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 { 7938991de90SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio); 7948991de90SMatthew Wilcox (Oracle) INIT_LIST_HEAD(&folio->_deferred_list); 795de53c05fSMatthew Wilcox (Oracle) folio_set_large_rmappable(folio); 7969a982250SKirill A. Shutemov } 7979a982250SKirill A. Shutemov 798a644b0abSMatthew Wilcox (Oracle) static inline bool is_transparent_hugepage(struct folio *folio) 799005ba37cSSean Christopherson { 800a644b0abSMatthew Wilcox (Oracle) if (!folio_test_large(folio)) 801fa1f68ccSZou Wei return false; 802005ba37cSSean Christopherson 803f04029f3SMatthew Wilcox (Oracle) return is_huge_zero_page(&folio->page) || 804de53c05fSMatthew Wilcox (Oracle) folio_test_large_rmappable(folio); 805005ba37cSSean Christopherson } 806005ba37cSSean Christopherson 80797d3d0f9SKirill A. Shutemov static unsigned long __thp_get_unmapped_area(struct file *filp, 80897d3d0f9SKirill A. Shutemov unsigned long addr, unsigned long len, 80974d2fad1SToshi Kani loff_t off, unsigned long flags, unsigned long size) 81074d2fad1SToshi Kani { 81174d2fad1SToshi Kani loff_t off_end = off + len; 81274d2fad1SToshi Kani loff_t off_align = round_up(off, size); 813*96204e15SRyan Roberts unsigned long len_pad, ret, off_sub; 81474d2fad1SToshi Kani 8154ef9ad19SYang Shi if (IS_ENABLED(CONFIG_32BIT) || in_compat_syscall()) 8164ef9ad19SYang Shi return 0; 8174ef9ad19SYang Shi 81874d2fad1SToshi Kani if (off_end <= off_align || (off_end - off_align) < size) 81974d2fad1SToshi Kani return 0; 82074d2fad1SToshi Kani 82174d2fad1SToshi Kani len_pad = len + size; 82274d2fad1SToshi Kani if (len_pad < len || (off + len_pad) < off) 82374d2fad1SToshi Kani return 0; 82474d2fad1SToshi Kani 82597d3d0f9SKirill A. Shutemov ret = current->mm->get_unmapped_area(filp, addr, len_pad, 82674d2fad1SToshi Kani off >> PAGE_SHIFT, flags); 82797d3d0f9SKirill A. Shutemov 82897d3d0f9SKirill A. Shutemov /* 82997d3d0f9SKirill A. Shutemov * The failure might be due to length padding. The caller will retry 83097d3d0f9SKirill A. Shutemov * without the padding. 83197d3d0f9SKirill A. Shutemov */ 83297d3d0f9SKirill A. Shutemov if (IS_ERR_VALUE(ret)) 83374d2fad1SToshi Kani return 0; 83474d2fad1SToshi Kani 83597d3d0f9SKirill A. Shutemov /* 83697d3d0f9SKirill A. Shutemov * Do not try to align to THP boundary if allocation at the address 83797d3d0f9SKirill A. Shutemov * hint succeeds. 83897d3d0f9SKirill A. Shutemov */ 83997d3d0f9SKirill A. Shutemov if (ret == addr) 84074d2fad1SToshi Kani return addr; 84197d3d0f9SKirill A. Shutemov 842*96204e15SRyan Roberts off_sub = (off - ret) & (size - 1); 843*96204e15SRyan Roberts 844*96204e15SRyan Roberts if (current->mm->get_unmapped_area == arch_get_unmapped_area_topdown && 845*96204e15SRyan Roberts !off_sub) 846*96204e15SRyan Roberts return ret + size; 847*96204e15SRyan Roberts 848*96204e15SRyan Roberts ret += off_sub; 84997d3d0f9SKirill A. Shutemov return ret; 85074d2fad1SToshi Kani } 85174d2fad1SToshi Kani 85274d2fad1SToshi Kani unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, 85374d2fad1SToshi Kani unsigned long len, unsigned long pgoff, unsigned long flags) 85474d2fad1SToshi Kani { 85597d3d0f9SKirill A. Shutemov unsigned long ret; 85674d2fad1SToshi Kani loff_t off = (loff_t)pgoff << PAGE_SHIFT; 85774d2fad1SToshi Kani 85897d3d0f9SKirill A. Shutemov ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE); 85997d3d0f9SKirill A. Shutemov if (ret) 86097d3d0f9SKirill A. Shutemov return ret; 8611854bc6eSWilliam Kucharski 86274d2fad1SToshi Kani return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); 86374d2fad1SToshi Kani } 86474d2fad1SToshi Kani EXPORT_SYMBOL_GPL(thp_get_unmapped_area); 86574d2fad1SToshi Kani 8662b740303SSouptick Joarder static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf, 8672b740303SSouptick Joarder struct page *page, gfp_t gfp) 86871e3aac0SAndrea Arcangeli { 86982b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 870cfe3236dSKefeng Wang struct folio *folio = page_folio(page); 87171e3aac0SAndrea Arcangeli pgtable_t pgtable; 87282b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 8732b740303SSouptick Joarder vm_fault_t ret = 0; 87471e3aac0SAndrea Arcangeli 875cfe3236dSKefeng Wang VM_BUG_ON_FOLIO(!folio_test_large(folio), folio); 87600501b53SJohannes Weiner 877cfe3236dSKefeng Wang if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) { 878cfe3236dSKefeng Wang folio_put(folio); 8796b251fc9SAndrea Arcangeli count_vm_event(THP_FAULT_FALLBACK); 88085b9f46eSDavid Rientjes count_vm_event(THP_FAULT_FALLBACK_CHARGE); 8816b251fc9SAndrea Arcangeli return VM_FAULT_FALLBACK; 8826b251fc9SAndrea Arcangeli } 883cfe3236dSKefeng Wang folio_throttle_swaprate(folio, gfp); 88471e3aac0SAndrea Arcangeli 8854cf58924SJoel Fernandes (Google) pgtable = pte_alloc_one(vma->vm_mm); 88600501b53SJohannes Weiner if (unlikely(!pgtable)) { 8876b31d595SMichal Hocko ret = VM_FAULT_OOM; 8886b31d595SMichal Hocko goto release; 88900501b53SJohannes Weiner } 89000501b53SJohannes Weiner 891c79b57e4SHuang Ying clear_huge_page(page, vmf->address, HPAGE_PMD_NR); 89252f37629SMinchan Kim /* 893cfe3236dSKefeng Wang * The memory barrier inside __folio_mark_uptodate makes sure that 89452f37629SMinchan Kim * clear_huge_page writes become visible before the set_pmd_at() 89552f37629SMinchan Kim * write. 89652f37629SMinchan Kim */ 897cfe3236dSKefeng Wang __folio_mark_uptodate(folio); 89871e3aac0SAndrea Arcangeli 89982b0f8c3SJan Kara vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 90082b0f8c3SJan Kara if (unlikely(!pmd_none(*vmf->pmd))) { 9016b31d595SMichal Hocko goto unlock_release; 90271e3aac0SAndrea Arcangeli } else { 90371e3aac0SAndrea Arcangeli pmd_t entry; 9046b251fc9SAndrea Arcangeli 9056b31d595SMichal Hocko ret = check_stable_address_space(vma->vm_mm); 9066b31d595SMichal Hocko if (ret) 9076b31d595SMichal Hocko goto unlock_release; 9086b31d595SMichal Hocko 9096b251fc9SAndrea Arcangeli /* Deliver the page fault to userland */ 9106b251fc9SAndrea Arcangeli if (userfaultfd_missing(vma)) { 91182b0f8c3SJan Kara spin_unlock(vmf->ptl); 912cfe3236dSKefeng Wang folio_put(folio); 913bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 9148fd5eda4SMiaohe Lin ret = handle_userfault(vmf, VM_UFFD_MISSING); 9158fd5eda4SMiaohe Lin VM_BUG_ON(ret & VM_FAULT_FALLBACK); 9168fd5eda4SMiaohe Lin return ret; 9176b251fc9SAndrea Arcangeli } 9186b251fc9SAndrea Arcangeli 9193122359aSKirill A. Shutemov entry = mk_huge_pmd(page, vma->vm_page_prot); 920f55e1014SLinus Torvalds entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 921cfe3236dSKefeng Wang folio_add_new_anon_rmap(folio, vma, haddr); 922cfe3236dSKefeng Wang folio_add_lru_vma(folio, vma); 92382b0f8c3SJan Kara pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); 92482b0f8c3SJan Kara set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 925fca40573SBibo Mao update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 926bae473a4SKirill A. Shutemov add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); 927c4812909SKirill A. Shutemov mm_inc_nr_ptes(vma->vm_mm); 92882b0f8c3SJan Kara spin_unlock(vmf->ptl); 9296b251fc9SAndrea Arcangeli count_vm_event(THP_FAULT_ALLOC); 9309d82c694SJohannes Weiner count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC); 93171e3aac0SAndrea Arcangeli } 93271e3aac0SAndrea Arcangeli 933aa2e878eSDavid Rientjes return 0; 9346b31d595SMichal Hocko unlock_release: 9356b31d595SMichal Hocko spin_unlock(vmf->ptl); 9366b31d595SMichal Hocko release: 9376b31d595SMichal Hocko if (pgtable) 9386b31d595SMichal Hocko pte_free(vma->vm_mm, pgtable); 939cfe3236dSKefeng Wang folio_put(folio); 9406b31d595SMichal Hocko return ret; 9416b31d595SMichal Hocko 94271e3aac0SAndrea Arcangeli } 94371e3aac0SAndrea Arcangeli 944444eb2a4SMel Gorman /* 94521440d7eSDavid Rientjes * always: directly stall for all thp allocations 94621440d7eSDavid Rientjes * defer: wake kswapd and fail if not immediately available 94721440d7eSDavid Rientjes * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise 94821440d7eSDavid Rientjes * fail if not immediately available 94921440d7eSDavid Rientjes * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately 95021440d7eSDavid Rientjes * available 95121440d7eSDavid Rientjes * never: never stall for any thp allocation 952444eb2a4SMel Gorman */ 953164cc4feSRik van Riel gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma) 9540bbbc0b3SAndrea Arcangeli { 955164cc4feSRik van Riel const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE); 95689c83fb5SMichal Hocko 957ac79f78dSDavid Rientjes /* Always do synchronous compaction */ 95821440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) 959a8282608SAndrea Arcangeli return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); 960ac79f78dSDavid Rientjes 961ac79f78dSDavid Rientjes /* Kick kcompactd and fail quickly */ 96221440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) 96319deb769SDavid Rientjes return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; 964ac79f78dSDavid Rientjes 965ac79f78dSDavid Rientjes /* Synchronous compaction if madvised, otherwise kick kcompactd */ 96621440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) 96719deb769SDavid Rientjes return GFP_TRANSHUGE_LIGHT | 96819deb769SDavid Rientjes (vma_madvised ? __GFP_DIRECT_RECLAIM : 969ac79f78dSDavid Rientjes __GFP_KSWAPD_RECLAIM); 970ac79f78dSDavid Rientjes 971ac79f78dSDavid Rientjes /* Only do synchronous compaction if madvised */ 97221440d7eSDavid Rientjes if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) 97319deb769SDavid Rientjes return GFP_TRANSHUGE_LIGHT | 97419deb769SDavid Rientjes (vma_madvised ? __GFP_DIRECT_RECLAIM : 0); 975ac79f78dSDavid Rientjes 97619deb769SDavid Rientjes return GFP_TRANSHUGE_LIGHT; 977444eb2a4SMel Gorman } 978444eb2a4SMel Gorman 979c4088ebdSKirill A. Shutemov /* Caller must hold page table lock. */ 9802efeb8daSMiaohe Lin static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 98197ae1749SKirill A. Shutemov struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 9825918d10aSKirill A. Shutemov struct page *zero_page) 983fc9fe822SKirill A. Shutemov { 984fc9fe822SKirill A. Shutemov pmd_t entry; 9857c414164SAndrew Morton if (!pmd_none(*pmd)) 9862efeb8daSMiaohe Lin return; 9875918d10aSKirill A. Shutemov entry = mk_pmd(zero_page, vma->vm_page_prot); 988fc9fe822SKirill A. Shutemov entry = pmd_mkhuge(entry); 9896b0b50b0SAneesh Kumar K.V pgtable_trans_huge_deposit(mm, pmd, pgtable); 990fc9fe822SKirill A. Shutemov set_pmd_at(mm, haddr, pmd, entry); 991c4812909SKirill A. Shutemov mm_inc_nr_ptes(mm); 992fc9fe822SKirill A. Shutemov } 993fc9fe822SKirill A. Shutemov 9942b740303SSouptick Joarder vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf) 99571e3aac0SAndrea Arcangeli { 99682b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 997077fcf11SAneesh Kumar K.V gfp_t gfp; 998cb196ee1SMatthew Wilcox (Oracle) struct folio *folio; 99982b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 100071e3aac0SAndrea Arcangeli 10013485b883SRyan Roberts if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER)) 1002c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 100371e3aac0SAndrea Arcangeli if (unlikely(anon_vma_prepare(vma))) 100471e3aac0SAndrea Arcangeli return VM_FAULT_OOM; 10054fa6893fSYang Shi khugepaged_enter_vma(vma, vma->vm_flags); 1006d2081b2bSYang Shi 100782b0f8c3SJan Kara if (!(vmf->flags & FAULT_FLAG_WRITE) && 1008bae473a4SKirill A. Shutemov !mm_forbids_zeropage(vma->vm_mm) && 100979da5407SKirill A. Shutemov transparent_hugepage_use_zero_page()) { 101080371957SKirill A. Shutemov pgtable_t pgtable; 10115918d10aSKirill A. Shutemov struct page *zero_page; 10122b740303SSouptick Joarder vm_fault_t ret; 10134cf58924SJoel Fernandes (Google) pgtable = pte_alloc_one(vma->vm_mm); 101480371957SKirill A. Shutemov if (unlikely(!pgtable)) 101580371957SKirill A. Shutemov return VM_FAULT_OOM; 10166fcb52a5SAaron Lu zero_page = mm_get_huge_zero_page(vma->vm_mm); 10175918d10aSKirill A. Shutemov if (unlikely(!zero_page)) { 1018bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 101997ae1749SKirill A. Shutemov count_vm_event(THP_FAULT_FALLBACK); 1020c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 102197ae1749SKirill A. Shutemov } 102282b0f8c3SJan Kara vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 10236b251fc9SAndrea Arcangeli ret = 0; 102482b0f8c3SJan Kara if (pmd_none(*vmf->pmd)) { 10256b31d595SMichal Hocko ret = check_stable_address_space(vma->vm_mm); 10266b31d595SMichal Hocko if (ret) { 10276b31d595SMichal Hocko spin_unlock(vmf->ptl); 1028bfe8cc1dSGerald Schaefer pte_free(vma->vm_mm, pgtable); 10296b31d595SMichal Hocko } else if (userfaultfd_missing(vma)) { 103082b0f8c3SJan Kara spin_unlock(vmf->ptl); 1031bfe8cc1dSGerald Schaefer pte_free(vma->vm_mm, pgtable); 103282b0f8c3SJan Kara ret = handle_userfault(vmf, VM_UFFD_MISSING); 10336b251fc9SAndrea Arcangeli VM_BUG_ON(ret & VM_FAULT_FALLBACK); 10346b251fc9SAndrea Arcangeli } else { 1035bae473a4SKirill A. Shutemov set_huge_zero_page(pgtable, vma->vm_mm, vma, 103682b0f8c3SJan Kara haddr, vmf->pmd, zero_page); 1037fca40573SBibo Mao update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 103882b0f8c3SJan Kara spin_unlock(vmf->ptl); 10396b251fc9SAndrea Arcangeli } 1040bfe8cc1dSGerald Schaefer } else { 104182b0f8c3SJan Kara spin_unlock(vmf->ptl); 1042bae473a4SKirill A. Shutemov pte_free(vma->vm_mm, pgtable); 1043bfe8cc1dSGerald Schaefer } 10446b251fc9SAndrea Arcangeli return ret; 104580371957SKirill A. Shutemov } 1046164cc4feSRik van Riel gfp = vma_thp_gfp_mask(vma); 1047cb196ee1SMatthew Wilcox (Oracle) folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true); 1048cb196ee1SMatthew Wilcox (Oracle) if (unlikely(!folio)) { 104981ab4201SAndi Kleen count_vm_event(THP_FAULT_FALLBACK); 1050c0292554SKirill A. Shutemov return VM_FAULT_FALLBACK; 105181ab4201SAndi Kleen } 1052cb196ee1SMatthew Wilcox (Oracle) return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp); 105371e3aac0SAndrea Arcangeli } 105471e3aac0SAndrea Arcangeli 1055ae18d6dcSMatthew Wilcox static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, 10563b6521f5SOliver O'Halloran pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, 10573b6521f5SOliver O'Halloran pgtable_t pgtable) 10585cad465dSMatthew Wilcox { 10595cad465dSMatthew Wilcox struct mm_struct *mm = vma->vm_mm; 10605cad465dSMatthew Wilcox pmd_t entry; 10615cad465dSMatthew Wilcox spinlock_t *ptl; 10625cad465dSMatthew Wilcox 10635cad465dSMatthew Wilcox ptl = pmd_lock(mm, pmd); 1064c6f3c5eeSAneesh Kumar K.V if (!pmd_none(*pmd)) { 1065c6f3c5eeSAneesh Kumar K.V if (write) { 1066c6f3c5eeSAneesh Kumar K.V if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) { 1067c6f3c5eeSAneesh Kumar K.V WARN_ON_ONCE(!is_huge_zero_pmd(*pmd)); 1068c6f3c5eeSAneesh Kumar K.V goto out_unlock; 1069c6f3c5eeSAneesh Kumar K.V } 1070c6f3c5eeSAneesh Kumar K.V entry = pmd_mkyoung(*pmd); 1071c6f3c5eeSAneesh Kumar K.V entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 1072c6f3c5eeSAneesh Kumar K.V if (pmdp_set_access_flags(vma, addr, pmd, entry, 1)) 1073c6f3c5eeSAneesh Kumar K.V update_mmu_cache_pmd(vma, addr, pmd); 1074c6f3c5eeSAneesh Kumar K.V } 1075c6f3c5eeSAneesh Kumar K.V 1076c6f3c5eeSAneesh Kumar K.V goto out_unlock; 1077c6f3c5eeSAneesh Kumar K.V } 1078c6f3c5eeSAneesh Kumar K.V 1079f25748e3SDan Williams entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); 1080f25748e3SDan Williams if (pfn_t_devmap(pfn)) 1081f25748e3SDan Williams entry = pmd_mkdevmap(entry); 10825cad465dSMatthew Wilcox if (write) { 1083f55e1014SLinus Torvalds entry = pmd_mkyoung(pmd_mkdirty(entry)); 1084f55e1014SLinus Torvalds entry = maybe_pmd_mkwrite(entry, vma); 10855cad465dSMatthew Wilcox } 10863b6521f5SOliver O'Halloran 10873b6521f5SOliver O'Halloran if (pgtable) { 10883b6521f5SOliver O'Halloran pgtable_trans_huge_deposit(mm, pmd, pgtable); 1089c4812909SKirill A. Shutemov mm_inc_nr_ptes(mm); 1090c6f3c5eeSAneesh Kumar K.V pgtable = NULL; 10913b6521f5SOliver O'Halloran } 10923b6521f5SOliver O'Halloran 10935cad465dSMatthew Wilcox set_pmd_at(mm, addr, pmd, entry); 10945cad465dSMatthew Wilcox update_mmu_cache_pmd(vma, addr, pmd); 1095c6f3c5eeSAneesh Kumar K.V 1096c6f3c5eeSAneesh Kumar K.V out_unlock: 10975cad465dSMatthew Wilcox spin_unlock(ptl); 1098c6f3c5eeSAneesh Kumar K.V if (pgtable) 1099c6f3c5eeSAneesh Kumar K.V pte_free(mm, pgtable); 11005cad465dSMatthew Wilcox } 11015cad465dSMatthew Wilcox 11029a9731b1SThomas Hellstrom (VMware) /** 11037b806d22SLorenzo Stoakes * vmf_insert_pfn_pmd - insert a pmd size pfn 11049a9731b1SThomas Hellstrom (VMware) * @vmf: Structure describing the fault 11059a9731b1SThomas Hellstrom (VMware) * @pfn: pfn to insert 11069a9731b1SThomas Hellstrom (VMware) * @write: whether it's a write fault 11079a9731b1SThomas Hellstrom (VMware) * 11087b806d22SLorenzo Stoakes * Insert a pmd size pfn. See vmf_insert_pfn() for additional info. 11099a9731b1SThomas Hellstrom (VMware) * 11109a9731b1SThomas Hellstrom (VMware) * Return: vm_fault_t value. 11119a9731b1SThomas Hellstrom (VMware) */ 11127b806d22SLorenzo Stoakes vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write) 11135cad465dSMatthew Wilcox { 1114fce86ff5SDan Williams unsigned long addr = vmf->address & PMD_MASK; 1115fce86ff5SDan Williams struct vm_area_struct *vma = vmf->vma; 11167b806d22SLorenzo Stoakes pgprot_t pgprot = vma->vm_page_prot; 11173b6521f5SOliver O'Halloran pgtable_t pgtable = NULL; 1118fce86ff5SDan Williams 11195cad465dSMatthew Wilcox /* 11205cad465dSMatthew Wilcox * If we had pmd_special, we could avoid all these restrictions, 11215cad465dSMatthew Wilcox * but we need to be consistent with PTEs and architectures that 11225cad465dSMatthew Wilcox * can't support a 'special' bit. 11235cad465dSMatthew Wilcox */ 1124e1fb4a08SDave Jiang BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 1125e1fb4a08SDave Jiang !pfn_t_devmap(pfn)); 11265cad465dSMatthew Wilcox BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 11275cad465dSMatthew Wilcox (VM_PFNMAP|VM_MIXEDMAP)); 11285cad465dSMatthew Wilcox BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 11295cad465dSMatthew Wilcox 11305cad465dSMatthew Wilcox if (addr < vma->vm_start || addr >= vma->vm_end) 11315cad465dSMatthew Wilcox return VM_FAULT_SIGBUS; 1132308a047cSBorislav Petkov 11333b6521f5SOliver O'Halloran if (arch_needs_pgtable_deposit()) { 11344cf58924SJoel Fernandes (Google) pgtable = pte_alloc_one(vma->vm_mm); 11353b6521f5SOliver O'Halloran if (!pgtable) 11363b6521f5SOliver O'Halloran return VM_FAULT_OOM; 11373b6521f5SOliver O'Halloran } 11383b6521f5SOliver O'Halloran 1139308a047cSBorislav Petkov track_pfn_insert(vma, &pgprot, pfn); 1140308a047cSBorislav Petkov 1141fce86ff5SDan Williams insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable); 1142ae18d6dcSMatthew Wilcox return VM_FAULT_NOPAGE; 11435cad465dSMatthew Wilcox } 11447b806d22SLorenzo Stoakes EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); 11455cad465dSMatthew Wilcox 1146a00cc7d9SMatthew Wilcox #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1147f55e1014SLinus Torvalds static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) 1148a00cc7d9SMatthew Wilcox { 1149f55e1014SLinus Torvalds if (likely(vma->vm_flags & VM_WRITE)) 1150a00cc7d9SMatthew Wilcox pud = pud_mkwrite(pud); 1151a00cc7d9SMatthew Wilcox return pud; 1152a00cc7d9SMatthew Wilcox } 1153a00cc7d9SMatthew Wilcox 1154a00cc7d9SMatthew Wilcox static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, 11557b806d22SLorenzo Stoakes pud_t *pud, pfn_t pfn, bool write) 1156a00cc7d9SMatthew Wilcox { 1157a00cc7d9SMatthew Wilcox struct mm_struct *mm = vma->vm_mm; 11587b806d22SLorenzo Stoakes pgprot_t prot = vma->vm_page_prot; 1159a00cc7d9SMatthew Wilcox pud_t entry; 1160a00cc7d9SMatthew Wilcox spinlock_t *ptl; 1161a00cc7d9SMatthew Wilcox 1162a00cc7d9SMatthew Wilcox ptl = pud_lock(mm, pud); 1163c6f3c5eeSAneesh Kumar K.V if (!pud_none(*pud)) { 1164c6f3c5eeSAneesh Kumar K.V if (write) { 1165c6f3c5eeSAneesh Kumar K.V if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) { 1166c6f3c5eeSAneesh Kumar K.V WARN_ON_ONCE(!is_huge_zero_pud(*pud)); 1167c6f3c5eeSAneesh Kumar K.V goto out_unlock; 1168c6f3c5eeSAneesh Kumar K.V } 1169c6f3c5eeSAneesh Kumar K.V entry = pud_mkyoung(*pud); 1170c6f3c5eeSAneesh Kumar K.V entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma); 1171c6f3c5eeSAneesh Kumar K.V if (pudp_set_access_flags(vma, addr, pud, entry, 1)) 1172c6f3c5eeSAneesh Kumar K.V update_mmu_cache_pud(vma, addr, pud); 1173c6f3c5eeSAneesh Kumar K.V } 1174c6f3c5eeSAneesh Kumar K.V goto out_unlock; 1175c6f3c5eeSAneesh Kumar K.V } 1176c6f3c5eeSAneesh Kumar K.V 1177a00cc7d9SMatthew Wilcox entry = pud_mkhuge(pfn_t_pud(pfn, prot)); 1178a00cc7d9SMatthew Wilcox if (pfn_t_devmap(pfn)) 1179a00cc7d9SMatthew Wilcox entry = pud_mkdevmap(entry); 1180a00cc7d9SMatthew Wilcox if (write) { 1181f55e1014SLinus Torvalds entry = pud_mkyoung(pud_mkdirty(entry)); 1182f55e1014SLinus Torvalds entry = maybe_pud_mkwrite(entry, vma); 1183a00cc7d9SMatthew Wilcox } 1184a00cc7d9SMatthew Wilcox set_pud_at(mm, addr, pud, entry); 1185a00cc7d9SMatthew Wilcox update_mmu_cache_pud(vma, addr, pud); 1186c6f3c5eeSAneesh Kumar K.V 1187c6f3c5eeSAneesh Kumar K.V out_unlock: 1188a00cc7d9SMatthew Wilcox spin_unlock(ptl); 1189a00cc7d9SMatthew Wilcox } 1190a00cc7d9SMatthew Wilcox 11919a9731b1SThomas Hellstrom (VMware) /** 11927b806d22SLorenzo Stoakes * vmf_insert_pfn_pud - insert a pud size pfn 11939a9731b1SThomas Hellstrom (VMware) * @vmf: Structure describing the fault 11949a9731b1SThomas Hellstrom (VMware) * @pfn: pfn to insert 11959a9731b1SThomas Hellstrom (VMware) * @write: whether it's a write fault 11969a9731b1SThomas Hellstrom (VMware) * 11977b806d22SLorenzo Stoakes * Insert a pud size pfn. See vmf_insert_pfn() for additional info. 11989a9731b1SThomas Hellstrom (VMware) * 11999a9731b1SThomas Hellstrom (VMware) * Return: vm_fault_t value. 12009a9731b1SThomas Hellstrom (VMware) */ 12017b806d22SLorenzo Stoakes vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write) 1202a00cc7d9SMatthew Wilcox { 1203fce86ff5SDan Williams unsigned long addr = vmf->address & PUD_MASK; 1204fce86ff5SDan Williams struct vm_area_struct *vma = vmf->vma; 12057b806d22SLorenzo Stoakes pgprot_t pgprot = vma->vm_page_prot; 1206fce86ff5SDan Williams 1207a00cc7d9SMatthew Wilcox /* 1208a00cc7d9SMatthew Wilcox * If we had pud_special, we could avoid all these restrictions, 1209a00cc7d9SMatthew Wilcox * but we need to be consistent with PTEs and architectures that 1210a00cc7d9SMatthew Wilcox * can't support a 'special' bit. 1211a00cc7d9SMatthew Wilcox */ 121262ec0d8cSDave Jiang BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 121362ec0d8cSDave Jiang !pfn_t_devmap(pfn)); 1214a00cc7d9SMatthew Wilcox BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 1215a00cc7d9SMatthew Wilcox (VM_PFNMAP|VM_MIXEDMAP)); 1216a00cc7d9SMatthew Wilcox BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 1217a00cc7d9SMatthew Wilcox 1218a00cc7d9SMatthew Wilcox if (addr < vma->vm_start || addr >= vma->vm_end) 1219a00cc7d9SMatthew Wilcox return VM_FAULT_SIGBUS; 1220a00cc7d9SMatthew Wilcox 1221a00cc7d9SMatthew Wilcox track_pfn_insert(vma, &pgprot, pfn); 1222a00cc7d9SMatthew Wilcox 12237b806d22SLorenzo Stoakes insert_pfn_pud(vma, addr, vmf->pud, pfn, write); 1224a00cc7d9SMatthew Wilcox return VM_FAULT_NOPAGE; 1225a00cc7d9SMatthew Wilcox } 12267b806d22SLorenzo Stoakes EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud); 1227a00cc7d9SMatthew Wilcox #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1228a00cc7d9SMatthew Wilcox 12293565fce3SDan Williams static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, 1230a69e4717SMiaohe Lin pmd_t *pmd, bool write) 12313565fce3SDan Williams { 12323565fce3SDan Williams pmd_t _pmd; 12333565fce3SDan Williams 1234a8f97366SKirill A. Shutemov _pmd = pmd_mkyoung(*pmd); 1235a69e4717SMiaohe Lin if (write) 1236a8f97366SKirill A. Shutemov _pmd = pmd_mkdirty(_pmd); 12373565fce3SDan Williams if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, 1238a69e4717SMiaohe Lin pmd, _pmd, write)) 12393565fce3SDan Williams update_mmu_cache_pmd(vma, addr, pmd); 12403565fce3SDan Williams } 12413565fce3SDan Williams 12423565fce3SDan Williams struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, 1243df06b37fSKeith Busch pmd_t *pmd, int flags, struct dev_pagemap **pgmap) 12443565fce3SDan Williams { 12453565fce3SDan Williams unsigned long pfn = pmd_pfn(*pmd); 12463565fce3SDan Williams struct mm_struct *mm = vma->vm_mm; 12473565fce3SDan Williams struct page *page; 12480f089235SLogan Gunthorpe int ret; 12493565fce3SDan Williams 12503565fce3SDan Williams assert_spin_locked(pmd_lockptr(mm, pmd)); 12513565fce3SDan Williams 1252f6f37321SLinus Torvalds if (flags & FOLL_WRITE && !pmd_write(*pmd)) 12533565fce3SDan Williams return NULL; 12543565fce3SDan Williams 12553565fce3SDan Williams if (pmd_present(*pmd) && pmd_devmap(*pmd)) 12563565fce3SDan Williams /* pass */; 12573565fce3SDan Williams else 12583565fce3SDan Williams return NULL; 12593565fce3SDan Williams 12603565fce3SDan Williams if (flags & FOLL_TOUCH) 1261a69e4717SMiaohe Lin touch_pmd(vma, addr, pmd, flags & FOLL_WRITE); 12623565fce3SDan Williams 12633565fce3SDan Williams /* 12643565fce3SDan Williams * device mapped pages can only be returned if the 12653565fce3SDan Williams * caller will manage the page reference count. 12663565fce3SDan Williams */ 12673faa52c0SJohn Hubbard if (!(flags & (FOLL_GET | FOLL_PIN))) 12683565fce3SDan Williams return ERR_PTR(-EEXIST); 12693565fce3SDan Williams 12703565fce3SDan Williams pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; 1271df06b37fSKeith Busch *pgmap = get_dev_pagemap(pfn, *pgmap); 1272df06b37fSKeith Busch if (!*pgmap) 12733565fce3SDan Williams return ERR_PTR(-EFAULT); 12743565fce3SDan Williams page = pfn_to_page(pfn); 12750f089235SLogan Gunthorpe ret = try_grab_page(page, flags); 12760f089235SLogan Gunthorpe if (ret) 12770f089235SLogan Gunthorpe page = ERR_PTR(ret); 12783565fce3SDan Williams 12793565fce3SDan Williams return page; 12803565fce3SDan Williams } 12813565fce3SDan Williams 128271e3aac0SAndrea Arcangeli int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, 128371e3aac0SAndrea Arcangeli pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 12848f34f1eaSPeter Xu struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma) 128571e3aac0SAndrea Arcangeli { 1286c4088ebdSKirill A. Shutemov spinlock_t *dst_ptl, *src_ptl; 128771e3aac0SAndrea Arcangeli struct page *src_page; 128896c772c2SDavid Hildenbrand struct folio *src_folio; 128971e3aac0SAndrea Arcangeli pmd_t pmd; 129012c9d70bSMatthew Wilcox pgtable_t pgtable = NULL; 1291628d47ceSKirill A. Shutemov int ret = -ENOMEM; 129271e3aac0SAndrea Arcangeli 1293628d47ceSKirill A. Shutemov /* Skip if can be re-fill on fault */ 12948f34f1eaSPeter Xu if (!vma_is_anonymous(dst_vma)) 1295628d47ceSKirill A. Shutemov return 0; 1296628d47ceSKirill A. Shutemov 12974cf58924SJoel Fernandes (Google) pgtable = pte_alloc_one(dst_mm); 129871e3aac0SAndrea Arcangeli if (unlikely(!pgtable)) 129971e3aac0SAndrea Arcangeli goto out; 130071e3aac0SAndrea Arcangeli 1301c4088ebdSKirill A. Shutemov dst_ptl = pmd_lock(dst_mm, dst_pmd); 1302c4088ebdSKirill A. Shutemov src_ptl = pmd_lockptr(src_mm, src_pmd); 1303c4088ebdSKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 130471e3aac0SAndrea Arcangeli 130571e3aac0SAndrea Arcangeli ret = -EAGAIN; 130671e3aac0SAndrea Arcangeli pmd = *src_pmd; 130784c3fc4eSZi Yan 130884c3fc4eSZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 130984c3fc4eSZi Yan if (unlikely(is_swap_pmd(pmd))) { 131084c3fc4eSZi Yan swp_entry_t entry = pmd_to_swp_entry(pmd); 131184c3fc4eSZi Yan 131284c3fc4eSZi Yan VM_BUG_ON(!is_pmd_migration_entry(pmd)); 13136c287605SDavid Hildenbrand if (!is_readable_migration_entry(entry)) { 13144dd845b5SAlistair Popple entry = make_readable_migration_entry( 13154dd845b5SAlistair Popple swp_offset(entry)); 131684c3fc4eSZi Yan pmd = swp_entry_to_pmd(entry); 1317ab6e3d09SNaoya Horiguchi if (pmd_swp_soft_dirty(*src_pmd)) 1318ab6e3d09SNaoya Horiguchi pmd = pmd_swp_mksoft_dirty(pmd); 13198f34f1eaSPeter Xu if (pmd_swp_uffd_wp(*src_pmd)) 13208f34f1eaSPeter Xu pmd = pmd_swp_mkuffd_wp(pmd); 132184c3fc4eSZi Yan set_pmd_at(src_mm, addr, src_pmd, pmd); 132284c3fc4eSZi Yan } 1323dd8a67f9SZi Yan add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 1324af5b0f6aSKirill A. Shutemov mm_inc_nr_ptes(dst_mm); 1325dd8a67f9SZi Yan pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 13268f34f1eaSPeter Xu if (!userfaultfd_wp(dst_vma)) 13278f34f1eaSPeter Xu pmd = pmd_swp_clear_uffd_wp(pmd); 132884c3fc4eSZi Yan set_pmd_at(dst_mm, addr, dst_pmd, pmd); 132984c3fc4eSZi Yan ret = 0; 133084c3fc4eSZi Yan goto out_unlock; 133184c3fc4eSZi Yan } 133284c3fc4eSZi Yan #endif 133384c3fc4eSZi Yan 1334628d47ceSKirill A. Shutemov if (unlikely(!pmd_trans_huge(pmd))) { 133571e3aac0SAndrea Arcangeli pte_free(dst_mm, pgtable); 133671e3aac0SAndrea Arcangeli goto out_unlock; 133771e3aac0SAndrea Arcangeli } 1338fc9fe822SKirill A. Shutemov /* 1339c4088ebdSKirill A. Shutemov * When page table lock is held, the huge zero pmd should not be 1340fc9fe822SKirill A. Shutemov * under splitting since we don't split the page itself, only pmd to 1341fc9fe822SKirill A. Shutemov * a page table. 1342fc9fe822SKirill A. Shutemov */ 1343fc9fe822SKirill A. Shutemov if (is_huge_zero_pmd(pmd)) { 134497ae1749SKirill A. Shutemov /* 134597ae1749SKirill A. Shutemov * get_huge_zero_page() will never allocate a new page here, 134697ae1749SKirill A. Shutemov * since we already have a zero page to copy. It just takes a 134797ae1749SKirill A. Shutemov * reference. 134897ae1749SKirill A. Shutemov */ 13495fc7a5f6SPeter Xu mm_get_huge_zero_page(dst_mm); 13505fc7a5f6SPeter Xu goto out_zero_page; 1351fc9fe822SKirill A. Shutemov } 1352de466bd6SMel Gorman 135371e3aac0SAndrea Arcangeli src_page = pmd_page(pmd); 1354309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(src_page), src_page); 135596c772c2SDavid Hildenbrand src_folio = page_folio(src_page); 1356d042035eSPeter Xu 135796c772c2SDavid Hildenbrand folio_get(src_folio); 135896c772c2SDavid Hildenbrand if (unlikely(folio_try_dup_anon_rmap_pmd(src_folio, src_page, src_vma))) { 1359fb3d824dSDavid Hildenbrand /* Page maybe pinned: split and retry the fault on PTEs. */ 136096c772c2SDavid Hildenbrand folio_put(src_folio); 1361d042035eSPeter Xu pte_free(dst_mm, pgtable); 1362d042035eSPeter Xu spin_unlock(src_ptl); 1363d042035eSPeter Xu spin_unlock(dst_ptl); 13648f34f1eaSPeter Xu __split_huge_pmd(src_vma, src_pmd, addr, false, NULL); 1365d042035eSPeter Xu return -EAGAIN; 1366d042035eSPeter Xu } 136771e3aac0SAndrea Arcangeli add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); 13685fc7a5f6SPeter Xu out_zero_page: 1369c4812909SKirill A. Shutemov mm_inc_nr_ptes(dst_mm); 13705c7fb56eSDan Williams pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); 137171e3aac0SAndrea Arcangeli pmdp_set_wrprotect(src_mm, addr, src_pmd); 13728f34f1eaSPeter Xu if (!userfaultfd_wp(dst_vma)) 13738f34f1eaSPeter Xu pmd = pmd_clear_uffd_wp(pmd); 137471e3aac0SAndrea Arcangeli pmd = pmd_mkold(pmd_wrprotect(pmd)); 137571e3aac0SAndrea Arcangeli set_pmd_at(dst_mm, addr, dst_pmd, pmd); 137671e3aac0SAndrea Arcangeli 137771e3aac0SAndrea Arcangeli ret = 0; 137871e3aac0SAndrea Arcangeli out_unlock: 1379c4088ebdSKirill A. Shutemov spin_unlock(src_ptl); 1380c4088ebdSKirill A. Shutemov spin_unlock(dst_ptl); 138171e3aac0SAndrea Arcangeli out: 138271e3aac0SAndrea Arcangeli return ret; 138371e3aac0SAndrea Arcangeli } 138471e3aac0SAndrea Arcangeli 1385a00cc7d9SMatthew Wilcox #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 1386a00cc7d9SMatthew Wilcox static void touch_pud(struct vm_area_struct *vma, unsigned long addr, 13875fe653e9SMiaohe Lin pud_t *pud, bool write) 1388a00cc7d9SMatthew Wilcox { 1389a00cc7d9SMatthew Wilcox pud_t _pud; 1390a00cc7d9SMatthew Wilcox 1391a8f97366SKirill A. Shutemov _pud = pud_mkyoung(*pud); 13925fe653e9SMiaohe Lin if (write) 1393a8f97366SKirill A. Shutemov _pud = pud_mkdirty(_pud); 1394a00cc7d9SMatthew Wilcox if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, 13955fe653e9SMiaohe Lin pud, _pud, write)) 1396a00cc7d9SMatthew Wilcox update_mmu_cache_pud(vma, addr, pud); 1397a00cc7d9SMatthew Wilcox } 1398a00cc7d9SMatthew Wilcox 1399a00cc7d9SMatthew Wilcox struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, 1400df06b37fSKeith Busch pud_t *pud, int flags, struct dev_pagemap **pgmap) 1401a00cc7d9SMatthew Wilcox { 1402a00cc7d9SMatthew Wilcox unsigned long pfn = pud_pfn(*pud); 1403a00cc7d9SMatthew Wilcox struct mm_struct *mm = vma->vm_mm; 1404a00cc7d9SMatthew Wilcox struct page *page; 14050f089235SLogan Gunthorpe int ret; 1406a00cc7d9SMatthew Wilcox 1407a00cc7d9SMatthew Wilcox assert_spin_locked(pud_lockptr(mm, pud)); 1408a00cc7d9SMatthew Wilcox 1409f6f37321SLinus Torvalds if (flags & FOLL_WRITE && !pud_write(*pud)) 1410a00cc7d9SMatthew Wilcox return NULL; 1411a00cc7d9SMatthew Wilcox 1412a00cc7d9SMatthew Wilcox if (pud_present(*pud) && pud_devmap(*pud)) 1413a00cc7d9SMatthew Wilcox /* pass */; 1414a00cc7d9SMatthew Wilcox else 1415a00cc7d9SMatthew Wilcox return NULL; 1416a00cc7d9SMatthew Wilcox 1417a00cc7d9SMatthew Wilcox if (flags & FOLL_TOUCH) 14185fe653e9SMiaohe Lin touch_pud(vma, addr, pud, flags & FOLL_WRITE); 1419a00cc7d9SMatthew Wilcox 1420a00cc7d9SMatthew Wilcox /* 1421a00cc7d9SMatthew Wilcox * device mapped pages can only be returned if the 1422a00cc7d9SMatthew Wilcox * caller will manage the page reference count. 14233faa52c0SJohn Hubbard * 14243faa52c0SJohn Hubbard * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here: 1425a00cc7d9SMatthew Wilcox */ 14263faa52c0SJohn Hubbard if (!(flags & (FOLL_GET | FOLL_PIN))) 1427a00cc7d9SMatthew Wilcox return ERR_PTR(-EEXIST); 1428a00cc7d9SMatthew Wilcox 1429a00cc7d9SMatthew Wilcox pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; 1430df06b37fSKeith Busch *pgmap = get_dev_pagemap(pfn, *pgmap); 1431df06b37fSKeith Busch if (!*pgmap) 1432a00cc7d9SMatthew Wilcox return ERR_PTR(-EFAULT); 1433a00cc7d9SMatthew Wilcox page = pfn_to_page(pfn); 14340f089235SLogan Gunthorpe 14350f089235SLogan Gunthorpe ret = try_grab_page(page, flags); 14360f089235SLogan Gunthorpe if (ret) 14370f089235SLogan Gunthorpe page = ERR_PTR(ret); 1438a00cc7d9SMatthew Wilcox 1439a00cc7d9SMatthew Wilcox return page; 1440a00cc7d9SMatthew Wilcox } 1441a00cc7d9SMatthew Wilcox 1442a00cc7d9SMatthew Wilcox int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, 1443a00cc7d9SMatthew Wilcox pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1444a00cc7d9SMatthew Wilcox struct vm_area_struct *vma) 1445a00cc7d9SMatthew Wilcox { 1446a00cc7d9SMatthew Wilcox spinlock_t *dst_ptl, *src_ptl; 1447a00cc7d9SMatthew Wilcox pud_t pud; 1448a00cc7d9SMatthew Wilcox int ret; 1449a00cc7d9SMatthew Wilcox 1450a00cc7d9SMatthew Wilcox dst_ptl = pud_lock(dst_mm, dst_pud); 1451a00cc7d9SMatthew Wilcox src_ptl = pud_lockptr(src_mm, src_pud); 1452a00cc7d9SMatthew Wilcox spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1453a00cc7d9SMatthew Wilcox 1454a00cc7d9SMatthew Wilcox ret = -EAGAIN; 1455a00cc7d9SMatthew Wilcox pud = *src_pud; 1456a00cc7d9SMatthew Wilcox if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) 1457a00cc7d9SMatthew Wilcox goto out_unlock; 1458a00cc7d9SMatthew Wilcox 1459a00cc7d9SMatthew Wilcox /* 1460a00cc7d9SMatthew Wilcox * When page table lock is held, the huge zero pud should not be 1461a00cc7d9SMatthew Wilcox * under splitting since we don't split the page itself, only pud to 1462a00cc7d9SMatthew Wilcox * a page table. 1463a00cc7d9SMatthew Wilcox */ 1464a00cc7d9SMatthew Wilcox if (is_huge_zero_pud(pud)) { 1465a00cc7d9SMatthew Wilcox /* No huge zero pud yet */ 1466a00cc7d9SMatthew Wilcox } 1467a00cc7d9SMatthew Wilcox 1468fb3d824dSDavid Hildenbrand /* 146996c772c2SDavid Hildenbrand * TODO: once we support anonymous pages, use 147096c772c2SDavid Hildenbrand * folio_try_dup_anon_rmap_*() and split if duplicating fails. 1471fb3d824dSDavid Hildenbrand */ 1472a00cc7d9SMatthew Wilcox pudp_set_wrprotect(src_mm, addr, src_pud); 1473a00cc7d9SMatthew Wilcox pud = pud_mkold(pud_wrprotect(pud)); 1474a00cc7d9SMatthew Wilcox set_pud_at(dst_mm, addr, dst_pud, pud); 1475a00cc7d9SMatthew Wilcox 1476a00cc7d9SMatthew Wilcox ret = 0; 1477a00cc7d9SMatthew Wilcox out_unlock: 1478a00cc7d9SMatthew Wilcox spin_unlock(src_ptl); 1479a00cc7d9SMatthew Wilcox spin_unlock(dst_ptl); 1480a00cc7d9SMatthew Wilcox return ret; 1481a00cc7d9SMatthew Wilcox } 1482a00cc7d9SMatthew Wilcox 1483a00cc7d9SMatthew Wilcox void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) 1484a00cc7d9SMatthew Wilcox { 1485a00cc7d9SMatthew Wilcox bool write = vmf->flags & FAULT_FLAG_WRITE; 1486a00cc7d9SMatthew Wilcox 1487a00cc7d9SMatthew Wilcox vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); 1488a00cc7d9SMatthew Wilcox if (unlikely(!pud_same(*vmf->pud, orig_pud))) 1489a00cc7d9SMatthew Wilcox goto unlock; 1490a00cc7d9SMatthew Wilcox 14915fe653e9SMiaohe Lin touch_pud(vmf->vma, vmf->address, vmf->pud, write); 1492a00cc7d9SMatthew Wilcox unlock: 1493a00cc7d9SMatthew Wilcox spin_unlock(vmf->ptl); 1494a00cc7d9SMatthew Wilcox } 1495a00cc7d9SMatthew Wilcox #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 1496a00cc7d9SMatthew Wilcox 14975db4f15cSYang Shi void huge_pmd_set_accessed(struct vm_fault *vmf) 1498a1dd450bSWill Deacon { 149920f664aaSMinchan Kim bool write = vmf->flags & FAULT_FLAG_WRITE; 1500a1dd450bSWill Deacon 150182b0f8c3SJan Kara vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1502a69e4717SMiaohe Lin if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd))) 1503a1dd450bSWill Deacon goto unlock; 1504a1dd450bSWill Deacon 1505a69e4717SMiaohe Lin touch_pmd(vmf->vma, vmf->address, vmf->pmd, write); 1506a1dd450bSWill Deacon 1507a1dd450bSWill Deacon unlock: 150882b0f8c3SJan Kara spin_unlock(vmf->ptl); 1509a1dd450bSWill Deacon } 1510a1dd450bSWill Deacon 15115db4f15cSYang Shi vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf) 151271e3aac0SAndrea Arcangeli { 1513c89357e2SDavid Hildenbrand const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; 151482b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 15152fad3d14SMatthew Wilcox (Oracle) struct folio *folio; 15163917c802SKirill A. Shutemov struct page *page; 151782b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 15185db4f15cSYang Shi pmd_t orig_pmd = vmf->orig_pmd; 151971e3aac0SAndrea Arcangeli 152082b0f8c3SJan Kara vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); 152181d1b09cSSasha Levin VM_BUG_ON_VMA(!vma->anon_vma, vma); 15223917c802SKirill A. Shutemov 152393b4796dSKirill A. Shutemov if (is_huge_zero_pmd(orig_pmd)) 15243917c802SKirill A. Shutemov goto fallback; 15253917c802SKirill A. Shutemov 152682b0f8c3SJan Kara spin_lock(vmf->ptl); 15273917c802SKirill A. Shutemov 15283917c802SKirill A. Shutemov if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 15293917c802SKirill A. Shutemov spin_unlock(vmf->ptl); 15303917c802SKirill A. Shutemov return 0; 15313917c802SKirill A. Shutemov } 153271e3aac0SAndrea Arcangeli 153371e3aac0SAndrea Arcangeli page = pmd_page(orig_pmd); 15342fad3d14SMatthew Wilcox (Oracle) folio = page_folio(page); 1535f6004e73SMiaohe Lin VM_BUG_ON_PAGE(!PageHead(page), page); 15363917c802SKirill A. Shutemov 15376c287605SDavid Hildenbrand /* Early check when only holding the PT lock. */ 15386c287605SDavid Hildenbrand if (PageAnonExclusive(page)) 15396c287605SDavid Hildenbrand goto reuse; 15406c287605SDavid Hildenbrand 15412fad3d14SMatthew Wilcox (Oracle) if (!folio_trylock(folio)) { 15422fad3d14SMatthew Wilcox (Oracle) folio_get(folio); 1543ba3c4ce6SHuang Ying spin_unlock(vmf->ptl); 15442fad3d14SMatthew Wilcox (Oracle) folio_lock(folio); 1545ba3c4ce6SHuang Ying spin_lock(vmf->ptl); 1546ba3c4ce6SHuang Ying if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { 15473917c802SKirill A. Shutemov spin_unlock(vmf->ptl); 15482fad3d14SMatthew Wilcox (Oracle) folio_unlock(folio); 15492fad3d14SMatthew Wilcox (Oracle) folio_put(folio); 15503917c802SKirill A. Shutemov return 0; 1551ba3c4ce6SHuang Ying } 15522fad3d14SMatthew Wilcox (Oracle) folio_put(folio); 1553ba3c4ce6SHuang Ying } 15543917c802SKirill A. Shutemov 15556c287605SDavid Hildenbrand /* Recheck after temporarily dropping the PT lock. */ 15566c287605SDavid Hildenbrand if (PageAnonExclusive(page)) { 15572fad3d14SMatthew Wilcox (Oracle) folio_unlock(folio); 15586c287605SDavid Hildenbrand goto reuse; 15596c287605SDavid Hildenbrand } 15606c287605SDavid Hildenbrand 15613917c802SKirill A. Shutemov /* 15622fad3d14SMatthew Wilcox (Oracle) * See do_wp_page(): we can only reuse the folio exclusively if 15632fad3d14SMatthew Wilcox (Oracle) * there are no additional references. Note that we always drain 15641fec6890SMatthew Wilcox (Oracle) * the LRU cache immediately after adding a THP. 15653917c802SKirill A. Shutemov */ 15662fad3d14SMatthew Wilcox (Oracle) if (folio_ref_count(folio) > 15672fad3d14SMatthew Wilcox (Oracle) 1 + folio_test_swapcache(folio) * folio_nr_pages(folio)) 15683bff7e3fSDavid Hildenbrand goto unlock_fallback; 15692fad3d14SMatthew Wilcox (Oracle) if (folio_test_swapcache(folio)) 15702fad3d14SMatthew Wilcox (Oracle) folio_free_swap(folio); 15712fad3d14SMatthew Wilcox (Oracle) if (folio_ref_count(folio) == 1) { 157271e3aac0SAndrea Arcangeli pmd_t entry; 15736c54dc6cSDavid Hildenbrand 157406968625SDavid Hildenbrand folio_move_anon_rmap(folio, vma); 15755ca43289SDavid Hildenbrand SetPageAnonExclusive(page); 15762fad3d14SMatthew Wilcox (Oracle) folio_unlock(folio); 15776c287605SDavid Hildenbrand reuse: 1578c89357e2SDavid Hildenbrand if (unlikely(unshare)) { 1579c89357e2SDavid Hildenbrand spin_unlock(vmf->ptl); 1580c89357e2SDavid Hildenbrand return 0; 1581c89357e2SDavid Hildenbrand } 158271e3aac0SAndrea Arcangeli entry = pmd_mkyoung(orig_pmd); 1583f55e1014SLinus Torvalds entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 158482b0f8c3SJan Kara if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) 158582b0f8c3SJan Kara update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 15863917c802SKirill A. Shutemov spin_unlock(vmf->ptl); 1587cb8d8633SDavid Hildenbrand return 0; 158871e3aac0SAndrea Arcangeli } 15893917c802SKirill A. Shutemov 15903bff7e3fSDavid Hildenbrand unlock_fallback: 15912fad3d14SMatthew Wilcox (Oracle) folio_unlock(folio); 159282b0f8c3SJan Kara spin_unlock(vmf->ptl); 15933917c802SKirill A. Shutemov fallback: 15943917c802SKirill A. Shutemov __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL); 15953917c802SKirill A. Shutemov return VM_FAULT_FALLBACK; 159671e3aac0SAndrea Arcangeli } 159771e3aac0SAndrea Arcangeli 1598c27f479eSDavid Hildenbrand static inline bool can_change_pmd_writable(struct vm_area_struct *vma, 1599c27f479eSDavid Hildenbrand unsigned long addr, pmd_t pmd) 1600c27f479eSDavid Hildenbrand { 1601c27f479eSDavid Hildenbrand struct page *page; 1602c27f479eSDavid Hildenbrand 1603c27f479eSDavid Hildenbrand if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE))) 1604c27f479eSDavid Hildenbrand return false; 1605c27f479eSDavid Hildenbrand 1606c27f479eSDavid Hildenbrand /* Don't touch entries that are not even readable (NUMA hinting). */ 1607c27f479eSDavid Hildenbrand if (pmd_protnone(pmd)) 1608c27f479eSDavid Hildenbrand return false; 1609c27f479eSDavid Hildenbrand 1610c27f479eSDavid Hildenbrand /* Do we need write faults for softdirty tracking? */ 1611c27f479eSDavid Hildenbrand if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd)) 1612c27f479eSDavid Hildenbrand return false; 1613c27f479eSDavid Hildenbrand 1614c27f479eSDavid Hildenbrand /* Do we need write faults for uffd-wp tracking? */ 1615c27f479eSDavid Hildenbrand if (userfaultfd_huge_pmd_wp(vma, pmd)) 1616c27f479eSDavid Hildenbrand return false; 1617c27f479eSDavid Hildenbrand 1618c27f479eSDavid Hildenbrand if (!(vma->vm_flags & VM_SHARED)) { 1619c27f479eSDavid Hildenbrand /* See can_change_pte_writable(). */ 1620c27f479eSDavid Hildenbrand page = vm_normal_page_pmd(vma, addr, pmd); 1621c27f479eSDavid Hildenbrand return page && PageAnon(page) && PageAnonExclusive(page); 1622c27f479eSDavid Hildenbrand } 1623c27f479eSDavid Hildenbrand 1624c27f479eSDavid Hildenbrand /* See can_change_pte_writable(). */ 1625c27f479eSDavid Hildenbrand return pmd_dirty(pmd); 1626c27f479eSDavid Hildenbrand } 1627c27f479eSDavid Hildenbrand 16285535be30SDavid Hildenbrand /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */ 16295535be30SDavid Hildenbrand static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page, 16305535be30SDavid Hildenbrand struct vm_area_struct *vma, 16315535be30SDavid Hildenbrand unsigned int flags) 16328310d48bSKeno Fischer { 16335535be30SDavid Hildenbrand /* If the pmd is writable, we can write to the page. */ 16345535be30SDavid Hildenbrand if (pmd_write(pmd)) 16355535be30SDavid Hildenbrand return true; 16365535be30SDavid Hildenbrand 16375535be30SDavid Hildenbrand /* Maybe FOLL_FORCE is set to override it? */ 16385535be30SDavid Hildenbrand if (!(flags & FOLL_FORCE)) 16395535be30SDavid Hildenbrand return false; 16405535be30SDavid Hildenbrand 16415535be30SDavid Hildenbrand /* But FOLL_FORCE has no effect on shared mappings */ 16425535be30SDavid Hildenbrand if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED)) 16435535be30SDavid Hildenbrand return false; 16445535be30SDavid Hildenbrand 16455535be30SDavid Hildenbrand /* ... or read-only private ones */ 16465535be30SDavid Hildenbrand if (!(vma->vm_flags & VM_MAYWRITE)) 16475535be30SDavid Hildenbrand return false; 16485535be30SDavid Hildenbrand 16495535be30SDavid Hildenbrand /* ... or already writable ones that just need to take a write fault */ 16505535be30SDavid Hildenbrand if (vma->vm_flags & VM_WRITE) 16515535be30SDavid Hildenbrand return false; 16525535be30SDavid Hildenbrand 16535535be30SDavid Hildenbrand /* 16545535be30SDavid Hildenbrand * See can_change_pte_writable(): we broke COW and could map the page 16555535be30SDavid Hildenbrand * writable if we have an exclusive anonymous page ... 16565535be30SDavid Hildenbrand */ 16575535be30SDavid Hildenbrand if (!page || !PageAnon(page) || !PageAnonExclusive(page)) 16585535be30SDavid Hildenbrand return false; 16595535be30SDavid Hildenbrand 16605535be30SDavid Hildenbrand /* ... and a write-fault isn't required for other reasons. */ 16615535be30SDavid Hildenbrand if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd)) 16625535be30SDavid Hildenbrand return false; 16635535be30SDavid Hildenbrand return !userfaultfd_huge_pmd_wp(vma, pmd); 16648310d48bSKeno Fischer } 16658310d48bSKeno Fischer 1666b676b293SDavid Rientjes struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, 166771e3aac0SAndrea Arcangeli unsigned long addr, 166871e3aac0SAndrea Arcangeli pmd_t *pmd, 166971e3aac0SAndrea Arcangeli unsigned int flags) 167071e3aac0SAndrea Arcangeli { 1671b676b293SDavid Rientjes struct mm_struct *mm = vma->vm_mm; 16725535be30SDavid Hildenbrand struct page *page; 16730f089235SLogan Gunthorpe int ret; 167471e3aac0SAndrea Arcangeli 1675c4088ebdSKirill A. Shutemov assert_spin_locked(pmd_lockptr(mm, pmd)); 167671e3aac0SAndrea Arcangeli 16775535be30SDavid Hildenbrand page = pmd_page(*pmd); 16785535be30SDavid Hildenbrand VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); 16795535be30SDavid Hildenbrand 16805535be30SDavid Hildenbrand if ((flags & FOLL_WRITE) && 16815535be30SDavid Hildenbrand !can_follow_write_pmd(*pmd, page, vma, flags)) 16825535be30SDavid Hildenbrand return NULL; 168371e3aac0SAndrea Arcangeli 168485facf25SKirill A. Shutemov /* Avoid dumping huge zero page */ 168585facf25SKirill A. Shutemov if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) 168685facf25SKirill A. Shutemov return ERR_PTR(-EFAULT); 168785facf25SKirill A. Shutemov 1688d74943a2SDavid Hildenbrand if (pmd_protnone(*pmd) && !gup_can_follow_protnone(vma, flags)) 16895535be30SDavid Hildenbrand return NULL; 16903faa52c0SJohn Hubbard 169184209e87SDavid Hildenbrand if (!pmd_write(*pmd) && gup_must_unshare(vma, flags, page)) 1692a7f22660SDavid Hildenbrand return ERR_PTR(-EMLINK); 1693a7f22660SDavid Hildenbrand 1694b6a2619cSDavid Hildenbrand VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && 1695b6a2619cSDavid Hildenbrand !PageAnonExclusive(page), page); 1696b6a2619cSDavid Hildenbrand 16970f089235SLogan Gunthorpe ret = try_grab_page(page, flags); 16980f089235SLogan Gunthorpe if (ret) 16990f089235SLogan Gunthorpe return ERR_PTR(ret); 17003faa52c0SJohn Hubbard 17013565fce3SDan Williams if (flags & FOLL_TOUCH) 1702a69e4717SMiaohe Lin touch_pmd(vma, addr, pmd, flags & FOLL_WRITE); 17033faa52c0SJohn Hubbard 170471e3aac0SAndrea Arcangeli page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; 1705ca120cf6SDan Williams VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); 170671e3aac0SAndrea Arcangeli 170771e3aac0SAndrea Arcangeli return page; 170871e3aac0SAndrea Arcangeli } 170971e3aac0SAndrea Arcangeli 1710d10e63f2SMel Gorman /* NUMA hinting page fault entry point for trans huge pmds */ 17115db4f15cSYang Shi vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf) 1712d10e63f2SMel Gorman { 171382b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 1714c5b5a3ddSYang Shi pmd_t oldpmd = vmf->orig_pmd; 1715c5b5a3ddSYang Shi pmd_t pmd; 1716667ffc31SKefeng Wang struct folio *folio; 171782b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 1718667ffc31SKefeng Wang int nid = NUMA_NO_NODE; 171933024536SHuang Ying int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK); 17206a56ccbcSDavid Hildenbrand bool migrated = false, writable = false; 17216688cc05SPeter Zijlstra int flags = 0; 1722d10e63f2SMel Gorman 172382b0f8c3SJan Kara vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1724c5b5a3ddSYang Shi if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) { 172582b0f8c3SJan Kara spin_unlock(vmf->ptl); 1726de466bd6SMel Gorman goto out; 1727de466bd6SMel Gorman } 1728de466bd6SMel Gorman 1729c5b5a3ddSYang Shi pmd = pmd_modify(oldpmd, vma->vm_page_prot); 17306a56ccbcSDavid Hildenbrand 17316a56ccbcSDavid Hildenbrand /* 17326a56ccbcSDavid Hildenbrand * Detect now whether the PMD could be writable; this information 17336a56ccbcSDavid Hildenbrand * is only valid while holding the PT lock. 17346a56ccbcSDavid Hildenbrand */ 17356a56ccbcSDavid Hildenbrand writable = pmd_write(pmd); 17366a56ccbcSDavid Hildenbrand if (!writable && vma_wants_manual_pte_write_upgrade(vma) && 17376a56ccbcSDavid Hildenbrand can_change_pmd_writable(vma, vmf->address, pmd)) 17386a56ccbcSDavid Hildenbrand writable = true; 17396a56ccbcSDavid Hildenbrand 1740667ffc31SKefeng Wang folio = vm_normal_folio_pmd(vma, haddr, pmd); 1741667ffc31SKefeng Wang if (!folio) 1742c5b5a3ddSYang Shi goto out_map; 1743c5b5a3ddSYang Shi 1744c5b5a3ddSYang Shi /* See similar comment in do_numa_page for explanation */ 17456a56ccbcSDavid Hildenbrand if (!writable) 1746c5b5a3ddSYang Shi flags |= TNF_NO_GROUP; 1747c5b5a3ddSYang Shi 1748667ffc31SKefeng Wang nid = folio_nid(folio); 174933024536SHuang Ying /* 175033024536SHuang Ying * For memory tiering mode, cpupid of slow memory page is used 175133024536SHuang Ying * to record page access time. So use default value. 175233024536SHuang Ying */ 1753667ffc31SKefeng Wang if (node_is_toptier(nid)) 1754c4a8d2faSKefeng Wang last_cpupid = folio_last_cpupid(folio); 1755cda6d936SKefeng Wang target_nid = numa_migrate_prep(folio, vma, haddr, nid, &flags); 1756c5b5a3ddSYang Shi if (target_nid == NUMA_NO_NODE) { 1757667ffc31SKefeng Wang folio_put(folio); 1758c5b5a3ddSYang Shi goto out_map; 1759c5b5a3ddSYang Shi } 1760c5b5a3ddSYang Shi 176182b0f8c3SJan Kara spin_unlock(vmf->ptl); 17626a56ccbcSDavid Hildenbrand writable = false; 17638b1b436dSPeter Zijlstra 1764667ffc31SKefeng Wang migrated = migrate_misplaced_folio(folio, vma, target_nid); 17656688cc05SPeter Zijlstra if (migrated) { 17666688cc05SPeter Zijlstra flags |= TNF_MIGRATED; 1767667ffc31SKefeng Wang nid = target_nid; 1768c5b5a3ddSYang Shi } else { 1769074c2381SMel Gorman flags |= TNF_MIGRATE_FAIL; 1770c5b5a3ddSYang Shi vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 1771c5b5a3ddSYang Shi if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) { 177282b0f8c3SJan Kara spin_unlock(vmf->ptl); 1773c5b5a3ddSYang Shi goto out; 1774c5b5a3ddSYang Shi } 1775c5b5a3ddSYang Shi goto out_map; 1776c5b5a3ddSYang Shi } 1777b8916634SMel Gorman 1778b8916634SMel Gorman out: 1779667ffc31SKefeng Wang if (nid != NUMA_NO_NODE) 1780667ffc31SKefeng Wang task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags); 17818191acbdSMel Gorman 1782d10e63f2SMel Gorman return 0; 1783c5b5a3ddSYang Shi 1784c5b5a3ddSYang Shi out_map: 1785c5b5a3ddSYang Shi /* Restore the PMD */ 1786c5b5a3ddSYang Shi pmd = pmd_modify(oldpmd, vma->vm_page_prot); 1787c5b5a3ddSYang Shi pmd = pmd_mkyoung(pmd); 17886a56ccbcSDavid Hildenbrand if (writable) 1789161e393cSRick Edgecombe pmd = pmd_mkwrite(pmd, vma); 1790c5b5a3ddSYang Shi set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); 1791c5b5a3ddSYang Shi update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); 1792c5b5a3ddSYang Shi spin_unlock(vmf->ptl); 1793c5b5a3ddSYang Shi goto out; 1794d10e63f2SMel Gorman } 1795d10e63f2SMel Gorman 1796319904adSHuang Ying /* 1797319904adSHuang Ying * Return true if we do MADV_FREE successfully on entire pmd page. 1798319904adSHuang Ying * Otherwise, return false. 1799319904adSHuang Ying */ 1800319904adSHuang Ying bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1801b8d3c4c3SMinchan Kim pmd_t *pmd, unsigned long addr, unsigned long next) 1802b8d3c4c3SMinchan Kim { 1803b8d3c4c3SMinchan Kim spinlock_t *ptl; 1804b8d3c4c3SMinchan Kim pmd_t orig_pmd; 1805fc986a38SKefeng Wang struct folio *folio; 1806b8d3c4c3SMinchan Kim struct mm_struct *mm = tlb->mm; 1807319904adSHuang Ying bool ret = false; 1808b8d3c4c3SMinchan Kim 1809ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 181007e32661SAneesh Kumar K.V 1811b6ec57f4SKirill A. Shutemov ptl = pmd_trans_huge_lock(pmd, vma); 1812b6ec57f4SKirill A. Shutemov if (!ptl) 181325eedabeSLinus Torvalds goto out_unlocked; 1814b8d3c4c3SMinchan Kim 1815b8d3c4c3SMinchan Kim orig_pmd = *pmd; 1816319904adSHuang Ying if (is_huge_zero_pmd(orig_pmd)) 1817b8d3c4c3SMinchan Kim goto out; 1818b8d3c4c3SMinchan Kim 181984c3fc4eSZi Yan if (unlikely(!pmd_present(orig_pmd))) { 182084c3fc4eSZi Yan VM_BUG_ON(thp_migration_supported() && 182184c3fc4eSZi Yan !is_pmd_migration_entry(orig_pmd)); 182284c3fc4eSZi Yan goto out; 182384c3fc4eSZi Yan } 182484c3fc4eSZi Yan 1825fc986a38SKefeng Wang folio = pfn_folio(pmd_pfn(orig_pmd)); 1826b8d3c4c3SMinchan Kim /* 1827fc986a38SKefeng Wang * If other processes are mapping this folio, we couldn't discard 1828fc986a38SKefeng Wang * the folio unless they all do MADV_FREE so let's skip the folio. 1829b8d3c4c3SMinchan Kim */ 183020b18aadSYin Fengwei if (folio_estimated_sharers(folio) != 1) 1831b8d3c4c3SMinchan Kim goto out; 1832b8d3c4c3SMinchan Kim 1833fc986a38SKefeng Wang if (!folio_trylock(folio)) 1834b8d3c4c3SMinchan Kim goto out; 1835b8d3c4c3SMinchan Kim 1836b8d3c4c3SMinchan Kim /* 1837b8d3c4c3SMinchan Kim * If user want to discard part-pages of THP, split it so MADV_FREE 1838b8d3c4c3SMinchan Kim * will deactivate only them. 1839b8d3c4c3SMinchan Kim */ 1840b8d3c4c3SMinchan Kim if (next - addr != HPAGE_PMD_SIZE) { 1841fc986a38SKefeng Wang folio_get(folio); 1842b8d3c4c3SMinchan Kim spin_unlock(ptl); 1843fc986a38SKefeng Wang split_folio(folio); 1844fc986a38SKefeng Wang folio_unlock(folio); 1845fc986a38SKefeng Wang folio_put(folio); 1846b8d3c4c3SMinchan Kim goto out_unlocked; 1847b8d3c4c3SMinchan Kim } 1848b8d3c4c3SMinchan Kim 1849fc986a38SKefeng Wang if (folio_test_dirty(folio)) 1850fc986a38SKefeng Wang folio_clear_dirty(folio); 1851fc986a38SKefeng Wang folio_unlock(folio); 1852b8d3c4c3SMinchan Kim 1853b8d3c4c3SMinchan Kim if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { 185458ceeb6bSKirill A. Shutemov pmdp_invalidate(vma, addr, pmd); 1855b8d3c4c3SMinchan Kim orig_pmd = pmd_mkold(orig_pmd); 1856b8d3c4c3SMinchan Kim orig_pmd = pmd_mkclean(orig_pmd); 1857b8d3c4c3SMinchan Kim 1858b8d3c4c3SMinchan Kim set_pmd_at(mm, addr, pmd, orig_pmd); 1859b8d3c4c3SMinchan Kim tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 1860b8d3c4c3SMinchan Kim } 1861802a3a92SShaohua Li 18626a6fe9ebSKefeng Wang folio_mark_lazyfree(folio); 1863319904adSHuang Ying ret = true; 1864b8d3c4c3SMinchan Kim out: 1865b8d3c4c3SMinchan Kim spin_unlock(ptl); 1866b8d3c4c3SMinchan Kim out_unlocked: 1867b8d3c4c3SMinchan Kim return ret; 1868b8d3c4c3SMinchan Kim } 1869b8d3c4c3SMinchan Kim 1870953c66c2SAneesh Kumar K.V static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) 1871953c66c2SAneesh Kumar K.V { 1872953c66c2SAneesh Kumar K.V pgtable_t pgtable; 1873953c66c2SAneesh Kumar K.V 1874953c66c2SAneesh Kumar K.V pgtable = pgtable_trans_huge_withdraw(mm, pmd); 1875953c66c2SAneesh Kumar K.V pte_free(mm, pgtable); 1876c4812909SKirill A. Shutemov mm_dec_nr_ptes(mm); 1877953c66c2SAneesh Kumar K.V } 1878953c66c2SAneesh Kumar K.V 187971e3aac0SAndrea Arcangeli int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 1880f21760b1SShaohua Li pmd_t *pmd, unsigned long addr) 188171e3aac0SAndrea Arcangeli { 1882f5c8ad47SDavid Miller pmd_t orig_pmd; 1883da146769SKirill A. Shutemov spinlock_t *ptl; 1884da146769SKirill A. Shutemov 1885ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 188607e32661SAneesh Kumar K.V 1887b6ec57f4SKirill A. Shutemov ptl = __pmd_trans_huge_lock(pmd, vma); 1888b6ec57f4SKirill A. Shutemov if (!ptl) 1889da146769SKirill A. Shutemov return 0; 1890a6bf2bb0SAneesh Kumar K.V /* 1891a6bf2bb0SAneesh Kumar K.V * For architectures like ppc64 we look at deposited pgtable 18928809aa2dSAneesh Kumar K.V * when calling pmdp_huge_get_and_clear. So do the 1893a6bf2bb0SAneesh Kumar K.V * pgtable_trans_huge_withdraw after finishing pmdp related 1894a6bf2bb0SAneesh Kumar K.V * operations. 1895a6bf2bb0SAneesh Kumar K.V */ 189693a98695SAneesh Kumar K.V orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd, 1897fcbe08d6SMartin Schwidefsky tlb->fullmm); 1898e5136e87SRick Edgecombe arch_check_zapped_pmd(vma, orig_pmd); 1899f21760b1SShaohua Li tlb_remove_pmd_tlb_entry(tlb, pmd, addr); 19002484ca9bSThomas Hellstrom (VMware) if (vma_is_special_huge(vma)) { 19013b6521f5SOliver O'Halloran if (arch_needs_pgtable_deposit()) 19023b6521f5SOliver O'Halloran zap_deposited_table(tlb->mm, pmd); 19034897c765SMatthew Wilcox spin_unlock(ptl); 1904da146769SKirill A. Shutemov } else if (is_huge_zero_pmd(orig_pmd)) { 1905c14a6eb4SOliver O'Halloran zap_deposited_table(tlb->mm, pmd); 1906bf929152SKirill A. Shutemov spin_unlock(ptl); 1907479f0abbSKirill A. Shutemov } else { 1908616b8371SZi Yan struct page *page = NULL; 1909616b8371SZi Yan int flush_needed = 1; 1910616b8371SZi Yan 1911616b8371SZi Yan if (pmd_present(orig_pmd)) { 1912616b8371SZi Yan page = pmd_page(orig_pmd); 1913a8e61d58SDavid Hildenbrand folio_remove_rmap_pmd(page_folio(page), page, vma); 1914309381feSSasha Levin VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); 1915309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 1916616b8371SZi Yan } else if (thp_migration_supported()) { 1917616b8371SZi Yan swp_entry_t entry; 1918616b8371SZi Yan 1919616b8371SZi Yan VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); 1920616b8371SZi Yan entry = pmd_to_swp_entry(orig_pmd); 1921af5cdaf8SAlistair Popple page = pfn_swap_entry_to_page(entry); 1922616b8371SZi Yan flush_needed = 0; 1923616b8371SZi Yan } else 1924616b8371SZi Yan WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); 1925616b8371SZi Yan 1926b5072380SKirill A. Shutemov if (PageAnon(page)) { 1927c14a6eb4SOliver O'Halloran zap_deposited_table(tlb->mm, pmd); 1928b5072380SKirill A. Shutemov add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); 1929b5072380SKirill A. Shutemov } else { 1930953c66c2SAneesh Kumar K.V if (arch_needs_pgtable_deposit()) 1931953c66c2SAneesh Kumar K.V zap_deposited_table(tlb->mm, pmd); 1932fadae295SYang Shi add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR); 1933b5072380SKirill A. Shutemov } 1934616b8371SZi Yan 1935bf929152SKirill A. Shutemov spin_unlock(ptl); 1936616b8371SZi Yan if (flush_needed) 1937e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); 1938479f0abbSKirill A. Shutemov } 1939da146769SKirill A. Shutemov return 1; 194071e3aac0SAndrea Arcangeli } 194171e3aac0SAndrea Arcangeli 19421dd38b6cSAneesh Kumar K.V #ifndef pmd_move_must_withdraw 19431dd38b6cSAneesh Kumar K.V static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, 19441dd38b6cSAneesh Kumar K.V spinlock_t *old_pmd_ptl, 19451dd38b6cSAneesh Kumar K.V struct vm_area_struct *vma) 19461dd38b6cSAneesh Kumar K.V { 19471dd38b6cSAneesh Kumar K.V /* 19481dd38b6cSAneesh Kumar K.V * With split pmd lock we also need to move preallocated 19491dd38b6cSAneesh Kumar K.V * PTE page table if new_pmd is on different PMD page table. 19501dd38b6cSAneesh Kumar K.V * 19511dd38b6cSAneesh Kumar K.V * We also don't deposit and withdraw tables for file pages. 19521dd38b6cSAneesh Kumar K.V */ 19531dd38b6cSAneesh Kumar K.V return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); 19541dd38b6cSAneesh Kumar K.V } 19551dd38b6cSAneesh Kumar K.V #endif 19561dd38b6cSAneesh Kumar K.V 1957ab6e3d09SNaoya Horiguchi static pmd_t move_soft_dirty_pmd(pmd_t pmd) 1958ab6e3d09SNaoya Horiguchi { 1959ab6e3d09SNaoya Horiguchi #ifdef CONFIG_MEM_SOFT_DIRTY 1960ab6e3d09SNaoya Horiguchi if (unlikely(is_pmd_migration_entry(pmd))) 1961ab6e3d09SNaoya Horiguchi pmd = pmd_swp_mksoft_dirty(pmd); 1962ab6e3d09SNaoya Horiguchi else if (pmd_present(pmd)) 1963ab6e3d09SNaoya Horiguchi pmd = pmd_mksoft_dirty(pmd); 1964ab6e3d09SNaoya Horiguchi #endif 1965ab6e3d09SNaoya Horiguchi return pmd; 1966ab6e3d09SNaoya Horiguchi } 1967ab6e3d09SNaoya Horiguchi 1968bf8616d5SHugh Dickins bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, 1969b8aa9d9dSWei Yang unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd) 197037a1c49aSAndrea Arcangeli { 1971bf929152SKirill A. Shutemov spinlock_t *old_ptl, *new_ptl; 197237a1c49aSAndrea Arcangeli pmd_t pmd; 197337a1c49aSAndrea Arcangeli struct mm_struct *mm = vma->vm_mm; 19745d190420SAaron Lu bool force_flush = false; 197537a1c49aSAndrea Arcangeli 197637a1c49aSAndrea Arcangeli /* 197737a1c49aSAndrea Arcangeli * The destination pmd shouldn't be established, free_pgtables() 1978a5be621eSHugh Dickins * should have released it; but move_page_tables() might have already 1979a5be621eSHugh Dickins * inserted a page table, if racing against shmem/file collapse. 198037a1c49aSAndrea Arcangeli */ 1981a5be621eSHugh Dickins if (!pmd_none(*new_pmd)) { 198237a1c49aSAndrea Arcangeli VM_BUG_ON(pmd_trans_huge(*new_pmd)); 19834b471e88SKirill A. Shutemov return false; 198437a1c49aSAndrea Arcangeli } 198537a1c49aSAndrea Arcangeli 1986bf929152SKirill A. Shutemov /* 1987bf929152SKirill A. Shutemov * We don't have to worry about the ordering of src and dst 1988c1e8d7c6SMichel Lespinasse * ptlocks because exclusive mmap_lock prevents deadlock. 1989bf929152SKirill A. Shutemov */ 1990b6ec57f4SKirill A. Shutemov old_ptl = __pmd_trans_huge_lock(old_pmd, vma); 1991b6ec57f4SKirill A. Shutemov if (old_ptl) { 1992bf929152SKirill A. Shutemov new_ptl = pmd_lockptr(mm, new_pmd); 1993bf929152SKirill A. Shutemov if (new_ptl != old_ptl) 1994bf929152SKirill A. Shutemov spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); 19958809aa2dSAneesh Kumar K.V pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); 1996eb66ae03SLinus Torvalds if (pmd_present(pmd)) 1997a2ce2666SAaron Lu force_flush = true; 199837a1c49aSAndrea Arcangeli VM_BUG_ON(!pmd_none(*new_pmd)); 19993592806cSKirill A. Shutemov 20001dd38b6cSAneesh Kumar K.V if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { 2001b3084f4dSAneesh Kumar K.V pgtable_t pgtable; 20023592806cSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); 20033592806cSKirill A. Shutemov pgtable_trans_huge_deposit(mm, new_pmd, pgtable); 20043592806cSKirill A. Shutemov } 2005ab6e3d09SNaoya Horiguchi pmd = move_soft_dirty_pmd(pmd); 2006ab6e3d09SNaoya Horiguchi set_pmd_at(mm, new_addr, new_pmd, pmd); 20075d190420SAaron Lu if (force_flush) 20087c38f181SMiaohe Lin flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE); 2009eb66ae03SLinus Torvalds if (new_ptl != old_ptl) 2010eb66ae03SLinus Torvalds spin_unlock(new_ptl); 2011bf929152SKirill A. Shutemov spin_unlock(old_ptl); 20124b471e88SKirill A. Shutemov return true; 201337a1c49aSAndrea Arcangeli } 20144b471e88SKirill A. Shutemov return false; 201537a1c49aSAndrea Arcangeli } 201637a1c49aSAndrea Arcangeli 2017f123d74aSMel Gorman /* 2018f123d74aSMel Gorman * Returns 2019f123d74aSMel Gorman * - 0 if PMD could not be locked 2020f0953a1bSIngo Molnar * - 1 if PMD was locked but protections unchanged and TLB flush unnecessary 2021e346e668SYang Shi * or if prot_numa but THP migration is not supported 2022f0953a1bSIngo Molnar * - HPAGE_PMD_NR if protections changed and TLB flush necessary 2023f123d74aSMel Gorman */ 20244a18419fSNadav Amit int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, 20254a18419fSNadav Amit pmd_t *pmd, unsigned long addr, pgprot_t newprot, 20264a18419fSNadav Amit unsigned long cp_flags) 2027cd7548abSJohannes Weiner { 2028cd7548abSJohannes Weiner struct mm_struct *mm = vma->vm_mm; 2029bf929152SKirill A. Shutemov spinlock_t *ptl; 2030c9fe6656SNadav Amit pmd_t oldpmd, entry; 203158705444SPeter Xu bool prot_numa = cp_flags & MM_CP_PROT_NUMA; 2032292924b2SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 2033292924b2SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 20346a56ccbcSDavid Hildenbrand int ret = 1; 2035cd7548abSJohannes Weiner 20364a18419fSNadav Amit tlb_change_page_size(tlb, HPAGE_PMD_SIZE); 20374a18419fSNadav Amit 2038e346e668SYang Shi if (prot_numa && !thp_migration_supported()) 2039e346e668SYang Shi return 1; 2040e346e668SYang Shi 2041b6ec57f4SKirill A. Shutemov ptl = __pmd_trans_huge_lock(pmd, vma); 20420a85e51dSKirill A. Shutemov if (!ptl) 20430a85e51dSKirill A. Shutemov return 0; 20440a85e51dSKirill A. Shutemov 204584c3fc4eSZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 204684c3fc4eSZi Yan if (is_swap_pmd(*pmd)) { 204784c3fc4eSZi Yan swp_entry_t entry = pmd_to_swp_entry(*pmd); 2048d986ba2bSKefeng Wang struct folio *folio = page_folio(pfn_swap_entry_to_page(entry)); 204924bf08c4SDavid Hildenbrand pmd_t newpmd; 205084c3fc4eSZi Yan 205184c3fc4eSZi Yan VM_BUG_ON(!is_pmd_migration_entry(*pmd)); 20524dd845b5SAlistair Popple if (is_writable_migration_entry(entry)) { 205384c3fc4eSZi Yan /* 205484c3fc4eSZi Yan * A protection check is difficult so 205584c3fc4eSZi Yan * just be safe and disable write 205684c3fc4eSZi Yan */ 2057d986ba2bSKefeng Wang if (folio_test_anon(folio)) 20586c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry(swp_offset(entry)); 20596c287605SDavid Hildenbrand else 20606c287605SDavid Hildenbrand entry = make_readable_migration_entry(swp_offset(entry)); 206184c3fc4eSZi Yan newpmd = swp_entry_to_pmd(entry); 2062ab6e3d09SNaoya Horiguchi if (pmd_swp_soft_dirty(*pmd)) 2063ab6e3d09SNaoya Horiguchi newpmd = pmd_swp_mksoft_dirty(newpmd); 206424bf08c4SDavid Hildenbrand } else { 206524bf08c4SDavid Hildenbrand newpmd = *pmd; 206684c3fc4eSZi Yan } 206724bf08c4SDavid Hildenbrand 206824bf08c4SDavid Hildenbrand if (uffd_wp) 206924bf08c4SDavid Hildenbrand newpmd = pmd_swp_mkuffd_wp(newpmd); 207024bf08c4SDavid Hildenbrand else if (uffd_wp_resolve) 207124bf08c4SDavid Hildenbrand newpmd = pmd_swp_clear_uffd_wp(newpmd); 207224bf08c4SDavid Hildenbrand if (!pmd_same(*pmd, newpmd)) 207324bf08c4SDavid Hildenbrand set_pmd_at(mm, addr, pmd, newpmd); 207484c3fc4eSZi Yan goto unlock; 207584c3fc4eSZi Yan } 207684c3fc4eSZi Yan #endif 207784c3fc4eSZi Yan 2078a1a3a2fcSHuang Ying if (prot_numa) { 2079d986ba2bSKefeng Wang struct folio *folio; 208033024536SHuang Ying bool toptier; 2081e944fd67SMel Gorman /* 2082e944fd67SMel Gorman * Avoid trapping faults against the zero page. The read-only 2083e944fd67SMel Gorman * data is likely to be read-cached on the local CPU and 2084e944fd67SMel Gorman * local/remote hits to the zero page are not interesting. 2085e944fd67SMel Gorman */ 2086a1a3a2fcSHuang Ying if (is_huge_zero_pmd(*pmd)) 20870a85e51dSKirill A. Shutemov goto unlock; 2088e944fd67SMel Gorman 2089a1a3a2fcSHuang Ying if (pmd_protnone(*pmd)) 20900a85e51dSKirill A. Shutemov goto unlock; 20910a85e51dSKirill A. Shutemov 2092d986ba2bSKefeng Wang folio = page_folio(pmd_page(*pmd)); 2093d986ba2bSKefeng Wang toptier = node_is_toptier(folio_nid(folio)); 2094a1a3a2fcSHuang Ying /* 2095a1a3a2fcSHuang Ying * Skip scanning top tier node if normal numa 2096a1a3a2fcSHuang Ying * balancing is disabled 2097a1a3a2fcSHuang Ying */ 2098a1a3a2fcSHuang Ying if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) && 209933024536SHuang Ying toptier) 2100a1a3a2fcSHuang Ying goto unlock; 210133024536SHuang Ying 210233024536SHuang Ying if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING && 210333024536SHuang Ying !toptier) 2104d986ba2bSKefeng Wang folio_xchg_access_time(folio, 2105d986ba2bSKefeng Wang jiffies_to_msecs(jiffies)); 2106a1a3a2fcSHuang Ying } 2107ced10803SKirill A. Shutemov /* 21083e4e28c5SMichel Lespinasse * In case prot_numa, we are under mmap_read_lock(mm). It's critical 2109ced10803SKirill A. Shutemov * to not clear pmd intermittently to avoid race with MADV_DONTNEED 21103e4e28c5SMichel Lespinasse * which is also under mmap_read_lock(mm): 2111ced10803SKirill A. Shutemov * 2112ced10803SKirill A. Shutemov * CPU0: CPU1: 2113ced10803SKirill A. Shutemov * change_huge_pmd(prot_numa=1) 2114ced10803SKirill A. Shutemov * pmdp_huge_get_and_clear_notify() 2115ced10803SKirill A. Shutemov * madvise_dontneed() 2116ced10803SKirill A. Shutemov * zap_pmd_range() 2117ced10803SKirill A. Shutemov * pmd_trans_huge(*pmd) == 0 (without ptl) 2118ced10803SKirill A. Shutemov * // skip the pmd 2119ced10803SKirill A. Shutemov * set_pmd_at(); 2120ced10803SKirill A. Shutemov * // pmd is re-established 2121ced10803SKirill A. Shutemov * 2122ced10803SKirill A. Shutemov * The race makes MADV_DONTNEED miss the huge pmd and don't clear it 2123ced10803SKirill A. Shutemov * which may break userspace. 2124ced10803SKirill A. Shutemov * 21254f831457SNadav Amit * pmdp_invalidate_ad() is required to make sure we don't miss 2126ced10803SKirill A. Shutemov * dirty/young flags set by hardware. 2127ced10803SKirill A. Shutemov */ 21284f831457SNadav Amit oldpmd = pmdp_invalidate_ad(vma, addr, pmd); 2129ced10803SKirill A. Shutemov 2130c9fe6656SNadav Amit entry = pmd_modify(oldpmd, newprot); 2131f1eb1bacSPeter Xu if (uffd_wp) 2132292924b2SPeter Xu entry = pmd_mkuffd_wp(entry); 2133f1eb1bacSPeter Xu else if (uffd_wp_resolve) 2134292924b2SPeter Xu /* 2135292924b2SPeter Xu * Leave the write bit to be handled by PF interrupt 2136292924b2SPeter Xu * handler, then things like COW could be properly 2137292924b2SPeter Xu * handled. 2138292924b2SPeter Xu */ 2139292924b2SPeter Xu entry = pmd_clear_uffd_wp(entry); 2140c27f479eSDavid Hildenbrand 2141c27f479eSDavid Hildenbrand /* See change_pte_range(). */ 2142c27f479eSDavid Hildenbrand if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) && 2143c27f479eSDavid Hildenbrand can_change_pmd_writable(vma, addr, entry)) 2144161e393cSRick Edgecombe entry = pmd_mkwrite(entry, vma); 2145c27f479eSDavid Hildenbrand 2146f123d74aSMel Gorman ret = HPAGE_PMD_NR; 214756eecdb9SAneesh Kumar K.V set_pmd_at(mm, addr, pmd, entry); 21484a18419fSNadav Amit 2149c9fe6656SNadav Amit if (huge_pmd_needs_flush(oldpmd, entry)) 21504a18419fSNadav Amit tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE); 21510a85e51dSKirill A. Shutemov unlock: 2152bf929152SKirill A. Shutemov spin_unlock(ptl); 2153cd7548abSJohannes Weiner return ret; 2154cd7548abSJohannes Weiner } 2155cd7548abSJohannes Weiner 2156adef4406SAndrea Arcangeli #ifdef CONFIG_USERFAULTFD 2157adef4406SAndrea Arcangeli /* 2158adef4406SAndrea Arcangeli * The PT lock for src_pmd and the mmap_lock for reading are held by 2159adef4406SAndrea Arcangeli * the caller, but it must return after releasing the page_table_lock. 2160adef4406SAndrea Arcangeli * Just move the page from src_pmd to dst_pmd if possible. 2161adef4406SAndrea Arcangeli * Return zero if succeeded in moving the page, -EAGAIN if it needs to be 2162adef4406SAndrea Arcangeli * repeated by the caller, or other errors in case of failure. 2163adef4406SAndrea Arcangeli */ 2164adef4406SAndrea Arcangeli int move_pages_huge_pmd(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd, pmd_t dst_pmdval, 2165adef4406SAndrea Arcangeli struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 2166adef4406SAndrea Arcangeli unsigned long dst_addr, unsigned long src_addr) 2167adef4406SAndrea Arcangeli { 2168adef4406SAndrea Arcangeli pmd_t _dst_pmd, src_pmdval; 2169adef4406SAndrea Arcangeli struct page *src_page; 2170adef4406SAndrea Arcangeli struct folio *src_folio; 2171adef4406SAndrea Arcangeli struct anon_vma *src_anon_vma; 2172adef4406SAndrea Arcangeli spinlock_t *src_ptl, *dst_ptl; 2173adef4406SAndrea Arcangeli pgtable_t src_pgtable; 2174adef4406SAndrea Arcangeli struct mmu_notifier_range range; 2175adef4406SAndrea Arcangeli int err = 0; 2176adef4406SAndrea Arcangeli 2177adef4406SAndrea Arcangeli src_pmdval = *src_pmd; 2178adef4406SAndrea Arcangeli src_ptl = pmd_lockptr(mm, src_pmd); 2179adef4406SAndrea Arcangeli 2180adef4406SAndrea Arcangeli lockdep_assert_held(src_ptl); 2181adef4406SAndrea Arcangeli mmap_assert_locked(mm); 2182adef4406SAndrea Arcangeli 2183adef4406SAndrea Arcangeli /* Sanity checks before the operation */ 2184adef4406SAndrea Arcangeli if (WARN_ON_ONCE(!pmd_none(dst_pmdval)) || WARN_ON_ONCE(src_addr & ~HPAGE_PMD_MASK) || 2185adef4406SAndrea Arcangeli WARN_ON_ONCE(dst_addr & ~HPAGE_PMD_MASK)) { 2186adef4406SAndrea Arcangeli spin_unlock(src_ptl); 2187adef4406SAndrea Arcangeli return -EINVAL; 2188adef4406SAndrea Arcangeli } 2189adef4406SAndrea Arcangeli 2190adef4406SAndrea Arcangeli if (!pmd_trans_huge(src_pmdval)) { 2191adef4406SAndrea Arcangeli spin_unlock(src_ptl); 2192adef4406SAndrea Arcangeli if (is_pmd_migration_entry(src_pmdval)) { 2193adef4406SAndrea Arcangeli pmd_migration_entry_wait(mm, &src_pmdval); 2194adef4406SAndrea Arcangeli return -EAGAIN; 2195adef4406SAndrea Arcangeli } 2196adef4406SAndrea Arcangeli return -ENOENT; 2197adef4406SAndrea Arcangeli } 2198adef4406SAndrea Arcangeli 2199adef4406SAndrea Arcangeli src_page = pmd_page(src_pmdval); 2200adef4406SAndrea Arcangeli if (unlikely(!PageAnonExclusive(src_page))) { 2201adef4406SAndrea Arcangeli spin_unlock(src_ptl); 2202adef4406SAndrea Arcangeli return -EBUSY; 2203adef4406SAndrea Arcangeli } 2204adef4406SAndrea Arcangeli 2205adef4406SAndrea Arcangeli src_folio = page_folio(src_page); 2206adef4406SAndrea Arcangeli folio_get(src_folio); 2207adef4406SAndrea Arcangeli spin_unlock(src_ptl); 2208adef4406SAndrea Arcangeli 2209adef4406SAndrea Arcangeli flush_cache_range(src_vma, src_addr, src_addr + HPAGE_PMD_SIZE); 2210adef4406SAndrea Arcangeli mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, src_addr, 2211adef4406SAndrea Arcangeli src_addr + HPAGE_PMD_SIZE); 2212adef4406SAndrea Arcangeli mmu_notifier_invalidate_range_start(&range); 2213adef4406SAndrea Arcangeli 2214adef4406SAndrea Arcangeli folio_lock(src_folio); 2215adef4406SAndrea Arcangeli 2216adef4406SAndrea Arcangeli /* 2217adef4406SAndrea Arcangeli * split_huge_page walks the anon_vma chain without the page 2218adef4406SAndrea Arcangeli * lock. Serialize against it with the anon_vma lock, the page 2219adef4406SAndrea Arcangeli * lock is not enough. 2220adef4406SAndrea Arcangeli */ 2221adef4406SAndrea Arcangeli src_anon_vma = folio_get_anon_vma(src_folio); 2222adef4406SAndrea Arcangeli if (!src_anon_vma) { 2223adef4406SAndrea Arcangeli err = -EAGAIN; 2224adef4406SAndrea Arcangeli goto unlock_folio; 2225adef4406SAndrea Arcangeli } 2226adef4406SAndrea Arcangeli anon_vma_lock_write(src_anon_vma); 2227adef4406SAndrea Arcangeli 2228adef4406SAndrea Arcangeli dst_ptl = pmd_lockptr(mm, dst_pmd); 2229adef4406SAndrea Arcangeli double_pt_lock(src_ptl, dst_ptl); 2230adef4406SAndrea Arcangeli if (unlikely(!pmd_same(*src_pmd, src_pmdval) || 2231adef4406SAndrea Arcangeli !pmd_same(*dst_pmd, dst_pmdval))) { 2232adef4406SAndrea Arcangeli err = -EAGAIN; 2233adef4406SAndrea Arcangeli goto unlock_ptls; 2234adef4406SAndrea Arcangeli } 2235adef4406SAndrea Arcangeli if (folio_maybe_dma_pinned(src_folio) || 2236adef4406SAndrea Arcangeli !PageAnonExclusive(&src_folio->page)) { 2237adef4406SAndrea Arcangeli err = -EBUSY; 2238adef4406SAndrea Arcangeli goto unlock_ptls; 2239adef4406SAndrea Arcangeli } 2240adef4406SAndrea Arcangeli 2241adef4406SAndrea Arcangeli if (WARN_ON_ONCE(!folio_test_head(src_folio)) || 2242adef4406SAndrea Arcangeli WARN_ON_ONCE(!folio_test_anon(src_folio))) { 2243adef4406SAndrea Arcangeli err = -EBUSY; 2244adef4406SAndrea Arcangeli goto unlock_ptls; 2245adef4406SAndrea Arcangeli } 2246adef4406SAndrea Arcangeli 2247adef4406SAndrea Arcangeli folio_move_anon_rmap(src_folio, dst_vma); 2248adef4406SAndrea Arcangeli WRITE_ONCE(src_folio->index, linear_page_index(dst_vma, dst_addr)); 2249adef4406SAndrea Arcangeli 2250adef4406SAndrea Arcangeli src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd); 2251adef4406SAndrea Arcangeli /* Folio got pinned from under us. Put it back and fail the move. */ 2252adef4406SAndrea Arcangeli if (folio_maybe_dma_pinned(src_folio)) { 2253adef4406SAndrea Arcangeli set_pmd_at(mm, src_addr, src_pmd, src_pmdval); 2254adef4406SAndrea Arcangeli err = -EBUSY; 2255adef4406SAndrea Arcangeli goto unlock_ptls; 2256adef4406SAndrea Arcangeli } 2257adef4406SAndrea Arcangeli 2258adef4406SAndrea Arcangeli _dst_pmd = mk_huge_pmd(&src_folio->page, dst_vma->vm_page_prot); 2259adef4406SAndrea Arcangeli /* Follow mremap() behavior and treat the entry dirty after the move */ 2260adef4406SAndrea Arcangeli _dst_pmd = pmd_mkwrite(pmd_mkdirty(_dst_pmd), dst_vma); 2261adef4406SAndrea Arcangeli set_pmd_at(mm, dst_addr, dst_pmd, _dst_pmd); 2262adef4406SAndrea Arcangeli 2263adef4406SAndrea Arcangeli src_pgtable = pgtable_trans_huge_withdraw(mm, src_pmd); 2264adef4406SAndrea Arcangeli pgtable_trans_huge_deposit(mm, dst_pmd, src_pgtable); 2265adef4406SAndrea Arcangeli unlock_ptls: 2266adef4406SAndrea Arcangeli double_pt_unlock(src_ptl, dst_ptl); 2267adef4406SAndrea Arcangeli anon_vma_unlock_write(src_anon_vma); 2268adef4406SAndrea Arcangeli put_anon_vma(src_anon_vma); 2269adef4406SAndrea Arcangeli unlock_folio: 2270adef4406SAndrea Arcangeli /* unblock rmap walks */ 2271adef4406SAndrea Arcangeli folio_unlock(src_folio); 2272adef4406SAndrea Arcangeli mmu_notifier_invalidate_range_end(&range); 2273adef4406SAndrea Arcangeli folio_put(src_folio); 2274adef4406SAndrea Arcangeli return err; 2275adef4406SAndrea Arcangeli } 2276adef4406SAndrea Arcangeli #endif /* CONFIG_USERFAULTFD */ 2277adef4406SAndrea Arcangeli 2278025c5b24SNaoya Horiguchi /* 22798f19b0c0SHuang Ying * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. 2280025c5b24SNaoya Horiguchi * 22818f19b0c0SHuang Ying * Note that if it returns page table lock pointer, this routine returns without 22828f19b0c0SHuang Ying * unlocking page table lock. So callers must unlock it. 2283025c5b24SNaoya Horiguchi */ 2284b6ec57f4SKirill A. Shutemov spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) 2285025c5b24SNaoya Horiguchi { 2286b6ec57f4SKirill A. Shutemov spinlock_t *ptl; 2287b6ec57f4SKirill A. Shutemov ptl = pmd_lock(vma->vm_mm, pmd); 228884c3fc4eSZi Yan if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || 228984c3fc4eSZi Yan pmd_devmap(*pmd))) 2290b6ec57f4SKirill A. Shutemov return ptl; 2291b6ec57f4SKirill A. Shutemov spin_unlock(ptl); 2292b6ec57f4SKirill A. Shutemov return NULL; 2293025c5b24SNaoya Horiguchi } 2294025c5b24SNaoya Horiguchi 2295a00cc7d9SMatthew Wilcox /* 2296d965e390SMiaohe Lin * Returns page table lock pointer if a given pud maps a thp, NULL otherwise. 2297a00cc7d9SMatthew Wilcox * 2298d965e390SMiaohe Lin * Note that if it returns page table lock pointer, this routine returns without 2299d965e390SMiaohe Lin * unlocking page table lock. So callers must unlock it. 2300a00cc7d9SMatthew Wilcox */ 2301a00cc7d9SMatthew Wilcox spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) 2302a00cc7d9SMatthew Wilcox { 2303a00cc7d9SMatthew Wilcox spinlock_t *ptl; 2304a00cc7d9SMatthew Wilcox 2305a00cc7d9SMatthew Wilcox ptl = pud_lock(vma->vm_mm, pud); 2306a00cc7d9SMatthew Wilcox if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) 2307a00cc7d9SMatthew Wilcox return ptl; 2308a00cc7d9SMatthew Wilcox spin_unlock(ptl); 2309a00cc7d9SMatthew Wilcox return NULL; 2310a00cc7d9SMatthew Wilcox } 2311a00cc7d9SMatthew Wilcox 2312a00cc7d9SMatthew Wilcox #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD 2313a00cc7d9SMatthew Wilcox int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, 2314a00cc7d9SMatthew Wilcox pud_t *pud, unsigned long addr) 2315a00cc7d9SMatthew Wilcox { 2316a00cc7d9SMatthew Wilcox spinlock_t *ptl; 2317a00cc7d9SMatthew Wilcox 2318a00cc7d9SMatthew Wilcox ptl = __pud_trans_huge_lock(pud, vma); 2319a00cc7d9SMatthew Wilcox if (!ptl) 2320a00cc7d9SMatthew Wilcox return 0; 232174929079SMiaohe Lin 2322f32928abSAneesh Kumar K.V pudp_huge_get_and_clear_full(vma, addr, pud, tlb->fullmm); 2323a00cc7d9SMatthew Wilcox tlb_remove_pud_tlb_entry(tlb, pud, addr); 23242484ca9bSThomas Hellstrom (VMware) if (vma_is_special_huge(vma)) { 2325a00cc7d9SMatthew Wilcox spin_unlock(ptl); 2326a00cc7d9SMatthew Wilcox /* No zero page support yet */ 2327a00cc7d9SMatthew Wilcox } else { 2328a00cc7d9SMatthew Wilcox /* No support for anonymous PUD pages yet */ 2329a00cc7d9SMatthew Wilcox BUG(); 2330a00cc7d9SMatthew Wilcox } 2331a00cc7d9SMatthew Wilcox return 1; 2332a00cc7d9SMatthew Wilcox } 2333a00cc7d9SMatthew Wilcox 2334a00cc7d9SMatthew Wilcox static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, 2335a00cc7d9SMatthew Wilcox unsigned long haddr) 2336a00cc7d9SMatthew Wilcox { 2337a00cc7d9SMatthew Wilcox VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); 2338a00cc7d9SMatthew Wilcox VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2339a00cc7d9SMatthew Wilcox VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); 2340a00cc7d9SMatthew Wilcox VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); 2341a00cc7d9SMatthew Wilcox 2342ce9311cfSYisheng Xie count_vm_event(THP_SPLIT_PUD); 2343a00cc7d9SMatthew Wilcox 2344ec8832d0SAlistair Popple pudp_huge_clear_flush(vma, haddr, pud); 2345a00cc7d9SMatthew Wilcox } 2346a00cc7d9SMatthew Wilcox 2347a00cc7d9SMatthew Wilcox void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, 2348a00cc7d9SMatthew Wilcox unsigned long address) 2349a00cc7d9SMatthew Wilcox { 2350a00cc7d9SMatthew Wilcox spinlock_t *ptl; 2351ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 2352a00cc7d9SMatthew Wilcox 23537d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 23546f4f13e8SJérôme Glisse address & HPAGE_PUD_MASK, 2355ac46d4f3SJérôme Glisse (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE); 2356ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 2357ac46d4f3SJérôme Glisse ptl = pud_lock(vma->vm_mm, pud); 2358a00cc7d9SMatthew Wilcox if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) 2359a00cc7d9SMatthew Wilcox goto out; 2360ac46d4f3SJérôme Glisse __split_huge_pud_locked(vma, pud, range.start); 2361a00cc7d9SMatthew Wilcox 2362a00cc7d9SMatthew Wilcox out: 2363a00cc7d9SMatthew Wilcox spin_unlock(ptl); 2364ec8832d0SAlistair Popple mmu_notifier_invalidate_range_end(&range); 2365a00cc7d9SMatthew Wilcox } 2366a00cc7d9SMatthew Wilcox #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 2367a00cc7d9SMatthew Wilcox 2368eef1b3baSKirill A. Shutemov static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 2369eef1b3baSKirill A. Shutemov unsigned long haddr, pmd_t *pmd) 2370eef1b3baSKirill A. Shutemov { 2371eef1b3baSKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 2372eef1b3baSKirill A. Shutemov pgtable_t pgtable; 237342b2af2cSDavid Hildenbrand pmd_t _pmd, old_pmd; 2374c9c1ee20SHugh Dickins unsigned long addr; 2375c9c1ee20SHugh Dickins pte_t *pte; 2376eef1b3baSKirill A. Shutemov int i; 2377eef1b3baSKirill A. Shutemov 23780f10851eSJérôme Glisse /* 23790f10851eSJérôme Glisse * Leave pmd empty until pte is filled note that it is fine to delay 23800f10851eSJérôme Glisse * notification until mmu_notifier_invalidate_range_end() as we are 23810f10851eSJérôme Glisse * replacing a zero pmd write protected page with a zero pte write 23820f10851eSJérôme Glisse * protected page. 23830f10851eSJérôme Glisse * 2384ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 23850f10851eSJérôme Glisse */ 238642b2af2cSDavid Hildenbrand old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd); 2387eef1b3baSKirill A. Shutemov 2388eef1b3baSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2389eef1b3baSKirill A. Shutemov pmd_populate(mm, &_pmd, pgtable); 2390eef1b3baSKirill A. Shutemov 2391c9c1ee20SHugh Dickins pte = pte_offset_map(&_pmd, haddr); 2392c9c1ee20SHugh Dickins VM_BUG_ON(!pte); 2393c9c1ee20SHugh Dickins for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2394c9c1ee20SHugh Dickins pte_t entry; 2395c9c1ee20SHugh Dickins 2396c9c1ee20SHugh Dickins entry = pfn_pte(my_zero_pfn(addr), vma->vm_page_prot); 2397eef1b3baSKirill A. Shutemov entry = pte_mkspecial(entry); 239842b2af2cSDavid Hildenbrand if (pmd_uffd_wp(old_pmd)) 239942b2af2cSDavid Hildenbrand entry = pte_mkuffd_wp(entry); 2400c33c7948SRyan Roberts VM_BUG_ON(!pte_none(ptep_get(pte))); 2401c9c1ee20SHugh Dickins set_pte_at(mm, addr, pte, entry); 2402c9c1ee20SHugh Dickins pte++; 2403eef1b3baSKirill A. Shutemov } 2404c9c1ee20SHugh Dickins pte_unmap(pte - 1); 2405eef1b3baSKirill A. Shutemov smp_wmb(); /* make pte visible before pmd */ 2406eef1b3baSKirill A. Shutemov pmd_populate(mm, pmd, pgtable); 2407eef1b3baSKirill A. Shutemov } 2408eef1b3baSKirill A. Shutemov 2409eef1b3baSKirill A. Shutemov static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, 2410ba988280SKirill A. Shutemov unsigned long haddr, bool freeze) 2411eef1b3baSKirill A. Shutemov { 2412eef1b3baSKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 241391b2978aSDavid Hildenbrand struct folio *folio; 2414eef1b3baSKirill A. Shutemov struct page *page; 2415eef1b3baSKirill A. Shutemov pgtable_t pgtable; 2416423ac9afSAneesh Kumar K.V pmd_t old_pmd, _pmd; 2417292924b2SPeter Xu bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false; 24180ccf7f16SPeter Xu bool anon_exclusive = false, dirty = false; 24192ac015e2SKirill A. Shutemov unsigned long addr; 2420c9c1ee20SHugh Dickins pte_t *pte; 2421eef1b3baSKirill A. Shutemov int i; 2422eef1b3baSKirill A. Shutemov 2423eef1b3baSKirill A. Shutemov VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); 2424eef1b3baSKirill A. Shutemov VM_BUG_ON_VMA(vma->vm_start > haddr, vma); 2425eef1b3baSKirill A. Shutemov VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); 242684c3fc4eSZi Yan VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) 242784c3fc4eSZi Yan && !pmd_devmap(*pmd)); 2428eef1b3baSKirill A. Shutemov 2429eef1b3baSKirill A. Shutemov count_vm_event(THP_SPLIT_PMD); 2430eef1b3baSKirill A. Shutemov 2431d21b9e57SKirill A. Shutemov if (!vma_is_anonymous(vma)) { 2432ec8832d0SAlistair Popple old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd); 2433953c66c2SAneesh Kumar K.V /* 2434953c66c2SAneesh Kumar K.V * We are going to unmap this huge page. So 2435953c66c2SAneesh Kumar K.V * just go ahead and zap it 2436953c66c2SAneesh Kumar K.V */ 2437953c66c2SAneesh Kumar K.V if (arch_needs_pgtable_deposit()) 2438953c66c2SAneesh Kumar K.V zap_deposited_table(mm, pmd); 24392484ca9bSThomas Hellstrom (VMware) if (vma_is_special_huge(vma)) 2440d21b9e57SKirill A. Shutemov return; 244199fa8a48SHugh Dickins if (unlikely(is_pmd_migration_entry(old_pmd))) { 244299fa8a48SHugh Dickins swp_entry_t entry; 244399fa8a48SHugh Dickins 244499fa8a48SHugh Dickins entry = pmd_to_swp_entry(old_pmd); 2445af5cdaf8SAlistair Popple page = pfn_swap_entry_to_page(entry); 244699fa8a48SHugh Dickins } else { 244799fa8a48SHugh Dickins page = pmd_page(old_pmd); 2448a8e61d58SDavid Hildenbrand folio = page_folio(page); 2449a8e61d58SDavid Hildenbrand if (!folio_test_dirty(folio) && pmd_dirty(old_pmd)) 2450db44c658SDavid Hildenbrand folio_mark_dirty(folio); 2451a8e61d58SDavid Hildenbrand if (!folio_test_referenced(folio) && pmd_young(old_pmd)) 2452a8e61d58SDavid Hildenbrand folio_set_referenced(folio); 2453a8e61d58SDavid Hildenbrand folio_remove_rmap_pmd(folio, page, vma); 2454a8e61d58SDavid Hildenbrand folio_put(folio); 245599fa8a48SHugh Dickins } 2456fadae295SYang Shi add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR); 2457eef1b3baSKirill A. Shutemov return; 245899fa8a48SHugh Dickins } 245999fa8a48SHugh Dickins 24603b77e8c8SHugh Dickins if (is_huge_zero_pmd(*pmd)) { 24614645b9feSJérôme Glisse /* 24624645b9feSJérôme Glisse * FIXME: Do we want to invalidate secondary mmu by calling 24631af5a810SAlistair Popple * mmu_notifier_arch_invalidate_secondary_tlbs() see comments below 24641af5a810SAlistair Popple * inside __split_huge_pmd() ? 24654645b9feSJérôme Glisse * 24664645b9feSJérôme Glisse * We are going from a zero huge page write protected to zero 24674645b9feSJérôme Glisse * small page also write protected so it does not seems useful 24684645b9feSJérôme Glisse * to invalidate secondary mmu at this time. 24694645b9feSJérôme Glisse */ 2470eef1b3baSKirill A. Shutemov return __split_huge_zero_page_pmd(vma, haddr, pmd); 2471eef1b3baSKirill A. Shutemov } 2472eef1b3baSKirill A. Shutemov 2473423ac9afSAneesh Kumar K.V /* 2474423ac9afSAneesh Kumar K.V * Up to this point the pmd is present and huge and userland has the 2475423ac9afSAneesh Kumar K.V * whole access to the hugepage during the split (which happens in 2476423ac9afSAneesh Kumar K.V * place). If we overwrite the pmd with the not-huge version pointing 2477423ac9afSAneesh Kumar K.V * to the pte here (which of course we could if all CPUs were bug 2478423ac9afSAneesh Kumar K.V * free), userland could trigger a small page size TLB miss on the 2479423ac9afSAneesh Kumar K.V * small sized TLB while the hugepage TLB entry is still established in 2480423ac9afSAneesh Kumar K.V * the huge TLB. Some CPU doesn't like that. 248142742d9bSAlexander A. Klimov * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum 248242742d9bSAlexander A. Klimov * 383 on page 105. Intel should be safe but is also warns that it's 2483423ac9afSAneesh Kumar K.V * only safe if the permission and cache attributes of the two entries 2484423ac9afSAneesh Kumar K.V * loaded in the two TLB is identical (which should be the case here). 2485423ac9afSAneesh Kumar K.V * But it is generally safer to never allow small and huge TLB entries 2486423ac9afSAneesh Kumar K.V * for the same virtual address to be loaded simultaneously. So instead 2487423ac9afSAneesh Kumar K.V * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the 2488423ac9afSAneesh Kumar K.V * current pmd notpresent (atomically because here the pmd_trans_huge 2489423ac9afSAneesh Kumar K.V * must remain set at all times on the pmd until the split is complete 2490423ac9afSAneesh Kumar K.V * for this pmd), then we flush the SMP TLB and finally we write the 2491423ac9afSAneesh Kumar K.V * non-huge version of the pmd entry with pmd_populate. 2492423ac9afSAneesh Kumar K.V */ 2493423ac9afSAneesh Kumar K.V old_pmd = pmdp_invalidate(vma, haddr, pmd); 2494423ac9afSAneesh Kumar K.V 2495423ac9afSAneesh Kumar K.V pmd_migration = is_pmd_migration_entry(old_pmd); 24962e83ee1dSPeter Xu if (unlikely(pmd_migration)) { 249784c3fc4eSZi Yan swp_entry_t entry; 249884c3fc4eSZi Yan 2499423ac9afSAneesh Kumar K.V entry = pmd_to_swp_entry(old_pmd); 2500af5cdaf8SAlistair Popple page = pfn_swap_entry_to_page(entry); 25014dd845b5SAlistair Popple write = is_writable_migration_entry(entry); 25026c287605SDavid Hildenbrand if (PageAnon(page)) 25036c287605SDavid Hildenbrand anon_exclusive = is_readable_exclusive_migration_entry(entry); 25042e346877SPeter Xu young = is_migration_entry_young(entry); 25052e346877SPeter Xu dirty = is_migration_entry_dirty(entry); 25062e83ee1dSPeter Xu soft_dirty = pmd_swp_soft_dirty(old_pmd); 2507f45ec5ffSPeter Xu uffd_wp = pmd_swp_uffd_wp(old_pmd); 25082e83ee1dSPeter Xu } else { 2509423ac9afSAneesh Kumar K.V page = pmd_page(old_pmd); 251091b2978aSDavid Hildenbrand folio = page_folio(page); 25110ccf7f16SPeter Xu if (pmd_dirty(old_pmd)) { 25120ccf7f16SPeter Xu dirty = true; 251391b2978aSDavid Hildenbrand folio_set_dirty(folio); 25140ccf7f16SPeter Xu } 2515423ac9afSAneesh Kumar K.V write = pmd_write(old_pmd); 2516423ac9afSAneesh Kumar K.V young = pmd_young(old_pmd); 2517423ac9afSAneesh Kumar K.V soft_dirty = pmd_soft_dirty(old_pmd); 2518292924b2SPeter Xu uffd_wp = pmd_uffd_wp(old_pmd); 25196c287605SDavid Hildenbrand 252091b2978aSDavid Hildenbrand VM_WARN_ON_FOLIO(!folio_ref_count(folio), folio); 252191b2978aSDavid Hildenbrand VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio); 25226c287605SDavid Hildenbrand 25236c287605SDavid Hildenbrand /* 25246c287605SDavid Hildenbrand * Without "freeze", we'll simply split the PMD, propagating the 25256c287605SDavid Hildenbrand * PageAnonExclusive() flag for each PTE by setting it for 25266c287605SDavid Hildenbrand * each subpage -- no need to (temporarily) clear. 25276c287605SDavid Hildenbrand * 25286c287605SDavid Hildenbrand * With "freeze" we want to replace mapped pages by 25296c287605SDavid Hildenbrand * migration entries right away. This is only possible if we 25306c287605SDavid Hildenbrand * managed to clear PageAnonExclusive() -- see 25316c287605SDavid Hildenbrand * set_pmd_migration_entry(). 25326c287605SDavid Hildenbrand * 25336c287605SDavid Hildenbrand * In case we cannot clear PageAnonExclusive(), split the PMD 25346c287605SDavid Hildenbrand * only and let try_to_migrate_one() fail later. 2535088b8aa5SDavid Hildenbrand * 2536e3b4b137SDavid Hildenbrand * See folio_try_share_anon_rmap_pmd(): invalidate PMD first. 25376c287605SDavid Hildenbrand */ 253891b2978aSDavid Hildenbrand anon_exclusive = PageAnonExclusive(page); 2539e3b4b137SDavid Hildenbrand if (freeze && anon_exclusive && 2540e3b4b137SDavid Hildenbrand folio_try_share_anon_rmap_pmd(folio, page)) 25416c287605SDavid Hildenbrand freeze = false; 254291b2978aSDavid Hildenbrand if (!freeze) { 254391b2978aSDavid Hildenbrand rmap_t rmap_flags = RMAP_NONE; 254491b2978aSDavid Hildenbrand 254591b2978aSDavid Hildenbrand folio_ref_add(folio, HPAGE_PMD_NR - 1); 254691b2978aSDavid Hildenbrand if (anon_exclusive) 254791b2978aSDavid Hildenbrand rmap_flags |= RMAP_EXCLUSIVE; 254891b2978aSDavid Hildenbrand folio_add_anon_rmap_ptes(folio, page, HPAGE_PMD_NR, 254991b2978aSDavid Hildenbrand vma, haddr, rmap_flags); 255091b2978aSDavid Hildenbrand } 25519d84604bSHugh Dickins } 2552eef1b3baSKirill A. Shutemov 2553423ac9afSAneesh Kumar K.V /* 2554423ac9afSAneesh Kumar K.V * Withdraw the table only after we mark the pmd entry invalid. 2555423ac9afSAneesh Kumar K.V * This's critical for some architectures (Power). 2556423ac9afSAneesh Kumar K.V */ 2557eef1b3baSKirill A. Shutemov pgtable = pgtable_trans_huge_withdraw(mm, pmd); 2558eef1b3baSKirill A. Shutemov pmd_populate(mm, &_pmd, pgtable); 2559eef1b3baSKirill A. Shutemov 2560c9c1ee20SHugh Dickins pte = pte_offset_map(&_pmd, haddr); 2561c9c1ee20SHugh Dickins VM_BUG_ON(!pte); 25622ac015e2SKirill A. Shutemov for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { 2563c9c1ee20SHugh Dickins pte_t entry; 2564eef1b3baSKirill A. Shutemov /* 2565eef1b3baSKirill A. Shutemov * Note that NUMA hinting access restrictions are not 2566eef1b3baSKirill A. Shutemov * transferred to avoid any possibility of altering 2567eef1b3baSKirill A. Shutemov * permissions across VMAs. 2568eef1b3baSKirill A. Shutemov */ 256984c3fc4eSZi Yan if (freeze || pmd_migration) { 2570ba988280SKirill A. Shutemov swp_entry_t swp_entry; 25714dd845b5SAlistair Popple if (write) 25724dd845b5SAlistair Popple swp_entry = make_writable_migration_entry( 25734dd845b5SAlistair Popple page_to_pfn(page + i)); 25746c287605SDavid Hildenbrand else if (anon_exclusive) 25756c287605SDavid Hildenbrand swp_entry = make_readable_exclusive_migration_entry( 25766c287605SDavid Hildenbrand page_to_pfn(page + i)); 25774dd845b5SAlistair Popple else 25784dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 25794dd845b5SAlistair Popple page_to_pfn(page + i)); 25802e346877SPeter Xu if (young) 25812e346877SPeter Xu swp_entry = make_migration_entry_young(swp_entry); 25822e346877SPeter Xu if (dirty) 25832e346877SPeter Xu swp_entry = make_migration_entry_dirty(swp_entry); 2584ba988280SKirill A. Shutemov entry = swp_entry_to_pte(swp_entry); 2585804dd150SAndrea Arcangeli if (soft_dirty) 2586804dd150SAndrea Arcangeli entry = pte_swp_mksoft_dirty(entry); 2587f45ec5ffSPeter Xu if (uffd_wp) 2588f45ec5ffSPeter Xu entry = pte_swp_mkuffd_wp(entry); 2589ba988280SKirill A. Shutemov } else { 25906d2329f8SAndrea Arcangeli entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); 25911462c52eSDavid Hildenbrand if (write) 2592161e393cSRick Edgecombe entry = pte_mkwrite(entry, vma); 2593eef1b3baSKirill A. Shutemov if (!young) 2594eef1b3baSKirill A. Shutemov entry = pte_mkold(entry); 2595e833bc50SPeter Xu /* NOTE: this may set soft-dirty too on some archs */ 2596e833bc50SPeter Xu if (dirty) 2597e833bc50SPeter Xu entry = pte_mkdirty(entry); 2598804dd150SAndrea Arcangeli if (soft_dirty) 2599804dd150SAndrea Arcangeli entry = pte_mksoft_dirty(entry); 2600292924b2SPeter Xu if (uffd_wp) 2601292924b2SPeter Xu entry = pte_mkuffd_wp(entry); 2602ba988280SKirill A. Shutemov } 2603c33c7948SRyan Roberts VM_BUG_ON(!pte_none(ptep_get(pte))); 26042ac015e2SKirill A. Shutemov set_pte_at(mm, addr, pte, entry); 2605c9c1ee20SHugh Dickins pte++; 2606eef1b3baSKirill A. Shutemov } 2607c9c1ee20SHugh Dickins pte_unmap(pte - 1); 2608eef1b3baSKirill A. Shutemov 2609cb67f428SHugh Dickins if (!pmd_migration) 2610a8e61d58SDavid Hildenbrand folio_remove_rmap_pmd(folio, page, vma); 261196d82debSHugh Dickins if (freeze) 261296d82debSHugh Dickins put_page(page); 2613eef1b3baSKirill A. Shutemov 2614eef1b3baSKirill A. Shutemov smp_wmb(); /* make pte visible before pmd */ 2615eef1b3baSKirill A. Shutemov pmd_populate(mm, pmd, pgtable); 2616eef1b3baSKirill A. Shutemov } 2617eef1b3baSKirill A. Shutemov 2618eef1b3baSKirill A. Shutemov void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, 2619af28a988SMatthew Wilcox (Oracle) unsigned long address, bool freeze, struct folio *folio) 2620eef1b3baSKirill A. Shutemov { 2621eef1b3baSKirill A. Shutemov spinlock_t *ptl; 2622ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 2623eef1b3baSKirill A. Shutemov 26247d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 26256f4f13e8SJérôme Glisse address & HPAGE_PMD_MASK, 2626ac46d4f3SJérôme Glisse (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE); 2627ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 2628ac46d4f3SJérôme Glisse ptl = pmd_lock(vma->vm_mm, pmd); 262933f4751eSNaoya Horiguchi 263033f4751eSNaoya Horiguchi /* 2631af28a988SMatthew Wilcox (Oracle) * If caller asks to setup a migration entry, we need a folio to check 2632af28a988SMatthew Wilcox (Oracle) * pmd against. Otherwise we can end up replacing wrong folio. 263333f4751eSNaoya Horiguchi */ 2634af28a988SMatthew Wilcox (Oracle) VM_BUG_ON(freeze && !folio); 263583a8441fSMatthew Wilcox (Oracle) VM_WARN_ON_ONCE(folio && !folio_test_locked(folio)); 263633f4751eSNaoya Horiguchi 26377f760917SDavid Hildenbrand if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) || 263883a8441fSMatthew Wilcox (Oracle) is_pmd_migration_entry(*pmd)) { 2639cea33328SMiaohe Lin /* 2640cea33328SMiaohe Lin * It's safe to call pmd_page when folio is set because it's 2641cea33328SMiaohe Lin * guaranteed that pmd is present. 2642cea33328SMiaohe Lin */ 264383a8441fSMatthew Wilcox (Oracle) if (folio && folio != page_folio(pmd_page(*pmd))) 264483a8441fSMatthew Wilcox (Oracle) goto out; 2645ac46d4f3SJérôme Glisse __split_huge_pmd_locked(vma, pmd, range.start, freeze); 264683a8441fSMatthew Wilcox (Oracle) } 26477f760917SDavid Hildenbrand 2648e90309c9SKirill A. Shutemov out: 2649eef1b3baSKirill A. Shutemov spin_unlock(ptl); 2650ec8832d0SAlistair Popple mmu_notifier_invalidate_range_end(&range); 2651eef1b3baSKirill A. Shutemov } 2652eef1b3baSKirill A. Shutemov 2653fec89c10SKirill A. Shutemov void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, 2654af28a988SMatthew Wilcox (Oracle) bool freeze, struct folio *folio) 265594fcc585SAndrea Arcangeli { 265650722804SZach O'Keefe pmd_t *pmd = mm_find_pmd(vma->vm_mm, address); 265794fcc585SAndrea Arcangeli 265850722804SZach O'Keefe if (!pmd) 2659f72e7dcdSHugh Dickins return; 2660f72e7dcdSHugh Dickins 2661af28a988SMatthew Wilcox (Oracle) __split_huge_pmd(vma, pmd, address, freeze, folio); 266294fcc585SAndrea Arcangeli } 266394fcc585SAndrea Arcangeli 266471f9e58eSMiaohe Lin static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address) 266571f9e58eSMiaohe Lin { 266671f9e58eSMiaohe Lin /* 266771f9e58eSMiaohe Lin * If the new address isn't hpage aligned and it could previously 266871f9e58eSMiaohe Lin * contain an hugepage: check if we need to split an huge pmd. 266971f9e58eSMiaohe Lin */ 267071f9e58eSMiaohe Lin if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) && 267171f9e58eSMiaohe Lin range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE), 267271f9e58eSMiaohe Lin ALIGN(address, HPAGE_PMD_SIZE))) 267371f9e58eSMiaohe Lin split_huge_pmd_address(vma, address, false, NULL); 267471f9e58eSMiaohe Lin } 267571f9e58eSMiaohe Lin 2676e1b9996bSKirill A. Shutemov void vma_adjust_trans_huge(struct vm_area_struct *vma, 267794fcc585SAndrea Arcangeli unsigned long start, 267894fcc585SAndrea Arcangeli unsigned long end, 267994fcc585SAndrea Arcangeli long adjust_next) 268094fcc585SAndrea Arcangeli { 268171f9e58eSMiaohe Lin /* Check if we need to split start first. */ 268271f9e58eSMiaohe Lin split_huge_pmd_if_needed(vma, start); 268371f9e58eSMiaohe Lin 268471f9e58eSMiaohe Lin /* Check if we need to split end next. */ 268571f9e58eSMiaohe Lin split_huge_pmd_if_needed(vma, end); 268694fcc585SAndrea Arcangeli 268794fcc585SAndrea Arcangeli /* 268868540502SMatthew Wilcox (Oracle) * If we're also updating the next vma vm_start, 268971f9e58eSMiaohe Lin * check if we need to split it. 269094fcc585SAndrea Arcangeli */ 269194fcc585SAndrea Arcangeli if (adjust_next > 0) { 269268540502SMatthew Wilcox (Oracle) struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end); 269394fcc585SAndrea Arcangeli unsigned long nstart = next->vm_start; 2694f9d86a60SWei Yang nstart += adjust_next; 269571f9e58eSMiaohe Lin split_huge_pmd_if_needed(next, nstart); 269694fcc585SAndrea Arcangeli } 269794fcc585SAndrea Arcangeli } 2698e9b61f19SKirill A. Shutemov 2699684555aaSMatthew Wilcox (Oracle) static void unmap_folio(struct folio *folio) 2700e9b61f19SKirill A. Shutemov { 2701a98a2f0cSAlistair Popple enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD | 27023027c6f8SBaolin Wang TTU_SYNC | TTU_BATCH_FLUSH; 2703e9b61f19SKirill A. Shutemov 2704684555aaSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_large(folio), folio); 2705e9b61f19SKirill A. Shutemov 2706a98a2f0cSAlistair Popple /* 2707a98a2f0cSAlistair Popple * Anon pages need migration entries to preserve them, but file 2708a98a2f0cSAlistair Popple * pages can simply be left unmapped, then faulted back on demand. 2709a98a2f0cSAlistair Popple * If that is ever changed (perhaps for mlock), update remap_page(). 2710a98a2f0cSAlistair Popple */ 27114b8554c5SMatthew Wilcox (Oracle) if (folio_test_anon(folio)) 27124b8554c5SMatthew Wilcox (Oracle) try_to_migrate(folio, ttu_flags); 2713a98a2f0cSAlistair Popple else 2714869f7ee6SMatthew Wilcox (Oracle) try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK); 27153027c6f8SBaolin Wang 27163027c6f8SBaolin Wang try_to_unmap_flush(); 2717bd56086fSKirill A. Shutemov } 2718bd56086fSKirill A. Shutemov 27194eecb8b9SMatthew Wilcox (Oracle) static void remap_page(struct folio *folio, unsigned long nr) 2720e9b61f19SKirill A. Shutemov { 27214eecb8b9SMatthew Wilcox (Oracle) int i = 0; 2722ab02c252SHugh Dickins 2723684555aaSMatthew Wilcox (Oracle) /* If unmap_folio() uses try_to_migrate() on file, remove this check */ 27244eecb8b9SMatthew Wilcox (Oracle) if (!folio_test_anon(folio)) 2725ab02c252SHugh Dickins return; 27264eecb8b9SMatthew Wilcox (Oracle) for (;;) { 27274eecb8b9SMatthew Wilcox (Oracle) remove_migration_ptes(folio, folio, true); 27284eecb8b9SMatthew Wilcox (Oracle) i += folio_nr_pages(folio); 27294eecb8b9SMatthew Wilcox (Oracle) if (i >= nr) 27304eecb8b9SMatthew Wilcox (Oracle) break; 27314eecb8b9SMatthew Wilcox (Oracle) folio = folio_next(folio); 2732e9b61f19SKirill A. Shutemov } 2733ace71a19SKirill A. Shutemov } 2734e9b61f19SKirill A. Shutemov 273594866635SAlex Shi static void lru_add_page_tail(struct page *head, struct page *tail, 273688dcb9a3SAlex Shi struct lruvec *lruvec, struct list_head *list) 273788dcb9a3SAlex Shi { 273894866635SAlex Shi VM_BUG_ON_PAGE(!PageHead(head), head); 273994866635SAlex Shi VM_BUG_ON_PAGE(PageCompound(tail), head); 274094866635SAlex Shi VM_BUG_ON_PAGE(PageLRU(tail), head); 27416168d0daSAlex Shi lockdep_assert_held(&lruvec->lru_lock); 274288dcb9a3SAlex Shi 27436dbb5741SAlex Shi if (list) { 274488dcb9a3SAlex Shi /* page reclaim is reclaiming a huge page */ 27456dbb5741SAlex Shi VM_WARN_ON(PageLRU(head)); 274694866635SAlex Shi get_page(tail); 274794866635SAlex Shi list_add_tail(&tail->lru, list); 274888dcb9a3SAlex Shi } else { 27496dbb5741SAlex Shi /* head is still on lru (and we have it frozen) */ 27506dbb5741SAlex Shi VM_WARN_ON(!PageLRU(head)); 275107ca7606SHugh Dickins if (PageUnevictable(tail)) 275207ca7606SHugh Dickins tail->mlock_count = 0; 275307ca7606SHugh Dickins else 27546dbb5741SAlex Shi list_add_tail(&tail->lru, &head->lru); 275507ca7606SHugh Dickins SetPageLRU(tail); 275688dcb9a3SAlex Shi } 275788dcb9a3SAlex Shi } 275888dcb9a3SAlex Shi 275907e09c48SDavid Hildenbrand static void __split_huge_page_tail(struct folio *folio, int tail, 2760e9b61f19SKirill A. Shutemov struct lruvec *lruvec, struct list_head *list) 2761e9b61f19SKirill A. Shutemov { 276207e09c48SDavid Hildenbrand struct page *head = &folio->page; 2763e9b61f19SKirill A. Shutemov struct page *page_tail = head + tail; 276407e09c48SDavid Hildenbrand /* 276507e09c48SDavid Hildenbrand * Careful: new_folio is not a "real" folio before we cleared PageTail. 276607e09c48SDavid Hildenbrand * Don't pass it around before clear_compound_head(). 276707e09c48SDavid Hildenbrand */ 276807e09c48SDavid Hildenbrand struct folio *new_folio = (struct folio *)page_tail; 2769e9b61f19SKirill A. Shutemov 27708df651c7SKirill A. Shutemov VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); 2771e9b61f19SKirill A. Shutemov 2772e9b61f19SKirill A. Shutemov /* 2773605ca5edSKonstantin Khlebnikov * Clone page flags before unfreezing refcount. 2774605ca5edSKonstantin Khlebnikov * 2775605ca5edSKonstantin Khlebnikov * After successful get_page_unless_zero() might follow flags change, 27768958b249SHaitao Shi * for example lock_page() which set PG_waiters. 27776c287605SDavid Hildenbrand * 27786c287605SDavid Hildenbrand * Note that for mapped sub-pages of an anonymous THP, 2779684555aaSMatthew Wilcox (Oracle) * PG_anon_exclusive has been cleared in unmap_folio() and is stored in 27806c287605SDavid Hildenbrand * the migration entry instead from where remap_page() will restore it. 27816c287605SDavid Hildenbrand * We can still have PG_anon_exclusive set on effectively unmapped and 27826c287605SDavid Hildenbrand * unreferenced sub-pages of an anonymous THP: we can simply drop 27836c287605SDavid Hildenbrand * PG_anon_exclusive (-> PG_mappedtodisk) for these here. 2784e9b61f19SKirill A. Shutemov */ 2785e9b61f19SKirill A. Shutemov page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 2786e9b61f19SKirill A. Shutemov page_tail->flags |= (head->flags & 2787e9b61f19SKirill A. Shutemov ((1L << PG_referenced) | 2788e9b61f19SKirill A. Shutemov (1L << PG_swapbacked) | 278938d8b4e6SHuang Ying (1L << PG_swapcache) | 2790e9b61f19SKirill A. Shutemov (1L << PG_mlocked) | 2791e9b61f19SKirill A. Shutemov (1L << PG_uptodate) | 2792e9b61f19SKirill A. Shutemov (1L << PG_active) | 27931899ad18SJohannes Weiner (1L << PG_workingset) | 2794e9b61f19SKirill A. Shutemov (1L << PG_locked) | 2795b8d3c4c3SMinchan Kim (1L << PG_unevictable) | 2796b0284cd2SCatalin Marinas #ifdef CONFIG_ARCH_USES_PG_ARCH_X 279772e6afa0SCatalin Marinas (1L << PG_arch_2) | 2798ef6458b1SPeter Collingbourne (1L << PG_arch_3) | 279972e6afa0SCatalin Marinas #endif 2800ec1c86b2SYu Zhao (1L << PG_dirty) | 2801ec1c86b2SYu Zhao LRU_GEN_MASK | LRU_REFS_MASK)); 2802e9b61f19SKirill A. Shutemov 2803cb67f428SHugh Dickins /* ->mapping in first and second tail page is replaced by other uses */ 2804173d9d9fSHugh Dickins VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, 2805173d9d9fSHugh Dickins page_tail); 2806173d9d9fSHugh Dickins page_tail->mapping = head->mapping; 2807173d9d9fSHugh Dickins page_tail->index = head->index + tail; 280871e2d666SMel Gorman 280971e2d666SMel Gorman /* 2810cfeed8ffSDavid Hildenbrand * page->private should not be set in tail pages. Fix up and warn once 2811cfeed8ffSDavid Hildenbrand * if private is unexpectedly set. 281271e2d666SMel Gorman */ 2813cfeed8ffSDavid Hildenbrand if (unlikely(page_tail->private)) { 2814cfeed8ffSDavid Hildenbrand VM_WARN_ON_ONCE_PAGE(true, page_tail); 2815b653db77SMatthew Wilcox (Oracle) page_tail->private = 0; 281671e2d666SMel Gorman } 281707e09c48SDavid Hildenbrand if (folio_test_swapcache(folio)) 281807e09c48SDavid Hildenbrand new_folio->swap.val = folio->swap.val + tail; 2819173d9d9fSHugh Dickins 2820605ca5edSKonstantin Khlebnikov /* Page flags must be visible before we make the page non-compound. */ 2821e9b61f19SKirill A. Shutemov smp_wmb(); 2822e9b61f19SKirill A. Shutemov 2823605ca5edSKonstantin Khlebnikov /* 2824605ca5edSKonstantin Khlebnikov * Clear PageTail before unfreezing page refcount. 2825605ca5edSKonstantin Khlebnikov * 2826605ca5edSKonstantin Khlebnikov * After successful get_page_unless_zero() might follow put_page() 2827605ca5edSKonstantin Khlebnikov * which needs correct compound_head(). 2828605ca5edSKonstantin Khlebnikov */ 2829e9b61f19SKirill A. Shutemov clear_compound_head(page_tail); 2830e9b61f19SKirill A. Shutemov 2831605ca5edSKonstantin Khlebnikov /* Finally unfreeze refcount. Additional reference from page cache. */ 2832b7542769SKefeng Wang page_ref_unfreeze(page_tail, 1 + (!folio_test_anon(folio) || 2833b7542769SKefeng Wang folio_test_swapcache(folio))); 2834605ca5edSKonstantin Khlebnikov 2835b7542769SKefeng Wang if (folio_test_young(folio)) 2836b7542769SKefeng Wang folio_set_young(new_folio); 2837b7542769SKefeng Wang if (folio_test_idle(folio)) 2838b7542769SKefeng Wang folio_set_idle(new_folio); 2839e9b61f19SKirill A. Shutemov 2840c8253011SKefeng Wang folio_xchg_last_cpupid(new_folio, folio_last_cpupid(folio)); 284194723aafSMichal Hocko 284294723aafSMichal Hocko /* 284394723aafSMichal Hocko * always add to the tail because some iterators expect new 284494723aafSMichal Hocko * pages to show after the currently processed elements - e.g. 284594723aafSMichal Hocko * migrate_pages 284694723aafSMichal Hocko */ 2847e9b61f19SKirill A. Shutemov lru_add_page_tail(head, page_tail, lruvec, list); 2848e9b61f19SKirill A. Shutemov } 2849e9b61f19SKirill A. Shutemov 2850baa355fdSKirill A. Shutemov static void __split_huge_page(struct page *page, struct list_head *list, 2851b6769834SAlex Shi pgoff_t end) 2852e9b61f19SKirill A. Shutemov { 2853e809c3feSMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 2854e809c3feSMatthew Wilcox (Oracle) struct page *head = &folio->page; 2855e9b61f19SKirill A. Shutemov struct lruvec *lruvec; 28564101196bSMatthew Wilcox (Oracle) struct address_space *swap_cache = NULL; 28574101196bSMatthew Wilcox (Oracle) unsigned long offset = 0; 28588cce5475SKirill A. Shutemov unsigned int nr = thp_nr_pages(head); 2859509f0069SHugh Dickins int i, nr_dropped = 0; 2860e9b61f19SKirill A. Shutemov 2861e9b61f19SKirill A. Shutemov /* complete memcg works before add pages to LRU */ 2862be6c8982SZhou Guanghui split_page_memcg(head, nr); 2863e9b61f19SKirill A. Shutemov 286407e09c48SDavid Hildenbrand if (folio_test_anon(folio) && folio_test_swapcache(folio)) { 286507e09c48SDavid Hildenbrand offset = swp_offset(folio->swap); 286607e09c48SDavid Hildenbrand swap_cache = swap_address_space(folio->swap); 28674101196bSMatthew Wilcox (Oracle) xa_lock(&swap_cache->i_pages); 28684101196bSMatthew Wilcox (Oracle) } 28694101196bSMatthew Wilcox (Oracle) 2870f0953a1bSIngo Molnar /* lock lru list/PageCompound, ref frozen by page_ref_freeze */ 2871e809c3feSMatthew Wilcox (Oracle) lruvec = folio_lruvec_lock(folio); 2872b6769834SAlex Shi 2873eac96c3eSYang Shi ClearPageHasHWPoisoned(head); 2874eac96c3eSYang Shi 28758cce5475SKirill A. Shutemov for (i = nr - 1; i >= 1; i--) { 287607e09c48SDavid Hildenbrand __split_huge_page_tail(folio, i, lruvec, list); 2877d144bf62SHugh Dickins /* Some pages can be beyond EOF: drop them from page cache */ 2878baa355fdSKirill A. Shutemov if (head[i].index >= end) { 2879fb5c2029SMatthew Wilcox (Oracle) struct folio *tail = page_folio(head + i); 2880fb5c2029SMatthew Wilcox (Oracle) 2881d144bf62SHugh Dickins if (shmem_mapping(head->mapping)) 2882509f0069SHugh Dickins nr_dropped++; 2883fb5c2029SMatthew Wilcox (Oracle) else if (folio_test_clear_dirty(tail)) 2884fb5c2029SMatthew Wilcox (Oracle) folio_account_cleaned(tail, 2885fb5c2029SMatthew Wilcox (Oracle) inode_to_wb(folio->mapping->host)); 2886fb5c2029SMatthew Wilcox (Oracle) __filemap_remove_folio(tail, NULL); 2887fb5c2029SMatthew Wilcox (Oracle) folio_put(tail); 28884101196bSMatthew Wilcox (Oracle) } else if (!PageAnon(page)) { 28894101196bSMatthew Wilcox (Oracle) __xa_store(&head->mapping->i_pages, head[i].index, 28904101196bSMatthew Wilcox (Oracle) head + i, 0); 28914101196bSMatthew Wilcox (Oracle) } else if (swap_cache) { 28924101196bSMatthew Wilcox (Oracle) __xa_store(&swap_cache->i_pages, offset + i, 28934101196bSMatthew Wilcox (Oracle) head + i, 0); 2894baa355fdSKirill A. Shutemov } 2895baa355fdSKirill A. Shutemov } 2896e9b61f19SKirill A. Shutemov 2897e9b61f19SKirill A. Shutemov ClearPageCompound(head); 28986168d0daSAlex Shi unlock_page_lruvec(lruvec); 2899b6769834SAlex Shi /* Caller disabled irqs, so they are still disabled here */ 2900f7da677bSVlastimil Babka 29018cce5475SKirill A. Shutemov split_page_owner(head, nr); 2902f7da677bSVlastimil Babka 2903baa355fdSKirill A. Shutemov /* See comment in __split_huge_page_tail() */ 2904baa355fdSKirill A. Shutemov if (PageAnon(head)) { 2905aa5dc07fSMatthew Wilcox /* Additional pin to swap cache */ 29064101196bSMatthew Wilcox (Oracle) if (PageSwapCache(head)) { 290738d8b4e6SHuang Ying page_ref_add(head, 2); 29084101196bSMatthew Wilcox (Oracle) xa_unlock(&swap_cache->i_pages); 29094101196bSMatthew Wilcox (Oracle) } else { 2910baa355fdSKirill A. Shutemov page_ref_inc(head); 29114101196bSMatthew Wilcox (Oracle) } 2912baa355fdSKirill A. Shutemov } else { 2913aa5dc07fSMatthew Wilcox /* Additional pin to page cache */ 2914baa355fdSKirill A. Shutemov page_ref_add(head, 2); 2915b93b0163SMatthew Wilcox xa_unlock(&head->mapping->i_pages); 2916baa355fdSKirill A. Shutemov } 2917b6769834SAlex Shi local_irq_enable(); 2918e9b61f19SKirill A. Shutemov 2919509f0069SHugh Dickins if (nr_dropped) 2920509f0069SHugh Dickins shmem_uncharge(head->mapping->host, nr_dropped); 29214eecb8b9SMatthew Wilcox (Oracle) remap_page(folio, nr); 2922e9b61f19SKirill A. Shutemov 292307e09c48SDavid Hildenbrand if (folio_test_swapcache(folio)) 292407e09c48SDavid Hildenbrand split_swap_cluster(folio->swap); 2925c4f9c701SHuang Ying 29268cce5475SKirill A. Shutemov for (i = 0; i < nr; i++) { 2927e9b61f19SKirill A. Shutemov struct page *subpage = head + i; 2928e9b61f19SKirill A. Shutemov if (subpage == page) 2929e9b61f19SKirill A. Shutemov continue; 2930e9b61f19SKirill A. Shutemov unlock_page(subpage); 2931e9b61f19SKirill A. Shutemov 2932e9b61f19SKirill A. Shutemov /* 2933e9b61f19SKirill A. Shutemov * Subpages may be freed if there wasn't any mapping 2934e9b61f19SKirill A. Shutemov * like if add_to_swap() is running on a lru page that 2935e9b61f19SKirill A. Shutemov * had its mapping zapped. And freeing these pages 2936e9b61f19SKirill A. Shutemov * requires taking the lru_lock so we do the put_page 2937e9b61f19SKirill A. Shutemov * of the tail pages after the split is complete. 2938e9b61f19SKirill A. Shutemov */ 29390b175468SMiaohe Lin free_page_and_swap_cache(subpage); 2940e9b61f19SKirill A. Shutemov } 2941e9b61f19SKirill A. Shutemov } 2942e9b61f19SKirill A. Shutemov 2943b8f593cdSHuang Ying /* Racy check whether the huge page can be split */ 2944d4b4084aSMatthew Wilcox (Oracle) bool can_split_folio(struct folio *folio, int *pextra_pins) 2945b8f593cdSHuang Ying { 2946b8f593cdSHuang Ying int extra_pins; 2947b8f593cdSHuang Ying 2948aa5dc07fSMatthew Wilcox /* Additional pins from page cache */ 2949d4b4084aSMatthew Wilcox (Oracle) if (folio_test_anon(folio)) 2950d4b4084aSMatthew Wilcox (Oracle) extra_pins = folio_test_swapcache(folio) ? 2951d4b4084aSMatthew Wilcox (Oracle) folio_nr_pages(folio) : 0; 2952b8f593cdSHuang Ying else 2953d4b4084aSMatthew Wilcox (Oracle) extra_pins = folio_nr_pages(folio); 2954b8f593cdSHuang Ying if (pextra_pins) 2955b8f593cdSHuang Ying *pextra_pins = extra_pins; 2956d4b4084aSMatthew Wilcox (Oracle) return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1; 2957b8f593cdSHuang Ying } 2958b8f593cdSHuang Ying 29596d0a07edSAndrea Arcangeli /* 2960e9b61f19SKirill A. Shutemov * This function splits huge page into normal pages. @page can point to any 2961e9b61f19SKirill A. Shutemov * subpage of huge page to split. Split doesn't change the position of @page. 2962e9b61f19SKirill A. Shutemov * 2963e9b61f19SKirill A. Shutemov * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. 2964e9b61f19SKirill A. Shutemov * The huge page must be locked. 2965e9b61f19SKirill A. Shutemov * 2966e9b61f19SKirill A. Shutemov * If @list is null, tail pages will be added to LRU list, otherwise, to @list. 2967e9b61f19SKirill A. Shutemov * 2968e9b61f19SKirill A. Shutemov * Both head page and tail pages will inherit mapping, flags, and so on from 2969e9b61f19SKirill A. Shutemov * the hugepage. 2970e9b61f19SKirill A. Shutemov * 2971e9b61f19SKirill A. Shutemov * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if 2972e9b61f19SKirill A. Shutemov * they are not mapped. 2973e9b61f19SKirill A. Shutemov * 2974e9b61f19SKirill A. Shutemov * Returns 0 if the hugepage is split successfully. 2975e9b61f19SKirill A. Shutemov * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under 2976e9b61f19SKirill A. Shutemov * us. 2977e9b61f19SKirill A. Shutemov */ 2978e9b61f19SKirill A. Shutemov int split_huge_page_to_list(struct page *page, struct list_head *list) 2979e9b61f19SKirill A. Shutemov { 29804eecb8b9SMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 2981f8baa6beSMatthew Wilcox (Oracle) struct deferred_split *ds_queue = get_deferred_split_queue(folio); 29823e9a13daSMatthew Wilcox (Oracle) XA_STATE(xas, &folio->mapping->i_pages, folio->index); 2983baa355fdSKirill A. Shutemov struct anon_vma *anon_vma = NULL; 2984baa355fdSKirill A. Shutemov struct address_space *mapping = NULL; 2985504e070dSYang Shi int extra_pins, ret; 2986006d3ff2SHugh Dickins pgoff_t end; 2987478d134eSXu Yu bool is_hzp; 2988e9b61f19SKirill A. Shutemov 29893e9a13daSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 29903e9a13daSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_large(folio), folio); 2991e9b61f19SKirill A. Shutemov 29923e9a13daSMatthew Wilcox (Oracle) is_hzp = is_huge_zero_page(&folio->page); 29934737edbbSNaoya Horiguchi if (is_hzp) { 29944737edbbSNaoya Horiguchi pr_warn_ratelimited("Called split_huge_page for huge zero page\n"); 2995478d134eSXu Yu return -EBUSY; 29964737edbbSNaoya Horiguchi } 2997478d134eSXu Yu 29983e9a13daSMatthew Wilcox (Oracle) if (folio_test_writeback(folio)) 299959807685SHuang Ying return -EBUSY; 300059807685SHuang Ying 30013e9a13daSMatthew Wilcox (Oracle) if (folio_test_anon(folio)) { 3002e9b61f19SKirill A. Shutemov /* 3003c1e8d7c6SMichel Lespinasse * The caller does not necessarily hold an mmap_lock that would 3004baa355fdSKirill A. Shutemov * prevent the anon_vma disappearing so we first we take a 3005baa355fdSKirill A. Shutemov * reference to it and then lock the anon_vma for write. This 30062f031c6fSMatthew Wilcox (Oracle) * is similar to folio_lock_anon_vma_read except the write lock 3007baa355fdSKirill A. Shutemov * is taken to serialise against parallel split or collapse 3008baa355fdSKirill A. Shutemov * operations. 3009e9b61f19SKirill A. Shutemov */ 301029eea9b5SMatthew Wilcox (Oracle) anon_vma = folio_get_anon_vma(folio); 3011e9b61f19SKirill A. Shutemov if (!anon_vma) { 3012e9b61f19SKirill A. Shutemov ret = -EBUSY; 3013e9b61f19SKirill A. Shutemov goto out; 3014e9b61f19SKirill A. Shutemov } 3015006d3ff2SHugh Dickins end = -1; 3016baa355fdSKirill A. Shutemov mapping = NULL; 3017e9b61f19SKirill A. Shutemov anon_vma_lock_write(anon_vma); 3018baa355fdSKirill A. Shutemov } else { 30196a3edd29SYin Fengwei gfp_t gfp; 30206a3edd29SYin Fengwei 30213e9a13daSMatthew Wilcox (Oracle) mapping = folio->mapping; 3022baa355fdSKirill A. Shutemov 3023baa355fdSKirill A. Shutemov /* Truncated ? */ 3024baa355fdSKirill A. Shutemov if (!mapping) { 3025baa355fdSKirill A. Shutemov ret = -EBUSY; 3026baa355fdSKirill A. Shutemov goto out; 3027baa355fdSKirill A. Shutemov } 3028baa355fdSKirill A. Shutemov 30296a3edd29SYin Fengwei gfp = current_gfp_context(mapping_gfp_mask(mapping) & 30306a3edd29SYin Fengwei GFP_RECLAIM_MASK); 30316a3edd29SYin Fengwei 30320201ebf2SDavid Howells if (!filemap_release_folio(folio, gfp)) { 30336a3edd29SYin Fengwei ret = -EBUSY; 30346a3edd29SYin Fengwei goto out; 30356a3edd29SYin Fengwei } 30366a3edd29SYin Fengwei 30373e9a13daSMatthew Wilcox (Oracle) xas_split_alloc(&xas, folio, folio_order(folio), gfp); 30386b24ca4aSMatthew Wilcox (Oracle) if (xas_error(&xas)) { 30396b24ca4aSMatthew Wilcox (Oracle) ret = xas_error(&xas); 30406b24ca4aSMatthew Wilcox (Oracle) goto out; 30416b24ca4aSMatthew Wilcox (Oracle) } 30426b24ca4aSMatthew Wilcox (Oracle) 3043baa355fdSKirill A. Shutemov anon_vma = NULL; 3044baa355fdSKirill A. Shutemov i_mmap_lock_read(mapping); 3045006d3ff2SHugh Dickins 3046006d3ff2SHugh Dickins /* 3047006d3ff2SHugh Dickins *__split_huge_page() may need to trim off pages beyond EOF: 3048006d3ff2SHugh Dickins * but on 32-bit, i_size_read() takes an irq-unsafe seqlock, 3049006d3ff2SHugh Dickins * which cannot be nested inside the page tree lock. So note 3050006d3ff2SHugh Dickins * end now: i_size itself may be changed at any moment, but 30513e9a13daSMatthew Wilcox (Oracle) * folio lock is good enough to serialize the trimming. 3052006d3ff2SHugh Dickins */ 3053006d3ff2SHugh Dickins end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); 3054d144bf62SHugh Dickins if (shmem_mapping(mapping)) 3055d144bf62SHugh Dickins end = shmem_fallocend(mapping->host, end); 3056baa355fdSKirill A. Shutemov } 3057e9b61f19SKirill A. Shutemov 3058e9b61f19SKirill A. Shutemov /* 3059684555aaSMatthew Wilcox (Oracle) * Racy check if we can split the page, before unmap_folio() will 3060e9b61f19SKirill A. Shutemov * split PMDs 3061e9b61f19SKirill A. Shutemov */ 3062d4b4084aSMatthew Wilcox (Oracle) if (!can_split_folio(folio, &extra_pins)) { 3063fd4a7ac3SBaolin Wang ret = -EAGAIN; 3064e9b61f19SKirill A. Shutemov goto out_unlock; 3065e9b61f19SKirill A. Shutemov } 3066e9b61f19SKirill A. Shutemov 3067684555aaSMatthew Wilcox (Oracle) unmap_folio(folio); 3068e9b61f19SKirill A. Shutemov 3069b6769834SAlex Shi /* block interrupt reentry in xa_lock and spinlock */ 3070b6769834SAlex Shi local_irq_disable(); 3071baa355fdSKirill A. Shutemov if (mapping) { 3072baa355fdSKirill A. Shutemov /* 30733e9a13daSMatthew Wilcox (Oracle) * Check if the folio is present in page cache. 30743e9a13daSMatthew Wilcox (Oracle) * We assume all tail are present too, if folio is there. 3075baa355fdSKirill A. Shutemov */ 30766b24ca4aSMatthew Wilcox (Oracle) xas_lock(&xas); 30776b24ca4aSMatthew Wilcox (Oracle) xas_reset(&xas); 30783e9a13daSMatthew Wilcox (Oracle) if (xas_load(&xas) != folio) 3079baa355fdSKirill A. Shutemov goto fail; 3080baa355fdSKirill A. Shutemov } 3081baa355fdSKirill A. Shutemov 30820139aa7bSJoonsoo Kim /* Prevent deferred_split_scan() touching ->_refcount */ 3083364c1eebSYang Shi spin_lock(&ds_queue->split_queue_lock); 30843e9a13daSMatthew Wilcox (Oracle) if (folio_ref_freeze(folio, 1 + extra_pins)) { 30854375a553SMatthew Wilcox (Oracle) if (!list_empty(&folio->_deferred_list)) { 3086364c1eebSYang Shi ds_queue->split_queue_len--; 30874375a553SMatthew Wilcox (Oracle) list_del(&folio->_deferred_list); 30889a982250SKirill A. Shutemov } 3089afb97172SWei Yang spin_unlock(&ds_queue->split_queue_lock); 309006d3eff6SKirill A. Shutemov if (mapping) { 30913e9a13daSMatthew Wilcox (Oracle) int nr = folio_nr_pages(folio); 3092bf9eceadSMuchun Song 30933e9a13daSMatthew Wilcox (Oracle) xas_split(&xas, folio, folio_order(folio)); 3094a48d5bdcSStefan Roesch if (folio_test_pmd_mappable(folio)) { 30953e9a13daSMatthew Wilcox (Oracle) if (folio_test_swapbacked(folio)) { 3096a48d5bdcSStefan Roesch __lruvec_stat_mod_folio(folio, 3097a48d5bdcSStefan Roesch NR_SHMEM_THPS, -nr); 30981ca7554dSMarek Szyprowski } else { 3099a48d5bdcSStefan Roesch __lruvec_stat_mod_folio(folio, 3100a48d5bdcSStefan Roesch NR_FILE_THPS, -nr); 31011ca7554dSMarek Szyprowski filemap_nr_thps_dec(mapping); 31021ca7554dSMarek Szyprowski } 310306d3eff6SKirill A. Shutemov } 3104a48d5bdcSStefan Roesch } 310506d3eff6SKirill A. Shutemov 3106b6769834SAlex Shi __split_huge_page(page, list, end); 3107e9b61f19SKirill A. Shutemov ret = 0; 3108baa355fdSKirill A. Shutemov } else { 3109364c1eebSYang Shi spin_unlock(&ds_queue->split_queue_lock); 3110504e070dSYang Shi fail: 3111504e070dSYang Shi if (mapping) 31126b24ca4aSMatthew Wilcox (Oracle) xas_unlock(&xas); 3113b6769834SAlex Shi local_irq_enable(); 31144eecb8b9SMatthew Wilcox (Oracle) remap_page(folio, folio_nr_pages(folio)); 3115fd4a7ac3SBaolin Wang ret = -EAGAIN; 3116e9b61f19SKirill A. Shutemov } 3117e9b61f19SKirill A. Shutemov 3118e9b61f19SKirill A. Shutemov out_unlock: 3119baa355fdSKirill A. Shutemov if (anon_vma) { 3120e9b61f19SKirill A. Shutemov anon_vma_unlock_write(anon_vma); 3121e9b61f19SKirill A. Shutemov put_anon_vma(anon_vma); 3122baa355fdSKirill A. Shutemov } 3123baa355fdSKirill A. Shutemov if (mapping) 3124baa355fdSKirill A. Shutemov i_mmap_unlock_read(mapping); 3125e9b61f19SKirill A. Shutemov out: 312669a37a8bSMatthew Wilcox (Oracle) xas_destroy(&xas); 3127e9b61f19SKirill A. Shutemov count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); 3128e9b61f19SKirill A. Shutemov return ret; 3129e9b61f19SKirill A. Shutemov } 31309a982250SKirill A. Shutemov 31318dc4a8f1SMatthew Wilcox (Oracle) void folio_undo_large_rmappable(struct folio *folio) 31329a982250SKirill A. Shutemov { 31338dc4a8f1SMatthew Wilcox (Oracle) struct deferred_split *ds_queue; 31349a982250SKirill A. Shutemov unsigned long flags; 31359a982250SKirill A. Shutemov 3136deedad80SYin Fengwei /* 3137deedad80SYin Fengwei * At this point, there is no one trying to add the folio to 3138deedad80SYin Fengwei * deferred_list. If folio is not in deferred_list, it's safe 3139deedad80SYin Fengwei * to check without acquiring the split_queue_lock. 3140deedad80SYin Fengwei */ 31418dc4a8f1SMatthew Wilcox (Oracle) if (data_race(list_empty(&folio->_deferred_list))) 31428dc4a8f1SMatthew Wilcox (Oracle) return; 31438dc4a8f1SMatthew Wilcox (Oracle) 31448dc4a8f1SMatthew Wilcox (Oracle) ds_queue = get_deferred_split_queue(folio); 3145364c1eebSYang Shi spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 31468991de90SMatthew Wilcox (Oracle) if (!list_empty(&folio->_deferred_list)) { 3147364c1eebSYang Shi ds_queue->split_queue_len--; 31489bcef597SBaolin Wang list_del_init(&folio->_deferred_list); 31499a982250SKirill A. Shutemov } 3150364c1eebSYang Shi spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 3151deedad80SYin Fengwei } 31529a982250SKirill A. Shutemov 3153f158ed61SMatthew Wilcox (Oracle) void deferred_split_folio(struct folio *folio) 31549a982250SKirill A. Shutemov { 3155f8baa6beSMatthew Wilcox (Oracle) struct deferred_split *ds_queue = get_deferred_split_queue(folio); 315687eaceb3SYang Shi #ifdef CONFIG_MEMCG 31578991de90SMatthew Wilcox (Oracle) struct mem_cgroup *memcg = folio_memcg(folio); 315887eaceb3SYang Shi #endif 31599a982250SKirill A. Shutemov unsigned long flags; 31609a982250SKirill A. Shutemov 31618991de90SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio); 31629a982250SKirill A. Shutemov 316387eaceb3SYang Shi /* 316487eaceb3SYang Shi * The try_to_unmap() in page reclaim path might reach here too, 316587eaceb3SYang Shi * this may cause a race condition to corrupt deferred split queue. 31668991de90SMatthew Wilcox (Oracle) * And, if page reclaim is already handling the same folio, it is 316787eaceb3SYang Shi * unnecessary to handle it again in shrinker. 316887eaceb3SYang Shi * 31698991de90SMatthew Wilcox (Oracle) * Check the swapcache flag to determine if the folio is being 31708991de90SMatthew Wilcox (Oracle) * handled by page reclaim since THP swap would add the folio into 317187eaceb3SYang Shi * swap cache before calling try_to_unmap(). 317287eaceb3SYang Shi */ 31738991de90SMatthew Wilcox (Oracle) if (folio_test_swapcache(folio)) 317487eaceb3SYang Shi return; 317587eaceb3SYang Shi 31768991de90SMatthew Wilcox (Oracle) if (!list_empty(&folio->_deferred_list)) 31779a982250SKirill A. Shutemov return; 31789a982250SKirill A. Shutemov 3179364c1eebSYang Shi spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 31808991de90SMatthew Wilcox (Oracle) if (list_empty(&folio->_deferred_list)) { 3181f9719a03SKirill A. Shutemov count_vm_event(THP_DEFERRED_SPLIT_PAGE); 31828991de90SMatthew Wilcox (Oracle) list_add_tail(&folio->_deferred_list, &ds_queue->split_queue); 3183364c1eebSYang Shi ds_queue->split_queue_len++; 318487eaceb3SYang Shi #ifdef CONFIG_MEMCG 318587eaceb3SYang Shi if (memcg) 31868991de90SMatthew Wilcox (Oracle) set_shrinker_bit(memcg, folio_nid(folio), 318754d91729SQi Zheng deferred_split_shrinker->id); 318887eaceb3SYang Shi #endif 31899a982250SKirill A. Shutemov } 3190364c1eebSYang Shi spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 31919a982250SKirill A. Shutemov } 31929a982250SKirill A. Shutemov 31939a982250SKirill A. Shutemov static unsigned long deferred_split_count(struct shrinker *shrink, 31949a982250SKirill A. Shutemov struct shrink_control *sc) 31959a982250SKirill A. Shutemov { 3196a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(sc->nid); 3197364c1eebSYang Shi struct deferred_split *ds_queue = &pgdata->deferred_split_queue; 319887eaceb3SYang Shi 319987eaceb3SYang Shi #ifdef CONFIG_MEMCG 320087eaceb3SYang Shi if (sc->memcg) 320187eaceb3SYang Shi ds_queue = &sc->memcg->deferred_split_queue; 320287eaceb3SYang Shi #endif 3203364c1eebSYang Shi return READ_ONCE(ds_queue->split_queue_len); 32049a982250SKirill A. Shutemov } 32059a982250SKirill A. Shutemov 32069a982250SKirill A. Shutemov static unsigned long deferred_split_scan(struct shrinker *shrink, 32079a982250SKirill A. Shutemov struct shrink_control *sc) 32089a982250SKirill A. Shutemov { 3209a3d0a918SKirill A. Shutemov struct pglist_data *pgdata = NODE_DATA(sc->nid); 3210364c1eebSYang Shi struct deferred_split *ds_queue = &pgdata->deferred_split_queue; 32119a982250SKirill A. Shutemov unsigned long flags; 32124375a553SMatthew Wilcox (Oracle) LIST_HEAD(list); 32134375a553SMatthew Wilcox (Oracle) struct folio *folio, *next; 32149a982250SKirill A. Shutemov int split = 0; 32159a982250SKirill A. Shutemov 321687eaceb3SYang Shi #ifdef CONFIG_MEMCG 321787eaceb3SYang Shi if (sc->memcg) 321887eaceb3SYang Shi ds_queue = &sc->memcg->deferred_split_queue; 321987eaceb3SYang Shi #endif 322087eaceb3SYang Shi 3221364c1eebSYang Shi spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 32229a982250SKirill A. Shutemov /* Take pin on all head pages to avoid freeing them under us */ 32234375a553SMatthew Wilcox (Oracle) list_for_each_entry_safe(folio, next, &ds_queue->split_queue, 32244375a553SMatthew Wilcox (Oracle) _deferred_list) { 32254375a553SMatthew Wilcox (Oracle) if (folio_try_get(folio)) { 32264375a553SMatthew Wilcox (Oracle) list_move(&folio->_deferred_list, &list); 3227e3ae1953SKirill A. Shutemov } else { 32284375a553SMatthew Wilcox (Oracle) /* We lost race with folio_put() */ 32294375a553SMatthew Wilcox (Oracle) list_del_init(&folio->_deferred_list); 3230364c1eebSYang Shi ds_queue->split_queue_len--; 32319a982250SKirill A. Shutemov } 3232e3ae1953SKirill A. Shutemov if (!--sc->nr_to_scan) 3233e3ae1953SKirill A. Shutemov break; 32349a982250SKirill A. Shutemov } 3235364c1eebSYang Shi spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 32369a982250SKirill A. Shutemov 32374375a553SMatthew Wilcox (Oracle) list_for_each_entry_safe(folio, next, &list, _deferred_list) { 32384375a553SMatthew Wilcox (Oracle) if (!folio_trylock(folio)) 3239fa41b900SKirill A. Shutemov goto next; 32409a982250SKirill A. Shutemov /* split_huge_page() removes page from list on success */ 32414375a553SMatthew Wilcox (Oracle) if (!split_folio(folio)) 32429a982250SKirill A. Shutemov split++; 32434375a553SMatthew Wilcox (Oracle) folio_unlock(folio); 3244fa41b900SKirill A. Shutemov next: 32454375a553SMatthew Wilcox (Oracle) folio_put(folio); 32469a982250SKirill A. Shutemov } 32479a982250SKirill A. Shutemov 3248364c1eebSYang Shi spin_lock_irqsave(&ds_queue->split_queue_lock, flags); 3249364c1eebSYang Shi list_splice_tail(&list, &ds_queue->split_queue); 3250364c1eebSYang Shi spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); 32519a982250SKirill A. Shutemov 3252cb8d68ecSKirill A. Shutemov /* 3253cb8d68ecSKirill A. Shutemov * Stop shrinker if we didn't split any page, but the queue is empty. 3254cb8d68ecSKirill A. Shutemov * This can happen if pages were freed under us. 3255cb8d68ecSKirill A. Shutemov */ 3256364c1eebSYang Shi if (!split && list_empty(&ds_queue->split_queue)) 3257cb8d68ecSKirill A. Shutemov return SHRINK_STOP; 3258cb8d68ecSKirill A. Shutemov return split; 32599a982250SKirill A. Shutemov } 32609a982250SKirill A. Shutemov 326149071d43SKirill A. Shutemov #ifdef CONFIG_DEBUG_FS 3262fa6c0231SZi Yan static void split_huge_pages_all(void) 326349071d43SKirill A. Shutemov { 326449071d43SKirill A. Shutemov struct zone *zone; 326549071d43SKirill A. Shutemov struct page *page; 3266630e7c5eSKefeng Wang struct folio *folio; 326749071d43SKirill A. Shutemov unsigned long pfn, max_zone_pfn; 326849071d43SKirill A. Shutemov unsigned long total = 0, split = 0; 326949071d43SKirill A. Shutemov 3270fa6c0231SZi Yan pr_debug("Split all THPs\n"); 3271a17206daSMiaohe Lin for_each_zone(zone) { 3272a17206daSMiaohe Lin if (!managed_zone(zone)) 3273a17206daSMiaohe Lin continue; 327449071d43SKirill A. Shutemov max_zone_pfn = zone_end_pfn(zone); 327549071d43SKirill A. Shutemov for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { 3276a17206daSMiaohe Lin int nr_pages; 327749071d43SKirill A. Shutemov 32782b7aa91bSNaoya Horiguchi page = pfn_to_online_page(pfn); 3279630e7c5eSKefeng Wang if (!page || PageTail(page)) 3280630e7c5eSKefeng Wang continue; 3281630e7c5eSKefeng Wang folio = page_folio(page); 3282630e7c5eSKefeng Wang if (!folio_try_get(folio)) 328349071d43SKirill A. Shutemov continue; 328449071d43SKirill A. Shutemov 3285630e7c5eSKefeng Wang if (unlikely(page_folio(page) != folio)) 328649071d43SKirill A. Shutemov goto next; 328749071d43SKirill A. Shutemov 3288630e7c5eSKefeng Wang if (zone != folio_zone(folio)) 3289630e7c5eSKefeng Wang goto next; 3290630e7c5eSKefeng Wang 3291630e7c5eSKefeng Wang if (!folio_test_large(folio) 3292630e7c5eSKefeng Wang || folio_test_hugetlb(folio) 3293630e7c5eSKefeng Wang || !folio_test_lru(folio)) 329449071d43SKirill A. Shutemov goto next; 329549071d43SKirill A. Shutemov 329649071d43SKirill A. Shutemov total++; 3297630e7c5eSKefeng Wang folio_lock(folio); 3298630e7c5eSKefeng Wang nr_pages = folio_nr_pages(folio); 3299630e7c5eSKefeng Wang if (!split_folio(folio)) 330049071d43SKirill A. Shutemov split++; 3301a17206daSMiaohe Lin pfn += nr_pages - 1; 3302630e7c5eSKefeng Wang folio_unlock(folio); 330349071d43SKirill A. Shutemov next: 3304630e7c5eSKefeng Wang folio_put(folio); 3305fa6c0231SZi Yan cond_resched(); 330649071d43SKirill A. Shutemov } 330749071d43SKirill A. Shutemov } 330849071d43SKirill A. Shutemov 3309fa6c0231SZi Yan pr_debug("%lu of %lu THP split\n", split, total); 331049071d43SKirill A. Shutemov } 3311fa6c0231SZi Yan 3312fa6c0231SZi Yan static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma) 3313fa6c0231SZi Yan { 3314fa6c0231SZi Yan return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) || 3315fa6c0231SZi Yan is_vm_hugetlb_page(vma); 3316fa6c0231SZi Yan } 3317fa6c0231SZi Yan 3318fa6c0231SZi Yan static int split_huge_pages_pid(int pid, unsigned long vaddr_start, 3319fa6c0231SZi Yan unsigned long vaddr_end) 3320fa6c0231SZi Yan { 3321fa6c0231SZi Yan int ret = 0; 3322fa6c0231SZi Yan struct task_struct *task; 3323fa6c0231SZi Yan struct mm_struct *mm; 3324fa6c0231SZi Yan unsigned long total = 0, split = 0; 3325fa6c0231SZi Yan unsigned long addr; 3326fa6c0231SZi Yan 3327fa6c0231SZi Yan vaddr_start &= PAGE_MASK; 3328fa6c0231SZi Yan vaddr_end &= PAGE_MASK; 3329fa6c0231SZi Yan 3330fa6c0231SZi Yan /* Find the task_struct from pid */ 3331fa6c0231SZi Yan rcu_read_lock(); 3332fa6c0231SZi Yan task = find_task_by_vpid(pid); 3333fa6c0231SZi Yan if (!task) { 3334fa6c0231SZi Yan rcu_read_unlock(); 3335fa6c0231SZi Yan ret = -ESRCH; 3336fa6c0231SZi Yan goto out; 3337fa6c0231SZi Yan } 3338fa6c0231SZi Yan get_task_struct(task); 3339fa6c0231SZi Yan rcu_read_unlock(); 3340fa6c0231SZi Yan 3341fa6c0231SZi Yan /* Find the mm_struct */ 3342fa6c0231SZi Yan mm = get_task_mm(task); 3343fa6c0231SZi Yan put_task_struct(task); 3344fa6c0231SZi Yan 3345fa6c0231SZi Yan if (!mm) { 3346fa6c0231SZi Yan ret = -EINVAL; 3347fa6c0231SZi Yan goto out; 3348fa6c0231SZi Yan } 3349fa6c0231SZi Yan 3350fa6c0231SZi Yan pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n", 3351fa6c0231SZi Yan pid, vaddr_start, vaddr_end); 3352fa6c0231SZi Yan 3353fa6c0231SZi Yan mmap_read_lock(mm); 3354fa6c0231SZi Yan /* 3355fa6c0231SZi Yan * always increase addr by PAGE_SIZE, since we could have a PTE page 3356fa6c0231SZi Yan * table filled with PTE-mapped THPs, each of which is distinct. 3357fa6c0231SZi Yan */ 3358fa6c0231SZi Yan for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) { 335974ba2b38SMiaohe Lin struct vm_area_struct *vma = vma_lookup(mm, addr); 3360fa6c0231SZi Yan struct page *page; 3361a644b0abSMatthew Wilcox (Oracle) struct folio *folio; 3362fa6c0231SZi Yan 336374ba2b38SMiaohe Lin if (!vma) 3364fa6c0231SZi Yan break; 3365fa6c0231SZi Yan 3366fa6c0231SZi Yan /* skip special VMA and hugetlb VMA */ 3367fa6c0231SZi Yan if (vma_not_suitable_for_thp_split(vma)) { 3368fa6c0231SZi Yan addr = vma->vm_end; 3369fa6c0231SZi Yan continue; 3370fa6c0231SZi Yan } 3371fa6c0231SZi Yan 3372fa6c0231SZi Yan /* FOLL_DUMP to ignore special (like zero) pages */ 337387d2762eSMiaohe Lin page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 3374fa6c0231SZi Yan 3375f7091ed6SHaiyue Wang if (IS_ERR_OR_NULL(page)) 3376fa6c0231SZi Yan continue; 3377fa6c0231SZi Yan 3378a644b0abSMatthew Wilcox (Oracle) folio = page_folio(page); 3379a644b0abSMatthew Wilcox (Oracle) if (!is_transparent_hugepage(folio)) 3380fa6c0231SZi Yan goto next; 3381fa6c0231SZi Yan 3382fa6c0231SZi Yan total++; 3383a644b0abSMatthew Wilcox (Oracle) if (!can_split_folio(folio, NULL)) 3384fa6c0231SZi Yan goto next; 3385fa6c0231SZi Yan 3386a644b0abSMatthew Wilcox (Oracle) if (!folio_trylock(folio)) 3387fa6c0231SZi Yan goto next; 3388fa6c0231SZi Yan 3389a644b0abSMatthew Wilcox (Oracle) if (!split_folio(folio)) 3390fa6c0231SZi Yan split++; 3391fa6c0231SZi Yan 3392a644b0abSMatthew Wilcox (Oracle) folio_unlock(folio); 3393fa6c0231SZi Yan next: 3394a644b0abSMatthew Wilcox (Oracle) folio_put(folio); 3395fa6c0231SZi Yan cond_resched(); 3396fa6c0231SZi Yan } 3397fa6c0231SZi Yan mmap_read_unlock(mm); 3398fa6c0231SZi Yan mmput(mm); 3399fa6c0231SZi Yan 3400fa6c0231SZi Yan pr_debug("%lu of %lu THP split\n", split, total); 3401fa6c0231SZi Yan 3402fa6c0231SZi Yan out: 3403fa6c0231SZi Yan return ret; 3404fa6c0231SZi Yan } 3405fa6c0231SZi Yan 3406fbe37501SZi Yan static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start, 3407fbe37501SZi Yan pgoff_t off_end) 3408fbe37501SZi Yan { 3409fbe37501SZi Yan struct filename *file; 3410fbe37501SZi Yan struct file *candidate; 3411fbe37501SZi Yan struct address_space *mapping; 3412fbe37501SZi Yan int ret = -EINVAL; 3413fbe37501SZi Yan pgoff_t index; 3414fbe37501SZi Yan int nr_pages = 1; 3415fbe37501SZi Yan unsigned long total = 0, split = 0; 3416fbe37501SZi Yan 3417fbe37501SZi Yan file = getname_kernel(file_path); 3418fbe37501SZi Yan if (IS_ERR(file)) 3419fbe37501SZi Yan return ret; 3420fbe37501SZi Yan 3421fbe37501SZi Yan candidate = file_open_name(file, O_RDONLY, 0); 3422fbe37501SZi Yan if (IS_ERR(candidate)) 3423fbe37501SZi Yan goto out; 3424fbe37501SZi Yan 3425fbe37501SZi Yan pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n", 3426fbe37501SZi Yan file_path, off_start, off_end); 3427fbe37501SZi Yan 3428fbe37501SZi Yan mapping = candidate->f_mapping; 3429fbe37501SZi Yan 3430fbe37501SZi Yan for (index = off_start; index < off_end; index += nr_pages) { 34311fb130b2SChristoph Hellwig struct folio *folio = filemap_get_folio(mapping, index); 3432fbe37501SZi Yan 3433fbe37501SZi Yan nr_pages = 1; 343466dabbb6SChristoph Hellwig if (IS_ERR(folio)) 3435fbe37501SZi Yan continue; 3436fbe37501SZi Yan 34379ee2c086SMatthew Wilcox (Oracle) if (!folio_test_large(folio)) 3438fbe37501SZi Yan goto next; 3439fbe37501SZi Yan 3440fbe37501SZi Yan total++; 34419ee2c086SMatthew Wilcox (Oracle) nr_pages = folio_nr_pages(folio); 3442fbe37501SZi Yan 34439ee2c086SMatthew Wilcox (Oracle) if (!folio_trylock(folio)) 3444fbe37501SZi Yan goto next; 3445fbe37501SZi Yan 34469ee2c086SMatthew Wilcox (Oracle) if (!split_folio(folio)) 3447fbe37501SZi Yan split++; 3448fbe37501SZi Yan 34499ee2c086SMatthew Wilcox (Oracle) folio_unlock(folio); 3450fbe37501SZi Yan next: 34519ee2c086SMatthew Wilcox (Oracle) folio_put(folio); 3452fbe37501SZi Yan cond_resched(); 3453fbe37501SZi Yan } 3454fbe37501SZi Yan 3455fbe37501SZi Yan filp_close(candidate, NULL); 3456fbe37501SZi Yan ret = 0; 3457fbe37501SZi Yan 3458fbe37501SZi Yan pr_debug("%lu of %lu file-backed THP split\n", split, total); 3459fbe37501SZi Yan out: 3460fbe37501SZi Yan putname(file); 3461fbe37501SZi Yan return ret; 3462fbe37501SZi Yan } 3463fbe37501SZi Yan 3464fa6c0231SZi Yan #define MAX_INPUT_BUF_SZ 255 3465fa6c0231SZi Yan 3466fa6c0231SZi Yan static ssize_t split_huge_pages_write(struct file *file, const char __user *buf, 3467fa6c0231SZi Yan size_t count, loff_t *ppops) 3468fa6c0231SZi Yan { 3469fa6c0231SZi Yan static DEFINE_MUTEX(split_debug_mutex); 3470fa6c0231SZi Yan ssize_t ret; 3471fbe37501SZi Yan /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */ 3472fbe37501SZi Yan char input_buf[MAX_INPUT_BUF_SZ]; 3473fa6c0231SZi Yan int pid; 3474fa6c0231SZi Yan unsigned long vaddr_start, vaddr_end; 3475fa6c0231SZi Yan 3476fa6c0231SZi Yan ret = mutex_lock_interruptible(&split_debug_mutex); 3477fa6c0231SZi Yan if (ret) 3478fa6c0231SZi Yan return ret; 3479fa6c0231SZi Yan 3480fa6c0231SZi Yan ret = -EFAULT; 3481fa6c0231SZi Yan 3482fa6c0231SZi Yan memset(input_buf, 0, MAX_INPUT_BUF_SZ); 3483fa6c0231SZi Yan if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ))) 3484fa6c0231SZi Yan goto out; 3485fa6c0231SZi Yan 3486fa6c0231SZi Yan input_buf[MAX_INPUT_BUF_SZ - 1] = '\0'; 3487fbe37501SZi Yan 3488fbe37501SZi Yan if (input_buf[0] == '/') { 3489fbe37501SZi Yan char *tok; 3490fbe37501SZi Yan char *buf = input_buf; 3491fbe37501SZi Yan char file_path[MAX_INPUT_BUF_SZ]; 3492fbe37501SZi Yan pgoff_t off_start = 0, off_end = 0; 3493fbe37501SZi Yan size_t input_len = strlen(input_buf); 3494fbe37501SZi Yan 3495fbe37501SZi Yan tok = strsep(&buf, ","); 3496fbe37501SZi Yan if (tok) { 34971212e00cSMatthew Wilcox (Oracle) strcpy(file_path, tok); 3498fbe37501SZi Yan } else { 3499fbe37501SZi Yan ret = -EINVAL; 3500fbe37501SZi Yan goto out; 3501fbe37501SZi Yan } 3502fbe37501SZi Yan 3503fbe37501SZi Yan ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end); 3504fbe37501SZi Yan if (ret != 2) { 3505fbe37501SZi Yan ret = -EINVAL; 3506fbe37501SZi Yan goto out; 3507fbe37501SZi Yan } 3508fbe37501SZi Yan ret = split_huge_pages_in_file(file_path, off_start, off_end); 3509fbe37501SZi Yan if (!ret) 3510fbe37501SZi Yan ret = input_len; 3511fbe37501SZi Yan 3512fbe37501SZi Yan goto out; 3513fbe37501SZi Yan } 3514fbe37501SZi Yan 3515fa6c0231SZi Yan ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end); 3516fa6c0231SZi Yan if (ret == 1 && pid == 1) { 3517fa6c0231SZi Yan split_huge_pages_all(); 3518fa6c0231SZi Yan ret = strlen(input_buf); 3519fa6c0231SZi Yan goto out; 3520fa6c0231SZi Yan } else if (ret != 3) { 3521fa6c0231SZi Yan ret = -EINVAL; 3522fa6c0231SZi Yan goto out; 3523fa6c0231SZi Yan } 3524fa6c0231SZi Yan 3525fa6c0231SZi Yan ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end); 3526fa6c0231SZi Yan if (!ret) 3527fa6c0231SZi Yan ret = strlen(input_buf); 3528fa6c0231SZi Yan out: 3529fa6c0231SZi Yan mutex_unlock(&split_debug_mutex); 3530fa6c0231SZi Yan return ret; 3531fa6c0231SZi Yan 3532fa6c0231SZi Yan } 3533fa6c0231SZi Yan 3534fa6c0231SZi Yan static const struct file_operations split_huge_pages_fops = { 3535fa6c0231SZi Yan .owner = THIS_MODULE, 3536fa6c0231SZi Yan .write = split_huge_pages_write, 3537fa6c0231SZi Yan .llseek = no_llseek, 3538fa6c0231SZi Yan }; 353949071d43SKirill A. Shutemov 354049071d43SKirill A. Shutemov static int __init split_huge_pages_debugfs(void) 354149071d43SKirill A. Shutemov { 3542d9f7979cSGreg Kroah-Hartman debugfs_create_file("split_huge_pages", 0200, NULL, NULL, 354349071d43SKirill A. Shutemov &split_huge_pages_fops); 354449071d43SKirill A. Shutemov return 0; 354549071d43SKirill A. Shutemov } 354649071d43SKirill A. Shutemov late_initcall(split_huge_pages_debugfs); 354749071d43SKirill A. Shutemov #endif 3548616b8371SZi Yan 3549616b8371SZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 35507f5abe60SDavid Hildenbrand int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, 3551616b8371SZi Yan struct page *page) 3552616b8371SZi Yan { 3553a8e61d58SDavid Hildenbrand struct folio *folio = page_folio(page); 3554616b8371SZi Yan struct vm_area_struct *vma = pvmw->vma; 3555616b8371SZi Yan struct mm_struct *mm = vma->vm_mm; 3556616b8371SZi Yan unsigned long address = pvmw->address; 35576c287605SDavid Hildenbrand bool anon_exclusive; 3558616b8371SZi Yan pmd_t pmdval; 3559616b8371SZi Yan swp_entry_t entry; 3560ab6e3d09SNaoya Horiguchi pmd_t pmdswp; 3561616b8371SZi Yan 3562616b8371SZi Yan if (!(pvmw->pmd && !pvmw->pte)) 35637f5abe60SDavid Hildenbrand return 0; 3564616b8371SZi Yan 3565616b8371SZi Yan flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); 35668a8683adSHuang Ying pmdval = pmdp_invalidate(vma, address, pvmw->pmd); 35676c287605SDavid Hildenbrand 3568e3b4b137SDavid Hildenbrand /* See folio_try_share_anon_rmap_pmd(): invalidate PMD first. */ 3569a8e61d58SDavid Hildenbrand anon_exclusive = folio_test_anon(folio) && PageAnonExclusive(page); 3570e3b4b137SDavid Hildenbrand if (anon_exclusive && folio_try_share_anon_rmap_pmd(folio, page)) { 35716c287605SDavid Hildenbrand set_pmd_at(mm, address, pvmw->pmd, pmdval); 35727f5abe60SDavid Hildenbrand return -EBUSY; 35736c287605SDavid Hildenbrand } 35746c287605SDavid Hildenbrand 3575616b8371SZi Yan if (pmd_dirty(pmdval)) 3576db44c658SDavid Hildenbrand folio_mark_dirty(folio); 35774dd845b5SAlistair Popple if (pmd_write(pmdval)) 35784dd845b5SAlistair Popple entry = make_writable_migration_entry(page_to_pfn(page)); 35796c287605SDavid Hildenbrand else if (anon_exclusive) 35806c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry(page_to_pfn(page)); 35814dd845b5SAlistair Popple else 35824dd845b5SAlistair Popple entry = make_readable_migration_entry(page_to_pfn(page)); 35832e346877SPeter Xu if (pmd_young(pmdval)) 35842e346877SPeter Xu entry = make_migration_entry_young(entry); 35852e346877SPeter Xu if (pmd_dirty(pmdval)) 35862e346877SPeter Xu entry = make_migration_entry_dirty(entry); 3587ab6e3d09SNaoya Horiguchi pmdswp = swp_entry_to_pmd(entry); 3588ab6e3d09SNaoya Horiguchi if (pmd_soft_dirty(pmdval)) 3589ab6e3d09SNaoya Horiguchi pmdswp = pmd_swp_mksoft_dirty(pmdswp); 359024bf08c4SDavid Hildenbrand if (pmd_uffd_wp(pmdval)) 359124bf08c4SDavid Hildenbrand pmdswp = pmd_swp_mkuffd_wp(pmdswp); 3592ab6e3d09SNaoya Horiguchi set_pmd_at(mm, address, pvmw->pmd, pmdswp); 3593a8e61d58SDavid Hildenbrand folio_remove_rmap_pmd(folio, page, vma); 3594a8e61d58SDavid Hildenbrand folio_put(folio); 3595283fd6feSAnshuman Khandual trace_set_migration_pmd(address, pmd_val(pmdswp)); 35967f5abe60SDavid Hildenbrand 35977f5abe60SDavid Hildenbrand return 0; 3598616b8371SZi Yan } 3599616b8371SZi Yan 3600616b8371SZi Yan void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) 3601616b8371SZi Yan { 360214d85a6eSDavid Hildenbrand struct folio *folio = page_folio(new); 3603616b8371SZi Yan struct vm_area_struct *vma = pvmw->vma; 3604616b8371SZi Yan struct mm_struct *mm = vma->vm_mm; 3605616b8371SZi Yan unsigned long address = pvmw->address; 36064fba8f2aSMiaohe Lin unsigned long haddr = address & HPAGE_PMD_MASK; 3607616b8371SZi Yan pmd_t pmde; 3608616b8371SZi Yan swp_entry_t entry; 3609616b8371SZi Yan 3610616b8371SZi Yan if (!(pvmw->pmd && !pvmw->pte)) 3611616b8371SZi Yan return; 3612616b8371SZi Yan 3613616b8371SZi Yan entry = pmd_to_swp_entry(*pvmw->pmd); 361414d85a6eSDavid Hildenbrand folio_get(folio); 36152e346877SPeter Xu pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot)); 3616ab6e3d09SNaoya Horiguchi if (pmd_swp_soft_dirty(*pvmw->pmd)) 3617ab6e3d09SNaoya Horiguchi pmde = pmd_mksoft_dirty(pmde); 36183c811f78SDavid Hildenbrand if (is_writable_migration_entry(entry)) 3619161e393cSRick Edgecombe pmde = pmd_mkwrite(pmde, vma); 36208f34f1eaSPeter Xu if (pmd_swp_uffd_wp(*pvmw->pmd)) 3621f1eb1bacSPeter Xu pmde = pmd_mkuffd_wp(pmde); 36222e346877SPeter Xu if (!is_migration_entry_young(entry)) 36232e346877SPeter Xu pmde = pmd_mkold(pmde); 36242e346877SPeter Xu /* NOTE: this may contain setting soft-dirty on some archs */ 362514d85a6eSDavid Hildenbrand if (folio_test_dirty(folio) && is_migration_entry_dirty(entry)) 36262e346877SPeter Xu pmde = pmd_mkdirty(pmde); 3627616b8371SZi Yan 362814d85a6eSDavid Hildenbrand if (folio_test_anon(folio)) { 3629395db7b1SDavid Hildenbrand rmap_t rmap_flags = RMAP_NONE; 36306c287605SDavid Hildenbrand 36316c287605SDavid Hildenbrand if (!is_readable_migration_entry(entry)) 36326c287605SDavid Hildenbrand rmap_flags |= RMAP_EXCLUSIVE; 36336c287605SDavid Hildenbrand 3634395db7b1SDavid Hildenbrand folio_add_anon_rmap_pmd(folio, new, vma, haddr, rmap_flags); 36356c287605SDavid Hildenbrand } else { 363614d85a6eSDavid Hildenbrand folio_add_file_rmap_pmd(folio, new, vma); 36376c287605SDavid Hildenbrand } 363814d85a6eSDavid Hildenbrand VM_BUG_ON(pmd_write(pmde) && folio_test_anon(folio) && !PageAnonExclusive(new)); 36394fba8f2aSMiaohe Lin set_pmd_at(mm, haddr, pvmw->pmd, pmde); 36405cbcf225SMuchun Song 36415cbcf225SMuchun Song /* No need to invalidate - it was non-present before */ 3642616b8371SZi Yan update_mmu_cache_pmd(vma, address, pvmw->pmd); 3643283fd6feSAnshuman Khandual trace_remove_migration_pmd(address, pmd_val(pmde)); 3644616b8371SZi Yan } 3645616b8371SZi Yan #endif 3646