120c8ccb1SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 271e3aac0SAndrea Arcangeli /* 371e3aac0SAndrea Arcangeli * Copyright (C) 2009 Red Hat, Inc. 471e3aac0SAndrea Arcangeli */ 571e3aac0SAndrea Arcangeli 6ae3a8c1cSAndrew Morton #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 7ae3a8c1cSAndrew Morton 871e3aac0SAndrea Arcangeli #include <linux/mm.h> 971e3aac0SAndrea Arcangeli #include <linux/sched.h> 10fa6c0231SZi Yan #include <linux/sched/mm.h> 11f7ccbae4SIngo Molnar #include <linux/sched/coredump.h> 126a3827d7SIngo Molnar #include <linux/sched/numa_balancing.h> 1371e3aac0SAndrea Arcangeli #include <linux/highmem.h> 1471e3aac0SAndrea Arcangeli #include <linux/hugetlb.h> 1571e3aac0SAndrea Arcangeli #include <linux/mmu_notifier.h> 1671e3aac0SAndrea Arcangeli #include <linux/rmap.h> 1771e3aac0SAndrea Arcangeli #include <linux/swap.h> 1897ae1749SKirill A. Shutemov #include <linux/shrinker.h> 19ba76149fSAndrea Arcangeli #include <linux/mm_inline.h> 20e9b61f19SKirill A. Shutemov #include <linux/swapops.h> 21fb5c2029SMatthew Wilcox (Oracle) #include <linux/backing-dev.h> 224897c765SMatthew Wilcox #include <linux/dax.h> 23ba76149fSAndrea Arcangeli #include <linux/khugepaged.h> 24878aee7dSAndrea Arcangeli #include <linux/freezer.h> 25f25748e3SDan Williams #include <linux/pfn_t.h> 26a664b2d8SAndrea Arcangeli #include <linux/mman.h> 273565fce3SDan Williams #include <linux/memremap.h> 28325adeb5SRalf Baechle #include <linux/pagemap.h> 2949071d43SKirill A. Shutemov #include <linux/debugfs.h> 304daae3b4SMel Gorman #include <linux/migrate.h> 3143b5fbbdSSasha Levin #include <linux/hashtable.h> 326b251fc9SAndrea Arcangeli #include <linux/userfaultfd_k.h> 3333c3fc71SVladimir Davydov #include <linux/page_idle.h> 34baa355fdSKirill A. Shutemov #include <linux/shmem_fs.h> 356b31d595SMichal Hocko #include <linux/oom.h> 3698fa15f3SAnshuman Khandual #include <linux/numa.h> 37f7da677bSVlastimil Babka #include <linux/page_owner.h> 38a1a3a2fcSHuang Ying #include <linux/sched/sysctl.h> 39467b171aSAneesh Kumar K.V #include <linux/memory-tiers.h> 404ef9ad19SYang Shi #include <linux/compat.h> 4197ae1749SKirill A. Shutemov 4271e3aac0SAndrea Arcangeli #include <asm/tlb.h> 4371e3aac0SAndrea Arcangeli #include <asm/pgalloc.h> 4471e3aac0SAndrea Arcangeli #include "internal.h" 45014bb1deSNeilBrown #include "swap.h" 4671e3aac0SAndrea Arcangeli 47283fd6feSAnshuman Khandual #define CREATE_TRACE_POINTS 48283fd6feSAnshuman Khandual #include <trace/events/thp.h> 49283fd6feSAnshuman Khandual 50ba76149fSAndrea Arcangeli /* 51b14d595aSMichael DeGuzis * By default, transparent hugepage support is disabled in order to avoid 52b14d595aSMichael DeGuzis * risking an increased memory footprint for applications that are not 53b14d595aSMichael DeGuzis * guaranteed to benefit from it. When transparent hugepage support is 54b14d595aSMichael DeGuzis * enabled, it is for all mappings, and khugepaged scans all mappings. 558bfa3f9aSJianguo Wu * Defrag is invoked by khugepaged hugepage allocations and by page faults 568bfa3f9aSJianguo Wu * for all hugepage allocations. 57ba76149fSAndrea Arcangeli */ 5871e3aac0SAndrea Arcangeli unsigned long transparent_hugepage_flags __read_mostly = 5913ece886SAndrea Arcangeli #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS 60ba76149fSAndrea Arcangeli (1<<TRANSPARENT_HUGEPAGE_FLAG)| 6113ece886SAndrea Arcangeli #endif 6213ece886SAndrea Arcangeli #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE 6313ece886SAndrea Arcangeli (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| 6413ece886SAndrea Arcangeli #endif 65444eb2a4SMel Gorman (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)| 6679da5407SKirill A. Shutemov (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| 6779da5407SKirill A. Shutemov (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 68ba76149fSAndrea Arcangeli 6954d91729SQi Zheng static struct shrinker *deferred_split_shrinker; 7054d91729SQi Zheng static unsigned long deferred_split_count(struct shrinker *shrink, 7154d91729SQi Zheng struct shrink_control *sc); 7254d91729SQi Zheng static unsigned long deferred_split_scan(struct shrinker *shrink, 7354d91729SQi Zheng struct shrink_control *sc); 74f000565aSAndrea Arcangeli 7597ae1749SKirill A. Shutemov static atomic_t huge_zero_refcount; 7656873f43SWang, Yalin struct page *huge_zero_page __read_mostly; 773b77e8c8SHugh Dickins unsigned long huge_zero_pfn __read_mostly = ~0UL; 783485b883SRyan Roberts unsigned long huge_anon_orders_always __read_mostly; 793485b883SRyan Roberts unsigned long huge_anon_orders_madvise __read_mostly; 803485b883SRyan Roberts unsigned long huge_anon_orders_inherit __read_mostly; 814a6c1297SKirill A. Shutemov 823485b883SRyan Roberts unsigned long __thp_vma_allowable_orders(struct vm_area_struct *vma, 833485b883SRyan Roberts unsigned long vm_flags, bool smaps, 843485b883SRyan Roberts bool in_pf, bool enforce_sysfs, 853485b883SRyan Roberts unsigned long orders) 867635d9cbSMichal Hocko { 873485b883SRyan Roberts /* Check the intersection of requested and supported orders. */ 883485b883SRyan Roberts orders &= vma_is_anonymous(vma) ? 893485b883SRyan Roberts THP_ORDERS_ALL_ANON : THP_ORDERS_ALL_FILE; 903485b883SRyan Roberts if (!orders) 913485b883SRyan Roberts return 0; 923485b883SRyan Roberts 939fec5168SYang Shi if (!vma->vm_mm) /* vdso */ 943485b883SRyan Roberts return 0; 959fec5168SYang Shi 967da4e2cbSYang Shi /* 977da4e2cbSYang Shi * Explicitly disabled through madvise or prctl, or some 987da4e2cbSYang Shi * architectures may disable THP for some mappings, for 997da4e2cbSYang Shi * example, s390 kvm. 1007da4e2cbSYang Shi * */ 1017da4e2cbSYang Shi if ((vm_flags & VM_NOHUGEPAGE) || 1027da4e2cbSYang Shi test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags)) 1033485b883SRyan Roberts return 0; 1047da4e2cbSYang Shi /* 1057da4e2cbSYang Shi * If the hardware/firmware marked hugepage support disabled. 1067da4e2cbSYang Shi */ 1073c556d24SPeter Xu if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED)) 1083485b883SRyan Roberts return 0; 1099fec5168SYang Shi 1107da4e2cbSYang Shi /* khugepaged doesn't collapse DAX vma, but page fault is fine. */ 1117da4e2cbSYang Shi if (vma_is_dax(vma)) 1123485b883SRyan Roberts return in_pf ? orders : 0; 1137da4e2cbSYang Shi 1147da4e2cbSYang Shi /* 1157a81751fSZach O'Keefe * khugepaged special VMA and hugetlb VMA. 1167da4e2cbSYang Shi * Must be checked after dax since some dax mappings may have 1177da4e2cbSYang Shi * VM_MIXEDMAP set. 1187da4e2cbSYang Shi */ 1197a81751fSZach O'Keefe if (!in_pf && !smaps && (vm_flags & VM_NO_KHUGEPAGED)) 1203485b883SRyan Roberts return 0; 1219fec5168SYang Shi 1227da4e2cbSYang Shi /* 1233485b883SRyan Roberts * Check alignment for file vma and size for both file and anon vma by 1243485b883SRyan Roberts * filtering out the unsuitable orders. 1257da4e2cbSYang Shi * 1267da4e2cbSYang Shi * Skip the check for page fault. Huge fault does the check in fault 1273485b883SRyan Roberts * handlers. 1287da4e2cbSYang Shi */ 1293485b883SRyan Roberts if (!in_pf) { 1303485b883SRyan Roberts int order = highest_order(orders); 1313485b883SRyan Roberts unsigned long addr; 1323485b883SRyan Roberts 1333485b883SRyan Roberts while (orders) { 1343485b883SRyan Roberts addr = vma->vm_end - (PAGE_SIZE << order); 1353485b883SRyan Roberts if (thp_vma_suitable_order(vma, addr, order)) 1363485b883SRyan Roberts break; 1373485b883SRyan Roberts order = next_order(&orders, order); 1383485b883SRyan Roberts } 1393485b883SRyan Roberts 1403485b883SRyan Roberts if (!orders) 1413485b883SRyan Roberts return 0; 1423485b883SRyan Roberts } 1439fec5168SYang Shi 1447da4e2cbSYang Shi /* 1457da4e2cbSYang Shi * Enabled via shmem mount options or sysfs settings. 1467da4e2cbSYang Shi * Must be done before hugepage flags check since shmem has its 1477da4e2cbSYang Shi * own flags. 1487da4e2cbSYang Shi */ 1497da4e2cbSYang Shi if (!in_pf && shmem_file(vma->vm_file)) 1502cf13384SDavid Stevens return shmem_is_huge(file_inode(vma->vm_file), vma->vm_pgoff, 1513485b883SRyan Roberts !enforce_sysfs, vma->vm_mm, vm_flags) 1523485b883SRyan Roberts ? orders : 0; 1539fec5168SYang Shi 1547a81751fSZach O'Keefe if (!vma_is_anonymous(vma)) { 1557a81751fSZach O'Keefe /* 1563485b883SRyan Roberts * Enforce sysfs THP requirements as necessary. Anonymous vmas 1573485b883SRyan Roberts * were already handled in thp_vma_allowable_orders(). 1583485b883SRyan Roberts */ 1593485b883SRyan Roberts if (enforce_sysfs && 1603485b883SRyan Roberts (!hugepage_global_enabled() || (!(vm_flags & VM_HUGEPAGE) && 1613485b883SRyan Roberts !hugepage_global_always()))) 1623485b883SRyan Roberts return 0; 1633485b883SRyan Roberts 1643485b883SRyan Roberts /* 1657a81751fSZach O'Keefe * Trust that ->huge_fault() handlers know what they are doing 1667a81751fSZach O'Keefe * in fault path. 1677a81751fSZach O'Keefe */ 1687a81751fSZach O'Keefe if (((in_pf || smaps)) && vma->vm_ops->huge_fault) 1693485b883SRyan Roberts return orders; 1707a81751fSZach O'Keefe /* Only regular file is valid in collapse path */ 1717a81751fSZach O'Keefe if (((!in_pf || smaps)) && file_thp_enabled(vma)) 1723485b883SRyan Roberts return orders; 1733485b883SRyan Roberts return 0; 1747a81751fSZach O'Keefe } 1759fec5168SYang Shi 1769fec5168SYang Shi if (vma_is_temporary_stack(vma)) 1773485b883SRyan Roberts return 0; 1789fec5168SYang Shi 1799fec5168SYang Shi /* 1809fec5168SYang Shi * THPeligible bit of smaps should show 1 for proper VMAs even 1819fec5168SYang Shi * though anon_vma is not initialized yet. 1827da4e2cbSYang Shi * 1837da4e2cbSYang Shi * Allow page fault since anon_vma may be not initialized until 1847da4e2cbSYang Shi * the first page fault. 1859fec5168SYang Shi */ 1869fec5168SYang Shi if (!vma->anon_vma) 1873485b883SRyan Roberts return (smaps || in_pf) ? orders : 0; 1889fec5168SYang Shi 1893485b883SRyan Roberts return orders; 1907635d9cbSMichal Hocko } 1917635d9cbSMichal Hocko 192aaa9705bSMiaohe Lin static bool get_huge_zero_page(void) 19397ae1749SKirill A. Shutemov { 19497ae1749SKirill A. Shutemov struct page *zero_page; 19597ae1749SKirill A. Shutemov retry: 19697ae1749SKirill A. Shutemov if (likely(atomic_inc_not_zero(&huge_zero_refcount))) 197aaa9705bSMiaohe Lin return true; 19897ae1749SKirill A. Shutemov 19997ae1749SKirill A. Shutemov zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, 20097ae1749SKirill A. Shutemov HPAGE_PMD_ORDER); 201d8a8e1f0SKirill A. Shutemov if (!zero_page) { 202d8a8e1f0SKirill A. Shutemov count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); 203aaa9705bSMiaohe Lin return false; 204d8a8e1f0SKirill A. Shutemov } 20597ae1749SKirill A. Shutemov preempt_disable(); 2065918d10aSKirill A. Shutemov if (cmpxchg(&huge_zero_page, NULL, zero_page)) { 20797ae1749SKirill A. Shutemov preempt_enable(); 2085ddacbe9SYu Zhao __free_pages(zero_page, compound_order(zero_page)); 20997ae1749SKirill A. Shutemov goto retry; 21097ae1749SKirill A. Shutemov } 2113b77e8c8SHugh Dickins WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page)); 21297ae1749SKirill A. Shutemov 21397ae1749SKirill A. Shutemov /* We take additional reference here. It will be put back by shrinker */ 21497ae1749SKirill A. Shutemov atomic_set(&huge_zero_refcount, 2); 21597ae1749SKirill A. Shutemov preempt_enable(); 216f4981502SLiu Shixin count_vm_event(THP_ZERO_PAGE_ALLOC); 217aaa9705bSMiaohe Lin return true; 21897ae1749SKirill A. Shutemov } 21997ae1749SKirill A. Shutemov 2206fcb52a5SAaron Lu static void put_huge_zero_page(void) 22197ae1749SKirill A. Shutemov { 22297ae1749SKirill A. Shutemov /* 22397ae1749SKirill A. Shutemov * Counter should never go to zero here. Only shrinker can put 22497ae1749SKirill A. Shutemov * last reference. 22597ae1749SKirill A. Shutemov */ 22697ae1749SKirill A. Shutemov BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); 22797ae1749SKirill A. Shutemov } 22897ae1749SKirill A. Shutemov 2296fcb52a5SAaron Lu struct page *mm_get_huge_zero_page(struct mm_struct *mm) 2306fcb52a5SAaron Lu { 2316fcb52a5SAaron Lu if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 2326fcb52a5SAaron Lu return READ_ONCE(huge_zero_page); 2336fcb52a5SAaron Lu 2346fcb52a5SAaron Lu if (!get_huge_zero_page()) 2356fcb52a5SAaron Lu return NULL; 2366fcb52a5SAaron Lu 2376fcb52a5SAaron Lu if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 2386fcb52a5SAaron Lu put_huge_zero_page(); 2396fcb52a5SAaron Lu 2406fcb52a5SAaron Lu return READ_ONCE(huge_zero_page); 2416fcb52a5SAaron Lu } 2426fcb52a5SAaron Lu 2436fcb52a5SAaron Lu void mm_put_huge_zero_page(struct mm_struct *mm) 2446fcb52a5SAaron Lu { 2456fcb52a5SAaron Lu if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) 2466fcb52a5SAaron Lu put_huge_zero_page(); 2476fcb52a5SAaron Lu } 2486fcb52a5SAaron Lu 24948896466SGlauber Costa static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, 25097ae1749SKirill A. Shutemov struct shrink_control *sc) 25197ae1749SKirill A. Shutemov { 25297ae1749SKirill A. Shutemov /* we can free zero page only if last reference remains */ 25397ae1749SKirill A. Shutemov return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; 25448896466SGlauber Costa } 25597ae1749SKirill A. Shutemov 25648896466SGlauber Costa static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, 25748896466SGlauber Costa struct shrink_control *sc) 25848896466SGlauber Costa { 25997ae1749SKirill A. Shutemov if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { 2605918d10aSKirill A. Shutemov struct page *zero_page = xchg(&huge_zero_page, NULL); 2615918d10aSKirill A. Shutemov BUG_ON(zero_page == NULL); 2623b77e8c8SHugh Dickins WRITE_ONCE(huge_zero_pfn, ~0UL); 2635ddacbe9SYu Zhao __free_pages(zero_page, compound_order(zero_page)); 26448896466SGlauber Costa return HPAGE_PMD_NR; 26597ae1749SKirill A. Shutemov } 26697ae1749SKirill A. Shutemov 26797ae1749SKirill A. Shutemov return 0; 26897ae1749SKirill A. Shutemov } 26997ae1749SKirill A. Shutemov 27054d91729SQi Zheng static struct shrinker *huge_zero_page_shrinker; 27197ae1749SKirill A. Shutemov 27271e3aac0SAndrea Arcangeli #ifdef CONFIG_SYSFS 27371e3aac0SAndrea Arcangeli static ssize_t enabled_show(struct kobject *kobj, 27471e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 27571e3aac0SAndrea Arcangeli { 276bfb0ffebSJoe Perches const char *output; 277bfb0ffebSJoe Perches 278444eb2a4SMel Gorman if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) 279bfb0ffebSJoe Perches output = "[always] madvise never"; 280bfb0ffebSJoe Perches else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 281bfb0ffebSJoe Perches &transparent_hugepage_flags)) 282bfb0ffebSJoe Perches output = "always [madvise] never"; 283444eb2a4SMel Gorman else 284bfb0ffebSJoe Perches output = "always madvise [never]"; 285bfb0ffebSJoe Perches 286bfb0ffebSJoe Perches return sysfs_emit(buf, "%s\n", output); 28771e3aac0SAndrea Arcangeli } 288444eb2a4SMel Gorman 28971e3aac0SAndrea Arcangeli static ssize_t enabled_store(struct kobject *kobj, 29071e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 29171e3aac0SAndrea Arcangeli const char *buf, size_t count) 29271e3aac0SAndrea Arcangeli { 29321440d7eSDavid Rientjes ssize_t ret = count; 294ba76149fSAndrea Arcangeli 295f42f2552SDavid Rientjes if (sysfs_streq(buf, "always")) { 29621440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 29721440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 298f42f2552SDavid Rientjes } else if (sysfs_streq(buf, "madvise")) { 29921440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 30021440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 301f42f2552SDavid Rientjes } else if (sysfs_streq(buf, "never")) { 30221440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); 30321440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); 30421440d7eSDavid Rientjes } else 30521440d7eSDavid Rientjes ret = -EINVAL; 306ba76149fSAndrea Arcangeli 307ba76149fSAndrea Arcangeli if (ret > 0) { 308b46e756fSKirill A. Shutemov int err = start_stop_khugepaged(); 309ba76149fSAndrea Arcangeli if (err) 310ba76149fSAndrea Arcangeli ret = err; 311ba76149fSAndrea Arcangeli } 312ba76149fSAndrea Arcangeli return ret; 31371e3aac0SAndrea Arcangeli } 31437139bb0SMiaohe Lin 31537139bb0SMiaohe Lin static struct kobj_attribute enabled_attr = __ATTR_RW(enabled); 31671e3aac0SAndrea Arcangeli 317b46e756fSKirill A. Shutemov ssize_t single_hugepage_flag_show(struct kobject *kobj, 31871e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf, 31971e3aac0SAndrea Arcangeli enum transparent_hugepage_flag flag) 32071e3aac0SAndrea Arcangeli { 321bfb0ffebSJoe Perches return sysfs_emit(buf, "%d\n", 322e27e6151SBen Hutchings !!test_bit(flag, &transparent_hugepage_flags)); 32371e3aac0SAndrea Arcangeli } 324e27e6151SBen Hutchings 325b46e756fSKirill A. Shutemov ssize_t single_hugepage_flag_store(struct kobject *kobj, 32671e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 32771e3aac0SAndrea Arcangeli const char *buf, size_t count, 32871e3aac0SAndrea Arcangeli enum transparent_hugepage_flag flag) 32971e3aac0SAndrea Arcangeli { 330e27e6151SBen Hutchings unsigned long value; 331e27e6151SBen Hutchings int ret; 332e27e6151SBen Hutchings 333e27e6151SBen Hutchings ret = kstrtoul(buf, 10, &value); 334e27e6151SBen Hutchings if (ret < 0) 335e27e6151SBen Hutchings return ret; 336e27e6151SBen Hutchings if (value > 1) 33771e3aac0SAndrea Arcangeli return -EINVAL; 33871e3aac0SAndrea Arcangeli 339e27e6151SBen Hutchings if (value) 340e27e6151SBen Hutchings set_bit(flag, &transparent_hugepage_flags); 341e27e6151SBen Hutchings else 342e27e6151SBen Hutchings clear_bit(flag, &transparent_hugepage_flags); 343e27e6151SBen Hutchings 34471e3aac0SAndrea Arcangeli return count; 34571e3aac0SAndrea Arcangeli } 34671e3aac0SAndrea Arcangeli 34771e3aac0SAndrea Arcangeli static ssize_t defrag_show(struct kobject *kobj, 34871e3aac0SAndrea Arcangeli struct kobj_attribute *attr, char *buf) 34971e3aac0SAndrea Arcangeli { 350bfb0ffebSJoe Perches const char *output; 351bfb0ffebSJoe Perches 352bfb0ffebSJoe Perches if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, 353bfb0ffebSJoe Perches &transparent_hugepage_flags)) 354bfb0ffebSJoe Perches output = "[always] defer defer+madvise madvise never"; 355bfb0ffebSJoe Perches else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, 356bfb0ffebSJoe Perches &transparent_hugepage_flags)) 357bfb0ffebSJoe Perches output = "always [defer] defer+madvise madvise never"; 358bfb0ffebSJoe Perches else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, 359bfb0ffebSJoe Perches &transparent_hugepage_flags)) 360bfb0ffebSJoe Perches output = "always defer [defer+madvise] madvise never"; 361bfb0ffebSJoe Perches else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, 362bfb0ffebSJoe Perches &transparent_hugepage_flags)) 363bfb0ffebSJoe Perches output = "always defer defer+madvise [madvise] never"; 364bfb0ffebSJoe Perches else 365bfb0ffebSJoe Perches output = "always defer defer+madvise madvise [never]"; 366bfb0ffebSJoe Perches 367bfb0ffebSJoe Perches return sysfs_emit(buf, "%s\n", output); 36871e3aac0SAndrea Arcangeli } 36921440d7eSDavid Rientjes 37071e3aac0SAndrea Arcangeli static ssize_t defrag_store(struct kobject *kobj, 37171e3aac0SAndrea Arcangeli struct kobj_attribute *attr, 37271e3aac0SAndrea Arcangeli const char *buf, size_t count) 37371e3aac0SAndrea Arcangeli { 374f42f2552SDavid Rientjes if (sysfs_streq(buf, "always")) { 37521440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 37621440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 37721440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 37821440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 379f42f2552SDavid Rientjes } else if (sysfs_streq(buf, "defer+madvise")) { 38021440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 38121440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 38221440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 38321440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 384f42f2552SDavid Rientjes } else if (sysfs_streq(buf, "defer")) { 3854fad7fb6SDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 3864fad7fb6SDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 3874fad7fb6SDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 3884fad7fb6SDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 389f42f2552SDavid Rientjes } else if (sysfs_streq(buf, "madvise")) { 39021440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 39121440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 39221440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 39321440d7eSDavid Rientjes set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 394f42f2552SDavid Rientjes } else if (sysfs_streq(buf, "never")) { 39521440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); 39621440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); 39721440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); 39821440d7eSDavid Rientjes clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); 39921440d7eSDavid Rientjes } else 40021440d7eSDavid Rientjes return -EINVAL; 40121440d7eSDavid Rientjes 40221440d7eSDavid Rientjes return count; 40371e3aac0SAndrea Arcangeli } 40437139bb0SMiaohe Lin static struct kobj_attribute defrag_attr = __ATTR_RW(defrag); 40571e3aac0SAndrea Arcangeli 40679da5407SKirill A. Shutemov static ssize_t use_zero_page_show(struct kobject *kobj, 40779da5407SKirill A. Shutemov struct kobj_attribute *attr, char *buf) 40879da5407SKirill A. Shutemov { 409b46e756fSKirill A. Shutemov return single_hugepage_flag_show(kobj, attr, buf, 41079da5407SKirill A. Shutemov TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 41179da5407SKirill A. Shutemov } 41279da5407SKirill A. Shutemov static ssize_t use_zero_page_store(struct kobject *kobj, 41379da5407SKirill A. Shutemov struct kobj_attribute *attr, const char *buf, size_t count) 41479da5407SKirill A. Shutemov { 415b46e756fSKirill A. Shutemov return single_hugepage_flag_store(kobj, attr, buf, count, 41679da5407SKirill A. Shutemov TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 41779da5407SKirill A. Shutemov } 41837139bb0SMiaohe Lin static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page); 41949920d28SHugh Dickins 42049920d28SHugh Dickins static ssize_t hpage_pmd_size_show(struct kobject *kobj, 42149920d28SHugh Dickins struct kobj_attribute *attr, char *buf) 42249920d28SHugh Dickins { 423ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE); 42449920d28SHugh Dickins } 42549920d28SHugh Dickins static struct kobj_attribute hpage_pmd_size_attr = 42649920d28SHugh Dickins __ATTR_RO(hpage_pmd_size); 42749920d28SHugh Dickins 42871e3aac0SAndrea Arcangeli static struct attribute *hugepage_attr[] = { 42971e3aac0SAndrea Arcangeli &enabled_attr.attr, 43071e3aac0SAndrea Arcangeli &defrag_attr.attr, 43179da5407SKirill A. Shutemov &use_zero_page_attr.attr, 43249920d28SHugh Dickins &hpage_pmd_size_attr.attr, 433396bcc52SMatthew Wilcox (Oracle) #ifdef CONFIG_SHMEM 4345a6e75f8SKirill A. Shutemov &shmem_enabled_attr.attr, 4355a6e75f8SKirill A. Shutemov #endif 43671e3aac0SAndrea Arcangeli NULL, 43771e3aac0SAndrea Arcangeli }; 43871e3aac0SAndrea Arcangeli 4398aa95a21SArvind Yadav static const struct attribute_group hugepage_attr_group = { 44071e3aac0SAndrea Arcangeli .attrs = hugepage_attr, 441ba76149fSAndrea Arcangeli }; 442ba76149fSAndrea Arcangeli 4433485b883SRyan Roberts static void hugepage_exit_sysfs(struct kobject *hugepage_kobj); 4443485b883SRyan Roberts static void thpsize_release(struct kobject *kobj); 4453485b883SRyan Roberts static DEFINE_SPINLOCK(huge_anon_orders_lock); 4463485b883SRyan Roberts static LIST_HEAD(thpsize_list); 4473485b883SRyan Roberts 4483485b883SRyan Roberts struct thpsize { 4493485b883SRyan Roberts struct kobject kobj; 4503485b883SRyan Roberts struct list_head node; 4513485b883SRyan Roberts int order; 4523485b883SRyan Roberts }; 4533485b883SRyan Roberts 4543485b883SRyan Roberts #define to_thpsize(kobj) container_of(kobj, struct thpsize, kobj) 4553485b883SRyan Roberts 4563485b883SRyan Roberts static ssize_t thpsize_enabled_show(struct kobject *kobj, 4573485b883SRyan Roberts struct kobj_attribute *attr, char *buf) 4583485b883SRyan Roberts { 4593485b883SRyan Roberts int order = to_thpsize(kobj)->order; 4603485b883SRyan Roberts const char *output; 4613485b883SRyan Roberts 4623485b883SRyan Roberts if (test_bit(order, &huge_anon_orders_always)) 4633485b883SRyan Roberts output = "[always] inherit madvise never"; 4643485b883SRyan Roberts else if (test_bit(order, &huge_anon_orders_inherit)) 4653485b883SRyan Roberts output = "always [inherit] madvise never"; 4663485b883SRyan Roberts else if (test_bit(order, &huge_anon_orders_madvise)) 4673485b883SRyan Roberts output = "always inherit [madvise] never"; 4683485b883SRyan Roberts else 4693485b883SRyan Roberts output = "always inherit madvise [never]"; 4703485b883SRyan Roberts 4713485b883SRyan Roberts return sysfs_emit(buf, "%s\n", output); 4723485b883SRyan Roberts } 4733485b883SRyan Roberts 4743485b883SRyan Roberts static ssize_t thpsize_enabled_store(struct kobject *kobj, 4753485b883SRyan Roberts struct kobj_attribute *attr, 4763485b883SRyan Roberts const char *buf, size_t count) 4773485b883SRyan Roberts { 4783485b883SRyan Roberts int order = to_thpsize(kobj)->order; 4793485b883SRyan Roberts ssize_t ret = count; 4803485b883SRyan Roberts 4813485b883SRyan Roberts if (sysfs_streq(buf, "always")) { 4823485b883SRyan Roberts spin_lock(&huge_anon_orders_lock); 4833485b883SRyan Roberts clear_bit(order, &huge_anon_orders_inherit); 4843485b883SRyan Roberts clear_bit(order, &huge_anon_orders_madvise); 4853485b883SRyan Roberts set_bit(order, &huge_anon_orders_always); 4863485b883SRyan Roberts spin_unlock(&huge_anon_orders_lock); 4873485b883SRyan Roberts } else if (sysfs_streq(buf, "inherit")) { 4883485b883SRyan Roberts spin_lock(&huge_anon_orders_lock); 4893485b883SRyan Roberts clear_bit(order, &huge_anon_orders_always); 4903485b883SRyan Roberts clear_bit(order, &huge_anon_orders_madvise); 4913485b883SRyan Roberts set_bit(order, &huge_anon_orders_inherit); 4923485b883SRyan Roberts spin_unlock(&huge_anon_orders_lock); 4933485b883SRyan Roberts } else if (sysfs_streq(buf, "madvise")) { 4943485b883SRyan Roberts spin_lock(&huge_anon_orders_lock); 4953485b883SRyan Roberts clear_bit(order, &huge_anon_orders_always); 4963485b883SRyan Roberts clear_bit(order, &huge_anon_orders_inherit); 4973485b883SRyan Roberts set_bit(order, &huge_anon_orders_madvise); 4983485b883SRyan Roberts spin_unlock(&huge_anon_orders_lock); 4993485b883SRyan Roberts } else if (sysfs_streq(buf, "never")) { 5003485b883SRyan Roberts spin_lock(&huge_anon_orders_lock); 5013485b883SRyan Roberts clear_bit(order, &huge_anon_orders_always); 5023485b883SRyan Roberts clear_bit(order, &huge_anon_orders_inherit); 5033485b883SRyan Roberts clear_bit(order, &huge_anon_orders_madvise); 5043485b883SRyan Roberts spin_unlock(&huge_anon_orders_lock); 5053485b883SRyan Roberts } else 5063485b883SRyan Roberts ret = -EINVAL; 5073485b883SRyan Roberts 5083485b883SRyan Roberts return ret; 5093485b883SRyan Roberts } 5103485b883SRyan Roberts 5113485b883SRyan Roberts static struct kobj_attribute thpsize_enabled_attr = 5123485b883SRyan Roberts __ATTR(enabled, 0644, thpsize_enabled_show, thpsize_enabled_store); 5133485b883SRyan Roberts 5143485b883SRyan Roberts static struct attribute *thpsize_attrs[] = { 5153485b883SRyan Roberts &thpsize_enabled_attr.attr, 5163485b883SRyan Roberts NULL, 5173485b883SRyan Roberts }; 5183485b883SRyan Roberts 5193485b883SRyan Roberts static const struct attribute_group thpsize_attr_group = { 5203485b883SRyan Roberts .attrs = thpsize_attrs, 5213485b883SRyan Roberts }; 5223485b883SRyan Roberts 5233485b883SRyan Roberts static const struct kobj_type thpsize_ktype = { 5243485b883SRyan Roberts .release = &thpsize_release, 5253485b883SRyan Roberts .sysfs_ops = &kobj_sysfs_ops, 5263485b883SRyan Roberts }; 5273485b883SRyan Roberts 5283485b883SRyan Roberts static struct thpsize *thpsize_create(int order, struct kobject *parent) 5293485b883SRyan Roberts { 5303485b883SRyan Roberts unsigned long size = (PAGE_SIZE << order) / SZ_1K; 5313485b883SRyan Roberts struct thpsize *thpsize; 5323485b883SRyan Roberts int ret; 5333485b883SRyan Roberts 5343485b883SRyan Roberts thpsize = kzalloc(sizeof(*thpsize), GFP_KERNEL); 5353485b883SRyan Roberts if (!thpsize) 5363485b883SRyan Roberts return ERR_PTR(-ENOMEM); 5373485b883SRyan Roberts 5383485b883SRyan Roberts ret = kobject_init_and_add(&thpsize->kobj, &thpsize_ktype, parent, 5393485b883SRyan Roberts "hugepages-%lukB", size); 5403485b883SRyan Roberts if (ret) { 5413485b883SRyan Roberts kfree(thpsize); 5423485b883SRyan Roberts return ERR_PTR(ret); 5433485b883SRyan Roberts } 5443485b883SRyan Roberts 5453485b883SRyan Roberts ret = sysfs_create_group(&thpsize->kobj, &thpsize_attr_group); 5463485b883SRyan Roberts if (ret) { 5473485b883SRyan Roberts kobject_put(&thpsize->kobj); 5483485b883SRyan Roberts return ERR_PTR(ret); 5493485b883SRyan Roberts } 5503485b883SRyan Roberts 5513485b883SRyan Roberts thpsize->order = order; 5523485b883SRyan Roberts return thpsize; 5533485b883SRyan Roberts } 5543485b883SRyan Roberts 5553485b883SRyan Roberts static void thpsize_release(struct kobject *kobj) 5563485b883SRyan Roberts { 5573485b883SRyan Roberts kfree(to_thpsize(kobj)); 5583485b883SRyan Roberts } 5593485b883SRyan Roberts 560569e5590SShaohua Li static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 561569e5590SShaohua Li { 562569e5590SShaohua Li int err; 5633485b883SRyan Roberts struct thpsize *thpsize; 5643485b883SRyan Roberts unsigned long orders; 5653485b883SRyan Roberts int order; 5663485b883SRyan Roberts 5673485b883SRyan Roberts /* 5683485b883SRyan Roberts * Default to setting PMD-sized THP to inherit the global setting and 5693485b883SRyan Roberts * disable all other sizes. powerpc's PMD_ORDER isn't a compile-time 5703485b883SRyan Roberts * constant so we have to do this here. 5713485b883SRyan Roberts */ 5723485b883SRyan Roberts huge_anon_orders_inherit = BIT(PMD_ORDER); 573569e5590SShaohua Li 574569e5590SShaohua Li *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); 575569e5590SShaohua Li if (unlikely(!*hugepage_kobj)) { 576ae3a8c1cSAndrew Morton pr_err("failed to create transparent hugepage kobject\n"); 577569e5590SShaohua Li return -ENOMEM; 578569e5590SShaohua Li } 579569e5590SShaohua Li 580569e5590SShaohua Li err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); 581569e5590SShaohua Li if (err) { 582ae3a8c1cSAndrew Morton pr_err("failed to register transparent hugepage group\n"); 583569e5590SShaohua Li goto delete_obj; 584569e5590SShaohua Li } 585569e5590SShaohua Li 586569e5590SShaohua Li err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); 587569e5590SShaohua Li if (err) { 588ae3a8c1cSAndrew Morton pr_err("failed to register transparent hugepage group\n"); 589569e5590SShaohua Li goto remove_hp_group; 590569e5590SShaohua Li } 591569e5590SShaohua Li 5923485b883SRyan Roberts orders = THP_ORDERS_ALL_ANON; 5933485b883SRyan Roberts order = highest_order(orders); 5943485b883SRyan Roberts while (orders) { 5953485b883SRyan Roberts thpsize = thpsize_create(order, *hugepage_kobj); 5963485b883SRyan Roberts if (IS_ERR(thpsize)) { 5973485b883SRyan Roberts pr_err("failed to create thpsize for order %d\n", order); 5983485b883SRyan Roberts err = PTR_ERR(thpsize); 5993485b883SRyan Roberts goto remove_all; 6003485b883SRyan Roberts } 6013485b883SRyan Roberts list_add(&thpsize->node, &thpsize_list); 6023485b883SRyan Roberts order = next_order(&orders, order); 6033485b883SRyan Roberts } 6043485b883SRyan Roberts 605569e5590SShaohua Li return 0; 606569e5590SShaohua Li 6073485b883SRyan Roberts remove_all: 6083485b883SRyan Roberts hugepage_exit_sysfs(*hugepage_kobj); 6093485b883SRyan Roberts return err; 610569e5590SShaohua Li remove_hp_group: 611569e5590SShaohua Li sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); 612569e5590SShaohua Li delete_obj: 613569e5590SShaohua Li kobject_put(*hugepage_kobj); 614569e5590SShaohua Li return err; 615569e5590SShaohua Li } 616569e5590SShaohua Li 617569e5590SShaohua Li static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) 618569e5590SShaohua Li { 6193485b883SRyan Roberts struct thpsize *thpsize, *tmp; 6203485b883SRyan Roberts 6213485b883SRyan Roberts list_for_each_entry_safe(thpsize, tmp, &thpsize_list, node) { 6223485b883SRyan Roberts list_del(&thpsize->node); 6233485b883SRyan Roberts kobject_put(&thpsize->kobj); 6243485b883SRyan Roberts } 6253485b883SRyan Roberts 626569e5590SShaohua Li sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); 627569e5590SShaohua Li sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); 628569e5590SShaohua Li kobject_put(hugepage_kobj); 629569e5590SShaohua Li } 630569e5590SShaohua Li #else 631569e5590SShaohua Li static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) 632569e5590SShaohua Li { 633569e5590SShaohua Li return 0; 634569e5590SShaohua Li } 635569e5590SShaohua Li 636569e5590SShaohua Li static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) 637569e5590SShaohua Li { 638569e5590SShaohua Li } 63971e3aac0SAndrea Arcangeli #endif /* CONFIG_SYSFS */ 64071e3aac0SAndrea Arcangeli 64154d91729SQi Zheng static int __init thp_shrinker_init(void) 64254d91729SQi Zheng { 64354d91729SQi Zheng huge_zero_page_shrinker = shrinker_alloc(0, "thp-zero"); 64454d91729SQi Zheng if (!huge_zero_page_shrinker) 64554d91729SQi Zheng return -ENOMEM; 64654d91729SQi Zheng 64754d91729SQi Zheng deferred_split_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE | 64854d91729SQi Zheng SHRINKER_MEMCG_AWARE | 64954d91729SQi Zheng SHRINKER_NONSLAB, 65054d91729SQi Zheng "thp-deferred_split"); 65154d91729SQi Zheng if (!deferred_split_shrinker) { 65254d91729SQi Zheng shrinker_free(huge_zero_page_shrinker); 65354d91729SQi Zheng return -ENOMEM; 65454d91729SQi Zheng } 65554d91729SQi Zheng 65654d91729SQi Zheng huge_zero_page_shrinker->count_objects = shrink_huge_zero_page_count; 65754d91729SQi Zheng huge_zero_page_shrinker->scan_objects = shrink_huge_zero_page_scan; 65854d91729SQi Zheng shrinker_register(huge_zero_page_shrinker); 65954d91729SQi Zheng 66054d91729SQi Zheng deferred_split_shrinker->count_objects = deferred_split_count; 66154d91729SQi Zheng deferred_split_shrinker->scan_objects = deferred_split_scan; 66254d91729SQi Zheng shrinker_register(deferred_split_shrinker); 66354d91729SQi Zheng 66454d91729SQi Zheng return 0; 66554d91729SQi Zheng } 66654d91729SQi Zheng 66754d91729SQi Zheng static void __init thp_shrinker_exit(void) 66854d91729SQi Zheng { 66954d91729SQi Zheng shrinker_free(huge_zero_page_shrinker); 67054d91729SQi Zheng shrinker_free(deferred_split_shrinker); 67154d91729SQi Zheng } 67254d91729SQi Zheng 67371e3aac0SAndrea Arcangeli static int __init hugepage_init(void) 67471e3aac0SAndrea Arcangeli { 67571e3aac0SAndrea Arcangeli int err; 676569e5590SShaohua Li struct kobject *hugepage_kobj; 67771e3aac0SAndrea Arcangeli 6784b7167b9SAndrea Arcangeli if (!has_transparent_hugepage()) { 6793c556d24SPeter Xu transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED; 680569e5590SShaohua Li return -EINVAL; 6814b7167b9SAndrea Arcangeli } 6824b7167b9SAndrea Arcangeli 683ff20c2e0SKirill A. Shutemov /* 684ff20c2e0SKirill A. Shutemov * hugepages can't be allocated by the buddy allocator 685ff20c2e0SKirill A. Shutemov */ 6865e0a760bSKirill A. Shutemov MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER > MAX_PAGE_ORDER); 687ff20c2e0SKirill A. Shutemov /* 688ff20c2e0SKirill A. Shutemov * we use page->mapping and page->index in second tail page 689ff20c2e0SKirill A. Shutemov * as list_head: assuming THP order >= 2 690ff20c2e0SKirill A. Shutemov */ 691ff20c2e0SKirill A. Shutemov MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); 692ff20c2e0SKirill A. Shutemov 693569e5590SShaohua Li err = hugepage_init_sysfs(&hugepage_kobj); 694569e5590SShaohua Li if (err) 69565ebb64fSKirill A. Shutemov goto err_sysfs; 696ba76149fSAndrea Arcangeli 697b46e756fSKirill A. Shutemov err = khugepaged_init(); 698ba76149fSAndrea Arcangeli if (err) 69965ebb64fSKirill A. Shutemov goto err_slab; 700ba76149fSAndrea Arcangeli 70154d91729SQi Zheng err = thp_shrinker_init(); 70265ebb64fSKirill A. Shutemov if (err) 70354d91729SQi Zheng goto err_shrinker; 70497ae1749SKirill A. Shutemov 70597562cd2SRik van Riel /* 70697562cd2SRik van Riel * By default disable transparent hugepages on smaller systems, 70797562cd2SRik van Riel * where the extra memory used could hurt more than TLB overhead 70897562cd2SRik van Riel * is likely to save. The admin can still enable it through /sys. 70997562cd2SRik van Riel */ 710ca79b0c2SArun KS if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) { 71197562cd2SRik van Riel transparent_hugepage_flags = 0; 71279553da2SKirill A. Shutemov return 0; 71379553da2SKirill A. Shutemov } 71497562cd2SRik van Riel 71579553da2SKirill A. Shutemov err = start_stop_khugepaged(); 71665ebb64fSKirill A. Shutemov if (err) 71765ebb64fSKirill A. Shutemov goto err_khugepaged; 718ba76149fSAndrea Arcangeli 719569e5590SShaohua Li return 0; 72065ebb64fSKirill A. Shutemov err_khugepaged: 72154d91729SQi Zheng thp_shrinker_exit(); 72254d91729SQi Zheng err_shrinker: 723b46e756fSKirill A. Shutemov khugepaged_destroy(); 72465ebb64fSKirill A. Shutemov err_slab: 725569e5590SShaohua Li hugepage_exit_sysfs(hugepage_kobj); 72665ebb64fSKirill A. Shutemov err_sysfs: 727ba76149fSAndrea Arcangeli return err; 72871e3aac0SAndrea Arcangeli } 729a64fb3cdSPaul Gortmaker subsys_initcall(hugepage_init); 73071e3aac0SAndrea Arcangeli 73171e3aac0SAndrea Arcangeli static int __init setup_transparent_hugepage(char *str) 73271e3aac0SAndrea Arcangeli { 73371e3aac0SAndrea Arcangeli int ret = 0; 73471e3aac0SAndrea Arcangeli if (!str) 73571e3aac0SAndrea Arcangeli goto out; 73671e3aac0SAndrea Arcangeli if (!strcmp(str, "always")) { 73771e3aac0SAndrea Arcangeli set_bit(TRANSPARENT_HUGEPAGE_FLAG, 73871e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 73971e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 74071e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 74171e3aac0SAndrea Arcangeli ret = 1; 74271e3aac0SAndrea Arcangeli } else if (!strcmp(str, "madvise")) { 74371e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 74471e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 74571e3aac0SAndrea Arcangeli set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 74671e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 74771e3aac0SAndrea Arcangeli ret = 1; 74871e3aac0SAndrea Arcangeli } else if (!strcmp(str, "never")) { 74971e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_FLAG, 75071e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 75171e3aac0SAndrea Arcangeli clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, 75271e3aac0SAndrea Arcangeli &transparent_hugepage_flags); 75371e3aac0SAndrea Arcangeli ret = 1; 75471e3aac0SAndrea Arcangeli } 75571e3aac0SAndrea Arcangeli out: 75671e3aac0SAndrea Arcangeli if (!ret) 757ae3a8c1cSAndrew Morton pr_warn("transparent_hugepage= cannot parse, ignored\n"); 75871e3aac0SAndrea Arcangeli return ret; 75971e3aac0SAndrea Arcangeli } 76071e3aac0SAndrea Arcangeli __setup("transparent_hugepage=", setup_transparent_hugepage); 76171e3aac0SAndrea Arcangeli 762f55e1014SLinus Torvalds pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) 76371e3aac0SAndrea Arcangeli { 764f55e1014SLinus Torvalds if (likely(vma->vm_flags & VM_WRITE)) 765161e393cSRick Edgecombe pmd = pmd_mkwrite(pmd, vma); 76671e3aac0SAndrea Arcangeli return pmd; 76771e3aac0SAndrea Arcangeli } 76871e3aac0SAndrea Arcangeli 76987eaceb3SYang Shi #ifdef CONFIG_MEMCG 770f8baa6beSMatthew Wilcox (Oracle) static inline 771f8baa6beSMatthew Wilcox (Oracle) struct deferred_split *get_deferred_split_queue(struct folio *folio) 7729a982250SKirill A. Shutemov { 773f8baa6beSMatthew Wilcox (Oracle) struct mem_cgroup *memcg = folio_memcg(folio); 774f8baa6beSMatthew Wilcox (Oracle) struct pglist_data *pgdat = NODE_DATA(folio_nid(folio)); 77587eaceb3SYang Shi 77687eaceb3SYang Shi if (memcg) 77787eaceb3SYang Shi return &memcg->deferred_split_queue; 77887eaceb3SYang Shi else 77987eaceb3SYang Shi return &pgdat->deferred_split_queue; 7809a982250SKirill A. Shutemov } 78187eaceb3SYang Shi #else 782f8baa6beSMatthew Wilcox (Oracle) static inline 783f8baa6beSMatthew Wilcox (Oracle) struct deferred_split *get_deferred_split_queue(struct folio *folio) 78487eaceb3SYang Shi { 785f8baa6beSMatthew Wilcox (Oracle) struct pglist_data *pgdat = NODE_DATA(folio_nid(folio)); 78687eaceb3SYang Shi 78787eaceb3SYang Shi return &pgdat->deferred_split_queue; 78887eaceb3SYang Shi } 78987eaceb3SYang Shi #endif 7909a982250SKirill A. Shutemov 791da6e7bf3SMatthew Wilcox (Oracle) void folio_prep_large_rmappable(struct folio *folio) 7929a982250SKirill A. Shutemov { 7938897277aSMatthew Wilcox (Oracle) if (!folio || !folio_test_large(folio)) 7948897277aSMatthew Wilcox (Oracle) return; 7958897277aSMatthew Wilcox (Oracle) if (folio_order(folio) > 1) 7968991de90SMatthew Wilcox (Oracle) INIT_LIST_HEAD(&folio->_deferred_list); 797de53c05fSMatthew Wilcox (Oracle) folio_set_large_rmappable(folio); 7989a982250SKirill A. Shutemov } 7999a982250SKirill A. Shutemov 800a644b0abSMatthew Wilcox (Oracle) static inline bool is_transparent_hugepage(struct folio *folio) 801005ba37cSSean Christopherson { 802a644b0abSMatthew Wilcox (Oracle) if (!folio_test_large(folio)) 803fa1f68ccSZou Wei return false; 804005ba37cSSean Christopherson 805f04029f3SMatthew Wilcox (Oracle) return is_huge_zero_page(&folio->page) || 806de53c05fSMatthew Wilcox (Oracle) folio_test_large_rmappable(folio); 807005ba37cSSean Christopherson } 808005ba37cSSean Christopherson 80997d3d0f9SKirill A. Shutemov static unsigned long __thp_get_unmapped_area(struct file *filp, 81097d3d0f9SKirill A. Shutemov unsigned long addr, unsigned long len, 81174d2fad1SToshi Kani loff_t off, unsigned long flags, unsigned long size) 81274d2fad1SToshi Kani { 81374d2fad1SToshi Kani loff_t off_end = off + len; 81474d2fad1SToshi Kani loff_t off_align = round_up(off, size); 81596204e15SRyan Roberts unsigned long len_pad, ret, off_sub; 81674d2fad1SToshi Kani 8174ef9ad19SYang Shi if (IS_ENABLED(CONFIG_32BIT) || in_compat_syscall()) 8184ef9ad19SYang Shi return 0; 8194ef9ad19SYang Shi 82074d2fad1SToshi Kani if (off_end <= off_align || (off_end - off_align) < size) 82174d2fad1SToshi Kani return 0; 82274d2fad1SToshi Kani 82374d2fad1SToshi Kani len_pad = len + size; 82474d2fad1SToshi Kani if (len_pad < len || (off + len_pad) < off) 82574d2fad1SToshi Kani return 0; 82674d2fad1SToshi Kani 82797d3d0f9SKirill A. Shutemov ret = current->mm->get_unmapped_area(filp, addr, len_pad, 82874d2fad1SToshi Kani off >> PAGE_SHIFT, flags); 82997d3d0f9SKirill A. Shutemov 83097d3d0f9SKirill A. Shutemov /* 83197d3d0f9SKirill A. Shutemov * The failure might be due to length padding. The caller will retry 83297d3d0f9SKirill A. Shutemov * without the padding. 83397d3d0f9SKirill A. Shutemov */ 83497d3d0f9SKirill A. Shutemov if (IS_ERR_VALUE(ret)) 83574d2fad1SToshi Kani return 0; 83674d2fad1SToshi Kani 83797d3d0f9SKirill A. Shutemov /* 83897d3d0f9SKirill A. Shutemov * Do not try to align to THP boundary if allocation at the address 83997d3d0f9SKirill A. Shutemov * hint succeeds. 84097d3d0f9SKirill A. Shutemov */ 84197d3d0f9SKirill A. Shutemov if (ret == addr) 84274d2fad1SToshi Kani return addr; 84397d3d0f9SKirill A. Shutemov 84496204e15SRyan Roberts off_sub = (off - ret) & (size - 1); 84596204e15SRyan Roberts 84696204e15SRyan Roberts if (current->mm->get_unmapped_area == arch_get_unmapped_area_topdown && 84796204e15SRyan Roberts !off_sub) 84896204e15SRyan Roberts return ret + size; 84996204e15SRyan Roberts 85096204e15SRyan Roberts ret += off_sub; 85197d3d0f9SKirill A. Shutemov return ret; 85274d2fad1SToshi Kani } 85374d2fad1SToshi Kani 85474d2fad1SToshi Kani unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, 85574d2fad1SToshi Kani unsigned long len, unsigned long pgoff, unsigned long flags) 85674d2fad1SToshi Kani { 85797d3d0f9SKirill A. Shutemov unsigned long ret; 85874d2fad1SToshi Kani loff_t off = (loff_t)pgoff << PAGE_SHIFT; 85974d2fad1SToshi Kani 86097d3d0f9SKirill A. Shutemov ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE); 86197d3d0f9SKirill A. Shutemov if (ret) 86297d3d0f9SKirill A. Shutemov return ret; 8631854bc6eSWilliam Kucharski 86474d2fad1SToshi Kani return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); 86574d2fad1SToshi Kani } 86674d2fad1SToshi Kani EXPORT_SYMBOL_GPL(thp_get_unmapped_area); 86774d2fad1SToshi Kani 8682b740303SSouptick Joarder static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf, 8692b740303SSouptick Joarder struct page *page, gfp_t gfp) 87071e3aac0SAndrea Arcangeli { 87182b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 872cfe3236dSKefeng Wang struct folio *folio = page_folio(page); 87371e3aac0SAndrea Arcangeli pgtable_t pgtable; 87482b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 8752b740303SSouptick Joarder vm_fault_t ret = 0; 87671e3aac0SAndrea Arcangeli 877cfe3236dSKefeng Wang VM_BUG_ON_FOLIO(!folio_test_large(folio), folio); 87800501b53SJohannes Weiner 879cfe3236dSKefeng Wang if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) { 880cfe3236dSKefeng Wang folio_put(folio); 8816b251fc9SAndrea Arcangeli count_vm_event(THP_FAULT_FALLBACK); 88285b9f46eSDavid Rientjes count_vm_event(THP_FAULT_FALLBACK_CHARGE); 8836b251fc9SAndrea Arcangeli return VM_FAULT_FALLBACK; 8846b251fc9SAndrea Arcangeli } 885cfe3236dSKefeng Wang folio_throttle_swaprate(folio, gfp); 88671e3aac0SAndrea Arcangeli 8874cf58924SJoel Fernandes (Google) pgtable = pte_alloc_one(vma->vm_mm); 88800501b53SJohannes Weiner if (unlikely(!pgtable)) { 8896b31d595SMichal Hocko ret = VM_FAULT_OOM; 8906b31d595SMichal Hocko goto release; 89100501b53SJohannes Weiner } 89200501b53SJohannes Weiner 893c79b57e4SHuang Ying clear_huge_page(page, vmf->address, HPAGE_PMD_NR); 89452f37629SMinchan Kim /* 895cfe3236dSKefeng Wang * The memory barrier inside __folio_mark_uptodate makes sure that 89652f37629SMinchan Kim * clear_huge_page writes become visible before the set_pmd_at() 89752f37629SMinchan Kim * write. 89852f37629SMinchan Kim */ 899cfe3236dSKefeng Wang __folio_mark_uptodate(folio); 90071e3aac0SAndrea Arcangeli 90182b0f8c3SJan Kara vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 90282b0f8c3SJan Kara if (unlikely(!pmd_none(*vmf->pmd))) { 9036b31d595SMichal Hocko goto unlock_release; 90471e3aac0SAndrea Arcangeli } else { 90571e3aac0SAndrea Arcangeli pmd_t entry; 9066b251fc9SAndrea Arcangeli 9076b31d595SMichal Hocko ret = check_stable_address_space(vma->vm_mm); 9086b31d595SMichal Hocko if (ret) 9096b31d595SMichal Hocko goto unlock_release; 9106b31d595SMichal Hocko 9116b251fc9SAndrea Arcangeli /* Deliver the page fault to userland */ 9126b251fc9SAndrea Arcangeli if (userfaultfd_missing(vma)) { 91382b0f8c3SJan Kara spin_unlock(vmf->ptl); 914cfe3236dSKefeng Wang folio_put(folio); 915bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 9168fd5eda4SMiaohe Lin ret = handle_userfault(vmf, VM_UFFD_MISSING); 9178fd5eda4SMiaohe Lin VM_BUG_ON(ret & VM_FAULT_FALLBACK); 9188fd5eda4SMiaohe Lin return ret; 9196b251fc9SAndrea Arcangeli } 9206b251fc9SAndrea Arcangeli 9213122359aSKirill A. Shutemov entry = mk_huge_pmd(page, vma->vm_page_prot); 922f55e1014SLinus Torvalds entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 923cfe3236dSKefeng Wang folio_add_new_anon_rmap(folio, vma, haddr); 924cfe3236dSKefeng Wang folio_add_lru_vma(folio, vma); 92582b0f8c3SJan Kara pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); 92682b0f8c3SJan Kara set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 927fca40573SBibo Mao update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 928bae473a4SKirill A. Shutemov add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 929c4812909SKirill A. Shutemov mm_inc_nr_ptes(vma->vm_mm); 93082b0f8c3SJan Kara spin_unlock(vmf->ptl); 9316b251fc9SAndrea Arcangeli count_vm_event(THP_FAULT_ALLOC); 9329d82c694SJohannes Weiner count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC); 93371e3aac0SAndrea Arcangeli } 93471e3aac0SAndrea Arcangeli 935aa2e878eSDavid Rientjes return 0; 9366b31d595SMichal Hocko unlock_release: 9376b31d595SMichal Hocko spin_unlock(vmf->ptl); 9386b31d595SMichal Hocko release: 9396b31d595SMichal Hocko if (pgtable) 9406b31d595SMichal Hocko pte_free(vma->vm_mm, pgtable); 941cfe3236dSKefeng Wang folio_put(folio); 9426b31d595SMichal Hocko return ret; 9436b31d595SMichal Hocko 94471e3aac0SAndrea Arcangeli } 94571e3aac0SAndrea Arcangeli 946444eb2a4SMel Gorman /* 94721440d7eSDavid Rientjes * always: directly stall for all thp allocations 94821440d7eSDavid Rientjes * defer: wake kswapd and fail if not immediately available 94921440d7eSDavid Rientjes * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise 95021440d7eSDavid Rientjes * fail if not immediately available 95121440d7eSDavid Rientjes * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately 95221440d7eSDavid Rientjes * available 95321440d7eSDavid Rientjes * never: never stall for any thp allocation 954444eb2a4SMel Gorman */ 955164cc4feSRik van Riel gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma) 9560bbbc0b3SAndrea Arcangeli { 957164cc4feSRik van Riel const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE); 95889c83fb5SMichal Hocko 959ac79f78dSDavid Rientjes /* Always do synchronous compaction */ 96021440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 961a8282608SAndrea Arcangeli return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); 962ac79f78dSDavid Rientjes 963ac79f78dSDavid Rientjes /* Kick kcompactd and fail quickly */ 96421440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 96519deb769SDavid Rientjes return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; 966ac79f78dSDavid Rientjes 967ac79f78dSDavid Rientjes /* Synchronous compaction if madvised, otherwise kick kcompactd */ 96821440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 96919deb769SDavid Rientjes return GFP_TRANSHUGE_LIGHT | 97019deb769SDavid Rientjes (vma_madvised ? __GFP_DIRECT_RECLAIM : 971ac79f78dSDavid Rientjes __GFP_KSWAPD_RECLAIM); 972ac79f78dSDavid Rientjes 973ac79f78dSDavid Rientjes /* Only do synchronous compaction if madvised */ 97421440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 97519deb769SDavid Rientjes return GFP_TRANSHUGE_LIGHT | 97619deb769SDavid Rientjes (vma_madvised ? __GFP_DIRECT_RECLAIM : 0); 977ac79f78dSDavid Rientjes 97819deb769SDavid Rientjes return GFP_TRANSHUGE_LIGHT; 979444eb2a4SMel Gorman } 980444eb2a4SMel Gorman 981c4088ebdSKirill A. Shutemov /* Caller must hold page table lock. */ 9822efeb8daSMiaohe Lin static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 98397ae1749SKirill A. Shutemov struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 9845918d10aSKirill A. Shutemov struct page *zero_page) 985fc9fe822SKirill A. Shutemov { 986fc9fe822SKirill A. Shutemov pmd_t entry; 9877c414164SAndrew Morton if (!pmd_none(*pmd)) 9882efeb8daSMiaohe Lin return; 9895918d10aSKirill A. Shutemov entry = mk_pmd(zero_page, vma->vm_page_prot); 990fc9fe822SKirill A. Shutemov entry = pmd_mkhuge(entry); 9916b0b50b0SAneesh Kumar K.V pgtable_trans_huge_deposit(mm, pmd, pgtable); 992fc9fe822SKirill A. Shutemov set_pmd_at(mm, haddr, pmd, entry); 993c4812909SKirill A. Shutemov mm_inc_nr_ptes(mm); 994fc9fe822SKirill A. Shutemov } 995fc9fe822SKirill A. Shutemov 9962b740303SSouptick Joarder vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf) 99771e3aac0SAndrea Arcangeli { 99882b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 999077fcf11SAneesh Kumar K.V gfp_t gfp; 1000cb196ee1SMatthew Wilcox (Oracle) struct folio *folio; 100182b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 100271e3aac0SAndrea Arcangeli 10033485b883SRyan Roberts if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER)) 1004c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 100571e3aac0SAndrea Arcangeli if (unlikely(anon_vma_prepare(vma))) 100671e3aac0SAndrea Arcangeli return VM_FAULT_OOM; 10074fa6893fSYang Shi khugepaged_enter_vma(vma, vma->vm_flags); 1008d2081b2bSYang Shi 100982b0f8c3SJan Kara if (!(vmf->flags & FAULT_FLAG_WRITE) && 1010bae473a4SKirill A. Shutemov !mm_forbids_zeropage(vma->vm_mm) && 101179da5407SKirill A. Shutemov transparent_hugepage_use_zero_page()) { 101280371957SKirill A. Shutemov pgtable_t pgtable; 10135918d10aSKirill A. Shutemov struct page *zero_page; 10142b740303SSouptick Joarder vm_fault_t ret; 10154cf58924SJoel Fernandes (Google) pgtable = pte_alloc_one(vma->vm_mm); 101680371957SKirill A. Shutemov if (unlikely(!pgtable)) 101780371957SKirill A. Shutemov return VM_FAULT_OOM; 10186fcb52a5SAaron Lu zero_page = mm_get_huge_zero_page(vma->vm_mm); 10195918d10aSKirill A. Shutemov if (unlikely(!zero_page)) { 1020bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 102197ae1749SKirill A. Shutemov count_vm_event(THP_FAULT_FALLBACK); 1022c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 102397ae1749SKirill A. Shutemov } 102482b0f8c3SJan Kara vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 10256b251fc9SAndrea Arcangeli ret = 0; 102682b0f8c3SJan Kara if (pmd_none(*vmf->pmd)) { 10276b31d595SMichal Hocko ret = check_stable_address_space(vma->vm_mm); 10286b31d595SMichal Hocko if (ret) { 10296b31d595SMichal Hocko spin_unlock(vmf->ptl); 1030bfe8cc1dSGerald Schaefer pte_free(vma->vm_mm, pgtable); 10316b31d595SMichal Hocko } else if (userfaultfd_missing(vma)) { 103282b0f8c3SJan Kara spin_unlock(vmf->ptl); 1033bfe8cc1dSGerald Schaefer pte_free(vma->vm_mm, pgtable); 103482b0f8c3SJan Kara ret = handle_userfault(vmf, VM_UFFD_MISSING); 10356b251fc9SAndrea Arcangeli VM_BUG_ON(ret & VM_FAULT_FALLBACK); 10366b251fc9SAndrea Arcangeli } else { 1037bae473a4SKirill A. Shutemov set_huge_zero_page(pgtable, vma->vm_mm, vma, 103882b0f8c3SJan Kara haddr, vmf->pmd, zero_page); 1039fca40573SBibo Mao update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 104082b0f8c3SJan Kara spin_unlock(vmf->ptl); 10416b251fc9SAndrea Arcangeli } 1042bfe8cc1dSGerald Schaefer } else { 104382b0f8c3SJan Kara spin_unlock(vmf->ptl); 1044bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 1045bfe8cc1dSGerald Schaefer } 10466b251fc9SAndrea Arcangeli return ret; 104780371957SKirill A. Shutemov } 1048164cc4feSRik van Riel gfp = vma_thp_gfp_mask(vma); 1049cb196ee1SMatthew Wilcox (Oracle) folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true); 1050cb196ee1SMatthew Wilcox (Oracle) if (unlikely(!folio)) { 105181ab4201SAndi Kleen count_vm_event(THP_FAULT_FALLBACK); 1052c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 105381ab4201SAndi Kleen } 1054cb196ee1SMatthew Wilcox (Oracle) return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp); 105571e3aac0SAndrea Arcangeli } 105671e3aac0SAndrea Arcangeli 1057ae18d6dcSMatthew Wilcox static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 10583b6521f5SOliver O'Halloran pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, 10593b6521f5SOliver O'Halloran pgtable_t pgtable) 10605cad465dSMatthew Wilcox { 10615cad465dSMatthew Wilcox struct mm_struct *mm = vma->vm_mm; 10625cad465dSMatthew Wilcox pmd_t entry; 10635cad465dSMatthew Wilcox spinlock_t *ptl; 10645cad465dSMatthew Wilcox 10655cad465dSMatthew Wilcox ptl = pmd_lock(mm, pmd); 1066c6f3c5eeSAneesh Kumar K.V if (!pmd_none(*pmd)) { 1067c6f3c5eeSAneesh Kumar K.V if (write) { 1068c6f3c5eeSAneesh Kumar K.V if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) { 1069c6f3c5eeSAneesh Kumar K.V WARN_ON_ONCE(!is_huge_zero_pmd(*pmd)); 1070c6f3c5eeSAneesh Kumar K.V goto out_unlock; 1071c6f3c5eeSAneesh Kumar K.V } 1072c6f3c5eeSAneesh Kumar K.V entry = pmd_mkyoung(*pmd); 1073c6f3c5eeSAneesh Kumar K.V entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1074c6f3c5eeSAneesh Kumar K.V if (pmdp_set_access_flags(vma, addr, pmd, entry, 1)) 1075c6f3c5eeSAneesh Kumar K.V update_mmu_cache_pmd(vma, addr, pmd); 1076c6f3c5eeSAneesh Kumar K.V } 1077c6f3c5eeSAneesh Kumar K.V 1078c6f3c5eeSAneesh Kumar K.V goto out_unlock; 1079c6f3c5eeSAneesh Kumar K.V } 1080c6f3c5eeSAneesh Kumar K.V 1081f25748e3SDan Williams entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); 1082f25748e3SDan Williams if (pfn_t_devmap(pfn)) 1083f25748e3SDan Williams entry = pmd_mkdevmap(entry); 10845cad465dSMatthew Wilcox if (write) { 1085f55e1014SLinus Torvalds entry = pmd_mkyoung(pmd_mkdirty(entry)); 1086f55e1014SLinus Torvalds entry = maybe_pmd_mkwrite(entry, vma); 10875cad465dSMatthew Wilcox } 10883b6521f5SOliver O'Halloran 10893b6521f5SOliver O'Halloran if (pgtable) { 10903b6521f5SOliver O'Halloran pgtable_trans_huge_deposit(mm, pmd, pgtable); 1091c4812909SKirill A. Shutemov mm_inc_nr_ptes(mm); 1092c6f3c5eeSAneesh Kumar K.V pgtable = NULL; 10933b6521f5SOliver O'Halloran } 10943b6521f5SOliver O'Halloran 10955cad465dSMatthew Wilcox set_pmd_at(mm, addr, pmd, entry); 10965cad465dSMatthew Wilcox update_mmu_cache_pmd(vma, addr, pmd); 1097c6f3c5eeSAneesh Kumar K.V 1098c6f3c5eeSAneesh Kumar K.V out_unlock: 10995cad465dSMatthew Wilcox spin_unlock(ptl); 1100c6f3c5eeSAneesh Kumar K.V if (pgtable) 1101c6f3c5eeSAneesh Kumar K.V pte_free(mm, pgtable); 11025cad465dSMatthew Wilcox } 11035cad465dSMatthew Wilcox 11049a9731b1SThomas Hellstrom (VMware) /** 11057b806d22SLorenzo Stoakes * vmf_insert_pfn_pmd - insert a pmd size pfn 11069a9731b1SThomas Hellstrom (VMware) * @vmf: Structure describing the fault 11079a9731b1SThomas Hellstrom (VMware) * @pfn: pfn to insert 11089a9731b1SThomas Hellstrom (VMware) * @write: whether it's a write fault 11099a9731b1SThomas Hellstrom (VMware) * 11107b806d22SLorenzo Stoakes * Insert a pmd size pfn. See vmf_insert_pfn() for additional info. 11119a9731b1SThomas Hellstrom (VMware) * 11129a9731b1SThomas Hellstrom (VMware) * Return: vm_fault_t value. 11139a9731b1SThomas Hellstrom (VMware) */ 11147b806d22SLorenzo Stoakes vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write) 11155cad465dSMatthew Wilcox { 1116fce86ff5SDan Williams unsigned long addr = vmf->address & PMD_MASK; 1117fce86ff5SDan Williams struct vm_area_struct *vma = vmf->vma; 11187b806d22SLorenzo Stoakes pgprot_t pgprot = vma->vm_page_prot; 11193b6521f5SOliver O'Halloran pgtable_t pgtable = NULL; 1120fce86ff5SDan Williams 11215cad465dSMatthew Wilcox /* 11225cad465dSMatthew Wilcox * If we had pmd_special, we could avoid all these restrictions, 11235cad465dSMatthew Wilcox * but we need to be consistent with PTEs and architectures that 11245cad465dSMatthew Wilcox * can't support a 'special' bit. 11255cad465dSMatthew Wilcox */ 1126e1fb4a08SDave Jiang BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 1127e1fb4a08SDave Jiang !pfn_t_devmap(pfn)); 11285cad465dSMatthew Wilcox BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 11295cad465dSMatthew Wilcox (VM_PFNMAP|VM_MIXEDMAP)); 11305cad465dSMatthew Wilcox BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 11315cad465dSMatthew Wilcox 11325cad465dSMatthew Wilcox if (addr < vma->vm_start || addr >= vma->vm_end) 11335cad465dSMatthew Wilcox return VM_FAULT_SIGBUS; 1134308a047cSBorislav Petkov 11353b6521f5SOliver O'Halloran if (arch_needs_pgtable_deposit()) { 11364cf58924SJoel Fernandes (Google) pgtable = pte_alloc_one(vma->vm_mm); 11373b6521f5SOliver O'Halloran if (!pgtable) 11383b6521f5SOliver O'Halloran return VM_FAULT_OOM; 11393b6521f5SOliver O'Halloran } 11403b6521f5SOliver O'Halloran 1141308a047cSBorislav Petkov track_pfn_insert(vma, &pgprot, pfn); 1142308a047cSBorislav Petkov 1143fce86ff5SDan Williams insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable); 1144ae18d6dcSMatthew Wilcox return VM_FAULT_NOPAGE; 11455cad465dSMatthew Wilcox } 11467b806d22SLorenzo Stoakes EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); 11475cad465dSMatthew Wilcox 1148a00cc7d9SMatthew Wilcox #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1149f55e1014SLinus Torvalds static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) 1150a00cc7d9SMatthew Wilcox { 1151f55e1014SLinus Torvalds if (likely(vma->vm_flags & VM_WRITE)) 1152a00cc7d9SMatthew Wilcox pud = pud_mkwrite(pud); 1153a00cc7d9SMatthew Wilcox return pud; 1154a00cc7d9SMatthew Wilcox } 1155a00cc7d9SMatthew Wilcox 1156a00cc7d9SMatthew Wilcox static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 11577b806d22SLorenzo Stoakes pud_t *pud, pfn_t pfn, bool write) 1158a00cc7d9SMatthew Wilcox { 1159a00cc7d9SMatthew Wilcox struct mm_struct *mm = vma->vm_mm; 11607b806d22SLorenzo Stoakes pgprot_t prot = vma->vm_page_prot; 1161a00cc7d9SMatthew Wilcox pud_t entry; 1162a00cc7d9SMatthew Wilcox spinlock_t *ptl; 1163a00cc7d9SMatthew Wilcox 1164a00cc7d9SMatthew Wilcox ptl = pud_lock(mm, pud); 1165c6f3c5eeSAneesh Kumar K.V if (!pud_none(*pud)) { 1166c6f3c5eeSAneesh Kumar K.V if (write) { 1167c6f3c5eeSAneesh Kumar K.V if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) { 1168c6f3c5eeSAneesh Kumar K.V WARN_ON_ONCE(!is_huge_zero_pud(*pud)); 1169c6f3c5eeSAneesh Kumar K.V goto out_unlock; 1170c6f3c5eeSAneesh Kumar K.V } 1171c6f3c5eeSAneesh Kumar K.V entry = pud_mkyoung(*pud); 1172c6f3c5eeSAneesh Kumar K.V entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma); 1173c6f3c5eeSAneesh Kumar K.V if (pudp_set_access_flags(vma, addr, pud, entry, 1)) 1174c6f3c5eeSAneesh Kumar K.V update_mmu_cache_pud(vma, addr, pud); 1175c6f3c5eeSAneesh Kumar K.V } 1176c6f3c5eeSAneesh Kumar K.V goto out_unlock; 1177c6f3c5eeSAneesh Kumar K.V } 1178c6f3c5eeSAneesh Kumar K.V 1179a00cc7d9SMatthew Wilcox entry = pud_mkhuge(pfn_t_pud(pfn, prot)); 1180a00cc7d9SMatthew Wilcox if (pfn_t_devmap(pfn)) 1181a00cc7d9SMatthew Wilcox entry = pud_mkdevmap(entry); 1182a00cc7d9SMatthew Wilcox if (write) { 1183f55e1014SLinus Torvalds entry = pud_mkyoung(pud_mkdirty(entry)); 1184f55e1014SLinus Torvalds entry = maybe_pud_mkwrite(entry, vma); 1185a00cc7d9SMatthew Wilcox } 1186a00cc7d9SMatthew Wilcox set_pud_at(mm, addr, pud, entry); 1187a00cc7d9SMatthew Wilcox update_mmu_cache_pud(vma, addr, pud); 1188c6f3c5eeSAneesh Kumar K.V 1189c6f3c5eeSAneesh Kumar K.V out_unlock: 1190a00cc7d9SMatthew Wilcox spin_unlock(ptl); 1191a00cc7d9SMatthew Wilcox } 1192a00cc7d9SMatthew Wilcox 11939a9731b1SThomas Hellstrom (VMware) /** 11947b806d22SLorenzo Stoakes * vmf_insert_pfn_pud - insert a pud size pfn 11959a9731b1SThomas Hellstrom (VMware) * @vmf: Structure describing the fault 11969a9731b1SThomas Hellstrom (VMware) * @pfn: pfn to insert 11979a9731b1SThomas Hellstrom (VMware) * @write: whether it's a write fault 11989a9731b1SThomas Hellstrom (VMware) * 11997b806d22SLorenzo Stoakes * Insert a pud size pfn. See vmf_insert_pfn() for additional info. 12009a9731b1SThomas Hellstrom (VMware) * 12019a9731b1SThomas Hellstrom (VMware) * Return: vm_fault_t value. 12029a9731b1SThomas Hellstrom (VMware) */ 12037b806d22SLorenzo Stoakes vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write) 1204a00cc7d9SMatthew Wilcox { 1205fce86ff5SDan Williams unsigned long addr = vmf->address & PUD_MASK; 1206fce86ff5SDan Williams struct vm_area_struct *vma = vmf->vma; 12077b806d22SLorenzo Stoakes pgprot_t pgprot = vma->vm_page_prot; 1208fce86ff5SDan Williams 1209a00cc7d9SMatthew Wilcox /* 1210a00cc7d9SMatthew Wilcox * If we had pud_special, we could avoid all these restrictions, 1211a00cc7d9SMatthew Wilcox * but we need to be consistent with PTEs and architectures that 1212a00cc7d9SMatthew Wilcox * can't support a 'special' bit. 1213a00cc7d9SMatthew Wilcox */ 121462ec0d8cSDave Jiang BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 121562ec0d8cSDave Jiang !pfn_t_devmap(pfn)); 1216a00cc7d9SMatthew Wilcox BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 1217a00cc7d9SMatthew Wilcox (VM_PFNMAP|VM_MIXEDMAP)); 1218a00cc7d9SMatthew Wilcox BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 1219a00cc7d9SMatthew Wilcox 1220a00cc7d9SMatthew Wilcox if (addr < vma->vm_start || addr >= vma->vm_end) 1221a00cc7d9SMatthew Wilcox return VM_FAULT_SIGBUS; 1222a00cc7d9SMatthew Wilcox 1223a00cc7d9SMatthew Wilcox track_pfn_insert(vma, &pgprot, pfn); 1224a00cc7d9SMatthew Wilcox 12257b806d22SLorenzo Stoakes insert_pfn_pud(vma, addr, vmf->pud, pfn, write); 1226a00cc7d9SMatthew Wilcox return VM_FAULT_NOPAGE; 1227a00cc7d9SMatthew Wilcox } 12287b806d22SLorenzo Stoakes EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud); 1229a00cc7d9SMatthew Wilcox #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1230a00cc7d9SMatthew Wilcox 12313565fce3SDan Williams static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, 1232a69e4717SMiaohe Lin pmd_t *pmd, bool write) 12333565fce3SDan Williams { 12343565fce3SDan Williams pmd_t _pmd; 12353565fce3SDan Williams 1236a8f97366SKirill A. Shutemov _pmd = pmd_mkyoung(*pmd); 1237a69e4717SMiaohe Lin if (write) 1238a8f97366SKirill A. Shutemov _pmd = pmd_mkdirty(_pmd); 12393565fce3SDan Williams if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 1240a69e4717SMiaohe Lin pmd, _pmd, write)) 12413565fce3SDan Williams update_mmu_cache_pmd(vma, addr, pmd); 12423565fce3SDan Williams } 12433565fce3SDan Williams 12443565fce3SDan Williams struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, 1245df06b37fSKeith Busch pmd_t *pmd, int flags, struct dev_pagemap **pgmap) 12463565fce3SDan Williams { 12473565fce3SDan Williams unsigned long pfn = pmd_pfn(*pmd); 12483565fce3SDan Williams struct mm_struct *mm = vma->vm_mm; 12493565fce3SDan Williams struct page *page; 12500f089235SLogan Gunthorpe int ret; 12513565fce3SDan Williams 12523565fce3SDan Williams assert_spin_locked(pmd_lockptr(mm, pmd)); 12533565fce3SDan Williams 1254f6f37321SLinus Torvalds if (flags & FOLL_WRITE && !pmd_write(*pmd)) 12553565fce3SDan Williams return NULL; 12563565fce3SDan Williams 12573565fce3SDan Williams if (pmd_present(*pmd) && pmd_devmap(*pmd)) 12583565fce3SDan Williams /* pass */; 12593565fce3SDan Williams else 12603565fce3SDan Williams return NULL; 12613565fce3SDan Williams 12623565fce3SDan Williams if (flags & FOLL_TOUCH) 1263a69e4717SMiaohe Lin touch_pmd(vma, addr, pmd, flags & FOLL_WRITE); 12643565fce3SDan Williams 12653565fce3SDan Williams /* 12663565fce3SDan Williams * device mapped pages can only be returned if the 12673565fce3SDan Williams * caller will manage the page reference count. 12683565fce3SDan Williams */ 12693faa52c0SJohn Hubbard if (!(flags & (FOLL_GET | FOLL_PIN))) 12703565fce3SDan Williams return ERR_PTR(-EEXIST); 12713565fce3SDan Williams 12723565fce3SDan Williams pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; 1273df06b37fSKeith Busch *pgmap = get_dev_pagemap(pfn, *pgmap); 1274df06b37fSKeith Busch if (!*pgmap) 12753565fce3SDan Williams return ERR_PTR(-EFAULT); 12763565fce3SDan Williams page = pfn_to_page(pfn); 12770f089235SLogan Gunthorpe ret = try_grab_page(page, flags); 12780f089235SLogan Gunthorpe if (ret) 12790f089235SLogan Gunthorpe page = ERR_PTR(ret); 12803565fce3SDan Williams 12813565fce3SDan Williams return page; 12823565fce3SDan Williams } 12833565fce3SDan Williams 128471e3aac0SAndrea Arcangeli int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 128571e3aac0SAndrea Arcangeli pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 12868f34f1eaSPeter Xu struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma) 128771e3aac0SAndrea Arcangeli { 1288c4088ebdSKirill A. Shutemov spinlock_t *dst_ptl, *src_ptl; 128971e3aac0SAndrea Arcangeli struct page *src_page; 129096c772c2SDavid Hildenbrand struct folio *src_folio; 129171e3aac0SAndrea Arcangeli pmd_t pmd; 129212c9d70bSMatthew Wilcox pgtable_t pgtable = NULL; 1293628d47ceSKirill A. Shutemov int ret = -ENOMEM; 129471e3aac0SAndrea Arcangeli 1295628d47ceSKirill A. Shutemov /* Skip if can be re-fill on fault */ 12968f34f1eaSPeter Xu if (!vma_is_anonymous(dst_vma)) 1297628d47ceSKirill A. Shutemov return 0; 1298628d47ceSKirill A. Shutemov 12994cf58924SJoel Fernandes (Google) pgtable = pte_alloc_one(dst_mm); 130071e3aac0SAndrea Arcangeli if (unlikely(!pgtable)) 130171e3aac0SAndrea Arcangeli goto out; 130271e3aac0SAndrea Arcangeli 1303c4088ebdSKirill A. Shutemov dst_ptl = pmd_lock(dst_mm, dst_pmd); 1304c4088ebdSKirill A. Shutemov src_ptl = pmd_lockptr(src_mm, src_pmd); 1305c4088ebdSKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 130671e3aac0SAndrea Arcangeli 130771e3aac0SAndrea Arcangeli ret = -EAGAIN; 130871e3aac0SAndrea Arcangeli pmd = *src_pmd; 130984c3fc4eSZi Yan 131084c3fc4eSZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 131184c3fc4eSZi Yan if (unlikely(is_swap_pmd(pmd))) { 131284c3fc4eSZi Yan swp_entry_t entry = pmd_to_swp_entry(pmd); 131384c3fc4eSZi Yan 131484c3fc4eSZi Yan VM_BUG_ON(!is_pmd_migration_entry(pmd)); 13156c287605SDavid Hildenbrand if (!is_readable_migration_entry(entry)) { 13164dd845b5SAlistair Popple entry = make_readable_migration_entry( 13174dd845b5SAlistair Popple swp_offset(entry)); 131884c3fc4eSZi Yan pmd = swp_entry_to_pmd(entry); 1319ab6e3d09SNaoya Horiguchi if (pmd_swp_soft_dirty(*src_pmd)) 1320ab6e3d09SNaoya Horiguchi pmd = pmd_swp_mksoft_dirty(pmd); 13218f34f1eaSPeter Xu if (pmd_swp_uffd_wp(*src_pmd)) 13228f34f1eaSPeter Xu pmd = pmd_swp_mkuffd_wp(pmd); 132384c3fc4eSZi Yan set_pmd_at(src_mm, addr, src_pmd, pmd); 132484c3fc4eSZi Yan } 1325dd8a67f9SZi Yan add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1326af5b0f6aSKirill A. Shutemov mm_inc_nr_ptes(dst_mm); 1327dd8a67f9SZi Yan pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 13288f34f1eaSPeter Xu if (!userfaultfd_wp(dst_vma)) 13298f34f1eaSPeter Xu pmd = pmd_swp_clear_uffd_wp(pmd); 133084c3fc4eSZi Yan set_pmd_at(dst_mm, addr, dst_pmd, pmd); 133184c3fc4eSZi Yan ret = 0; 133284c3fc4eSZi Yan goto out_unlock; 133384c3fc4eSZi Yan } 133484c3fc4eSZi Yan #endif 133584c3fc4eSZi Yan 1336628d47ceSKirill A. Shutemov if (unlikely(!pmd_trans_huge(pmd))) { 133771e3aac0SAndrea Arcangeli pte_free(dst_mm, pgtable); 133871e3aac0SAndrea Arcangeli goto out_unlock; 133971e3aac0SAndrea Arcangeli } 1340fc9fe822SKirill A. Shutemov /* 1341c4088ebdSKirill A. Shutemov * When page table lock is held, the huge zero pmd should not be 1342fc9fe822SKirill A. Shutemov * under splitting since we don't split the page itself, only pmd to 1343fc9fe822SKirill A. Shutemov * a page table. 1344fc9fe822SKirill A. Shutemov */ 1345fc9fe822SKirill A. Shutemov if (is_huge_zero_pmd(pmd)) { 134697ae1749SKirill A. Shutemov /* 134797ae1749SKirill A. Shutemov * get_huge_zero_page() will never allocate a new page here, 134897ae1749SKirill A. Shutemov * since we already have a zero page to copy. It just takes a 134997ae1749SKirill A. Shutemov * reference. 135097ae1749SKirill A. Shutemov */ 13515fc7a5f6SPeter Xu mm_get_huge_zero_page(dst_mm); 13525fc7a5f6SPeter Xu goto out_zero_page; 1353fc9fe822SKirill A. Shutemov } 1354de466bd6SMel Gorman 135571e3aac0SAndrea Arcangeli src_page = pmd_page(pmd); 1356309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 135796c772c2SDavid Hildenbrand src_folio = page_folio(src_page); 1358d042035eSPeter Xu 135996c772c2SDavid Hildenbrand folio_get(src_folio); 136096c772c2SDavid Hildenbrand if (unlikely(folio_try_dup_anon_rmap_pmd(src_folio, src_page, src_vma))) { 1361fb3d824dSDavid Hildenbrand /* Page maybe pinned: split and retry the fault on PTEs. */ 136296c772c2SDavid Hildenbrand folio_put(src_folio); 1363d042035eSPeter Xu pte_free(dst_mm, pgtable); 1364d042035eSPeter Xu spin_unlock(src_ptl); 1365d042035eSPeter Xu spin_unlock(dst_ptl); 13668f34f1eaSPeter Xu __split_huge_pmd(src_vma, src_pmd, addr, false, NULL); 1367d042035eSPeter Xu return -EAGAIN; 1368d042035eSPeter Xu } 136971e3aac0SAndrea Arcangeli add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 13705fc7a5f6SPeter Xu out_zero_page: 1371c4812909SKirill A. Shutemov mm_inc_nr_ptes(dst_mm); 13725c7fb56eSDan Williams pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 137371e3aac0SAndrea Arcangeli pmdp_set_wrprotect(src_mm, addr, src_pmd); 13748f34f1eaSPeter Xu if (!userfaultfd_wp(dst_vma)) 13758f34f1eaSPeter Xu pmd = pmd_clear_uffd_wp(pmd); 137671e3aac0SAndrea Arcangeli pmd = pmd_mkold(pmd_wrprotect(pmd)); 137771e3aac0SAndrea Arcangeli set_pmd_at(dst_mm, addr, dst_pmd, pmd); 137871e3aac0SAndrea Arcangeli 137971e3aac0SAndrea Arcangeli ret = 0; 138071e3aac0SAndrea Arcangeli out_unlock: 1381c4088ebdSKirill A. Shutemov spin_unlock(src_ptl); 1382c4088ebdSKirill A. Shutemov spin_unlock(dst_ptl); 138371e3aac0SAndrea Arcangeli out: 138471e3aac0SAndrea Arcangeli return ret; 138571e3aac0SAndrea Arcangeli } 138671e3aac0SAndrea Arcangeli 1387a00cc7d9SMatthew Wilcox #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1388a00cc7d9SMatthew Wilcox static void touch_pud(struct vm_area_struct *vma, unsigned long addr, 13895fe653e9SMiaohe Lin pud_t *pud, bool write) 1390a00cc7d9SMatthew Wilcox { 1391a00cc7d9SMatthew Wilcox pud_t _pud; 1392a00cc7d9SMatthew Wilcox 1393a8f97366SKirill A. Shutemov _pud = pud_mkyoung(*pud); 13945fe653e9SMiaohe Lin if (write) 1395a8f97366SKirill A. Shutemov _pud = pud_mkdirty(_pud); 1396a00cc7d9SMatthew Wilcox if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, 13975fe653e9SMiaohe Lin pud, _pud, write)) 1398a00cc7d9SMatthew Wilcox update_mmu_cache_pud(vma, addr, pud); 1399a00cc7d9SMatthew Wilcox } 1400a00cc7d9SMatthew Wilcox 1401a00cc7d9SMatthew Wilcox struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, 1402df06b37fSKeith Busch pud_t *pud, int flags, struct dev_pagemap **pgmap) 1403a00cc7d9SMatthew Wilcox { 1404a00cc7d9SMatthew Wilcox unsigned long pfn = pud_pfn(*pud); 1405a00cc7d9SMatthew Wilcox struct mm_struct *mm = vma->vm_mm; 1406a00cc7d9SMatthew Wilcox struct page *page; 14070f089235SLogan Gunthorpe int ret; 1408a00cc7d9SMatthew Wilcox 1409a00cc7d9SMatthew Wilcox assert_spin_locked(pud_lockptr(mm, pud)); 1410a00cc7d9SMatthew Wilcox 1411f6f37321SLinus Torvalds if (flags & FOLL_WRITE && !pud_write(*pud)) 1412a00cc7d9SMatthew Wilcox return NULL; 1413a00cc7d9SMatthew Wilcox 1414a00cc7d9SMatthew Wilcox if (pud_present(*pud) && pud_devmap(*pud)) 1415a00cc7d9SMatthew Wilcox /* pass */; 1416a00cc7d9SMatthew Wilcox else 1417a00cc7d9SMatthew Wilcox return NULL; 1418a00cc7d9SMatthew Wilcox 1419a00cc7d9SMatthew Wilcox if (flags & FOLL_TOUCH) 14205fe653e9SMiaohe Lin touch_pud(vma, addr, pud, flags & FOLL_WRITE); 1421a00cc7d9SMatthew Wilcox 1422a00cc7d9SMatthew Wilcox /* 1423a00cc7d9SMatthew Wilcox * device mapped pages can only be returned if the 1424a00cc7d9SMatthew Wilcox * caller will manage the page reference count. 14253faa52c0SJohn Hubbard * 14263faa52c0SJohn Hubbard * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here: 1427a00cc7d9SMatthew Wilcox */ 14283faa52c0SJohn Hubbard if (!(flags & (FOLL_GET | FOLL_PIN))) 1429a00cc7d9SMatthew Wilcox return ERR_PTR(-EEXIST); 1430a00cc7d9SMatthew Wilcox 1431a00cc7d9SMatthew Wilcox pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; 1432df06b37fSKeith Busch *pgmap = get_dev_pagemap(pfn, *pgmap); 1433df06b37fSKeith Busch if (!*pgmap) 1434a00cc7d9SMatthew Wilcox return ERR_PTR(-EFAULT); 1435a00cc7d9SMatthew Wilcox page = pfn_to_page(pfn); 14360f089235SLogan Gunthorpe 14370f089235SLogan Gunthorpe ret = try_grab_page(page, flags); 14380f089235SLogan Gunthorpe if (ret) 14390f089235SLogan Gunthorpe page = ERR_PTR(ret); 1440a00cc7d9SMatthew Wilcox 1441a00cc7d9SMatthew Wilcox return page; 1442a00cc7d9SMatthew Wilcox } 1443a00cc7d9SMatthew Wilcox 1444a00cc7d9SMatthew Wilcox int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1445a00cc7d9SMatthew Wilcox pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1446a00cc7d9SMatthew Wilcox struct vm_area_struct *vma) 1447a00cc7d9SMatthew Wilcox { 1448a00cc7d9SMatthew Wilcox spinlock_t *dst_ptl, *src_ptl; 1449a00cc7d9SMatthew Wilcox pud_t pud; 1450a00cc7d9SMatthew Wilcox int ret; 1451a00cc7d9SMatthew Wilcox 1452a00cc7d9SMatthew Wilcox dst_ptl = pud_lock(dst_mm, dst_pud); 1453a00cc7d9SMatthew Wilcox src_ptl = pud_lockptr(src_mm, src_pud); 1454a00cc7d9SMatthew Wilcox spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1455a00cc7d9SMatthew Wilcox 1456a00cc7d9SMatthew Wilcox ret = -EAGAIN; 1457a00cc7d9SMatthew Wilcox pud = *src_pud; 1458a00cc7d9SMatthew Wilcox if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) 1459a00cc7d9SMatthew Wilcox goto out_unlock; 1460a00cc7d9SMatthew Wilcox 1461a00cc7d9SMatthew Wilcox /* 1462a00cc7d9SMatthew Wilcox * When page table lock is held, the huge zero pud should not be 1463a00cc7d9SMatthew Wilcox * under splitting since we don't split the page itself, only pud to 1464a00cc7d9SMatthew Wilcox * a page table. 1465a00cc7d9SMatthew Wilcox */ 1466a00cc7d9SMatthew Wilcox if (is_huge_zero_pud(pud)) { 1467a00cc7d9SMatthew Wilcox /* No huge zero pud yet */ 1468a00cc7d9SMatthew Wilcox } 1469a00cc7d9SMatthew Wilcox 1470fb3d824dSDavid Hildenbrand /* 147196c772c2SDavid Hildenbrand * TODO: once we support anonymous pages, use 147296c772c2SDavid Hildenbrand * folio_try_dup_anon_rmap_*() and split if duplicating fails. 1473fb3d824dSDavid Hildenbrand */ 1474a00cc7d9SMatthew Wilcox pudp_set_wrprotect(src_mm, addr, src_pud); 1475a00cc7d9SMatthew Wilcox pud = pud_mkold(pud_wrprotect(pud)); 1476a00cc7d9SMatthew Wilcox set_pud_at(dst_mm, addr, dst_pud, pud); 1477a00cc7d9SMatthew Wilcox 1478a00cc7d9SMatthew Wilcox ret = 0; 1479a00cc7d9SMatthew Wilcox out_unlock: 1480a00cc7d9SMatthew Wilcox spin_unlock(src_ptl); 1481a00cc7d9SMatthew Wilcox spin_unlock(dst_ptl); 1482a00cc7d9SMatthew Wilcox return ret; 1483a00cc7d9SMatthew Wilcox } 1484a00cc7d9SMatthew Wilcox 1485a00cc7d9SMatthew Wilcox void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) 1486a00cc7d9SMatthew Wilcox { 1487a00cc7d9SMatthew Wilcox bool write = vmf->flags & FAULT_FLAG_WRITE; 1488a00cc7d9SMatthew Wilcox 1489a00cc7d9SMatthew Wilcox vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); 1490a00cc7d9SMatthew Wilcox if (unlikely(!pud_same(*vmf->pud, orig_pud))) 1491a00cc7d9SMatthew Wilcox goto unlock; 1492a00cc7d9SMatthew Wilcox 14935fe653e9SMiaohe Lin touch_pud(vmf->vma, vmf->address, vmf->pud, write); 1494a00cc7d9SMatthew Wilcox unlock: 1495a00cc7d9SMatthew Wilcox spin_unlock(vmf->ptl); 1496a00cc7d9SMatthew Wilcox } 1497a00cc7d9SMatthew Wilcox #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1498a00cc7d9SMatthew Wilcox 14995db4f15cSYang Shi void huge_pmd_set_accessed(struct vm_fault *vmf) 1500a1dd450bSWill Deacon { 150120f664aaSMinchan Kim bool write = vmf->flags & FAULT_FLAG_WRITE; 1502a1dd450bSWill Deacon 150382b0f8c3SJan Kara vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1504a69e4717SMiaohe Lin if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd))) 1505a1dd450bSWill Deacon goto unlock; 1506a1dd450bSWill Deacon 1507a69e4717SMiaohe Lin touch_pmd(vmf->vma, vmf->address, vmf->pmd, write); 1508a1dd450bSWill Deacon 1509a1dd450bSWill Deacon unlock: 151082b0f8c3SJan Kara spin_unlock(vmf->ptl); 1511a1dd450bSWill Deacon } 1512a1dd450bSWill Deacon 15135db4f15cSYang Shi vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf) 151471e3aac0SAndrea Arcangeli { 1515c89357e2SDavid Hildenbrand const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; 151682b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 15172fad3d14SMatthew Wilcox (Oracle) struct folio *folio; 15183917c802SKirill A. Shutemov struct page *page; 151982b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 15205db4f15cSYang Shi pmd_t orig_pmd = vmf->orig_pmd; 152171e3aac0SAndrea Arcangeli 152282b0f8c3SJan Kara vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); 152381d1b09cSSasha Levin VM_BUG_ON_VMA(!vma->anon_vma, vma); 15243917c802SKirill A. Shutemov 152593b4796dSKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) 15263917c802SKirill A. Shutemov goto fallback; 15273917c802SKirill A. Shutemov 152882b0f8c3SJan Kara spin_lock(vmf->ptl); 15293917c802SKirill A. Shutemov 15303917c802SKirill A. Shutemov if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 15313917c802SKirill A. Shutemov spin_unlock(vmf->ptl); 15323917c802SKirill A. Shutemov return 0; 15333917c802SKirill A. Shutemov } 153471e3aac0SAndrea Arcangeli 153571e3aac0SAndrea Arcangeli page = pmd_page(orig_pmd); 15362fad3d14SMatthew Wilcox (Oracle) folio = page_folio(page); 1537f6004e73SMiaohe Lin VM_BUG_ON_PAGE(!PageHead(page), page); 15383917c802SKirill A. Shutemov 15396c287605SDavid Hildenbrand /* Early check when only holding the PT lock. */ 15406c287605SDavid Hildenbrand if (PageAnonExclusive(page)) 15416c287605SDavid Hildenbrand goto reuse; 15426c287605SDavid Hildenbrand 15432fad3d14SMatthew Wilcox (Oracle) if (!folio_trylock(folio)) { 15442fad3d14SMatthew Wilcox (Oracle) folio_get(folio); 1545ba3c4ce6SHuang Ying spin_unlock(vmf->ptl); 15462fad3d14SMatthew Wilcox (Oracle) folio_lock(folio); 1547ba3c4ce6SHuang Ying spin_lock(vmf->ptl); 1548ba3c4ce6SHuang Ying if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 15493917c802SKirill A. Shutemov spin_unlock(vmf->ptl); 15502fad3d14SMatthew Wilcox (Oracle) folio_unlock(folio); 15512fad3d14SMatthew Wilcox (Oracle) folio_put(folio); 15523917c802SKirill A. Shutemov return 0; 1553ba3c4ce6SHuang Ying } 15542fad3d14SMatthew Wilcox (Oracle) folio_put(folio); 1555ba3c4ce6SHuang Ying } 15563917c802SKirill A. Shutemov 15576c287605SDavid Hildenbrand /* Recheck after temporarily dropping the PT lock. */ 15586c287605SDavid Hildenbrand if (PageAnonExclusive(page)) { 15592fad3d14SMatthew Wilcox (Oracle) folio_unlock(folio); 15606c287605SDavid Hildenbrand goto reuse; 15616c287605SDavid Hildenbrand } 15626c287605SDavid Hildenbrand 15633917c802SKirill A. Shutemov /* 15642fad3d14SMatthew Wilcox (Oracle) * See do_wp_page(): we can only reuse the folio exclusively if 15652fad3d14SMatthew Wilcox (Oracle) * there are no additional references. Note that we always drain 15661fec6890SMatthew Wilcox (Oracle) * the LRU cache immediately after adding a THP. 15673917c802SKirill A. Shutemov */ 15682fad3d14SMatthew Wilcox (Oracle) if (folio_ref_count(folio) > 15692fad3d14SMatthew Wilcox (Oracle) 1 + folio_test_swapcache(folio) * folio_nr_pages(folio)) 15703bff7e3fSDavid Hildenbrand goto unlock_fallback; 15712fad3d14SMatthew Wilcox (Oracle) if (folio_test_swapcache(folio)) 15722fad3d14SMatthew Wilcox (Oracle) folio_free_swap(folio); 15732fad3d14SMatthew Wilcox (Oracle) if (folio_ref_count(folio) == 1) { 157471e3aac0SAndrea Arcangeli pmd_t entry; 15756c54dc6cSDavid Hildenbrand 157606968625SDavid Hildenbrand folio_move_anon_rmap(folio, vma); 15775ca43289SDavid Hildenbrand SetPageAnonExclusive(page); 15782fad3d14SMatthew Wilcox (Oracle) folio_unlock(folio); 15796c287605SDavid Hildenbrand reuse: 1580c89357e2SDavid Hildenbrand if (unlikely(unshare)) { 1581c89357e2SDavid Hildenbrand spin_unlock(vmf->ptl); 1582c89357e2SDavid Hildenbrand return 0; 1583c89357e2SDavid Hildenbrand } 158471e3aac0SAndrea Arcangeli entry = pmd_mkyoung(orig_pmd); 1585f55e1014SLinus Torvalds entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 158682b0f8c3SJan Kara if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) 158782b0f8c3SJan Kara update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 15883917c802SKirill A. Shutemov spin_unlock(vmf->ptl); 1589cb8d8633SDavid Hildenbrand return 0; 159071e3aac0SAndrea Arcangeli } 15913917c802SKirill A. Shutemov 15923bff7e3fSDavid Hildenbrand unlock_fallback: 15932fad3d14SMatthew Wilcox (Oracle) folio_unlock(folio); 159482b0f8c3SJan Kara spin_unlock(vmf->ptl); 15953917c802SKirill A. Shutemov fallback: 15963917c802SKirill A. Shutemov __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL); 15973917c802SKirill A. Shutemov return VM_FAULT_FALLBACK; 159871e3aac0SAndrea Arcangeli } 159971e3aac0SAndrea Arcangeli 1600c27f479eSDavid Hildenbrand static inline bool can_change_pmd_writable(struct vm_area_struct *vma, 1601c27f479eSDavid Hildenbrand unsigned long addr, pmd_t pmd) 1602c27f479eSDavid Hildenbrand { 1603c27f479eSDavid Hildenbrand struct page *page; 1604c27f479eSDavid Hildenbrand 1605c27f479eSDavid Hildenbrand if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE))) 1606c27f479eSDavid Hildenbrand return false; 1607c27f479eSDavid Hildenbrand 1608c27f479eSDavid Hildenbrand /* Don't touch entries that are not even readable (NUMA hinting). */ 1609c27f479eSDavid Hildenbrand if (pmd_protnone(pmd)) 1610c27f479eSDavid Hildenbrand return false; 1611c27f479eSDavid Hildenbrand 1612c27f479eSDavid Hildenbrand /* Do we need write faults for softdirty tracking? */ 1613c27f479eSDavid Hildenbrand if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd)) 1614c27f479eSDavid Hildenbrand return false; 1615c27f479eSDavid Hildenbrand 1616c27f479eSDavid Hildenbrand /* Do we need write faults for uffd-wp tracking? */ 1617c27f479eSDavid Hildenbrand if (userfaultfd_huge_pmd_wp(vma, pmd)) 1618c27f479eSDavid Hildenbrand return false; 1619c27f479eSDavid Hildenbrand 1620c27f479eSDavid Hildenbrand if (!(vma->vm_flags & VM_SHARED)) { 1621c27f479eSDavid Hildenbrand /* See can_change_pte_writable(). */ 1622c27f479eSDavid Hildenbrand page = vm_normal_page_pmd(vma, addr, pmd); 1623c27f479eSDavid Hildenbrand return page && PageAnon(page) && PageAnonExclusive(page); 1624c27f479eSDavid Hildenbrand } 1625c27f479eSDavid Hildenbrand 1626c27f479eSDavid Hildenbrand /* See can_change_pte_writable(). */ 1627c27f479eSDavid Hildenbrand return pmd_dirty(pmd); 1628c27f479eSDavid Hildenbrand } 1629c27f479eSDavid Hildenbrand 16305535be30SDavid Hildenbrand /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */ 16315535be30SDavid Hildenbrand static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page, 16325535be30SDavid Hildenbrand struct vm_area_struct *vma, 16335535be30SDavid Hildenbrand unsigned int flags) 16348310d48bSKeno Fischer { 16355535be30SDavid Hildenbrand /* If the pmd is writable, we can write to the page. */ 16365535be30SDavid Hildenbrand if (pmd_write(pmd)) 16375535be30SDavid Hildenbrand return true; 16385535be30SDavid Hildenbrand 16395535be30SDavid Hildenbrand /* Maybe FOLL_FORCE is set to override it? */ 16405535be30SDavid Hildenbrand if (!(flags & FOLL_FORCE)) 16415535be30SDavid Hildenbrand return false; 16425535be30SDavid Hildenbrand 16435535be30SDavid Hildenbrand /* But FOLL_FORCE has no effect on shared mappings */ 16445535be30SDavid Hildenbrand if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED)) 16455535be30SDavid Hildenbrand return false; 16465535be30SDavid Hildenbrand 16475535be30SDavid Hildenbrand /* ... or read-only private ones */ 16485535be30SDavid Hildenbrand if (!(vma->vm_flags & VM_MAYWRITE)) 16495535be30SDavid Hildenbrand return false; 16505535be30SDavid Hildenbrand 16515535be30SDavid Hildenbrand /* ... or already writable ones that just need to take a write fault */ 16525535be30SDavid Hildenbrand if (vma->vm_flags & VM_WRITE) 16535535be30SDavid Hildenbrand return false; 16545535be30SDavid Hildenbrand 16555535be30SDavid Hildenbrand /* 16565535be30SDavid Hildenbrand * See can_change_pte_writable(): we broke COW and could map the page 16575535be30SDavid Hildenbrand * writable if we have an exclusive anonymous page ... 16585535be30SDavid Hildenbrand */ 16595535be30SDavid Hildenbrand if (!page || !PageAnon(page) || !PageAnonExclusive(page)) 16605535be30SDavid Hildenbrand return false; 16615535be30SDavid Hildenbrand 16625535be30SDavid Hildenbrand /* ... and a write-fault isn't required for other reasons. */ 16635535be30SDavid Hildenbrand if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd)) 16645535be30SDavid Hildenbrand return false; 16655535be30SDavid Hildenbrand return !userfaultfd_huge_pmd_wp(vma, pmd); 16668310d48bSKeno Fischer } 16678310d48bSKeno Fischer 1668b676b293SDavid Rientjes struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 166971e3aac0SAndrea Arcangeli unsigned long addr, 167071e3aac0SAndrea Arcangeli pmd_t *pmd, 167171e3aac0SAndrea Arcangeli unsigned int flags) 167271e3aac0SAndrea Arcangeli { 1673b676b293SDavid Rientjes struct mm_struct *mm = vma->vm_mm; 16745535be30SDavid Hildenbrand struct page *page; 16750f089235SLogan Gunthorpe int ret; 167671e3aac0SAndrea Arcangeli 1677c4088ebdSKirill A. Shutemov assert_spin_locked(pmd_lockptr(mm, pmd)); 167871e3aac0SAndrea Arcangeli 16795535be30SDavid Hildenbrand page = pmd_page(*pmd); 16805535be30SDavid Hildenbrand VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); 16815535be30SDavid Hildenbrand 16825535be30SDavid Hildenbrand if ((flags & FOLL_WRITE) && 16835535be30SDavid Hildenbrand !can_follow_write_pmd(*pmd, page, vma, flags)) 16845535be30SDavid Hildenbrand return NULL; 168571e3aac0SAndrea Arcangeli 168685facf25SKirill A. Shutemov /* Avoid dumping huge zero page */ 168785facf25SKirill A. Shutemov if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 168885facf25SKirill A. Shutemov return ERR_PTR(-EFAULT); 168985facf25SKirill A. Shutemov 1690d74943a2SDavid Hildenbrand if (pmd_protnone(*pmd) && !gup_can_follow_protnone(vma, flags)) 16915535be30SDavid Hildenbrand return NULL; 16923faa52c0SJohn Hubbard 169384209e87SDavid Hildenbrand if (!pmd_write(*pmd) && gup_must_unshare(vma, flags, page)) 1694a7f22660SDavid Hildenbrand return ERR_PTR(-EMLINK); 1695a7f22660SDavid Hildenbrand 1696b6a2619cSDavid Hildenbrand VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && 1697b6a2619cSDavid Hildenbrand !PageAnonExclusive(page), page); 1698b6a2619cSDavid Hildenbrand 16990f089235SLogan Gunthorpe ret = try_grab_page(page, flags); 17000f089235SLogan Gunthorpe if (ret) 17010f089235SLogan Gunthorpe return ERR_PTR(ret); 17023faa52c0SJohn Hubbard 17033565fce3SDan Williams if (flags & FOLL_TOUCH) 1704a69e4717SMiaohe Lin touch_pmd(vma, addr, pmd, flags & FOLL_WRITE); 17053faa52c0SJohn Hubbard 170671e3aac0SAndrea Arcangeli page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1707ca120cf6SDan Williams VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); 170871e3aac0SAndrea Arcangeli 170971e3aac0SAndrea Arcangeli return page; 171071e3aac0SAndrea Arcangeli } 171171e3aac0SAndrea Arcangeli 1712d10e63f2SMel Gorman /* NUMA hinting page fault entry point for trans huge pmds */ 17135db4f15cSYang Shi vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf) 1714d10e63f2SMel Gorman { 171582b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 1716c5b5a3ddSYang Shi pmd_t oldpmd = vmf->orig_pmd; 1717c5b5a3ddSYang Shi pmd_t pmd; 1718667ffc31SKefeng Wang struct folio *folio; 171982b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1720667ffc31SKefeng Wang int nid = NUMA_NO_NODE; 172133024536SHuang Ying int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK); 17226a56ccbcSDavid Hildenbrand bool migrated = false, writable = false; 17236688cc05SPeter Zijlstra int flags = 0; 1724d10e63f2SMel Gorman 172582b0f8c3SJan Kara vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1726c5b5a3ddSYang Shi if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) { 172782b0f8c3SJan Kara spin_unlock(vmf->ptl); 1728de466bd6SMel Gorman goto out; 1729de466bd6SMel Gorman } 1730de466bd6SMel Gorman 1731c5b5a3ddSYang Shi pmd = pmd_modify(oldpmd, vma->vm_page_prot); 17326a56ccbcSDavid Hildenbrand 17336a56ccbcSDavid Hildenbrand /* 17346a56ccbcSDavid Hildenbrand * Detect now whether the PMD could be writable; this information 17356a56ccbcSDavid Hildenbrand * is only valid while holding the PT lock. 17366a56ccbcSDavid Hildenbrand */ 17376a56ccbcSDavid Hildenbrand writable = pmd_write(pmd); 17386a56ccbcSDavid Hildenbrand if (!writable && vma_wants_manual_pte_write_upgrade(vma) && 17396a56ccbcSDavid Hildenbrand can_change_pmd_writable(vma, vmf->address, pmd)) 17406a56ccbcSDavid Hildenbrand writable = true; 17416a56ccbcSDavid Hildenbrand 1742667ffc31SKefeng Wang folio = vm_normal_folio_pmd(vma, haddr, pmd); 1743667ffc31SKefeng Wang if (!folio) 1744c5b5a3ddSYang Shi goto out_map; 1745c5b5a3ddSYang Shi 1746c5b5a3ddSYang Shi /* See similar comment in do_numa_page for explanation */ 17476a56ccbcSDavid Hildenbrand if (!writable) 1748c5b5a3ddSYang Shi flags |= TNF_NO_GROUP; 1749c5b5a3ddSYang Shi 1750667ffc31SKefeng Wang nid = folio_nid(folio); 175133024536SHuang Ying /* 175233024536SHuang Ying * For memory tiering mode, cpupid of slow memory page is used 175333024536SHuang Ying * to record page access time. So use default value. 175433024536SHuang Ying */ 1755667ffc31SKefeng Wang if (node_is_toptier(nid)) 1756c4a8d2faSKefeng Wang last_cpupid = folio_last_cpupid(folio); 1757cda6d936SKefeng Wang target_nid = numa_migrate_prep(folio, vma, haddr, nid, &flags); 1758c5b5a3ddSYang Shi if (target_nid == NUMA_NO_NODE) { 1759667ffc31SKefeng Wang folio_put(folio); 1760c5b5a3ddSYang Shi goto out_map; 1761c5b5a3ddSYang Shi } 1762c5b5a3ddSYang Shi 176382b0f8c3SJan Kara spin_unlock(vmf->ptl); 17646a56ccbcSDavid Hildenbrand writable = false; 17658b1b436dSPeter Zijlstra 1766667ffc31SKefeng Wang migrated = migrate_misplaced_folio(folio, vma, target_nid); 17676688cc05SPeter Zijlstra if (migrated) { 17686688cc05SPeter Zijlstra flags |= TNF_MIGRATED; 1769667ffc31SKefeng Wang nid = target_nid; 1770c5b5a3ddSYang Shi } else { 1771074c2381SMel Gorman flags |= TNF_MIGRATE_FAIL; 1772c5b5a3ddSYang Shi vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1773c5b5a3ddSYang Shi if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) { 177482b0f8c3SJan Kara spin_unlock(vmf->ptl); 1775c5b5a3ddSYang Shi goto out; 1776c5b5a3ddSYang Shi } 1777c5b5a3ddSYang Shi goto out_map; 1778c5b5a3ddSYang Shi } 1779b8916634SMel Gorman 1780b8916634SMel Gorman out: 1781667ffc31SKefeng Wang if (nid != NUMA_NO_NODE) 1782667ffc31SKefeng Wang task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags); 17838191acbdSMel Gorman 1784d10e63f2SMel Gorman return 0; 1785c5b5a3ddSYang Shi 1786c5b5a3ddSYang Shi out_map: 1787c5b5a3ddSYang Shi /* Restore the PMD */ 1788c5b5a3ddSYang Shi pmd = pmd_modify(oldpmd, vma->vm_page_prot); 1789c5b5a3ddSYang Shi pmd = pmd_mkyoung(pmd); 17906a56ccbcSDavid Hildenbrand if (writable) 1791161e393cSRick Edgecombe pmd = pmd_mkwrite(pmd, vma); 1792c5b5a3ddSYang Shi set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); 1793c5b5a3ddSYang Shi update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1794c5b5a3ddSYang Shi spin_unlock(vmf->ptl); 1795c5b5a3ddSYang Shi goto out; 1796d10e63f2SMel Gorman } 1797d10e63f2SMel Gorman 1798319904adSHuang Ying /* 1799319904adSHuang Ying * Return true if we do MADV_FREE successfully on entire pmd page. 1800319904adSHuang Ying * Otherwise, return false. 1801319904adSHuang Ying */ 1802319904adSHuang Ying bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1803b8d3c4c3SMinchan Kim pmd_t *pmd, unsigned long addr, unsigned long next) 1804b8d3c4c3SMinchan Kim { 1805b8d3c4c3SMinchan Kim spinlock_t *ptl; 1806b8d3c4c3SMinchan Kim pmd_t orig_pmd; 1807fc986a38SKefeng Wang struct folio *folio; 1808b8d3c4c3SMinchan Kim struct mm_struct *mm = tlb->mm; 1809319904adSHuang Ying bool ret = false; 1810b8d3c4c3SMinchan Kim 1811ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 181207e32661SAneesh Kumar K.V 1813b6ec57f4SKirill A. Shutemov ptl = pmd_trans_huge_lock(pmd, vma); 1814b6ec57f4SKirill A. Shutemov if (!ptl) 181525eedabeSLinus Torvalds goto out_unlocked; 1816b8d3c4c3SMinchan Kim 1817b8d3c4c3SMinchan Kim orig_pmd = *pmd; 1818319904adSHuang Ying if (is_huge_zero_pmd(orig_pmd)) 1819b8d3c4c3SMinchan Kim goto out; 1820b8d3c4c3SMinchan Kim 182184c3fc4eSZi Yan if (unlikely(!pmd_present(orig_pmd))) { 182284c3fc4eSZi Yan VM_BUG_ON(thp_migration_supported() && 182384c3fc4eSZi Yan !is_pmd_migration_entry(orig_pmd)); 182484c3fc4eSZi Yan goto out; 182584c3fc4eSZi Yan } 182684c3fc4eSZi Yan 1827fc986a38SKefeng Wang folio = pfn_folio(pmd_pfn(orig_pmd)); 1828b8d3c4c3SMinchan Kim /* 1829fc986a38SKefeng Wang * If other processes are mapping this folio, we couldn't discard 1830fc986a38SKefeng Wang * the folio unless they all do MADV_FREE so let's skip the folio. 1831b8d3c4c3SMinchan Kim */ 183220b18aadSYin Fengwei if (folio_estimated_sharers(folio) != 1) 1833b8d3c4c3SMinchan Kim goto out; 1834b8d3c4c3SMinchan Kim 1835fc986a38SKefeng Wang if (!folio_trylock(folio)) 1836b8d3c4c3SMinchan Kim goto out; 1837b8d3c4c3SMinchan Kim 1838b8d3c4c3SMinchan Kim /* 1839b8d3c4c3SMinchan Kim * If user want to discard part-pages of THP, split it so MADV_FREE 1840b8d3c4c3SMinchan Kim * will deactivate only them. 1841b8d3c4c3SMinchan Kim */ 1842b8d3c4c3SMinchan Kim if (next - addr != HPAGE_PMD_SIZE) { 1843fc986a38SKefeng Wang folio_get(folio); 1844b8d3c4c3SMinchan Kim spin_unlock(ptl); 1845fc986a38SKefeng Wang split_folio(folio); 1846fc986a38SKefeng Wang folio_unlock(folio); 1847fc986a38SKefeng Wang folio_put(folio); 1848b8d3c4c3SMinchan Kim goto out_unlocked; 1849b8d3c4c3SMinchan Kim } 1850b8d3c4c3SMinchan Kim 1851fc986a38SKefeng Wang if (folio_test_dirty(folio)) 1852fc986a38SKefeng Wang folio_clear_dirty(folio); 1853fc986a38SKefeng Wang folio_unlock(folio); 1854b8d3c4c3SMinchan Kim 1855b8d3c4c3SMinchan Kim if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { 185658ceeb6bSKirill A. Shutemov pmdp_invalidate(vma, addr, pmd); 1857b8d3c4c3SMinchan Kim orig_pmd = pmd_mkold(orig_pmd); 1858b8d3c4c3SMinchan Kim orig_pmd = pmd_mkclean(orig_pmd); 1859b8d3c4c3SMinchan Kim 1860b8d3c4c3SMinchan Kim set_pmd_at(mm, addr, pmd, orig_pmd); 1861b8d3c4c3SMinchan Kim tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1862b8d3c4c3SMinchan Kim } 1863802a3a92SShaohua Li 18646a6fe9ebSKefeng Wang folio_mark_lazyfree(folio); 1865319904adSHuang Ying ret = true; 1866b8d3c4c3SMinchan Kim out: 1867b8d3c4c3SMinchan Kim spin_unlock(ptl); 1868b8d3c4c3SMinchan Kim out_unlocked: 1869b8d3c4c3SMinchan Kim return ret; 1870b8d3c4c3SMinchan Kim } 1871b8d3c4c3SMinchan Kim 1872953c66c2SAneesh Kumar K.V static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) 1873953c66c2SAneesh Kumar K.V { 1874953c66c2SAneesh Kumar K.V pgtable_t pgtable; 1875953c66c2SAneesh Kumar K.V 1876953c66c2SAneesh Kumar K.V pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1877953c66c2SAneesh Kumar K.V pte_free(mm, pgtable); 1878c4812909SKirill A. Shutemov mm_dec_nr_ptes(mm); 1879953c66c2SAneesh Kumar K.V } 1880953c66c2SAneesh Kumar K.V 188171e3aac0SAndrea Arcangeli int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1882f21760b1SShaohua Li pmd_t *pmd, unsigned long addr) 188371e3aac0SAndrea Arcangeli { 1884f5c8ad47SDavid Miller pmd_t orig_pmd; 1885da146769SKirill A. Shutemov spinlock_t *ptl; 1886da146769SKirill A. Shutemov 1887ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 188807e32661SAneesh Kumar K.V 1889b6ec57f4SKirill A. Shutemov ptl = __pmd_trans_huge_lock(pmd, vma); 1890b6ec57f4SKirill A. Shutemov if (!ptl) 1891da146769SKirill A. Shutemov return 0; 1892a6bf2bb0SAneesh Kumar K.V /* 1893a6bf2bb0SAneesh Kumar K.V * For architectures like ppc64 we look at deposited pgtable 18948809aa2dSAneesh Kumar K.V * when calling pmdp_huge_get_and_clear. So do the 1895a6bf2bb0SAneesh Kumar K.V * pgtable_trans_huge_withdraw after finishing pmdp related 1896a6bf2bb0SAneesh Kumar K.V * operations. 1897a6bf2bb0SAneesh Kumar K.V */ 189893a98695SAneesh Kumar K.V orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd, 1899fcbe08d6SMartin Schwidefsky tlb->fullmm); 1900e5136e87SRick Edgecombe arch_check_zapped_pmd(vma, orig_pmd); 1901f21760b1SShaohua Li tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 19022484ca9bSThomas Hellstrom (VMware) if (vma_is_special_huge(vma)) { 19033b6521f5SOliver O'Halloran if (arch_needs_pgtable_deposit()) 19043b6521f5SOliver O'Halloran zap_deposited_table(tlb->mm, pmd); 19054897c765SMatthew Wilcox spin_unlock(ptl); 1906da146769SKirill A. Shutemov } else if (is_huge_zero_pmd(orig_pmd)) { 1907c14a6eb4SOliver O'Halloran zap_deposited_table(tlb->mm, pmd); 1908bf929152SKirill A. Shutemov spin_unlock(ptl); 1909479f0abbSKirill A. Shutemov } else { 19100103b27aSKefeng Wang struct folio *folio = NULL; 1911616b8371SZi Yan int flush_needed = 1; 1912616b8371SZi Yan 1913616b8371SZi Yan if (pmd_present(orig_pmd)) { 19140103b27aSKefeng Wang struct page *page = pmd_page(orig_pmd); 19150103b27aSKefeng Wang 19160103b27aSKefeng Wang folio = page_folio(page); 19170103b27aSKefeng Wang folio_remove_rmap_pmd(folio, page, vma); 1918309381feSSasha Levin VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 1919309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 1920616b8371SZi Yan } else if (thp_migration_supported()) { 1921616b8371SZi Yan swp_entry_t entry; 1922616b8371SZi Yan 1923616b8371SZi Yan VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); 1924616b8371SZi Yan entry = pmd_to_swp_entry(orig_pmd); 19250103b27aSKefeng Wang folio = pfn_swap_entry_folio(entry); 1926616b8371SZi Yan flush_needed = 0; 1927616b8371SZi Yan } else 1928616b8371SZi Yan WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); 1929616b8371SZi Yan 19300103b27aSKefeng Wang if (folio_test_anon(folio)) { 1931c14a6eb4SOliver O'Halloran zap_deposited_table(tlb->mm, pmd); 1932b5072380SKirill A. Shutemov add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1933b5072380SKirill A. Shutemov } else { 1934953c66c2SAneesh Kumar K.V if (arch_needs_pgtable_deposit()) 1935953c66c2SAneesh Kumar K.V zap_deposited_table(tlb->mm, pmd); 19366b27cc6cSKefeng Wang add_mm_counter(tlb->mm, mm_counter_file(folio), 19370103b27aSKefeng Wang -HPAGE_PMD_NR); 1938b5072380SKirill A. Shutemov } 1939616b8371SZi Yan 1940bf929152SKirill A. Shutemov spin_unlock(ptl); 1941616b8371SZi Yan if (flush_needed) 19420103b27aSKefeng Wang tlb_remove_page_size(tlb, &folio->page, HPAGE_PMD_SIZE); 1943479f0abbSKirill A. Shutemov } 1944da146769SKirill A. Shutemov return 1; 194571e3aac0SAndrea Arcangeli } 194671e3aac0SAndrea Arcangeli 19471dd38b6cSAneesh Kumar K.V #ifndef pmd_move_must_withdraw 19481dd38b6cSAneesh Kumar K.V static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, 19491dd38b6cSAneesh Kumar K.V spinlock_t *old_pmd_ptl, 19501dd38b6cSAneesh Kumar K.V struct vm_area_struct *vma) 19511dd38b6cSAneesh Kumar K.V { 19521dd38b6cSAneesh Kumar K.V /* 19531dd38b6cSAneesh Kumar K.V * With split pmd lock we also need to move preallocated 19541dd38b6cSAneesh Kumar K.V * PTE page table if new_pmd is on different PMD page table. 19551dd38b6cSAneesh Kumar K.V * 19561dd38b6cSAneesh Kumar K.V * We also don't deposit and withdraw tables for file pages. 19571dd38b6cSAneesh Kumar K.V */ 19581dd38b6cSAneesh Kumar K.V return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); 19591dd38b6cSAneesh Kumar K.V } 19601dd38b6cSAneesh Kumar K.V #endif 19611dd38b6cSAneesh Kumar K.V 1962ab6e3d09SNaoya Horiguchi static pmd_t move_soft_dirty_pmd(pmd_t pmd) 1963ab6e3d09SNaoya Horiguchi { 1964ab6e3d09SNaoya Horiguchi #ifdef CONFIG_MEM_SOFT_DIRTY 1965ab6e3d09SNaoya Horiguchi if (unlikely(is_pmd_migration_entry(pmd))) 1966ab6e3d09SNaoya Horiguchi pmd = pmd_swp_mksoft_dirty(pmd); 1967ab6e3d09SNaoya Horiguchi else if (pmd_present(pmd)) 1968ab6e3d09SNaoya Horiguchi pmd = pmd_mksoft_dirty(pmd); 1969ab6e3d09SNaoya Horiguchi #endif 1970ab6e3d09SNaoya Horiguchi return pmd; 1971ab6e3d09SNaoya Horiguchi } 1972ab6e3d09SNaoya Horiguchi 1973bf8616d5SHugh Dickins bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, 1974b8aa9d9dSWei Yang unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd) 197537a1c49aSAndrea Arcangeli { 1976bf929152SKirill A. Shutemov spinlock_t *old_ptl, *new_ptl; 197737a1c49aSAndrea Arcangeli pmd_t pmd; 197837a1c49aSAndrea Arcangeli struct mm_struct *mm = vma->vm_mm; 19795d190420SAaron Lu bool force_flush = false; 198037a1c49aSAndrea Arcangeli 198137a1c49aSAndrea Arcangeli /* 198237a1c49aSAndrea Arcangeli * The destination pmd shouldn't be established, free_pgtables() 1983a5be621eSHugh Dickins * should have released it; but move_page_tables() might have already 1984a5be621eSHugh Dickins * inserted a page table, if racing against shmem/file collapse. 198537a1c49aSAndrea Arcangeli */ 1986a5be621eSHugh Dickins if (!pmd_none(*new_pmd)) { 198737a1c49aSAndrea Arcangeli VM_BUG_ON(pmd_trans_huge(*new_pmd)); 19884b471e88SKirill A. Shutemov return false; 198937a1c49aSAndrea Arcangeli } 199037a1c49aSAndrea Arcangeli 1991bf929152SKirill A. Shutemov /* 1992bf929152SKirill A. Shutemov * We don't have to worry about the ordering of src and dst 1993c1e8d7c6SMichel Lespinasse * ptlocks because exclusive mmap_lock prevents deadlock. 1994bf929152SKirill A. Shutemov */ 1995b6ec57f4SKirill A. Shutemov old_ptl = __pmd_trans_huge_lock(old_pmd, vma); 1996b6ec57f4SKirill A. Shutemov if (old_ptl) { 1997bf929152SKirill A. Shutemov new_ptl = pmd_lockptr(mm, new_pmd); 1998bf929152SKirill A. Shutemov if (new_ptl != old_ptl) 1999bf929152SKirill A. Shutemov spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 20008809aa2dSAneesh Kumar K.V pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 2001eb66ae03SLinus Torvalds if (pmd_present(pmd)) 2002a2ce2666SAaron Lu force_flush = true; 200337a1c49aSAndrea Arcangeli VM_BUG_ON(!pmd_none(*new_pmd)); 20043592806cSKirill A. Shutemov 20051dd38b6cSAneesh Kumar K.V if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { 2006b3084f4dSAneesh Kumar K.V pgtable_t pgtable; 20073592806cSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 20083592806cSKirill A. Shutemov pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 20093592806cSKirill A. Shutemov } 2010ab6e3d09SNaoya Horiguchi pmd = move_soft_dirty_pmd(pmd); 2011ab6e3d09SNaoya Horiguchi set_pmd_at(mm, new_addr, new_pmd, pmd); 20125d190420SAaron Lu if (force_flush) 20137c38f181SMiaohe Lin flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE); 2014eb66ae03SLinus Torvalds if (new_ptl != old_ptl) 2015eb66ae03SLinus Torvalds spin_unlock(new_ptl); 2016bf929152SKirill A. Shutemov spin_unlock(old_ptl); 20174b471e88SKirill A. Shutemov return true; 201837a1c49aSAndrea Arcangeli } 20194b471e88SKirill A. Shutemov return false; 202037a1c49aSAndrea Arcangeli } 202137a1c49aSAndrea Arcangeli 2022f123d74aSMel Gorman /* 2023f123d74aSMel Gorman * Returns 2024f123d74aSMel Gorman * - 0 if PMD could not be locked 2025f0953a1bSIngo Molnar * - 1 if PMD was locked but protections unchanged and TLB flush unnecessary 2026e346e668SYang Shi * or if prot_numa but THP migration is not supported 2027f0953a1bSIngo Molnar * - HPAGE_PMD_NR if protections changed and TLB flush necessary 2028f123d74aSMel Gorman */ 20294a18419fSNadav Amit int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 20304a18419fSNadav Amit pmd_t *pmd, unsigned long addr, pgprot_t newprot, 20314a18419fSNadav Amit unsigned long cp_flags) 2032cd7548abSJohannes Weiner { 2033cd7548abSJohannes Weiner struct mm_struct *mm = vma->vm_mm; 2034bf929152SKirill A. Shutemov spinlock_t *ptl; 2035c9fe6656SNadav Amit pmd_t oldpmd, entry; 203658705444SPeter Xu bool prot_numa = cp_flags & MM_CP_PROT_NUMA; 2037292924b2SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 2038292924b2SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 20396a56ccbcSDavid Hildenbrand int ret = 1; 2040cd7548abSJohannes Weiner 20414a18419fSNadav Amit tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 20424a18419fSNadav Amit 2043e346e668SYang Shi if (prot_numa && !thp_migration_supported()) 2044e346e668SYang Shi return 1; 2045e346e668SYang Shi 2046b6ec57f4SKirill A. Shutemov ptl = __pmd_trans_huge_lock(pmd, vma); 20470a85e51dSKirill A. Shutemov if (!ptl) 20480a85e51dSKirill A. Shutemov return 0; 20490a85e51dSKirill A. Shutemov 205084c3fc4eSZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 205184c3fc4eSZi Yan if (is_swap_pmd(*pmd)) { 205284c3fc4eSZi Yan swp_entry_t entry = pmd_to_swp_entry(*pmd); 20535662400aSMatthew Wilcox (Oracle) struct folio *folio = pfn_swap_entry_folio(entry); 205424bf08c4SDavid Hildenbrand pmd_t newpmd; 205584c3fc4eSZi Yan 205684c3fc4eSZi Yan VM_BUG_ON(!is_pmd_migration_entry(*pmd)); 20574dd845b5SAlistair Popple if (is_writable_migration_entry(entry)) { 205884c3fc4eSZi Yan /* 205984c3fc4eSZi Yan * A protection check is difficult so 206084c3fc4eSZi Yan * just be safe and disable write 206184c3fc4eSZi Yan */ 2062d986ba2bSKefeng Wang if (folio_test_anon(folio)) 20636c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry(swp_offset(entry)); 20646c287605SDavid Hildenbrand else 20656c287605SDavid Hildenbrand entry = make_readable_migration_entry(swp_offset(entry)); 206684c3fc4eSZi Yan newpmd = swp_entry_to_pmd(entry); 2067ab6e3d09SNaoya Horiguchi if (pmd_swp_soft_dirty(*pmd)) 2068ab6e3d09SNaoya Horiguchi newpmd = pmd_swp_mksoft_dirty(newpmd); 206924bf08c4SDavid Hildenbrand } else { 207024bf08c4SDavid Hildenbrand newpmd = *pmd; 207184c3fc4eSZi Yan } 207224bf08c4SDavid Hildenbrand 207324bf08c4SDavid Hildenbrand if (uffd_wp) 207424bf08c4SDavid Hildenbrand newpmd = pmd_swp_mkuffd_wp(newpmd); 207524bf08c4SDavid Hildenbrand else if (uffd_wp_resolve) 207624bf08c4SDavid Hildenbrand newpmd = pmd_swp_clear_uffd_wp(newpmd); 207724bf08c4SDavid Hildenbrand if (!pmd_same(*pmd, newpmd)) 207824bf08c4SDavid Hildenbrand set_pmd_at(mm, addr, pmd, newpmd); 207984c3fc4eSZi Yan goto unlock; 208084c3fc4eSZi Yan } 208184c3fc4eSZi Yan #endif 208284c3fc4eSZi Yan 2083a1a3a2fcSHuang Ying if (prot_numa) { 2084d986ba2bSKefeng Wang struct folio *folio; 208533024536SHuang Ying bool toptier; 2086e944fd67SMel Gorman /* 2087e944fd67SMel Gorman * Avoid trapping faults against the zero page. The read-only 2088e944fd67SMel Gorman * data is likely to be read-cached on the local CPU and 2089e944fd67SMel Gorman * local/remote hits to the zero page are not interesting. 2090e944fd67SMel Gorman */ 2091a1a3a2fcSHuang Ying if (is_huge_zero_pmd(*pmd)) 20920a85e51dSKirill A. Shutemov goto unlock; 2093e944fd67SMel Gorman 2094a1a3a2fcSHuang Ying if (pmd_protnone(*pmd)) 20950a85e51dSKirill A. Shutemov goto unlock; 20960a85e51dSKirill A. Shutemov 2097d986ba2bSKefeng Wang folio = page_folio(pmd_page(*pmd)); 2098d986ba2bSKefeng Wang toptier = node_is_toptier(folio_nid(folio)); 2099a1a3a2fcSHuang Ying /* 2100a1a3a2fcSHuang Ying * Skip scanning top tier node if normal numa 2101a1a3a2fcSHuang Ying * balancing is disabled 2102a1a3a2fcSHuang Ying */ 2103a1a3a2fcSHuang Ying if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) && 210433024536SHuang Ying toptier) 2105a1a3a2fcSHuang Ying goto unlock; 210633024536SHuang Ying 210733024536SHuang Ying if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING && 210833024536SHuang Ying !toptier) 2109d986ba2bSKefeng Wang folio_xchg_access_time(folio, 2110d986ba2bSKefeng Wang jiffies_to_msecs(jiffies)); 2111a1a3a2fcSHuang Ying } 2112ced10803SKirill A. Shutemov /* 21133e4e28c5SMichel Lespinasse * In case prot_numa, we are under mmap_read_lock(mm). It's critical 2114ced10803SKirill A. Shutemov * to not clear pmd intermittently to avoid race with MADV_DONTNEED 21153e4e28c5SMichel Lespinasse * which is also under mmap_read_lock(mm): 2116ced10803SKirill A. Shutemov * 2117ced10803SKirill A. Shutemov * CPU0: CPU1: 2118ced10803SKirill A. Shutemov * change_huge_pmd(prot_numa=1) 2119ced10803SKirill A. Shutemov * pmdp_huge_get_and_clear_notify() 2120ced10803SKirill A. Shutemov * madvise_dontneed() 2121ced10803SKirill A. Shutemov * zap_pmd_range() 2122ced10803SKirill A. Shutemov * pmd_trans_huge(*pmd) == 0 (without ptl) 2123ced10803SKirill A. Shutemov * // skip the pmd 2124ced10803SKirill A. Shutemov * set_pmd_at(); 2125ced10803SKirill A. Shutemov * // pmd is re-established 2126ced10803SKirill A. Shutemov * 2127ced10803SKirill A. Shutemov * The race makes MADV_DONTNEED miss the huge pmd and don't clear it 2128ced10803SKirill A. Shutemov * which may break userspace. 2129ced10803SKirill A. Shutemov * 21304f831457SNadav Amit * pmdp_invalidate_ad() is required to make sure we don't miss 2131ced10803SKirill A. Shutemov * dirty/young flags set by hardware. 2132ced10803SKirill A. Shutemov */ 21334f831457SNadav Amit oldpmd = pmdp_invalidate_ad(vma, addr, pmd); 2134ced10803SKirill A. Shutemov 2135c9fe6656SNadav Amit entry = pmd_modify(oldpmd, newprot); 2136f1eb1bacSPeter Xu if (uffd_wp) 2137292924b2SPeter Xu entry = pmd_mkuffd_wp(entry); 2138f1eb1bacSPeter Xu else if (uffd_wp_resolve) 2139292924b2SPeter Xu /* 2140292924b2SPeter Xu * Leave the write bit to be handled by PF interrupt 2141292924b2SPeter Xu * handler, then things like COW could be properly 2142292924b2SPeter Xu * handled. 2143292924b2SPeter Xu */ 2144292924b2SPeter Xu entry = pmd_clear_uffd_wp(entry); 2145c27f479eSDavid Hildenbrand 2146c27f479eSDavid Hildenbrand /* See change_pte_range(). */ 2147c27f479eSDavid Hildenbrand if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) && 2148c27f479eSDavid Hildenbrand can_change_pmd_writable(vma, addr, entry)) 2149161e393cSRick Edgecombe entry = pmd_mkwrite(entry, vma); 2150c27f479eSDavid Hildenbrand 2151f123d74aSMel Gorman ret = HPAGE_PMD_NR; 215256eecdb9SAneesh Kumar K.V set_pmd_at(mm, addr, pmd, entry); 21534a18419fSNadav Amit 2154c9fe6656SNadav Amit if (huge_pmd_needs_flush(oldpmd, entry)) 21554a18419fSNadav Amit tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE); 21560a85e51dSKirill A. Shutemov unlock: 2157bf929152SKirill A. Shutemov spin_unlock(ptl); 2158cd7548abSJohannes Weiner return ret; 2159cd7548abSJohannes Weiner } 2160cd7548abSJohannes Weiner 2161adef4406SAndrea Arcangeli #ifdef CONFIG_USERFAULTFD 2162adef4406SAndrea Arcangeli /* 2163867a43a3SLokesh Gidra * The PT lock for src_pmd and dst_vma/src_vma (for reading) are locked by 2164adef4406SAndrea Arcangeli * the caller, but it must return after releasing the page_table_lock. 2165adef4406SAndrea Arcangeli * Just move the page from src_pmd to dst_pmd if possible. 2166adef4406SAndrea Arcangeli * Return zero if succeeded in moving the page, -EAGAIN if it needs to be 2167adef4406SAndrea Arcangeli * repeated by the caller, or other errors in case of failure. 2168adef4406SAndrea Arcangeli */ 2169adef4406SAndrea Arcangeli int move_pages_huge_pmd(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd, pmd_t dst_pmdval, 2170adef4406SAndrea Arcangeli struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 2171adef4406SAndrea Arcangeli unsigned long dst_addr, unsigned long src_addr) 2172adef4406SAndrea Arcangeli { 2173adef4406SAndrea Arcangeli pmd_t _dst_pmd, src_pmdval; 2174adef4406SAndrea Arcangeli struct page *src_page; 2175adef4406SAndrea Arcangeli struct folio *src_folio; 2176adef4406SAndrea Arcangeli struct anon_vma *src_anon_vma; 2177adef4406SAndrea Arcangeli spinlock_t *src_ptl, *dst_ptl; 2178adef4406SAndrea Arcangeli pgtable_t src_pgtable; 2179adef4406SAndrea Arcangeli struct mmu_notifier_range range; 2180adef4406SAndrea Arcangeli int err = 0; 2181adef4406SAndrea Arcangeli 2182adef4406SAndrea Arcangeli src_pmdval = *src_pmd; 2183adef4406SAndrea Arcangeli src_ptl = pmd_lockptr(mm, src_pmd); 2184adef4406SAndrea Arcangeli 2185adef4406SAndrea Arcangeli lockdep_assert_held(src_ptl); 2186867a43a3SLokesh Gidra vma_assert_locked(src_vma); 2187867a43a3SLokesh Gidra vma_assert_locked(dst_vma); 2188adef4406SAndrea Arcangeli 2189adef4406SAndrea Arcangeli /* Sanity checks before the operation */ 2190adef4406SAndrea Arcangeli if (WARN_ON_ONCE(!pmd_none(dst_pmdval)) || WARN_ON_ONCE(src_addr & ~HPAGE_PMD_MASK) || 2191adef4406SAndrea Arcangeli WARN_ON_ONCE(dst_addr & ~HPAGE_PMD_MASK)) { 2192adef4406SAndrea Arcangeli spin_unlock(src_ptl); 2193adef4406SAndrea Arcangeli return -EINVAL; 2194adef4406SAndrea Arcangeli } 2195adef4406SAndrea Arcangeli 2196adef4406SAndrea Arcangeli if (!pmd_trans_huge(src_pmdval)) { 2197adef4406SAndrea Arcangeli spin_unlock(src_ptl); 2198adef4406SAndrea Arcangeli if (is_pmd_migration_entry(src_pmdval)) { 2199adef4406SAndrea Arcangeli pmd_migration_entry_wait(mm, &src_pmdval); 2200adef4406SAndrea Arcangeli return -EAGAIN; 2201adef4406SAndrea Arcangeli } 2202adef4406SAndrea Arcangeli return -ENOENT; 2203adef4406SAndrea Arcangeli } 2204adef4406SAndrea Arcangeli 2205adef4406SAndrea Arcangeli src_page = pmd_page(src_pmdval); 2206eb1521daSSuren Baghdasaryan 2207eb1521daSSuren Baghdasaryan if (!is_huge_zero_pmd(src_pmdval)) { 2208adef4406SAndrea Arcangeli if (unlikely(!PageAnonExclusive(src_page))) { 2209adef4406SAndrea Arcangeli spin_unlock(src_ptl); 2210adef4406SAndrea Arcangeli return -EBUSY; 2211adef4406SAndrea Arcangeli } 2212adef4406SAndrea Arcangeli 2213adef4406SAndrea Arcangeli src_folio = page_folio(src_page); 2214adef4406SAndrea Arcangeli folio_get(src_folio); 2215eb1521daSSuren Baghdasaryan } else 2216eb1521daSSuren Baghdasaryan src_folio = NULL; 2217eb1521daSSuren Baghdasaryan 2218adef4406SAndrea Arcangeli spin_unlock(src_ptl); 2219adef4406SAndrea Arcangeli 2220adef4406SAndrea Arcangeli flush_cache_range(src_vma, src_addr, src_addr + HPAGE_PMD_SIZE); 2221adef4406SAndrea Arcangeli mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, src_addr, 2222adef4406SAndrea Arcangeli src_addr + HPAGE_PMD_SIZE); 2223adef4406SAndrea Arcangeli mmu_notifier_invalidate_range_start(&range); 2224adef4406SAndrea Arcangeli 2225eb1521daSSuren Baghdasaryan if (src_folio) { 2226adef4406SAndrea Arcangeli folio_lock(src_folio); 2227adef4406SAndrea Arcangeli 2228adef4406SAndrea Arcangeli /* 2229adef4406SAndrea Arcangeli * split_huge_page walks the anon_vma chain without the page 2230adef4406SAndrea Arcangeli * lock. Serialize against it with the anon_vma lock, the page 2231adef4406SAndrea Arcangeli * lock is not enough. 2232adef4406SAndrea Arcangeli */ 2233adef4406SAndrea Arcangeli src_anon_vma = folio_get_anon_vma(src_folio); 2234adef4406SAndrea Arcangeli if (!src_anon_vma) { 2235adef4406SAndrea Arcangeli err = -EAGAIN; 2236adef4406SAndrea Arcangeli goto unlock_folio; 2237adef4406SAndrea Arcangeli } 2238adef4406SAndrea Arcangeli anon_vma_lock_write(src_anon_vma); 2239eb1521daSSuren Baghdasaryan } else 2240eb1521daSSuren Baghdasaryan src_anon_vma = NULL; 2241adef4406SAndrea Arcangeli 2242adef4406SAndrea Arcangeli dst_ptl = pmd_lockptr(mm, dst_pmd); 2243adef4406SAndrea Arcangeli double_pt_lock(src_ptl, dst_ptl); 2244adef4406SAndrea Arcangeli if (unlikely(!pmd_same(*src_pmd, src_pmdval) || 2245adef4406SAndrea Arcangeli !pmd_same(*dst_pmd, dst_pmdval))) { 2246adef4406SAndrea Arcangeli err = -EAGAIN; 2247adef4406SAndrea Arcangeli goto unlock_ptls; 2248adef4406SAndrea Arcangeli } 2249eb1521daSSuren Baghdasaryan if (src_folio) { 2250adef4406SAndrea Arcangeli if (folio_maybe_dma_pinned(src_folio) || 2251adef4406SAndrea Arcangeli !PageAnonExclusive(&src_folio->page)) { 2252adef4406SAndrea Arcangeli err = -EBUSY; 2253adef4406SAndrea Arcangeli goto unlock_ptls; 2254adef4406SAndrea Arcangeli } 2255adef4406SAndrea Arcangeli 2256adef4406SAndrea Arcangeli if (WARN_ON_ONCE(!folio_test_head(src_folio)) || 2257adef4406SAndrea Arcangeli WARN_ON_ONCE(!folio_test_anon(src_folio))) { 2258adef4406SAndrea Arcangeli err = -EBUSY; 2259adef4406SAndrea Arcangeli goto unlock_ptls; 2260adef4406SAndrea Arcangeli } 2261adef4406SAndrea Arcangeli 2262adef4406SAndrea Arcangeli src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd); 2263adef4406SAndrea Arcangeli /* Folio got pinned from under us. Put it back and fail the move. */ 2264adef4406SAndrea Arcangeli if (folio_maybe_dma_pinned(src_folio)) { 2265adef4406SAndrea Arcangeli set_pmd_at(mm, src_addr, src_pmd, src_pmdval); 2266adef4406SAndrea Arcangeli err = -EBUSY; 2267adef4406SAndrea Arcangeli goto unlock_ptls; 2268adef4406SAndrea Arcangeli } 2269adef4406SAndrea Arcangeli 2270*c0205eafSLokesh Gidra folio_move_anon_rmap(src_folio, dst_vma); 2271*c0205eafSLokesh Gidra WRITE_ONCE(src_folio->index, linear_page_index(dst_vma, dst_addr)); 2272*c0205eafSLokesh Gidra 2273adef4406SAndrea Arcangeli _dst_pmd = mk_huge_pmd(&src_folio->page, dst_vma->vm_page_prot); 2274adef4406SAndrea Arcangeli /* Follow mremap() behavior and treat the entry dirty after the move */ 2275adef4406SAndrea Arcangeli _dst_pmd = pmd_mkwrite(pmd_mkdirty(_dst_pmd), dst_vma); 2276eb1521daSSuren Baghdasaryan } else { 2277eb1521daSSuren Baghdasaryan src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd); 2278eb1521daSSuren Baghdasaryan _dst_pmd = mk_huge_pmd(src_page, dst_vma->vm_page_prot); 2279eb1521daSSuren Baghdasaryan } 2280adef4406SAndrea Arcangeli set_pmd_at(mm, dst_addr, dst_pmd, _dst_pmd); 2281adef4406SAndrea Arcangeli 2282adef4406SAndrea Arcangeli src_pgtable = pgtable_trans_huge_withdraw(mm, src_pmd); 2283adef4406SAndrea Arcangeli pgtable_trans_huge_deposit(mm, dst_pmd, src_pgtable); 2284adef4406SAndrea Arcangeli unlock_ptls: 2285adef4406SAndrea Arcangeli double_pt_unlock(src_ptl, dst_ptl); 2286eb1521daSSuren Baghdasaryan if (src_anon_vma) { 2287adef4406SAndrea Arcangeli anon_vma_unlock_write(src_anon_vma); 2288adef4406SAndrea Arcangeli put_anon_vma(src_anon_vma); 2289eb1521daSSuren Baghdasaryan } 2290adef4406SAndrea Arcangeli unlock_folio: 2291adef4406SAndrea Arcangeli /* unblock rmap walks */ 2292eb1521daSSuren Baghdasaryan if (src_folio) 2293adef4406SAndrea Arcangeli folio_unlock(src_folio); 2294adef4406SAndrea Arcangeli mmu_notifier_invalidate_range_end(&range); 2295eb1521daSSuren Baghdasaryan if (src_folio) 2296adef4406SAndrea Arcangeli folio_put(src_folio); 2297adef4406SAndrea Arcangeli return err; 2298adef4406SAndrea Arcangeli } 2299adef4406SAndrea Arcangeli #endif /* CONFIG_USERFAULTFD */ 2300adef4406SAndrea Arcangeli 2301025c5b24SNaoya Horiguchi /* 23028f19b0c0SHuang Ying * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. 2303025c5b24SNaoya Horiguchi * 23048f19b0c0SHuang Ying * Note that if it returns page table lock pointer, this routine returns without 23058f19b0c0SHuang Ying * unlocking page table lock. So callers must unlock it. 2306025c5b24SNaoya Horiguchi */ 2307b6ec57f4SKirill A. Shutemov spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 2308025c5b24SNaoya Horiguchi { 2309b6ec57f4SKirill A. Shutemov spinlock_t *ptl; 2310b6ec57f4SKirill A. Shutemov ptl = pmd_lock(vma->vm_mm, pmd); 231184c3fc4eSZi Yan if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || 231284c3fc4eSZi Yan pmd_devmap(*pmd))) 2313b6ec57f4SKirill A. Shutemov return ptl; 2314b6ec57f4SKirill A. Shutemov spin_unlock(ptl); 2315b6ec57f4SKirill A. Shutemov return NULL; 2316025c5b24SNaoya Horiguchi } 2317025c5b24SNaoya Horiguchi 2318a00cc7d9SMatthew Wilcox /* 2319d965e390SMiaohe Lin * Returns page table lock pointer if a given pud maps a thp, NULL otherwise. 2320a00cc7d9SMatthew Wilcox * 2321d965e390SMiaohe Lin * Note that if it returns page table lock pointer, this routine returns without 2322d965e390SMiaohe Lin * unlocking page table lock. So callers must unlock it. 2323a00cc7d9SMatthew Wilcox */ 2324a00cc7d9SMatthew Wilcox spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) 2325a00cc7d9SMatthew Wilcox { 2326a00cc7d9SMatthew Wilcox spinlock_t *ptl; 2327a00cc7d9SMatthew Wilcox 2328a00cc7d9SMatthew Wilcox ptl = pud_lock(vma->vm_mm, pud); 2329a00cc7d9SMatthew Wilcox if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) 2330a00cc7d9SMatthew Wilcox return ptl; 2331a00cc7d9SMatthew Wilcox spin_unlock(ptl); 2332a00cc7d9SMatthew Wilcox return NULL; 2333a00cc7d9SMatthew Wilcox } 2334a00cc7d9SMatthew Wilcox 2335a00cc7d9SMatthew Wilcox #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 2336a00cc7d9SMatthew Wilcox int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, 2337a00cc7d9SMatthew Wilcox pud_t *pud, unsigned long addr) 2338a00cc7d9SMatthew Wilcox { 2339a00cc7d9SMatthew Wilcox spinlock_t *ptl; 2340a00cc7d9SMatthew Wilcox 2341a00cc7d9SMatthew Wilcox ptl = __pud_trans_huge_lock(pud, vma); 2342a00cc7d9SMatthew Wilcox if (!ptl) 2343a00cc7d9SMatthew Wilcox return 0; 234474929079SMiaohe Lin 2345f32928abSAneesh Kumar K.V pudp_huge_get_and_clear_full(vma, addr, pud, tlb->fullmm); 2346a00cc7d9SMatthew Wilcox tlb_remove_pud_tlb_entry(tlb, pud, addr); 23472484ca9bSThomas Hellstrom (VMware) if (vma_is_special_huge(vma)) { 2348a00cc7d9SMatthew Wilcox spin_unlock(ptl); 2349a00cc7d9SMatthew Wilcox /* No zero page support yet */ 2350a00cc7d9SMatthew Wilcox } else { 2351a00cc7d9SMatthew Wilcox /* No support for anonymous PUD pages yet */ 2352a00cc7d9SMatthew Wilcox BUG(); 2353a00cc7d9SMatthew Wilcox } 2354a00cc7d9SMatthew Wilcox return 1; 2355a00cc7d9SMatthew Wilcox } 2356a00cc7d9SMatthew Wilcox 2357a00cc7d9SMatthew Wilcox static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, 2358a00cc7d9SMatthew Wilcox unsigned long haddr) 2359a00cc7d9SMatthew Wilcox { 2360a00cc7d9SMatthew Wilcox VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); 2361a00cc7d9SMatthew Wilcox VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2362a00cc7d9SMatthew Wilcox VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); 2363a00cc7d9SMatthew Wilcox VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); 2364a00cc7d9SMatthew Wilcox 2365ce9311cfSYisheng Xie count_vm_event(THP_SPLIT_PUD); 2366a00cc7d9SMatthew Wilcox 2367ec8832d0SAlistair Popple pudp_huge_clear_flush(vma, haddr, pud); 2368a00cc7d9SMatthew Wilcox } 2369a00cc7d9SMatthew Wilcox 2370a00cc7d9SMatthew Wilcox void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, 2371a00cc7d9SMatthew Wilcox unsigned long address) 2372a00cc7d9SMatthew Wilcox { 2373a00cc7d9SMatthew Wilcox spinlock_t *ptl; 2374ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 2375a00cc7d9SMatthew Wilcox 23767d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 23776f4f13e8SJérôme Glisse address & HPAGE_PUD_MASK, 2378ac46d4f3SJérôme Glisse (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE); 2379ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 2380ac46d4f3SJérôme Glisse ptl = pud_lock(vma->vm_mm, pud); 2381a00cc7d9SMatthew Wilcox if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) 2382a00cc7d9SMatthew Wilcox goto out; 2383ac46d4f3SJérôme Glisse __split_huge_pud_locked(vma, pud, range.start); 2384a00cc7d9SMatthew Wilcox 2385a00cc7d9SMatthew Wilcox out: 2386a00cc7d9SMatthew Wilcox spin_unlock(ptl); 2387ec8832d0SAlistair Popple mmu_notifier_invalidate_range_end(&range); 2388a00cc7d9SMatthew Wilcox } 2389a00cc7d9SMatthew Wilcox #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 2390a00cc7d9SMatthew Wilcox 2391eef1b3baSKirill A. Shutemov static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 2392eef1b3baSKirill A. Shutemov unsigned long haddr, pmd_t *pmd) 2393eef1b3baSKirill A. Shutemov { 2394eef1b3baSKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 2395eef1b3baSKirill A. Shutemov pgtable_t pgtable; 239642b2af2cSDavid Hildenbrand pmd_t _pmd, old_pmd; 2397c9c1ee20SHugh Dickins unsigned long addr; 2398c9c1ee20SHugh Dickins pte_t *pte; 2399eef1b3baSKirill A. Shutemov int i; 2400eef1b3baSKirill A. Shutemov 24010f10851eSJérôme Glisse /* 24020f10851eSJérôme Glisse * Leave pmd empty until pte is filled note that it is fine to delay 24030f10851eSJérôme Glisse * notification until mmu_notifier_invalidate_range_end() as we are 24040f10851eSJérôme Glisse * replacing a zero pmd write protected page with a zero pte write 24050f10851eSJérôme Glisse * protected page. 24060f10851eSJérôme Glisse * 2407ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 24080f10851eSJérôme Glisse */ 240942b2af2cSDavid Hildenbrand old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd); 2410eef1b3baSKirill A. Shutemov 2411eef1b3baSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2412eef1b3baSKirill A. Shutemov pmd_populate(mm, &_pmd, pgtable); 2413eef1b3baSKirill A. Shutemov 2414c9c1ee20SHugh Dickins pte = pte_offset_map(&_pmd, haddr); 2415c9c1ee20SHugh Dickins VM_BUG_ON(!pte); 2416c9c1ee20SHugh Dickins for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2417c9c1ee20SHugh Dickins pte_t entry; 2418c9c1ee20SHugh Dickins 2419c9c1ee20SHugh Dickins entry = pfn_pte(my_zero_pfn(addr), vma->vm_page_prot); 2420eef1b3baSKirill A. Shutemov entry = pte_mkspecial(entry); 242142b2af2cSDavid Hildenbrand if (pmd_uffd_wp(old_pmd)) 242242b2af2cSDavid Hildenbrand entry = pte_mkuffd_wp(entry); 2423c33c7948SRyan Roberts VM_BUG_ON(!pte_none(ptep_get(pte))); 2424c9c1ee20SHugh Dickins set_pte_at(mm, addr, pte, entry); 2425c9c1ee20SHugh Dickins pte++; 2426eef1b3baSKirill A. Shutemov } 2427c9c1ee20SHugh Dickins pte_unmap(pte - 1); 2428eef1b3baSKirill A. Shutemov smp_wmb(); /* make pte visible before pmd */ 2429eef1b3baSKirill A. Shutemov pmd_populate(mm, pmd, pgtable); 2430eef1b3baSKirill A. Shutemov } 2431eef1b3baSKirill A. Shutemov 2432eef1b3baSKirill A. Shutemov static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, 2433ba988280SKirill A. Shutemov unsigned long haddr, bool freeze) 2434eef1b3baSKirill A. Shutemov { 2435eef1b3baSKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 243691b2978aSDavid Hildenbrand struct folio *folio; 2437eef1b3baSKirill A. Shutemov struct page *page; 2438eef1b3baSKirill A. Shutemov pgtable_t pgtable; 2439423ac9afSAneesh Kumar K.V pmd_t old_pmd, _pmd; 2440292924b2SPeter Xu bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false; 24410ccf7f16SPeter Xu bool anon_exclusive = false, dirty = false; 24422ac015e2SKirill A. Shutemov unsigned long addr; 2443c9c1ee20SHugh Dickins pte_t *pte; 2444eef1b3baSKirill A. Shutemov int i; 2445eef1b3baSKirill A. Shutemov 2446eef1b3baSKirill A. Shutemov VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); 2447eef1b3baSKirill A. Shutemov VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2448eef1b3baSKirill A. Shutemov VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); 244984c3fc4eSZi Yan VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) 245084c3fc4eSZi Yan && !pmd_devmap(*pmd)); 2451eef1b3baSKirill A. Shutemov 2452eef1b3baSKirill A. Shutemov count_vm_event(THP_SPLIT_PMD); 2453eef1b3baSKirill A. Shutemov 2454d21b9e57SKirill A. Shutemov if (!vma_is_anonymous(vma)) { 2455ec8832d0SAlistair Popple old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd); 2456953c66c2SAneesh Kumar K.V /* 2457953c66c2SAneesh Kumar K.V * We are going to unmap this huge page. So 2458953c66c2SAneesh Kumar K.V * just go ahead and zap it 2459953c66c2SAneesh Kumar K.V */ 2460953c66c2SAneesh Kumar K.V if (arch_needs_pgtable_deposit()) 2461953c66c2SAneesh Kumar K.V zap_deposited_table(mm, pmd); 24622484ca9bSThomas Hellstrom (VMware) if (vma_is_special_huge(vma)) 2463d21b9e57SKirill A. Shutemov return; 246499fa8a48SHugh Dickins if (unlikely(is_pmd_migration_entry(old_pmd))) { 246599fa8a48SHugh Dickins swp_entry_t entry; 246699fa8a48SHugh Dickins 246799fa8a48SHugh Dickins entry = pmd_to_swp_entry(old_pmd); 2468439992ffSKefeng Wang folio = pfn_swap_entry_folio(entry); 246999fa8a48SHugh Dickins } else { 247099fa8a48SHugh Dickins page = pmd_page(old_pmd); 2471a8e61d58SDavid Hildenbrand folio = page_folio(page); 2472a8e61d58SDavid Hildenbrand if (!folio_test_dirty(folio) && pmd_dirty(old_pmd)) 2473db44c658SDavid Hildenbrand folio_mark_dirty(folio); 2474a8e61d58SDavid Hildenbrand if (!folio_test_referenced(folio) && pmd_young(old_pmd)) 2475a8e61d58SDavid Hildenbrand folio_set_referenced(folio); 2476a8e61d58SDavid Hildenbrand folio_remove_rmap_pmd(folio, page, vma); 2477a8e61d58SDavid Hildenbrand folio_put(folio); 247899fa8a48SHugh Dickins } 24796b27cc6cSKefeng Wang add_mm_counter(mm, mm_counter_file(folio), -HPAGE_PMD_NR); 2480eef1b3baSKirill A. Shutemov return; 248199fa8a48SHugh Dickins } 248299fa8a48SHugh Dickins 24833b77e8c8SHugh Dickins if (is_huge_zero_pmd(*pmd)) { 24844645b9feSJérôme Glisse /* 24854645b9feSJérôme Glisse * FIXME: Do we want to invalidate secondary mmu by calling 24861af5a810SAlistair Popple * mmu_notifier_arch_invalidate_secondary_tlbs() see comments below 24871af5a810SAlistair Popple * inside __split_huge_pmd() ? 24884645b9feSJérôme Glisse * 24894645b9feSJérôme Glisse * We are going from a zero huge page write protected to zero 24904645b9feSJérôme Glisse * small page also write protected so it does not seems useful 24914645b9feSJérôme Glisse * to invalidate secondary mmu at this time. 24924645b9feSJérôme Glisse */ 2493eef1b3baSKirill A. Shutemov return __split_huge_zero_page_pmd(vma, haddr, pmd); 2494eef1b3baSKirill A. Shutemov } 2495eef1b3baSKirill A. Shutemov 2496423ac9afSAneesh Kumar K.V /* 2497423ac9afSAneesh Kumar K.V * Up to this point the pmd is present and huge and userland has the 2498423ac9afSAneesh Kumar K.V * whole access to the hugepage during the split (which happens in 2499423ac9afSAneesh Kumar K.V * place). If we overwrite the pmd with the not-huge version pointing 2500423ac9afSAneesh Kumar K.V * to the pte here (which of course we could if all CPUs were bug 2501423ac9afSAneesh Kumar K.V * free), userland could trigger a small page size TLB miss on the 2502423ac9afSAneesh Kumar K.V * small sized TLB while the hugepage TLB entry is still established in 2503423ac9afSAneesh Kumar K.V * the huge TLB. Some CPU doesn't like that. 250442742d9bSAlexander A. Klimov * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum 250542742d9bSAlexander A. Klimov * 383 on page 105. Intel should be safe but is also warns that it's 2506423ac9afSAneesh Kumar K.V * only safe if the permission and cache attributes of the two entries 2507423ac9afSAneesh Kumar K.V * loaded in the two TLB is identical (which should be the case here). 2508423ac9afSAneesh Kumar K.V * But it is generally safer to never allow small and huge TLB entries 2509423ac9afSAneesh Kumar K.V * for the same virtual address to be loaded simultaneously. So instead 2510423ac9afSAneesh Kumar K.V * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the 2511423ac9afSAneesh Kumar K.V * current pmd notpresent (atomically because here the pmd_trans_huge 2512423ac9afSAneesh Kumar K.V * must remain set at all times on the pmd until the split is complete 2513423ac9afSAneesh Kumar K.V * for this pmd), then we flush the SMP TLB and finally we write the 2514423ac9afSAneesh Kumar K.V * non-huge version of the pmd entry with pmd_populate. 2515423ac9afSAneesh Kumar K.V */ 2516423ac9afSAneesh Kumar K.V old_pmd = pmdp_invalidate(vma, haddr, pmd); 2517423ac9afSAneesh Kumar K.V 2518423ac9afSAneesh Kumar K.V pmd_migration = is_pmd_migration_entry(old_pmd); 25192e83ee1dSPeter Xu if (unlikely(pmd_migration)) { 252084c3fc4eSZi Yan swp_entry_t entry; 252184c3fc4eSZi Yan 2522423ac9afSAneesh Kumar K.V entry = pmd_to_swp_entry(old_pmd); 2523af5cdaf8SAlistair Popple page = pfn_swap_entry_to_page(entry); 25244dd845b5SAlistair Popple write = is_writable_migration_entry(entry); 25256c287605SDavid Hildenbrand if (PageAnon(page)) 25266c287605SDavid Hildenbrand anon_exclusive = is_readable_exclusive_migration_entry(entry); 25272e346877SPeter Xu young = is_migration_entry_young(entry); 25282e346877SPeter Xu dirty = is_migration_entry_dirty(entry); 25292e83ee1dSPeter Xu soft_dirty = pmd_swp_soft_dirty(old_pmd); 2530f45ec5ffSPeter Xu uffd_wp = pmd_swp_uffd_wp(old_pmd); 25312e83ee1dSPeter Xu } else { 2532423ac9afSAneesh Kumar K.V page = pmd_page(old_pmd); 253391b2978aSDavid Hildenbrand folio = page_folio(page); 25340ccf7f16SPeter Xu if (pmd_dirty(old_pmd)) { 25350ccf7f16SPeter Xu dirty = true; 253691b2978aSDavid Hildenbrand folio_set_dirty(folio); 25370ccf7f16SPeter Xu } 2538423ac9afSAneesh Kumar K.V write = pmd_write(old_pmd); 2539423ac9afSAneesh Kumar K.V young = pmd_young(old_pmd); 2540423ac9afSAneesh Kumar K.V soft_dirty = pmd_soft_dirty(old_pmd); 2541292924b2SPeter Xu uffd_wp = pmd_uffd_wp(old_pmd); 25426c287605SDavid Hildenbrand 254391b2978aSDavid Hildenbrand VM_WARN_ON_FOLIO(!folio_ref_count(folio), folio); 254491b2978aSDavid Hildenbrand VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio); 25456c287605SDavid Hildenbrand 25466c287605SDavid Hildenbrand /* 25476c287605SDavid Hildenbrand * Without "freeze", we'll simply split the PMD, propagating the 25486c287605SDavid Hildenbrand * PageAnonExclusive() flag for each PTE by setting it for 25496c287605SDavid Hildenbrand * each subpage -- no need to (temporarily) clear. 25506c287605SDavid Hildenbrand * 25516c287605SDavid Hildenbrand * With "freeze" we want to replace mapped pages by 25526c287605SDavid Hildenbrand * migration entries right away. This is only possible if we 25536c287605SDavid Hildenbrand * managed to clear PageAnonExclusive() -- see 25546c287605SDavid Hildenbrand * set_pmd_migration_entry(). 25556c287605SDavid Hildenbrand * 25566c287605SDavid Hildenbrand * In case we cannot clear PageAnonExclusive(), split the PMD 25576c287605SDavid Hildenbrand * only and let try_to_migrate_one() fail later. 2558088b8aa5SDavid Hildenbrand * 2559e3b4b137SDavid Hildenbrand * See folio_try_share_anon_rmap_pmd(): invalidate PMD first. 25606c287605SDavid Hildenbrand */ 256191b2978aSDavid Hildenbrand anon_exclusive = PageAnonExclusive(page); 2562e3b4b137SDavid Hildenbrand if (freeze && anon_exclusive && 2563e3b4b137SDavid Hildenbrand folio_try_share_anon_rmap_pmd(folio, page)) 25646c287605SDavid Hildenbrand freeze = false; 256591b2978aSDavid Hildenbrand if (!freeze) { 256691b2978aSDavid Hildenbrand rmap_t rmap_flags = RMAP_NONE; 256791b2978aSDavid Hildenbrand 256891b2978aSDavid Hildenbrand folio_ref_add(folio, HPAGE_PMD_NR - 1); 256991b2978aSDavid Hildenbrand if (anon_exclusive) 257091b2978aSDavid Hildenbrand rmap_flags |= RMAP_EXCLUSIVE; 257191b2978aSDavid Hildenbrand folio_add_anon_rmap_ptes(folio, page, HPAGE_PMD_NR, 257291b2978aSDavid Hildenbrand vma, haddr, rmap_flags); 257391b2978aSDavid Hildenbrand } 25749d84604bSHugh Dickins } 2575eef1b3baSKirill A. Shutemov 2576423ac9afSAneesh Kumar K.V /* 2577423ac9afSAneesh Kumar K.V * Withdraw the table only after we mark the pmd entry invalid. 2578423ac9afSAneesh Kumar K.V * This's critical for some architectures (Power). 2579423ac9afSAneesh Kumar K.V */ 2580eef1b3baSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2581eef1b3baSKirill A. Shutemov pmd_populate(mm, &_pmd, pgtable); 2582eef1b3baSKirill A. Shutemov 2583c9c1ee20SHugh Dickins pte = pte_offset_map(&_pmd, haddr); 2584c9c1ee20SHugh Dickins VM_BUG_ON(!pte); 25852bdba986SRyan Roberts 2586eef1b3baSKirill A. Shutemov /* 25872bdba986SRyan Roberts * Note that NUMA hinting access restrictions are not transferred to 25882bdba986SRyan Roberts * avoid any possibility of altering permissions across VMAs. 2589eef1b3baSKirill A. Shutemov */ 259084c3fc4eSZi Yan if (freeze || pmd_migration) { 25912bdba986SRyan Roberts for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 25922bdba986SRyan Roberts pte_t entry; 2593ba988280SKirill A. Shutemov swp_entry_t swp_entry; 25942bdba986SRyan Roberts 25954dd845b5SAlistair Popple if (write) 25964dd845b5SAlistair Popple swp_entry = make_writable_migration_entry( 25974dd845b5SAlistair Popple page_to_pfn(page + i)); 25986c287605SDavid Hildenbrand else if (anon_exclusive) 25996c287605SDavid Hildenbrand swp_entry = make_readable_exclusive_migration_entry( 26006c287605SDavid Hildenbrand page_to_pfn(page + i)); 26014dd845b5SAlistair Popple else 26024dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 26034dd845b5SAlistair Popple page_to_pfn(page + i)); 26042e346877SPeter Xu if (young) 26052e346877SPeter Xu swp_entry = make_migration_entry_young(swp_entry); 26062e346877SPeter Xu if (dirty) 26072e346877SPeter Xu swp_entry = make_migration_entry_dirty(swp_entry); 2608ba988280SKirill A. Shutemov entry = swp_entry_to_pte(swp_entry); 2609804dd150SAndrea Arcangeli if (soft_dirty) 2610804dd150SAndrea Arcangeli entry = pte_swp_mksoft_dirty(entry); 2611f45ec5ffSPeter Xu if (uffd_wp) 2612f45ec5ffSPeter Xu entry = pte_swp_mkuffd_wp(entry); 26132bdba986SRyan Roberts 26142bdba986SRyan Roberts VM_WARN_ON(!pte_none(ptep_get(pte + i))); 26152bdba986SRyan Roberts set_pte_at(mm, addr, pte + i, entry); 26162bdba986SRyan Roberts } 2617ba988280SKirill A. Shutemov } else { 26182bdba986SRyan Roberts pte_t entry; 26192bdba986SRyan Roberts 26202bdba986SRyan Roberts entry = mk_pte(page, READ_ONCE(vma->vm_page_prot)); 26211462c52eSDavid Hildenbrand if (write) 2622161e393cSRick Edgecombe entry = pte_mkwrite(entry, vma); 2623eef1b3baSKirill A. Shutemov if (!young) 2624eef1b3baSKirill A. Shutemov entry = pte_mkold(entry); 2625e833bc50SPeter Xu /* NOTE: this may set soft-dirty too on some archs */ 2626e833bc50SPeter Xu if (dirty) 2627e833bc50SPeter Xu entry = pte_mkdirty(entry); 2628804dd150SAndrea Arcangeli if (soft_dirty) 2629804dd150SAndrea Arcangeli entry = pte_mksoft_dirty(entry); 2630292924b2SPeter Xu if (uffd_wp) 2631292924b2SPeter Xu entry = pte_mkuffd_wp(entry); 26322bdba986SRyan Roberts 26332bdba986SRyan Roberts for (i = 0; i < HPAGE_PMD_NR; i++) 26342bdba986SRyan Roberts VM_WARN_ON(!pte_none(ptep_get(pte + i))); 26352bdba986SRyan Roberts 26362bdba986SRyan Roberts set_ptes(mm, haddr, pte, entry, HPAGE_PMD_NR); 2637ba988280SKirill A. Shutemov } 26382bdba986SRyan Roberts pte_unmap(pte); 2639eef1b3baSKirill A. Shutemov 2640cb67f428SHugh Dickins if (!pmd_migration) 2641a8e61d58SDavid Hildenbrand folio_remove_rmap_pmd(folio, page, vma); 264296d82debSHugh Dickins if (freeze) 264396d82debSHugh Dickins put_page(page); 2644eef1b3baSKirill A. Shutemov 2645eef1b3baSKirill A. Shutemov smp_wmb(); /* make pte visible before pmd */ 2646eef1b3baSKirill A. Shutemov pmd_populate(mm, pmd, pgtable); 2647eef1b3baSKirill A. Shutemov } 2648eef1b3baSKirill A. Shutemov 2649eef1b3baSKirill A. Shutemov void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 2650af28a988SMatthew Wilcox (Oracle) unsigned long address, bool freeze, struct folio *folio) 2651eef1b3baSKirill A. Shutemov { 2652eef1b3baSKirill A. Shutemov spinlock_t *ptl; 2653ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 2654eef1b3baSKirill A. Shutemov 26557d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 26566f4f13e8SJérôme Glisse address & HPAGE_PMD_MASK, 2657ac46d4f3SJérôme Glisse (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE); 2658ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 2659ac46d4f3SJérôme Glisse ptl = pmd_lock(vma->vm_mm, pmd); 266033f4751eSNaoya Horiguchi 266133f4751eSNaoya Horiguchi /* 2662af28a988SMatthew Wilcox (Oracle) * If caller asks to setup a migration entry, we need a folio to check 2663af28a988SMatthew Wilcox (Oracle) * pmd against. Otherwise we can end up replacing wrong folio. 266433f4751eSNaoya Horiguchi */ 2665af28a988SMatthew Wilcox (Oracle) VM_BUG_ON(freeze && !folio); 266683a8441fSMatthew Wilcox (Oracle) VM_WARN_ON_ONCE(folio && !folio_test_locked(folio)); 266733f4751eSNaoya Horiguchi 26687f760917SDavid Hildenbrand if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) || 266983a8441fSMatthew Wilcox (Oracle) is_pmd_migration_entry(*pmd)) { 2670cea33328SMiaohe Lin /* 2671cea33328SMiaohe Lin * It's safe to call pmd_page when folio is set because it's 2672cea33328SMiaohe Lin * guaranteed that pmd is present. 2673cea33328SMiaohe Lin */ 267483a8441fSMatthew Wilcox (Oracle) if (folio && folio != page_folio(pmd_page(*pmd))) 267583a8441fSMatthew Wilcox (Oracle) goto out; 2676ac46d4f3SJérôme Glisse __split_huge_pmd_locked(vma, pmd, range.start, freeze); 267783a8441fSMatthew Wilcox (Oracle) } 26787f760917SDavid Hildenbrand 2679e90309c9SKirill A. Shutemov out: 2680eef1b3baSKirill A. Shutemov spin_unlock(ptl); 2681ec8832d0SAlistair Popple mmu_notifier_invalidate_range_end(&range); 2682eef1b3baSKirill A. Shutemov } 2683eef1b3baSKirill A. Shutemov 2684fec89c10SKirill A. Shutemov void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, 2685af28a988SMatthew Wilcox (Oracle) bool freeze, struct folio *folio) 268694fcc585SAndrea Arcangeli { 268750722804SZach O'Keefe pmd_t *pmd = mm_find_pmd(vma->vm_mm, address); 268894fcc585SAndrea Arcangeli 268950722804SZach O'Keefe if (!pmd) 2690f72e7dcdSHugh Dickins return; 2691f72e7dcdSHugh Dickins 2692af28a988SMatthew Wilcox (Oracle) __split_huge_pmd(vma, pmd, address, freeze, folio); 269394fcc585SAndrea Arcangeli } 269494fcc585SAndrea Arcangeli 269571f9e58eSMiaohe Lin static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address) 269671f9e58eSMiaohe Lin { 269771f9e58eSMiaohe Lin /* 269871f9e58eSMiaohe Lin * If the new address isn't hpage aligned and it could previously 269971f9e58eSMiaohe Lin * contain an hugepage: check if we need to split an huge pmd. 270071f9e58eSMiaohe Lin */ 270171f9e58eSMiaohe Lin if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) && 270271f9e58eSMiaohe Lin range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE), 270371f9e58eSMiaohe Lin ALIGN(address, HPAGE_PMD_SIZE))) 270471f9e58eSMiaohe Lin split_huge_pmd_address(vma, address, false, NULL); 270571f9e58eSMiaohe Lin } 270671f9e58eSMiaohe Lin 2707e1b9996bSKirill A. Shutemov void vma_adjust_trans_huge(struct vm_area_struct *vma, 270894fcc585SAndrea Arcangeli unsigned long start, 270994fcc585SAndrea Arcangeli unsigned long end, 271094fcc585SAndrea Arcangeli long adjust_next) 271194fcc585SAndrea Arcangeli { 271271f9e58eSMiaohe Lin /* Check if we need to split start first. */ 271371f9e58eSMiaohe Lin split_huge_pmd_if_needed(vma, start); 271471f9e58eSMiaohe Lin 271571f9e58eSMiaohe Lin /* Check if we need to split end next. */ 271671f9e58eSMiaohe Lin split_huge_pmd_if_needed(vma, end); 271794fcc585SAndrea Arcangeli 271894fcc585SAndrea Arcangeli /* 271968540502SMatthew Wilcox (Oracle) * If we're also updating the next vma vm_start, 272071f9e58eSMiaohe Lin * check if we need to split it. 272194fcc585SAndrea Arcangeli */ 272294fcc585SAndrea Arcangeli if (adjust_next > 0) { 272368540502SMatthew Wilcox (Oracle) struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end); 272494fcc585SAndrea Arcangeli unsigned long nstart = next->vm_start; 2725f9d86a60SWei Yang nstart += adjust_next; 272671f9e58eSMiaohe Lin split_huge_pmd_if_needed(next, nstart); 272794fcc585SAndrea Arcangeli } 272894fcc585SAndrea Arcangeli } 2729e9b61f19SKirill A. Shutemov 2730684555aaSMatthew Wilcox (Oracle) static void unmap_folio(struct folio *folio) 2731e9b61f19SKirill A. Shutemov { 2732319a624eSZi Yan enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SYNC | 2733319a624eSZi Yan TTU_BATCH_FLUSH; 2734e9b61f19SKirill A. Shutemov 2735684555aaSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_large(folio), folio); 2736e9b61f19SKirill A. Shutemov 2737319a624eSZi Yan if (folio_test_pmd_mappable(folio)) 2738319a624eSZi Yan ttu_flags |= TTU_SPLIT_HUGE_PMD; 2739319a624eSZi Yan 2740a98a2f0cSAlistair Popple /* 2741a98a2f0cSAlistair Popple * Anon pages need migration entries to preserve them, but file 2742a98a2f0cSAlistair Popple * pages can simply be left unmapped, then faulted back on demand. 2743a98a2f0cSAlistair Popple * If that is ever changed (perhaps for mlock), update remap_page(). 2744a98a2f0cSAlistair Popple */ 27454b8554c5SMatthew Wilcox (Oracle) if (folio_test_anon(folio)) 27464b8554c5SMatthew Wilcox (Oracle) try_to_migrate(folio, ttu_flags); 2747a98a2f0cSAlistair Popple else 2748869f7ee6SMatthew Wilcox (Oracle) try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK); 27493027c6f8SBaolin Wang 27503027c6f8SBaolin Wang try_to_unmap_flush(); 2751bd56086fSKirill A. Shutemov } 2752bd56086fSKirill A. Shutemov 27534eecb8b9SMatthew Wilcox (Oracle) static void remap_page(struct folio *folio, unsigned long nr) 2754e9b61f19SKirill A. Shutemov { 27554eecb8b9SMatthew Wilcox (Oracle) int i = 0; 2756ab02c252SHugh Dickins 2757684555aaSMatthew Wilcox (Oracle) /* If unmap_folio() uses try_to_migrate() on file, remove this check */ 27584eecb8b9SMatthew Wilcox (Oracle) if (!folio_test_anon(folio)) 2759ab02c252SHugh Dickins return; 27604eecb8b9SMatthew Wilcox (Oracle) for (;;) { 27614eecb8b9SMatthew Wilcox (Oracle) remove_migration_ptes(folio, folio, true); 27624eecb8b9SMatthew Wilcox (Oracle) i += folio_nr_pages(folio); 27634eecb8b9SMatthew Wilcox (Oracle) if (i >= nr) 27644eecb8b9SMatthew Wilcox (Oracle) break; 27654eecb8b9SMatthew Wilcox (Oracle) folio = folio_next(folio); 2766e9b61f19SKirill A. Shutemov } 2767ace71a19SKirill A. Shutemov } 2768e9b61f19SKirill A. Shutemov 276994866635SAlex Shi static void lru_add_page_tail(struct page *head, struct page *tail, 277088dcb9a3SAlex Shi struct lruvec *lruvec, struct list_head *list) 277188dcb9a3SAlex Shi { 277294866635SAlex Shi VM_BUG_ON_PAGE(!PageHead(head), head); 277394866635SAlex Shi VM_BUG_ON_PAGE(PageLRU(tail), head); 27746168d0daSAlex Shi lockdep_assert_held(&lruvec->lru_lock); 277588dcb9a3SAlex Shi 27766dbb5741SAlex Shi if (list) { 277788dcb9a3SAlex Shi /* page reclaim is reclaiming a huge page */ 27786dbb5741SAlex Shi VM_WARN_ON(PageLRU(head)); 277994866635SAlex Shi get_page(tail); 278094866635SAlex Shi list_add_tail(&tail->lru, list); 278188dcb9a3SAlex Shi } else { 27826dbb5741SAlex Shi /* head is still on lru (and we have it frozen) */ 27836dbb5741SAlex Shi VM_WARN_ON(!PageLRU(head)); 278407ca7606SHugh Dickins if (PageUnevictable(tail)) 278507ca7606SHugh Dickins tail->mlock_count = 0; 278607ca7606SHugh Dickins else 27876dbb5741SAlex Shi list_add_tail(&tail->lru, &head->lru); 278807ca7606SHugh Dickins SetPageLRU(tail); 278988dcb9a3SAlex Shi } 279088dcb9a3SAlex Shi } 279188dcb9a3SAlex Shi 279207e09c48SDavid Hildenbrand static void __split_huge_page_tail(struct folio *folio, int tail, 2793c010d47fSZi Yan struct lruvec *lruvec, struct list_head *list, 2794c010d47fSZi Yan unsigned int new_order) 2795e9b61f19SKirill A. Shutemov { 279607e09c48SDavid Hildenbrand struct page *head = &folio->page; 2797e9b61f19SKirill A. Shutemov struct page *page_tail = head + tail; 279807e09c48SDavid Hildenbrand /* 279907e09c48SDavid Hildenbrand * Careful: new_folio is not a "real" folio before we cleared PageTail. 280007e09c48SDavid Hildenbrand * Don't pass it around before clear_compound_head(). 280107e09c48SDavid Hildenbrand */ 280207e09c48SDavid Hildenbrand struct folio *new_folio = (struct folio *)page_tail; 2803e9b61f19SKirill A. Shutemov 28048df651c7SKirill A. Shutemov VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); 2805e9b61f19SKirill A. Shutemov 2806e9b61f19SKirill A. Shutemov /* 2807605ca5edSKonstantin Khlebnikov * Clone page flags before unfreezing refcount. 2808605ca5edSKonstantin Khlebnikov * 2809605ca5edSKonstantin Khlebnikov * After successful get_page_unless_zero() might follow flags change, 28108958b249SHaitao Shi * for example lock_page() which set PG_waiters. 28116c287605SDavid Hildenbrand * 28126c287605SDavid Hildenbrand * Note that for mapped sub-pages of an anonymous THP, 2813684555aaSMatthew Wilcox (Oracle) * PG_anon_exclusive has been cleared in unmap_folio() and is stored in 28146c287605SDavid Hildenbrand * the migration entry instead from where remap_page() will restore it. 28156c287605SDavid Hildenbrand * We can still have PG_anon_exclusive set on effectively unmapped and 28166c287605SDavid Hildenbrand * unreferenced sub-pages of an anonymous THP: we can simply drop 28176c287605SDavid Hildenbrand * PG_anon_exclusive (-> PG_mappedtodisk) for these here. 2818e9b61f19SKirill A. Shutemov */ 2819e9b61f19SKirill A. Shutemov page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 2820e9b61f19SKirill A. Shutemov page_tail->flags |= (head->flags & 2821e9b61f19SKirill A. Shutemov ((1L << PG_referenced) | 2822e9b61f19SKirill A. Shutemov (1L << PG_swapbacked) | 282338d8b4e6SHuang Ying (1L << PG_swapcache) | 2824e9b61f19SKirill A. Shutemov (1L << PG_mlocked) | 2825e9b61f19SKirill A. Shutemov (1L << PG_uptodate) | 2826e9b61f19SKirill A. Shutemov (1L << PG_active) | 28271899ad18SJohannes Weiner (1L << PG_workingset) | 2828e9b61f19SKirill A. Shutemov (1L << PG_locked) | 2829b8d3c4c3SMinchan Kim (1L << PG_unevictable) | 2830b0284cd2SCatalin Marinas #ifdef CONFIG_ARCH_USES_PG_ARCH_X 283172e6afa0SCatalin Marinas (1L << PG_arch_2) | 2832ef6458b1SPeter Collingbourne (1L << PG_arch_3) | 283372e6afa0SCatalin Marinas #endif 2834ec1c86b2SYu Zhao (1L << PG_dirty) | 2835ec1c86b2SYu Zhao LRU_GEN_MASK | LRU_REFS_MASK)); 2836e9b61f19SKirill A. Shutemov 2837cb67f428SHugh Dickins /* ->mapping in first and second tail page is replaced by other uses */ 2838173d9d9fSHugh Dickins VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, 2839173d9d9fSHugh Dickins page_tail); 2840173d9d9fSHugh Dickins page_tail->mapping = head->mapping; 2841173d9d9fSHugh Dickins page_tail->index = head->index + tail; 284271e2d666SMel Gorman 284371e2d666SMel Gorman /* 2844cfeed8ffSDavid Hildenbrand * page->private should not be set in tail pages. Fix up and warn once 2845cfeed8ffSDavid Hildenbrand * if private is unexpectedly set. 284671e2d666SMel Gorman */ 2847cfeed8ffSDavid Hildenbrand if (unlikely(page_tail->private)) { 2848cfeed8ffSDavid Hildenbrand VM_WARN_ON_ONCE_PAGE(true, page_tail); 2849b653db77SMatthew Wilcox (Oracle) page_tail->private = 0; 285071e2d666SMel Gorman } 285107e09c48SDavid Hildenbrand if (folio_test_swapcache(folio)) 285207e09c48SDavid Hildenbrand new_folio->swap.val = folio->swap.val + tail; 2853173d9d9fSHugh Dickins 2854605ca5edSKonstantin Khlebnikov /* Page flags must be visible before we make the page non-compound. */ 2855e9b61f19SKirill A. Shutemov smp_wmb(); 2856e9b61f19SKirill A. Shutemov 2857605ca5edSKonstantin Khlebnikov /* 2858605ca5edSKonstantin Khlebnikov * Clear PageTail before unfreezing page refcount. 2859605ca5edSKonstantin Khlebnikov * 2860605ca5edSKonstantin Khlebnikov * After successful get_page_unless_zero() might follow put_page() 2861605ca5edSKonstantin Khlebnikov * which needs correct compound_head(). 2862605ca5edSKonstantin Khlebnikov */ 2863e9b61f19SKirill A. Shutemov clear_compound_head(page_tail); 2864c010d47fSZi Yan if (new_order) { 2865c010d47fSZi Yan prep_compound_page(page_tail, new_order); 2866c010d47fSZi Yan folio_prep_large_rmappable(new_folio); 2867c010d47fSZi Yan } 2868e9b61f19SKirill A. Shutemov 2869605ca5edSKonstantin Khlebnikov /* Finally unfreeze refcount. Additional reference from page cache. */ 2870c010d47fSZi Yan page_ref_unfreeze(page_tail, 2871c010d47fSZi Yan 1 + ((!folio_test_anon(folio) || folio_test_swapcache(folio)) ? 2872c010d47fSZi Yan folio_nr_pages(new_folio) : 0)); 2873605ca5edSKonstantin Khlebnikov 2874b7542769SKefeng Wang if (folio_test_young(folio)) 2875b7542769SKefeng Wang folio_set_young(new_folio); 2876b7542769SKefeng Wang if (folio_test_idle(folio)) 2877b7542769SKefeng Wang folio_set_idle(new_folio); 2878e9b61f19SKirill A. Shutemov 2879c8253011SKefeng Wang folio_xchg_last_cpupid(new_folio, folio_last_cpupid(folio)); 288094723aafSMichal Hocko 288194723aafSMichal Hocko /* 288294723aafSMichal Hocko * always add to the tail because some iterators expect new 288394723aafSMichal Hocko * pages to show after the currently processed elements - e.g. 288494723aafSMichal Hocko * migrate_pages 288594723aafSMichal Hocko */ 2886e9b61f19SKirill A. Shutemov lru_add_page_tail(head, page_tail, lruvec, list); 2887e9b61f19SKirill A. Shutemov } 2888e9b61f19SKirill A. Shutemov 2889baa355fdSKirill A. Shutemov static void __split_huge_page(struct page *page, struct list_head *list, 2890c010d47fSZi Yan pgoff_t end, unsigned int new_order) 2891e9b61f19SKirill A. Shutemov { 2892e809c3feSMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 2893e809c3feSMatthew Wilcox (Oracle) struct page *head = &folio->page; 2894e9b61f19SKirill A. Shutemov struct lruvec *lruvec; 28954101196bSMatthew Wilcox (Oracle) struct address_space *swap_cache = NULL; 28964101196bSMatthew Wilcox (Oracle) unsigned long offset = 0; 2897509f0069SHugh Dickins int i, nr_dropped = 0; 2898c010d47fSZi Yan unsigned int new_nr = 1 << new_order; 2899502003bbSZi Yan int order = folio_order(folio); 2900502003bbSZi Yan unsigned int nr = 1 << order; 2901e9b61f19SKirill A. Shutemov 2902e9b61f19SKirill A. Shutemov /* complete memcg works before add pages to LRU */ 2903c010d47fSZi Yan split_page_memcg(head, order, new_order); 2904e9b61f19SKirill A. Shutemov 290507e09c48SDavid Hildenbrand if (folio_test_anon(folio) && folio_test_swapcache(folio)) { 290607e09c48SDavid Hildenbrand offset = swp_offset(folio->swap); 290707e09c48SDavid Hildenbrand swap_cache = swap_address_space(folio->swap); 29084101196bSMatthew Wilcox (Oracle) xa_lock(&swap_cache->i_pages); 29094101196bSMatthew Wilcox (Oracle) } 29104101196bSMatthew Wilcox (Oracle) 2911f0953a1bSIngo Molnar /* lock lru list/PageCompound, ref frozen by page_ref_freeze */ 2912e809c3feSMatthew Wilcox (Oracle) lruvec = folio_lruvec_lock(folio); 2913b6769834SAlex Shi 2914eac96c3eSYang Shi ClearPageHasHWPoisoned(head); 2915eac96c3eSYang Shi 2916c010d47fSZi Yan for (i = nr - new_nr; i >= new_nr; i -= new_nr) { 2917c010d47fSZi Yan __split_huge_page_tail(folio, i, lruvec, list, new_order); 2918d144bf62SHugh Dickins /* Some pages can be beyond EOF: drop them from page cache */ 2919baa355fdSKirill A. Shutemov if (head[i].index >= end) { 2920fb5c2029SMatthew Wilcox (Oracle) struct folio *tail = page_folio(head + i); 2921fb5c2029SMatthew Wilcox (Oracle) 2922435a7554SMatthew Wilcox (Oracle) if (shmem_mapping(folio->mapping)) 2923509f0069SHugh Dickins nr_dropped++; 2924fb5c2029SMatthew Wilcox (Oracle) else if (folio_test_clear_dirty(tail)) 2925fb5c2029SMatthew Wilcox (Oracle) folio_account_cleaned(tail, 2926fb5c2029SMatthew Wilcox (Oracle) inode_to_wb(folio->mapping->host)); 2927fb5c2029SMatthew Wilcox (Oracle) __filemap_remove_folio(tail, NULL); 2928fb5c2029SMatthew Wilcox (Oracle) folio_put(tail); 29294101196bSMatthew Wilcox (Oracle) } else if (!PageAnon(page)) { 2930435a7554SMatthew Wilcox (Oracle) __xa_store(&folio->mapping->i_pages, head[i].index, 29314101196bSMatthew Wilcox (Oracle) head + i, 0); 29324101196bSMatthew Wilcox (Oracle) } else if (swap_cache) { 29334101196bSMatthew Wilcox (Oracle) __xa_store(&swap_cache->i_pages, offset + i, 29344101196bSMatthew Wilcox (Oracle) head + i, 0); 2935baa355fdSKirill A. Shutemov } 2936baa355fdSKirill A. Shutemov } 2937e9b61f19SKirill A. Shutemov 2938c010d47fSZi Yan if (!new_order) 2939e9b61f19SKirill A. Shutemov ClearPageCompound(head); 2940c010d47fSZi Yan else { 2941c010d47fSZi Yan struct folio *new_folio = (struct folio *)head; 2942c010d47fSZi Yan 2943c010d47fSZi Yan folio_set_order(new_folio, new_order); 2944c010d47fSZi Yan } 29456168d0daSAlex Shi unlock_page_lruvec(lruvec); 2946b6769834SAlex Shi /* Caller disabled irqs, so they are still disabled here */ 2947f7da677bSVlastimil Babka 2948c010d47fSZi Yan split_page_owner(head, order, new_order); 2949f7da677bSVlastimil Babka 2950baa355fdSKirill A. Shutemov /* See comment in __split_huge_page_tail() */ 2951435a7554SMatthew Wilcox (Oracle) if (folio_test_anon(folio)) { 2952aa5dc07fSMatthew Wilcox /* Additional pin to swap cache */ 2953435a7554SMatthew Wilcox (Oracle) if (folio_test_swapcache(folio)) { 2954435a7554SMatthew Wilcox (Oracle) folio_ref_add(folio, 1 + new_nr); 29554101196bSMatthew Wilcox (Oracle) xa_unlock(&swap_cache->i_pages); 29564101196bSMatthew Wilcox (Oracle) } else { 2957435a7554SMatthew Wilcox (Oracle) folio_ref_inc(folio); 29584101196bSMatthew Wilcox (Oracle) } 2959baa355fdSKirill A. Shutemov } else { 2960aa5dc07fSMatthew Wilcox /* Additional pin to page cache */ 2961435a7554SMatthew Wilcox (Oracle) folio_ref_add(folio, 1 + new_nr); 2962435a7554SMatthew Wilcox (Oracle) xa_unlock(&folio->mapping->i_pages); 2963baa355fdSKirill A. Shutemov } 2964b6769834SAlex Shi local_irq_enable(); 2965e9b61f19SKirill A. Shutemov 2966509f0069SHugh Dickins if (nr_dropped) 2967435a7554SMatthew Wilcox (Oracle) shmem_uncharge(folio->mapping->host, nr_dropped); 29684eecb8b9SMatthew Wilcox (Oracle) remap_page(folio, nr); 2969e9b61f19SKirill A. Shutemov 297007e09c48SDavid Hildenbrand if (folio_test_swapcache(folio)) 297107e09c48SDavid Hildenbrand split_swap_cluster(folio->swap); 2972c4f9c701SHuang Ying 2973c010d47fSZi Yan /* 2974c010d47fSZi Yan * set page to its compound_head when split to non order-0 pages, so 2975c010d47fSZi Yan * we can skip unlocking it below, since PG_locked is transferred to 2976c010d47fSZi Yan * the compound_head of the page and the caller will unlock it. 2977c010d47fSZi Yan */ 2978c010d47fSZi Yan if (new_order) 2979c010d47fSZi Yan page = compound_head(page); 2980c010d47fSZi Yan 2981c010d47fSZi Yan for (i = 0; i < nr; i += new_nr) { 2982e9b61f19SKirill A. Shutemov struct page *subpage = head + i; 2983435a7554SMatthew Wilcox (Oracle) struct folio *new_folio = page_folio(subpage); 2984e9b61f19SKirill A. Shutemov if (subpage == page) 2985e9b61f19SKirill A. Shutemov continue; 2986435a7554SMatthew Wilcox (Oracle) folio_unlock(new_folio); 2987e9b61f19SKirill A. Shutemov 2988e9b61f19SKirill A. Shutemov /* 2989e9b61f19SKirill A. Shutemov * Subpages may be freed if there wasn't any mapping 2990e9b61f19SKirill A. Shutemov * like if add_to_swap() is running on a lru page that 2991e9b61f19SKirill A. Shutemov * had its mapping zapped. And freeing these pages 2992e9b61f19SKirill A. Shutemov * requires taking the lru_lock so we do the put_page 2993e9b61f19SKirill A. Shutemov * of the tail pages after the split is complete. 2994e9b61f19SKirill A. Shutemov */ 29950b175468SMiaohe Lin free_page_and_swap_cache(subpage); 2996e9b61f19SKirill A. Shutemov } 2997e9b61f19SKirill A. Shutemov } 2998e9b61f19SKirill A. Shutemov 2999b8f593cdSHuang Ying /* Racy check whether the huge page can be split */ 3000d4b4084aSMatthew Wilcox (Oracle) bool can_split_folio(struct folio *folio, int *pextra_pins) 3001b8f593cdSHuang Ying { 3002b8f593cdSHuang Ying int extra_pins; 3003b8f593cdSHuang Ying 3004aa5dc07fSMatthew Wilcox /* Additional pins from page cache */ 3005d4b4084aSMatthew Wilcox (Oracle) if (folio_test_anon(folio)) 3006d4b4084aSMatthew Wilcox (Oracle) extra_pins = folio_test_swapcache(folio) ? 3007d4b4084aSMatthew Wilcox (Oracle) folio_nr_pages(folio) : 0; 3008b8f593cdSHuang Ying else 3009d4b4084aSMatthew Wilcox (Oracle) extra_pins = folio_nr_pages(folio); 3010b8f593cdSHuang Ying if (pextra_pins) 3011b8f593cdSHuang Ying *pextra_pins = extra_pins; 3012d4b4084aSMatthew Wilcox (Oracle) return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1; 3013b8f593cdSHuang Ying } 3014b8f593cdSHuang Ying 30156d0a07edSAndrea Arcangeli /* 3016c010d47fSZi Yan * This function splits huge page into pages in @new_order. @page can point to 3017c010d47fSZi Yan * any subpage of huge page to split. Split doesn't change the position of 3018c010d47fSZi Yan * @page. 3019c010d47fSZi Yan * 3020c010d47fSZi Yan * NOTE: order-1 anonymous folio is not supported because _deferred_list, 3021c010d47fSZi Yan * which is used by partially mapped folios, is stored in subpage 2 and an 3022c010d47fSZi Yan * order-1 folio only has subpage 0 and 1. File-backed order-1 folios are OK, 3023c010d47fSZi Yan * since they do not use _deferred_list. 3024e9b61f19SKirill A. Shutemov * 3025e9b61f19SKirill A. Shutemov * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. 3026e9b61f19SKirill A. Shutemov * The huge page must be locked. 3027e9b61f19SKirill A. Shutemov * 3028e9b61f19SKirill A. Shutemov * If @list is null, tail pages will be added to LRU list, otherwise, to @list. 3029e9b61f19SKirill A. Shutemov * 3030c010d47fSZi Yan * Pages in new_order will inherit mapping, flags, and so on from the hugepage. 3031e9b61f19SKirill A. Shutemov * 3032c010d47fSZi Yan * GUP pin and PG_locked transferred to @page or the compound page @page belongs 3033c010d47fSZi Yan * to. Rest subpages can be freed if they are not mapped. 3034e9b61f19SKirill A. Shutemov * 3035e9b61f19SKirill A. Shutemov * Returns 0 if the hugepage is split successfully. 3036e9b61f19SKirill A. Shutemov * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under 3037e9b61f19SKirill A. Shutemov * us. 3038e9b61f19SKirill A. Shutemov */ 3039c010d47fSZi Yan int split_huge_page_to_list_to_order(struct page *page, struct list_head *list, 3040c010d47fSZi Yan unsigned int new_order) 3041e9b61f19SKirill A. Shutemov { 30424eecb8b9SMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 3043f8baa6beSMatthew Wilcox (Oracle) struct deferred_split *ds_queue = get_deferred_split_queue(folio); 3044c010d47fSZi Yan /* reset xarray order to new order after split */ 3045c010d47fSZi Yan XA_STATE_ORDER(xas, &folio->mapping->i_pages, folio->index, new_order); 3046baa355fdSKirill A. Shutemov struct anon_vma *anon_vma = NULL; 3047baa355fdSKirill A. Shutemov struct address_space *mapping = NULL; 3048504e070dSYang Shi int extra_pins, ret; 3049006d3ff2SHugh Dickins pgoff_t end; 3050478d134eSXu Yu bool is_hzp; 3051e9b61f19SKirill A. Shutemov 30523e9a13daSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 30533e9a13daSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_large(folio), folio); 3054e9b61f19SKirill A. Shutemov 30551412ecb3SZi Yan if (new_order >= folio_order(folio)) 30561412ecb3SZi Yan return -EINVAL; 30571412ecb3SZi Yan 3058c010d47fSZi Yan /* Cannot split anonymous THP to order-1 */ 3059c010d47fSZi Yan if (new_order == 1 && folio_test_anon(folio)) { 3060c010d47fSZi Yan VM_WARN_ONCE(1, "Cannot split to order-1 folio"); 3061c010d47fSZi Yan return -EINVAL; 3062c010d47fSZi Yan } 3063c010d47fSZi Yan 3064c010d47fSZi Yan if (new_order) { 3065c010d47fSZi Yan /* Only swapping a whole PMD-mapped folio is supported */ 3066c010d47fSZi Yan if (folio_test_swapcache(folio)) 3067c010d47fSZi Yan return -EINVAL; 3068c010d47fSZi Yan /* Split shmem folio to non-zero order not supported */ 3069c010d47fSZi Yan if (shmem_mapping(folio->mapping)) { 3070c010d47fSZi Yan VM_WARN_ONCE(1, 3071c010d47fSZi Yan "Cannot split shmem folio to non-0 order"); 3072c010d47fSZi Yan return -EINVAL; 3073c010d47fSZi Yan } 3074c010d47fSZi Yan /* No split if the file system does not support large folio */ 3075c010d47fSZi Yan if (!mapping_large_folio_support(folio->mapping)) { 3076c010d47fSZi Yan VM_WARN_ONCE(1, 3077c010d47fSZi Yan "Cannot split file folio to non-0 order"); 3078c010d47fSZi Yan return -EINVAL; 3079c010d47fSZi Yan } 3080c010d47fSZi Yan } 3081c010d47fSZi Yan 3082c010d47fSZi Yan 30833e9a13daSMatthew Wilcox (Oracle) is_hzp = is_huge_zero_page(&folio->page); 30844737edbbSNaoya Horiguchi if (is_hzp) { 30854737edbbSNaoya Horiguchi pr_warn_ratelimited("Called split_huge_page for huge zero page\n"); 3086478d134eSXu Yu return -EBUSY; 30874737edbbSNaoya Horiguchi } 3088478d134eSXu Yu 30893e9a13daSMatthew Wilcox (Oracle) if (folio_test_writeback(folio)) 309059807685SHuang Ying return -EBUSY; 309159807685SHuang Ying 30923e9a13daSMatthew Wilcox (Oracle) if (folio_test_anon(folio)) { 3093e9b61f19SKirill A. Shutemov /* 3094c1e8d7c6SMichel Lespinasse * The caller does not necessarily hold an mmap_lock that would 3095baa355fdSKirill A. Shutemov * prevent the anon_vma disappearing so we first we take a 3096baa355fdSKirill A. Shutemov * reference to it and then lock the anon_vma for write. This 30972f031c6fSMatthew Wilcox (Oracle) * is similar to folio_lock_anon_vma_read except the write lock 3098baa355fdSKirill A. Shutemov * is taken to serialise against parallel split or collapse 3099baa355fdSKirill A. Shutemov * operations. 3100e9b61f19SKirill A. Shutemov */ 310129eea9b5SMatthew Wilcox (Oracle) anon_vma = folio_get_anon_vma(folio); 3102e9b61f19SKirill A. Shutemov if (!anon_vma) { 3103e9b61f19SKirill A. Shutemov ret = -EBUSY; 3104e9b61f19SKirill A. Shutemov goto out; 3105e9b61f19SKirill A. Shutemov } 3106006d3ff2SHugh Dickins end = -1; 3107baa355fdSKirill A. Shutemov mapping = NULL; 3108e9b61f19SKirill A. Shutemov anon_vma_lock_write(anon_vma); 3109baa355fdSKirill A. Shutemov } else { 31106a3edd29SYin Fengwei gfp_t gfp; 31116a3edd29SYin Fengwei 31123e9a13daSMatthew Wilcox (Oracle) mapping = folio->mapping; 3113baa355fdSKirill A. Shutemov 3114baa355fdSKirill A. Shutemov /* Truncated ? */ 3115baa355fdSKirill A. Shutemov if (!mapping) { 3116baa355fdSKirill A. Shutemov ret = -EBUSY; 3117baa355fdSKirill A. Shutemov goto out; 3118baa355fdSKirill A. Shutemov } 3119baa355fdSKirill A. Shutemov 31206a3edd29SYin Fengwei gfp = current_gfp_context(mapping_gfp_mask(mapping) & 31216a3edd29SYin Fengwei GFP_RECLAIM_MASK); 31226a3edd29SYin Fengwei 31230201ebf2SDavid Howells if (!filemap_release_folio(folio, gfp)) { 31246a3edd29SYin Fengwei ret = -EBUSY; 31256a3edd29SYin Fengwei goto out; 31266a3edd29SYin Fengwei } 31276a3edd29SYin Fengwei 31283e9a13daSMatthew Wilcox (Oracle) xas_split_alloc(&xas, folio, folio_order(folio), gfp); 31296b24ca4aSMatthew Wilcox (Oracle) if (xas_error(&xas)) { 31306b24ca4aSMatthew Wilcox (Oracle) ret = xas_error(&xas); 31316b24ca4aSMatthew Wilcox (Oracle) goto out; 31326b24ca4aSMatthew Wilcox (Oracle) } 31336b24ca4aSMatthew Wilcox (Oracle) 3134baa355fdSKirill A. Shutemov anon_vma = NULL; 3135baa355fdSKirill A. Shutemov i_mmap_lock_read(mapping); 3136006d3ff2SHugh Dickins 3137006d3ff2SHugh Dickins /* 3138006d3ff2SHugh Dickins *__split_huge_page() may need to trim off pages beyond EOF: 3139006d3ff2SHugh Dickins * but on 32-bit, i_size_read() takes an irq-unsafe seqlock, 3140006d3ff2SHugh Dickins * which cannot be nested inside the page tree lock. So note 3141006d3ff2SHugh Dickins * end now: i_size itself may be changed at any moment, but 31423e9a13daSMatthew Wilcox (Oracle) * folio lock is good enough to serialize the trimming. 3143006d3ff2SHugh Dickins */ 3144006d3ff2SHugh Dickins end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); 3145d144bf62SHugh Dickins if (shmem_mapping(mapping)) 3146d144bf62SHugh Dickins end = shmem_fallocend(mapping->host, end); 3147baa355fdSKirill A. Shutemov } 3148e9b61f19SKirill A. Shutemov 3149e9b61f19SKirill A. Shutemov /* 3150684555aaSMatthew Wilcox (Oracle) * Racy check if we can split the page, before unmap_folio() will 3151e9b61f19SKirill A. Shutemov * split PMDs 3152e9b61f19SKirill A. Shutemov */ 3153d4b4084aSMatthew Wilcox (Oracle) if (!can_split_folio(folio, &extra_pins)) { 3154fd4a7ac3SBaolin Wang ret = -EAGAIN; 3155e9b61f19SKirill A. Shutemov goto out_unlock; 3156e9b61f19SKirill A. Shutemov } 3157e9b61f19SKirill A. Shutemov 3158684555aaSMatthew Wilcox (Oracle) unmap_folio(folio); 3159e9b61f19SKirill A. Shutemov 3160b6769834SAlex Shi /* block interrupt reentry in xa_lock and spinlock */ 3161b6769834SAlex Shi local_irq_disable(); 3162baa355fdSKirill A. Shutemov if (mapping) { 3163baa355fdSKirill A. Shutemov /* 31643e9a13daSMatthew Wilcox (Oracle) * Check if the folio is present in page cache. 31653e9a13daSMatthew Wilcox (Oracle) * We assume all tail are present too, if folio is there. 3166baa355fdSKirill A. Shutemov */ 31676b24ca4aSMatthew Wilcox (Oracle) xas_lock(&xas); 31686b24ca4aSMatthew Wilcox (Oracle) xas_reset(&xas); 31693e9a13daSMatthew Wilcox (Oracle) if (xas_load(&xas) != folio) 3170baa355fdSKirill A. Shutemov goto fail; 3171baa355fdSKirill A. Shutemov } 3172baa355fdSKirill A. Shutemov 31730139aa7bSJoonsoo Kim /* Prevent deferred_split_scan() touching ->_refcount */ 3174364c1eebSYang Shi spin_lock(&ds_queue->split_queue_lock); 31753e9a13daSMatthew Wilcox (Oracle) if (folio_ref_freeze(folio, 1 + extra_pins)) { 31768897277aSMatthew Wilcox (Oracle) if (folio_order(folio) > 1 && 31778897277aSMatthew Wilcox (Oracle) !list_empty(&folio->_deferred_list)) { 3178364c1eebSYang Shi ds_queue->split_queue_len--; 3179c010d47fSZi Yan /* 3180c010d47fSZi Yan * Reinitialize page_deferred_list after removing the 3181c010d47fSZi Yan * page from the split_queue, otherwise a subsequent 3182c010d47fSZi Yan * split will see list corruption when checking the 3183c010d47fSZi Yan * page_deferred_list. 3184c010d47fSZi Yan */ 3185c010d47fSZi Yan list_del_init(&folio->_deferred_list); 31869a982250SKirill A. Shutemov } 3187afb97172SWei Yang spin_unlock(&ds_queue->split_queue_lock); 318806d3eff6SKirill A. Shutemov if (mapping) { 31893e9a13daSMatthew Wilcox (Oracle) int nr = folio_nr_pages(folio); 3190bf9eceadSMuchun Song 31913e9a13daSMatthew Wilcox (Oracle) xas_split(&xas, folio, folio_order(folio)); 3192c010d47fSZi Yan if (folio_test_pmd_mappable(folio) && 3193c010d47fSZi Yan new_order < HPAGE_PMD_ORDER) { 31943e9a13daSMatthew Wilcox (Oracle) if (folio_test_swapbacked(folio)) { 3195a48d5bdcSStefan Roesch __lruvec_stat_mod_folio(folio, 3196a48d5bdcSStefan Roesch NR_SHMEM_THPS, -nr); 31971ca7554dSMarek Szyprowski } else { 3198a48d5bdcSStefan Roesch __lruvec_stat_mod_folio(folio, 3199a48d5bdcSStefan Roesch NR_FILE_THPS, -nr); 32001ca7554dSMarek Szyprowski filemap_nr_thps_dec(mapping); 32011ca7554dSMarek Szyprowski } 320206d3eff6SKirill A. Shutemov } 3203a48d5bdcSStefan Roesch } 320406d3eff6SKirill A. Shutemov 3205c010d47fSZi Yan __split_huge_page(page, list, end, new_order); 3206e9b61f19SKirill A. Shutemov ret = 0; 3207baa355fdSKirill A. Shutemov } else { 3208364c1eebSYang Shi spin_unlock(&ds_queue->split_queue_lock); 3209504e070dSYang Shi fail: 3210504e070dSYang Shi if (mapping) 32116b24ca4aSMatthew Wilcox (Oracle) xas_unlock(&xas); 3212b6769834SAlex Shi local_irq_enable(); 32134eecb8b9SMatthew Wilcox (Oracle) remap_page(folio, folio_nr_pages(folio)); 3214fd4a7ac3SBaolin Wang ret = -EAGAIN; 3215e9b61f19SKirill A. Shutemov } 3216e9b61f19SKirill A. Shutemov 3217e9b61f19SKirill A. Shutemov out_unlock: 3218baa355fdSKirill A. Shutemov if (anon_vma) { 3219e9b61f19SKirill A. Shutemov anon_vma_unlock_write(anon_vma); 3220e9b61f19SKirill A. Shutemov put_anon_vma(anon_vma); 3221baa355fdSKirill A. Shutemov } 3222baa355fdSKirill A. Shutemov if (mapping) 3223baa355fdSKirill A. Shutemov i_mmap_unlock_read(mapping); 3224e9b61f19SKirill A. Shutemov out: 322569a37a8bSMatthew Wilcox (Oracle) xas_destroy(&xas); 3226e9b61f19SKirill A. Shutemov count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); 3227e9b61f19SKirill A. Shutemov return ret; 3228e9b61f19SKirill A. Shutemov } 32299a982250SKirill A. Shutemov 32308dc4a8f1SMatthew Wilcox (Oracle) void folio_undo_large_rmappable(struct folio *folio) 32319a982250SKirill A. Shutemov { 32328dc4a8f1SMatthew Wilcox (Oracle) struct deferred_split *ds_queue; 32339a982250SKirill A. Shutemov unsigned long flags; 32349a982250SKirill A. Shutemov 32358897277aSMatthew Wilcox (Oracle) if (folio_order(folio) <= 1) 32368897277aSMatthew Wilcox (Oracle) return; 32378897277aSMatthew Wilcox (Oracle) 3238deedad80SYin Fengwei /* 3239deedad80SYin Fengwei * At this point, there is no one trying to add the folio to 3240deedad80SYin Fengwei * deferred_list. If folio is not in deferred_list, it's safe 3241deedad80SYin Fengwei * to check without acquiring the split_queue_lock. 3242deedad80SYin Fengwei */ 32438dc4a8f1SMatthew Wilcox (Oracle) if (data_race(list_empty(&folio->_deferred_list))) 32448dc4a8f1SMatthew Wilcox (Oracle) return; 32458dc4a8f1SMatthew Wilcox (Oracle) 32468dc4a8f1SMatthew Wilcox (Oracle) ds_queue = get_deferred_split_queue(folio); 3247364c1eebSYang Shi spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 32488991de90SMatthew Wilcox (Oracle) if (!list_empty(&folio->_deferred_list)) { 3249364c1eebSYang Shi ds_queue->split_queue_len--; 32509bcef597SBaolin Wang list_del_init(&folio->_deferred_list); 32519a982250SKirill A. Shutemov } 3252364c1eebSYang Shi spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 3253deedad80SYin Fengwei } 32549a982250SKirill A. Shutemov 3255f158ed61SMatthew Wilcox (Oracle) void deferred_split_folio(struct folio *folio) 32569a982250SKirill A. Shutemov { 3257f8baa6beSMatthew Wilcox (Oracle) struct deferred_split *ds_queue = get_deferred_split_queue(folio); 325887eaceb3SYang Shi #ifdef CONFIG_MEMCG 32598991de90SMatthew Wilcox (Oracle) struct mem_cgroup *memcg = folio_memcg(folio); 326087eaceb3SYang Shi #endif 32619a982250SKirill A. Shutemov unsigned long flags; 32629a982250SKirill A. Shutemov 32638897277aSMatthew Wilcox (Oracle) /* 32648897277aSMatthew Wilcox (Oracle) * Order 1 folios have no space for a deferred list, but we also 32658897277aSMatthew Wilcox (Oracle) * won't waste much memory by not adding them to the deferred list. 32668897277aSMatthew Wilcox (Oracle) */ 32678897277aSMatthew Wilcox (Oracle) if (folio_order(folio) <= 1) 32688897277aSMatthew Wilcox (Oracle) return; 32699a982250SKirill A. Shutemov 327087eaceb3SYang Shi /* 327187eaceb3SYang Shi * The try_to_unmap() in page reclaim path might reach here too, 327287eaceb3SYang Shi * this may cause a race condition to corrupt deferred split queue. 32738991de90SMatthew Wilcox (Oracle) * And, if page reclaim is already handling the same folio, it is 327487eaceb3SYang Shi * unnecessary to handle it again in shrinker. 327587eaceb3SYang Shi * 32768991de90SMatthew Wilcox (Oracle) * Check the swapcache flag to determine if the folio is being 32778991de90SMatthew Wilcox (Oracle) * handled by page reclaim since THP swap would add the folio into 327887eaceb3SYang Shi * swap cache before calling try_to_unmap(). 327987eaceb3SYang Shi */ 32808991de90SMatthew Wilcox (Oracle) if (folio_test_swapcache(folio)) 328187eaceb3SYang Shi return; 328287eaceb3SYang Shi 32838991de90SMatthew Wilcox (Oracle) if (!list_empty(&folio->_deferred_list)) 32849a982250SKirill A. Shutemov return; 32859a982250SKirill A. Shutemov 3286364c1eebSYang Shi spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 32878991de90SMatthew Wilcox (Oracle) if (list_empty(&folio->_deferred_list)) { 3288f9719a03SKirill A. Shutemov count_vm_event(THP_DEFERRED_SPLIT_PAGE); 32898991de90SMatthew Wilcox (Oracle) list_add_tail(&folio->_deferred_list, &ds_queue->split_queue); 3290364c1eebSYang Shi ds_queue->split_queue_len++; 329187eaceb3SYang Shi #ifdef CONFIG_MEMCG 329287eaceb3SYang Shi if (memcg) 32938991de90SMatthew Wilcox (Oracle) set_shrinker_bit(memcg, folio_nid(folio), 329454d91729SQi Zheng deferred_split_shrinker->id); 329587eaceb3SYang Shi #endif 32969a982250SKirill A. Shutemov } 3297364c1eebSYang Shi spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 32989a982250SKirill A. Shutemov } 32999a982250SKirill A. Shutemov 33009a982250SKirill A. Shutemov static unsigned long deferred_split_count(struct shrinker *shrink, 33019a982250SKirill A. Shutemov struct shrink_control *sc) 33029a982250SKirill A. Shutemov { 3303a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(sc->nid); 3304364c1eebSYang Shi struct deferred_split *ds_queue = &pgdata->deferred_split_queue; 330587eaceb3SYang Shi 330687eaceb3SYang Shi #ifdef CONFIG_MEMCG 330787eaceb3SYang Shi if (sc->memcg) 330887eaceb3SYang Shi ds_queue = &sc->memcg->deferred_split_queue; 330987eaceb3SYang Shi #endif 3310364c1eebSYang Shi return READ_ONCE(ds_queue->split_queue_len); 33119a982250SKirill A. Shutemov } 33129a982250SKirill A. Shutemov 33139a982250SKirill A. Shutemov static unsigned long deferred_split_scan(struct shrinker *shrink, 33149a982250SKirill A. Shutemov struct shrink_control *sc) 33159a982250SKirill A. Shutemov { 3316a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(sc->nid); 3317364c1eebSYang Shi struct deferred_split *ds_queue = &pgdata->deferred_split_queue; 33189a982250SKirill A. Shutemov unsigned long flags; 33194375a553SMatthew Wilcox (Oracle) LIST_HEAD(list); 33204375a553SMatthew Wilcox (Oracle) struct folio *folio, *next; 33219a982250SKirill A. Shutemov int split = 0; 33229a982250SKirill A. Shutemov 332387eaceb3SYang Shi #ifdef CONFIG_MEMCG 332487eaceb3SYang Shi if (sc->memcg) 332587eaceb3SYang Shi ds_queue = &sc->memcg->deferred_split_queue; 332687eaceb3SYang Shi #endif 332787eaceb3SYang Shi 3328364c1eebSYang Shi spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 33299a982250SKirill A. Shutemov /* Take pin on all head pages to avoid freeing them under us */ 33304375a553SMatthew Wilcox (Oracle) list_for_each_entry_safe(folio, next, &ds_queue->split_queue, 33314375a553SMatthew Wilcox (Oracle) _deferred_list) { 33324375a553SMatthew Wilcox (Oracle) if (folio_try_get(folio)) { 33334375a553SMatthew Wilcox (Oracle) list_move(&folio->_deferred_list, &list); 3334e3ae1953SKirill A. Shutemov } else { 33354375a553SMatthew Wilcox (Oracle) /* We lost race with folio_put() */ 33364375a553SMatthew Wilcox (Oracle) list_del_init(&folio->_deferred_list); 3337364c1eebSYang Shi ds_queue->split_queue_len--; 33389a982250SKirill A. Shutemov } 3339e3ae1953SKirill A. Shutemov if (!--sc->nr_to_scan) 3340e3ae1953SKirill A. Shutemov break; 33419a982250SKirill A. Shutemov } 3342364c1eebSYang Shi spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 33439a982250SKirill A. Shutemov 33444375a553SMatthew Wilcox (Oracle) list_for_each_entry_safe(folio, next, &list, _deferred_list) { 33454375a553SMatthew Wilcox (Oracle) if (!folio_trylock(folio)) 3346fa41b900SKirill A. Shutemov goto next; 33479a982250SKirill A. Shutemov /* split_huge_page() removes page from list on success */ 33484375a553SMatthew Wilcox (Oracle) if (!split_folio(folio)) 33499a982250SKirill A. Shutemov split++; 33504375a553SMatthew Wilcox (Oracle) folio_unlock(folio); 3351fa41b900SKirill A. Shutemov next: 33524375a553SMatthew Wilcox (Oracle) folio_put(folio); 33539a982250SKirill A. Shutemov } 33549a982250SKirill A. Shutemov 3355364c1eebSYang Shi spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 3356364c1eebSYang Shi list_splice_tail(&list, &ds_queue->split_queue); 3357364c1eebSYang Shi spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 33589a982250SKirill A. Shutemov 3359cb8d68ecSKirill A. Shutemov /* 3360cb8d68ecSKirill A. Shutemov * Stop shrinker if we didn't split any page, but the queue is empty. 3361cb8d68ecSKirill A. Shutemov * This can happen if pages were freed under us. 3362cb8d68ecSKirill A. Shutemov */ 3363364c1eebSYang Shi if (!split && list_empty(&ds_queue->split_queue)) 3364cb8d68ecSKirill A. Shutemov return SHRINK_STOP; 3365cb8d68ecSKirill A. Shutemov return split; 33669a982250SKirill A. Shutemov } 33679a982250SKirill A. Shutemov 336849071d43SKirill A. Shutemov #ifdef CONFIG_DEBUG_FS 3369fa6c0231SZi Yan static void split_huge_pages_all(void) 337049071d43SKirill A. Shutemov { 337149071d43SKirill A. Shutemov struct zone *zone; 337249071d43SKirill A. Shutemov struct page *page; 3373630e7c5eSKefeng Wang struct folio *folio; 337449071d43SKirill A. Shutemov unsigned long pfn, max_zone_pfn; 337549071d43SKirill A. Shutemov unsigned long total = 0, split = 0; 337649071d43SKirill A. Shutemov 3377fa6c0231SZi Yan pr_debug("Split all THPs\n"); 3378a17206daSMiaohe Lin for_each_zone(zone) { 3379a17206daSMiaohe Lin if (!managed_zone(zone)) 3380a17206daSMiaohe Lin continue; 338149071d43SKirill A. Shutemov max_zone_pfn = zone_end_pfn(zone); 338249071d43SKirill A. Shutemov for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { 3383a17206daSMiaohe Lin int nr_pages; 338449071d43SKirill A. Shutemov 33852b7aa91bSNaoya Horiguchi page = pfn_to_online_page(pfn); 3386630e7c5eSKefeng Wang if (!page || PageTail(page)) 3387630e7c5eSKefeng Wang continue; 3388630e7c5eSKefeng Wang folio = page_folio(page); 3389630e7c5eSKefeng Wang if (!folio_try_get(folio)) 339049071d43SKirill A. Shutemov continue; 339149071d43SKirill A. Shutemov 3392630e7c5eSKefeng Wang if (unlikely(page_folio(page) != folio)) 339349071d43SKirill A. Shutemov goto next; 339449071d43SKirill A. Shutemov 3395630e7c5eSKefeng Wang if (zone != folio_zone(folio)) 3396630e7c5eSKefeng Wang goto next; 3397630e7c5eSKefeng Wang 3398630e7c5eSKefeng Wang if (!folio_test_large(folio) 3399630e7c5eSKefeng Wang || folio_test_hugetlb(folio) 3400630e7c5eSKefeng Wang || !folio_test_lru(folio)) 340149071d43SKirill A. Shutemov goto next; 340249071d43SKirill A. Shutemov 340349071d43SKirill A. Shutemov total++; 3404630e7c5eSKefeng Wang folio_lock(folio); 3405630e7c5eSKefeng Wang nr_pages = folio_nr_pages(folio); 3406630e7c5eSKefeng Wang if (!split_folio(folio)) 340749071d43SKirill A. Shutemov split++; 3408a17206daSMiaohe Lin pfn += nr_pages - 1; 3409630e7c5eSKefeng Wang folio_unlock(folio); 341049071d43SKirill A. Shutemov next: 3411630e7c5eSKefeng Wang folio_put(folio); 3412fa6c0231SZi Yan cond_resched(); 341349071d43SKirill A. Shutemov } 341449071d43SKirill A. Shutemov } 341549071d43SKirill A. Shutemov 3416fa6c0231SZi Yan pr_debug("%lu of %lu THP split\n", split, total); 341749071d43SKirill A. Shutemov } 3418fa6c0231SZi Yan 3419fa6c0231SZi Yan static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma) 3420fa6c0231SZi Yan { 3421fa6c0231SZi Yan return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) || 3422fa6c0231SZi Yan is_vm_hugetlb_page(vma); 3423fa6c0231SZi Yan } 3424fa6c0231SZi Yan 3425fa6c0231SZi Yan static int split_huge_pages_pid(int pid, unsigned long vaddr_start, 3426fc4d1823SZi Yan unsigned long vaddr_end, unsigned int new_order) 3427fa6c0231SZi Yan { 3428fa6c0231SZi Yan int ret = 0; 3429fa6c0231SZi Yan struct task_struct *task; 3430fa6c0231SZi Yan struct mm_struct *mm; 3431fa6c0231SZi Yan unsigned long total = 0, split = 0; 3432fa6c0231SZi Yan unsigned long addr; 3433fa6c0231SZi Yan 3434fa6c0231SZi Yan vaddr_start &= PAGE_MASK; 3435fa6c0231SZi Yan vaddr_end &= PAGE_MASK; 3436fa6c0231SZi Yan 3437fa6c0231SZi Yan /* Find the task_struct from pid */ 3438fa6c0231SZi Yan rcu_read_lock(); 3439fa6c0231SZi Yan task = find_task_by_vpid(pid); 3440fa6c0231SZi Yan if (!task) { 3441fa6c0231SZi Yan rcu_read_unlock(); 3442fa6c0231SZi Yan ret = -ESRCH; 3443fa6c0231SZi Yan goto out; 3444fa6c0231SZi Yan } 3445fa6c0231SZi Yan get_task_struct(task); 3446fa6c0231SZi Yan rcu_read_unlock(); 3447fa6c0231SZi Yan 3448fa6c0231SZi Yan /* Find the mm_struct */ 3449fa6c0231SZi Yan mm = get_task_mm(task); 3450fa6c0231SZi Yan put_task_struct(task); 3451fa6c0231SZi Yan 3452fa6c0231SZi Yan if (!mm) { 3453fa6c0231SZi Yan ret = -EINVAL; 3454fa6c0231SZi Yan goto out; 3455fa6c0231SZi Yan } 3456fa6c0231SZi Yan 3457fa6c0231SZi Yan pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n", 3458fa6c0231SZi Yan pid, vaddr_start, vaddr_end); 3459fa6c0231SZi Yan 3460fa6c0231SZi Yan mmap_read_lock(mm); 3461fa6c0231SZi Yan /* 3462fa6c0231SZi Yan * always increase addr by PAGE_SIZE, since we could have a PTE page 3463fa6c0231SZi Yan * table filled with PTE-mapped THPs, each of which is distinct. 3464fa6c0231SZi Yan */ 3465fa6c0231SZi Yan for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) { 346674ba2b38SMiaohe Lin struct vm_area_struct *vma = vma_lookup(mm, addr); 3467fa6c0231SZi Yan struct page *page; 3468a644b0abSMatthew Wilcox (Oracle) struct folio *folio; 3469fa6c0231SZi Yan 347074ba2b38SMiaohe Lin if (!vma) 3471fa6c0231SZi Yan break; 3472fa6c0231SZi Yan 3473fa6c0231SZi Yan /* skip special VMA and hugetlb VMA */ 3474fa6c0231SZi Yan if (vma_not_suitable_for_thp_split(vma)) { 3475fa6c0231SZi Yan addr = vma->vm_end; 3476fa6c0231SZi Yan continue; 3477fa6c0231SZi Yan } 3478fa6c0231SZi Yan 3479fa6c0231SZi Yan /* FOLL_DUMP to ignore special (like zero) pages */ 348087d2762eSMiaohe Lin page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 3481fa6c0231SZi Yan 3482f7091ed6SHaiyue Wang if (IS_ERR_OR_NULL(page)) 3483fa6c0231SZi Yan continue; 3484fa6c0231SZi Yan 3485a644b0abSMatthew Wilcox (Oracle) folio = page_folio(page); 3486a644b0abSMatthew Wilcox (Oracle) if (!is_transparent_hugepage(folio)) 3487fa6c0231SZi Yan goto next; 3488fa6c0231SZi Yan 34892394aef6SZi Yan if (new_order >= folio_order(folio)) 34902394aef6SZi Yan goto next; 34912394aef6SZi Yan 3492fa6c0231SZi Yan total++; 3493fc4d1823SZi Yan /* 3494fc4d1823SZi Yan * For folios with private, split_huge_page_to_list_to_order() 3495fc4d1823SZi Yan * will try to drop it before split and then check if the folio 3496fc4d1823SZi Yan * can be split or not. So skip the check here. 3497fc4d1823SZi Yan */ 3498fc4d1823SZi Yan if (!folio_test_private(folio) && 3499fc4d1823SZi Yan !can_split_folio(folio, NULL)) 3500fa6c0231SZi Yan goto next; 3501fa6c0231SZi Yan 3502a644b0abSMatthew Wilcox (Oracle) if (!folio_trylock(folio)) 3503fa6c0231SZi Yan goto next; 3504fa6c0231SZi Yan 3505fc4d1823SZi Yan if (!split_folio_to_order(folio, new_order)) 3506fa6c0231SZi Yan split++; 3507fa6c0231SZi Yan 3508a644b0abSMatthew Wilcox (Oracle) folio_unlock(folio); 3509fa6c0231SZi Yan next: 3510a644b0abSMatthew Wilcox (Oracle) folio_put(folio); 3511fa6c0231SZi Yan cond_resched(); 3512fa6c0231SZi Yan } 3513fa6c0231SZi Yan mmap_read_unlock(mm); 3514fa6c0231SZi Yan mmput(mm); 3515fa6c0231SZi Yan 3516fa6c0231SZi Yan pr_debug("%lu of %lu THP split\n", split, total); 3517fa6c0231SZi Yan 3518fa6c0231SZi Yan out: 3519fa6c0231SZi Yan return ret; 3520fa6c0231SZi Yan } 3521fa6c0231SZi Yan 3522fbe37501SZi Yan static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start, 3523fc4d1823SZi Yan pgoff_t off_end, unsigned int new_order) 3524fbe37501SZi Yan { 3525fbe37501SZi Yan struct filename *file; 3526fbe37501SZi Yan struct file *candidate; 3527fbe37501SZi Yan struct address_space *mapping; 3528fbe37501SZi Yan int ret = -EINVAL; 3529fbe37501SZi Yan pgoff_t index; 3530fbe37501SZi Yan int nr_pages = 1; 3531fbe37501SZi Yan unsigned long total = 0, split = 0; 3532fbe37501SZi Yan 3533fbe37501SZi Yan file = getname_kernel(file_path); 3534fbe37501SZi Yan if (IS_ERR(file)) 3535fbe37501SZi Yan return ret; 3536fbe37501SZi Yan 3537fbe37501SZi Yan candidate = file_open_name(file, O_RDONLY, 0); 3538fbe37501SZi Yan if (IS_ERR(candidate)) 3539fbe37501SZi Yan goto out; 3540fbe37501SZi Yan 3541fbe37501SZi Yan pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n", 3542fbe37501SZi Yan file_path, off_start, off_end); 3543fbe37501SZi Yan 3544fbe37501SZi Yan mapping = candidate->f_mapping; 3545fbe37501SZi Yan 3546fbe37501SZi Yan for (index = off_start; index < off_end; index += nr_pages) { 35471fb130b2SChristoph Hellwig struct folio *folio = filemap_get_folio(mapping, index); 3548fbe37501SZi Yan 3549fbe37501SZi Yan nr_pages = 1; 355066dabbb6SChristoph Hellwig if (IS_ERR(folio)) 3551fbe37501SZi Yan continue; 3552fbe37501SZi Yan 35539ee2c086SMatthew Wilcox (Oracle) if (!folio_test_large(folio)) 3554fbe37501SZi Yan goto next; 3555fbe37501SZi Yan 3556fbe37501SZi Yan total++; 35579ee2c086SMatthew Wilcox (Oracle) nr_pages = folio_nr_pages(folio); 3558fbe37501SZi Yan 35592394aef6SZi Yan if (new_order >= folio_order(folio)) 35602394aef6SZi Yan goto next; 35612394aef6SZi Yan 35629ee2c086SMatthew Wilcox (Oracle) if (!folio_trylock(folio)) 3563fbe37501SZi Yan goto next; 3564fbe37501SZi Yan 3565fc4d1823SZi Yan if (!split_folio_to_order(folio, new_order)) 3566fbe37501SZi Yan split++; 3567fbe37501SZi Yan 35689ee2c086SMatthew Wilcox (Oracle) folio_unlock(folio); 3569fbe37501SZi Yan next: 35709ee2c086SMatthew Wilcox (Oracle) folio_put(folio); 3571fbe37501SZi Yan cond_resched(); 3572fbe37501SZi Yan } 3573fbe37501SZi Yan 3574fbe37501SZi Yan filp_close(candidate, NULL); 3575fbe37501SZi Yan ret = 0; 3576fbe37501SZi Yan 3577fbe37501SZi Yan pr_debug("%lu of %lu file-backed THP split\n", split, total); 3578fbe37501SZi Yan out: 3579fbe37501SZi Yan putname(file); 3580fbe37501SZi Yan return ret; 3581fbe37501SZi Yan } 3582fbe37501SZi Yan 3583fa6c0231SZi Yan #define MAX_INPUT_BUF_SZ 255 3584fa6c0231SZi Yan 3585fa6c0231SZi Yan static ssize_t split_huge_pages_write(struct file *file, const char __user *buf, 3586fa6c0231SZi Yan size_t count, loff_t *ppops) 3587fa6c0231SZi Yan { 3588fa6c0231SZi Yan static DEFINE_MUTEX(split_debug_mutex); 3589fa6c0231SZi Yan ssize_t ret; 3590fc4d1823SZi Yan /* 3591fc4d1823SZi Yan * hold pid, start_vaddr, end_vaddr, new_order or 3592fc4d1823SZi Yan * file_path, off_start, off_end, new_order 3593fc4d1823SZi Yan */ 3594fbe37501SZi Yan char input_buf[MAX_INPUT_BUF_SZ]; 3595fa6c0231SZi Yan int pid; 3596fa6c0231SZi Yan unsigned long vaddr_start, vaddr_end; 3597fc4d1823SZi Yan unsigned int new_order = 0; 3598fa6c0231SZi Yan 3599fa6c0231SZi Yan ret = mutex_lock_interruptible(&split_debug_mutex); 3600fa6c0231SZi Yan if (ret) 3601fa6c0231SZi Yan return ret; 3602fa6c0231SZi Yan 3603fa6c0231SZi Yan ret = -EFAULT; 3604fa6c0231SZi Yan 3605fa6c0231SZi Yan memset(input_buf, 0, MAX_INPUT_BUF_SZ); 3606fa6c0231SZi Yan if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ))) 3607fa6c0231SZi Yan goto out; 3608fa6c0231SZi Yan 3609fa6c0231SZi Yan input_buf[MAX_INPUT_BUF_SZ - 1] = '\0'; 3610fbe37501SZi Yan 3611fbe37501SZi Yan if (input_buf[0] == '/') { 3612fbe37501SZi Yan char *tok; 3613fbe37501SZi Yan char *buf = input_buf; 3614fbe37501SZi Yan char file_path[MAX_INPUT_BUF_SZ]; 3615fbe37501SZi Yan pgoff_t off_start = 0, off_end = 0; 3616fbe37501SZi Yan size_t input_len = strlen(input_buf); 3617fbe37501SZi Yan 3618fbe37501SZi Yan tok = strsep(&buf, ","); 3619fbe37501SZi Yan if (tok) { 36201212e00cSMatthew Wilcox (Oracle) strcpy(file_path, tok); 3621fbe37501SZi Yan } else { 3622fbe37501SZi Yan ret = -EINVAL; 3623fbe37501SZi Yan goto out; 3624fbe37501SZi Yan } 3625fbe37501SZi Yan 3626fc4d1823SZi Yan ret = sscanf(buf, "0x%lx,0x%lx,%d", &off_start, &off_end, &new_order); 3627fc4d1823SZi Yan if (ret != 2 && ret != 3) { 3628fbe37501SZi Yan ret = -EINVAL; 3629fbe37501SZi Yan goto out; 3630fbe37501SZi Yan } 3631fc4d1823SZi Yan ret = split_huge_pages_in_file(file_path, off_start, off_end, new_order); 3632fbe37501SZi Yan if (!ret) 3633fbe37501SZi Yan ret = input_len; 3634fbe37501SZi Yan 3635fbe37501SZi Yan goto out; 3636fbe37501SZi Yan } 3637fbe37501SZi Yan 3638fc4d1823SZi Yan ret = sscanf(input_buf, "%d,0x%lx,0x%lx,%d", &pid, &vaddr_start, &vaddr_end, &new_order); 3639fa6c0231SZi Yan if (ret == 1 && pid == 1) { 3640fa6c0231SZi Yan split_huge_pages_all(); 3641fa6c0231SZi Yan ret = strlen(input_buf); 3642fa6c0231SZi Yan goto out; 3643fc4d1823SZi Yan } else if (ret != 3 && ret != 4) { 3644fa6c0231SZi Yan ret = -EINVAL; 3645fa6c0231SZi Yan goto out; 3646fa6c0231SZi Yan } 3647fa6c0231SZi Yan 3648fc4d1823SZi Yan ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end, new_order); 3649fa6c0231SZi Yan if (!ret) 3650fa6c0231SZi Yan ret = strlen(input_buf); 3651fa6c0231SZi Yan out: 3652fa6c0231SZi Yan mutex_unlock(&split_debug_mutex); 3653fa6c0231SZi Yan return ret; 3654fa6c0231SZi Yan 3655fa6c0231SZi Yan } 3656fa6c0231SZi Yan 3657fa6c0231SZi Yan static const struct file_operations split_huge_pages_fops = { 3658fa6c0231SZi Yan .owner = THIS_MODULE, 3659fa6c0231SZi Yan .write = split_huge_pages_write, 3660fa6c0231SZi Yan .llseek = no_llseek, 3661fa6c0231SZi Yan }; 366249071d43SKirill A. Shutemov 366349071d43SKirill A. Shutemov static int __init split_huge_pages_debugfs(void) 366449071d43SKirill A. Shutemov { 3665d9f7979cSGreg Kroah-Hartman debugfs_create_file("split_huge_pages", 0200, NULL, NULL, 366649071d43SKirill A. Shutemov &split_huge_pages_fops); 366749071d43SKirill A. Shutemov return 0; 366849071d43SKirill A. Shutemov } 366949071d43SKirill A. Shutemov late_initcall(split_huge_pages_debugfs); 367049071d43SKirill A. Shutemov #endif 3671616b8371SZi Yan 3672616b8371SZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 36737f5abe60SDavid Hildenbrand int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, 3674616b8371SZi Yan struct page *page) 3675616b8371SZi Yan { 3676a8e61d58SDavid Hildenbrand struct folio *folio = page_folio(page); 3677616b8371SZi Yan struct vm_area_struct *vma = pvmw->vma; 3678616b8371SZi Yan struct mm_struct *mm = vma->vm_mm; 3679616b8371SZi Yan unsigned long address = pvmw->address; 36806c287605SDavid Hildenbrand bool anon_exclusive; 3681616b8371SZi Yan pmd_t pmdval; 3682616b8371SZi Yan swp_entry_t entry; 3683ab6e3d09SNaoya Horiguchi pmd_t pmdswp; 3684616b8371SZi Yan 3685616b8371SZi Yan if (!(pvmw->pmd && !pvmw->pte)) 36867f5abe60SDavid Hildenbrand return 0; 3687616b8371SZi Yan 3688616b8371SZi Yan flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); 36898a8683adSHuang Ying pmdval = pmdp_invalidate(vma, address, pvmw->pmd); 36906c287605SDavid Hildenbrand 3691e3b4b137SDavid Hildenbrand /* See folio_try_share_anon_rmap_pmd(): invalidate PMD first. */ 3692a8e61d58SDavid Hildenbrand anon_exclusive = folio_test_anon(folio) && PageAnonExclusive(page); 3693e3b4b137SDavid Hildenbrand if (anon_exclusive && folio_try_share_anon_rmap_pmd(folio, page)) { 36946c287605SDavid Hildenbrand set_pmd_at(mm, address, pvmw->pmd, pmdval); 36957f5abe60SDavid Hildenbrand return -EBUSY; 36966c287605SDavid Hildenbrand } 36976c287605SDavid Hildenbrand 3698616b8371SZi Yan if (pmd_dirty(pmdval)) 3699db44c658SDavid Hildenbrand folio_mark_dirty(folio); 37004dd845b5SAlistair Popple if (pmd_write(pmdval)) 37014dd845b5SAlistair Popple entry = make_writable_migration_entry(page_to_pfn(page)); 37026c287605SDavid Hildenbrand else if (anon_exclusive) 37036c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry(page_to_pfn(page)); 37044dd845b5SAlistair Popple else 37054dd845b5SAlistair Popple entry = make_readable_migration_entry(page_to_pfn(page)); 37062e346877SPeter Xu if (pmd_young(pmdval)) 37072e346877SPeter Xu entry = make_migration_entry_young(entry); 37082e346877SPeter Xu if (pmd_dirty(pmdval)) 37092e346877SPeter Xu entry = make_migration_entry_dirty(entry); 3710ab6e3d09SNaoya Horiguchi pmdswp = swp_entry_to_pmd(entry); 3711ab6e3d09SNaoya Horiguchi if (pmd_soft_dirty(pmdval)) 3712ab6e3d09SNaoya Horiguchi pmdswp = pmd_swp_mksoft_dirty(pmdswp); 371324bf08c4SDavid Hildenbrand if (pmd_uffd_wp(pmdval)) 371424bf08c4SDavid Hildenbrand pmdswp = pmd_swp_mkuffd_wp(pmdswp); 3715ab6e3d09SNaoya Horiguchi set_pmd_at(mm, address, pvmw->pmd, pmdswp); 3716a8e61d58SDavid Hildenbrand folio_remove_rmap_pmd(folio, page, vma); 3717a8e61d58SDavid Hildenbrand folio_put(folio); 3718283fd6feSAnshuman Khandual trace_set_migration_pmd(address, pmd_val(pmdswp)); 37197f5abe60SDavid Hildenbrand 37207f5abe60SDavid Hildenbrand return 0; 3721616b8371SZi Yan } 3722616b8371SZi Yan 3723616b8371SZi Yan void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) 3724616b8371SZi Yan { 372514d85a6eSDavid Hildenbrand struct folio *folio = page_folio(new); 3726616b8371SZi Yan struct vm_area_struct *vma = pvmw->vma; 3727616b8371SZi Yan struct mm_struct *mm = vma->vm_mm; 3728616b8371SZi Yan unsigned long address = pvmw->address; 37294fba8f2aSMiaohe Lin unsigned long haddr = address & HPAGE_PMD_MASK; 3730616b8371SZi Yan pmd_t pmde; 3731616b8371SZi Yan swp_entry_t entry; 3732616b8371SZi Yan 3733616b8371SZi Yan if (!(pvmw->pmd && !pvmw->pte)) 3734616b8371SZi Yan return; 3735616b8371SZi Yan 3736616b8371SZi Yan entry = pmd_to_swp_entry(*pvmw->pmd); 373714d85a6eSDavid Hildenbrand folio_get(folio); 37382e346877SPeter Xu pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot)); 3739ab6e3d09SNaoya Horiguchi if (pmd_swp_soft_dirty(*pvmw->pmd)) 3740ab6e3d09SNaoya Horiguchi pmde = pmd_mksoft_dirty(pmde); 37413c811f78SDavid Hildenbrand if (is_writable_migration_entry(entry)) 3742161e393cSRick Edgecombe pmde = pmd_mkwrite(pmde, vma); 37438f34f1eaSPeter Xu if (pmd_swp_uffd_wp(*pvmw->pmd)) 3744f1eb1bacSPeter Xu pmde = pmd_mkuffd_wp(pmde); 37452e346877SPeter Xu if (!is_migration_entry_young(entry)) 37462e346877SPeter Xu pmde = pmd_mkold(pmde); 37472e346877SPeter Xu /* NOTE: this may contain setting soft-dirty on some archs */ 374814d85a6eSDavid Hildenbrand if (folio_test_dirty(folio) && is_migration_entry_dirty(entry)) 37492e346877SPeter Xu pmde = pmd_mkdirty(pmde); 3750616b8371SZi Yan 375114d85a6eSDavid Hildenbrand if (folio_test_anon(folio)) { 3752395db7b1SDavid Hildenbrand rmap_t rmap_flags = RMAP_NONE; 37536c287605SDavid Hildenbrand 37546c287605SDavid Hildenbrand if (!is_readable_migration_entry(entry)) 37556c287605SDavid Hildenbrand rmap_flags |= RMAP_EXCLUSIVE; 37566c287605SDavid Hildenbrand 3757395db7b1SDavid Hildenbrand folio_add_anon_rmap_pmd(folio, new, vma, haddr, rmap_flags); 37586c287605SDavid Hildenbrand } else { 375914d85a6eSDavid Hildenbrand folio_add_file_rmap_pmd(folio, new, vma); 37606c287605SDavid Hildenbrand } 376114d85a6eSDavid Hildenbrand VM_BUG_ON(pmd_write(pmde) && folio_test_anon(folio) && !PageAnonExclusive(new)); 37624fba8f2aSMiaohe Lin set_pmd_at(mm, haddr, pvmw->pmd, pmde); 37635cbcf225SMuchun Song 37645cbcf225SMuchun Song /* No need to invalidate - it was non-present before */ 3765616b8371SZi Yan update_mmu_cache_pmd(vma, address, pvmw->pmd); 3766283fd6feSAnshuman Khandual trace_remove_migration_pmd(address, pmd_val(pmde)); 3767616b8371SZi Yan } 3768616b8371SZi Yan #endif 3769