9202916ecc0SJérôme Glisse } 9212be7fa10SMatthew Wilcox (Oracle) return writeout(mapping, src); 922b969c4abSMel Gorman } 9238351a6e4SChristoph Lameter 9248351a6e4SChristoph Lameter /* 9258351a6e4SChristoph Lameter * Buffers may be managed in a filesystem specific way. 9268351a6e4SChristoph Lameter * We must have no buffers or drop them. 9278351a6e4SChristoph Lameter */ 9280201ebf2SDavid Howells if (!filemap_release_folio(src, GFP_KERNEL)) 929806031bbSMel Gorman return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; 9308351a6e4SChristoph Lameter 93154184650SMatthew Wilcox (Oracle) return migrate_folio(mapping, dst, src, mode); 9328351a6e4SChristoph Lameter } 9338351a6e4SChristoph Lameter 9341d8b85ccSChristoph Lameter /* 935e24f0b8fSChristoph Lameter * Move a page to a newly allocated page 936e24f0b8fSChristoph Lameter * The page is locked and all ptes have been successfully removed. 937b20a3503SChristoph Lameter * 938e24f0b8fSChristoph Lameter * The new page will have replaced the old page if this function 939e24f0b8fSChristoph Lameter * is successful. 940894bc310SLee Schermerhorn * 941894bc310SLee Schermerhorn * Return value: 942894bc310SLee Schermerhorn * < 0 - error code 94378bd5209SRafael Aquini * MIGRATEPAGE_SUCCESS - success 944b20a3503SChristoph Lameter */ 945e7e3ffebSMatthew Wilcox (Oracle) static int move_to_new_folio(struct folio *dst, struct folio *src, 9465c3f9a67SHugh Dickins enum migrate_mode mode) 947b20a3503SChristoph Lameter { 948bda807d4SMinchan Kim int rc = -EAGAIN; 949e7e3ffebSMatthew Wilcox (Oracle) bool is_lru = !__PageMovable(&src->page); 950b20a3503SChristoph Lameter 951e7e3ffebSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(src), src); 952e7e3ffebSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst); 953b20a3503SChristoph Lameter 954bda807d4SMinchan Kim if (likely(is_lru)) { 95568f2736aSMatthew Wilcox (Oracle) struct address_space *mapping = folio_mapping(src); 95668f2736aSMatthew Wilcox (Oracle) 957b20a3503SChristoph Lameter if (!mapping) 95854184650SMatthew Wilcox (Oracle) rc = migrate_folio(mapping, dst, src, mode); 9595490da4fSMatthew Wilcox (Oracle) else if (mapping->a_ops->migrate_folio) 960b20a3503SChristoph Lameter /* 9615490da4fSMatthew Wilcox (Oracle) * Most folios have a mapping and most filesystems 9625490da4fSMatthew Wilcox (Oracle) * provide a migrate_folio callback. Anonymous folios 963bda807d4SMinchan Kim * are part of swap space which also has its own 9645490da4fSMatthew Wilcox (Oracle) * migrate_folio callback. This is the most common path 965bda807d4SMinchan Kim * for page migration. 966b20a3503SChristoph Lameter */ 9675490da4fSMatthew Wilcox (Oracle) rc = mapping->a_ops->migrate_folio(mapping, dst, src, 9685490da4fSMatthew Wilcox (Oracle) mode); 9698351a6e4SChristoph Lameter else 9708faa8ef5SMatthew Wilcox (Oracle) rc = fallback_migrate_folio(mapping, dst, src, mode); 971bda807d4SMinchan Kim } else { 97268f2736aSMatthew Wilcox (Oracle) const struct movable_operations *mops; 97368f2736aSMatthew Wilcox (Oracle) 974bda807d4SMinchan Kim /* 975bda807d4SMinchan Kim * In case of non-lru page, it could be released after 976bda807d4SMinchan Kim * isolation step. In that case, we shouldn't try migration. 977bda807d4SMinchan Kim */ 978e7e3ffebSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_isolated(src), src); 979e7e3ffebSMatthew Wilcox (Oracle) if (!folio_test_movable(src)) { 980bda807d4SMinchan Kim rc = MIGRATEPAGE_SUCCESS; 981e7e3ffebSMatthew Wilcox (Oracle) folio_clear_isolated(src); 982bda807d4SMinchan Kim goto out; 983bda807d4SMinchan Kim } 984bda807d4SMinchan Kim 985da707a6dSVishal Moola (Oracle) mops = folio_movable_ops(src); 98668f2736aSMatthew Wilcox (Oracle) rc = mops->migrate_page(&dst->page, &src->page, mode); 987bda807d4SMinchan Kim WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && 988e7e3ffebSMatthew Wilcox (Oracle) !folio_test_isolated(src)); 989bda807d4SMinchan Kim } 990b20a3503SChristoph Lameter 9915c3f9a67SHugh Dickins /* 992e7e3ffebSMatthew Wilcox (Oracle) * When successful, old pagecache src->mapping must be cleared before 993e7e3ffebSMatthew Wilcox (Oracle) * src is freed; but stats require that PageAnon be left as PageAnon. 9945c3f9a67SHugh Dickins */ 9955c3f9a67SHugh Dickins if (rc == MIGRATEPAGE_SUCCESS) { 996e7e3ffebSMatthew Wilcox (Oracle) if (__PageMovable(&src->page)) { 997e7e3ffebSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_isolated(src), src); 998bda807d4SMinchan Kim 999bda807d4SMinchan Kim /* 1000bda807d4SMinchan Kim * We clear PG_movable under page_lock so any compactor 1001bda807d4SMinchan Kim * cannot try to migrate this page. 1002bda807d4SMinchan Kim */ 1003e7e3ffebSMatthew Wilcox (Oracle) folio_clear_isolated(src); 1004bda807d4SMinchan Kim } 1005bda807d4SMinchan Kim 1006bda807d4SMinchan Kim /* 1007e7e3ffebSMatthew Wilcox (Oracle) * Anonymous and movable src->mapping will be cleared by 1008bda807d4SMinchan Kim * free_pages_prepare so don't reset it here for keeping 1009bda807d4SMinchan Kim * the type to work PageAnon, for example. 1010bda807d4SMinchan Kim */ 1011e7e3ffebSMatthew Wilcox (Oracle) if (!folio_mapping_flags(src)) 1012e7e3ffebSMatthew Wilcox (Oracle) src->mapping = NULL; 1013d2b2c6ddSLars Persson 1014e7e3ffebSMatthew Wilcox (Oracle) if (likely(!folio_is_zone_device(dst))) 1015e7e3ffebSMatthew Wilcox (Oracle) flush_dcache_folio(dst); 10163fe2011fSMel Gorman } 1017bda807d4SMinchan Kim out: 1018e24f0b8fSChristoph Lameter return rc; 1019e24f0b8fSChristoph Lameter } 1020e24f0b8fSChristoph Lameter 102164c8902eSHuang Ying /* 102264c8902eSHuang Ying * To record some information during migration, we use some unused 102364c8902eSHuang Ying * fields (mapping and private) of struct folio of the newly allocated 102464c8902eSHuang Ying * destination folio. This is safe because nobody is using them 102564c8902eSHuang Ying * except us. 102664c8902eSHuang Ying */ 1027e77d587aSLinus Torvalds union migration_ptr { 1028e77d587aSLinus Torvalds struct anon_vma *anon_vma; 1029e77d587aSLinus Torvalds struct address_space *mapping; 1030e77d587aSLinus Torvalds }; 103164c8902eSHuang Ying static void __migrate_folio_record(struct folio *dst, 103264c8902eSHuang Ying unsigned long page_was_mapped, 103364c8902eSHuang Ying struct anon_vma *anon_vma) 1034e24f0b8fSChristoph Lameter { 1035e77d587aSLinus Torvalds union migration_ptr ptr = { .anon_vma = anon_vma }; 1036e77d587aSLinus Torvalds dst->mapping = ptr.mapping; 103764c8902eSHuang Ying dst->private = (void *)page_was_mapped; 103864c8902eSHuang Ying } 103964c8902eSHuang Ying 104064c8902eSHuang Ying static void __migrate_folio_extract(struct folio *dst, 104164c8902eSHuang Ying int *page_was_mappedp, 104264c8902eSHuang Ying struct anon_vma **anon_vmap) 104364c8902eSHuang Ying { 1044e77d587aSLinus Torvalds union migration_ptr ptr = { .mapping = dst->mapping }; 1045e77d587aSLinus Torvalds *anon_vmap = ptr.anon_vma; 104664c8902eSHuang Ying *page_was_mappedp = (unsigned long)dst->private; 104764c8902eSHuang Ying dst->mapping = NULL; 104864c8902eSHuang Ying dst->private = NULL; 104964c8902eSHuang Ying } 105064c8902eSHuang Ying 10515dfab109SHuang Ying /* Restore the source folio to the original state upon failure */ 10525dfab109SHuang Ying static void migrate_folio_undo_src(struct folio *src, 10535dfab109SHuang Ying int page_was_mapped, 10545dfab109SHuang Ying struct anon_vma *anon_vma, 1055ebe75e47SHuang Ying bool locked, 10565dfab109SHuang Ying struct list_head *ret) 10575dfab109SHuang Ying { 10585dfab109SHuang Ying if (page_was_mapped) 10595dfab109SHuang Ying remove_migration_ptes(src, src, false); 10605dfab109SHuang Ying /* Drop an anon_vma reference if we took one */ 10615dfab109SHuang Ying if (anon_vma) 10625dfab109SHuang Ying put_anon_vma(anon_vma); 1063ebe75e47SHuang Ying if (locked) 10645dfab109SHuang Ying folio_unlock(src); 1065ebe75e47SHuang Ying if (ret) 10665dfab109SHuang Ying list_move_tail(&src->lru, ret); 10675dfab109SHuang Ying } 10685dfab109SHuang Ying 10695dfab109SHuang Ying /* Restore the destination folio to the original state upon failure */ 10704e096ae1SMatthew Wilcox (Oracle) static void migrate_folio_undo_dst(struct folio *dst, bool locked, 10714e096ae1SMatthew Wilcox (Oracle) free_folio_t put_new_folio, unsigned long private) 10725dfab109SHuang Ying { 1073ebe75e47SHuang Ying if (locked) 10745dfab109SHuang Ying folio_unlock(dst); 10754e096ae1SMatthew Wilcox (Oracle) if (put_new_folio) 10764e096ae1SMatthew Wilcox (Oracle) put_new_folio(dst, private); 10775dfab109SHuang Ying else 10785dfab109SHuang Ying folio_put(dst); 10795dfab109SHuang Ying } 10805dfab109SHuang Ying 108164c8902eSHuang Ying /* Cleanup src folio upon migration success */ 108264c8902eSHuang Ying static void migrate_folio_done(struct folio *src, 108364c8902eSHuang Ying enum migrate_reason reason) 108464c8902eSHuang Ying { 108564c8902eSHuang Ying /* 108664c8902eSHuang Ying * Compaction can migrate also non-LRU pages which are 108764c8902eSHuang Ying * not accounted to NR_ISOLATED_*. They can be recognized 108864c8902eSHuang Ying * as __PageMovable 108964c8902eSHuang Ying */ 109064c8902eSHuang Ying if (likely(!__folio_test_movable(src))) 109164c8902eSHuang Ying mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON + 109264c8902eSHuang Ying folio_is_file_lru(src), -folio_nr_pages(src)); 109364c8902eSHuang Ying 109464c8902eSHuang Ying if (reason != MR_MEMORY_FAILURE) 109564c8902eSHuang Ying /* We release the page in page_handle_poison. */ 109664c8902eSHuang Ying folio_put(src); 109764c8902eSHuang Ying } 109864c8902eSHuang Ying 1099ebe75e47SHuang Ying /* Obtain the lock on page, remove all ptes. */ 11004e096ae1SMatthew Wilcox (Oracle) static int migrate_folio_unmap(new_folio_t get_new_folio, 11014e096ae1SMatthew Wilcox (Oracle) free_folio_t put_new_folio, unsigned long private, 11024e096ae1SMatthew Wilcox (Oracle) struct folio *src, struct folio **dstp, enum migrate_mode mode, 11032ef7dbb2SHuang Ying enum migrate_reason reason, struct list_head *ret) 1104e24f0b8fSChristoph Lameter { 1105ebe75e47SHuang Ying struct folio *dst; 11060dabec93SMinchan Kim int rc = -EAGAIN; 110764c8902eSHuang Ying int page_was_mapped = 0; 11083f6c8272SMel Gorman struct anon_vma *anon_vma = NULL; 1109682a71a1SMatthew Wilcox (Oracle) bool is_lru = !__PageMovable(&src->page); 1110ebe75e47SHuang Ying bool locked = false; 1111ebe75e47SHuang Ying bool dst_locked = false; 1112ebe75e47SHuang Ying 1113ebe75e47SHuang Ying if (folio_ref_count(src) == 1) { 1114ebe75e47SHuang Ying /* Folio was freed from under us. So we are done. */ 1115ebe75e47SHuang Ying folio_clear_active(src); 1116ebe75e47SHuang Ying folio_clear_unevictable(src); 1117ebe75e47SHuang Ying /* free_pages_prepare() will clear PG_isolated. */ 1118ebe75e47SHuang Ying list_del(&src->lru); 1119ebe75e47SHuang Ying migrate_folio_done(src, reason); 1120ebe75e47SHuang Ying return MIGRATEPAGE_SUCCESS; 1121ebe75e47SHuang Ying } 1122ebe75e47SHuang Ying 11234e096ae1SMatthew Wilcox (Oracle) dst = get_new_folio(src, private); 11244e096ae1SMatthew Wilcox (Oracle) if (!dst) 1125ebe75e47SHuang Ying return -ENOMEM; 1126ebe75e47SHuang Ying *dstp = dst; 1127ebe75e47SHuang Ying 1128ebe75e47SHuang Ying dst->private = NULL; 112995a402c3SChristoph Lameter 1130682a71a1SMatthew Wilcox (Oracle) if (!folio_trylock(src)) { 11312ef7dbb2SHuang Ying if (mode == MIGRATE_ASYNC) 11320dabec93SMinchan Kim goto out; 11333e7d3449SMel Gorman 11343e7d3449SMel Gorman /* 11353e7d3449SMel Gorman * It's not safe for direct compaction to call lock_page. 11363e7d3449SMel Gorman * For example, during page readahead pages are added locked 11373e7d3449SMel Gorman * to the LRU. Later, when the IO completes the pages are 11383e7d3449SMel Gorman * marked uptodate and unlocked. However, the queueing 11393e7d3449SMel Gorman * could be merging multiple pages for one bio (e.g. 1140d4388340SMatthew Wilcox (Oracle) * mpage_readahead). If an allocation happens for the 11413e7d3449SMel Gorman * second or third page, the process can end up locking 11423e7d3449SMel Gorman * the same page twice and deadlocking. Rather than 11433e7d3449SMel Gorman * trying to be clever about what pages can be locked, 11443e7d3449SMel Gorman * avoid the use of lock_page for direct compaction 11453e7d3449SMel Gorman * altogether. 11463e7d3449SMel Gorman */ 11473e7d3449SMel Gorman if (current->flags & PF_MEMALLOC) 11480dabec93SMinchan Kim goto out; 11493e7d3449SMel Gorman 11504bb6dc79SDouglas Anderson /* 11514bb6dc79SDouglas Anderson * In "light" mode, we can wait for transient locks (eg 11524bb6dc79SDouglas Anderson * inserting a page into the page table), but it's not 11534bb6dc79SDouglas Anderson * worth waiting for I/O. 11544bb6dc79SDouglas Anderson */ 11554bb6dc79SDouglas Anderson if (mode == MIGRATE_SYNC_LIGHT && !folio_test_uptodate(src)) 11564bb6dc79SDouglas Anderson goto out; 11574bb6dc79SDouglas Anderson 1158682a71a1SMatthew Wilcox (Oracle) folio_lock(src); 1159e24f0b8fSChristoph Lameter } 1160ebe75e47SHuang Ying locked = true; 1161e24f0b8fSChristoph Lameter 1162682a71a1SMatthew Wilcox (Oracle) if (folio_test_writeback(src)) { 116311bc82d6SAndrea Arcangeli /* 1164fed5b64aSJianguo Wu * Only in the case of a full synchronous migration is it 1165a6bc32b8SMel Gorman * necessary to wait for PageWriteback. In the async case, 1166a6bc32b8SMel Gorman * the retry loop is too short and in the sync-light case, 1167a6bc32b8SMel Gorman * the overhead of stalling is too much 116811bc82d6SAndrea Arcangeli */ 11692916ecc0SJérôme Glisse switch (mode) { 11702916ecc0SJérôme Glisse case MIGRATE_SYNC: 11712916ecc0SJérôme Glisse case MIGRATE_SYNC_NO_COPY: 11722916ecc0SJérôme Glisse break; 11732916ecc0SJérôme Glisse default: 117411bc82d6SAndrea Arcangeli rc = -EBUSY; 1175ebe75e47SHuang Ying goto out; 117611bc82d6SAndrea Arcangeli } 1177682a71a1SMatthew Wilcox (Oracle) folio_wait_writeback(src); 1178e24f0b8fSChristoph Lameter } 117903f15c86SHugh Dickins 1180e24f0b8fSChristoph Lameter /* 1181682a71a1SMatthew Wilcox (Oracle) * By try_to_migrate(), src->mapcount goes down to 0 here. In this case, 1182682a71a1SMatthew Wilcox (Oracle) * we cannot notice that anon_vma is freed while we migrate a page. 11831ce82b69SHugh Dickins * This get_anon_vma() delays freeing anon_vma pointer until the end 1184dc386d4dSKAMEZAWA Hiroyuki * of migration. File cache pages are no problem because of page_lock() 1185989f89c5SKAMEZAWA Hiroyuki * File Caches may use write_page() or lock_page() in migration, then, 1186989f89c5SKAMEZAWA Hiroyuki * just care Anon page here. 11873fe2011fSMel Gorman * 118829eea9b5SMatthew Wilcox (Oracle) * Only folio_get_anon_vma() understands the subtleties of 118903f15c86SHugh Dickins * getting a hold on an anon_vma from outside one of its mms. 119003f15c86SHugh Dickins * But if we cannot get anon_vma, then we won't need it anyway, 119103f15c86SHugh Dickins * because that implies that the anon page is no longer mapped 119203f15c86SHugh Dickins * (and cannot be remapped so long as we hold the page lock). 11933fe2011fSMel Gorman */ 1194682a71a1SMatthew Wilcox (Oracle) if (folio_test_anon(src) && !folio_test_ksm(src)) 119529eea9b5SMatthew Wilcox (Oracle) anon_vma = folio_get_anon_vma(src); 119662e1c553SShaohua Li 11977db7671fSHugh Dickins /* 11987db7671fSHugh Dickins * Block others from accessing the new page when we get around to 11997db7671fSHugh Dickins * establishing additional references. We are usually the only one 1200682a71a1SMatthew Wilcox (Oracle) * holding a reference to dst at this point. We used to have a BUG 1201682a71a1SMatthew Wilcox (Oracle) * here if folio_trylock(dst) fails, but would like to allow for 1202682a71a1SMatthew Wilcox (Oracle) * cases where there might be a race with the previous use of dst. 12037db7671fSHugh Dickins * This is much like races on refcount of oldpage: just don't BUG(). 12047db7671fSHugh Dickins */ 1205682a71a1SMatthew Wilcox (Oracle) if (unlikely(!folio_trylock(dst))) 1206ebe75e47SHuang Ying goto out; 1207ebe75e47SHuang Ying dst_locked = true; 12087db7671fSHugh Dickins 1209bda807d4SMinchan Kim if (unlikely(!is_lru)) { 121064c8902eSHuang Ying __migrate_folio_record(dst, page_was_mapped, anon_vma); 121164c8902eSHuang Ying return MIGRATEPAGE_UNMAP; 1212bda807d4SMinchan Kim } 1213bda807d4SMinchan Kim 1214dc386d4dSKAMEZAWA Hiroyuki /* 121562e1c553SShaohua Li * Corner case handling: 121662e1c553SShaohua Li * 1. When a new swap-cache page is read into, it is added to the LRU 121762e1c553SShaohua Li * and treated as swapcache but it has no rmap yet. 1218682a71a1SMatthew Wilcox (Oracle) * Calling try_to_unmap() against a src->mapping==NULL page will 121962e1c553SShaohua Li * trigger a BUG. So handle it here. 1220d12b8951SYang Shi * 2. An orphaned page (see truncate_cleanup_page) might have 122162e1c553SShaohua Li * fs-private metadata. The page can be picked up due to memory 122262e1c553SShaohua Li * offlining. Everywhere else except page reclaim, the page is 122362e1c553SShaohua Li * invisible to the vm, so the page can not be migrated. So try to 122462e1c553SShaohua Li * free the metadata, so the page can be freed. 1225dc386d4dSKAMEZAWA Hiroyuki */ 1226682a71a1SMatthew Wilcox (Oracle) if (!src->mapping) { 1227682a71a1SMatthew Wilcox (Oracle) if (folio_test_private(src)) { 1228682a71a1SMatthew Wilcox (Oracle) try_to_free_buffers(src); 1229ebe75e47SHuang Ying goto out; 123062e1c553SShaohua Li } 1231682a71a1SMatthew Wilcox (Oracle) } else if (folio_mapped(src)) { 12327db7671fSHugh Dickins /* Establish migration ptes */ 1233682a71a1SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_test_anon(src) && 1234682a71a1SMatthew Wilcox (Oracle) !folio_test_ksm(src) && !anon_vma, src); 1235fb3592c4SHuang Ying try_to_migrate(src, mode == MIGRATE_ASYNC ? TTU_BATCH_FLUSH : 0); 123664c8902eSHuang Ying page_was_mapped = 1; 12372ebba6b7SHugh Dickins } 1238dc386d4dSKAMEZAWA Hiroyuki 123964c8902eSHuang Ying if (!folio_mapped(src)) { 124064c8902eSHuang Ying __migrate_folio_record(dst, page_was_mapped, anon_vma); 124164c8902eSHuang Ying return MIGRATEPAGE_UNMAP; 124264c8902eSHuang Ying } 124364c8902eSHuang Ying 124464c8902eSHuang Ying out: 124580562ba0SHuang Ying /* 124680562ba0SHuang Ying * A folio that has not been unmapped will be restored to 124780562ba0SHuang Ying * right list unless we want to retry. 124880562ba0SHuang Ying */ 1249fb3592c4SHuang Ying if (rc == -EAGAIN) 1250ebe75e47SHuang Ying ret = NULL; 125180562ba0SHuang Ying 1252ebe75e47SHuang Ying migrate_folio_undo_src(src, page_was_mapped, anon_vma, locked, ret); 12534e096ae1SMatthew Wilcox (Oracle) migrate_folio_undo_dst(dst, dst_locked, put_new_folio, private); 125480562ba0SHuang Ying 125580562ba0SHuang Ying return rc; 125680562ba0SHuang Ying } 125780562ba0SHuang Ying 1258ebe75e47SHuang Ying /* Migrate the folio to the newly allocated folio in dst. */ 12594e096ae1SMatthew Wilcox (Oracle) static int migrate_folio_move(free_folio_t put_new_folio, unsigned long private, 1260ebe75e47SHuang Ying struct folio *src, struct folio *dst, 1261ebe75e47SHuang Ying enum migrate_mode mode, enum migrate_reason reason, 1262ebe75e47SHuang Ying struct list_head *ret) 126364c8902eSHuang Ying { 126464c8902eSHuang Ying int rc; 126564c8902eSHuang Ying int page_was_mapped = 0; 126664c8902eSHuang Ying struct anon_vma *anon_vma = NULL; 126764c8902eSHuang Ying bool is_lru = !__PageMovable(&src->page); 12685dfab109SHuang Ying struct list_head *prev; 126964c8902eSHuang Ying 127064c8902eSHuang Ying __migrate_folio_extract(dst, &page_was_mapped, &anon_vma); 12715dfab109SHuang Ying prev = dst->lru.prev; 12725dfab109SHuang Ying list_del(&dst->lru); 127364c8902eSHuang Ying 1274682a71a1SMatthew Wilcox (Oracle) rc = move_to_new_folio(dst, src, mode); 1275ebe75e47SHuang Ying if (rc) 1276ebe75e47SHuang Ying goto out; 12775dfab109SHuang Ying 127864c8902eSHuang Ying if (unlikely(!is_lru)) 127964c8902eSHuang Ying goto out_unlock_both; 1280e24f0b8fSChristoph Lameter 1281c3096e67SHugh Dickins /* 1282682a71a1SMatthew Wilcox (Oracle) * When successful, push dst to LRU immediately: so that if it 1283c3096e67SHugh Dickins * turns out to be an mlocked page, remove_migration_ptes() will 1284682a71a1SMatthew Wilcox (Oracle) * automatically build up the correct dst->mlock_count for it. 1285c3096e67SHugh Dickins * 1286c3096e67SHugh Dickins * We would like to do something similar for the old page, when 1287c3096e67SHugh Dickins * unsuccessful, and other cases when a page has been temporarily 1288c3096e67SHugh Dickins * isolated from the unevictable LRU: but this case is the easiest. 1289c3096e67SHugh Dickins */ 1290682a71a1SMatthew Wilcox (Oracle) folio_add_lru(dst); 12915c3f9a67SHugh Dickins if (page_was_mapped) 1292c3096e67SHugh Dickins lru_add_drain(); 1293c3096e67SHugh Dickins 12945c3f9a67SHugh Dickins if (page_was_mapped) 1295ebe75e47SHuang Ying remove_migration_ptes(src, dst, false); 12963f6c8272SMel Gorman 12977db7671fSHugh Dickins out_unlock_both: 1298682a71a1SMatthew Wilcox (Oracle) folio_unlock(dst); 1299ebe75e47SHuang Ying set_page_owner_migrate_reason(&dst->page, reason); 1300c6c919ebSMinchan Kim /* 1301682a71a1SMatthew Wilcox (Oracle) * If migration is successful, decrease refcount of dst, 1302c6c919ebSMinchan Kim * which will not free the page because new page owner increased 1303c3096e67SHugh Dickins * refcounter. 1304c6c919ebSMinchan Kim */ 1305682a71a1SMatthew Wilcox (Oracle) folio_put(dst); 1306c6c919ebSMinchan Kim 1307ebe75e47SHuang Ying /* 1308ebe75e47SHuang Ying * A folio that has been migrated has all references removed 1309ebe75e47SHuang Ying * and will be freed. 1310ebe75e47SHuang Ying */ 1311ebe75e47SHuang Ying list_del(&src->lru); 1312ebe75e47SHuang Ying /* Drop an anon_vma reference if we took one */ 1313ebe75e47SHuang Ying if (anon_vma) 1314ebe75e47SHuang Ying put_anon_vma(anon_vma); 1315ebe75e47SHuang Ying folio_unlock(src); 1316ebe75e47SHuang Ying migrate_folio_done(src, reason); 1317ebe75e47SHuang Ying 1318ebe75e47SHuang Ying return rc; 1319ebe75e47SHuang Ying out: 1320ebe75e47SHuang Ying /* 1321ebe75e47SHuang Ying * A folio that has not been migrated will be restored to 1322ebe75e47SHuang Ying * right list unless we want to retry. 1323ebe75e47SHuang Ying */ 1324ebe75e47SHuang Ying if (rc == -EAGAIN) { 1325ebe75e47SHuang Ying list_add(&dst->lru, prev); 1326ebe75e47SHuang Ying __migrate_folio_record(dst, page_was_mapped, anon_vma); 13270dabec93SMinchan Kim return rc; 13280dabec93SMinchan Kim } 132995a402c3SChristoph Lameter 1330ebe75e47SHuang Ying migrate_folio_undo_src(src, page_was_mapped, anon_vma, true, ret); 13314e096ae1SMatthew Wilcox (Oracle) migrate_folio_undo_dst(dst, true, put_new_folio, private); 133268711a74SDavid Rientjes 1333e24f0b8fSChristoph Lameter return rc; 1334e24f0b8fSChristoph Lameter } 1335b20a3503SChristoph Lameter 1336e24f0b8fSChristoph Lameter /* 1337290408d4SNaoya Horiguchi * Counterpart of unmap_and_move_page() for hugepage migration. 1338290408d4SNaoya Horiguchi * 1339290408d4SNaoya Horiguchi * This function doesn't wait the completion of hugepage I/O 1340290408d4SNaoya Horiguchi * because there is no race between I/O and migration for hugepage. 1341290408d4SNaoya Horiguchi * Note that currently hugepage I/O occurs only in direct I/O 1342290408d4SNaoya Horiguchi * where no lock is held and PG_writeback is irrelevant, 1343290408d4SNaoya Horiguchi * and writeback status of all subpages are counted in the reference 1344290408d4SNaoya Horiguchi * count of the head page (i.e. if all subpages of a 2MB hugepage are 1345290408d4SNaoya Horiguchi * under direct I/O, the reference of the head page is 512 and a bit more.) 1346290408d4SNaoya Horiguchi * This means that when we try to migrate hugepage whose subpages are 1347290408d4SNaoya Horiguchi * doing direct I/O, some references remain after try_to_unmap() and 1348290408d4SNaoya Horiguchi * hugepage migration fails without data corruption. 1349290408d4SNaoya Horiguchi * 1350290408d4SNaoya Horiguchi * There is also no race when direct I/O is issued on the page under migration, 1351290408d4SNaoya Horiguchi * because then pte is replaced with migration swap entry and direct I/O code 1352290408d4SNaoya Horiguchi * will wait in the page fault for migration to complete. 1353290408d4SNaoya Horiguchi */ 13544e096ae1SMatthew Wilcox (Oracle) static int unmap_and_move_huge_page(new_folio_t get_new_folio, 13554e096ae1SMatthew Wilcox (Oracle) free_folio_t put_new_folio, unsigned long private, 13564e096ae1SMatthew Wilcox (Oracle) struct folio *src, int force, enum migrate_mode mode, 13574e096ae1SMatthew Wilcox (Oracle) int reason, struct list_head *ret) 1358290408d4SNaoya Horiguchi { 13594e096ae1SMatthew Wilcox (Oracle) struct folio *dst; 13602def7424SHugh Dickins int rc = -EAGAIN; 13612ebba6b7SHugh Dickins int page_was_mapped = 0; 1362290408d4SNaoya Horiguchi struct anon_vma *anon_vma = NULL; 1363c0d0381aSMike Kravetz struct address_space *mapping = NULL; 1364290408d4SNaoya Horiguchi 1365c33db292SMatthew Wilcox (Oracle) if (folio_ref_count(src) == 1) { 136671a64f61SMuchun Song /* page was freed from under us. So we are done. */ 1367ea8e72f4SSidhartha Kumar folio_putback_active_hugetlb(src); 136871a64f61SMuchun Song return MIGRATEPAGE_SUCCESS; 136971a64f61SMuchun Song } 137071a64f61SMuchun Song 13714e096ae1SMatthew Wilcox (Oracle) dst = get_new_folio(src, private); 13724e096ae1SMatthew Wilcox (Oracle) if (!dst) 1373290408d4SNaoya Horiguchi return -ENOMEM; 1374290408d4SNaoya Horiguchi 1375c33db292SMatthew Wilcox (Oracle) if (!folio_trylock(src)) { 13762916ecc0SJérôme Glisse if (!force) 1377290408d4SNaoya Horiguchi goto out; 13782916ecc0SJérôme Glisse switch (mode) { 13792916ecc0SJérôme Glisse case MIGRATE_SYNC: 13802916ecc0SJérôme Glisse case MIGRATE_SYNC_NO_COPY: 13812916ecc0SJérôme Glisse break; 13822916ecc0SJérôme Glisse default: 13832916ecc0SJérôme Glisse goto out; 13842916ecc0SJérôme Glisse } 1385c33db292SMatthew Wilcox (Oracle) folio_lock(src); 1386290408d4SNaoya Horiguchi } 1387290408d4SNaoya Horiguchi 1388cb6acd01SMike Kravetz /* 1389cb6acd01SMike Kravetz * Check for pages which are in the process of being freed. Without 1390c33db292SMatthew Wilcox (Oracle) * folio_mapping() set, hugetlbfs specific move page routine will not 1391cb6acd01SMike Kravetz * be called and we could leak usage counts for subpools. 1392cb6acd01SMike Kravetz */ 1393345c62d1SSidhartha Kumar if (hugetlb_folio_subpool(src) && !folio_mapping(src)) { 1394cb6acd01SMike Kravetz rc = -EBUSY; 1395cb6acd01SMike Kravetz goto out_unlock; 1396cb6acd01SMike Kravetz } 1397cb6acd01SMike Kravetz 1398c33db292SMatthew Wilcox (Oracle) if (folio_test_anon(src)) 139929eea9b5SMatthew Wilcox (Oracle) anon_vma = folio_get_anon_vma(src); 1400290408d4SNaoya Horiguchi 1401c33db292SMatthew Wilcox (Oracle) if (unlikely(!folio_trylock(dst))) 14027db7671fSHugh Dickins goto put_anon; 14037db7671fSHugh Dickins 1404c33db292SMatthew Wilcox (Oracle) if (folio_mapped(src)) { 1405a98a2f0cSAlistair Popple enum ttu_flags ttu = 0; 1406336bf30eSMike Kravetz 1407c33db292SMatthew Wilcox (Oracle) if (!folio_test_anon(src)) { 1408c0d0381aSMike Kravetz /* 1409336bf30eSMike Kravetz * In shared mappings, try_to_unmap could potentially 1410336bf30eSMike Kravetz * call huge_pmd_unshare. Because of this, take 1411336bf30eSMike Kravetz * semaphore in write mode here and set TTU_RMAP_LOCKED 1412336bf30eSMike Kravetz * to let lower levels know we have taken the lock. 1413c0d0381aSMike Kravetz */ 14144e096ae1SMatthew Wilcox (Oracle) mapping = hugetlb_page_mapping_lock_write(&src->page); 1415c0d0381aSMike Kravetz if (unlikely(!mapping)) 1416c0d0381aSMike Kravetz goto unlock_put_anon; 1417c0d0381aSMike Kravetz 14185202978bSMiaohe Lin ttu = TTU_RMAP_LOCKED; 1419336bf30eSMike Kravetz } 1420336bf30eSMike Kravetz 14214b8554c5SMatthew Wilcox (Oracle) try_to_migrate(src, ttu); 14222ebba6b7SHugh Dickins page_was_mapped = 1; 1423336bf30eSMike Kravetz 14245202978bSMiaohe Lin if (ttu & TTU_RMAP_LOCKED) 1425336bf30eSMike Kravetz i_mmap_unlock_write(mapping); 14262ebba6b7SHugh Dickins } 1427290408d4SNaoya Horiguchi 1428c33db292SMatthew Wilcox (Oracle) if (!folio_mapped(src)) 1429e7e3ffebSMatthew Wilcox (Oracle) rc = move_to_new_folio(dst, src, mode); 1430290408d4SNaoya Horiguchi 1431336bf30eSMike Kravetz if (page_was_mapped) 14324eecb8b9SMatthew Wilcox (Oracle) remove_migration_ptes(src, 14334eecb8b9SMatthew Wilcox (Oracle) rc == MIGRATEPAGE_SUCCESS ? dst : src, false); 1434290408d4SNaoya Horiguchi 1435c0d0381aSMike Kravetz unlock_put_anon: 1436c33db292SMatthew Wilcox (Oracle) folio_unlock(dst); 14377db7671fSHugh Dickins 14387db7671fSHugh Dickins put_anon: 1439fd4a4663SHugh Dickins if (anon_vma) 14409e60109fSPeter Zijlstra put_anon_vma(anon_vma); 14418e6ac7faSAneesh Kumar K.V 14422def7424SHugh Dickins if (rc == MIGRATEPAGE_SUCCESS) { 1443345c62d1SSidhartha Kumar move_hugetlb_state(src, dst, reason); 14444e096ae1SMatthew Wilcox (Oracle) put_new_folio = NULL; 14452def7424SHugh Dickins } 14468e6ac7faSAneesh Kumar K.V 1447cb6acd01SMike Kravetz out_unlock: 1448c33db292SMatthew Wilcox (Oracle) folio_unlock(src); 144909761333SHillf Danton out: 1450dd4ae78aSYang Shi if (rc == MIGRATEPAGE_SUCCESS) 1451ea8e72f4SSidhartha Kumar folio_putback_active_hugetlb(src); 1452a04840c6SMiaohe Lin else if (rc != -EAGAIN) 1453c33db292SMatthew Wilcox (Oracle) list_move_tail(&src->lru, ret); 145468711a74SDavid Rientjes 145568711a74SDavid Rientjes /* 145668711a74SDavid Rientjes * If migration was not successful and there's a freeing callback, use 145768711a74SDavid Rientjes * it. Otherwise, put_page() will drop the reference grabbed during 145868711a74SDavid Rientjes * isolation. 145968711a74SDavid Rientjes */ 14604e096ae1SMatthew Wilcox (Oracle) if (put_new_folio) 14614e096ae1SMatthew Wilcox (Oracle) put_new_folio(dst, private); 146268711a74SDavid Rientjes else 1463ea8e72f4SSidhartha Kumar folio_putback_active_hugetlb(dst); 146468711a74SDavid Rientjes 1465290408d4SNaoya Horiguchi return rc; 1466290408d4SNaoya Horiguchi } 1467290408d4SNaoya Horiguchi 1468eaec4e63SHuang Ying static inline int try_split_folio(struct folio *folio, struct list_head *split_folios) 1469d532e2e5SYang Shi { 14709c62ff00SHuang Ying int rc; 1471d532e2e5SYang Shi 1472eaec4e63SHuang Ying folio_lock(folio); 1473eaec4e63SHuang Ying rc = split_folio_to_list(folio, split_folios); 1474eaec4e63SHuang Ying folio_unlock(folio); 1475e6fa8a79SHuang Ying if (!rc) 1476eaec4e63SHuang Ying list_move_tail(&folio->lru, split_folios); 1477d532e2e5SYang Shi 1478d532e2e5SYang Shi return rc; 1479d532e2e5SYang Shi } 1480d532e2e5SYang Shi 148142012e04SHuang Ying #ifdef CONFIG_TRANSPARENT_HUGEPAGE 148242012e04SHuang Ying #define NR_MAX_BATCHED_MIGRATION HPAGE_PMD_NR 148342012e04SHuang Ying #else 148442012e04SHuang Ying #define NR_MAX_BATCHED_MIGRATION 512 148542012e04SHuang Ying #endif 1486e5bfff8bSHuang Ying #define NR_MAX_MIGRATE_PAGES_RETRY 10 14872ef7dbb2SHuang Ying #define NR_MAX_MIGRATE_ASYNC_RETRY 3 14882ef7dbb2SHuang Ying #define NR_MAX_MIGRATE_SYNC_RETRY \ 14892ef7dbb2SHuang Ying (NR_MAX_MIGRATE_PAGES_RETRY - NR_MAX_MIGRATE_ASYNC_RETRY) 1490e5bfff8bSHuang Ying 14915b855937SHuang Ying struct migrate_pages_stats { 14925b855937SHuang Ying int nr_succeeded; /* Normal and large folios migrated successfully, in 14935b855937SHuang Ying units of base pages */ 14945b855937SHuang Ying int nr_failed_pages; /* Normal and large folios failed to be migrated, in 14955b855937SHuang Ying units of base pages. Untried folios aren't counted */ 14965b855937SHuang Ying int nr_thp_succeeded; /* THP migrated successfully */ 14975b855937SHuang Ying int nr_thp_failed; /* THP failed to be migrated */ 14985b855937SHuang Ying int nr_thp_split; /* THP split before migrating */ 14995b855937SHuang Ying }; 15005b855937SHuang Ying 1501290408d4SNaoya Horiguchi /* 1502e5bfff8bSHuang Ying * Returns the number of hugetlb folios that were not migrated, or an error code 1503e5bfff8bSHuang Ying * after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no hugetlb folios are movable 1504e5bfff8bSHuang Ying * any more because the list has become empty or no retryable hugetlb folios 1505e5bfff8bSHuang Ying * exist any more. It is caller's responsibility to call putback_movable_pages() 1506e5bfff8bSHuang Ying * only if ret != 0. 1507e5bfff8bSHuang Ying */ 15084e096ae1SMatthew Wilcox (Oracle) static int migrate_hugetlbs(struct list_head *from, new_folio_t get_new_folio, 15094e096ae1SMatthew Wilcox (Oracle) free_folio_t put_new_folio, unsigned long private, 1510e5bfff8bSHuang Ying enum migrate_mode mode, int reason, 1511e5bfff8bSHuang Ying struct migrate_pages_stats *stats, 1512e5bfff8bSHuang Ying struct list_head *ret_folios) 1513e5bfff8bSHuang Ying { 1514e5bfff8bSHuang Ying int retry = 1; 1515e5bfff8bSHuang Ying int nr_failed = 0; 1516e5bfff8bSHuang Ying int nr_retry_pages = 0; 1517e5bfff8bSHuang Ying int pass = 0; 1518e5bfff8bSHuang Ying struct folio *folio, *folio2; 1519e5bfff8bSHuang Ying int rc, nr_pages; 1520e5bfff8bSHuang Ying 1521e5bfff8bSHuang Ying for (pass = 0; pass < NR_MAX_MIGRATE_PAGES_RETRY && retry; pass++) { 1522e5bfff8bSHuang Ying retry = 0; 1523e5bfff8bSHuang Ying nr_retry_pages = 0; 1524e5bfff8bSHuang Ying 1525e5bfff8bSHuang Ying list_for_each_entry_safe(folio, folio2, from, lru) { 1526e5bfff8bSHuang Ying if (!folio_test_hugetlb(folio)) 1527e5bfff8bSHuang Ying continue; 1528e5bfff8bSHuang Ying 1529e5bfff8bSHuang Ying nr_pages = folio_nr_pages(folio); 1530e5bfff8bSHuang Ying 1531e5bfff8bSHuang Ying cond_resched(); 1532e5bfff8bSHuang Ying 15336f7d760eSHuang Ying /* 15346f7d760eSHuang Ying * Migratability of hugepages depends on architectures and 15356f7d760eSHuang Ying * their size. This check is necessary because some callers 15366f7d760eSHuang Ying * of hugepage migration like soft offline and memory 15376f7d760eSHuang Ying * hotremove don't walk through page tables or check whether 15386f7d760eSHuang Ying * the hugepage is pmd-based or not before kicking migration. 15396f7d760eSHuang Ying */ 15406f7d760eSHuang Ying if (!hugepage_migration_supported(folio_hstate(folio))) { 15416f7d760eSHuang Ying nr_failed++; 15426f7d760eSHuang Ying stats->nr_failed_pages += nr_pages; 15436f7d760eSHuang Ying list_move_tail(&folio->lru, ret_folios); 15446f7d760eSHuang Ying continue; 15456f7d760eSHuang Ying } 15466f7d760eSHuang Ying 15474e096ae1SMatthew Wilcox (Oracle) rc = unmap_and_move_huge_page(get_new_folio, 15484e096ae1SMatthew Wilcox (Oracle) put_new_folio, private, 15494e096ae1SMatthew Wilcox (Oracle) folio, pass > 2, mode, 1550e5bfff8bSHuang Ying reason, ret_folios); 1551e5bfff8bSHuang Ying /* 1552e5bfff8bSHuang Ying * The rules are: 1553e5bfff8bSHuang Ying * Success: hugetlb folio will be put back 1554e5bfff8bSHuang Ying * -EAGAIN: stay on the from list 1555e5bfff8bSHuang Ying * -ENOMEM: stay on the from list 1556e5bfff8bSHuang Ying * Other errno: put on ret_folios list 1557e5bfff8bSHuang Ying */ 1558e5bfff8bSHuang Ying switch(rc) { 1559e5bfff8bSHuang Ying case -ENOMEM: 1560e5bfff8bSHuang Ying /* 1561e5bfff8bSHuang Ying * When memory is low, don't bother to try to migrate 1562e5bfff8bSHuang Ying * other folios, just exit. 1563e5bfff8bSHuang Ying */ 1564e5bfff8bSHuang Ying stats->nr_failed_pages += nr_pages + nr_retry_pages; 1565e5bfff8bSHuang Ying return -ENOMEM; 1566e5bfff8bSHuang Ying case -EAGAIN: 1567e5bfff8bSHuang Ying retry++; 1568e5bfff8bSHuang Ying nr_retry_pages += nr_pages; 1569e5bfff8bSHuang Ying break; 1570e5bfff8bSHuang Ying case MIGRATEPAGE_SUCCESS: 1571e5bfff8bSHuang Ying stats->nr_succeeded += nr_pages; 1572e5bfff8bSHuang Ying break; 1573e5bfff8bSHuang Ying default: 1574e5bfff8bSHuang Ying /* 1575e5bfff8bSHuang Ying * Permanent failure (-EBUSY, etc.): 1576e5bfff8bSHuang Ying * unlike -EAGAIN case, the failed folio is 1577e5bfff8bSHuang Ying * removed from migration folio list and not 1578e5bfff8bSHuang Ying * retried in the next outer loop. 1579e5bfff8bSHuang Ying */ 1580e5bfff8bSHuang Ying nr_failed++; 1581e5bfff8bSHuang Ying stats->nr_failed_pages += nr_pages; 1582e5bfff8bSHuang Ying break; 1583e5bfff8bSHuang Ying } 1584e5bfff8bSHuang Ying } 1585e5bfff8bSHuang Ying } 1586e5bfff8bSHuang Ying /* 1587e5bfff8bSHuang Ying * nr_failed is number of hugetlb folios failed to be migrated. After 1588e5bfff8bSHuang Ying * NR_MAX_MIGRATE_PAGES_RETRY attempts, give up and count retried hugetlb 1589e5bfff8bSHuang Ying * folios as failed. 1590e5bfff8bSHuang Ying */ 1591e5bfff8bSHuang Ying nr_failed += retry; 1592e5bfff8bSHuang Ying stats->nr_failed_pages += nr_retry_pages; 1593e5bfff8bSHuang Ying 1594e5bfff8bSHuang Ying return nr_failed; 1595e5bfff8bSHuang Ying } 1596e5bfff8bSHuang Ying 15975dfab109SHuang Ying /* 15985dfab109SHuang Ying * migrate_pages_batch() first unmaps folios in the from list as many as 15995dfab109SHuang Ying * possible, then move the unmapped folios. 1600fb3592c4SHuang Ying * 1601fb3592c4SHuang Ying * We only batch migration if mode == MIGRATE_ASYNC to avoid to wait a 1602fb3592c4SHuang Ying * lock or bit when we have locked more than one folio. Which may cause 1603fb3592c4SHuang Ying * deadlock (e.g., for loop device). So, if mode != MIGRATE_ASYNC, the 1604fb3592c4SHuang Ying * length of the from list must be <= 1. 16055dfab109SHuang Ying */ 16064e096ae1SMatthew Wilcox (Oracle) static int migrate_pages_batch(struct list_head *from, 16074e096ae1SMatthew Wilcox (Oracle) new_folio_t get_new_folio, free_folio_t put_new_folio, 16084e096ae1SMatthew Wilcox (Oracle) unsigned long private, enum migrate_mode mode, int reason, 16094e096ae1SMatthew Wilcox (Oracle) struct list_head *ret_folios, struct list_head *split_folios, 16104e096ae1SMatthew Wilcox (Oracle) struct migrate_pages_stats *stats, int nr_pass) 161142012e04SHuang Ying { 1612a21d2133SHuang Ying int retry = 1; 161342012e04SHuang Ying int thp_retry = 1; 161442012e04SHuang Ying int nr_failed = 0; 161542012e04SHuang Ying int nr_retry_pages = 0; 161642012e04SHuang Ying int pass = 0; 161742012e04SHuang Ying bool is_thp = false; 16185dfab109SHuang Ying struct folio *folio, *folio2, *dst = NULL, *dst2; 1619a21d2133SHuang Ying int rc, rc_saved = 0, nr_pages; 16205dfab109SHuang Ying LIST_HEAD(unmap_folios); 16215dfab109SHuang Ying LIST_HEAD(dst_folios); 162242012e04SHuang Ying bool nosplit = (reason == MR_NUMA_MISPLACED); 162342012e04SHuang Ying 1624fb3592c4SHuang Ying VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC && 1625fb3592c4SHuang Ying !list_empty(from) && !list_is_singular(from)); 1626a21d2133SHuang Ying 1627124abcedSHuang Ying for (pass = 0; pass < nr_pass && retry; pass++) { 162842012e04SHuang Ying retry = 0; 162942012e04SHuang Ying thp_retry = 0; 163042012e04SHuang Ying nr_retry_pages = 0; 163142012e04SHuang Ying 163242012e04SHuang Ying list_for_each_entry_safe(folio, folio2, from, lru) { 1633124abcedSHuang Ying is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio); 163442012e04SHuang Ying nr_pages = folio_nr_pages(folio); 163542012e04SHuang Ying 163642012e04SHuang Ying cond_resched(); 163742012e04SHuang Ying 16386f7d760eSHuang Ying /* 16396f7d760eSHuang Ying * Large folio migration might be unsupported or 16406f7d760eSHuang Ying * the allocation might be failed so we should retry 16416f7d760eSHuang Ying * on the same folio with the large folio split 16426f7d760eSHuang Ying * to normal folios. 16436f7d760eSHuang Ying * 16446f7d760eSHuang Ying * Split folios are put in split_folios, and 16456f7d760eSHuang Ying * we will migrate them after the rest of the 16466f7d760eSHuang Ying * list is processed. 16476f7d760eSHuang Ying */ 16486f7d760eSHuang Ying if (!thp_migration_supported() && is_thp) { 1649124abcedSHuang Ying nr_failed++; 16506f7d760eSHuang Ying stats->nr_thp_failed++; 1651a21d2133SHuang Ying if (!try_split_folio(folio, split_folios)) { 16526f7d760eSHuang Ying stats->nr_thp_split++; 16536f7d760eSHuang Ying continue; 16546f7d760eSHuang Ying } 16556f7d760eSHuang Ying stats->nr_failed_pages += nr_pages; 16566f7d760eSHuang Ying list_move_tail(&folio->lru, ret_folios); 16576f7d760eSHuang Ying continue; 16586f7d760eSHuang Ying } 16596f7d760eSHuang Ying 16604e096ae1SMatthew Wilcox (Oracle) rc = migrate_folio_unmap(get_new_folio, put_new_folio, 16614e096ae1SMatthew Wilcox (Oracle) private, folio, &dst, mode, reason, 16624e096ae1SMatthew Wilcox (Oracle) ret_folios); 166342012e04SHuang Ying /* 166442012e04SHuang Ying * The rules are: 166542012e04SHuang Ying * Success: folio will be freed 16665dfab109SHuang Ying * Unmap: folio will be put on unmap_folios list, 16675dfab109SHuang Ying * dst folio put on dst_folios list 166842012e04SHuang Ying * -EAGAIN: stay on the from list 166942012e04SHuang Ying * -ENOMEM: stay on the from list 167042012e04SHuang Ying * Other errno: put on ret_folios list 167142012e04SHuang Ying */ 167242012e04SHuang Ying switch(rc) { 167342012e04SHuang Ying case -ENOMEM: 167442012e04SHuang Ying /* 167542012e04SHuang Ying * When memory is low, don't bother to try to migrate 16765dfab109SHuang Ying * other folios, move unmapped folios, then exit. 167742012e04SHuang Ying */ 1678124abcedSHuang Ying nr_failed++; 167942012e04SHuang Ying stats->nr_thp_failed += is_thp; 168042012e04SHuang Ying /* Large folio NUMA faulting doesn't split to retry. */ 1681124abcedSHuang Ying if (folio_test_large(folio) && !nosplit) { 1682a21d2133SHuang Ying int ret = try_split_folio(folio, split_folios); 168342012e04SHuang Ying 168442012e04SHuang Ying if (!ret) { 168542012e04SHuang Ying stats->nr_thp_split += is_thp; 168642012e04SHuang Ying break; 168742012e04SHuang Ying } else if (reason == MR_LONGTERM_PIN && 168842012e04SHuang Ying ret == -EAGAIN) { 168942012e04SHuang Ying /* 169042012e04SHuang Ying * Try again to split large folio to 169142012e04SHuang Ying * mitigate the failure of longterm pinning. 169242012e04SHuang Ying */ 1693124abcedSHuang Ying retry++; 169442012e04SHuang Ying thp_retry += is_thp; 169542012e04SHuang Ying nr_retry_pages += nr_pages; 1696851ae642SHuang Ying /* Undo duplicated failure counting. */ 1697124abcedSHuang Ying nr_failed--; 1698851ae642SHuang Ying stats->nr_thp_failed -= is_thp; 169942012e04SHuang Ying break; 170042012e04SHuang Ying } 170142012e04SHuang Ying } 170242012e04SHuang Ying 170342012e04SHuang Ying stats->nr_failed_pages += nr_pages + nr_retry_pages; 170442012e04SHuang Ying /* nr_failed isn't updated for not used */ 170542012e04SHuang Ying stats->nr_thp_failed += thp_retry; 17065dfab109SHuang Ying rc_saved = rc; 17075dfab109SHuang Ying if (list_empty(&unmap_folios)) 170842012e04SHuang Ying goto out; 17095dfab109SHuang Ying else 17105dfab109SHuang Ying goto move; 171142012e04SHuang Ying case -EAGAIN: 171242012e04SHuang Ying retry++; 1713124abcedSHuang Ying thp_retry += is_thp; 171442012e04SHuang Ying nr_retry_pages += nr_pages; 171542012e04SHuang Ying break; 171642012e04SHuang Ying case MIGRATEPAGE_SUCCESS: 171742012e04SHuang Ying stats->nr_succeeded += nr_pages; 171842012e04SHuang Ying stats->nr_thp_succeeded += is_thp; 171942012e04SHuang Ying break; 17205dfab109SHuang Ying case MIGRATEPAGE_UNMAP: 17215dfab109SHuang Ying list_move_tail(&folio->lru, &unmap_folios); 17225dfab109SHuang Ying list_add_tail(&dst->lru, &dst_folios); 17235dfab109SHuang Ying break; 172442012e04SHuang Ying default: 172542012e04SHuang Ying /* 172642012e04SHuang Ying * Permanent failure (-EBUSY, etc.): 172742012e04SHuang Ying * unlike -EAGAIN case, the failed folio is 172842012e04SHuang Ying * removed from migration folio list and not 172942012e04SHuang Ying * retried in the next outer loop. 173042012e04SHuang Ying */ 173142012e04SHuang Ying nr_failed++; 1732124abcedSHuang Ying stats->nr_thp_failed += is_thp; 173342012e04SHuang Ying stats->nr_failed_pages += nr_pages; 173442012e04SHuang Ying break; 173542012e04SHuang Ying } 173642012e04SHuang Ying } 173742012e04SHuang Ying } 173842012e04SHuang Ying nr_failed += retry; 173942012e04SHuang Ying stats->nr_thp_failed += thp_retry; 174042012e04SHuang Ying stats->nr_failed_pages += nr_retry_pages; 17415dfab109SHuang Ying move: 17427e12beb8SHuang Ying /* Flush TLBs for all unmapped folios */ 17437e12beb8SHuang Ying try_to_unmap_flush(); 17447e12beb8SHuang Ying 17455dfab109SHuang Ying retry = 1; 1746124abcedSHuang Ying for (pass = 0; pass < nr_pass && retry; pass++) { 17475dfab109SHuang Ying retry = 0; 17485dfab109SHuang Ying thp_retry = 0; 17495dfab109SHuang Ying nr_retry_pages = 0; 17505dfab109SHuang Ying 17515dfab109SHuang Ying dst = list_first_entry(&dst_folios, struct folio, lru); 17525dfab109SHuang Ying dst2 = list_next_entry(dst, lru); 17535dfab109SHuang Ying list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) { 1754124abcedSHuang Ying is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio); 17555dfab109SHuang Ying nr_pages = folio_nr_pages(folio); 17565dfab109SHuang Ying 17575dfab109SHuang Ying cond_resched(); 17585dfab109SHuang Ying 17594e096ae1SMatthew Wilcox (Oracle) rc = migrate_folio_move(put_new_folio, private, 17605dfab109SHuang Ying folio, dst, mode, 17615dfab109SHuang Ying reason, ret_folios); 17625dfab109SHuang Ying /* 17635dfab109SHuang Ying * The rules are: 17645dfab109SHuang Ying * Success: folio will be freed 17655dfab109SHuang Ying * -EAGAIN: stay on the unmap_folios list 17665dfab109SHuang Ying * Other errno: put on ret_folios list 17675dfab109SHuang Ying */ 17685dfab109SHuang Ying switch(rc) { 17695dfab109SHuang Ying case -EAGAIN: 17705dfab109SHuang Ying retry++; 1771124abcedSHuang Ying thp_retry += is_thp; 17725dfab109SHuang Ying nr_retry_pages += nr_pages; 17735dfab109SHuang Ying break; 17745dfab109SHuang Ying case MIGRATEPAGE_SUCCESS: 17755dfab109SHuang Ying stats->nr_succeeded += nr_pages; 17765dfab109SHuang Ying stats->nr_thp_succeeded += is_thp; 17775dfab109SHuang Ying break; 17785dfab109SHuang Ying default: 17795dfab109SHuang Ying nr_failed++; 1780124abcedSHuang Ying stats->nr_thp_failed += is_thp; 17815dfab109SHuang Ying stats->nr_failed_pages += nr_pages; 17825dfab109SHuang Ying break; 17835dfab109SHuang Ying } 17845dfab109SHuang Ying dst = dst2; 17855dfab109SHuang Ying dst2 = list_next_entry(dst, lru); 17865dfab109SHuang Ying } 17875dfab109SHuang Ying } 17885dfab109SHuang Ying nr_failed += retry; 17895dfab109SHuang Ying stats->nr_thp_failed += thp_retry; 17905dfab109SHuang Ying stats->nr_failed_pages += nr_retry_pages; 17915dfab109SHuang Ying 1792124abcedSHuang Ying rc = rc_saved ? : nr_failed; 17935dfab109SHuang Ying out: 17945dfab109SHuang Ying /* Cleanup remaining folios */ 17955dfab109SHuang Ying dst = list_first_entry(&dst_folios, struct folio, lru); 17965dfab109SHuang Ying dst2 = list_next_entry(dst, lru); 17975dfab109SHuang Ying list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) { 17985dfab109SHuang Ying int page_was_mapped = 0; 17995dfab109SHuang Ying struct anon_vma *anon_vma = NULL; 18005dfab109SHuang Ying 18015dfab109SHuang Ying __migrate_folio_extract(dst, &page_was_mapped, &anon_vma); 18025dfab109SHuang Ying migrate_folio_undo_src(folio, page_was_mapped, anon_vma, 1803ebe75e47SHuang Ying true, ret_folios); 18045dfab109SHuang Ying list_del(&dst->lru); 18054e096ae1SMatthew Wilcox (Oracle) migrate_folio_undo_dst(dst, true, put_new_folio, private); 18065dfab109SHuang Ying dst = dst2; 18075dfab109SHuang Ying dst2 = list_next_entry(dst, lru); 18085dfab109SHuang Ying } 18095dfab109SHuang Ying 181042012e04SHuang Ying return rc; 181142012e04SHuang Ying } 181242012e04SHuang Ying 18134e096ae1SMatthew Wilcox (Oracle) static int migrate_pages_sync(struct list_head *from, new_folio_t get_new_folio, 18144e096ae1SMatthew Wilcox (Oracle) free_folio_t put_new_folio, unsigned long private, 18154e096ae1SMatthew Wilcox (Oracle) enum migrate_mode mode, int reason, 18164e096ae1SMatthew Wilcox (Oracle) struct list_head *ret_folios, struct list_head *split_folios, 18174e096ae1SMatthew Wilcox (Oracle) struct migrate_pages_stats *stats) 18182ef7dbb2SHuang Ying { 18192ef7dbb2SHuang Ying int rc, nr_failed = 0; 18202ef7dbb2SHuang Ying LIST_HEAD(folios); 18212ef7dbb2SHuang Ying struct migrate_pages_stats astats; 18222ef7dbb2SHuang Ying 18232ef7dbb2SHuang Ying memset(&astats, 0, sizeof(astats)); 18242ef7dbb2SHuang Ying /* Try to migrate in batch with MIGRATE_ASYNC mode firstly */ 18254e096ae1SMatthew Wilcox (Oracle) rc = migrate_pages_batch(from, get_new_folio, put_new_folio, private, MIGRATE_ASYNC, 18262ef7dbb2SHuang Ying reason, &folios, split_folios, &astats, 18272ef7dbb2SHuang Ying NR_MAX_MIGRATE_ASYNC_RETRY); 18282ef7dbb2SHuang Ying stats->nr_succeeded += astats.nr_succeeded; 18292ef7dbb2SHuang Ying stats->nr_thp_succeeded += astats.nr_thp_succeeded; 18302ef7dbb2SHuang Ying stats->nr_thp_split += astats.nr_thp_split; 18312ef7dbb2SHuang Ying if (rc < 0) { 18322ef7dbb2SHuang Ying stats->nr_failed_pages += astats.nr_failed_pages; 18332ef7dbb2SHuang Ying stats->nr_thp_failed += astats.nr_thp_failed; 18342ef7dbb2SHuang Ying list_splice_tail(&folios, ret_folios); 18352ef7dbb2SHuang Ying return rc; 18362ef7dbb2SHuang Ying } 18372ef7dbb2SHuang Ying stats->nr_thp_failed += astats.nr_thp_split; 18382ef7dbb2SHuang Ying nr_failed += astats.nr_thp_split; 18392ef7dbb2SHuang Ying /* 18402ef7dbb2SHuang Ying * Fall back to migrate all failed folios one by one synchronously. All 18412ef7dbb2SHuang Ying * failed folios except split THPs will be retried, so their failure 18422ef7dbb2SHuang Ying * isn't counted 18432ef7dbb2SHuang Ying */ 18442ef7dbb2SHuang Ying list_splice_tail_init(&folios, from); 18452ef7dbb2SHuang Ying while (!list_empty(from)) { 18462ef7dbb2SHuang Ying list_move(from->next, &folios); 18474e096ae1SMatthew Wilcox (Oracle) rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio, 18482ef7dbb2SHuang Ying private, mode, reason, ret_folios, 18492ef7dbb2SHuang Ying split_folios, stats, NR_MAX_MIGRATE_SYNC_RETRY); 18502ef7dbb2SHuang Ying list_splice_tail_init(&folios, ret_folios); 18512ef7dbb2SHuang Ying if (rc < 0) 18522ef7dbb2SHuang Ying return rc; 18532ef7dbb2SHuang Ying nr_failed += rc; 18542ef7dbb2SHuang Ying } 18552ef7dbb2SHuang Ying 18562ef7dbb2SHuang Ying return nr_failed; 18572ef7dbb2SHuang Ying } 18582ef7dbb2SHuang Ying 1859e24f0b8fSChristoph Lameter /* 1860eaec4e63SHuang Ying * migrate_pages - migrate the folios specified in a list, to the free folios 1861c73e5c9cSSrivatsa S. Bhat * supplied as the target for the page migration 1862e24f0b8fSChristoph Lameter * 1863eaec4e63SHuang Ying * @from: The list of folios to be migrated. 18644e096ae1SMatthew Wilcox (Oracle) * @get_new_folio: The function used to allocate free folios to be used 1865eaec4e63SHuang Ying * as the target of the folio migration. 18664e096ae1SMatthew Wilcox (Oracle) * @put_new_folio: The function used to free target folios if migration 186768711a74SDavid Rientjes * fails, or NULL if no special handling is necessary. 18684e096ae1SMatthew Wilcox (Oracle) * @private: Private data to be passed on to get_new_folio() 1869c73e5c9cSSrivatsa S. Bhat * @mode: The migration mode that specifies the constraints for 1870eaec4e63SHuang Ying * folio migration, if any. 1871eaec4e63SHuang Ying * @reason: The reason for folio migration. 1872eaec4e63SHuang Ying * @ret_succeeded: Set to the number of folios migrated successfully if 18735ac95884SYang Shi * the caller passes a non-NULL pointer. 1874e24f0b8fSChristoph Lameter * 1875e5bfff8bSHuang Ying * The function returns after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no folios 1876e5bfff8bSHuang Ying * are movable any more because the list has become empty or no retryable folios 1877e5bfff8bSHuang Ying * exist any more. It is caller's responsibility to call putback_movable_pages() 1878e5bfff8bSHuang Ying * only if ret != 0. 1879e24f0b8fSChristoph Lameter * 1880eaec4e63SHuang Ying * Returns the number of {normal folio, large folio, hugetlb} that were not 1881eaec4e63SHuang Ying * migrated, or an error code. The number of large folio splits will be 1882eaec4e63SHuang Ying * considered as the number of non-migrated large folio, no matter how many 1883eaec4e63SHuang Ying * split folios of the large folio are migrated successfully. 1884e24f0b8fSChristoph Lameter */ 18854e096ae1SMatthew Wilcox (Oracle) int migrate_pages(struct list_head *from, new_folio_t get_new_folio, 18864e096ae1SMatthew Wilcox (Oracle) free_folio_t put_new_folio, unsigned long private, 18875ac95884SYang Shi enum migrate_mode mode, int reason, unsigned int *ret_succeeded) 1888e24f0b8fSChristoph Lameter { 188942012e04SHuang Ying int rc, rc_gather; 18902ef7dbb2SHuang Ying int nr_pages; 1891eaec4e63SHuang Ying struct folio *folio, *folio2; 189242012e04SHuang Ying LIST_HEAD(folios); 1893eaec4e63SHuang Ying LIST_HEAD(ret_folios); 1894a21d2133SHuang Ying LIST_HEAD(split_folios); 18955b855937SHuang Ying struct migrate_pages_stats stats; 18962d1db3b1SChristoph Lameter 18977bc1aec5SLiam Mark trace_mm_migrate_pages_start(mode, reason); 18987bc1aec5SLiam Mark 18995b855937SHuang Ying memset(&stats, 0, sizeof(stats)); 1900e24f0b8fSChristoph Lameter 19014e096ae1SMatthew Wilcox (Oracle) rc_gather = migrate_hugetlbs(from, get_new_folio, put_new_folio, private, 190242012e04SHuang Ying mode, reason, &stats, &ret_folios); 190342012e04SHuang Ying if (rc_gather < 0) 190495a402c3SChristoph Lameter goto out; 1905fb3592c4SHuang Ying 190642012e04SHuang Ying again: 190742012e04SHuang Ying nr_pages = 0; 1908b20a3503SChristoph Lameter list_for_each_entry_safe(folio, folio2, from, lru) { 1909e5bfff8bSHuang Ying /* Retried hugetlb folios will be kept in list */ 1910e5bfff8bSHuang Ying if (folio_test_hugetlb(folio)) { 1911e5bfff8bSHuang Ying list_move_tail(&folio->lru, &ret_folios); 1912e5bfff8bSHuang Ying continue; 1913eaec4e63SHuang Ying } 1914f430893bSMiaohe Lin 191542012e04SHuang Ying nr_pages += folio_nr_pages(folio); 19162ef7dbb2SHuang Ying if (nr_pages >= NR_MAX_BATCHED_MIGRATION) 1917e24f0b8fSChristoph Lameter break; 1918b20a3503SChristoph Lameter } 19192ef7dbb2SHuang Ying if (nr_pages >= NR_MAX_BATCHED_MIGRATION) 1920fb3592c4SHuang Ying list_cut_before(&folios, from, &folio2->lru); 192142012e04SHuang Ying else 192242012e04SHuang Ying list_splice_init(from, &folios); 19232ef7dbb2SHuang Ying if (mode == MIGRATE_ASYNC) 19244e096ae1SMatthew Wilcox (Oracle) rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio, 19254e096ae1SMatthew Wilcox (Oracle) private, mode, reason, &ret_folios, 19264e096ae1SMatthew Wilcox (Oracle) &split_folios, &stats, 1927a21d2133SHuang Ying NR_MAX_MIGRATE_PAGES_RETRY); 19282ef7dbb2SHuang Ying else 19294e096ae1SMatthew Wilcox (Oracle) rc = migrate_pages_sync(&folios, get_new_folio, put_new_folio, 19304e096ae1SMatthew Wilcox (Oracle) private, mode, reason, &ret_folios, 19314e096ae1SMatthew Wilcox (Oracle) &split_folios, &stats); 193242012e04SHuang Ying list_splice_tail_init(&folios, &ret_folios); 193342012e04SHuang Ying if (rc < 0) { 193442012e04SHuang Ying rc_gather = rc; 1935a21d2133SHuang Ying list_splice_tail(&split_folios, &ret_folios); 1936b20a3503SChristoph Lameter goto out; 1937b20a3503SChristoph Lameter } 1938a21d2133SHuang Ying if (!list_empty(&split_folios)) { 1939a21d2133SHuang Ying /* 1940a21d2133SHuang Ying * Failure isn't counted since all split folios of a large folio 1941a21d2133SHuang Ying * is counted as 1 failure already. And, we only try to migrate 1942a21d2133SHuang Ying * with minimal effort, force MIGRATE_ASYNC mode and retry once. 1943a21d2133SHuang Ying */ 19444e096ae1SMatthew Wilcox (Oracle) migrate_pages_batch(&split_folios, get_new_folio, 19454e096ae1SMatthew Wilcox (Oracle) put_new_folio, private, MIGRATE_ASYNC, reason, 19464e096ae1SMatthew Wilcox (Oracle) &ret_folios, NULL, &stats, 1); 1947a21d2133SHuang Ying list_splice_tail_init(&split_folios, &ret_folios); 1948a21d2133SHuang Ying } 194942012e04SHuang Ying rc_gather += rc; 195042012e04SHuang Ying if (!list_empty(from)) 195142012e04SHuang Ying goto again; 195295a402c3SChristoph Lameter out: 1953dd4ae78aSYang Shi /* 1954eaec4e63SHuang Ying * Put the permanent failure folio back to migration list, they 1955dd4ae78aSYang Shi * will be put back to the right list by the caller. 1956dd4ae78aSYang Shi */ 1957eaec4e63SHuang Ying list_splice(&ret_folios, from); 1958dd4ae78aSYang Shi 195903e5f82eSBaolin Wang /* 1960eaec4e63SHuang Ying * Return 0 in case all split folios of fail-to-migrate large folios 1961eaec4e63SHuang Ying * are migrated successfully. 196203e5f82eSBaolin Wang */ 196303e5f82eSBaolin Wang if (list_empty(from)) 196442012e04SHuang Ying rc_gather = 0; 196503e5f82eSBaolin Wang 19665b855937SHuang Ying count_vm_events(PGMIGRATE_SUCCESS, stats.nr_succeeded); 19675b855937SHuang Ying count_vm_events(PGMIGRATE_FAIL, stats.nr_failed_pages); 19685b855937SHuang Ying count_vm_events(THP_MIGRATION_SUCCESS, stats.nr_thp_succeeded); 19695b855937SHuang Ying count_vm_events(THP_MIGRATION_FAIL, stats.nr_thp_failed); 19705b855937SHuang Ying count_vm_events(THP_MIGRATION_SPLIT, stats.nr_thp_split); 19715b855937SHuang Ying trace_mm_migrate_pages(stats.nr_succeeded, stats.nr_failed_pages, 19725b855937SHuang Ying stats.nr_thp_succeeded, stats.nr_thp_failed, 19735b855937SHuang Ying stats.nr_thp_split, mode, reason); 19747b2a2d4aSMel Gorman 19755ac95884SYang Shi if (ret_succeeded) 19765b855937SHuang Ying *ret_succeeded = stats.nr_succeeded; 19775ac95884SYang Shi 197842012e04SHuang Ying return rc_gather; 1979b20a3503SChristoph Lameter } 1980b20a3503SChristoph Lameter 19814e096ae1SMatthew Wilcox (Oracle) struct folio *alloc_migration_target(struct folio *src, unsigned long private) 1982b4b38223SJoonsoo Kim { 198319fc7bedSJoonsoo Kim struct migration_target_control *mtc; 198419fc7bedSJoonsoo Kim gfp_t gfp_mask; 1985b4b38223SJoonsoo Kim unsigned int order = 0; 198619fc7bedSJoonsoo Kim int nid; 198719fc7bedSJoonsoo Kim int zidx; 198819fc7bedSJoonsoo Kim 198919fc7bedSJoonsoo Kim mtc = (struct migration_target_control *)private; 199019fc7bedSJoonsoo Kim gfp_mask = mtc->gfp_mask; 199119fc7bedSJoonsoo Kim nid = mtc->nid; 199219fc7bedSJoonsoo Kim if (nid == NUMA_NO_NODE) 19934e096ae1SMatthew Wilcox (Oracle) nid = folio_nid(src); 1994b4b38223SJoonsoo Kim 19954e096ae1SMatthew Wilcox (Oracle) if (folio_test_hugetlb(src)) { 19964e096ae1SMatthew Wilcox (Oracle) struct hstate *h = folio_hstate(src); 1997d92bbc27SJoonsoo Kim 199819fc7bedSJoonsoo Kim gfp_mask = htlb_modify_alloc_mask(h, gfp_mask); 19994e096ae1SMatthew Wilcox (Oracle) return alloc_hugetlb_folio_nodemask(h, nid, 2000e37d3e83SSidhartha Kumar mtc->nmask, gfp_mask); 2001d92bbc27SJoonsoo Kim } 2002b4b38223SJoonsoo Kim 20034e096ae1SMatthew Wilcox (Oracle) if (folio_test_large(src)) { 20049933a0c8SJoonsoo Kim /* 20059933a0c8SJoonsoo Kim * clear __GFP_RECLAIM to make the migration callback 20069933a0c8SJoonsoo Kim * consistent with regular THP allocations. 20079933a0c8SJoonsoo Kim */ 20089933a0c8SJoonsoo Kim gfp_mask &= ~__GFP_RECLAIM; 2009b4b38223SJoonsoo Kim gfp_mask |= GFP_TRANSHUGE; 20104e096ae1SMatthew Wilcox (Oracle) order = folio_order(src); 2011b4b38223SJoonsoo Kim } 20124e096ae1SMatthew Wilcox (Oracle) zidx = zone_idx(folio_zone(src)); 201319fc7bedSJoonsoo Kim if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE) 2014b4b38223SJoonsoo Kim gfp_mask |= __GFP_HIGHMEM; 2015b4b38223SJoonsoo Kim 20164e096ae1SMatthew Wilcox (Oracle) return __folio_alloc(gfp_mask, order, nid, mtc->nmask); 2017b4b38223SJoonsoo Kim } 2018b4b38223SJoonsoo Kim 2019742755a1SChristoph Lameter #ifdef CONFIG_NUMA 2020742755a1SChristoph Lameter 2021a49bd4d7SMichal Hocko static int store_status(int __user *status, int start, int value, int nr) 2022742755a1SChristoph Lameter { 2023a49bd4d7SMichal Hocko while (nr-- > 0) { 2024a49bd4d7SMichal Hocko if (put_user(value, status + start)) 2025a49bd4d7SMichal Hocko return -EFAULT; 2026a49bd4d7SMichal Hocko start++; 2027a49bd4d7SMichal Hocko } 2028742755a1SChristoph Lameter 2029a49bd4d7SMichal Hocko return 0; 2030a49bd4d7SMichal Hocko } 2031742755a1SChristoph Lameter 2032a49bd4d7SMichal Hocko static int do_move_pages_to_node(struct mm_struct *mm, 2033a49bd4d7SMichal Hocko struct list_head *pagelist, int node) 2034a49bd4d7SMichal Hocko { 2035a49bd4d7SMichal Hocko int err; 2036a0976311SJoonsoo Kim struct migration_target_control mtc = { 2037a0976311SJoonsoo Kim .nid = node, 2038a0976311SJoonsoo Kim .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 2039a0976311SJoonsoo Kim }; 2040742755a1SChristoph Lameter 2041a0976311SJoonsoo Kim err = migrate_pages(pagelist, alloc_migration_target, NULL, 20425ac95884SYang Shi (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL); 2043a49bd4d7SMichal Hocko if (err) 2044a49bd4d7SMichal Hocko putback_movable_pages(pagelist); 2045a49bd4d7SMichal Hocko return err; 2046742755a1SChristoph Lameter } 2047742755a1SChristoph Lameter 2048742755a1SChristoph Lameter /* 2049a49bd4d7SMichal Hocko * Resolves the given address to a struct page, isolates it from the LRU and 2050a49bd4d7SMichal Hocko * puts it to the given pagelist. 2051e0153fc2SYang Shi * Returns: 2052e0153fc2SYang Shi * errno - if the page cannot be found/isolated 2053e0153fc2SYang Shi * 0 - when it doesn't have to be migrated because it is already on the 2054e0153fc2SYang Shi * target node 2055e0153fc2SYang Shi * 1 - when it has been queued 2056742755a1SChristoph Lameter */ 2057428e106aSKirill A. Shutemov static int add_page_for_migration(struct mm_struct *mm, const void __user *p, 2058a49bd4d7SMichal Hocko int node, struct list_head *pagelist, bool migrate_all) 2059742755a1SChristoph Lameter { 2060742755a1SChristoph Lameter struct vm_area_struct *vma; 2061428e106aSKirill A. Shutemov unsigned long addr; 2062742755a1SChristoph Lameter struct page *page; 2063a49bd4d7SMichal Hocko int err; 20649747b9e9SBaolin Wang bool isolated; 2065742755a1SChristoph Lameter 2066d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 2067428e106aSKirill A. Shutemov addr = (unsigned long)untagged_addr_remote(mm, p); 2068428e106aSKirill A. Shutemov 2069742755a1SChristoph Lameter err = -EFAULT; 2070cb1c37b1SMiaohe Lin vma = vma_lookup(mm, addr); 2071cb1c37b1SMiaohe Lin if (!vma || !vma_migratable(vma)) 2072a49bd4d7SMichal Hocko goto out; 2073742755a1SChristoph Lameter 2074d899844eSKirill A. Shutemov /* FOLL_DUMP to ignore special (like zero) pages */ 207587d2762eSMiaohe Lin page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 207689f5b7daSLinus Torvalds 207789f5b7daSLinus Torvalds err = PTR_ERR(page); 207889f5b7daSLinus Torvalds if (IS_ERR(page)) 2079a49bd4d7SMichal Hocko goto out; 208089f5b7daSLinus Torvalds 2081742755a1SChristoph Lameter err = -ENOENT; 2082f7091ed6SHaiyue Wang if (!page) 2083a49bd4d7SMichal Hocko goto out; 2084742755a1SChristoph Lameter 2085f7091ed6SHaiyue Wang if (is_zone_device_page(page)) 2086f7091ed6SHaiyue Wang goto out_putpage; 2087f7091ed6SHaiyue Wang 2088a49bd4d7SMichal Hocko err = 0; 2089a49bd4d7SMichal Hocko if (page_to_nid(page) == node) 2090a49bd4d7SMichal Hocko goto out_putpage; 2091742755a1SChristoph Lameter 2092742755a1SChristoph Lameter err = -EACCES; 2093a49bd4d7SMichal Hocko if (page_mapcount(page) > 1 && !migrate_all) 2094a49bd4d7SMichal Hocko goto out_putpage; 2095742755a1SChristoph Lameter 2096e632a938SNaoya Horiguchi if (PageHuge(page)) { 2097e8db67ebSNaoya Horiguchi if (PageHead(page)) { 20989747b9e9SBaolin Wang isolated = isolate_hugetlb(page_folio(page), pagelist); 20999747b9e9SBaolin Wang err = isolated ? 1 : -EBUSY; 2100e8db67ebSNaoya Horiguchi } 2101a49bd4d7SMichal Hocko } else { 2102a49bd4d7SMichal Hocko struct page *head; 2103e632a938SNaoya Horiguchi 2104e8db67ebSNaoya Horiguchi head = compound_head(page); 2105f7f9c00dSBaolin Wang isolated = isolate_lru_page(head); 2106f7f9c00dSBaolin Wang if (!isolated) { 2107f7f9c00dSBaolin Wang err = -EBUSY; 2108a49bd4d7SMichal Hocko goto out_putpage; 2109f7f9c00dSBaolin Wang } 2110a49bd4d7SMichal Hocko 2111e0153fc2SYang Shi err = 1; 2112a49bd4d7SMichal Hocko list_add_tail(&head->lru, pagelist); 2113e8db67ebSNaoya Horiguchi mod_node_page_state(page_pgdat(head), 21149de4f22aSHuang Ying NR_ISOLATED_ANON + page_is_file_lru(head), 21156c357848SMatthew Wilcox (Oracle) thp_nr_pages(head)); 21166d9c285aSKOSAKI Motohiro } 2117a49bd4d7SMichal Hocko out_putpage: 2118742755a1SChristoph Lameter /* 2119742755a1SChristoph Lameter * Either remove the duplicate refcount from 2120742755a1SChristoph Lameter * isolate_lru_page() or drop the page ref if it was 2121742755a1SChristoph Lameter * not isolated. 2122742755a1SChristoph Lameter */ 2123742755a1SChristoph Lameter put_page(page); 2124a49bd4d7SMichal Hocko out: 2125d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 2126742755a1SChristoph Lameter return err; 2127742755a1SChristoph Lameter } 2128742755a1SChristoph Lameter 21297ca8783aSWei Yang static int move_pages_and_store_status(struct mm_struct *mm, int node, 21307ca8783aSWei Yang struct list_head *pagelist, int __user *status, 21317ca8783aSWei Yang int start, int i, unsigned long nr_pages) 21327ca8783aSWei Yang { 21337ca8783aSWei Yang int err; 21347ca8783aSWei Yang 21355d7ae891SWei Yang if (list_empty(pagelist)) 21365d7ae891SWei Yang return 0; 21375d7ae891SWei Yang 21387ca8783aSWei Yang err = do_move_pages_to_node(mm, pagelist, node); 21397ca8783aSWei Yang if (err) { 21407ca8783aSWei Yang /* 21417ca8783aSWei Yang * Positive err means the number of failed 21427ca8783aSWei Yang * pages to migrate. Since we are going to 21437ca8783aSWei Yang * abort and return the number of non-migrated 2144ab9dd4f8SLong Li * pages, so need to include the rest of the 21457ca8783aSWei Yang * nr_pages that have not been attempted as 21467ca8783aSWei Yang * well. 21477ca8783aSWei Yang */ 21487ca8783aSWei Yang if (err > 0) 2149a7504ed1SHuang Ying err += nr_pages - i; 21507ca8783aSWei Yang return err; 21517ca8783aSWei Yang } 21527ca8783aSWei Yang return store_status(status, start, node, i - start); 21537ca8783aSWei Yang } 21547ca8783aSWei Yang 2155742755a1SChristoph Lameter /* 21565e9a0f02SBrice Goglin * Migrate an array of page address onto an array of nodes and fill 21575e9a0f02SBrice Goglin * the corresponding array of status. 21585e9a0f02SBrice Goglin */ 21593268c63eSChristoph Lameter static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, 21605e9a0f02SBrice Goglin unsigned long nr_pages, 21615e9a0f02SBrice Goglin const void __user * __user *pages, 21625e9a0f02SBrice Goglin const int __user *nodes, 21635e9a0f02SBrice Goglin int __user *status, int flags) 21645e9a0f02SBrice Goglin { 2165a49bd4d7SMichal Hocko int current_node = NUMA_NO_NODE; 2166a49bd4d7SMichal Hocko LIST_HEAD(pagelist); 2167a49bd4d7SMichal Hocko int start, i; 2168a49bd4d7SMichal Hocko int err = 0, err1; 216935282a2dSBrice Goglin 2170361a2a22SMinchan Kim lru_cache_disable(); 217135282a2dSBrice Goglin 2172a49bd4d7SMichal Hocko for (i = start = 0; i < nr_pages; i++) { 21735e9a0f02SBrice Goglin const void __user *p; 21745e9a0f02SBrice Goglin int node; 21755e9a0f02SBrice Goglin 21763140a227SBrice Goglin err = -EFAULT; 2177a49bd4d7SMichal Hocko if (get_user(p, pages + i)) 2178a49bd4d7SMichal Hocko goto out_flush; 2179a49bd4d7SMichal Hocko if (get_user(node, nodes + i)) 2180a49bd4d7SMichal Hocko goto out_flush; 21815e9a0f02SBrice Goglin 21825e9a0f02SBrice Goglin err = -ENODEV; 21836f5a55f1SLinus Torvalds if (node < 0 || node >= MAX_NUMNODES) 2184a49bd4d7SMichal Hocko goto out_flush; 2185389162c2SLai Jiangshan if (!node_state(node, N_MEMORY)) 2186a49bd4d7SMichal Hocko goto out_flush; 21875e9a0f02SBrice Goglin 21885e9a0f02SBrice Goglin err = -EACCES; 21895e9a0f02SBrice Goglin if (!node_isset(node, task_nodes)) 2190a49bd4d7SMichal Hocko goto out_flush; 21915e9a0f02SBrice Goglin 2192a49bd4d7SMichal Hocko if (current_node == NUMA_NO_NODE) { 2193a49bd4d7SMichal Hocko current_node = node; 2194a49bd4d7SMichal Hocko start = i; 2195a49bd4d7SMichal Hocko } else if (node != current_node) { 21967ca8783aSWei Yang err = move_pages_and_store_status(mm, current_node, 21977ca8783aSWei Yang &pagelist, status, start, i, nr_pages); 2198a49bd4d7SMichal Hocko if (err) 2199a49bd4d7SMichal Hocko goto out; 2200a49bd4d7SMichal Hocko start = i; 2201a49bd4d7SMichal Hocko current_node = node; 22025e9a0f02SBrice Goglin } 22035e9a0f02SBrice Goglin 2204a49bd4d7SMichal Hocko /* 2205a49bd4d7SMichal Hocko * Errors in the page lookup or isolation are not fatal and we simply 2206a49bd4d7SMichal Hocko * report them via status 2207a49bd4d7SMichal Hocko */ 2208428e106aSKirill A. Shutemov err = add_page_for_migration(mm, p, current_node, &pagelist, 2209428e106aSKirill A. Shutemov flags & MPOL_MF_MOVE_ALL); 2210e0153fc2SYang Shi 2211d08221a0SWei Yang if (err > 0) { 2212e0153fc2SYang Shi /* The page is successfully queued for migration */ 2213e0153fc2SYang Shi continue; 2214e0153fc2SYang Shi } 22153140a227SBrice Goglin 2216d08221a0SWei Yang /* 221765462462SJohn Hubbard * The move_pages() man page does not have an -EEXIST choice, so 221865462462SJohn Hubbard * use -EFAULT instead. 221965462462SJohn Hubbard */ 222065462462SJohn Hubbard if (err == -EEXIST) 222165462462SJohn Hubbard err = -EFAULT; 222265462462SJohn Hubbard 222365462462SJohn Hubbard /* 2224d08221a0SWei Yang * If the page is already on the target node (!err), store the 2225d08221a0SWei Yang * node, otherwise, store the err. 2226d08221a0SWei Yang */ 2227d08221a0SWei Yang err = store_status(status, i, err ? : current_node, 1); 2228a49bd4d7SMichal Hocko if (err) 2229a49bd4d7SMichal Hocko goto out_flush; 22303140a227SBrice Goglin 22317ca8783aSWei Yang err = move_pages_and_store_status(mm, current_node, &pagelist, 22327ca8783aSWei Yang status, start, i, nr_pages); 2233a7504ed1SHuang Ying if (err) { 2234a7504ed1SHuang Ying /* We have accounted for page i */ 2235a7504ed1SHuang Ying if (err > 0) 2236a7504ed1SHuang Ying err--; 2237a49bd4d7SMichal Hocko goto out; 2238a7504ed1SHuang Ying } 2239a49bd4d7SMichal Hocko current_node = NUMA_NO_NODE; 22403140a227SBrice Goglin } 2241a49bd4d7SMichal Hocko out_flush: 2242a49bd4d7SMichal Hocko /* Make sure we do not overwrite the existing error */ 22437ca8783aSWei Yang err1 = move_pages_and_store_status(mm, current_node, &pagelist, 22447ca8783aSWei Yang status, start, i, nr_pages); 2245dfe9aa23SWei Yang if (err >= 0) 2246a49bd4d7SMichal Hocko err = err1; 22475e9a0f02SBrice Goglin out: 2248361a2a22SMinchan Kim lru_cache_enable(); 22495e9a0f02SBrice Goglin return err; 22505e9a0f02SBrice Goglin } 22515e9a0f02SBrice Goglin 22525e9a0f02SBrice Goglin /* 22532f007e74SBrice Goglin * Determine the nodes of an array of pages and store it in an array of status. 2254742755a1SChristoph Lameter */ 225580bba129SBrice Goglin static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, 225680bba129SBrice Goglin const void __user **pages, int *status) 2257742755a1SChristoph Lameter { 22582f007e74SBrice Goglin unsigned long i; 2259742755a1SChristoph Lameter 2260d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 22612f007e74SBrice Goglin 22622f007e74SBrice Goglin for (i = 0; i < nr_pages; i++) { 226380bba129SBrice Goglin unsigned long addr = (unsigned long)(*pages); 22642f007e74SBrice Goglin struct vm_area_struct *vma; 22652f007e74SBrice Goglin struct page *page; 2266c095adbcSKOSAKI Motohiro int err = -EFAULT; 22672f007e74SBrice Goglin 2268059b8b48SLiam Howlett vma = vma_lookup(mm, addr); 2269059b8b48SLiam Howlett if (!vma) 2270742755a1SChristoph Lameter goto set_status; 2271742755a1SChristoph Lameter 2272d899844eSKirill A. Shutemov /* FOLL_DUMP to ignore special (like zero) pages */ 227316fd6b31SBaolin Wang page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 227489f5b7daSLinus Torvalds 227589f5b7daSLinus Torvalds err = PTR_ERR(page); 227689f5b7daSLinus Torvalds if (IS_ERR(page)) 227789f5b7daSLinus Torvalds goto set_status; 227889f5b7daSLinus Torvalds 2279f7091ed6SHaiyue Wang err = -ENOENT; 2280f7091ed6SHaiyue Wang if (!page) 2281f7091ed6SHaiyue Wang goto set_status; 2282f7091ed6SHaiyue Wang 2283f7091ed6SHaiyue Wang if (!is_zone_device_page(page)) 22844cd61484SMiaohe Lin err = page_to_nid(page); 2285f7091ed6SHaiyue Wang 22864cd61484SMiaohe Lin put_page(page); 2287742755a1SChristoph Lameter set_status: 228880bba129SBrice Goglin *status = err; 228980bba129SBrice Goglin 229080bba129SBrice Goglin pages++; 229180bba129SBrice Goglin status++; 229280bba129SBrice Goglin } 229380bba129SBrice Goglin 2294d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 229580bba129SBrice Goglin } 229680bba129SBrice Goglin 22975b1b561bSArnd Bergmann static int get_compat_pages_array(const void __user *chunk_pages[], 22985b1b561bSArnd Bergmann const void __user * __user *pages, 22995b1b561bSArnd Bergmann unsigned long chunk_nr) 23005b1b561bSArnd Bergmann { 23015b1b561bSArnd Bergmann compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages; 23025b1b561bSArnd Bergmann compat_uptr_t p; 23035b1b561bSArnd Bergmann int i; 23045b1b561bSArnd Bergmann 23055b1b561bSArnd Bergmann for (i = 0; i < chunk_nr; i++) { 23065b1b561bSArnd Bergmann if (get_user(p, pages32 + i)) 23075b1b561bSArnd Bergmann return -EFAULT; 23085b1b561bSArnd Bergmann chunk_pages[i] = compat_ptr(p); 23095b1b561bSArnd Bergmann } 23105b1b561bSArnd Bergmann 23115b1b561bSArnd Bergmann return 0; 23125b1b561bSArnd Bergmann } 23135b1b561bSArnd Bergmann 231480bba129SBrice Goglin /* 231580bba129SBrice Goglin * Determine the nodes of a user array of pages and store it in 231680bba129SBrice Goglin * a user array of status. 231780bba129SBrice Goglin */ 231880bba129SBrice Goglin static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, 231980bba129SBrice Goglin const void __user * __user *pages, 232080bba129SBrice Goglin int __user *status) 232180bba129SBrice Goglin { 23223eefb826SMiaohe Lin #define DO_PAGES_STAT_CHUNK_NR 16UL 232380bba129SBrice Goglin const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; 232480bba129SBrice Goglin int chunk_status[DO_PAGES_STAT_CHUNK_NR]; 232580bba129SBrice Goglin 232687b8d1adSH. Peter Anvin while (nr_pages) { 23273eefb826SMiaohe Lin unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR); 232887b8d1adSH. Peter Anvin 23295b1b561bSArnd Bergmann if (in_compat_syscall()) { 23305b1b561bSArnd Bergmann if (get_compat_pages_array(chunk_pages, pages, 23315b1b561bSArnd Bergmann chunk_nr)) 233287b8d1adSH. Peter Anvin break; 23335b1b561bSArnd Bergmann } else { 23345b1b561bSArnd Bergmann if (copy_from_user(chunk_pages, pages, 23355b1b561bSArnd Bergmann chunk_nr * sizeof(*chunk_pages))) 23365b1b561bSArnd Bergmann break; 23375b1b561bSArnd Bergmann } 233880bba129SBrice Goglin 233980bba129SBrice Goglin do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); 234080bba129SBrice Goglin 234187b8d1adSH. Peter Anvin if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) 234287b8d1adSH. Peter Anvin break; 2343742755a1SChristoph Lameter 234487b8d1adSH. Peter Anvin pages += chunk_nr; 234587b8d1adSH. Peter Anvin status += chunk_nr; 234687b8d1adSH. Peter Anvin nr_pages -= chunk_nr; 234787b8d1adSH. Peter Anvin } 234887b8d1adSH. Peter Anvin return nr_pages ? -EFAULT : 0; 2349742755a1SChristoph Lameter } 2350742755a1SChristoph Lameter 23514dc200ceSMiaohe Lin static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes) 23524dc200ceSMiaohe Lin { 23534dc200ceSMiaohe Lin struct task_struct *task; 23544dc200ceSMiaohe Lin struct mm_struct *mm; 23554dc200ceSMiaohe Lin 23564dc200ceSMiaohe Lin /* 23574dc200ceSMiaohe Lin * There is no need to check if current process has the right to modify 23584dc200ceSMiaohe Lin * the specified process when they are same. 23594dc200ceSMiaohe Lin */ 23604dc200ceSMiaohe Lin if (!pid) { 23614dc200ceSMiaohe Lin mmget(current->mm); 23624dc200ceSMiaohe Lin *mem_nodes = cpuset_mems_allowed(current); 23634dc200ceSMiaohe Lin return current->mm; 23644dc200ceSMiaohe Lin } 23654dc200ceSMiaohe Lin 23664dc200ceSMiaohe Lin /* Find the mm_struct */ 23674dc200ceSMiaohe Lin rcu_read_lock(); 23684dc200ceSMiaohe Lin task = find_task_by_vpid(pid); 23694dc200ceSMiaohe Lin if (!task) { 23704dc200ceSMiaohe Lin rcu_read_unlock(); 23714dc200ceSMiaohe Lin return ERR_PTR(-ESRCH); 23724dc200ceSMiaohe Lin } 23734dc200ceSMiaohe Lin get_task_struct(task); 23744dc200ceSMiaohe Lin 23754dc200ceSMiaohe Lin /* 23764dc200ceSMiaohe Lin * Check if this process has the right to modify the specified 23774dc200ceSMiaohe Lin * process. Use the regular "ptrace_may_access()" checks. 23784dc200ceSMiaohe Lin */ 23794dc200ceSMiaohe Lin if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { 23804dc200ceSMiaohe Lin rcu_read_unlock(); 23814dc200ceSMiaohe Lin mm = ERR_PTR(-EPERM); 23824dc200ceSMiaohe Lin goto out; 23834dc200ceSMiaohe Lin } 23844dc200ceSMiaohe Lin rcu_read_unlock(); 23854dc200ceSMiaohe Lin 23864dc200ceSMiaohe Lin mm = ERR_PTR(security_task_movememory(task)); 23874dc200ceSMiaohe Lin if (IS_ERR(mm)) 23884dc200ceSMiaohe Lin goto out; 23894dc200ceSMiaohe Lin *mem_nodes = cpuset_mems_allowed(task); 23904dc200ceSMiaohe Lin mm = get_task_mm(task); 23914dc200ceSMiaohe Lin out: 23924dc200ceSMiaohe Lin put_task_struct(task); 23934dc200ceSMiaohe Lin if (!mm) 23944dc200ceSMiaohe Lin mm = ERR_PTR(-EINVAL); 23954dc200ceSMiaohe Lin return mm; 23964dc200ceSMiaohe Lin } 23974dc200ceSMiaohe Lin 2398742755a1SChristoph Lameter /* 2399742755a1SChristoph Lameter * Move a list of pages in the address space of the currently executing 2400742755a1SChristoph Lameter * process. 2401742755a1SChristoph Lameter */ 24027addf443SDominik Brodowski static int kernel_move_pages(pid_t pid, unsigned long nr_pages, 24037addf443SDominik Brodowski const void __user * __user *pages, 24047addf443SDominik Brodowski const int __user *nodes, 24057addf443SDominik Brodowski int __user *status, int flags) 2406742755a1SChristoph Lameter { 2407742755a1SChristoph Lameter struct mm_struct *mm; 24085e9a0f02SBrice Goglin int err; 24093268c63eSChristoph Lameter nodemask_t task_nodes; 2410742755a1SChristoph Lameter 2411742755a1SChristoph Lameter /* Check flags */ 2412742755a1SChristoph Lameter if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) 2413742755a1SChristoph Lameter return -EINVAL; 2414742755a1SChristoph Lameter 2415742755a1SChristoph Lameter if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) 2416742755a1SChristoph Lameter return -EPERM; 2417742755a1SChristoph Lameter 24184dc200ceSMiaohe Lin mm = find_mm_struct(pid, &task_nodes); 24194dc200ceSMiaohe Lin if (IS_ERR(mm)) 24204dc200ceSMiaohe Lin return PTR_ERR(mm); 24216e8b09eaSSasha Levin 24223268c63eSChristoph Lameter if (nodes) 24233268c63eSChristoph Lameter err = do_pages_move(mm, task_nodes, nr_pages, pages, 24243268c63eSChristoph Lameter nodes, status, flags); 24253268c63eSChristoph Lameter else 24265e9a0f02SBrice Goglin err = do_pages_stat(mm, nr_pages, pages, status); 24273268c63eSChristoph Lameter 24283268c63eSChristoph Lameter mmput(mm); 24293268c63eSChristoph Lameter return err; 2430742755a1SChristoph Lameter } 2431742755a1SChristoph Lameter 24327addf443SDominik Brodowski SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, 24337addf443SDominik Brodowski const void __user * __user *, pages, 24347addf443SDominik Brodowski const int __user *, nodes, 24357addf443SDominik Brodowski int __user *, status, int, flags) 24367addf443SDominik Brodowski { 24377addf443SDominik Brodowski return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); 24387addf443SDominik Brodowski } 24397addf443SDominik Brodowski 24407039e1dbSPeter Zijlstra #ifdef CONFIG_NUMA_BALANCING 24417039e1dbSPeter Zijlstra /* 24427039e1dbSPeter Zijlstra * Returns true if this is a safe migration target node for misplaced NUMA 2443bc53008eSWei Yang * pages. Currently it only checks the watermarks which is crude. 24447039e1dbSPeter Zijlstra */ 24457039e1dbSPeter Zijlstra static bool migrate_balanced_pgdat(struct pglist_data *pgdat, 24463abef4e6SMel Gorman unsigned long nr_migrate_pages) 24477039e1dbSPeter Zijlstra { 24487039e1dbSPeter Zijlstra int z; 2449599d0c95SMel Gorman 24507039e1dbSPeter Zijlstra for (z = pgdat->nr_zones - 1; z >= 0; z--) { 24517039e1dbSPeter Zijlstra struct zone *zone = pgdat->node_zones + z; 24527039e1dbSPeter Zijlstra 2453bc53008eSWei Yang if (!managed_zone(zone)) 24547039e1dbSPeter Zijlstra continue; 24557039e1dbSPeter Zijlstra 24567039e1dbSPeter Zijlstra /* Avoid waking kswapd by allocating pages_to_migrate pages. */ 24577039e1dbSPeter Zijlstra if (!zone_watermark_ok(zone, 0, 24587039e1dbSPeter Zijlstra high_wmark_pages(zone) + 24597039e1dbSPeter Zijlstra nr_migrate_pages, 2460bfe9d006SHuang Ying ZONE_MOVABLE, 0)) 24617039e1dbSPeter Zijlstra continue; 24627039e1dbSPeter Zijlstra return true; 24637039e1dbSPeter Zijlstra } 24647039e1dbSPeter Zijlstra return false; 24657039e1dbSPeter Zijlstra } 24667039e1dbSPeter Zijlstra 24674e096ae1SMatthew Wilcox (Oracle) static struct folio *alloc_misplaced_dst_folio(struct folio *src, 2468666feb21SMichal Hocko unsigned long data) 24697039e1dbSPeter Zijlstra { 24707039e1dbSPeter Zijlstra int nid = (int) data; 24714e096ae1SMatthew Wilcox (Oracle) int order = folio_order(src); 2472c185e494SMatthew Wilcox (Oracle) gfp_t gfp = __GFP_THISNODE; 24737039e1dbSPeter Zijlstra 2474c185e494SMatthew Wilcox (Oracle) if (order > 0) 2475c185e494SMatthew Wilcox (Oracle) gfp |= GFP_TRANSHUGE_LIGHT; 2476c185e494SMatthew Wilcox (Oracle) else { 2477c185e494SMatthew Wilcox (Oracle) gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY | 2478c185e494SMatthew Wilcox (Oracle) __GFP_NOWARN; 2479c185e494SMatthew Wilcox (Oracle) gfp &= ~__GFP_RECLAIM; 24807039e1dbSPeter Zijlstra } 24814e096ae1SMatthew Wilcox (Oracle) return __folio_alloc_node(gfp, order, nid); 2482c5b5a3ddSYang Shi } 2483c5b5a3ddSYang Shi 24842ac9e99fSKefeng Wang static int numamigrate_isolate_folio(pg_data_t *pgdat, struct folio *folio) 2485b32967ffSMel Gorman { 24862ac9e99fSKefeng Wang int nr_pages = folio_nr_pages(folio); 2487b32967ffSMel Gorman 2488b32967ffSMel Gorman /* Avoid migrating to a node that is nearly full */ 2489c574bbe9SHuang Ying if (!migrate_balanced_pgdat(pgdat, nr_pages)) { 2490c574bbe9SHuang Ying int z; 2491c574bbe9SHuang Ying 2492c574bbe9SHuang Ying if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)) 2493340ef390SHugh Dickins return 0; 2494c574bbe9SHuang Ying for (z = pgdat->nr_zones - 1; z >= 0; z--) { 2495bc53008eSWei Yang if (managed_zone(pgdat->node_zones + z)) 2496c574bbe9SHuang Ying break; 2497c574bbe9SHuang Ying } 24982ac9e99fSKefeng Wang wakeup_kswapd(pgdat->node_zones + z, 0, 24992ac9e99fSKefeng Wang folio_order(folio), ZONE_MOVABLE); 2500c574bbe9SHuang Ying return 0; 2501c574bbe9SHuang Ying } 2502b32967ffSMel Gorman 25032ac9e99fSKefeng Wang if (!folio_isolate_lru(folio)) 2504340ef390SHugh Dickins return 0; 2505340ef390SHugh Dickins 25062ac9e99fSKefeng Wang node_stat_mod_folio(folio, NR_ISOLATED_ANON + folio_is_file_lru(folio), 25072b9b624fSBaolin Wang nr_pages); 2508b32967ffSMel Gorman 2509b32967ffSMel Gorman /* 25102ac9e99fSKefeng Wang * Isolating the folio has taken another reference, so the 25112ac9e99fSKefeng Wang * caller's reference can be safely dropped without the folio 2512340ef390SHugh Dickins * disappearing underneath us during migration. 2513b32967ffSMel Gorman */ 25142ac9e99fSKefeng Wang folio_put(folio); 2515340ef390SHugh Dickins return 1; 2516b32967ffSMel Gorman } 2517b32967ffSMel Gorman 2518a8f60772SMel Gorman /* 2519*73eab3caSKefeng Wang * Attempt to migrate a misplaced folio to the specified destination 25207039e1dbSPeter Zijlstra * node. Caller is expected to have an elevated reference count on 2521*73eab3caSKefeng Wang * the folio that will be dropped by this function before returning. 25227039e1dbSPeter Zijlstra */ 2523*73eab3caSKefeng Wang int migrate_misplaced_folio(struct folio *folio, struct vm_area_struct *vma, 25241bc115d8SMel Gorman int node) 25257039e1dbSPeter Zijlstra { 2526a8f60772SMel Gorman pg_data_t *pgdat = NODE_DATA(node); 2527340ef390SHugh Dickins int isolated; 2528b32967ffSMel Gorman int nr_remaining; 2529e39bb6beSHuang Ying unsigned int nr_succeeded; 25307039e1dbSPeter Zijlstra LIST_HEAD(migratepages); 2531*73eab3caSKefeng Wang int nr_pages = folio_nr_pages(folio); 2532c5b5a3ddSYang Shi 2533c5b5a3ddSYang Shi /* 2534*73eab3caSKefeng Wang * Don't migrate file folios that are mapped in multiple processes 25351bc115d8SMel Gorman * with execute permissions as they are probably shared libraries. 2536*73eab3caSKefeng Wang * To check if the folio is shared, ideally we want to make sure 2537*73eab3caSKefeng Wang * every page is mapped to the same process. Doing that is very 2538*73eab3caSKefeng Wang * expensive, so check the estimated mapcount of the folio instead. 25397039e1dbSPeter Zijlstra */ 2540*73eab3caSKefeng Wang if (folio_estimated_sharers(folio) != 1 && folio_is_file_lru(folio) && 25417ee820eeSMiaohe Lin (vma->vm_flags & VM_EXEC)) 25427039e1dbSPeter Zijlstra goto out; 25437039e1dbSPeter Zijlstra 2544a8f60772SMel Gorman /* 2545*73eab3caSKefeng Wang * Also do not migrate dirty folios as not all filesystems can move 2546*73eab3caSKefeng Wang * dirty folios in MIGRATE_ASYNC mode which is a waste of cycles. 254709a913a7SMel Gorman */ 2548*73eab3caSKefeng Wang if (folio_is_file_lru(folio) && folio_test_dirty(folio)) 254909a913a7SMel Gorman goto out; 255009a913a7SMel Gorman 2551*73eab3caSKefeng Wang isolated = numamigrate_isolate_folio(pgdat, folio); 2552b32967ffSMel Gorman if (!isolated) 25537039e1dbSPeter Zijlstra goto out; 25547039e1dbSPeter Zijlstra 2555*73eab3caSKefeng Wang list_add(&folio->lru, &migratepages); 25564e096ae1SMatthew Wilcox (Oracle) nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_folio, 2557c185e494SMatthew Wilcox (Oracle) NULL, node, MIGRATE_ASYNC, 2558c185e494SMatthew Wilcox (Oracle) MR_NUMA_MISPLACED, &nr_succeeded); 25597039e1dbSPeter Zijlstra if (nr_remaining) { 256059c82b70SJoonsoo Kim if (!list_empty(&migratepages)) { 2561*73eab3caSKefeng Wang list_del(&folio->lru); 2562*73eab3caSKefeng Wang node_stat_mod_folio(folio, NR_ISOLATED_ANON + 2563*73eab3caSKefeng Wang folio_is_file_lru(folio), -nr_pages); 2564*73eab3caSKefeng Wang folio_putback_lru(folio); 256559c82b70SJoonsoo Kim } 25667039e1dbSPeter Zijlstra isolated = 0; 2567e39bb6beSHuang Ying } 2568e39bb6beSHuang Ying if (nr_succeeded) { 2569e39bb6beSHuang Ying count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded); 2570*73eab3caSKefeng Wang if (!node_is_toptier(folio_nid(folio)) && node_is_toptier(node)) 2571e39bb6beSHuang Ying mod_node_page_state(pgdat, PGPROMOTE_SUCCESS, 2572e39bb6beSHuang Ying nr_succeeded); 2573e39bb6beSHuang Ying } 25747039e1dbSPeter Zijlstra BUG_ON(!list_empty(&migratepages)); 25757039e1dbSPeter Zijlstra return isolated; 2576340ef390SHugh Dickins 2577340ef390SHugh Dickins out: 2578*73eab3caSKefeng Wang folio_put(folio); 2579340ef390SHugh Dickins return 0; 25807039e1dbSPeter Zijlstra } 2581220018d3SMel Gorman #endif /* CONFIG_NUMA_BALANCING */ 25827d6e2d96SOscar Salvador #endif /* CONFIG_NUMA */ 2583