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 1607e2a5e5aSKefeng Wang * __folio_test_movable because LRU folio's mapping cannot 1617e2a5e5aSKefeng Wang * have 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); 408*fc346d0aSCharan Teja Kalla long entries, i; 4098763cb45SJérôme Glisse 4106c5240aeSChristoph Lameter if (!mapping) { 4110e8c7d0fSChristoph Lameter /* Anonymous page without mapping */ 4123417013eSMatthew Wilcox (Oracle) if (folio_ref_count(folio) != expected_count) 4136c5240aeSChristoph Lameter return -EAGAIN; 414cf4b769aSHugh Dickins 415cf4b769aSHugh Dickins /* No turning back from here */ 4163417013eSMatthew Wilcox (Oracle) newfolio->index = folio->index; 4173417013eSMatthew Wilcox (Oracle) newfolio->mapping = folio->mapping; 4183417013eSMatthew Wilcox (Oracle) if (folio_test_swapbacked(folio)) 4193417013eSMatthew Wilcox (Oracle) __folio_set_swapbacked(newfolio); 420cf4b769aSHugh Dickins 42178bd5209SRafael Aquini return MIGRATEPAGE_SUCCESS; 4226c5240aeSChristoph Lameter } 4236c5240aeSChristoph Lameter 4243417013eSMatthew Wilcox (Oracle) oldzone = folio_zone(folio); 4253417013eSMatthew Wilcox (Oracle) newzone = folio_zone(newfolio); 42642cb14b1SHugh Dickins 42789eb946aSMatthew Wilcox xas_lock_irq(&xas); 4283417013eSMatthew Wilcox (Oracle) if (!folio_ref_freeze(folio, expected_count)) { 42989eb946aSMatthew Wilcox xas_unlock_irq(&xas); 430e286781dSNick Piggin return -EAGAIN; 431e286781dSNick Piggin } 432e286781dSNick Piggin 433b20a3503SChristoph Lameter /* 4343417013eSMatthew Wilcox (Oracle) * Now we know that no one else is looking at the folio: 435cf4b769aSHugh Dickins * no turning back from here. 436b20a3503SChristoph Lameter */ 4373417013eSMatthew Wilcox (Oracle) newfolio->index = folio->index; 4383417013eSMatthew Wilcox (Oracle) newfolio->mapping = folio->mapping; 4393417013eSMatthew Wilcox (Oracle) folio_ref_add(newfolio, nr); /* add cache reference */ 4403417013eSMatthew Wilcox (Oracle) if (folio_test_swapbacked(folio)) { 4413417013eSMatthew Wilcox (Oracle) __folio_set_swapbacked(newfolio); 4423417013eSMatthew Wilcox (Oracle) if (folio_test_swapcache(folio)) { 4433417013eSMatthew Wilcox (Oracle) folio_set_swapcache(newfolio); 4443417013eSMatthew Wilcox (Oracle) newfolio->private = folio_get_private(folio); 445b20a3503SChristoph Lameter } 446*fc346d0aSCharan Teja Kalla entries = nr; 4476326fec1SNicholas Piggin } else { 4483417013eSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio); 449*fc346d0aSCharan Teja Kalla entries = 1; 4506326fec1SNicholas Piggin } 451b20a3503SChristoph Lameter 45242cb14b1SHugh Dickins /* Move dirty while page refs frozen and newpage not yet exposed */ 4533417013eSMatthew Wilcox (Oracle) dirty = folio_test_dirty(folio); 45442cb14b1SHugh Dickins if (dirty) { 4553417013eSMatthew Wilcox (Oracle) folio_clear_dirty(folio); 4563417013eSMatthew Wilcox (Oracle) folio_set_dirty(newfolio); 45742cb14b1SHugh Dickins } 45842cb14b1SHugh Dickins 459*fc346d0aSCharan Teja Kalla /* Swap cache still stores N entries instead of a high-order entry */ 460*fc346d0aSCharan Teja Kalla for (i = 0; i < entries; i++) { 4613417013eSMatthew Wilcox (Oracle) xas_store(&xas, newfolio); 462*fc346d0aSCharan Teja Kalla xas_next(&xas); 463*fc346d0aSCharan Teja Kalla } 4647cf9c2c7SNick Piggin 4657cf9c2c7SNick Piggin /* 466937a94c9SJacobo Giralt * Drop cache reference from old page by unfreezing 467937a94c9SJacobo Giralt * to one less reference. 4687cf9c2c7SNick Piggin * We know this isn't the last reference. 4697cf9c2c7SNick Piggin */ 4703417013eSMatthew Wilcox (Oracle) folio_ref_unfreeze(folio, expected_count - nr); 4717cf9c2c7SNick Piggin 47289eb946aSMatthew Wilcox xas_unlock(&xas); 47342cb14b1SHugh Dickins /* Leave irq disabled to prevent preemption while updating stats */ 47442cb14b1SHugh Dickins 4750e8c7d0fSChristoph Lameter /* 4760e8c7d0fSChristoph Lameter * If moved to a different zone then also account 4770e8c7d0fSChristoph Lameter * the page for that zone. Other VM counters will be 4780e8c7d0fSChristoph Lameter * taken care of when we establish references to the 4790e8c7d0fSChristoph Lameter * new page and drop references to the old page. 4800e8c7d0fSChristoph Lameter * 4810e8c7d0fSChristoph Lameter * Note that anonymous pages are accounted for 4824b9d0fabSMel Gorman * via NR_FILE_PAGES and NR_ANON_MAPPED if they 4830e8c7d0fSChristoph Lameter * are mapped to swap space. 4840e8c7d0fSChristoph Lameter */ 48542cb14b1SHugh Dickins if (newzone != oldzone) { 4860d1c2072SJohannes Weiner struct lruvec *old_lruvec, *new_lruvec; 4870d1c2072SJohannes Weiner struct mem_cgroup *memcg; 4880d1c2072SJohannes Weiner 4893417013eSMatthew Wilcox (Oracle) memcg = folio_memcg(folio); 4900d1c2072SJohannes Weiner old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat); 4910d1c2072SJohannes Weiner new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat); 4920d1c2072SJohannes Weiner 4935c447d27SShakeel Butt __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr); 4945c447d27SShakeel Butt __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr); 4953417013eSMatthew Wilcox (Oracle) if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) { 4965c447d27SShakeel Butt __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr); 4975c447d27SShakeel Butt __mod_lruvec_state(new_lruvec, NR_SHMEM, nr); 4980b52c420SJan Glauber 4990b52c420SJan Glauber if (folio_test_pmd_mappable(folio)) { 5000b52c420SJan Glauber __mod_lruvec_state(old_lruvec, NR_SHMEM_THPS, -nr); 5010b52c420SJan Glauber __mod_lruvec_state(new_lruvec, NR_SHMEM_THPS, nr); 5020b52c420SJan Glauber } 5034b02108aSKOSAKI Motohiro } 504b6038942SShakeel Butt #ifdef CONFIG_SWAP 5053417013eSMatthew Wilcox (Oracle) if (folio_test_swapcache(folio)) { 506b6038942SShakeel Butt __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr); 507b6038942SShakeel Butt __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr); 508b6038942SShakeel Butt } 509b6038942SShakeel Butt #endif 510f56753acSChristoph Hellwig if (dirty && mapping_can_writeback(mapping)) { 5115c447d27SShakeel Butt __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr); 5125c447d27SShakeel Butt __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr); 5135c447d27SShakeel Butt __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr); 5145c447d27SShakeel Butt __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr); 51542cb14b1SHugh Dickins } 51642cb14b1SHugh Dickins } 51742cb14b1SHugh Dickins local_irq_enable(); 518b20a3503SChristoph Lameter 51978bd5209SRafael Aquini return MIGRATEPAGE_SUCCESS; 520b20a3503SChristoph Lameter } 5213417013eSMatthew Wilcox (Oracle) EXPORT_SYMBOL(folio_migrate_mapping); 522b20a3503SChristoph Lameter 523b20a3503SChristoph Lameter /* 524290408d4SNaoya Horiguchi * The expected number of remaining references is the same as that 5253417013eSMatthew Wilcox (Oracle) * of folio_migrate_mapping(). 526290408d4SNaoya Horiguchi */ 527290408d4SNaoya Horiguchi int migrate_huge_page_move_mapping(struct address_space *mapping, 528b890ec2aSMatthew Wilcox (Oracle) struct folio *dst, struct folio *src) 529290408d4SNaoya Horiguchi { 530b890ec2aSMatthew Wilcox (Oracle) XA_STATE(xas, &mapping->i_pages, folio_index(src)); 531290408d4SNaoya Horiguchi int expected_count; 532290408d4SNaoya Horiguchi 53389eb946aSMatthew Wilcox xas_lock_irq(&xas); 534a08c7193SSidhartha Kumar expected_count = folio_expected_refs(mapping, src); 535b890ec2aSMatthew Wilcox (Oracle) if (!folio_ref_freeze(src, expected_count)) { 53689eb946aSMatthew Wilcox xas_unlock_irq(&xas); 537290408d4SNaoya Horiguchi return -EAGAIN; 538290408d4SNaoya Horiguchi } 539290408d4SNaoya Horiguchi 540b890ec2aSMatthew Wilcox (Oracle) dst->index = src->index; 541b890ec2aSMatthew Wilcox (Oracle) dst->mapping = src->mapping; 5426a93ca8fSJohannes Weiner 543a08c7193SSidhartha Kumar folio_ref_add(dst, folio_nr_pages(dst)); 544290408d4SNaoya Horiguchi 545b890ec2aSMatthew Wilcox (Oracle) xas_store(&xas, dst); 546290408d4SNaoya Horiguchi 547a08c7193SSidhartha Kumar folio_ref_unfreeze(src, expected_count - folio_nr_pages(src)); 548290408d4SNaoya Horiguchi 54989eb946aSMatthew Wilcox xas_unlock_irq(&xas); 5506a93ca8fSJohannes Weiner 55178bd5209SRafael Aquini return MIGRATEPAGE_SUCCESS; 552290408d4SNaoya Horiguchi } 553290408d4SNaoya Horiguchi 554290408d4SNaoya Horiguchi /* 55519138349SMatthew Wilcox (Oracle) * Copy the flags and some other ancillary information 556b20a3503SChristoph Lameter */ 55719138349SMatthew Wilcox (Oracle) void folio_migrate_flags(struct folio *newfolio, struct folio *folio) 558b20a3503SChristoph Lameter { 5597851a45cSRik van Riel int cpupid; 5607851a45cSRik van Riel 56119138349SMatthew Wilcox (Oracle) if (folio_test_error(folio)) 56219138349SMatthew Wilcox (Oracle) folio_set_error(newfolio); 56319138349SMatthew Wilcox (Oracle) if (folio_test_referenced(folio)) 56419138349SMatthew Wilcox (Oracle) folio_set_referenced(newfolio); 56519138349SMatthew Wilcox (Oracle) if (folio_test_uptodate(folio)) 56619138349SMatthew Wilcox (Oracle) folio_mark_uptodate(newfolio); 56719138349SMatthew Wilcox (Oracle) if (folio_test_clear_active(folio)) { 56819138349SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio); 56919138349SMatthew Wilcox (Oracle) folio_set_active(newfolio); 57019138349SMatthew Wilcox (Oracle) } else if (folio_test_clear_unevictable(folio)) 57119138349SMatthew Wilcox (Oracle) folio_set_unevictable(newfolio); 57219138349SMatthew Wilcox (Oracle) if (folio_test_workingset(folio)) 57319138349SMatthew Wilcox (Oracle) folio_set_workingset(newfolio); 57419138349SMatthew Wilcox (Oracle) if (folio_test_checked(folio)) 57519138349SMatthew Wilcox (Oracle) folio_set_checked(newfolio); 5766c287605SDavid Hildenbrand /* 5776c287605SDavid Hildenbrand * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via 5786c287605SDavid Hildenbrand * migration entries. We can still have PG_anon_exclusive set on an 5796c287605SDavid Hildenbrand * effectively unmapped and unreferenced first sub-pages of an 5806c287605SDavid Hildenbrand * anonymous THP: we can simply copy it here via PG_mappedtodisk. 5816c287605SDavid Hildenbrand */ 58219138349SMatthew Wilcox (Oracle) if (folio_test_mappedtodisk(folio)) 58319138349SMatthew Wilcox (Oracle) folio_set_mappedtodisk(newfolio); 584b20a3503SChristoph Lameter 5853417013eSMatthew Wilcox (Oracle) /* Move dirty on pages not done by folio_migrate_mapping() */ 58619138349SMatthew Wilcox (Oracle) if (folio_test_dirty(folio)) 58719138349SMatthew Wilcox (Oracle) folio_set_dirty(newfolio); 588b20a3503SChristoph Lameter 58919138349SMatthew Wilcox (Oracle) if (folio_test_young(folio)) 59019138349SMatthew Wilcox (Oracle) folio_set_young(newfolio); 59119138349SMatthew Wilcox (Oracle) if (folio_test_idle(folio)) 59219138349SMatthew Wilcox (Oracle) folio_set_idle(newfolio); 59333c3fc71SVladimir Davydov 5947851a45cSRik van Riel /* 5957851a45cSRik van Riel * Copy NUMA information to the new page, to prevent over-eager 5967851a45cSRik van Riel * future migrations of this same page. 5977851a45cSRik van Riel */ 5984e694fe4SKefeng Wang cpupid = folio_xchg_last_cpupid(folio, -1); 59933024536SHuang Ying /* 60033024536SHuang Ying * For memory tiering mode, when migrate between slow and fast 60133024536SHuang Ying * memory node, reset cpupid, because that is used to record 60233024536SHuang Ying * page access time in slow memory node. 60333024536SHuang Ying */ 60433024536SHuang Ying if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) { 6054e694fe4SKefeng Wang bool f_toptier = node_is_toptier(folio_nid(folio)); 6064e694fe4SKefeng Wang bool t_toptier = node_is_toptier(folio_nid(newfolio)); 60733024536SHuang Ying 60833024536SHuang Ying if (f_toptier != t_toptier) 60933024536SHuang Ying cpupid = -1; 61033024536SHuang Ying } 6114e694fe4SKefeng Wang folio_xchg_last_cpupid(newfolio, cpupid); 6127851a45cSRik van Riel 61319138349SMatthew Wilcox (Oracle) folio_migrate_ksm(newfolio, folio); 614c8d6553bSHugh Dickins /* 615c8d6553bSHugh Dickins * Please do not reorder this without considering how mm/ksm.c's 616c8d6553bSHugh Dickins * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). 617c8d6553bSHugh Dickins */ 61819138349SMatthew Wilcox (Oracle) if (folio_test_swapcache(folio)) 61919138349SMatthew Wilcox (Oracle) folio_clear_swapcache(folio); 62019138349SMatthew Wilcox (Oracle) folio_clear_private(folio); 621ad2fa371SMuchun Song 622ad2fa371SMuchun Song /* page->private contains hugetlb specific flags */ 62319138349SMatthew Wilcox (Oracle) if (!folio_test_hugetlb(folio)) 62419138349SMatthew Wilcox (Oracle) folio->private = NULL; 625b20a3503SChristoph Lameter 626b20a3503SChristoph Lameter /* 627b20a3503SChristoph Lameter * If any waiters have accumulated on the new page then 628b20a3503SChristoph Lameter * wake them up. 629b20a3503SChristoph Lameter */ 63019138349SMatthew Wilcox (Oracle) if (folio_test_writeback(newfolio)) 63119138349SMatthew Wilcox (Oracle) folio_end_writeback(newfolio); 632d435edcaSVlastimil Babka 6336aeff241SYang Shi /* 6346aeff241SYang Shi * PG_readahead shares the same bit with PG_reclaim. The above 6356aeff241SYang Shi * end_page_writeback() may clear PG_readahead mistakenly, so set the 6366aeff241SYang Shi * bit after that. 6376aeff241SYang Shi */ 63819138349SMatthew Wilcox (Oracle) if (folio_test_readahead(folio)) 63919138349SMatthew Wilcox (Oracle) folio_set_readahead(newfolio); 6406aeff241SYang Shi 64119138349SMatthew Wilcox (Oracle) folio_copy_owner(newfolio, folio); 64274485cf2SJohannes Weiner 643d21bba2bSMatthew Wilcox (Oracle) mem_cgroup_migrate(folio, newfolio); 644b20a3503SChristoph Lameter } 64519138349SMatthew Wilcox (Oracle) EXPORT_SYMBOL(folio_migrate_flags); 6462916ecc0SJérôme Glisse 647715cbfd6SMatthew Wilcox (Oracle) void folio_migrate_copy(struct folio *newfolio, struct folio *folio) 6482916ecc0SJérôme Glisse { 649715cbfd6SMatthew Wilcox (Oracle) folio_copy(newfolio, folio); 650715cbfd6SMatthew Wilcox (Oracle) folio_migrate_flags(newfolio, folio); 6512916ecc0SJérôme Glisse } 652715cbfd6SMatthew Wilcox (Oracle) EXPORT_SYMBOL(folio_migrate_copy); 653b20a3503SChristoph Lameter 6541d8b85ccSChristoph Lameter /************************************************************ 6551d8b85ccSChristoph Lameter * Migration functions 6561d8b85ccSChristoph Lameter ***********************************************************/ 6571d8b85ccSChristoph Lameter 65816ce101dSAlistair Popple int migrate_folio_extra(struct address_space *mapping, struct folio *dst, 65916ce101dSAlistair Popple struct folio *src, enum migrate_mode mode, int extra_count) 66016ce101dSAlistair Popple { 66116ce101dSAlistair Popple int rc; 66216ce101dSAlistair Popple 66316ce101dSAlistair Popple BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */ 66416ce101dSAlistair Popple 66516ce101dSAlistair Popple rc = folio_migrate_mapping(mapping, dst, src, extra_count); 66616ce101dSAlistair Popple 66716ce101dSAlistair Popple if (rc != MIGRATEPAGE_SUCCESS) 66816ce101dSAlistair Popple return rc; 66916ce101dSAlistair Popple 67016ce101dSAlistair Popple if (mode != MIGRATE_SYNC_NO_COPY) 67116ce101dSAlistair Popple folio_migrate_copy(dst, src); 67216ce101dSAlistair Popple else 67316ce101dSAlistair Popple folio_migrate_flags(dst, src); 67416ce101dSAlistair Popple return MIGRATEPAGE_SUCCESS; 67516ce101dSAlistair Popple } 67616ce101dSAlistair Popple 67754184650SMatthew Wilcox (Oracle) /** 67854184650SMatthew Wilcox (Oracle) * migrate_folio() - Simple folio migration. 67954184650SMatthew Wilcox (Oracle) * @mapping: The address_space containing the folio. 68054184650SMatthew Wilcox (Oracle) * @dst: The folio to migrate the data to. 68154184650SMatthew Wilcox (Oracle) * @src: The folio containing the current data. 68254184650SMatthew Wilcox (Oracle) * @mode: How to migrate the page. 683b20a3503SChristoph Lameter * 68454184650SMatthew Wilcox (Oracle) * Common logic to directly migrate a single LRU folio suitable for 68554184650SMatthew Wilcox (Oracle) * folios that do not use PagePrivate/PagePrivate2. 68654184650SMatthew Wilcox (Oracle) * 68754184650SMatthew Wilcox (Oracle) * Folios are locked upon entry and exit. 688b20a3503SChristoph Lameter */ 68954184650SMatthew Wilcox (Oracle) int migrate_folio(struct address_space *mapping, struct folio *dst, 69054184650SMatthew Wilcox (Oracle) struct folio *src, enum migrate_mode mode) 691b20a3503SChristoph Lameter { 69216ce101dSAlistair Popple return migrate_folio_extra(mapping, dst, src, mode, 0); 693b20a3503SChristoph Lameter } 69454184650SMatthew Wilcox (Oracle) EXPORT_SYMBOL(migrate_folio); 695b20a3503SChristoph Lameter 696925c86a1SChristoph Hellwig #ifdef CONFIG_BUFFER_HEAD 69784ade7c1SJan Kara /* Returns true if all buffers are successfully locked */ 69884ade7c1SJan Kara static bool buffer_migrate_lock_buffers(struct buffer_head *head, 69984ade7c1SJan Kara enum migrate_mode mode) 70084ade7c1SJan Kara { 70184ade7c1SJan Kara struct buffer_head *bh = head; 7024bb6dc79SDouglas Anderson struct buffer_head *failed_bh; 70384ade7c1SJan Kara 70484ade7c1SJan Kara do { 7054bb6dc79SDouglas Anderson if (!trylock_buffer(bh)) { 7064bb6dc79SDouglas Anderson if (mode == MIGRATE_ASYNC) 7074bb6dc79SDouglas Anderson goto unlock; 7084bb6dc79SDouglas Anderson if (mode == MIGRATE_SYNC_LIGHT && !buffer_uptodate(bh)) 7094bb6dc79SDouglas Anderson goto unlock; 71084ade7c1SJan Kara lock_buffer(bh); 7114bb6dc79SDouglas Anderson } 71284ade7c1SJan Kara 7134bb6dc79SDouglas Anderson bh = bh->b_this_page; 71484ade7c1SJan Kara } while (bh != head); 71584ade7c1SJan Kara 71684ade7c1SJan Kara return true; 71784ade7c1SJan Kara 7184bb6dc79SDouglas Anderson unlock: 7194bb6dc79SDouglas Anderson /* We failed to lock the buffer and cannot stall. */ 7204bb6dc79SDouglas Anderson failed_bh = bh; 72184ade7c1SJan Kara bh = head; 72284ade7c1SJan Kara while (bh != failed_bh) { 72384ade7c1SJan Kara unlock_buffer(bh); 72484ade7c1SJan Kara bh = bh->b_this_page; 72584ade7c1SJan Kara } 72684ade7c1SJan Kara 7274bb6dc79SDouglas Anderson return false; 72884ade7c1SJan Kara } 72984ade7c1SJan Kara 73067235182SMatthew Wilcox (Oracle) static int __buffer_migrate_folio(struct address_space *mapping, 73167235182SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode, 73289cb0888SJan Kara bool check_refs) 7331d8b85ccSChristoph Lameter { 7341d8b85ccSChristoph Lameter struct buffer_head *bh, *head; 7351d8b85ccSChristoph Lameter int rc; 736cc4f11e6SJan Kara int expected_count; 7371d8b85ccSChristoph Lameter 73867235182SMatthew Wilcox (Oracle) head = folio_buffers(src); 73967235182SMatthew Wilcox (Oracle) if (!head) 74054184650SMatthew Wilcox (Oracle) return migrate_folio(mapping, dst, src, mode); 7411d8b85ccSChristoph Lameter 742cc4f11e6SJan Kara /* Check whether page does not have extra refs before we do more work */ 743108ca835SMatthew Wilcox (Oracle) expected_count = folio_expected_refs(mapping, src); 74467235182SMatthew Wilcox (Oracle) if (folio_ref_count(src) != expected_count) 745cc4f11e6SJan Kara return -EAGAIN; 746cc4f11e6SJan Kara 747cc4f11e6SJan Kara if (!buffer_migrate_lock_buffers(head, mode)) 748cc4f11e6SJan Kara return -EAGAIN; 7491d8b85ccSChristoph Lameter 75089cb0888SJan Kara if (check_refs) { 75189cb0888SJan Kara bool busy; 75289cb0888SJan Kara bool invalidated = false; 75389cb0888SJan Kara 75489cb0888SJan Kara recheck_buffers: 75589cb0888SJan Kara busy = false; 75689cb0888SJan Kara spin_lock(&mapping->private_lock); 75789cb0888SJan Kara bh = head; 75889cb0888SJan Kara do { 75989cb0888SJan Kara if (atomic_read(&bh->b_count)) { 76089cb0888SJan Kara busy = true; 76189cb0888SJan Kara break; 76289cb0888SJan Kara } 76389cb0888SJan Kara bh = bh->b_this_page; 76489cb0888SJan Kara } while (bh != head); 76589cb0888SJan Kara if (busy) { 76689cb0888SJan Kara if (invalidated) { 76789cb0888SJan Kara rc = -EAGAIN; 76889cb0888SJan Kara goto unlock_buffers; 76989cb0888SJan Kara } 770ebdf4de5SJan Kara spin_unlock(&mapping->private_lock); 77189cb0888SJan Kara invalidate_bh_lrus(); 77289cb0888SJan Kara invalidated = true; 77389cb0888SJan Kara goto recheck_buffers; 77489cb0888SJan Kara } 77589cb0888SJan Kara } 77689cb0888SJan Kara 77767235182SMatthew Wilcox (Oracle) rc = folio_migrate_mapping(mapping, dst, src, 0); 77878bd5209SRafael Aquini if (rc != MIGRATEPAGE_SUCCESS) 779cc4f11e6SJan Kara goto unlock_buffers; 7801d8b85ccSChristoph Lameter 78167235182SMatthew Wilcox (Oracle) folio_attach_private(dst, folio_detach_private(src)); 7821d8b85ccSChristoph Lameter 7831d8b85ccSChristoph Lameter bh = head; 7841d8b85ccSChristoph Lameter do { 785d5db4f9dSMatthew Wilcox (Oracle) folio_set_bh(bh, dst, bh_offset(bh)); 7861d8b85ccSChristoph Lameter bh = bh->b_this_page; 7871d8b85ccSChristoph Lameter } while (bh != head); 7881d8b85ccSChristoph Lameter 7892916ecc0SJérôme Glisse if (mode != MIGRATE_SYNC_NO_COPY) 79067235182SMatthew Wilcox (Oracle) folio_migrate_copy(dst, src); 7912916ecc0SJérôme Glisse else 79267235182SMatthew Wilcox (Oracle) folio_migrate_flags(dst, src); 7931d8b85ccSChristoph Lameter 794cc4f11e6SJan Kara rc = MIGRATEPAGE_SUCCESS; 795cc4f11e6SJan Kara unlock_buffers: 796ebdf4de5SJan Kara if (check_refs) 797ebdf4de5SJan Kara spin_unlock(&mapping->private_lock); 7981d8b85ccSChristoph Lameter bh = head; 7991d8b85ccSChristoph Lameter do { 8001d8b85ccSChristoph Lameter unlock_buffer(bh); 8011d8b85ccSChristoph Lameter bh = bh->b_this_page; 8021d8b85ccSChristoph Lameter } while (bh != head); 8031d8b85ccSChristoph Lameter 804cc4f11e6SJan Kara return rc; 8051d8b85ccSChristoph Lameter } 80689cb0888SJan Kara 80767235182SMatthew Wilcox (Oracle) /** 80867235182SMatthew Wilcox (Oracle) * buffer_migrate_folio() - Migration function for folios with buffers. 80967235182SMatthew Wilcox (Oracle) * @mapping: The address space containing @src. 81067235182SMatthew Wilcox (Oracle) * @dst: The folio to migrate to. 81167235182SMatthew Wilcox (Oracle) * @src: The folio to migrate from. 81267235182SMatthew Wilcox (Oracle) * @mode: How to migrate the folio. 81367235182SMatthew Wilcox (Oracle) * 81467235182SMatthew Wilcox (Oracle) * This function can only be used if the underlying filesystem guarantees 81567235182SMatthew Wilcox (Oracle) * that no other references to @src exist. For example attached buffer 81667235182SMatthew Wilcox (Oracle) * heads are accessed only under the folio lock. If your filesystem cannot 81767235182SMatthew Wilcox (Oracle) * provide this guarantee, buffer_migrate_folio_norefs() may be more 81867235182SMatthew Wilcox (Oracle) * appropriate. 81967235182SMatthew Wilcox (Oracle) * 82067235182SMatthew Wilcox (Oracle) * Return: 0 on success or a negative errno on failure. 82189cb0888SJan Kara */ 82267235182SMatthew Wilcox (Oracle) int buffer_migrate_folio(struct address_space *mapping, 82367235182SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode) 82489cb0888SJan Kara { 82567235182SMatthew Wilcox (Oracle) return __buffer_migrate_folio(mapping, dst, src, mode, false); 82689cb0888SJan Kara } 82767235182SMatthew Wilcox (Oracle) EXPORT_SYMBOL(buffer_migrate_folio); 82889cb0888SJan Kara 82967235182SMatthew Wilcox (Oracle) /** 83067235182SMatthew Wilcox (Oracle) * buffer_migrate_folio_norefs() - Migration function for folios with buffers. 83167235182SMatthew Wilcox (Oracle) * @mapping: The address space containing @src. 83267235182SMatthew Wilcox (Oracle) * @dst: The folio to migrate to. 83367235182SMatthew Wilcox (Oracle) * @src: The folio to migrate from. 83467235182SMatthew Wilcox (Oracle) * @mode: How to migrate the folio. 83567235182SMatthew Wilcox (Oracle) * 83667235182SMatthew Wilcox (Oracle) * Like buffer_migrate_folio() except that this variant is more careful 83767235182SMatthew Wilcox (Oracle) * and checks that there are also no buffer head references. This function 83867235182SMatthew Wilcox (Oracle) * is the right one for mappings where buffer heads are directly looked 83967235182SMatthew Wilcox (Oracle) * up and referenced (such as block device mappings). 84067235182SMatthew Wilcox (Oracle) * 84167235182SMatthew Wilcox (Oracle) * Return: 0 on success or a negative errno on failure. 84289cb0888SJan Kara */ 84367235182SMatthew Wilcox (Oracle) int buffer_migrate_folio_norefs(struct address_space *mapping, 84467235182SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode) 84589cb0888SJan Kara { 84667235182SMatthew Wilcox (Oracle) return __buffer_migrate_folio(mapping, dst, src, mode, true); 84789cb0888SJan Kara } 848e26355e2SJan Kara EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs); 849925c86a1SChristoph Hellwig #endif /* CONFIG_BUFFER_HEAD */ 8501d8b85ccSChristoph Lameter 8512ec810d5SMatthew Wilcox (Oracle) int filemap_migrate_folio(struct address_space *mapping, 8522ec810d5SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode) 8532ec810d5SMatthew Wilcox (Oracle) { 8542ec810d5SMatthew Wilcox (Oracle) int ret; 8552ec810d5SMatthew Wilcox (Oracle) 8562ec810d5SMatthew Wilcox (Oracle) ret = folio_migrate_mapping(mapping, dst, src, 0); 8572ec810d5SMatthew Wilcox (Oracle) if (ret != MIGRATEPAGE_SUCCESS) 8582ec810d5SMatthew Wilcox (Oracle) return ret; 8592ec810d5SMatthew Wilcox (Oracle) 8602ec810d5SMatthew Wilcox (Oracle) if (folio_get_private(src)) 8612ec810d5SMatthew Wilcox (Oracle) folio_attach_private(dst, folio_detach_private(src)); 8622ec810d5SMatthew Wilcox (Oracle) 8632ec810d5SMatthew Wilcox (Oracle) if (mode != MIGRATE_SYNC_NO_COPY) 8642ec810d5SMatthew Wilcox (Oracle) folio_migrate_copy(dst, src); 8652ec810d5SMatthew Wilcox (Oracle) else 8662ec810d5SMatthew Wilcox (Oracle) folio_migrate_flags(dst, src); 8672ec810d5SMatthew Wilcox (Oracle) return MIGRATEPAGE_SUCCESS; 8682ec810d5SMatthew Wilcox (Oracle) } 8692ec810d5SMatthew Wilcox (Oracle) EXPORT_SYMBOL_GPL(filemap_migrate_folio); 8702ec810d5SMatthew Wilcox (Oracle) 87104e62a29SChristoph Lameter /* 8722be7fa10SMatthew Wilcox (Oracle) * Writeback a folio to clean the dirty state 87304e62a29SChristoph Lameter */ 8742be7fa10SMatthew Wilcox (Oracle) static int writeout(struct address_space *mapping, struct folio *folio) 87504e62a29SChristoph Lameter { 87604e62a29SChristoph Lameter struct writeback_control wbc = { 87704e62a29SChristoph Lameter .sync_mode = WB_SYNC_NONE, 87804e62a29SChristoph Lameter .nr_to_write = 1, 87904e62a29SChristoph Lameter .range_start = 0, 88004e62a29SChristoph Lameter .range_end = LLONG_MAX, 88104e62a29SChristoph Lameter .for_reclaim = 1 88204e62a29SChristoph Lameter }; 88304e62a29SChristoph Lameter int rc; 88404e62a29SChristoph Lameter 88504e62a29SChristoph Lameter if (!mapping->a_ops->writepage) 88604e62a29SChristoph Lameter /* No write method for the address space */ 88704e62a29SChristoph Lameter return -EINVAL; 88804e62a29SChristoph Lameter 8892be7fa10SMatthew Wilcox (Oracle) if (!folio_clear_dirty_for_io(folio)) 89004e62a29SChristoph Lameter /* Someone else already triggered a write */ 89104e62a29SChristoph Lameter return -EAGAIN; 89204e62a29SChristoph Lameter 89304e62a29SChristoph Lameter /* 8942be7fa10SMatthew Wilcox (Oracle) * A dirty folio may imply that the underlying filesystem has 8952be7fa10SMatthew Wilcox (Oracle) * the folio on some queue. So the folio must be clean for 8962be7fa10SMatthew Wilcox (Oracle) * migration. Writeout may mean we lose the lock and the 8972be7fa10SMatthew Wilcox (Oracle) * folio state is no longer what we checked for earlier. 89804e62a29SChristoph Lameter * At this point we know that the migration attempt cannot 89904e62a29SChristoph Lameter * be successful. 90004e62a29SChristoph Lameter */ 9014eecb8b9SMatthew Wilcox (Oracle) remove_migration_ptes(folio, folio, false); 90204e62a29SChristoph Lameter 9032be7fa10SMatthew Wilcox (Oracle) rc = mapping->a_ops->writepage(&folio->page, &wbc); 90404e62a29SChristoph Lameter 90504e62a29SChristoph Lameter if (rc != AOP_WRITEPAGE_ACTIVATE) 90604e62a29SChristoph Lameter /* unlocked. Relock */ 9072be7fa10SMatthew Wilcox (Oracle) folio_lock(folio); 90804e62a29SChristoph Lameter 909bda8550dSHugh Dickins return (rc < 0) ? -EIO : -EAGAIN; 91004e62a29SChristoph Lameter } 91104e62a29SChristoph Lameter 91204e62a29SChristoph Lameter /* 91304e62a29SChristoph Lameter * Default handling if a filesystem does not provide a migration function. 91404e62a29SChristoph Lameter */ 9158faa8ef5SMatthew Wilcox (Oracle) static int fallback_migrate_folio(struct address_space *mapping, 9168faa8ef5SMatthew Wilcox (Oracle) struct folio *dst, struct folio *src, enum migrate_mode mode) 9178351a6e4SChristoph Lameter { 9188faa8ef5SMatthew Wilcox (Oracle) if (folio_test_dirty(src)) { 9198faa8ef5SMatthew Wilcox (Oracle) /* Only writeback folios in full synchronous migration */ 9202916ecc0SJérôme Glisse switch (mode) { 9212916ecc0SJérôme Glisse case MIGRATE_SYNC: 9222916ecc0SJérôme Glisse case MIGRATE_SYNC_NO_COPY: 9232916ecc0SJérôme Glisse break; 9242916ecc0SJérôme Glisse default: 925b969c4abSMel Gorman return -EBUSY; 9262916ecc0SJérôme Glisse } 9272be7fa10SMatthew Wilcox (Oracle) return writeout(mapping, src); 928b969c4abSMel Gorman } 9298351a6e4SChristoph Lameter 9308351a6e4SChristoph Lameter /* 9318351a6e4SChristoph Lameter * Buffers may be managed in a filesystem specific way. 9328351a6e4SChristoph Lameter * We must have no buffers or drop them. 9338351a6e4SChristoph Lameter */ 9340201ebf2SDavid Howells if (!filemap_release_folio(src, GFP_KERNEL)) 935806031bbSMel Gorman return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; 9368351a6e4SChristoph Lameter 93754184650SMatthew Wilcox (Oracle) return migrate_folio(mapping, dst, src, mode); 9388351a6e4SChristoph Lameter } 9398351a6e4SChristoph Lameter 9401d8b85ccSChristoph Lameter /* 941e24f0b8fSChristoph Lameter * Move a page to a newly allocated page 942e24f0b8fSChristoph Lameter * The page is locked and all ptes have been successfully removed. 943b20a3503SChristoph Lameter * 944e24f0b8fSChristoph Lameter * The new page will have replaced the old page if this function 945e24f0b8fSChristoph Lameter * is successful. 946894bc310SLee Schermerhorn * 947894bc310SLee Schermerhorn * Return value: 948894bc310SLee Schermerhorn * < 0 - error code 94978bd5209SRafael Aquini * MIGRATEPAGE_SUCCESS - success 950b20a3503SChristoph Lameter */ 951e7e3ffebSMatthew Wilcox (Oracle) static int move_to_new_folio(struct folio *dst, struct folio *src, 9525c3f9a67SHugh Dickins enum migrate_mode mode) 953b20a3503SChristoph Lameter { 954bda807d4SMinchan Kim int rc = -EAGAIN; 9557e2a5e5aSKefeng Wang bool is_lru = !__folio_test_movable(src); 956b20a3503SChristoph Lameter 957e7e3ffebSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(src), src); 958e7e3ffebSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst); 959b20a3503SChristoph Lameter 960bda807d4SMinchan Kim if (likely(is_lru)) { 96168f2736aSMatthew Wilcox (Oracle) struct address_space *mapping = folio_mapping(src); 96268f2736aSMatthew Wilcox (Oracle) 963b20a3503SChristoph Lameter if (!mapping) 96454184650SMatthew Wilcox (Oracle) rc = migrate_folio(mapping, dst, src, mode); 9655490da4fSMatthew Wilcox (Oracle) else if (mapping->a_ops->migrate_folio) 966b20a3503SChristoph Lameter /* 9675490da4fSMatthew Wilcox (Oracle) * Most folios have a mapping and most filesystems 9685490da4fSMatthew Wilcox (Oracle) * provide a migrate_folio callback. Anonymous folios 969bda807d4SMinchan Kim * are part of swap space which also has its own 9705490da4fSMatthew Wilcox (Oracle) * migrate_folio callback. This is the most common path 971bda807d4SMinchan Kim * for page migration. 972b20a3503SChristoph Lameter */ 9735490da4fSMatthew Wilcox (Oracle) rc = mapping->a_ops->migrate_folio(mapping, dst, src, 9745490da4fSMatthew Wilcox (Oracle) mode); 9758351a6e4SChristoph Lameter else 9768faa8ef5SMatthew Wilcox (Oracle) rc = fallback_migrate_folio(mapping, dst, src, mode); 977bda807d4SMinchan Kim } else { 97868f2736aSMatthew Wilcox (Oracle) const struct movable_operations *mops; 97968f2736aSMatthew Wilcox (Oracle) 980bda807d4SMinchan Kim /* 981bda807d4SMinchan Kim * In case of non-lru page, it could be released after 982bda807d4SMinchan Kim * isolation step. In that case, we shouldn't try migration. 983bda807d4SMinchan Kim */ 984e7e3ffebSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_isolated(src), src); 985e7e3ffebSMatthew Wilcox (Oracle) if (!folio_test_movable(src)) { 986bda807d4SMinchan Kim rc = MIGRATEPAGE_SUCCESS; 987e7e3ffebSMatthew Wilcox (Oracle) folio_clear_isolated(src); 988bda807d4SMinchan Kim goto out; 989bda807d4SMinchan Kim } 990bda807d4SMinchan Kim 991da707a6dSVishal Moola (Oracle) mops = folio_movable_ops(src); 99268f2736aSMatthew Wilcox (Oracle) rc = mops->migrate_page(&dst->page, &src->page, mode); 993bda807d4SMinchan Kim WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && 994e7e3ffebSMatthew Wilcox (Oracle) !folio_test_isolated(src)); 995bda807d4SMinchan Kim } 996b20a3503SChristoph Lameter 9975c3f9a67SHugh Dickins /* 998e7e3ffebSMatthew Wilcox (Oracle) * When successful, old pagecache src->mapping must be cleared before 999e7e3ffebSMatthew Wilcox (Oracle) * src is freed; but stats require that PageAnon be left as PageAnon. 10005c3f9a67SHugh Dickins */ 10015c3f9a67SHugh Dickins if (rc == MIGRATEPAGE_SUCCESS) { 10027e2a5e5aSKefeng Wang if (__folio_test_movable(src)) { 1003e7e3ffebSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_isolated(src), src); 1004bda807d4SMinchan Kim 1005bda807d4SMinchan Kim /* 1006bda807d4SMinchan Kim * We clear PG_movable under page_lock so any compactor 1007bda807d4SMinchan Kim * cannot try to migrate this page. 1008bda807d4SMinchan Kim */ 1009e7e3ffebSMatthew Wilcox (Oracle) folio_clear_isolated(src); 1010bda807d4SMinchan Kim } 1011bda807d4SMinchan Kim 1012bda807d4SMinchan Kim /* 1013e7e3ffebSMatthew Wilcox (Oracle) * Anonymous and movable src->mapping will be cleared by 1014bda807d4SMinchan Kim * free_pages_prepare so don't reset it here for keeping 1015bda807d4SMinchan Kim * the type to work PageAnon, for example. 1016bda807d4SMinchan Kim */ 1017e7e3ffebSMatthew Wilcox (Oracle) if (!folio_mapping_flags(src)) 1018e7e3ffebSMatthew Wilcox (Oracle) src->mapping = NULL; 1019d2b2c6ddSLars Persson 1020e7e3ffebSMatthew Wilcox (Oracle) if (likely(!folio_is_zone_device(dst))) 1021e7e3ffebSMatthew Wilcox (Oracle) flush_dcache_folio(dst); 10223fe2011fSMel Gorman } 1023bda807d4SMinchan Kim out: 1024e24f0b8fSChristoph Lameter return rc; 1025e24f0b8fSChristoph Lameter } 1026e24f0b8fSChristoph Lameter 102764c8902eSHuang Ying /* 102864c8902eSHuang Ying * To record some information during migration, we use some unused 102964c8902eSHuang Ying * fields (mapping and private) of struct folio of the newly allocated 103064c8902eSHuang Ying * destination folio. This is safe because nobody is using them 103164c8902eSHuang Ying * except us. 103264c8902eSHuang Ying */ 1033e77d587aSLinus Torvalds union migration_ptr { 1034e77d587aSLinus Torvalds struct anon_vma *anon_vma; 1035e77d587aSLinus Torvalds struct address_space *mapping; 1036e77d587aSLinus Torvalds }; 1037eebb3dabSBaolin Wang 1038eebb3dabSBaolin Wang enum { 1039eebb3dabSBaolin Wang PAGE_WAS_MAPPED = BIT(0), 1040eebb3dabSBaolin Wang PAGE_WAS_MLOCKED = BIT(1), 1041eebb3dabSBaolin Wang }; 1042eebb3dabSBaolin Wang 104364c8902eSHuang Ying static void __migrate_folio_record(struct folio *dst, 1044eebb3dabSBaolin Wang unsigned long old_page_state, 104564c8902eSHuang Ying struct anon_vma *anon_vma) 1046e24f0b8fSChristoph Lameter { 1047e77d587aSLinus Torvalds union migration_ptr ptr = { .anon_vma = anon_vma }; 1048e77d587aSLinus Torvalds dst->mapping = ptr.mapping; 1049eebb3dabSBaolin Wang dst->private = (void *)old_page_state; 105064c8902eSHuang Ying } 105164c8902eSHuang Ying 105264c8902eSHuang Ying static void __migrate_folio_extract(struct folio *dst, 1053eebb3dabSBaolin Wang int *old_page_state, 105464c8902eSHuang Ying struct anon_vma **anon_vmap) 105564c8902eSHuang Ying { 1056e77d587aSLinus Torvalds union migration_ptr ptr = { .mapping = dst->mapping }; 1057e77d587aSLinus Torvalds *anon_vmap = ptr.anon_vma; 1058eebb3dabSBaolin Wang *old_page_state = (unsigned long)dst->private; 105964c8902eSHuang Ying dst->mapping = NULL; 106064c8902eSHuang Ying dst->private = NULL; 106164c8902eSHuang Ying } 106264c8902eSHuang Ying 10635dfab109SHuang Ying /* Restore the source folio to the original state upon failure */ 10645dfab109SHuang Ying static void migrate_folio_undo_src(struct folio *src, 10655dfab109SHuang Ying int page_was_mapped, 10665dfab109SHuang Ying struct anon_vma *anon_vma, 1067ebe75e47SHuang Ying bool locked, 10685dfab109SHuang Ying struct list_head *ret) 10695dfab109SHuang Ying { 10705dfab109SHuang Ying if (page_was_mapped) 10715dfab109SHuang Ying remove_migration_ptes(src, src, false); 10725dfab109SHuang Ying /* Drop an anon_vma reference if we took one */ 10735dfab109SHuang Ying if (anon_vma) 10745dfab109SHuang Ying put_anon_vma(anon_vma); 1075ebe75e47SHuang Ying if (locked) 10765dfab109SHuang Ying folio_unlock(src); 1077ebe75e47SHuang Ying if (ret) 10785dfab109SHuang Ying list_move_tail(&src->lru, ret); 10795dfab109SHuang Ying } 10805dfab109SHuang Ying 10815dfab109SHuang Ying /* Restore the destination folio to the original state upon failure */ 10824e096ae1SMatthew Wilcox (Oracle) static void migrate_folio_undo_dst(struct folio *dst, bool locked, 10834e096ae1SMatthew Wilcox (Oracle) free_folio_t put_new_folio, unsigned long private) 10845dfab109SHuang Ying { 1085ebe75e47SHuang Ying if (locked) 10865dfab109SHuang Ying folio_unlock(dst); 10874e096ae1SMatthew Wilcox (Oracle) if (put_new_folio) 10884e096ae1SMatthew Wilcox (Oracle) put_new_folio(dst, private); 10895dfab109SHuang Ying else 10905dfab109SHuang Ying folio_put(dst); 10915dfab109SHuang Ying } 10925dfab109SHuang Ying 109364c8902eSHuang Ying /* Cleanup src folio upon migration success */ 109464c8902eSHuang Ying static void migrate_folio_done(struct folio *src, 109564c8902eSHuang Ying enum migrate_reason reason) 109664c8902eSHuang Ying { 109764c8902eSHuang Ying /* 109864c8902eSHuang Ying * Compaction can migrate also non-LRU pages which are 109964c8902eSHuang Ying * not accounted to NR_ISOLATED_*. They can be recognized 11007e2a5e5aSKefeng Wang * as __folio_test_movable 110164c8902eSHuang Ying */ 110264c8902eSHuang Ying if (likely(!__folio_test_movable(src))) 110364c8902eSHuang Ying mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON + 110464c8902eSHuang Ying folio_is_file_lru(src), -folio_nr_pages(src)); 110564c8902eSHuang Ying 110664c8902eSHuang Ying if (reason != MR_MEMORY_FAILURE) 110764c8902eSHuang Ying /* We release the page in page_handle_poison. */ 110864c8902eSHuang Ying folio_put(src); 110964c8902eSHuang Ying } 111064c8902eSHuang Ying 1111ebe75e47SHuang Ying /* Obtain the lock on page, remove all ptes. */ 11124e096ae1SMatthew Wilcox (Oracle) static int migrate_folio_unmap(new_folio_t get_new_folio, 11134e096ae1SMatthew Wilcox (Oracle) free_folio_t put_new_folio, unsigned long private, 11144e096ae1SMatthew Wilcox (Oracle) struct folio *src, struct folio **dstp, enum migrate_mode mode, 11152ef7dbb2SHuang Ying enum migrate_reason reason, struct list_head *ret) 1116e24f0b8fSChristoph Lameter { 1117ebe75e47SHuang Ying struct folio *dst; 11180dabec93SMinchan Kim int rc = -EAGAIN; 1119eebb3dabSBaolin Wang int old_page_state = 0; 11203f6c8272SMel Gorman struct anon_vma *anon_vma = NULL; 11217e2a5e5aSKefeng Wang bool is_lru = !__folio_test_movable(src); 1122ebe75e47SHuang Ying bool locked = false; 1123ebe75e47SHuang Ying bool dst_locked = false; 1124ebe75e47SHuang Ying 1125ebe75e47SHuang Ying if (folio_ref_count(src) == 1) { 1126ebe75e47SHuang Ying /* Folio was freed from under us. So we are done. */ 1127ebe75e47SHuang Ying folio_clear_active(src); 1128ebe75e47SHuang Ying folio_clear_unevictable(src); 1129ebe75e47SHuang Ying /* free_pages_prepare() will clear PG_isolated. */ 1130ebe75e47SHuang Ying list_del(&src->lru); 1131ebe75e47SHuang Ying migrate_folio_done(src, reason); 1132ebe75e47SHuang Ying return MIGRATEPAGE_SUCCESS; 1133ebe75e47SHuang Ying } 1134ebe75e47SHuang Ying 11354e096ae1SMatthew Wilcox (Oracle) dst = get_new_folio(src, private); 11364e096ae1SMatthew Wilcox (Oracle) if (!dst) 1137ebe75e47SHuang Ying return -ENOMEM; 1138ebe75e47SHuang Ying *dstp = dst; 1139ebe75e47SHuang Ying 1140ebe75e47SHuang Ying dst->private = NULL; 114195a402c3SChristoph Lameter 1142682a71a1SMatthew Wilcox (Oracle) if (!folio_trylock(src)) { 11432ef7dbb2SHuang Ying if (mode == MIGRATE_ASYNC) 11440dabec93SMinchan Kim goto out; 11453e7d3449SMel Gorman 11463e7d3449SMel Gorman /* 11473e7d3449SMel Gorman * It's not safe for direct compaction to call lock_page. 11483e7d3449SMel Gorman * For example, during page readahead pages are added locked 11493e7d3449SMel Gorman * to the LRU. Later, when the IO completes the pages are 11503e7d3449SMel Gorman * marked uptodate and unlocked. However, the queueing 11513e7d3449SMel Gorman * could be merging multiple pages for one bio (e.g. 1152d4388340SMatthew Wilcox (Oracle) * mpage_readahead). If an allocation happens for the 11533e7d3449SMel Gorman * second or third page, the process can end up locking 11543e7d3449SMel Gorman * the same page twice and deadlocking. Rather than 11553e7d3449SMel Gorman * trying to be clever about what pages can be locked, 11563e7d3449SMel Gorman * avoid the use of lock_page for direct compaction 11573e7d3449SMel Gorman * altogether. 11583e7d3449SMel Gorman */ 11593e7d3449SMel Gorman if (current->flags & PF_MEMALLOC) 11600dabec93SMinchan Kim goto out; 11613e7d3449SMel Gorman 11624bb6dc79SDouglas Anderson /* 11634bb6dc79SDouglas Anderson * In "light" mode, we can wait for transient locks (eg 11644bb6dc79SDouglas Anderson * inserting a page into the page table), but it's not 11654bb6dc79SDouglas Anderson * worth waiting for I/O. 11664bb6dc79SDouglas Anderson */ 11674bb6dc79SDouglas Anderson if (mode == MIGRATE_SYNC_LIGHT && !folio_test_uptodate(src)) 11684bb6dc79SDouglas Anderson goto out; 11694bb6dc79SDouglas Anderson 1170682a71a1SMatthew Wilcox (Oracle) folio_lock(src); 1171e24f0b8fSChristoph Lameter } 1172ebe75e47SHuang Ying locked = true; 1173eebb3dabSBaolin Wang if (folio_test_mlocked(src)) 1174eebb3dabSBaolin Wang old_page_state |= PAGE_WAS_MLOCKED; 1175e24f0b8fSChristoph Lameter 1176682a71a1SMatthew Wilcox (Oracle) if (folio_test_writeback(src)) { 117711bc82d6SAndrea Arcangeli /* 1178fed5b64aSJianguo Wu * Only in the case of a full synchronous migration is it 1179a6bc32b8SMel Gorman * necessary to wait for PageWriteback. In the async case, 1180a6bc32b8SMel Gorman * the retry loop is too short and in the sync-light case, 1181a6bc32b8SMel Gorman * the overhead of stalling is too much 118211bc82d6SAndrea Arcangeli */ 11832916ecc0SJérôme Glisse switch (mode) { 11842916ecc0SJérôme Glisse case MIGRATE_SYNC: 11852916ecc0SJérôme Glisse case MIGRATE_SYNC_NO_COPY: 11862916ecc0SJérôme Glisse break; 11872916ecc0SJérôme Glisse default: 118811bc82d6SAndrea Arcangeli rc = -EBUSY; 1189ebe75e47SHuang Ying goto out; 119011bc82d6SAndrea Arcangeli } 1191682a71a1SMatthew Wilcox (Oracle) folio_wait_writeback(src); 1192e24f0b8fSChristoph Lameter } 119303f15c86SHugh Dickins 1194e24f0b8fSChristoph Lameter /* 1195682a71a1SMatthew Wilcox (Oracle) * By try_to_migrate(), src->mapcount goes down to 0 here. In this case, 1196682a71a1SMatthew Wilcox (Oracle) * we cannot notice that anon_vma is freed while we migrate a page. 11971ce82b69SHugh Dickins * This get_anon_vma() delays freeing anon_vma pointer until the end 1198dc386d4dSKAMEZAWA Hiroyuki * of migration. File cache pages are no problem because of page_lock() 1199989f89c5SKAMEZAWA Hiroyuki * File Caches may use write_page() or lock_page() in migration, then, 1200989f89c5SKAMEZAWA Hiroyuki * just care Anon page here. 12013fe2011fSMel Gorman * 120229eea9b5SMatthew Wilcox (Oracle) * Only folio_get_anon_vma() understands the subtleties of 120303f15c86SHugh Dickins * getting a hold on an anon_vma from outside one of its mms. 120403f15c86SHugh Dickins * But if we cannot get anon_vma, then we won't need it anyway, 120503f15c86SHugh Dickins * because that implies that the anon page is no longer mapped 120603f15c86SHugh Dickins * (and cannot be remapped so long as we hold the page lock). 12073fe2011fSMel Gorman */ 1208682a71a1SMatthew Wilcox (Oracle) if (folio_test_anon(src) && !folio_test_ksm(src)) 120929eea9b5SMatthew Wilcox (Oracle) anon_vma = folio_get_anon_vma(src); 121062e1c553SShaohua Li 12117db7671fSHugh Dickins /* 12127db7671fSHugh Dickins * Block others from accessing the new page when we get around to 12137db7671fSHugh Dickins * establishing additional references. We are usually the only one 1214682a71a1SMatthew Wilcox (Oracle) * holding a reference to dst at this point. We used to have a BUG 1215682a71a1SMatthew Wilcox (Oracle) * here if folio_trylock(dst) fails, but would like to allow for 1216682a71a1SMatthew Wilcox (Oracle) * cases where there might be a race with the previous use of dst. 12177db7671fSHugh Dickins * This is much like races on refcount of oldpage: just don't BUG(). 12187db7671fSHugh Dickins */ 1219682a71a1SMatthew Wilcox (Oracle) if (unlikely(!folio_trylock(dst))) 1220ebe75e47SHuang Ying goto out; 1221ebe75e47SHuang Ying dst_locked = true; 12227db7671fSHugh Dickins 1223bda807d4SMinchan Kim if (unlikely(!is_lru)) { 1224eebb3dabSBaolin Wang __migrate_folio_record(dst, old_page_state, anon_vma); 122564c8902eSHuang Ying return MIGRATEPAGE_UNMAP; 1226bda807d4SMinchan Kim } 1227bda807d4SMinchan Kim 1228dc386d4dSKAMEZAWA Hiroyuki /* 122962e1c553SShaohua Li * Corner case handling: 123062e1c553SShaohua Li * 1. When a new swap-cache page is read into, it is added to the LRU 123162e1c553SShaohua Li * and treated as swapcache but it has no rmap yet. 1232682a71a1SMatthew Wilcox (Oracle) * Calling try_to_unmap() against a src->mapping==NULL page will 123362e1c553SShaohua Li * trigger a BUG. So handle it here. 1234d12b8951SYang Shi * 2. An orphaned page (see truncate_cleanup_page) might have 123562e1c553SShaohua Li * fs-private metadata. The page can be picked up due to memory 123662e1c553SShaohua Li * offlining. Everywhere else except page reclaim, the page is 123762e1c553SShaohua Li * invisible to the vm, so the page can not be migrated. So try to 123862e1c553SShaohua Li * free the metadata, so the page can be freed. 1239dc386d4dSKAMEZAWA Hiroyuki */ 1240682a71a1SMatthew Wilcox (Oracle) if (!src->mapping) { 1241682a71a1SMatthew Wilcox (Oracle) if (folio_test_private(src)) { 1242682a71a1SMatthew Wilcox (Oracle) try_to_free_buffers(src); 1243ebe75e47SHuang Ying goto out; 124462e1c553SShaohua Li } 1245682a71a1SMatthew Wilcox (Oracle) } else if (folio_mapped(src)) { 12467db7671fSHugh Dickins /* Establish migration ptes */ 1247682a71a1SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_test_anon(src) && 1248682a71a1SMatthew Wilcox (Oracle) !folio_test_ksm(src) && !anon_vma, src); 1249fb3592c4SHuang Ying try_to_migrate(src, mode == MIGRATE_ASYNC ? TTU_BATCH_FLUSH : 0); 1250eebb3dabSBaolin Wang old_page_state |= PAGE_WAS_MAPPED; 12512ebba6b7SHugh Dickins } 1252dc386d4dSKAMEZAWA Hiroyuki 125364c8902eSHuang Ying if (!folio_mapped(src)) { 1254eebb3dabSBaolin Wang __migrate_folio_record(dst, old_page_state, anon_vma); 125564c8902eSHuang Ying return MIGRATEPAGE_UNMAP; 125664c8902eSHuang Ying } 125764c8902eSHuang Ying 125864c8902eSHuang Ying out: 125980562ba0SHuang Ying /* 126080562ba0SHuang Ying * A folio that has not been unmapped will be restored to 126180562ba0SHuang Ying * right list unless we want to retry. 126280562ba0SHuang Ying */ 1263fb3592c4SHuang Ying if (rc == -EAGAIN) 1264ebe75e47SHuang Ying ret = NULL; 126580562ba0SHuang Ying 1266eebb3dabSBaolin Wang migrate_folio_undo_src(src, old_page_state & PAGE_WAS_MAPPED, 1267eebb3dabSBaolin Wang anon_vma, locked, ret); 12684e096ae1SMatthew Wilcox (Oracle) migrate_folio_undo_dst(dst, dst_locked, put_new_folio, private); 126980562ba0SHuang Ying 127080562ba0SHuang Ying return rc; 127180562ba0SHuang Ying } 127280562ba0SHuang Ying 1273ebe75e47SHuang Ying /* Migrate the folio to the newly allocated folio in dst. */ 12744e096ae1SMatthew Wilcox (Oracle) static int migrate_folio_move(free_folio_t put_new_folio, unsigned long private, 1275ebe75e47SHuang Ying struct folio *src, struct folio *dst, 1276ebe75e47SHuang Ying enum migrate_mode mode, enum migrate_reason reason, 1277ebe75e47SHuang Ying struct list_head *ret) 127864c8902eSHuang Ying { 127964c8902eSHuang Ying int rc; 1280eebb3dabSBaolin Wang int old_page_state = 0; 128164c8902eSHuang Ying struct anon_vma *anon_vma = NULL; 12827e2a5e5aSKefeng Wang bool is_lru = !__folio_test_movable(src); 12835dfab109SHuang Ying struct list_head *prev; 128464c8902eSHuang Ying 1285eebb3dabSBaolin Wang __migrate_folio_extract(dst, &old_page_state, &anon_vma); 12865dfab109SHuang Ying prev = dst->lru.prev; 12875dfab109SHuang Ying list_del(&dst->lru); 128864c8902eSHuang Ying 1289682a71a1SMatthew Wilcox (Oracle) rc = move_to_new_folio(dst, src, mode); 1290ebe75e47SHuang Ying if (rc) 1291ebe75e47SHuang Ying goto out; 12925dfab109SHuang Ying 129364c8902eSHuang Ying if (unlikely(!is_lru)) 129464c8902eSHuang Ying goto out_unlock_both; 1295e24f0b8fSChristoph Lameter 1296c3096e67SHugh Dickins /* 1297682a71a1SMatthew Wilcox (Oracle) * When successful, push dst to LRU immediately: so that if it 1298c3096e67SHugh Dickins * turns out to be an mlocked page, remove_migration_ptes() will 1299682a71a1SMatthew Wilcox (Oracle) * automatically build up the correct dst->mlock_count for it. 1300c3096e67SHugh Dickins * 1301c3096e67SHugh Dickins * We would like to do something similar for the old page, when 1302c3096e67SHugh Dickins * unsuccessful, and other cases when a page has been temporarily 1303c3096e67SHugh Dickins * isolated from the unevictable LRU: but this case is the easiest. 1304c3096e67SHugh Dickins */ 1305682a71a1SMatthew Wilcox (Oracle) folio_add_lru(dst); 1306eebb3dabSBaolin Wang if (old_page_state & PAGE_WAS_MLOCKED) 1307c3096e67SHugh Dickins lru_add_drain(); 1308c3096e67SHugh Dickins 1309eebb3dabSBaolin Wang if (old_page_state & PAGE_WAS_MAPPED) 1310ebe75e47SHuang Ying remove_migration_ptes(src, dst, false); 13113f6c8272SMel Gorman 13127db7671fSHugh Dickins out_unlock_both: 1313682a71a1SMatthew Wilcox (Oracle) folio_unlock(dst); 1314ebe75e47SHuang Ying set_page_owner_migrate_reason(&dst->page, reason); 1315c6c919ebSMinchan Kim /* 1316682a71a1SMatthew Wilcox (Oracle) * If migration is successful, decrease refcount of dst, 1317c6c919ebSMinchan Kim * which will not free the page because new page owner increased 1318c3096e67SHugh Dickins * refcounter. 1319c6c919ebSMinchan Kim */ 1320682a71a1SMatthew Wilcox (Oracle) folio_put(dst); 1321c6c919ebSMinchan Kim 1322ebe75e47SHuang Ying /* 1323ebe75e47SHuang Ying * A folio that has been migrated has all references removed 1324ebe75e47SHuang Ying * and will be freed. 1325ebe75e47SHuang Ying */ 1326ebe75e47SHuang Ying list_del(&src->lru); 1327ebe75e47SHuang Ying /* Drop an anon_vma reference if we took one */ 1328ebe75e47SHuang Ying if (anon_vma) 1329ebe75e47SHuang Ying put_anon_vma(anon_vma); 1330ebe75e47SHuang Ying folio_unlock(src); 1331ebe75e47SHuang Ying migrate_folio_done(src, reason); 1332ebe75e47SHuang Ying 1333ebe75e47SHuang Ying return rc; 1334ebe75e47SHuang Ying out: 1335ebe75e47SHuang Ying /* 1336ebe75e47SHuang Ying * A folio that has not been migrated will be restored to 1337ebe75e47SHuang Ying * right list unless we want to retry. 1338ebe75e47SHuang Ying */ 1339ebe75e47SHuang Ying if (rc == -EAGAIN) { 1340ebe75e47SHuang Ying list_add(&dst->lru, prev); 1341eebb3dabSBaolin Wang __migrate_folio_record(dst, old_page_state, anon_vma); 13420dabec93SMinchan Kim return rc; 13430dabec93SMinchan Kim } 134495a402c3SChristoph Lameter 1345eebb3dabSBaolin Wang migrate_folio_undo_src(src, old_page_state & PAGE_WAS_MAPPED, 1346eebb3dabSBaolin Wang anon_vma, true, ret); 13474e096ae1SMatthew Wilcox (Oracle) migrate_folio_undo_dst(dst, true, put_new_folio, private); 134868711a74SDavid Rientjes 1349e24f0b8fSChristoph Lameter return rc; 1350e24f0b8fSChristoph Lameter } 1351b20a3503SChristoph Lameter 1352e24f0b8fSChristoph Lameter /* 1353290408d4SNaoya Horiguchi * Counterpart of unmap_and_move_page() for hugepage migration. 1354290408d4SNaoya Horiguchi * 1355290408d4SNaoya Horiguchi * This function doesn't wait the completion of hugepage I/O 1356290408d4SNaoya Horiguchi * because there is no race between I/O and migration for hugepage. 1357290408d4SNaoya Horiguchi * Note that currently hugepage I/O occurs only in direct I/O 1358290408d4SNaoya Horiguchi * where no lock is held and PG_writeback is irrelevant, 1359290408d4SNaoya Horiguchi * and writeback status of all subpages are counted in the reference 1360290408d4SNaoya Horiguchi * count of the head page (i.e. if all subpages of a 2MB hugepage are 1361290408d4SNaoya Horiguchi * under direct I/O, the reference of the head page is 512 and a bit more.) 1362290408d4SNaoya Horiguchi * This means that when we try to migrate hugepage whose subpages are 1363290408d4SNaoya Horiguchi * doing direct I/O, some references remain after try_to_unmap() and 1364290408d4SNaoya Horiguchi * hugepage migration fails without data corruption. 1365290408d4SNaoya Horiguchi * 1366290408d4SNaoya Horiguchi * There is also no race when direct I/O is issued on the page under migration, 1367290408d4SNaoya Horiguchi * because then pte is replaced with migration swap entry and direct I/O code 1368290408d4SNaoya Horiguchi * will wait in the page fault for migration to complete. 1369290408d4SNaoya Horiguchi */ 13704e096ae1SMatthew Wilcox (Oracle) static int unmap_and_move_huge_page(new_folio_t get_new_folio, 13714e096ae1SMatthew Wilcox (Oracle) free_folio_t put_new_folio, unsigned long private, 13724e096ae1SMatthew Wilcox (Oracle) struct folio *src, int force, enum migrate_mode mode, 13734e096ae1SMatthew Wilcox (Oracle) int reason, struct list_head *ret) 1374290408d4SNaoya Horiguchi { 13754e096ae1SMatthew Wilcox (Oracle) struct folio *dst; 13762def7424SHugh Dickins int rc = -EAGAIN; 13772ebba6b7SHugh Dickins int page_was_mapped = 0; 1378290408d4SNaoya Horiguchi struct anon_vma *anon_vma = NULL; 1379c0d0381aSMike Kravetz struct address_space *mapping = NULL; 1380290408d4SNaoya Horiguchi 1381c33db292SMatthew Wilcox (Oracle) if (folio_ref_count(src) == 1) { 138271a64f61SMuchun Song /* page was freed from under us. So we are done. */ 1383ea8e72f4SSidhartha Kumar folio_putback_active_hugetlb(src); 138471a64f61SMuchun Song return MIGRATEPAGE_SUCCESS; 138571a64f61SMuchun Song } 138671a64f61SMuchun Song 13874e096ae1SMatthew Wilcox (Oracle) dst = get_new_folio(src, private); 13884e096ae1SMatthew Wilcox (Oracle) if (!dst) 1389290408d4SNaoya Horiguchi return -ENOMEM; 1390290408d4SNaoya Horiguchi 1391c33db292SMatthew Wilcox (Oracle) if (!folio_trylock(src)) { 13922916ecc0SJérôme Glisse if (!force) 1393290408d4SNaoya Horiguchi goto out; 13942916ecc0SJérôme Glisse switch (mode) { 13952916ecc0SJérôme Glisse case MIGRATE_SYNC: 13962916ecc0SJérôme Glisse case MIGRATE_SYNC_NO_COPY: 13972916ecc0SJérôme Glisse break; 13982916ecc0SJérôme Glisse default: 13992916ecc0SJérôme Glisse goto out; 14002916ecc0SJérôme Glisse } 1401c33db292SMatthew Wilcox (Oracle) folio_lock(src); 1402290408d4SNaoya Horiguchi } 1403290408d4SNaoya Horiguchi 1404cb6acd01SMike Kravetz /* 1405cb6acd01SMike Kravetz * Check for pages which are in the process of being freed. Without 1406c33db292SMatthew Wilcox (Oracle) * folio_mapping() set, hugetlbfs specific move page routine will not 1407cb6acd01SMike Kravetz * be called and we could leak usage counts for subpools. 1408cb6acd01SMike Kravetz */ 1409345c62d1SSidhartha Kumar if (hugetlb_folio_subpool(src) && !folio_mapping(src)) { 1410cb6acd01SMike Kravetz rc = -EBUSY; 1411cb6acd01SMike Kravetz goto out_unlock; 1412cb6acd01SMike Kravetz } 1413cb6acd01SMike Kravetz 1414c33db292SMatthew Wilcox (Oracle) if (folio_test_anon(src)) 141529eea9b5SMatthew Wilcox (Oracle) anon_vma = folio_get_anon_vma(src); 1416290408d4SNaoya Horiguchi 1417c33db292SMatthew Wilcox (Oracle) if (unlikely(!folio_trylock(dst))) 14187db7671fSHugh Dickins goto put_anon; 14197db7671fSHugh Dickins 1420c33db292SMatthew Wilcox (Oracle) if (folio_mapped(src)) { 1421a98a2f0cSAlistair Popple enum ttu_flags ttu = 0; 1422336bf30eSMike Kravetz 1423c33db292SMatthew Wilcox (Oracle) if (!folio_test_anon(src)) { 1424c0d0381aSMike Kravetz /* 1425336bf30eSMike Kravetz * In shared mappings, try_to_unmap could potentially 1426336bf30eSMike Kravetz * call huge_pmd_unshare. Because of this, take 1427336bf30eSMike Kravetz * semaphore in write mode here and set TTU_RMAP_LOCKED 1428336bf30eSMike Kravetz * to let lower levels know we have taken the lock. 1429c0d0381aSMike Kravetz */ 14304e096ae1SMatthew Wilcox (Oracle) mapping = hugetlb_page_mapping_lock_write(&src->page); 1431c0d0381aSMike Kravetz if (unlikely(!mapping)) 1432c0d0381aSMike Kravetz goto unlock_put_anon; 1433c0d0381aSMike Kravetz 14345202978bSMiaohe Lin ttu = TTU_RMAP_LOCKED; 1435336bf30eSMike Kravetz } 1436336bf30eSMike Kravetz 14374b8554c5SMatthew Wilcox (Oracle) try_to_migrate(src, ttu); 14382ebba6b7SHugh Dickins page_was_mapped = 1; 1439336bf30eSMike Kravetz 14405202978bSMiaohe Lin if (ttu & TTU_RMAP_LOCKED) 1441336bf30eSMike Kravetz i_mmap_unlock_write(mapping); 14422ebba6b7SHugh Dickins } 1443290408d4SNaoya Horiguchi 1444c33db292SMatthew Wilcox (Oracle) if (!folio_mapped(src)) 1445e7e3ffebSMatthew Wilcox (Oracle) rc = move_to_new_folio(dst, src, mode); 1446290408d4SNaoya Horiguchi 1447336bf30eSMike Kravetz if (page_was_mapped) 14484eecb8b9SMatthew Wilcox (Oracle) remove_migration_ptes(src, 14494eecb8b9SMatthew Wilcox (Oracle) rc == MIGRATEPAGE_SUCCESS ? dst : src, false); 1450290408d4SNaoya Horiguchi 1451c0d0381aSMike Kravetz unlock_put_anon: 1452c33db292SMatthew Wilcox (Oracle) folio_unlock(dst); 14537db7671fSHugh Dickins 14547db7671fSHugh Dickins put_anon: 1455fd4a4663SHugh Dickins if (anon_vma) 14569e60109fSPeter Zijlstra put_anon_vma(anon_vma); 14578e6ac7faSAneesh Kumar K.V 14582def7424SHugh Dickins if (rc == MIGRATEPAGE_SUCCESS) { 1459345c62d1SSidhartha Kumar move_hugetlb_state(src, dst, reason); 14604e096ae1SMatthew Wilcox (Oracle) put_new_folio = NULL; 14612def7424SHugh Dickins } 14628e6ac7faSAneesh Kumar K.V 1463cb6acd01SMike Kravetz out_unlock: 1464c33db292SMatthew Wilcox (Oracle) folio_unlock(src); 146509761333SHillf Danton out: 1466dd4ae78aSYang Shi if (rc == MIGRATEPAGE_SUCCESS) 1467ea8e72f4SSidhartha Kumar folio_putback_active_hugetlb(src); 1468a04840c6SMiaohe Lin else if (rc != -EAGAIN) 1469c33db292SMatthew Wilcox (Oracle) list_move_tail(&src->lru, ret); 147068711a74SDavid Rientjes 147168711a74SDavid Rientjes /* 147268711a74SDavid Rientjes * If migration was not successful and there's a freeing callback, use 147368711a74SDavid Rientjes * it. Otherwise, put_page() will drop the reference grabbed during 147468711a74SDavid Rientjes * isolation. 147568711a74SDavid Rientjes */ 14764e096ae1SMatthew Wilcox (Oracle) if (put_new_folio) 14774e096ae1SMatthew Wilcox (Oracle) put_new_folio(dst, private); 147868711a74SDavid Rientjes else 1479ea8e72f4SSidhartha Kumar folio_putback_active_hugetlb(dst); 148068711a74SDavid Rientjes 1481290408d4SNaoya Horiguchi return rc; 1482290408d4SNaoya Horiguchi } 1483290408d4SNaoya Horiguchi 1484eaec4e63SHuang Ying static inline int try_split_folio(struct folio *folio, struct list_head *split_folios) 1485d532e2e5SYang Shi { 14869c62ff00SHuang Ying int rc; 1487d532e2e5SYang Shi 1488eaec4e63SHuang Ying folio_lock(folio); 1489eaec4e63SHuang Ying rc = split_folio_to_list(folio, split_folios); 1490eaec4e63SHuang Ying folio_unlock(folio); 1491e6fa8a79SHuang Ying if (!rc) 1492eaec4e63SHuang Ying list_move_tail(&folio->lru, split_folios); 1493d532e2e5SYang Shi 1494d532e2e5SYang Shi return rc; 1495d532e2e5SYang Shi } 1496d532e2e5SYang Shi 149742012e04SHuang Ying #ifdef CONFIG_TRANSPARENT_HUGEPAGE 149842012e04SHuang Ying #define NR_MAX_BATCHED_MIGRATION HPAGE_PMD_NR 149942012e04SHuang Ying #else 150042012e04SHuang Ying #define NR_MAX_BATCHED_MIGRATION 512 150142012e04SHuang Ying #endif 1502e5bfff8bSHuang Ying #define NR_MAX_MIGRATE_PAGES_RETRY 10 15032ef7dbb2SHuang Ying #define NR_MAX_MIGRATE_ASYNC_RETRY 3 15042ef7dbb2SHuang Ying #define NR_MAX_MIGRATE_SYNC_RETRY \ 15052ef7dbb2SHuang Ying (NR_MAX_MIGRATE_PAGES_RETRY - NR_MAX_MIGRATE_ASYNC_RETRY) 1506e5bfff8bSHuang Ying 15075b855937SHuang Ying struct migrate_pages_stats { 15085b855937SHuang Ying int nr_succeeded; /* Normal and large folios migrated successfully, in 15095b855937SHuang Ying units of base pages */ 15105b855937SHuang Ying int nr_failed_pages; /* Normal and large folios failed to be migrated, in 15115b855937SHuang Ying units of base pages. Untried folios aren't counted */ 15125b855937SHuang Ying int nr_thp_succeeded; /* THP migrated successfully */ 15135b855937SHuang Ying int nr_thp_failed; /* THP failed to be migrated */ 15145b855937SHuang Ying int nr_thp_split; /* THP split before migrating */ 1515a259945eSZi Yan int nr_split; /* Large folio (include THP) split before migrating */ 15165b855937SHuang Ying }; 15175b855937SHuang Ying 1518290408d4SNaoya Horiguchi /* 1519e5bfff8bSHuang Ying * Returns the number of hugetlb folios that were not migrated, or an error code 1520e5bfff8bSHuang Ying * after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no hugetlb folios are movable 1521e5bfff8bSHuang Ying * any more because the list has become empty or no retryable hugetlb folios 1522e5bfff8bSHuang Ying * exist any more. It is caller's responsibility to call putback_movable_pages() 1523e5bfff8bSHuang Ying * only if ret != 0. 1524e5bfff8bSHuang Ying */ 15254e096ae1SMatthew Wilcox (Oracle) static int migrate_hugetlbs(struct list_head *from, new_folio_t get_new_folio, 15264e096ae1SMatthew Wilcox (Oracle) free_folio_t put_new_folio, unsigned long private, 1527e5bfff8bSHuang Ying enum migrate_mode mode, int reason, 1528e5bfff8bSHuang Ying struct migrate_pages_stats *stats, 1529e5bfff8bSHuang Ying struct list_head *ret_folios) 1530e5bfff8bSHuang Ying { 1531e5bfff8bSHuang Ying int retry = 1; 1532e5bfff8bSHuang Ying int nr_failed = 0; 1533e5bfff8bSHuang Ying int nr_retry_pages = 0; 1534e5bfff8bSHuang Ying int pass = 0; 1535e5bfff8bSHuang Ying struct folio *folio, *folio2; 1536e5bfff8bSHuang Ying int rc, nr_pages; 1537e5bfff8bSHuang Ying 1538e5bfff8bSHuang Ying for (pass = 0; pass < NR_MAX_MIGRATE_PAGES_RETRY && retry; pass++) { 1539e5bfff8bSHuang Ying retry = 0; 1540e5bfff8bSHuang Ying nr_retry_pages = 0; 1541e5bfff8bSHuang Ying 1542e5bfff8bSHuang Ying list_for_each_entry_safe(folio, folio2, from, lru) { 1543e5bfff8bSHuang Ying if (!folio_test_hugetlb(folio)) 1544e5bfff8bSHuang Ying continue; 1545e5bfff8bSHuang Ying 1546e5bfff8bSHuang Ying nr_pages = folio_nr_pages(folio); 1547e5bfff8bSHuang Ying 1548e5bfff8bSHuang Ying cond_resched(); 1549e5bfff8bSHuang Ying 15506f7d760eSHuang Ying /* 15516f7d760eSHuang Ying * Migratability of hugepages depends on architectures and 15526f7d760eSHuang Ying * their size. This check is necessary because some callers 15536f7d760eSHuang Ying * of hugepage migration like soft offline and memory 15546f7d760eSHuang Ying * hotremove don't walk through page tables or check whether 15556f7d760eSHuang Ying * the hugepage is pmd-based or not before kicking migration. 15566f7d760eSHuang Ying */ 15576f7d760eSHuang Ying if (!hugepage_migration_supported(folio_hstate(folio))) { 15586f7d760eSHuang Ying nr_failed++; 15596f7d760eSHuang Ying stats->nr_failed_pages += nr_pages; 15606f7d760eSHuang Ying list_move_tail(&folio->lru, ret_folios); 15616f7d760eSHuang Ying continue; 15626f7d760eSHuang Ying } 15636f7d760eSHuang Ying 15644e096ae1SMatthew Wilcox (Oracle) rc = unmap_and_move_huge_page(get_new_folio, 15654e096ae1SMatthew Wilcox (Oracle) put_new_folio, private, 15664e096ae1SMatthew Wilcox (Oracle) folio, pass > 2, mode, 1567e5bfff8bSHuang Ying reason, ret_folios); 1568e5bfff8bSHuang Ying /* 1569e5bfff8bSHuang Ying * The rules are: 1570e5bfff8bSHuang Ying * Success: hugetlb folio will be put back 1571e5bfff8bSHuang Ying * -EAGAIN: stay on the from list 1572e5bfff8bSHuang Ying * -ENOMEM: stay on the from list 1573e5bfff8bSHuang Ying * Other errno: put on ret_folios list 1574e5bfff8bSHuang Ying */ 1575e5bfff8bSHuang Ying switch(rc) { 1576e5bfff8bSHuang Ying case -ENOMEM: 1577e5bfff8bSHuang Ying /* 1578e5bfff8bSHuang Ying * When memory is low, don't bother to try to migrate 1579e5bfff8bSHuang Ying * other folios, just exit. 1580e5bfff8bSHuang Ying */ 1581e5bfff8bSHuang Ying stats->nr_failed_pages += nr_pages + nr_retry_pages; 1582e5bfff8bSHuang Ying return -ENOMEM; 1583e5bfff8bSHuang Ying case -EAGAIN: 1584e5bfff8bSHuang Ying retry++; 1585e5bfff8bSHuang Ying nr_retry_pages += nr_pages; 1586e5bfff8bSHuang Ying break; 1587e5bfff8bSHuang Ying case MIGRATEPAGE_SUCCESS: 1588e5bfff8bSHuang Ying stats->nr_succeeded += nr_pages; 1589e5bfff8bSHuang Ying break; 1590e5bfff8bSHuang Ying default: 1591e5bfff8bSHuang Ying /* 1592e5bfff8bSHuang Ying * Permanent failure (-EBUSY, etc.): 1593e5bfff8bSHuang Ying * unlike -EAGAIN case, the failed folio is 1594e5bfff8bSHuang Ying * removed from migration folio list and not 1595e5bfff8bSHuang Ying * retried in the next outer loop. 1596e5bfff8bSHuang Ying */ 1597e5bfff8bSHuang Ying nr_failed++; 1598e5bfff8bSHuang Ying stats->nr_failed_pages += nr_pages; 1599e5bfff8bSHuang Ying break; 1600e5bfff8bSHuang Ying } 1601e5bfff8bSHuang Ying } 1602e5bfff8bSHuang Ying } 1603e5bfff8bSHuang Ying /* 1604e5bfff8bSHuang Ying * nr_failed is number of hugetlb folios failed to be migrated. After 1605e5bfff8bSHuang Ying * NR_MAX_MIGRATE_PAGES_RETRY attempts, give up and count retried hugetlb 1606e5bfff8bSHuang Ying * folios as failed. 1607e5bfff8bSHuang Ying */ 1608e5bfff8bSHuang Ying nr_failed += retry; 1609e5bfff8bSHuang Ying stats->nr_failed_pages += nr_retry_pages; 1610e5bfff8bSHuang Ying 1611e5bfff8bSHuang Ying return nr_failed; 1612e5bfff8bSHuang Ying } 1613e5bfff8bSHuang Ying 16145dfab109SHuang Ying /* 16155dfab109SHuang Ying * migrate_pages_batch() first unmaps folios in the from list as many as 16165dfab109SHuang Ying * possible, then move the unmapped folios. 1617fb3592c4SHuang Ying * 1618fb3592c4SHuang Ying * We only batch migration if mode == MIGRATE_ASYNC to avoid to wait a 1619fb3592c4SHuang Ying * lock or bit when we have locked more than one folio. Which may cause 1620fb3592c4SHuang Ying * deadlock (e.g., for loop device). So, if mode != MIGRATE_ASYNC, the 1621fb3592c4SHuang Ying * length of the from list must be <= 1. 16225dfab109SHuang Ying */ 16234e096ae1SMatthew Wilcox (Oracle) static int migrate_pages_batch(struct list_head *from, 16244e096ae1SMatthew Wilcox (Oracle) new_folio_t get_new_folio, free_folio_t put_new_folio, 16254e096ae1SMatthew Wilcox (Oracle) unsigned long private, enum migrate_mode mode, int reason, 16264e096ae1SMatthew Wilcox (Oracle) struct list_head *ret_folios, struct list_head *split_folios, 16274e096ae1SMatthew Wilcox (Oracle) struct migrate_pages_stats *stats, int nr_pass) 162842012e04SHuang Ying { 1629a21d2133SHuang Ying int retry = 1; 163042012e04SHuang Ying int thp_retry = 1; 163142012e04SHuang Ying int nr_failed = 0; 163242012e04SHuang Ying int nr_retry_pages = 0; 163342012e04SHuang Ying int pass = 0; 163442012e04SHuang Ying bool is_thp = false; 1635a259945eSZi Yan bool is_large = false; 16365dfab109SHuang Ying struct folio *folio, *folio2, *dst = NULL, *dst2; 1637a21d2133SHuang Ying int rc, rc_saved = 0, nr_pages; 16385dfab109SHuang Ying LIST_HEAD(unmap_folios); 16395dfab109SHuang Ying LIST_HEAD(dst_folios); 164042012e04SHuang Ying bool nosplit = (reason == MR_NUMA_MISPLACED); 164142012e04SHuang Ying 1642fb3592c4SHuang Ying VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC && 1643fb3592c4SHuang Ying !list_empty(from) && !list_is_singular(from)); 1644a21d2133SHuang Ying 1645124abcedSHuang Ying for (pass = 0; pass < nr_pass && retry; pass++) { 164642012e04SHuang Ying retry = 0; 164742012e04SHuang Ying thp_retry = 0; 164842012e04SHuang Ying nr_retry_pages = 0; 164942012e04SHuang Ying 165042012e04SHuang Ying list_for_each_entry_safe(folio, folio2, from, lru) { 1651a259945eSZi Yan is_large = folio_test_large(folio); 1652a259945eSZi Yan is_thp = is_large && folio_test_pmd_mappable(folio); 165342012e04SHuang Ying nr_pages = folio_nr_pages(folio); 165442012e04SHuang Ying 165542012e04SHuang Ying cond_resched(); 165642012e04SHuang Ying 16576f7d760eSHuang Ying /* 16586f7d760eSHuang Ying * Large folio migration might be unsupported or 16596f7d760eSHuang Ying * the allocation might be failed so we should retry 16606f7d760eSHuang Ying * on the same folio with the large folio split 16616f7d760eSHuang Ying * to normal folios. 16626f7d760eSHuang Ying * 16636f7d760eSHuang Ying * Split folios are put in split_folios, and 16646f7d760eSHuang Ying * we will migrate them after the rest of the 16656f7d760eSHuang Ying * list is processed. 16666f7d760eSHuang Ying */ 16676f7d760eSHuang Ying if (!thp_migration_supported() && is_thp) { 1668124abcedSHuang Ying nr_failed++; 16696f7d760eSHuang Ying stats->nr_thp_failed++; 1670a21d2133SHuang Ying if (!try_split_folio(folio, split_folios)) { 16716f7d760eSHuang Ying stats->nr_thp_split++; 1672a259945eSZi Yan stats->nr_split++; 16736f7d760eSHuang Ying continue; 16746f7d760eSHuang Ying } 16756f7d760eSHuang Ying stats->nr_failed_pages += nr_pages; 16766f7d760eSHuang Ying list_move_tail(&folio->lru, ret_folios); 16776f7d760eSHuang Ying continue; 16786f7d760eSHuang Ying } 16796f7d760eSHuang Ying 16804e096ae1SMatthew Wilcox (Oracle) rc = migrate_folio_unmap(get_new_folio, put_new_folio, 16814e096ae1SMatthew Wilcox (Oracle) private, folio, &dst, mode, reason, 16824e096ae1SMatthew Wilcox (Oracle) ret_folios); 168342012e04SHuang Ying /* 168442012e04SHuang Ying * The rules are: 168542012e04SHuang Ying * Success: folio will be freed 16865dfab109SHuang Ying * Unmap: folio will be put on unmap_folios list, 16875dfab109SHuang Ying * dst folio put on dst_folios list 168842012e04SHuang Ying * -EAGAIN: stay on the from list 168942012e04SHuang Ying * -ENOMEM: stay on the from list 169042012e04SHuang Ying * Other errno: put on ret_folios list 169142012e04SHuang Ying */ 169242012e04SHuang Ying switch(rc) { 169342012e04SHuang Ying case -ENOMEM: 169442012e04SHuang Ying /* 169542012e04SHuang Ying * When memory is low, don't bother to try to migrate 16965dfab109SHuang Ying * other folios, move unmapped folios, then exit. 169742012e04SHuang Ying */ 1698124abcedSHuang Ying nr_failed++; 169942012e04SHuang Ying stats->nr_thp_failed += is_thp; 170042012e04SHuang Ying /* Large folio NUMA faulting doesn't split to retry. */ 1701a259945eSZi Yan if (is_large && !nosplit) { 1702a21d2133SHuang Ying int ret = try_split_folio(folio, split_folios); 170342012e04SHuang Ying 170442012e04SHuang Ying if (!ret) { 170542012e04SHuang Ying stats->nr_thp_split += is_thp; 170649cac03aSZi Yan stats->nr_split++; 170742012e04SHuang Ying break; 170842012e04SHuang Ying } else if (reason == MR_LONGTERM_PIN && 170942012e04SHuang Ying ret == -EAGAIN) { 171042012e04SHuang Ying /* 171142012e04SHuang Ying * Try again to split large folio to 171242012e04SHuang Ying * mitigate the failure of longterm pinning. 171342012e04SHuang Ying */ 1714124abcedSHuang Ying retry++; 171542012e04SHuang Ying thp_retry += is_thp; 171642012e04SHuang Ying nr_retry_pages += nr_pages; 1717851ae642SHuang Ying /* Undo duplicated failure counting. */ 1718124abcedSHuang Ying nr_failed--; 1719851ae642SHuang Ying stats->nr_thp_failed -= is_thp; 172042012e04SHuang Ying break; 172142012e04SHuang Ying } 172242012e04SHuang Ying } 172342012e04SHuang Ying 172442012e04SHuang Ying stats->nr_failed_pages += nr_pages + nr_retry_pages; 172542012e04SHuang Ying /* nr_failed isn't updated for not used */ 172642012e04SHuang Ying stats->nr_thp_failed += thp_retry; 17275dfab109SHuang Ying rc_saved = rc; 17285dfab109SHuang Ying if (list_empty(&unmap_folios)) 172942012e04SHuang Ying goto out; 17305dfab109SHuang Ying else 17315dfab109SHuang Ying goto move; 173242012e04SHuang Ying case -EAGAIN: 173342012e04SHuang Ying retry++; 1734124abcedSHuang Ying thp_retry += is_thp; 173542012e04SHuang Ying nr_retry_pages += nr_pages; 173642012e04SHuang Ying break; 173742012e04SHuang Ying case MIGRATEPAGE_SUCCESS: 173842012e04SHuang Ying stats->nr_succeeded += nr_pages; 173942012e04SHuang Ying stats->nr_thp_succeeded += is_thp; 174042012e04SHuang Ying break; 17415dfab109SHuang Ying case MIGRATEPAGE_UNMAP: 17425dfab109SHuang Ying list_move_tail(&folio->lru, &unmap_folios); 17435dfab109SHuang Ying list_add_tail(&dst->lru, &dst_folios); 17445dfab109SHuang Ying break; 174542012e04SHuang Ying default: 174642012e04SHuang Ying /* 174742012e04SHuang Ying * Permanent failure (-EBUSY, etc.): 174842012e04SHuang Ying * unlike -EAGAIN case, the failed folio is 174942012e04SHuang Ying * removed from migration folio list and not 175042012e04SHuang Ying * retried in the next outer loop. 175142012e04SHuang Ying */ 175242012e04SHuang Ying nr_failed++; 1753124abcedSHuang Ying stats->nr_thp_failed += is_thp; 175442012e04SHuang Ying stats->nr_failed_pages += nr_pages; 175542012e04SHuang Ying break; 175642012e04SHuang Ying } 175742012e04SHuang Ying } 175842012e04SHuang Ying } 175942012e04SHuang Ying nr_failed += retry; 176042012e04SHuang Ying stats->nr_thp_failed += thp_retry; 176142012e04SHuang Ying stats->nr_failed_pages += nr_retry_pages; 17625dfab109SHuang Ying move: 17637e12beb8SHuang Ying /* Flush TLBs for all unmapped folios */ 17647e12beb8SHuang Ying try_to_unmap_flush(); 17657e12beb8SHuang Ying 17665dfab109SHuang Ying retry = 1; 1767124abcedSHuang Ying for (pass = 0; pass < nr_pass && retry; pass++) { 17685dfab109SHuang Ying retry = 0; 17695dfab109SHuang Ying thp_retry = 0; 17705dfab109SHuang Ying nr_retry_pages = 0; 17715dfab109SHuang Ying 17725dfab109SHuang Ying dst = list_first_entry(&dst_folios, struct folio, lru); 17735dfab109SHuang Ying dst2 = list_next_entry(dst, lru); 17745dfab109SHuang Ying list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) { 1775124abcedSHuang Ying is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio); 17765dfab109SHuang Ying nr_pages = folio_nr_pages(folio); 17775dfab109SHuang Ying 17785dfab109SHuang Ying cond_resched(); 17795dfab109SHuang Ying 17804e096ae1SMatthew Wilcox (Oracle) rc = migrate_folio_move(put_new_folio, private, 17815dfab109SHuang Ying folio, dst, mode, 17825dfab109SHuang Ying reason, ret_folios); 17835dfab109SHuang Ying /* 17845dfab109SHuang Ying * The rules are: 17855dfab109SHuang Ying * Success: folio will be freed 17865dfab109SHuang Ying * -EAGAIN: stay on the unmap_folios list 17875dfab109SHuang Ying * Other errno: put on ret_folios list 17885dfab109SHuang Ying */ 17895dfab109SHuang Ying switch(rc) { 17905dfab109SHuang Ying case -EAGAIN: 17915dfab109SHuang Ying retry++; 1792124abcedSHuang Ying thp_retry += is_thp; 17935dfab109SHuang Ying nr_retry_pages += nr_pages; 17945dfab109SHuang Ying break; 17955dfab109SHuang Ying case MIGRATEPAGE_SUCCESS: 17965dfab109SHuang Ying stats->nr_succeeded += nr_pages; 17975dfab109SHuang Ying stats->nr_thp_succeeded += is_thp; 17985dfab109SHuang Ying break; 17995dfab109SHuang Ying default: 18005dfab109SHuang Ying nr_failed++; 1801124abcedSHuang Ying stats->nr_thp_failed += is_thp; 18025dfab109SHuang Ying stats->nr_failed_pages += nr_pages; 18035dfab109SHuang Ying break; 18045dfab109SHuang Ying } 18055dfab109SHuang Ying dst = dst2; 18065dfab109SHuang Ying dst2 = list_next_entry(dst, lru); 18075dfab109SHuang Ying } 18085dfab109SHuang Ying } 18095dfab109SHuang Ying nr_failed += retry; 18105dfab109SHuang Ying stats->nr_thp_failed += thp_retry; 18115dfab109SHuang Ying stats->nr_failed_pages += nr_retry_pages; 18125dfab109SHuang Ying 1813124abcedSHuang Ying rc = rc_saved ? : nr_failed; 18145dfab109SHuang Ying out: 18155dfab109SHuang Ying /* Cleanup remaining folios */ 18165dfab109SHuang Ying dst = list_first_entry(&dst_folios, struct folio, lru); 18175dfab109SHuang Ying dst2 = list_next_entry(dst, lru); 18185dfab109SHuang Ying list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) { 1819eebb3dabSBaolin Wang int old_page_state = 0; 18205dfab109SHuang Ying struct anon_vma *anon_vma = NULL; 18215dfab109SHuang Ying 1822eebb3dabSBaolin Wang __migrate_folio_extract(dst, &old_page_state, &anon_vma); 1823eebb3dabSBaolin Wang migrate_folio_undo_src(folio, old_page_state & PAGE_WAS_MAPPED, 1824eebb3dabSBaolin Wang anon_vma, true, ret_folios); 18255dfab109SHuang Ying list_del(&dst->lru); 18264e096ae1SMatthew Wilcox (Oracle) migrate_folio_undo_dst(dst, true, put_new_folio, private); 18275dfab109SHuang Ying dst = dst2; 18285dfab109SHuang Ying dst2 = list_next_entry(dst, lru); 18295dfab109SHuang Ying } 18305dfab109SHuang Ying 183142012e04SHuang Ying return rc; 183242012e04SHuang Ying } 183342012e04SHuang Ying 18344e096ae1SMatthew Wilcox (Oracle) static int migrate_pages_sync(struct list_head *from, new_folio_t get_new_folio, 18354e096ae1SMatthew Wilcox (Oracle) free_folio_t put_new_folio, unsigned long private, 18364e096ae1SMatthew Wilcox (Oracle) enum migrate_mode mode, int reason, 18374e096ae1SMatthew Wilcox (Oracle) struct list_head *ret_folios, struct list_head *split_folios, 18384e096ae1SMatthew Wilcox (Oracle) struct migrate_pages_stats *stats) 18392ef7dbb2SHuang Ying { 18402ef7dbb2SHuang Ying int rc, nr_failed = 0; 18412ef7dbb2SHuang Ying LIST_HEAD(folios); 18422ef7dbb2SHuang Ying struct migrate_pages_stats astats; 18432ef7dbb2SHuang Ying 18442ef7dbb2SHuang Ying memset(&astats, 0, sizeof(astats)); 18452ef7dbb2SHuang Ying /* Try to migrate in batch with MIGRATE_ASYNC mode firstly */ 18464e096ae1SMatthew Wilcox (Oracle) rc = migrate_pages_batch(from, get_new_folio, put_new_folio, private, MIGRATE_ASYNC, 18472ef7dbb2SHuang Ying reason, &folios, split_folios, &astats, 18482ef7dbb2SHuang Ying NR_MAX_MIGRATE_ASYNC_RETRY); 18492ef7dbb2SHuang Ying stats->nr_succeeded += astats.nr_succeeded; 18502ef7dbb2SHuang Ying stats->nr_thp_succeeded += astats.nr_thp_succeeded; 18512ef7dbb2SHuang Ying stats->nr_thp_split += astats.nr_thp_split; 1852a259945eSZi Yan stats->nr_split += astats.nr_split; 18532ef7dbb2SHuang Ying if (rc < 0) { 18542ef7dbb2SHuang Ying stats->nr_failed_pages += astats.nr_failed_pages; 18552ef7dbb2SHuang Ying stats->nr_thp_failed += astats.nr_thp_failed; 18562ef7dbb2SHuang Ying list_splice_tail(&folios, ret_folios); 18572ef7dbb2SHuang Ying return rc; 18582ef7dbb2SHuang Ying } 18592ef7dbb2SHuang Ying stats->nr_thp_failed += astats.nr_thp_split; 1860a259945eSZi Yan /* 1861a259945eSZi Yan * Do not count rc, as pages will be retried below. 1862a259945eSZi Yan * Count nr_split only, since it includes nr_thp_split. 1863a259945eSZi Yan */ 1864a259945eSZi Yan nr_failed += astats.nr_split; 18652ef7dbb2SHuang Ying /* 18662ef7dbb2SHuang Ying * Fall back to migrate all failed folios one by one synchronously. All 18672ef7dbb2SHuang Ying * failed folios except split THPs will be retried, so their failure 18682ef7dbb2SHuang Ying * isn't counted 18692ef7dbb2SHuang Ying */ 18702ef7dbb2SHuang Ying list_splice_tail_init(&folios, from); 18712ef7dbb2SHuang Ying while (!list_empty(from)) { 18722ef7dbb2SHuang Ying list_move(from->next, &folios); 18734e096ae1SMatthew Wilcox (Oracle) rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio, 18742ef7dbb2SHuang Ying private, mode, reason, ret_folios, 18752ef7dbb2SHuang Ying split_folios, stats, NR_MAX_MIGRATE_SYNC_RETRY); 18762ef7dbb2SHuang Ying list_splice_tail_init(&folios, ret_folios); 18772ef7dbb2SHuang Ying if (rc < 0) 18782ef7dbb2SHuang Ying return rc; 18792ef7dbb2SHuang Ying nr_failed += rc; 18802ef7dbb2SHuang Ying } 18812ef7dbb2SHuang Ying 18822ef7dbb2SHuang Ying return nr_failed; 18832ef7dbb2SHuang Ying } 18842ef7dbb2SHuang Ying 1885e24f0b8fSChristoph Lameter /* 1886eaec4e63SHuang Ying * migrate_pages - migrate the folios specified in a list, to the free folios 1887c73e5c9cSSrivatsa S. Bhat * supplied as the target for the page migration 1888e24f0b8fSChristoph Lameter * 1889eaec4e63SHuang Ying * @from: The list of folios to be migrated. 18904e096ae1SMatthew Wilcox (Oracle) * @get_new_folio: The function used to allocate free folios to be used 1891eaec4e63SHuang Ying * as the target of the folio migration. 18924e096ae1SMatthew Wilcox (Oracle) * @put_new_folio: The function used to free target folios if migration 189368711a74SDavid Rientjes * fails, or NULL if no special handling is necessary. 18944e096ae1SMatthew Wilcox (Oracle) * @private: Private data to be passed on to get_new_folio() 1895c73e5c9cSSrivatsa S. Bhat * @mode: The migration mode that specifies the constraints for 1896eaec4e63SHuang Ying * folio migration, if any. 1897eaec4e63SHuang Ying * @reason: The reason for folio migration. 1898eaec4e63SHuang Ying * @ret_succeeded: Set to the number of folios migrated successfully if 18995ac95884SYang Shi * the caller passes a non-NULL pointer. 1900e24f0b8fSChristoph Lameter * 1901e5bfff8bSHuang Ying * The function returns after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no folios 1902e5bfff8bSHuang Ying * are movable any more because the list has become empty or no retryable folios 1903e5bfff8bSHuang Ying * exist any more. It is caller's responsibility to call putback_movable_pages() 1904e5bfff8bSHuang Ying * only if ret != 0. 1905e24f0b8fSChristoph Lameter * 1906eaec4e63SHuang Ying * Returns the number of {normal folio, large folio, hugetlb} that were not 1907eaec4e63SHuang Ying * migrated, or an error code. The number of large folio splits will be 1908eaec4e63SHuang Ying * considered as the number of non-migrated large folio, no matter how many 1909eaec4e63SHuang Ying * split folios of the large folio are migrated successfully. 1910e24f0b8fSChristoph Lameter */ 19114e096ae1SMatthew Wilcox (Oracle) int migrate_pages(struct list_head *from, new_folio_t get_new_folio, 19124e096ae1SMatthew Wilcox (Oracle) free_folio_t put_new_folio, unsigned long private, 19135ac95884SYang Shi enum migrate_mode mode, int reason, unsigned int *ret_succeeded) 1914e24f0b8fSChristoph Lameter { 191542012e04SHuang Ying int rc, rc_gather; 19162ef7dbb2SHuang Ying int nr_pages; 1917eaec4e63SHuang Ying struct folio *folio, *folio2; 191842012e04SHuang Ying LIST_HEAD(folios); 1919eaec4e63SHuang Ying LIST_HEAD(ret_folios); 1920a21d2133SHuang Ying LIST_HEAD(split_folios); 19215b855937SHuang Ying struct migrate_pages_stats stats; 19222d1db3b1SChristoph Lameter 19237bc1aec5SLiam Mark trace_mm_migrate_pages_start(mode, reason); 19247bc1aec5SLiam Mark 19255b855937SHuang Ying memset(&stats, 0, sizeof(stats)); 1926e24f0b8fSChristoph Lameter 19274e096ae1SMatthew Wilcox (Oracle) rc_gather = migrate_hugetlbs(from, get_new_folio, put_new_folio, private, 192842012e04SHuang Ying mode, reason, &stats, &ret_folios); 192942012e04SHuang Ying if (rc_gather < 0) 193095a402c3SChristoph Lameter goto out; 1931fb3592c4SHuang Ying 193242012e04SHuang Ying again: 193342012e04SHuang Ying nr_pages = 0; 1934b20a3503SChristoph Lameter list_for_each_entry_safe(folio, folio2, from, lru) { 1935e5bfff8bSHuang Ying /* Retried hugetlb folios will be kept in list */ 1936e5bfff8bSHuang Ying if (folio_test_hugetlb(folio)) { 1937e5bfff8bSHuang Ying list_move_tail(&folio->lru, &ret_folios); 1938e5bfff8bSHuang Ying continue; 1939eaec4e63SHuang Ying } 1940f430893bSMiaohe Lin 194142012e04SHuang Ying nr_pages += folio_nr_pages(folio); 19422ef7dbb2SHuang Ying if (nr_pages >= NR_MAX_BATCHED_MIGRATION) 1943e24f0b8fSChristoph Lameter break; 1944b20a3503SChristoph Lameter } 19452ef7dbb2SHuang Ying if (nr_pages >= NR_MAX_BATCHED_MIGRATION) 1946fb3592c4SHuang Ying list_cut_before(&folios, from, &folio2->lru); 194742012e04SHuang Ying else 194842012e04SHuang Ying list_splice_init(from, &folios); 19492ef7dbb2SHuang Ying if (mode == MIGRATE_ASYNC) 19504e096ae1SMatthew Wilcox (Oracle) rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio, 19514e096ae1SMatthew Wilcox (Oracle) private, mode, reason, &ret_folios, 19524e096ae1SMatthew Wilcox (Oracle) &split_folios, &stats, 1953a21d2133SHuang Ying NR_MAX_MIGRATE_PAGES_RETRY); 19542ef7dbb2SHuang Ying else 19554e096ae1SMatthew Wilcox (Oracle) rc = migrate_pages_sync(&folios, get_new_folio, put_new_folio, 19564e096ae1SMatthew Wilcox (Oracle) private, mode, reason, &ret_folios, 19574e096ae1SMatthew Wilcox (Oracle) &split_folios, &stats); 195842012e04SHuang Ying list_splice_tail_init(&folios, &ret_folios); 195942012e04SHuang Ying if (rc < 0) { 196042012e04SHuang Ying rc_gather = rc; 1961a21d2133SHuang Ying list_splice_tail(&split_folios, &ret_folios); 1962b20a3503SChristoph Lameter goto out; 1963b20a3503SChristoph Lameter } 1964a21d2133SHuang Ying if (!list_empty(&split_folios)) { 1965a21d2133SHuang Ying /* 1966a21d2133SHuang Ying * Failure isn't counted since all split folios of a large folio 1967a21d2133SHuang Ying * is counted as 1 failure already. And, we only try to migrate 1968a21d2133SHuang Ying * with minimal effort, force MIGRATE_ASYNC mode and retry once. 1969a21d2133SHuang Ying */ 19704e096ae1SMatthew Wilcox (Oracle) migrate_pages_batch(&split_folios, get_new_folio, 19714e096ae1SMatthew Wilcox (Oracle) put_new_folio, private, MIGRATE_ASYNC, reason, 19724e096ae1SMatthew Wilcox (Oracle) &ret_folios, NULL, &stats, 1); 1973a21d2133SHuang Ying list_splice_tail_init(&split_folios, &ret_folios); 1974a21d2133SHuang Ying } 197542012e04SHuang Ying rc_gather += rc; 197642012e04SHuang Ying if (!list_empty(from)) 197742012e04SHuang Ying goto again; 197895a402c3SChristoph Lameter out: 1979dd4ae78aSYang Shi /* 1980eaec4e63SHuang Ying * Put the permanent failure folio back to migration list, they 1981dd4ae78aSYang Shi * will be put back to the right list by the caller. 1982dd4ae78aSYang Shi */ 1983eaec4e63SHuang Ying list_splice(&ret_folios, from); 1984dd4ae78aSYang Shi 198503e5f82eSBaolin Wang /* 1986eaec4e63SHuang Ying * Return 0 in case all split folios of fail-to-migrate large folios 1987eaec4e63SHuang Ying * are migrated successfully. 198803e5f82eSBaolin Wang */ 198903e5f82eSBaolin Wang if (list_empty(from)) 199042012e04SHuang Ying rc_gather = 0; 199103e5f82eSBaolin Wang 19925b855937SHuang Ying count_vm_events(PGMIGRATE_SUCCESS, stats.nr_succeeded); 19935b855937SHuang Ying count_vm_events(PGMIGRATE_FAIL, stats.nr_failed_pages); 19945b855937SHuang Ying count_vm_events(THP_MIGRATION_SUCCESS, stats.nr_thp_succeeded); 19955b855937SHuang Ying count_vm_events(THP_MIGRATION_FAIL, stats.nr_thp_failed); 19965b855937SHuang Ying count_vm_events(THP_MIGRATION_SPLIT, stats.nr_thp_split); 19975b855937SHuang Ying trace_mm_migrate_pages(stats.nr_succeeded, stats.nr_failed_pages, 19985b855937SHuang Ying stats.nr_thp_succeeded, stats.nr_thp_failed, 199949cac03aSZi Yan stats.nr_thp_split, stats.nr_split, mode, 200049cac03aSZi Yan reason); 20017b2a2d4aSMel Gorman 20025ac95884SYang Shi if (ret_succeeded) 20035b855937SHuang Ying *ret_succeeded = stats.nr_succeeded; 20045ac95884SYang Shi 200542012e04SHuang Ying return rc_gather; 2006b20a3503SChristoph Lameter } 2007b20a3503SChristoph Lameter 20084e096ae1SMatthew Wilcox (Oracle) struct folio *alloc_migration_target(struct folio *src, unsigned long private) 2009b4b38223SJoonsoo Kim { 201019fc7bedSJoonsoo Kim struct migration_target_control *mtc; 201119fc7bedSJoonsoo Kim gfp_t gfp_mask; 2012b4b38223SJoonsoo Kim unsigned int order = 0; 201319fc7bedSJoonsoo Kim int nid; 201419fc7bedSJoonsoo Kim int zidx; 201519fc7bedSJoonsoo Kim 201619fc7bedSJoonsoo Kim mtc = (struct migration_target_control *)private; 201719fc7bedSJoonsoo Kim gfp_mask = mtc->gfp_mask; 201819fc7bedSJoonsoo Kim nid = mtc->nid; 201919fc7bedSJoonsoo Kim if (nid == NUMA_NO_NODE) 20204e096ae1SMatthew Wilcox (Oracle) nid = folio_nid(src); 2021b4b38223SJoonsoo Kim 20224e096ae1SMatthew Wilcox (Oracle) if (folio_test_hugetlb(src)) { 20234e096ae1SMatthew Wilcox (Oracle) struct hstate *h = folio_hstate(src); 2024d92bbc27SJoonsoo Kim 202519fc7bedSJoonsoo Kim gfp_mask = htlb_modify_alloc_mask(h, gfp_mask); 20264e096ae1SMatthew Wilcox (Oracle) return alloc_hugetlb_folio_nodemask(h, nid, 2027e37d3e83SSidhartha Kumar mtc->nmask, gfp_mask); 2028d92bbc27SJoonsoo Kim } 2029b4b38223SJoonsoo Kim 20304e096ae1SMatthew Wilcox (Oracle) if (folio_test_large(src)) { 20319933a0c8SJoonsoo Kim /* 20329933a0c8SJoonsoo Kim * clear __GFP_RECLAIM to make the migration callback 20339933a0c8SJoonsoo Kim * consistent with regular THP allocations. 20349933a0c8SJoonsoo Kim */ 20359933a0c8SJoonsoo Kim gfp_mask &= ~__GFP_RECLAIM; 2036b4b38223SJoonsoo Kim gfp_mask |= GFP_TRANSHUGE; 20374e096ae1SMatthew Wilcox (Oracle) order = folio_order(src); 2038b4b38223SJoonsoo Kim } 20394e096ae1SMatthew Wilcox (Oracle) zidx = zone_idx(folio_zone(src)); 204019fc7bedSJoonsoo Kim if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE) 2041b4b38223SJoonsoo Kim gfp_mask |= __GFP_HIGHMEM; 2042b4b38223SJoonsoo Kim 20434e096ae1SMatthew Wilcox (Oracle) return __folio_alloc(gfp_mask, order, nid, mtc->nmask); 2044b4b38223SJoonsoo Kim } 2045b4b38223SJoonsoo Kim 2046742755a1SChristoph Lameter #ifdef CONFIG_NUMA 2047742755a1SChristoph Lameter 2048a49bd4d7SMichal Hocko static int store_status(int __user *status, int start, int value, int nr) 2049742755a1SChristoph Lameter { 2050a49bd4d7SMichal Hocko while (nr-- > 0) { 2051a49bd4d7SMichal Hocko if (put_user(value, status + start)) 2052a49bd4d7SMichal Hocko return -EFAULT; 2053a49bd4d7SMichal Hocko start++; 2054a49bd4d7SMichal Hocko } 2055742755a1SChristoph Lameter 2056a49bd4d7SMichal Hocko return 0; 2057a49bd4d7SMichal Hocko } 2058742755a1SChristoph Lameter 2059ec47e250SGregory Price static int do_move_pages_to_node(struct list_head *pagelist, int node) 2060a49bd4d7SMichal Hocko { 2061a49bd4d7SMichal Hocko int err; 2062a0976311SJoonsoo Kim struct migration_target_control mtc = { 2063a0976311SJoonsoo Kim .nid = node, 2064a0976311SJoonsoo Kim .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 2065a0976311SJoonsoo Kim }; 2066742755a1SChristoph Lameter 2067a0976311SJoonsoo Kim err = migrate_pages(pagelist, alloc_migration_target, NULL, 20685ac95884SYang Shi (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL); 2069a49bd4d7SMichal Hocko if (err) 2070a49bd4d7SMichal Hocko putback_movable_pages(pagelist); 2071a49bd4d7SMichal Hocko return err; 2072742755a1SChristoph Lameter } 2073742755a1SChristoph Lameter 2074742755a1SChristoph Lameter /* 2075a49bd4d7SMichal Hocko * Resolves the given address to a struct page, isolates it from the LRU and 2076a49bd4d7SMichal Hocko * puts it to the given pagelist. 2077e0153fc2SYang Shi * Returns: 2078e0153fc2SYang Shi * errno - if the page cannot be found/isolated 2079e0153fc2SYang Shi * 0 - when it doesn't have to be migrated because it is already on the 2080e0153fc2SYang Shi * target node 2081e0153fc2SYang Shi * 1 - when it has been queued 2082742755a1SChristoph Lameter */ 2083428e106aSKirill A. Shutemov static int add_page_for_migration(struct mm_struct *mm, const void __user *p, 2084a49bd4d7SMichal Hocko int node, struct list_head *pagelist, bool migrate_all) 2085742755a1SChristoph Lameter { 2086742755a1SChristoph Lameter struct vm_area_struct *vma; 2087428e106aSKirill A. Shutemov unsigned long addr; 2088742755a1SChristoph Lameter struct page *page; 2089d64cfccbSKefeng Wang struct folio *folio; 2090a49bd4d7SMichal Hocko int err; 2091742755a1SChristoph Lameter 2092d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 2093428e106aSKirill A. Shutemov addr = (unsigned long)untagged_addr_remote(mm, p); 2094428e106aSKirill A. Shutemov 2095742755a1SChristoph Lameter err = -EFAULT; 2096cb1c37b1SMiaohe Lin vma = vma_lookup(mm, addr); 2097cb1c37b1SMiaohe Lin if (!vma || !vma_migratable(vma)) 2098a49bd4d7SMichal Hocko goto out; 2099742755a1SChristoph Lameter 2100d899844eSKirill A. Shutemov /* FOLL_DUMP to ignore special (like zero) pages */ 210187d2762eSMiaohe Lin page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 210289f5b7daSLinus Torvalds 210389f5b7daSLinus Torvalds err = PTR_ERR(page); 210489f5b7daSLinus Torvalds if (IS_ERR(page)) 2105a49bd4d7SMichal Hocko goto out; 210689f5b7daSLinus Torvalds 2107742755a1SChristoph Lameter err = -ENOENT; 2108f7091ed6SHaiyue Wang if (!page) 2109a49bd4d7SMichal Hocko goto out; 2110742755a1SChristoph Lameter 2111d64cfccbSKefeng Wang folio = page_folio(page); 2112d64cfccbSKefeng Wang if (folio_is_zone_device(folio)) 2113d64cfccbSKefeng Wang goto out_putfolio; 2114f7091ed6SHaiyue Wang 2115a49bd4d7SMichal Hocko err = 0; 2116d64cfccbSKefeng Wang if (folio_nid(folio) == node) 2117d64cfccbSKefeng Wang goto out_putfolio; 2118742755a1SChristoph Lameter 2119742755a1SChristoph Lameter err = -EACCES; 2120a49bd4d7SMichal Hocko if (page_mapcount(page) > 1 && !migrate_all) 2121d64cfccbSKefeng Wang goto out_putfolio; 2122742755a1SChristoph Lameter 2123f7f9c00dSBaolin Wang err = -EBUSY; 2124fa1df3f6SKefeng Wang if (folio_test_hugetlb(folio)) { 2125fa1df3f6SKefeng Wang if (isolate_hugetlb(folio, pagelist)) 2126fa1df3f6SKefeng Wang err = 1; 2127fa1df3f6SKefeng Wang } else { 2128fa1df3f6SKefeng Wang if (!folio_isolate_lru(folio)) 2129d64cfccbSKefeng Wang goto out_putfolio; 2130a49bd4d7SMichal Hocko 2131e0153fc2SYang Shi err = 1; 2132d64cfccbSKefeng Wang list_add_tail(&folio->lru, pagelist); 2133d64cfccbSKefeng Wang node_stat_mod_folio(folio, 2134d64cfccbSKefeng Wang NR_ISOLATED_ANON + folio_is_file_lru(folio), 2135d64cfccbSKefeng Wang folio_nr_pages(folio)); 21366d9c285aSKOSAKI Motohiro } 2137d64cfccbSKefeng Wang out_putfolio: 2138742755a1SChristoph Lameter /* 2139d64cfccbSKefeng Wang * Either remove the duplicate refcount from folio_isolate_lru() 2140d64cfccbSKefeng Wang * or drop the folio ref if it was not isolated. 2141742755a1SChristoph Lameter */ 2142d64cfccbSKefeng Wang folio_put(folio); 2143a49bd4d7SMichal Hocko out: 2144d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 2145742755a1SChristoph Lameter return err; 2146742755a1SChristoph Lameter } 2147742755a1SChristoph Lameter 2148ec47e250SGregory Price static int move_pages_and_store_status(int node, 21497ca8783aSWei Yang struct list_head *pagelist, int __user *status, 21507ca8783aSWei Yang int start, int i, unsigned long nr_pages) 21517ca8783aSWei Yang { 21527ca8783aSWei Yang int err; 21537ca8783aSWei Yang 21545d7ae891SWei Yang if (list_empty(pagelist)) 21555d7ae891SWei Yang return 0; 21565d7ae891SWei Yang 2157ec47e250SGregory Price err = do_move_pages_to_node(pagelist, node); 21587ca8783aSWei Yang if (err) { 21597ca8783aSWei Yang /* 21607ca8783aSWei Yang * Positive err means the number of failed 21617ca8783aSWei Yang * pages to migrate. Since we are going to 21627ca8783aSWei Yang * abort and return the number of non-migrated 2163ab9dd4f8SLong Li * pages, so need to include the rest of the 21647ca8783aSWei Yang * nr_pages that have not been attempted as 21657ca8783aSWei Yang * well. 21667ca8783aSWei Yang */ 21677ca8783aSWei Yang if (err > 0) 2168a7504ed1SHuang Ying err += nr_pages - i; 21697ca8783aSWei Yang return err; 21707ca8783aSWei Yang } 21717ca8783aSWei Yang return store_status(status, start, node, i - start); 21727ca8783aSWei Yang } 21737ca8783aSWei Yang 2174742755a1SChristoph Lameter /* 21755e9a0f02SBrice Goglin * Migrate an array of page address onto an array of nodes and fill 21765e9a0f02SBrice Goglin * the corresponding array of status. 21775e9a0f02SBrice Goglin */ 21783268c63eSChristoph Lameter static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, 21795e9a0f02SBrice Goglin unsigned long nr_pages, 21805e9a0f02SBrice Goglin const void __user * __user *pages, 21815e9a0f02SBrice Goglin const int __user *nodes, 21825e9a0f02SBrice Goglin int __user *status, int flags) 21835e9a0f02SBrice Goglin { 2184229e2253SGregory Price compat_uptr_t __user *compat_pages = (void __user *)pages; 2185a49bd4d7SMichal Hocko int current_node = NUMA_NO_NODE; 2186a49bd4d7SMichal Hocko LIST_HEAD(pagelist); 2187a49bd4d7SMichal Hocko int start, i; 2188a49bd4d7SMichal Hocko int err = 0, err1; 218935282a2dSBrice Goglin 2190361a2a22SMinchan Kim lru_cache_disable(); 219135282a2dSBrice Goglin 2192a49bd4d7SMichal Hocko for (i = start = 0; i < nr_pages; i++) { 21935e9a0f02SBrice Goglin const void __user *p; 21945e9a0f02SBrice Goglin int node; 21955e9a0f02SBrice Goglin 21963140a227SBrice Goglin err = -EFAULT; 2197229e2253SGregory Price if (in_compat_syscall()) { 2198229e2253SGregory Price compat_uptr_t cp; 2199229e2253SGregory Price 2200229e2253SGregory Price if (get_user(cp, compat_pages + i)) 2201229e2253SGregory Price goto out_flush; 2202229e2253SGregory Price 2203229e2253SGregory Price p = compat_ptr(cp); 2204229e2253SGregory Price } else { 2205a49bd4d7SMichal Hocko if (get_user(p, pages + i)) 2206a49bd4d7SMichal Hocko goto out_flush; 2207229e2253SGregory Price } 2208a49bd4d7SMichal Hocko if (get_user(node, nodes + i)) 2209a49bd4d7SMichal Hocko goto out_flush; 22105e9a0f02SBrice Goglin 22115e9a0f02SBrice Goglin err = -ENODEV; 22126f5a55f1SLinus Torvalds if (node < 0 || node >= MAX_NUMNODES) 2213a49bd4d7SMichal Hocko goto out_flush; 2214389162c2SLai Jiangshan if (!node_state(node, N_MEMORY)) 2215a49bd4d7SMichal Hocko goto out_flush; 22165e9a0f02SBrice Goglin 22175e9a0f02SBrice Goglin err = -EACCES; 22185e9a0f02SBrice Goglin if (!node_isset(node, task_nodes)) 2219a49bd4d7SMichal Hocko goto out_flush; 22205e9a0f02SBrice Goglin 2221a49bd4d7SMichal Hocko if (current_node == NUMA_NO_NODE) { 2222a49bd4d7SMichal Hocko current_node = node; 2223a49bd4d7SMichal Hocko start = i; 2224a49bd4d7SMichal Hocko } else if (node != current_node) { 2225ec47e250SGregory Price err = move_pages_and_store_status(current_node, 22267ca8783aSWei Yang &pagelist, status, start, i, nr_pages); 2227a49bd4d7SMichal Hocko if (err) 2228a49bd4d7SMichal Hocko goto out; 2229a49bd4d7SMichal Hocko start = i; 2230a49bd4d7SMichal Hocko current_node = node; 22315e9a0f02SBrice Goglin } 22325e9a0f02SBrice Goglin 2233a49bd4d7SMichal Hocko /* 2234a49bd4d7SMichal Hocko * Errors in the page lookup or isolation are not fatal and we simply 2235a49bd4d7SMichal Hocko * report them via status 2236a49bd4d7SMichal Hocko */ 2237428e106aSKirill A. Shutemov err = add_page_for_migration(mm, p, current_node, &pagelist, 2238428e106aSKirill A. Shutemov flags & MPOL_MF_MOVE_ALL); 2239e0153fc2SYang Shi 2240d08221a0SWei Yang if (err > 0) { 2241e0153fc2SYang Shi /* The page is successfully queued for migration */ 2242e0153fc2SYang Shi continue; 2243e0153fc2SYang Shi } 22443140a227SBrice Goglin 2245d08221a0SWei Yang /* 224665462462SJohn Hubbard * The move_pages() man page does not have an -EEXIST choice, so 224765462462SJohn Hubbard * use -EFAULT instead. 224865462462SJohn Hubbard */ 224965462462SJohn Hubbard if (err == -EEXIST) 225065462462SJohn Hubbard err = -EFAULT; 225165462462SJohn Hubbard 225265462462SJohn Hubbard /* 2253d08221a0SWei Yang * If the page is already on the target node (!err), store the 2254d08221a0SWei Yang * node, otherwise, store the err. 2255d08221a0SWei Yang */ 2256d08221a0SWei Yang err = store_status(status, i, err ? : current_node, 1); 2257a49bd4d7SMichal Hocko if (err) 2258a49bd4d7SMichal Hocko goto out_flush; 22593140a227SBrice Goglin 2260ec47e250SGregory Price err = move_pages_and_store_status(current_node, &pagelist, 22617ca8783aSWei Yang status, start, i, nr_pages); 2262a7504ed1SHuang Ying if (err) { 2263a7504ed1SHuang Ying /* We have accounted for page i */ 2264a7504ed1SHuang Ying if (err > 0) 2265a7504ed1SHuang Ying err--; 2266a49bd4d7SMichal Hocko goto out; 2267a7504ed1SHuang Ying } 2268a49bd4d7SMichal Hocko current_node = NUMA_NO_NODE; 22693140a227SBrice Goglin } 2270a49bd4d7SMichal Hocko out_flush: 2271a49bd4d7SMichal Hocko /* Make sure we do not overwrite the existing error */ 2272ec47e250SGregory Price err1 = move_pages_and_store_status(current_node, &pagelist, 22737ca8783aSWei Yang status, start, i, nr_pages); 2274dfe9aa23SWei Yang if (err >= 0) 2275a49bd4d7SMichal Hocko err = err1; 22765e9a0f02SBrice Goglin out: 2277361a2a22SMinchan Kim lru_cache_enable(); 22785e9a0f02SBrice Goglin return err; 22795e9a0f02SBrice Goglin } 22805e9a0f02SBrice Goglin 22815e9a0f02SBrice Goglin /* 22822f007e74SBrice Goglin * Determine the nodes of an array of pages and store it in an array of status. 2283742755a1SChristoph Lameter */ 228480bba129SBrice Goglin static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, 228580bba129SBrice Goglin const void __user **pages, int *status) 2286742755a1SChristoph Lameter { 22872f007e74SBrice Goglin unsigned long i; 2288742755a1SChristoph Lameter 2289d8ed45c5SMichel Lespinasse mmap_read_lock(mm); 22902f007e74SBrice Goglin 22912f007e74SBrice Goglin for (i = 0; i < nr_pages; i++) { 229280bba129SBrice Goglin unsigned long addr = (unsigned long)(*pages); 22932f007e74SBrice Goglin struct vm_area_struct *vma; 22942f007e74SBrice Goglin struct page *page; 2295c095adbcSKOSAKI Motohiro int err = -EFAULT; 22962f007e74SBrice Goglin 2297059b8b48SLiam Howlett vma = vma_lookup(mm, addr); 2298059b8b48SLiam Howlett if (!vma) 2299742755a1SChristoph Lameter goto set_status; 2300742755a1SChristoph Lameter 2301d899844eSKirill A. Shutemov /* FOLL_DUMP to ignore special (like zero) pages */ 230216fd6b31SBaolin Wang page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 230389f5b7daSLinus Torvalds 230489f5b7daSLinus Torvalds err = PTR_ERR(page); 230589f5b7daSLinus Torvalds if (IS_ERR(page)) 230689f5b7daSLinus Torvalds goto set_status; 230789f5b7daSLinus Torvalds 2308f7091ed6SHaiyue Wang err = -ENOENT; 2309f7091ed6SHaiyue Wang if (!page) 2310f7091ed6SHaiyue Wang goto set_status; 2311f7091ed6SHaiyue Wang 2312f7091ed6SHaiyue Wang if (!is_zone_device_page(page)) 23134cd61484SMiaohe Lin err = page_to_nid(page); 2314f7091ed6SHaiyue Wang 23154cd61484SMiaohe Lin put_page(page); 2316742755a1SChristoph Lameter set_status: 231780bba129SBrice Goglin *status = err; 231880bba129SBrice Goglin 231980bba129SBrice Goglin pages++; 232080bba129SBrice Goglin status++; 232180bba129SBrice Goglin } 232280bba129SBrice Goglin 2323d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 232480bba129SBrice Goglin } 232580bba129SBrice Goglin 23265b1b561bSArnd Bergmann static int get_compat_pages_array(const void __user *chunk_pages[], 23275b1b561bSArnd Bergmann const void __user * __user *pages, 23285b1b561bSArnd Bergmann unsigned long chunk_nr) 23295b1b561bSArnd Bergmann { 23305b1b561bSArnd Bergmann compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages; 23315b1b561bSArnd Bergmann compat_uptr_t p; 23325b1b561bSArnd Bergmann int i; 23335b1b561bSArnd Bergmann 23345b1b561bSArnd Bergmann for (i = 0; i < chunk_nr; i++) { 23355b1b561bSArnd Bergmann if (get_user(p, pages32 + i)) 23365b1b561bSArnd Bergmann return -EFAULT; 23375b1b561bSArnd Bergmann chunk_pages[i] = compat_ptr(p); 23385b1b561bSArnd Bergmann } 23395b1b561bSArnd Bergmann 23405b1b561bSArnd Bergmann return 0; 23415b1b561bSArnd Bergmann } 23425b1b561bSArnd Bergmann 234380bba129SBrice Goglin /* 234480bba129SBrice Goglin * Determine the nodes of a user array of pages and store it in 234580bba129SBrice Goglin * a user array of status. 234680bba129SBrice Goglin */ 234780bba129SBrice Goglin static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, 234880bba129SBrice Goglin const void __user * __user *pages, 234980bba129SBrice Goglin int __user *status) 235080bba129SBrice Goglin { 23513eefb826SMiaohe Lin #define DO_PAGES_STAT_CHUNK_NR 16UL 235280bba129SBrice Goglin const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; 235380bba129SBrice Goglin int chunk_status[DO_PAGES_STAT_CHUNK_NR]; 235480bba129SBrice Goglin 235587b8d1adSH. Peter Anvin while (nr_pages) { 23563eefb826SMiaohe Lin unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR); 235787b8d1adSH. Peter Anvin 23585b1b561bSArnd Bergmann if (in_compat_syscall()) { 23595b1b561bSArnd Bergmann if (get_compat_pages_array(chunk_pages, pages, 23605b1b561bSArnd Bergmann chunk_nr)) 236187b8d1adSH. Peter Anvin break; 23625b1b561bSArnd Bergmann } else { 23635b1b561bSArnd Bergmann if (copy_from_user(chunk_pages, pages, 23645b1b561bSArnd Bergmann chunk_nr * sizeof(*chunk_pages))) 23655b1b561bSArnd Bergmann break; 23665b1b561bSArnd Bergmann } 236780bba129SBrice Goglin 236880bba129SBrice Goglin do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); 236980bba129SBrice Goglin 237087b8d1adSH. Peter Anvin if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) 237187b8d1adSH. Peter Anvin break; 2372742755a1SChristoph Lameter 237387b8d1adSH. Peter Anvin pages += chunk_nr; 237487b8d1adSH. Peter Anvin status += chunk_nr; 237587b8d1adSH. Peter Anvin nr_pages -= chunk_nr; 237687b8d1adSH. Peter Anvin } 237787b8d1adSH. Peter Anvin return nr_pages ? -EFAULT : 0; 2378742755a1SChristoph Lameter } 2379742755a1SChristoph Lameter 23804dc200ceSMiaohe Lin static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes) 23814dc200ceSMiaohe Lin { 23824dc200ceSMiaohe Lin struct task_struct *task; 23834dc200ceSMiaohe Lin struct mm_struct *mm; 23844dc200ceSMiaohe Lin 23854dc200ceSMiaohe Lin /* 23864dc200ceSMiaohe Lin * There is no need to check if current process has the right to modify 23874dc200ceSMiaohe Lin * the specified process when they are same. 23884dc200ceSMiaohe Lin */ 23894dc200ceSMiaohe Lin if (!pid) { 23904dc200ceSMiaohe Lin mmget(current->mm); 23914dc200ceSMiaohe Lin *mem_nodes = cpuset_mems_allowed(current); 23924dc200ceSMiaohe Lin return current->mm; 23934dc200ceSMiaohe Lin } 23944dc200ceSMiaohe Lin 23954dc200ceSMiaohe Lin /* Find the mm_struct */ 23964dc200ceSMiaohe Lin rcu_read_lock(); 23974dc200ceSMiaohe Lin task = find_task_by_vpid(pid); 23984dc200ceSMiaohe Lin if (!task) { 23994dc200ceSMiaohe Lin rcu_read_unlock(); 24004dc200ceSMiaohe Lin return ERR_PTR(-ESRCH); 24014dc200ceSMiaohe Lin } 24024dc200ceSMiaohe Lin get_task_struct(task); 24034dc200ceSMiaohe Lin 24044dc200ceSMiaohe Lin /* 24054dc200ceSMiaohe Lin * Check if this process has the right to modify the specified 24064dc200ceSMiaohe Lin * process. Use the regular "ptrace_may_access()" checks. 24074dc200ceSMiaohe Lin */ 24084dc200ceSMiaohe Lin if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { 24094dc200ceSMiaohe Lin rcu_read_unlock(); 24104dc200ceSMiaohe Lin mm = ERR_PTR(-EPERM); 24114dc200ceSMiaohe Lin goto out; 24124dc200ceSMiaohe Lin } 24134dc200ceSMiaohe Lin rcu_read_unlock(); 24144dc200ceSMiaohe Lin 24154dc200ceSMiaohe Lin mm = ERR_PTR(security_task_movememory(task)); 24164dc200ceSMiaohe Lin if (IS_ERR(mm)) 24174dc200ceSMiaohe Lin goto out; 24184dc200ceSMiaohe Lin *mem_nodes = cpuset_mems_allowed(task); 24194dc200ceSMiaohe Lin mm = get_task_mm(task); 24204dc200ceSMiaohe Lin out: 24214dc200ceSMiaohe Lin put_task_struct(task); 24224dc200ceSMiaohe Lin if (!mm) 24234dc200ceSMiaohe Lin mm = ERR_PTR(-EINVAL); 24244dc200ceSMiaohe Lin return mm; 24254dc200ceSMiaohe Lin } 24264dc200ceSMiaohe Lin 2427742755a1SChristoph Lameter /* 2428742755a1SChristoph Lameter * Move a list of pages in the address space of the currently executing 2429742755a1SChristoph Lameter * process. 2430742755a1SChristoph Lameter */ 24317addf443SDominik Brodowski static int kernel_move_pages(pid_t pid, unsigned long nr_pages, 24327addf443SDominik Brodowski const void __user * __user *pages, 24337addf443SDominik Brodowski const int __user *nodes, 24347addf443SDominik Brodowski int __user *status, int flags) 2435742755a1SChristoph Lameter { 2436742755a1SChristoph Lameter struct mm_struct *mm; 24375e9a0f02SBrice Goglin int err; 24383268c63eSChristoph Lameter nodemask_t task_nodes; 2439742755a1SChristoph Lameter 2440742755a1SChristoph Lameter /* Check flags */ 2441742755a1SChristoph Lameter if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) 2442742755a1SChristoph Lameter return -EINVAL; 2443742755a1SChristoph Lameter 2444742755a1SChristoph Lameter if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) 2445742755a1SChristoph Lameter return -EPERM; 2446742755a1SChristoph Lameter 24474dc200ceSMiaohe Lin mm = find_mm_struct(pid, &task_nodes); 24484dc200ceSMiaohe Lin if (IS_ERR(mm)) 24494dc200ceSMiaohe Lin return PTR_ERR(mm); 24506e8b09eaSSasha Levin 24513268c63eSChristoph Lameter if (nodes) 24523268c63eSChristoph Lameter err = do_pages_move(mm, task_nodes, nr_pages, pages, 24533268c63eSChristoph Lameter nodes, status, flags); 24543268c63eSChristoph Lameter else 24555e9a0f02SBrice Goglin err = do_pages_stat(mm, nr_pages, pages, status); 24563268c63eSChristoph Lameter 24573268c63eSChristoph Lameter mmput(mm); 24583268c63eSChristoph Lameter return err; 2459742755a1SChristoph Lameter } 2460742755a1SChristoph Lameter 24617addf443SDominik Brodowski SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, 24627addf443SDominik Brodowski const void __user * __user *, pages, 24637addf443SDominik Brodowski const int __user *, nodes, 24647addf443SDominik Brodowski int __user *, status, int, flags) 24657addf443SDominik Brodowski { 24667addf443SDominik Brodowski return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); 24677addf443SDominik Brodowski } 24687addf443SDominik Brodowski 24697039e1dbSPeter Zijlstra #ifdef CONFIG_NUMA_BALANCING 24707039e1dbSPeter Zijlstra /* 24717039e1dbSPeter Zijlstra * Returns true if this is a safe migration target node for misplaced NUMA 2472bc53008eSWei Yang * pages. Currently it only checks the watermarks which is crude. 24737039e1dbSPeter Zijlstra */ 24747039e1dbSPeter Zijlstra static bool migrate_balanced_pgdat(struct pglist_data *pgdat, 24753abef4e6SMel Gorman unsigned long nr_migrate_pages) 24767039e1dbSPeter Zijlstra { 24777039e1dbSPeter Zijlstra int z; 2478599d0c95SMel Gorman 24797039e1dbSPeter Zijlstra for (z = pgdat->nr_zones - 1; z >= 0; z--) { 24807039e1dbSPeter Zijlstra struct zone *zone = pgdat->node_zones + z; 24817039e1dbSPeter Zijlstra 2482bc53008eSWei Yang if (!managed_zone(zone)) 24837039e1dbSPeter Zijlstra continue; 24847039e1dbSPeter Zijlstra 24857039e1dbSPeter Zijlstra /* Avoid waking kswapd by allocating pages_to_migrate pages. */ 24867039e1dbSPeter Zijlstra if (!zone_watermark_ok(zone, 0, 24877039e1dbSPeter Zijlstra high_wmark_pages(zone) + 24887039e1dbSPeter Zijlstra nr_migrate_pages, 2489bfe9d006SHuang Ying ZONE_MOVABLE, 0)) 24907039e1dbSPeter Zijlstra continue; 24917039e1dbSPeter Zijlstra return true; 24927039e1dbSPeter Zijlstra } 24937039e1dbSPeter Zijlstra return false; 24947039e1dbSPeter Zijlstra } 24957039e1dbSPeter Zijlstra 24964e096ae1SMatthew Wilcox (Oracle) static struct folio *alloc_misplaced_dst_folio(struct folio *src, 2497666feb21SMichal Hocko unsigned long data) 24987039e1dbSPeter Zijlstra { 24997039e1dbSPeter Zijlstra int nid = (int) data; 25004e096ae1SMatthew Wilcox (Oracle) int order = folio_order(src); 2501c185e494SMatthew Wilcox (Oracle) gfp_t gfp = __GFP_THISNODE; 25027039e1dbSPeter Zijlstra 2503c185e494SMatthew Wilcox (Oracle) if (order > 0) 2504c185e494SMatthew Wilcox (Oracle) gfp |= GFP_TRANSHUGE_LIGHT; 2505c185e494SMatthew Wilcox (Oracle) else { 2506c185e494SMatthew Wilcox (Oracle) gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY | 2507c185e494SMatthew Wilcox (Oracle) __GFP_NOWARN; 2508c185e494SMatthew Wilcox (Oracle) gfp &= ~__GFP_RECLAIM; 25097039e1dbSPeter Zijlstra } 25104e096ae1SMatthew Wilcox (Oracle) return __folio_alloc_node(gfp, order, nid); 2511c5b5a3ddSYang Shi } 2512c5b5a3ddSYang Shi 25132ac9e99fSKefeng Wang static int numamigrate_isolate_folio(pg_data_t *pgdat, struct folio *folio) 2514b32967ffSMel Gorman { 25152ac9e99fSKefeng Wang int nr_pages = folio_nr_pages(folio); 2516b32967ffSMel Gorman 2517b32967ffSMel Gorman /* Avoid migrating to a node that is nearly full */ 2518c574bbe9SHuang Ying if (!migrate_balanced_pgdat(pgdat, nr_pages)) { 2519c574bbe9SHuang Ying int z; 2520c574bbe9SHuang Ying 2521c574bbe9SHuang Ying if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)) 2522340ef390SHugh Dickins return 0; 2523c574bbe9SHuang Ying for (z = pgdat->nr_zones - 1; z >= 0; z--) { 2524bc53008eSWei Yang if (managed_zone(pgdat->node_zones + z)) 2525c574bbe9SHuang Ying break; 2526c574bbe9SHuang Ying } 25272ac9e99fSKefeng Wang wakeup_kswapd(pgdat->node_zones + z, 0, 25282ac9e99fSKefeng Wang folio_order(folio), ZONE_MOVABLE); 2529c574bbe9SHuang Ying return 0; 2530c574bbe9SHuang Ying } 2531b32967ffSMel Gorman 25322ac9e99fSKefeng Wang if (!folio_isolate_lru(folio)) 2533340ef390SHugh Dickins return 0; 2534340ef390SHugh Dickins 25352ac9e99fSKefeng Wang node_stat_mod_folio(folio, NR_ISOLATED_ANON + folio_is_file_lru(folio), 25362b9b624fSBaolin Wang nr_pages); 2537b32967ffSMel Gorman 2538b32967ffSMel Gorman /* 25392ac9e99fSKefeng Wang * Isolating the folio has taken another reference, so the 25402ac9e99fSKefeng Wang * caller's reference can be safely dropped without the folio 2541340ef390SHugh Dickins * disappearing underneath us during migration. 2542b32967ffSMel Gorman */ 25432ac9e99fSKefeng Wang folio_put(folio); 2544340ef390SHugh Dickins return 1; 2545b32967ffSMel Gorman } 2546b32967ffSMel Gorman 2547a8f60772SMel Gorman /* 254873eab3caSKefeng Wang * Attempt to migrate a misplaced folio to the specified destination 25497039e1dbSPeter Zijlstra * node. Caller is expected to have an elevated reference count on 255073eab3caSKefeng Wang * the folio that will be dropped by this function before returning. 25517039e1dbSPeter Zijlstra */ 255273eab3caSKefeng Wang int migrate_misplaced_folio(struct folio *folio, struct vm_area_struct *vma, 25531bc115d8SMel Gorman int node) 25547039e1dbSPeter Zijlstra { 2555a8f60772SMel Gorman pg_data_t *pgdat = NODE_DATA(node); 2556340ef390SHugh Dickins int isolated; 2557b32967ffSMel Gorman int nr_remaining; 2558e39bb6beSHuang Ying unsigned int nr_succeeded; 25597039e1dbSPeter Zijlstra LIST_HEAD(migratepages); 256073eab3caSKefeng Wang int nr_pages = folio_nr_pages(folio); 2561c5b5a3ddSYang Shi 2562c5b5a3ddSYang Shi /* 256373eab3caSKefeng Wang * Don't migrate file folios that are mapped in multiple processes 25641bc115d8SMel Gorman * with execute permissions as they are probably shared libraries. 256573eab3caSKefeng Wang * To check if the folio is shared, ideally we want to make sure 256673eab3caSKefeng Wang * every page is mapped to the same process. Doing that is very 256773eab3caSKefeng Wang * expensive, so check the estimated mapcount of the folio instead. 25687039e1dbSPeter Zijlstra */ 256973eab3caSKefeng Wang if (folio_estimated_sharers(folio) != 1 && folio_is_file_lru(folio) && 25707ee820eeSMiaohe Lin (vma->vm_flags & VM_EXEC)) 25717039e1dbSPeter Zijlstra goto out; 25727039e1dbSPeter Zijlstra 2573a8f60772SMel Gorman /* 257473eab3caSKefeng Wang * Also do not migrate dirty folios as not all filesystems can move 257573eab3caSKefeng Wang * dirty folios in MIGRATE_ASYNC mode which is a waste of cycles. 257609a913a7SMel Gorman */ 257773eab3caSKefeng Wang if (folio_is_file_lru(folio) && folio_test_dirty(folio)) 257809a913a7SMel Gorman goto out; 257909a913a7SMel Gorman 258073eab3caSKefeng Wang isolated = numamigrate_isolate_folio(pgdat, folio); 2581b32967ffSMel Gorman if (!isolated) 25827039e1dbSPeter Zijlstra goto out; 25837039e1dbSPeter Zijlstra 258473eab3caSKefeng Wang list_add(&folio->lru, &migratepages); 25854e096ae1SMatthew Wilcox (Oracle) nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_folio, 2586c185e494SMatthew Wilcox (Oracle) NULL, node, MIGRATE_ASYNC, 2587c185e494SMatthew Wilcox (Oracle) MR_NUMA_MISPLACED, &nr_succeeded); 25887039e1dbSPeter Zijlstra if (nr_remaining) { 258959c82b70SJoonsoo Kim if (!list_empty(&migratepages)) { 259073eab3caSKefeng Wang list_del(&folio->lru); 259173eab3caSKefeng Wang node_stat_mod_folio(folio, NR_ISOLATED_ANON + 259273eab3caSKefeng Wang folio_is_file_lru(folio), -nr_pages); 259373eab3caSKefeng Wang folio_putback_lru(folio); 259459c82b70SJoonsoo Kim } 25957039e1dbSPeter Zijlstra isolated = 0; 2596e39bb6beSHuang Ying } 2597e39bb6beSHuang Ying if (nr_succeeded) { 2598e39bb6beSHuang Ying count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded); 259973eab3caSKefeng Wang if (!node_is_toptier(folio_nid(folio)) && node_is_toptier(node)) 2600e39bb6beSHuang Ying mod_node_page_state(pgdat, PGPROMOTE_SUCCESS, 2601e39bb6beSHuang Ying nr_succeeded); 2602e39bb6beSHuang Ying } 26037039e1dbSPeter Zijlstra BUG_ON(!list_empty(&migratepages)); 26047039e1dbSPeter Zijlstra return isolated; 2605340ef390SHugh Dickins 2606340ef390SHugh Dickins out: 260773eab3caSKefeng Wang folio_put(folio); 2608340ef390SHugh Dickins return 0; 26097039e1dbSPeter Zijlstra } 2610220018d3SMel Gorman #endif /* CONFIG_NUMA_BALANCING */ 26117d6e2d96SOscar Salvador #endif /* CONFIG_NUMA */ 2612