1d61ea1cbSPeter Xu 2457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 31da177e4SLinus Torvalds /* 41da177e4SLinus Torvalds * linux/mm/memory.c 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 71da177e4SLinus Torvalds */ 81da177e4SLinus Torvalds 91da177e4SLinus Torvalds /* 101da177e4SLinus Torvalds * demand-loading started 01.12.91 - seems it is high on the list of 111da177e4SLinus Torvalds * things wanted, and it should be easy to implement. - Linus 121da177e4SLinus Torvalds */ 131da177e4SLinus Torvalds 141da177e4SLinus Torvalds /* 151da177e4SLinus Torvalds * Ok, demand-loading was easy, shared pages a little bit tricker. Shared 161da177e4SLinus Torvalds * pages started 02.12.91, seems to work. - Linus. 171da177e4SLinus Torvalds * 181da177e4SLinus Torvalds * Tested sharing by executing about 30 /bin/sh: under the old kernel it 191da177e4SLinus Torvalds * would have taken more than the 6M I have free, but it worked well as 201da177e4SLinus Torvalds * far as I could see. 211da177e4SLinus Torvalds * 221da177e4SLinus Torvalds * Also corrected some "invalidate()"s - I wasn't doing enough of them. 231da177e4SLinus Torvalds */ 241da177e4SLinus Torvalds 251da177e4SLinus Torvalds /* 261da177e4SLinus Torvalds * Real VM (paging to/from disk) started 18.12.91. Much more work and 271da177e4SLinus Torvalds * thought has to go into this. Oh, well.. 281da177e4SLinus Torvalds * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why. 291da177e4SLinus Torvalds * Found it. Everything seems to work now. 301da177e4SLinus Torvalds * 20.12.91 - Ok, making the swap-device changeable like the root. 311da177e4SLinus Torvalds */ 321da177e4SLinus Torvalds 331da177e4SLinus Torvalds /* 341da177e4SLinus Torvalds * 05.04.94 - Multi-page memory management added for v1.1. 351da177e4SLinus Torvalds * Idea by Alex Bligh (alex@cconcepts.co.uk) 361da177e4SLinus Torvalds * 371da177e4SLinus Torvalds * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG 381da177e4SLinus Torvalds * (Gerhard.Wichert@pdb.siemens.de) 391da177e4SLinus Torvalds * 401da177e4SLinus Torvalds * Aug/Sep 2004 Changed to four level page tables (Andi Kleen) 411da177e4SLinus Torvalds */ 421da177e4SLinus Torvalds 431da177e4SLinus Torvalds #include <linux/kernel_stat.h> 441da177e4SLinus Torvalds #include <linux/mm.h> 4536090defSArnd Bergmann #include <linux/mm_inline.h> 466e84f315SIngo Molnar #include <linux/sched/mm.h> 47f7ccbae4SIngo Molnar #include <linux/sched/coredump.h> 486a3827d7SIngo Molnar #include <linux/sched/numa_balancing.h> 4929930025SIngo Molnar #include <linux/sched/task.h> 501da177e4SLinus Torvalds #include <linux/hugetlb.h> 511da177e4SLinus Torvalds #include <linux/mman.h> 521da177e4SLinus Torvalds #include <linux/swap.h> 531da177e4SLinus Torvalds #include <linux/highmem.h> 541da177e4SLinus Torvalds #include <linux/pagemap.h> 555042db43SJérôme Glisse #include <linux/memremap.h> 56b073d7f8SAlexander Potapenko #include <linux/kmsan.h> 579a840895SHugh Dickins #include <linux/ksm.h> 581da177e4SLinus Torvalds #include <linux/rmap.h> 59b95f1b31SPaul Gortmaker #include <linux/export.h> 600ff92245SShailabh Nagar #include <linux/delayacct.h> 611da177e4SLinus Torvalds #include <linux/init.h> 6201c8f1c4SDan Williams #include <linux/pfn_t.h> 63edc79b2aSPeter Zijlstra #include <linux/writeback.h> 648a9f3ccdSBalbir Singh #include <linux/memcontrol.h> 65cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 663dc14741SHugh Dickins #include <linux/swapops.h> 673dc14741SHugh Dickins #include <linux/elf.h> 685a0e3ad6STejun Heo #include <linux/gfp.h> 694daae3b4SMel Gorman #include <linux/migrate.h> 702fbc57c5SAndy Shevchenko #include <linux/string.h> 71467b171aSAneesh Kumar K.V #include <linux/memory-tiers.h> 721592eef0SKirill A. Shutemov #include <linux/debugfs.h> 736b251fc9SAndrea Arcangeli #include <linux/userfaultfd_k.h> 74bc2466e4SJan Kara #include <linux/dax.h> 756b31d595SMichal Hocko #include <linux/oom.h> 7698fa15f3SAnshuman Khandual #include <linux/numa.h> 77bce617edSPeter Xu #include <linux/perf_event.h> 78bce617edSPeter Xu #include <linux/ptrace.h> 79e80d3909SJoerg Roedel #include <linux/vmalloc.h> 8033024536SHuang Ying #include <linux/sched/sysctl.h> 811da177e4SLinus Torvalds 82b3d1411bSJoel Fernandes (Google) #include <trace/events/kmem.h> 83b3d1411bSJoel Fernandes (Google) 846952b61dSAlexey Dobriyan #include <asm/io.h> 8533a709b2SDave Hansen #include <asm/mmu_context.h> 861da177e4SLinus Torvalds #include <asm/pgalloc.h> 877c0f6ba6SLinus Torvalds #include <linux/uaccess.h> 881da177e4SLinus Torvalds #include <asm/tlb.h> 891da177e4SLinus Torvalds #include <asm/tlbflush.h> 901da177e4SLinus Torvalds 91e80d3909SJoerg Roedel #include "pgalloc-track.h" 9242b77728SJan Beulich #include "internal.h" 93014bb1deSNeilBrown #include "swap.h" 9442b77728SJan Beulich 95af27d940SArnd Bergmann #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST) 9690572890SPeter Zijlstra #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid. 9775980e97SPeter Zijlstra #endif 9875980e97SPeter Zijlstra 99a9ee6cf5SMike Rapoport #ifndef CONFIG_NUMA 1001da177e4SLinus Torvalds unsigned long max_mapnr; 1011da177e4SLinus Torvalds EXPORT_SYMBOL(max_mapnr); 102166f61b9STobin C Harding 103166f61b9STobin C Harding struct page *mem_map; 1041da177e4SLinus Torvalds EXPORT_SYMBOL(mem_map); 1051da177e4SLinus Torvalds #endif 1061da177e4SLinus Torvalds 1075c041f5dSPeter Xu static vm_fault_t do_fault(struct vm_fault *vmf); 1082bad466cSPeter Xu static vm_fault_t do_anonymous_page(struct vm_fault *vmf); 1092bad466cSPeter Xu static bool vmf_pte_changed(struct vm_fault *vmf); 1102bad466cSPeter Xu 1112bad466cSPeter Xu /* 1122bad466cSPeter Xu * Return true if the original pte was a uffd-wp pte marker (so the pte was 1132bad466cSPeter Xu * wr-protected). 1142bad466cSPeter Xu */ 1152bad466cSPeter Xu static bool vmf_orig_pte_uffd_wp(struct vm_fault *vmf) 1162bad466cSPeter Xu { 1172bad466cSPeter Xu if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID)) 1182bad466cSPeter Xu return false; 1192bad466cSPeter Xu 1202bad466cSPeter Xu return pte_marker_uffd_wp(vmf->orig_pte); 1212bad466cSPeter Xu } 1225c041f5dSPeter Xu 1231da177e4SLinus Torvalds /* 1241da177e4SLinus Torvalds * A number of key systems in x86 including ioremap() rely on the assumption 1251da177e4SLinus Torvalds * that high_memory defines the upper bound on direct map memory, then end 126e99fb98dSKefeng Wang * of ZONE_NORMAL. 1271da177e4SLinus Torvalds */ 1281da177e4SLinus Torvalds void *high_memory; 1291da177e4SLinus Torvalds EXPORT_SYMBOL(high_memory); 1301da177e4SLinus Torvalds 13132a93233SIngo Molnar /* 13232a93233SIngo Molnar * Randomize the address space (stacks, mmaps, brk, etc.). 13332a93233SIngo Molnar * 13432a93233SIngo Molnar * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization, 13532a93233SIngo Molnar * as ancient (libc5 based) binaries can segfault. ) 13632a93233SIngo Molnar */ 13732a93233SIngo Molnar int randomize_va_space __read_mostly = 13832a93233SIngo Molnar #ifdef CONFIG_COMPAT_BRK 13932a93233SIngo Molnar 1; 14032a93233SIngo Molnar #else 14132a93233SIngo Molnar 2; 14232a93233SIngo Molnar #endif 143a62eaf15SAndi Kleen 14446bdb427SWill Deacon #ifndef arch_wants_old_prefaulted_pte 14546bdb427SWill Deacon static inline bool arch_wants_old_prefaulted_pte(void) 14646bdb427SWill Deacon { 14746bdb427SWill Deacon /* 14846bdb427SWill Deacon * Transitioning a PTE from 'old' to 'young' can be expensive on 14946bdb427SWill Deacon * some architectures, even if it's performed in hardware. By 15046bdb427SWill Deacon * default, "false" means prefaulted entries will be 'young'. 15146bdb427SWill Deacon */ 15246bdb427SWill Deacon return false; 15346bdb427SWill Deacon } 15446bdb427SWill Deacon #endif 15546bdb427SWill Deacon 156a62eaf15SAndi Kleen static int __init disable_randmaps(char *s) 157a62eaf15SAndi Kleen { 158a62eaf15SAndi Kleen randomize_va_space = 0; 1599b41046cSOGAWA Hirofumi return 1; 160a62eaf15SAndi Kleen } 161a62eaf15SAndi Kleen __setup("norandmaps", disable_randmaps); 162a62eaf15SAndi Kleen 16362eede62SHugh Dickins unsigned long zero_pfn __read_mostly; 1640b70068eSArd Biesheuvel EXPORT_SYMBOL(zero_pfn); 1650b70068eSArd Biesheuvel 166166f61b9STobin C Harding unsigned long highest_memmap_pfn __read_mostly; 167166f61b9STobin C Harding 168a13ea5b7SHugh Dickins /* 169a13ea5b7SHugh Dickins * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init() 170a13ea5b7SHugh Dickins */ 171a13ea5b7SHugh Dickins static int __init init_zero_pfn(void) 172a13ea5b7SHugh Dickins { 173a13ea5b7SHugh Dickins zero_pfn = page_to_pfn(ZERO_PAGE(0)); 174a13ea5b7SHugh Dickins return 0; 175a13ea5b7SHugh Dickins } 176e720e7d0SIlya Lipnitskiy early_initcall(init_zero_pfn); 177a62eaf15SAndi Kleen 178f1a79412SShakeel Butt void mm_trace_rss_stat(struct mm_struct *mm, int member) 179b3d1411bSJoel Fernandes (Google) { 180f1a79412SShakeel Butt trace_rss_stat(mm, member); 181b3d1411bSJoel Fernandes (Google) } 182d559db08SKAMEZAWA Hiroyuki 1831da177e4SLinus Torvalds /* 1841da177e4SLinus Torvalds * Note: this doesn't free the actual pages themselves. That 1851da177e4SLinus Torvalds * has been handled earlier when unmapping all the memory regions. 1861da177e4SLinus Torvalds */ 1879e1b32caSBenjamin Herrenschmidt static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, 1889e1b32caSBenjamin Herrenschmidt unsigned long addr) 1891da177e4SLinus Torvalds { 1902f569afdSMartin Schwidefsky pgtable_t token = pmd_pgtable(*pmd); 1911da177e4SLinus Torvalds pmd_clear(pmd); 1929e1b32caSBenjamin Herrenschmidt pte_free_tlb(tlb, token, addr); 193c4812909SKirill A. Shutemov mm_dec_nr_ptes(tlb->mm); 1941da177e4SLinus Torvalds } 1951da177e4SLinus Torvalds 196e0da382cSHugh Dickins static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 197e0da382cSHugh Dickins unsigned long addr, unsigned long end, 198e0da382cSHugh Dickins unsigned long floor, unsigned long ceiling) 1991da177e4SLinus Torvalds { 2001da177e4SLinus Torvalds pmd_t *pmd; 2011da177e4SLinus Torvalds unsigned long next; 202e0da382cSHugh Dickins unsigned long start; 2031da177e4SLinus Torvalds 204e0da382cSHugh Dickins start = addr; 2051da177e4SLinus Torvalds pmd = pmd_offset(pud, addr); 2061da177e4SLinus Torvalds do { 2071da177e4SLinus Torvalds next = pmd_addr_end(addr, end); 2081da177e4SLinus Torvalds if (pmd_none_or_clear_bad(pmd)) 2091da177e4SLinus Torvalds continue; 2109e1b32caSBenjamin Herrenschmidt free_pte_range(tlb, pmd, addr); 2111da177e4SLinus Torvalds } while (pmd++, addr = next, addr != end); 2121da177e4SLinus Torvalds 213e0da382cSHugh Dickins start &= PUD_MASK; 214e0da382cSHugh Dickins if (start < floor) 215e0da382cSHugh Dickins return; 216e0da382cSHugh Dickins if (ceiling) { 217e0da382cSHugh Dickins ceiling &= PUD_MASK; 218e0da382cSHugh Dickins if (!ceiling) 219e0da382cSHugh Dickins return; 2201da177e4SLinus Torvalds } 221e0da382cSHugh Dickins if (end - 1 > ceiling - 1) 222e0da382cSHugh Dickins return; 223e0da382cSHugh Dickins 224e0da382cSHugh Dickins pmd = pmd_offset(pud, start); 225e0da382cSHugh Dickins pud_clear(pud); 2269e1b32caSBenjamin Herrenschmidt pmd_free_tlb(tlb, pmd, start); 227dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(tlb->mm); 2281da177e4SLinus Torvalds } 2291da177e4SLinus Torvalds 230c2febafcSKirill A. Shutemov static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d, 231e0da382cSHugh Dickins unsigned long addr, unsigned long end, 232e0da382cSHugh Dickins unsigned long floor, unsigned long ceiling) 2331da177e4SLinus Torvalds { 2341da177e4SLinus Torvalds pud_t *pud; 2351da177e4SLinus Torvalds unsigned long next; 236e0da382cSHugh Dickins unsigned long start; 2371da177e4SLinus Torvalds 238e0da382cSHugh Dickins start = addr; 239c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 2401da177e4SLinus Torvalds do { 2411da177e4SLinus Torvalds next = pud_addr_end(addr, end); 2421da177e4SLinus Torvalds if (pud_none_or_clear_bad(pud)) 2431da177e4SLinus Torvalds continue; 244e0da382cSHugh Dickins free_pmd_range(tlb, pud, addr, next, floor, ceiling); 2451da177e4SLinus Torvalds } while (pud++, addr = next, addr != end); 2461da177e4SLinus Torvalds 247c2febafcSKirill A. Shutemov start &= P4D_MASK; 248c2febafcSKirill A. Shutemov if (start < floor) 249c2febafcSKirill A. Shutemov return; 250c2febafcSKirill A. Shutemov if (ceiling) { 251c2febafcSKirill A. Shutemov ceiling &= P4D_MASK; 252c2febafcSKirill A. Shutemov if (!ceiling) 253c2febafcSKirill A. Shutemov return; 254c2febafcSKirill A. Shutemov } 255c2febafcSKirill A. Shutemov if (end - 1 > ceiling - 1) 256c2febafcSKirill A. Shutemov return; 257c2febafcSKirill A. Shutemov 258c2febafcSKirill A. Shutemov pud = pud_offset(p4d, start); 259c2febafcSKirill A. Shutemov p4d_clear(p4d); 260c2febafcSKirill A. Shutemov pud_free_tlb(tlb, pud, start); 261b4e98d9aSKirill A. Shutemov mm_dec_nr_puds(tlb->mm); 262c2febafcSKirill A. Shutemov } 263c2febafcSKirill A. Shutemov 264c2febafcSKirill A. Shutemov static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd, 265c2febafcSKirill A. Shutemov unsigned long addr, unsigned long end, 266c2febafcSKirill A. Shutemov unsigned long floor, unsigned long ceiling) 267c2febafcSKirill A. Shutemov { 268c2febafcSKirill A. Shutemov p4d_t *p4d; 269c2febafcSKirill A. Shutemov unsigned long next; 270c2febafcSKirill A. Shutemov unsigned long start; 271c2febafcSKirill A. Shutemov 272c2febafcSKirill A. Shutemov start = addr; 273c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 274c2febafcSKirill A. Shutemov do { 275c2febafcSKirill A. Shutemov next = p4d_addr_end(addr, end); 276c2febafcSKirill A. Shutemov if (p4d_none_or_clear_bad(p4d)) 277c2febafcSKirill A. Shutemov continue; 278c2febafcSKirill A. Shutemov free_pud_range(tlb, p4d, addr, next, floor, ceiling); 279c2febafcSKirill A. Shutemov } while (p4d++, addr = next, addr != end); 280c2febafcSKirill A. Shutemov 281e0da382cSHugh Dickins start &= PGDIR_MASK; 282e0da382cSHugh Dickins if (start < floor) 283e0da382cSHugh Dickins return; 284e0da382cSHugh Dickins if (ceiling) { 285e0da382cSHugh Dickins ceiling &= PGDIR_MASK; 286e0da382cSHugh Dickins if (!ceiling) 287e0da382cSHugh Dickins return; 2881da177e4SLinus Torvalds } 289e0da382cSHugh Dickins if (end - 1 > ceiling - 1) 290e0da382cSHugh Dickins return; 291e0da382cSHugh Dickins 292c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, start); 293e0da382cSHugh Dickins pgd_clear(pgd); 294c2febafcSKirill A. Shutemov p4d_free_tlb(tlb, p4d, start); 2951da177e4SLinus Torvalds } 2961da177e4SLinus Torvalds 2971da177e4SLinus Torvalds /* 298e0da382cSHugh Dickins * This function frees user-level page tables of a process. 2991da177e4SLinus Torvalds */ 30042b77728SJan Beulich void free_pgd_range(struct mmu_gather *tlb, 301e0da382cSHugh Dickins unsigned long addr, unsigned long end, 302e0da382cSHugh Dickins unsigned long floor, unsigned long ceiling) 3031da177e4SLinus Torvalds { 3041da177e4SLinus Torvalds pgd_t *pgd; 3051da177e4SLinus Torvalds unsigned long next; 3061da177e4SLinus Torvalds 307e0da382cSHugh Dickins /* 308e0da382cSHugh Dickins * The next few lines have given us lots of grief... 309e0da382cSHugh Dickins * 310e0da382cSHugh Dickins * Why are we testing PMD* at this top level? Because often 311e0da382cSHugh Dickins * there will be no work to do at all, and we'd prefer not to 312e0da382cSHugh Dickins * go all the way down to the bottom just to discover that. 313e0da382cSHugh Dickins * 314e0da382cSHugh Dickins * Why all these "- 1"s? Because 0 represents both the bottom 315e0da382cSHugh Dickins * of the address space and the top of it (using -1 for the 316e0da382cSHugh Dickins * top wouldn't help much: the masks would do the wrong thing). 317e0da382cSHugh Dickins * The rule is that addr 0 and floor 0 refer to the bottom of 318e0da382cSHugh Dickins * the address space, but end 0 and ceiling 0 refer to the top 319e0da382cSHugh Dickins * Comparisons need to use "end - 1" and "ceiling - 1" (though 320e0da382cSHugh Dickins * that end 0 case should be mythical). 321e0da382cSHugh Dickins * 322e0da382cSHugh Dickins * Wherever addr is brought up or ceiling brought down, we must 323e0da382cSHugh Dickins * be careful to reject "the opposite 0" before it confuses the 324e0da382cSHugh Dickins * subsequent tests. But what about where end is brought down 325e0da382cSHugh Dickins * by PMD_SIZE below? no, end can't go down to 0 there. 326e0da382cSHugh Dickins * 327e0da382cSHugh Dickins * Whereas we round start (addr) and ceiling down, by different 328e0da382cSHugh Dickins * masks at different levels, in order to test whether a table 329e0da382cSHugh Dickins * now has no other vmas using it, so can be freed, we don't 330e0da382cSHugh Dickins * bother to round floor or end up - the tests don't need that. 331e0da382cSHugh Dickins */ 332e0da382cSHugh Dickins 333e0da382cSHugh Dickins addr &= PMD_MASK; 334e0da382cSHugh Dickins if (addr < floor) { 335e0da382cSHugh Dickins addr += PMD_SIZE; 336e0da382cSHugh Dickins if (!addr) 337e0da382cSHugh Dickins return; 338e0da382cSHugh Dickins } 339e0da382cSHugh Dickins if (ceiling) { 340e0da382cSHugh Dickins ceiling &= PMD_MASK; 341e0da382cSHugh Dickins if (!ceiling) 342e0da382cSHugh Dickins return; 343e0da382cSHugh Dickins } 344e0da382cSHugh Dickins if (end - 1 > ceiling - 1) 345e0da382cSHugh Dickins end -= PMD_SIZE; 346e0da382cSHugh Dickins if (addr > end - 1) 347e0da382cSHugh Dickins return; 34807e32661SAneesh Kumar K.V /* 34907e32661SAneesh Kumar K.V * We add page table cache pages with PAGE_SIZE, 35007e32661SAneesh Kumar K.V * (see pte_free_tlb()), flush the tlb if we need 35107e32661SAneesh Kumar K.V */ 352ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, PAGE_SIZE); 35342b77728SJan Beulich pgd = pgd_offset(tlb->mm, addr); 3541da177e4SLinus Torvalds do { 3551da177e4SLinus Torvalds next = pgd_addr_end(addr, end); 3561da177e4SLinus Torvalds if (pgd_none_or_clear_bad(pgd)) 3571da177e4SLinus Torvalds continue; 358c2febafcSKirill A. Shutemov free_p4d_range(tlb, pgd, addr, next, floor, ceiling); 3591da177e4SLinus Torvalds } while (pgd++, addr = next, addr != end); 360e0da382cSHugh Dickins } 361e0da382cSHugh Dickins 362fd892593SLiam R. Howlett void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas, 363763ecb03SLiam R. Howlett struct vm_area_struct *vma, unsigned long floor, 36498e51a22SSuren Baghdasaryan unsigned long ceiling, bool mm_wr_locked) 365e0da382cSHugh Dickins { 366763ecb03SLiam R. Howlett do { 367e0da382cSHugh Dickins unsigned long addr = vma->vm_start; 368763ecb03SLiam R. Howlett struct vm_area_struct *next; 369763ecb03SLiam R. Howlett 370763ecb03SLiam R. Howlett /* 371763ecb03SLiam R. Howlett * Note: USER_PGTABLES_CEILING may be passed as ceiling and may 372763ecb03SLiam R. Howlett * be 0. This will underflow and is okay. 373763ecb03SLiam R. Howlett */ 374fd892593SLiam R. Howlett next = mas_find(mas, ceiling - 1); 375d2406291SPeng Zhang if (unlikely(xa_is_zero(next))) 376d2406291SPeng Zhang next = NULL; 377e0da382cSHugh Dickins 3788f4f8c16SHugh Dickins /* 37925d9e2d1Snpiggin@suse.de * Hide vma from rmap and truncate_pagecache before freeing 38025d9e2d1Snpiggin@suse.de * pgtables 3818f4f8c16SHugh Dickins */ 38298e51a22SSuren Baghdasaryan if (mm_wr_locked) 38398e51a22SSuren Baghdasaryan vma_start_write(vma); 3845beb4930SRik van Riel unlink_anon_vmas(vma); 3858f4f8c16SHugh Dickins unlink_file_vma(vma); 3868f4f8c16SHugh Dickins 3879da61aefSDavid Gibson if (is_vm_hugetlb_page(vma)) { 3883bf5ee95SHugh Dickins hugetlb_free_pgd_range(tlb, addr, vma->vm_end, 3893bf5ee95SHugh Dickins floor, next ? next->vm_start : ceiling); 3903bf5ee95SHugh Dickins } else { 3913bf5ee95SHugh Dickins /* 3923bf5ee95SHugh Dickins * Optimization: gather nearby vmas into one call down 3933bf5ee95SHugh Dickins */ 3943bf5ee95SHugh Dickins while (next && next->vm_start <= vma->vm_end + PMD_SIZE 3954866920bSDavid Gibson && !is_vm_hugetlb_page(next)) { 396e0da382cSHugh Dickins vma = next; 397fd892593SLiam R. Howlett next = mas_find(mas, ceiling - 1); 398d2406291SPeng Zhang if (unlikely(xa_is_zero(next))) 399d2406291SPeng Zhang next = NULL; 40098e51a22SSuren Baghdasaryan if (mm_wr_locked) 40198e51a22SSuren Baghdasaryan vma_start_write(vma); 4025beb4930SRik van Riel unlink_anon_vmas(vma); 4038f4f8c16SHugh Dickins unlink_file_vma(vma); 404e0da382cSHugh Dickins } 4053bf5ee95SHugh Dickins free_pgd_range(tlb, addr, vma->vm_end, 406e0da382cSHugh Dickins floor, next ? next->vm_start : ceiling); 4073bf5ee95SHugh Dickins } 408e0da382cSHugh Dickins vma = next; 409763ecb03SLiam R. Howlett } while (vma); 4101da177e4SLinus Torvalds } 4111da177e4SLinus Torvalds 41203c4f204SQi Zheng void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte) 4131da177e4SLinus Torvalds { 41403c4f204SQi Zheng spinlock_t *ptl = pmd_lock(mm, pmd); 4151bb3630eSHugh Dickins 41603c4f204SQi Zheng if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 41703c4f204SQi Zheng mm_inc_nr_ptes(mm); 418362a61adSNick Piggin /* 419362a61adSNick Piggin * Ensure all pte setup (eg. pte page lock and page clearing) are 420362a61adSNick Piggin * visible before the pte is made visible to other CPUs by being 421362a61adSNick Piggin * put into page tables. 422362a61adSNick Piggin * 423362a61adSNick Piggin * The other side of the story is the pointer chasing in the page 424362a61adSNick Piggin * table walking code (when walking the page table without locking; 425362a61adSNick Piggin * ie. most of the time). Fortunately, these data accesses consist 426362a61adSNick Piggin * of a chain of data-dependent loads, meaning most CPUs (alpha 427362a61adSNick Piggin * being the notable exception) will already guarantee loads are 428362a61adSNick Piggin * seen in-order. See the alpha page table accessors for the 429bb7cdd38SWill Deacon * smp_rmb() barriers in page table walking code. 430362a61adSNick Piggin */ 431362a61adSNick Piggin smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */ 432ed33b5a6SQi Zheng pmd_populate(mm, pmd, *pte); 433ed33b5a6SQi Zheng *pte = NULL; 4344b471e88SKirill A. Shutemov } 435c4088ebdSKirill A. Shutemov spin_unlock(ptl); 436ed33b5a6SQi Zheng } 437ed33b5a6SQi Zheng 438ed33b5a6SQi Zheng int __pte_alloc(struct mm_struct *mm, pmd_t *pmd) 439ed33b5a6SQi Zheng { 440ed33b5a6SQi Zheng pgtable_t new = pte_alloc_one(mm); 441ed33b5a6SQi Zheng if (!new) 442ed33b5a6SQi Zheng return -ENOMEM; 4438ac1f832SAndrea Arcangeli 44403c4f204SQi Zheng pmd_install(mm, pmd, &new); 4452f569afdSMartin Schwidefsky if (new) 4462f569afdSMartin Schwidefsky pte_free(mm, new); 4471bb3630eSHugh Dickins return 0; 4481da177e4SLinus Torvalds } 4491da177e4SLinus Torvalds 4504cf58924SJoel Fernandes (Google) int __pte_alloc_kernel(pmd_t *pmd) 4511da177e4SLinus Torvalds { 4524cf58924SJoel Fernandes (Google) pte_t *new = pte_alloc_one_kernel(&init_mm); 4531da177e4SLinus Torvalds if (!new) 4541bb3630eSHugh Dickins return -ENOMEM; 4551da177e4SLinus Torvalds 456872fec16SHugh Dickins spin_lock(&init_mm.page_table_lock); 4578ac1f832SAndrea Arcangeli if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ 458ed33b5a6SQi Zheng smp_wmb(); /* See comment in pmd_install() */ 459872fec16SHugh Dickins pmd_populate_kernel(&init_mm, pmd, new); 4602f569afdSMartin Schwidefsky new = NULL; 4614b471e88SKirill A. Shutemov } 462872fec16SHugh Dickins spin_unlock(&init_mm.page_table_lock); 4632f569afdSMartin Schwidefsky if (new) 4642f569afdSMartin Schwidefsky pte_free_kernel(&init_mm, new); 4651bb3630eSHugh Dickins return 0; 4661da177e4SLinus Torvalds } 4671da177e4SLinus Torvalds 468d559db08SKAMEZAWA Hiroyuki static inline void init_rss_vec(int *rss) 469ae859762SHugh Dickins { 470d559db08SKAMEZAWA Hiroyuki memset(rss, 0, sizeof(int) * NR_MM_COUNTERS); 471d559db08SKAMEZAWA Hiroyuki } 472d559db08SKAMEZAWA Hiroyuki 473d559db08SKAMEZAWA Hiroyuki static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss) 474d559db08SKAMEZAWA Hiroyuki { 475d559db08SKAMEZAWA Hiroyuki int i; 476d559db08SKAMEZAWA Hiroyuki 477d559db08SKAMEZAWA Hiroyuki for (i = 0; i < NR_MM_COUNTERS; i++) 478d559db08SKAMEZAWA Hiroyuki if (rss[i]) 479d559db08SKAMEZAWA Hiroyuki add_mm_counter(mm, i, rss[i]); 480ae859762SHugh Dickins } 481ae859762SHugh Dickins 4821da177e4SLinus Torvalds /* 4836aab341eSLinus Torvalds * This function is called to print an error when a bad pte 4846aab341eSLinus Torvalds * is found. For example, we might have a PFN-mapped pte in 4856aab341eSLinus Torvalds * a region that doesn't allow it. 486b5810039SNick Piggin * 487b5810039SNick Piggin * The calling function must still handle the error. 488b5810039SNick Piggin */ 4893dc14741SHugh Dickins static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr, 4903dc14741SHugh Dickins pte_t pte, struct page *page) 491b5810039SNick Piggin { 4923dc14741SHugh Dickins pgd_t *pgd = pgd_offset(vma->vm_mm, addr); 493c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, addr); 494c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, addr); 4953dc14741SHugh Dickins pmd_t *pmd = pmd_offset(pud, addr); 4963dc14741SHugh Dickins struct address_space *mapping; 4973dc14741SHugh Dickins pgoff_t index; 498d936cf9bSHugh Dickins static unsigned long resume; 499d936cf9bSHugh Dickins static unsigned long nr_shown; 500d936cf9bSHugh Dickins static unsigned long nr_unshown; 501d936cf9bSHugh Dickins 502d936cf9bSHugh Dickins /* 503d936cf9bSHugh Dickins * Allow a burst of 60 reports, then keep quiet for that minute; 504d936cf9bSHugh Dickins * or allow a steady drip of one report per second. 505d936cf9bSHugh Dickins */ 506d936cf9bSHugh Dickins if (nr_shown == 60) { 507d936cf9bSHugh Dickins if (time_before(jiffies, resume)) { 508d936cf9bSHugh Dickins nr_unshown++; 509d936cf9bSHugh Dickins return; 510d936cf9bSHugh Dickins } 511d936cf9bSHugh Dickins if (nr_unshown) { 5121170532bSJoe Perches pr_alert("BUG: Bad page map: %lu messages suppressed\n", 513d936cf9bSHugh Dickins nr_unshown); 514d936cf9bSHugh Dickins nr_unshown = 0; 515d936cf9bSHugh Dickins } 516d936cf9bSHugh Dickins nr_shown = 0; 517d936cf9bSHugh Dickins } 518d936cf9bSHugh Dickins if (nr_shown++ == 0) 519d936cf9bSHugh Dickins resume = jiffies + 60 * HZ; 5203dc14741SHugh Dickins 5213dc14741SHugh Dickins mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL; 5223dc14741SHugh Dickins index = linear_page_index(vma, addr); 5233dc14741SHugh Dickins 5241170532bSJoe Perches pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n", 5253dc14741SHugh Dickins current->comm, 5263dc14741SHugh Dickins (long long)pte_val(pte), (long long)pmd_val(*pmd)); 527718a3821SWu Fengguang if (page) 528f0b791a3SDave Hansen dump_page(page, "bad pte"); 5296aa9b8b2SKefeng Wang pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n", 5303dc14741SHugh Dickins (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index); 5317e0a1265SMatthew Wilcox (Oracle) pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n", 5322682582aSKonstantin Khlebnikov vma->vm_file, 5332682582aSKonstantin Khlebnikov vma->vm_ops ? vma->vm_ops->fault : NULL, 5342682582aSKonstantin Khlebnikov vma->vm_file ? vma->vm_file->f_op->mmap : NULL, 5357e0a1265SMatthew Wilcox (Oracle) mapping ? mapping->a_ops->read_folio : NULL); 536b5810039SNick Piggin dump_stack(); 537373d4d09SRusty Russell add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 538b5810039SNick Piggin } 539b5810039SNick Piggin 540b5810039SNick Piggin /* 5417e675137SNick Piggin * vm_normal_page -- This function gets the "struct page" associated with a pte. 5426aab341eSLinus Torvalds * 5437e675137SNick Piggin * "Special" mappings do not wish to be associated with a "struct page" (either 5447e675137SNick Piggin * it doesn't exist, or it exists but they don't want to touch it). In this 5457e675137SNick Piggin * case, NULL is returned here. "Normal" mappings do have a struct page. 546b379d790SJared Hulbert * 5477e675137SNick Piggin * There are 2 broad cases. Firstly, an architecture may define a pte_special() 5487e675137SNick Piggin * pte bit, in which case this function is trivial. Secondly, an architecture 5497e675137SNick Piggin * may not have a spare pte bit, which requires a more complicated scheme, 5507e675137SNick Piggin * described below. 5517e675137SNick Piggin * 5527e675137SNick Piggin * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a 5537e675137SNick Piggin * special mapping (even if there are underlying and valid "struct pages"). 5547e675137SNick Piggin * COWed pages of a VM_PFNMAP are always normal. 5556aab341eSLinus Torvalds * 556b379d790SJared Hulbert * The way we recognize COWed pages within VM_PFNMAP mappings is through the 557b379d790SJared Hulbert * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit 5587e675137SNick Piggin * set, and the vm_pgoff will point to the first PFN mapped: thus every special 5597e675137SNick Piggin * mapping will always honor the rule 5606aab341eSLinus Torvalds * 5616aab341eSLinus Torvalds * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT) 5626aab341eSLinus Torvalds * 5637e675137SNick Piggin * And for normal mappings this is false. 564b379d790SJared Hulbert * 5657e675137SNick Piggin * This restricts such mappings to be a linear translation from virtual address 5667e675137SNick Piggin * to pfn. To get around this restriction, we allow arbitrary mappings so long 5677e675137SNick Piggin * as the vma is not a COW mapping; in that case, we know that all ptes are 5687e675137SNick Piggin * special (because none can have been COWed). 569b379d790SJared Hulbert * 570b379d790SJared Hulbert * 5717e675137SNick Piggin * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP. 5727e675137SNick Piggin * 573b379d790SJared Hulbert * VM_MIXEDMAP mappings can likewise contain memory with or without "struct 574b379d790SJared Hulbert * page" backing, however the difference is that _all_ pages with a struct 575b379d790SJared Hulbert * page (that is, those where pfn_valid is true) are refcounted and considered 576b379d790SJared Hulbert * normal pages by the VM. The disadvantage is that pages are refcounted 577b379d790SJared Hulbert * (which can be slower and simply not an option for some PFNMAP users). The 578b379d790SJared Hulbert * advantage is that we don't have to follow the strict linearity rule of 579b379d790SJared Hulbert * PFNMAP mappings in order to support COWable mappings. 580b379d790SJared Hulbert * 581ee498ed7SHugh Dickins */ 58225b2995aSChristoph Hellwig struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, 58325b2995aSChristoph Hellwig pte_t pte) 584ee498ed7SHugh Dickins { 58522b31eecSHugh Dickins unsigned long pfn = pte_pfn(pte); 5867e675137SNick Piggin 58700b3a331SLaurent Dufour if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) { 588b38af472SHugh Dickins if (likely(!pte_special(pte))) 58922b31eecSHugh Dickins goto check_pfn; 590667a0a06SDavid Vrabel if (vma->vm_ops && vma->vm_ops->find_special_page) 591667a0a06SDavid Vrabel return vma->vm_ops->find_special_page(vma, addr); 592a13ea5b7SHugh Dickins if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) 593a13ea5b7SHugh Dickins return NULL; 594df6ad698SJérôme Glisse if (is_zero_pfn(pfn)) 595df6ad698SJérôme Glisse return NULL; 596e1fb4a08SDave Jiang if (pte_devmap(pte)) 5973218f871SAlex Sierra /* 5983218f871SAlex Sierra * NOTE: New users of ZONE_DEVICE will not set pte_devmap() 5993218f871SAlex Sierra * and will have refcounts incremented on their struct pages 6003218f871SAlex Sierra * when they are inserted into PTEs, thus they are safe to 6013218f871SAlex Sierra * return here. Legacy ZONE_DEVICE pages that set pte_devmap() 6023218f871SAlex Sierra * do not have refcounts. Example of legacy ZONE_DEVICE is 6033218f871SAlex Sierra * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers. 6043218f871SAlex Sierra */ 605e1fb4a08SDave Jiang return NULL; 606e1fb4a08SDave Jiang 60722b31eecSHugh Dickins print_bad_pte(vma, addr, pte, NULL); 6087e675137SNick Piggin return NULL; 6097e675137SNick Piggin } 6107e675137SNick Piggin 61100b3a331SLaurent Dufour /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */ 6127e675137SNick Piggin 613b379d790SJared Hulbert if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 614b379d790SJared Hulbert if (vma->vm_flags & VM_MIXEDMAP) { 615b379d790SJared Hulbert if (!pfn_valid(pfn)) 616b379d790SJared Hulbert return NULL; 617b379d790SJared Hulbert goto out; 618b379d790SJared Hulbert } else { 6197e675137SNick Piggin unsigned long off; 6207e675137SNick Piggin off = (addr - vma->vm_start) >> PAGE_SHIFT; 6216aab341eSLinus Torvalds if (pfn == vma->vm_pgoff + off) 6226aab341eSLinus Torvalds return NULL; 62367121172SLinus Torvalds if (!is_cow_mapping(vma->vm_flags)) 624fb155c16SLinus Torvalds return NULL; 6256aab341eSLinus Torvalds } 626b379d790SJared Hulbert } 6276aab341eSLinus Torvalds 628b38af472SHugh Dickins if (is_zero_pfn(pfn)) 629b38af472SHugh Dickins return NULL; 63000b3a331SLaurent Dufour 63122b31eecSHugh Dickins check_pfn: 63222b31eecSHugh Dickins if (unlikely(pfn > highest_memmap_pfn)) { 63322b31eecSHugh Dickins print_bad_pte(vma, addr, pte, NULL); 63422b31eecSHugh Dickins return NULL; 63522b31eecSHugh Dickins } 6366aab341eSLinus Torvalds 6376aab341eSLinus Torvalds /* 6387e675137SNick Piggin * NOTE! We still have PageReserved() pages in the page tables. 6397e675137SNick Piggin * eg. VDSO mappings can cause them to exist. 6406aab341eSLinus Torvalds */ 641b379d790SJared Hulbert out: 6426aab341eSLinus Torvalds return pfn_to_page(pfn); 643ee498ed7SHugh Dickins } 644ee498ed7SHugh Dickins 645318e9342SVishal Moola (Oracle) struct folio *vm_normal_folio(struct vm_area_struct *vma, unsigned long addr, 646318e9342SVishal Moola (Oracle) pte_t pte) 647318e9342SVishal Moola (Oracle) { 648318e9342SVishal Moola (Oracle) struct page *page = vm_normal_page(vma, addr, pte); 649318e9342SVishal Moola (Oracle) 650318e9342SVishal Moola (Oracle) if (page) 651318e9342SVishal Moola (Oracle) return page_folio(page); 652318e9342SVishal Moola (Oracle) return NULL; 653318e9342SVishal Moola (Oracle) } 654318e9342SVishal Moola (Oracle) 65528093f9fSGerald Schaefer #ifdef CONFIG_TRANSPARENT_HUGEPAGE 65628093f9fSGerald Schaefer struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr, 65728093f9fSGerald Schaefer pmd_t pmd) 65828093f9fSGerald Schaefer { 65928093f9fSGerald Schaefer unsigned long pfn = pmd_pfn(pmd); 66028093f9fSGerald Schaefer 66128093f9fSGerald Schaefer /* 66228093f9fSGerald Schaefer * There is no pmd_special() but there may be special pmds, e.g. 66328093f9fSGerald Schaefer * in a direct-access (dax) mapping, so let's just replicate the 66400b3a331SLaurent Dufour * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here. 66528093f9fSGerald Schaefer */ 66628093f9fSGerald Schaefer if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { 66728093f9fSGerald Schaefer if (vma->vm_flags & VM_MIXEDMAP) { 66828093f9fSGerald Schaefer if (!pfn_valid(pfn)) 66928093f9fSGerald Schaefer return NULL; 67028093f9fSGerald Schaefer goto out; 67128093f9fSGerald Schaefer } else { 67228093f9fSGerald Schaefer unsigned long off; 67328093f9fSGerald Schaefer off = (addr - vma->vm_start) >> PAGE_SHIFT; 67428093f9fSGerald Schaefer if (pfn == vma->vm_pgoff + off) 67528093f9fSGerald Schaefer return NULL; 67628093f9fSGerald Schaefer if (!is_cow_mapping(vma->vm_flags)) 67728093f9fSGerald Schaefer return NULL; 67828093f9fSGerald Schaefer } 67928093f9fSGerald Schaefer } 68028093f9fSGerald Schaefer 681e1fb4a08SDave Jiang if (pmd_devmap(pmd)) 682e1fb4a08SDave Jiang return NULL; 6833cde287bSYu Zhao if (is_huge_zero_pmd(pmd)) 68428093f9fSGerald Schaefer return NULL; 68528093f9fSGerald Schaefer if (unlikely(pfn > highest_memmap_pfn)) 68628093f9fSGerald Schaefer return NULL; 68728093f9fSGerald Schaefer 68828093f9fSGerald Schaefer /* 68928093f9fSGerald Schaefer * NOTE! We still have PageReserved() pages in the page tables. 69028093f9fSGerald Schaefer * eg. VDSO mappings can cause them to exist. 69128093f9fSGerald Schaefer */ 69228093f9fSGerald Schaefer out: 69328093f9fSGerald Schaefer return pfn_to_page(pfn); 69428093f9fSGerald Schaefer } 69565610453SKefeng Wang 69665610453SKefeng Wang struct folio *vm_normal_folio_pmd(struct vm_area_struct *vma, 69765610453SKefeng Wang unsigned long addr, pmd_t pmd) 69865610453SKefeng Wang { 69965610453SKefeng Wang struct page *page = vm_normal_page_pmd(vma, addr, pmd); 70065610453SKefeng Wang 70165610453SKefeng Wang if (page) 70265610453SKefeng Wang return page_folio(page); 70365610453SKefeng Wang return NULL; 70465610453SKefeng Wang } 70528093f9fSGerald Schaefer #endif 70628093f9fSGerald Schaefer 707b756a3b5SAlistair Popple static void restore_exclusive_pte(struct vm_area_struct *vma, 708b756a3b5SAlistair Popple struct page *page, unsigned long address, 709b756a3b5SAlistair Popple pte_t *ptep) 710b756a3b5SAlistair Popple { 711b832a354SDavid Hildenbrand struct folio *folio = page_folio(page); 712c33c7948SRyan Roberts pte_t orig_pte; 713b756a3b5SAlistair Popple pte_t pte; 714b756a3b5SAlistair Popple swp_entry_t entry; 715b756a3b5SAlistair Popple 716c33c7948SRyan Roberts orig_pte = ptep_get(ptep); 717b756a3b5SAlistair Popple pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot))); 718c33c7948SRyan Roberts if (pte_swp_soft_dirty(orig_pte)) 719b756a3b5SAlistair Popple pte = pte_mksoft_dirty(pte); 720b756a3b5SAlistair Popple 721c33c7948SRyan Roberts entry = pte_to_swp_entry(orig_pte); 722c33c7948SRyan Roberts if (pte_swp_uffd_wp(orig_pte)) 723b756a3b5SAlistair Popple pte = pte_mkuffd_wp(pte); 724b756a3b5SAlistair Popple else if (is_writable_device_exclusive_entry(entry)) 725b756a3b5SAlistair Popple pte = maybe_mkwrite(pte_mkdirty(pte), vma); 726b756a3b5SAlistair Popple 727b832a354SDavid Hildenbrand VM_BUG_ON_FOLIO(pte_write(pte) && (!folio_test_anon(folio) && 728b832a354SDavid Hildenbrand PageAnonExclusive(page)), folio); 7296c287605SDavid Hildenbrand 730b756a3b5SAlistair Popple /* 731b756a3b5SAlistair Popple * No need to take a page reference as one was already 732b756a3b5SAlistair Popple * created when the swap entry was made. 733b756a3b5SAlistair Popple */ 734b832a354SDavid Hildenbrand if (folio_test_anon(folio)) 735b832a354SDavid Hildenbrand folio_add_anon_rmap_pte(folio, page, vma, address, RMAP_NONE); 736b756a3b5SAlistair Popple else 737b756a3b5SAlistair Popple /* 738b756a3b5SAlistair Popple * Currently device exclusive access only supports anonymous 739b756a3b5SAlistair Popple * memory so the entry shouldn't point to a filebacked page. 740b756a3b5SAlistair Popple */ 7414d8ff640SMiaohe Lin WARN_ON_ONCE(1); 742b756a3b5SAlistair Popple 7431eba86c0SPasha Tatashin set_pte_at(vma->vm_mm, address, ptep, pte); 7441eba86c0SPasha Tatashin 745b756a3b5SAlistair Popple /* 746b756a3b5SAlistair Popple * No need to invalidate - it was non-present before. However 747b756a3b5SAlistair Popple * secondary CPUs may have mappings that need invalidating. 748b756a3b5SAlistair Popple */ 749b756a3b5SAlistair Popple update_mmu_cache(vma, address, ptep); 750b756a3b5SAlistair Popple } 751b756a3b5SAlistair Popple 752b756a3b5SAlistair Popple /* 753b756a3b5SAlistair Popple * Tries to restore an exclusive pte if the page lock can be acquired without 754b756a3b5SAlistair Popple * sleeping. 755b756a3b5SAlistair Popple */ 756b756a3b5SAlistair Popple static int 757b756a3b5SAlistair Popple try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma, 758b756a3b5SAlistair Popple unsigned long addr) 759b756a3b5SAlistair Popple { 760c33c7948SRyan Roberts swp_entry_t entry = pte_to_swp_entry(ptep_get(src_pte)); 761b756a3b5SAlistair Popple struct page *page = pfn_swap_entry_to_page(entry); 762b756a3b5SAlistair Popple 763b756a3b5SAlistair Popple if (trylock_page(page)) { 764b756a3b5SAlistair Popple restore_exclusive_pte(vma, page, addr, src_pte); 765b756a3b5SAlistair Popple unlock_page(page); 766b756a3b5SAlistair Popple return 0; 767b756a3b5SAlistair Popple } 768b756a3b5SAlistair Popple 769b756a3b5SAlistair Popple return -EBUSY; 770b756a3b5SAlistair Popple } 771b756a3b5SAlistair Popple 772ee498ed7SHugh Dickins /* 7731da177e4SLinus Torvalds * copy one vm_area from one task to the other. Assumes the page tables 7741da177e4SLinus Torvalds * already present in the new task to be cleared in the whole range 7751da177e4SLinus Torvalds * covered by this vma. 7761da177e4SLinus Torvalds */ 7771da177e4SLinus Torvalds 778df3a57d1SLinus Torvalds static unsigned long 779df3a57d1SLinus Torvalds copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, 7808f34f1eaSPeter Xu pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma, 7818f34f1eaSPeter Xu struct vm_area_struct *src_vma, unsigned long addr, int *rss) 7821da177e4SLinus Torvalds { 7838f34f1eaSPeter Xu unsigned long vm_flags = dst_vma->vm_flags; 784c33c7948SRyan Roberts pte_t orig_pte = ptep_get(src_pte); 785c33c7948SRyan Roberts pte_t pte = orig_pte; 78608e7795eSDavid Hildenbrand struct folio *folio; 7871da177e4SLinus Torvalds struct page *page; 788c33c7948SRyan Roberts swp_entry_t entry = pte_to_swp_entry(orig_pte); 7890697212aSChristoph Lameter 7902022b4d1SHugh Dickins if (likely(!non_swap_entry(entry))) { 791570a335bSHugh Dickins if (swap_duplicate(entry) < 0) 7929a5cc85cSAlistair Popple return -EIO; 793570a335bSHugh Dickins 7941da177e4SLinus Torvalds /* make sure dst_mm is on swapoff's mmlist. */ 7951da177e4SLinus Torvalds if (unlikely(list_empty(&dst_mm->mmlist))) { 7961da177e4SLinus Torvalds spin_lock(&mmlist_lock); 797f412ac08SHugh Dickins if (list_empty(&dst_mm->mmlist)) 798f412ac08SHugh Dickins list_add(&dst_mm->mmlist, 799f412ac08SHugh Dickins &src_mm->mmlist); 8001da177e4SLinus Torvalds spin_unlock(&mmlist_lock); 8011da177e4SLinus Torvalds } 8021493a191SDavid Hildenbrand /* Mark the swap entry as shared. */ 803c33c7948SRyan Roberts if (pte_swp_exclusive(orig_pte)) { 804c33c7948SRyan Roberts pte = pte_swp_clear_exclusive(orig_pte); 8051493a191SDavid Hildenbrand set_pte_at(src_mm, addr, src_pte, pte); 8061493a191SDavid Hildenbrand } 807b084d435SKAMEZAWA Hiroyuki rss[MM_SWAPENTS]++; 8082022b4d1SHugh Dickins } else if (is_migration_entry(entry)) { 809530c2a0dSKefeng Wang folio = pfn_swap_entry_folio(entry); 8109f9f1acdSKonstantin Khlebnikov 811a23f517bSKefeng Wang rss[mm_counter(folio)]++; 8129f9f1acdSKonstantin Khlebnikov 8136c287605SDavid Hildenbrand if (!is_readable_migration_entry(entry) && 8140697212aSChristoph Lameter is_cow_mapping(vm_flags)) { 8150697212aSChristoph Lameter /* 8166c287605SDavid Hildenbrand * COW mappings require pages in both parent and child 8176c287605SDavid Hildenbrand * to be set to read. A previously exclusive entry is 8186c287605SDavid Hildenbrand * now shared. 8190697212aSChristoph Lameter */ 8204dd845b5SAlistair Popple entry = make_readable_migration_entry( 8214dd845b5SAlistair Popple swp_offset(entry)); 8220697212aSChristoph Lameter pte = swp_entry_to_pte(entry); 823c33c7948SRyan Roberts if (pte_swp_soft_dirty(orig_pte)) 824c3d16e16SCyrill Gorcunov pte = pte_swp_mksoft_dirty(pte); 825c33c7948SRyan Roberts if (pte_swp_uffd_wp(orig_pte)) 826f45ec5ffSPeter Xu pte = pte_swp_mkuffd_wp(pte); 8270697212aSChristoph Lameter set_pte_at(src_mm, addr, src_pte, pte); 8280697212aSChristoph Lameter } 8295042db43SJérôme Glisse } else if (is_device_private_entry(entry)) { 830af5cdaf8SAlistair Popple page = pfn_swap_entry_to_page(entry); 83108e7795eSDavid Hildenbrand folio = page_folio(page); 8325042db43SJérôme Glisse 8335042db43SJérôme Glisse /* 8345042db43SJérôme Glisse * Update rss count even for unaddressable pages, as 8355042db43SJérôme Glisse * they should treated just like normal pages in this 8365042db43SJérôme Glisse * respect. 8375042db43SJérôme Glisse * 8385042db43SJérôme Glisse * We will likely want to have some new rss counters 8395042db43SJérôme Glisse * for unaddressable pages, at some point. But for now 8405042db43SJérôme Glisse * keep things as they are. 8415042db43SJérôme Glisse */ 84208e7795eSDavid Hildenbrand folio_get(folio); 843a23f517bSKefeng Wang rss[mm_counter(folio)]++; 844fb3d824dSDavid Hildenbrand /* Cannot fail as these pages cannot get pinned. */ 84508e7795eSDavid Hildenbrand folio_try_dup_anon_rmap_pte(folio, page, src_vma); 8465042db43SJérôme Glisse 8475042db43SJérôme Glisse /* 8485042db43SJérôme Glisse * We do not preserve soft-dirty information, because so 8495042db43SJérôme Glisse * far, checkpoint/restore is the only feature that 8505042db43SJérôme Glisse * requires that. And checkpoint/restore does not work 8515042db43SJérôme Glisse * when a device driver is involved (you cannot easily 8525042db43SJérôme Glisse * save and restore device driver state). 8535042db43SJérôme Glisse */ 8544dd845b5SAlistair Popple if (is_writable_device_private_entry(entry) && 8555042db43SJérôme Glisse is_cow_mapping(vm_flags)) { 8564dd845b5SAlistair Popple entry = make_readable_device_private_entry( 8574dd845b5SAlistair Popple swp_offset(entry)); 8585042db43SJérôme Glisse pte = swp_entry_to_pte(entry); 859c33c7948SRyan Roberts if (pte_swp_uffd_wp(orig_pte)) 860f45ec5ffSPeter Xu pte = pte_swp_mkuffd_wp(pte); 8615042db43SJérôme Glisse set_pte_at(src_mm, addr, src_pte, pte); 8625042db43SJérôme Glisse } 863b756a3b5SAlistair Popple } else if (is_device_exclusive_entry(entry)) { 864b756a3b5SAlistair Popple /* 865b756a3b5SAlistair Popple * Make device exclusive entries present by restoring the 866b756a3b5SAlistair Popple * original entry then copying as for a present pte. Device 867b756a3b5SAlistair Popple * exclusive entries currently only support private writable 868b756a3b5SAlistair Popple * (ie. COW) mappings. 869b756a3b5SAlistair Popple */ 870b756a3b5SAlistair Popple VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags)); 871b756a3b5SAlistair Popple if (try_restore_exclusive_pte(src_pte, src_vma, addr)) 872b756a3b5SAlistair Popple return -EBUSY; 873b756a3b5SAlistair Popple return -ENOENT; 874c56d1b62SPeter Xu } else if (is_pte_marker_entry(entry)) { 875af19487fSAxel Rasmussen pte_marker marker = copy_pte_marker(entry, dst_vma); 876af19487fSAxel Rasmussen 877af19487fSAxel Rasmussen if (marker) 878af19487fSAxel Rasmussen set_pte_at(dst_mm, addr, dst_pte, 879af19487fSAxel Rasmussen make_pte_marker(marker)); 880c56d1b62SPeter Xu return 0; 8811da177e4SLinus Torvalds } 8828f34f1eaSPeter Xu if (!userfaultfd_wp(dst_vma)) 8838f34f1eaSPeter Xu pte = pte_swp_clear_uffd_wp(pte); 884df3a57d1SLinus Torvalds set_pte_at(dst_mm, addr, dst_pte, pte); 885df3a57d1SLinus Torvalds return 0; 8861da177e4SLinus Torvalds } 8871da177e4SLinus Torvalds 88870e806e4SPeter Xu /* 889b51ad4f8SDavid Hildenbrand * Copy a present and normal page. 89070e806e4SPeter Xu * 891b51ad4f8SDavid Hildenbrand * NOTE! The usual case is that this isn't required; 892b51ad4f8SDavid Hildenbrand * instead, the caller can just increase the page refcount 893b51ad4f8SDavid Hildenbrand * and re-use the pte the traditional way. 89470e806e4SPeter Xu * 89570e806e4SPeter Xu * And if we need a pre-allocated page but don't yet have 89670e806e4SPeter Xu * one, return a negative error to let the preallocation 89770e806e4SPeter Xu * code know so that it can do so outside the page table 89870e806e4SPeter Xu * lock. 89970e806e4SPeter Xu */ 90070e806e4SPeter Xu static inline int 901c78f4636SPeter Xu copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 902c78f4636SPeter Xu pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss, 903edf50470SMatthew Wilcox (Oracle) struct folio **prealloc, struct page *page) 90470e806e4SPeter Xu { 905edf50470SMatthew Wilcox (Oracle) struct folio *new_folio; 906b51ad4f8SDavid Hildenbrand pte_t pte; 90770e806e4SPeter Xu 908edf50470SMatthew Wilcox (Oracle) new_folio = *prealloc; 909edf50470SMatthew Wilcox (Oracle) if (!new_folio) 91070e806e4SPeter Xu return -EAGAIN; 91170e806e4SPeter Xu 91270e806e4SPeter Xu /* 91370e806e4SPeter Xu * We have a prealloc page, all good! Take it 91470e806e4SPeter Xu * over and copy the page & arm it. 91570e806e4SPeter Xu */ 91670e806e4SPeter Xu *prealloc = NULL; 917edf50470SMatthew Wilcox (Oracle) copy_user_highpage(&new_folio->page, page, addr, src_vma); 918edf50470SMatthew Wilcox (Oracle) __folio_mark_uptodate(new_folio); 919edf50470SMatthew Wilcox (Oracle) folio_add_new_anon_rmap(new_folio, dst_vma, addr); 920edf50470SMatthew Wilcox (Oracle) folio_add_lru_vma(new_folio, dst_vma); 921edf50470SMatthew Wilcox (Oracle) rss[MM_ANONPAGES]++; 92270e806e4SPeter Xu 92370e806e4SPeter Xu /* All done, just insert the new page copy in the child */ 924edf50470SMatthew Wilcox (Oracle) pte = mk_pte(&new_folio->page, dst_vma->vm_page_prot); 925c78f4636SPeter Xu pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma); 926c33c7948SRyan Roberts if (userfaultfd_pte_wp(dst_vma, ptep_get(src_pte))) 9278f34f1eaSPeter Xu /* Uffd-wp needs to be delivered to dest pte as well */ 928f1eb1bacSPeter Xu pte = pte_mkuffd_wp(pte); 929c78f4636SPeter Xu set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte); 93070e806e4SPeter Xu return 0; 93170e806e4SPeter Xu } 93270e806e4SPeter Xu 933*f8d93776SDavid Hildenbrand static __always_inline void __copy_present_ptes(struct vm_area_struct *dst_vma, 93423ed1908SDavid Hildenbrand struct vm_area_struct *src_vma, pte_t *dst_pte, pte_t *src_pte, 935*f8d93776SDavid Hildenbrand pte_t pte, unsigned long addr, int nr) 93623ed1908SDavid Hildenbrand { 93723ed1908SDavid Hildenbrand struct mm_struct *src_mm = src_vma->vm_mm; 93823ed1908SDavid Hildenbrand 93923ed1908SDavid Hildenbrand /* If it's a COW mapping, write protect it both processes. */ 94023ed1908SDavid Hildenbrand if (is_cow_mapping(src_vma->vm_flags) && pte_write(pte)) { 941*f8d93776SDavid Hildenbrand wrprotect_ptes(src_mm, addr, src_pte, nr); 94223ed1908SDavid Hildenbrand pte = pte_wrprotect(pte); 94323ed1908SDavid Hildenbrand } 94423ed1908SDavid Hildenbrand 94523ed1908SDavid Hildenbrand /* If it's a shared mapping, mark it clean in the child. */ 94623ed1908SDavid Hildenbrand if (src_vma->vm_flags & VM_SHARED) 94723ed1908SDavid Hildenbrand pte = pte_mkclean(pte); 94823ed1908SDavid Hildenbrand pte = pte_mkold(pte); 94923ed1908SDavid Hildenbrand 95023ed1908SDavid Hildenbrand if (!userfaultfd_wp(dst_vma)) 95123ed1908SDavid Hildenbrand pte = pte_clear_uffd_wp(pte); 95223ed1908SDavid Hildenbrand 953*f8d93776SDavid Hildenbrand set_ptes(dst_vma->vm_mm, addr, dst_pte, pte, nr); 95423ed1908SDavid Hildenbrand } 95523ed1908SDavid Hildenbrand 95670e806e4SPeter Xu /* 957*f8d93776SDavid Hildenbrand * Detect a PTE batch: consecutive (present) PTEs that map consecutive 958*f8d93776SDavid Hildenbrand * pages of the same folio. 959*f8d93776SDavid Hildenbrand * 960*f8d93776SDavid Hildenbrand * All PTEs inside a PTE batch have the same PTE bits set, excluding the PFN. 961*f8d93776SDavid Hildenbrand */ 962*f8d93776SDavid Hildenbrand static inline int folio_pte_batch(struct folio *folio, unsigned long addr, 963*f8d93776SDavid Hildenbrand pte_t *start_ptep, pte_t pte, int max_nr) 964*f8d93776SDavid Hildenbrand { 965*f8d93776SDavid Hildenbrand unsigned long folio_end_pfn = folio_pfn(folio) + folio_nr_pages(folio); 966*f8d93776SDavid Hildenbrand const pte_t *end_ptep = start_ptep + max_nr; 967*f8d93776SDavid Hildenbrand pte_t expected_pte = pte_next_pfn(pte); 968*f8d93776SDavid Hildenbrand pte_t *ptep = start_ptep + 1; 969*f8d93776SDavid Hildenbrand 970*f8d93776SDavid Hildenbrand VM_WARN_ON_FOLIO(!pte_present(pte), folio); 971*f8d93776SDavid Hildenbrand 972*f8d93776SDavid Hildenbrand while (ptep != end_ptep) { 973*f8d93776SDavid Hildenbrand pte = ptep_get(ptep); 974*f8d93776SDavid Hildenbrand 975*f8d93776SDavid Hildenbrand if (!pte_same(pte, expected_pte)) 976*f8d93776SDavid Hildenbrand break; 977*f8d93776SDavid Hildenbrand 978*f8d93776SDavid Hildenbrand /* 979*f8d93776SDavid Hildenbrand * Stop immediately once we reached the end of the folio. In 980*f8d93776SDavid Hildenbrand * corner cases the next PFN might fall into a different 981*f8d93776SDavid Hildenbrand * folio. 982*f8d93776SDavid Hildenbrand */ 983*f8d93776SDavid Hildenbrand if (pte_pfn(pte) == folio_end_pfn) 984*f8d93776SDavid Hildenbrand break; 985*f8d93776SDavid Hildenbrand 986*f8d93776SDavid Hildenbrand expected_pte = pte_next_pfn(expected_pte); 987*f8d93776SDavid Hildenbrand ptep++; 988*f8d93776SDavid Hildenbrand } 989*f8d93776SDavid Hildenbrand 990*f8d93776SDavid Hildenbrand return ptep - start_ptep; 991*f8d93776SDavid Hildenbrand } 992*f8d93776SDavid Hildenbrand 993*f8d93776SDavid Hildenbrand /* 994*f8d93776SDavid Hildenbrand * Copy one present PTE, trying to batch-process subsequent PTEs that map 995*f8d93776SDavid Hildenbrand * consecutive pages of the same folio by copying them as well. 996*f8d93776SDavid Hildenbrand * 997*f8d93776SDavid Hildenbrand * Returns -EAGAIN if one preallocated page is required to copy the next PTE. 998*f8d93776SDavid Hildenbrand * Otherwise, returns the number of copied PTEs (at least 1). 99970e806e4SPeter Xu */ 100070e806e4SPeter Xu static inline int 1001*f8d93776SDavid Hildenbrand copy_present_ptes(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 100253723298SDavid Hildenbrand pte_t *dst_pte, pte_t *src_pte, pte_t pte, unsigned long addr, 1003*f8d93776SDavid Hildenbrand int max_nr, int *rss, struct folio **prealloc) 1004df3a57d1SLinus Torvalds { 1005df3a57d1SLinus Torvalds struct page *page; 100614ddee41SMatthew Wilcox (Oracle) struct folio *folio; 1007*f8d93776SDavid Hildenbrand int err, nr; 1008df3a57d1SLinus Torvalds 1009c78f4636SPeter Xu page = vm_normal_page(src_vma, addr, pte); 101023ed1908SDavid Hildenbrand if (unlikely(!page)) 101123ed1908SDavid Hildenbrand goto copy_pte; 101223ed1908SDavid Hildenbrand 101314ddee41SMatthew Wilcox (Oracle) folio = page_folio(page); 1014*f8d93776SDavid Hildenbrand 1015*f8d93776SDavid Hildenbrand /* 1016*f8d93776SDavid Hildenbrand * If we likely have to copy, just don't bother with batching. Make 1017*f8d93776SDavid Hildenbrand * sure that the common "small folio" case is as fast as possible 1018*f8d93776SDavid Hildenbrand * by keeping the batching logic separate. 1019*f8d93776SDavid Hildenbrand */ 1020*f8d93776SDavid Hildenbrand if (unlikely(!*prealloc && folio_test_large(folio) && max_nr != 1)) { 1021*f8d93776SDavid Hildenbrand nr = folio_pte_batch(folio, addr, src_pte, pte, max_nr); 1022*f8d93776SDavid Hildenbrand folio_ref_add(folio, nr); 1023*f8d93776SDavid Hildenbrand if (folio_test_anon(folio)) { 1024*f8d93776SDavid Hildenbrand if (unlikely(folio_try_dup_anon_rmap_ptes(folio, page, 1025*f8d93776SDavid Hildenbrand nr, src_vma))) { 1026*f8d93776SDavid Hildenbrand folio_ref_sub(folio, nr); 1027*f8d93776SDavid Hildenbrand return -EAGAIN; 1028*f8d93776SDavid Hildenbrand } 1029*f8d93776SDavid Hildenbrand rss[MM_ANONPAGES] += nr; 1030*f8d93776SDavid Hildenbrand VM_WARN_ON_FOLIO(PageAnonExclusive(page), folio); 1031*f8d93776SDavid Hildenbrand } else { 1032*f8d93776SDavid Hildenbrand folio_dup_file_rmap_ptes(folio, page, nr); 1033*f8d93776SDavid Hildenbrand rss[mm_counter_file(folio)] += nr; 1034*f8d93776SDavid Hildenbrand } 1035*f8d93776SDavid Hildenbrand __copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte, pte, 1036*f8d93776SDavid Hildenbrand addr, nr); 1037*f8d93776SDavid Hildenbrand return nr; 1038*f8d93776SDavid Hildenbrand } 1039*f8d93776SDavid Hildenbrand 104023ed1908SDavid Hildenbrand folio_get(folio); 104123ed1908SDavid Hildenbrand if (folio_test_anon(folio)) { 1042b51ad4f8SDavid Hildenbrand /* 1043b51ad4f8SDavid Hildenbrand * If this page may have been pinned by the parent process, 1044b51ad4f8SDavid Hildenbrand * copy the page immediately for the child so that we'll always 1045b51ad4f8SDavid Hildenbrand * guarantee the pinned page won't be randomly replaced in the 1046b51ad4f8SDavid Hildenbrand * future. 1047b51ad4f8SDavid Hildenbrand */ 104808e7795eSDavid Hildenbrand if (unlikely(folio_try_dup_anon_rmap_pte(folio, page, src_vma))) { 1049fb3d824dSDavid Hildenbrand /* Page may be pinned, we have to copy. */ 105014ddee41SMatthew Wilcox (Oracle) folio_put(folio); 1051*f8d93776SDavid Hildenbrand err = copy_present_page(dst_vma, src_vma, dst_pte, src_pte, 1052b51ad4f8SDavid Hildenbrand addr, rss, prealloc, page); 1053*f8d93776SDavid Hildenbrand return err ? err : 1; 1054fb3d824dSDavid Hildenbrand } 1055edf50470SMatthew Wilcox (Oracle) rss[MM_ANONPAGES]++; 105623ed1908SDavid Hildenbrand VM_WARN_ON_FOLIO(PageAnonExclusive(page), folio); 105723ed1908SDavid Hildenbrand } else { 1058d8ef5e31SDavid Hildenbrand folio_dup_file_rmap_pte(folio, page); 10596b27cc6cSKefeng Wang rss[mm_counter_file(folio)]++; 106070e806e4SPeter Xu } 106170e806e4SPeter Xu 106223ed1908SDavid Hildenbrand copy_pte: 1063*f8d93776SDavid Hildenbrand __copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte, pte, addr, 1); 1064*f8d93776SDavid Hildenbrand return 1; 10656aab341eSLinus Torvalds } 1066ae859762SHugh Dickins 1067294de6d8SKefeng Wang static inline struct folio *folio_prealloc(struct mm_struct *src_mm, 1068294de6d8SKefeng Wang struct vm_area_struct *vma, unsigned long addr, bool need_zero) 106970e806e4SPeter Xu { 1070edf50470SMatthew Wilcox (Oracle) struct folio *new_folio; 107170e806e4SPeter Xu 1072294de6d8SKefeng Wang if (need_zero) 1073294de6d8SKefeng Wang new_folio = vma_alloc_zeroed_movable_folio(vma, addr); 1074294de6d8SKefeng Wang else 1075294de6d8SKefeng Wang new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma, 1076294de6d8SKefeng Wang addr, false); 1077294de6d8SKefeng Wang 1078edf50470SMatthew Wilcox (Oracle) if (!new_folio) 107970e806e4SPeter Xu return NULL; 108070e806e4SPeter Xu 1081edf50470SMatthew Wilcox (Oracle) if (mem_cgroup_charge(new_folio, src_mm, GFP_KERNEL)) { 1082edf50470SMatthew Wilcox (Oracle) folio_put(new_folio); 108370e806e4SPeter Xu return NULL; 108470e806e4SPeter Xu } 1085e601ded4SKefeng Wang folio_throttle_swaprate(new_folio, GFP_KERNEL); 108670e806e4SPeter Xu 1087edf50470SMatthew Wilcox (Oracle) return new_folio; 10881da177e4SLinus Torvalds } 10891da177e4SLinus Torvalds 1090c78f4636SPeter Xu static int 1091c78f4636SPeter Xu copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1092c78f4636SPeter Xu pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, 1093c78f4636SPeter Xu unsigned long end) 10941da177e4SLinus Torvalds { 1095c78f4636SPeter Xu struct mm_struct *dst_mm = dst_vma->vm_mm; 1096c78f4636SPeter Xu struct mm_struct *src_mm = src_vma->vm_mm; 1097c36987e2SDaisuke Nishimura pte_t *orig_src_pte, *orig_dst_pte; 10981da177e4SLinus Torvalds pte_t *src_pte, *dst_pte; 1099c33c7948SRyan Roberts pte_t ptent; 1100c74df32cSHugh Dickins spinlock_t *src_ptl, *dst_ptl; 1101*f8d93776SDavid Hildenbrand int progress, max_nr, ret = 0; 1102d559db08SKAMEZAWA Hiroyuki int rss[NR_MM_COUNTERS]; 1103570a335bSHugh Dickins swp_entry_t entry = (swp_entry_t){0}; 1104edf50470SMatthew Wilcox (Oracle) struct folio *prealloc = NULL; 1105*f8d93776SDavid Hildenbrand int nr; 11061da177e4SLinus Torvalds 11071da177e4SLinus Torvalds again: 110870e806e4SPeter Xu progress = 0; 1109d559db08SKAMEZAWA Hiroyuki init_rss_vec(rss); 1110d559db08SKAMEZAWA Hiroyuki 11113db82b93SHugh Dickins /* 11123db82b93SHugh Dickins * copy_pmd_range()'s prior pmd_none_or_clear_bad(src_pmd), and the 11133db82b93SHugh Dickins * error handling here, assume that exclusive mmap_lock on dst and src 11143db82b93SHugh Dickins * protects anon from unexpected THP transitions; with shmem and file 11153db82b93SHugh Dickins * protected by mmap_lock-less collapse skipping areas with anon_vma 11163db82b93SHugh Dickins * (whereas vma_needs_copy() skips areas without anon_vma). A rework 11173db82b93SHugh Dickins * can remove such assumptions later, but this is good enough for now. 11183db82b93SHugh Dickins */ 1119c74df32cSHugh Dickins dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl); 112070e806e4SPeter Xu if (!dst_pte) { 112170e806e4SPeter Xu ret = -ENOMEM; 112270e806e4SPeter Xu goto out; 112370e806e4SPeter Xu } 11243db82b93SHugh Dickins src_pte = pte_offset_map_nolock(src_mm, src_pmd, addr, &src_ptl); 11253db82b93SHugh Dickins if (!src_pte) { 11263db82b93SHugh Dickins pte_unmap_unlock(dst_pte, dst_ptl); 11273db82b93SHugh Dickins /* ret == 0 */ 11283db82b93SHugh Dickins goto out; 11293db82b93SHugh Dickins } 1130f20dc5f7SIngo Molnar spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 1131c36987e2SDaisuke Nishimura orig_src_pte = src_pte; 1132c36987e2SDaisuke Nishimura orig_dst_pte = dst_pte; 11336606c3e0SZachary Amsden arch_enter_lazy_mmu_mode(); 11341da177e4SLinus Torvalds 11351da177e4SLinus Torvalds do { 1136*f8d93776SDavid Hildenbrand nr = 1; 1137*f8d93776SDavid Hildenbrand 11381da177e4SLinus Torvalds /* 11391da177e4SLinus Torvalds * We are holding two locks at this point - either of them 11401da177e4SLinus Torvalds * could generate latencies in another task on another CPU. 11411da177e4SLinus Torvalds */ 1142e040f218SHugh Dickins if (progress >= 32) { 1143e040f218SHugh Dickins progress = 0; 1144e040f218SHugh Dickins if (need_resched() || 114595c354feSNick Piggin spin_needbreak(src_ptl) || spin_needbreak(dst_ptl)) 11461da177e4SLinus Torvalds break; 1147e040f218SHugh Dickins } 1148c33c7948SRyan Roberts ptent = ptep_get(src_pte); 1149c33c7948SRyan Roberts if (pte_none(ptent)) { 11501da177e4SLinus Torvalds progress++; 11511da177e4SLinus Torvalds continue; 11521da177e4SLinus Torvalds } 1153c33c7948SRyan Roberts if (unlikely(!pte_present(ptent))) { 11549a5cc85cSAlistair Popple ret = copy_nonpresent_pte(dst_mm, src_mm, 115579a1971cSLinus Torvalds dst_pte, src_pte, 11568f34f1eaSPeter Xu dst_vma, src_vma, 11578f34f1eaSPeter Xu addr, rss); 11589a5cc85cSAlistair Popple if (ret == -EIO) { 1159c33c7948SRyan Roberts entry = pte_to_swp_entry(ptep_get(src_pte)); 1160570a335bSHugh Dickins break; 1161b756a3b5SAlistair Popple } else if (ret == -EBUSY) { 1162b756a3b5SAlistair Popple break; 1163b756a3b5SAlistair Popple } else if (!ret) { 11641da177e4SLinus Torvalds progress += 8; 116579a1971cSLinus Torvalds continue; 116679a1971cSLinus Torvalds } 116753723298SDavid Hildenbrand ptent = ptep_get(src_pte); 116853723298SDavid Hildenbrand VM_WARN_ON_ONCE(!pte_present(ptent)); 1169b756a3b5SAlistair Popple 1170b756a3b5SAlistair Popple /* 1171b756a3b5SAlistair Popple * Device exclusive entry restored, continue by copying 1172b756a3b5SAlistair Popple * the now present pte. 1173b756a3b5SAlistair Popple */ 1174b756a3b5SAlistair Popple WARN_ON_ONCE(ret != -ENOENT); 1175b756a3b5SAlistair Popple } 1176*f8d93776SDavid Hildenbrand /* copy_present_ptes() will clear `*prealloc' if consumed */ 1177*f8d93776SDavid Hildenbrand max_nr = (end - addr) / PAGE_SIZE; 1178*f8d93776SDavid Hildenbrand ret = copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte, 1179*f8d93776SDavid Hildenbrand ptent, addr, max_nr, rss, &prealloc); 118070e806e4SPeter Xu /* 118170e806e4SPeter Xu * If we need a pre-allocated page for this pte, drop the 118270e806e4SPeter Xu * locks, allocate, and try again. 118370e806e4SPeter Xu */ 118470e806e4SPeter Xu if (unlikely(ret == -EAGAIN)) 118570e806e4SPeter Xu break; 118670e806e4SPeter Xu if (unlikely(prealloc)) { 118770e806e4SPeter Xu /* 118870e806e4SPeter Xu * pre-alloc page cannot be reused by next time so as 118970e806e4SPeter Xu * to strictly follow mempolicy (e.g., alloc_page_vma() 119070e806e4SPeter Xu * will allocate page according to address). This 119170e806e4SPeter Xu * could only happen if one pinned pte changed. 119270e806e4SPeter Xu */ 1193edf50470SMatthew Wilcox (Oracle) folio_put(prealloc); 119470e806e4SPeter Xu prealloc = NULL; 119570e806e4SPeter Xu } 1196*f8d93776SDavid Hildenbrand nr = ret; 1197*f8d93776SDavid Hildenbrand progress += 8 * nr; 1198*f8d93776SDavid Hildenbrand } while (dst_pte += nr, src_pte += nr, addr += PAGE_SIZE * nr, 1199*f8d93776SDavid Hildenbrand addr != end); 12001da177e4SLinus Torvalds 12016606c3e0SZachary Amsden arch_leave_lazy_mmu_mode(); 12023db82b93SHugh Dickins pte_unmap_unlock(orig_src_pte, src_ptl); 1203d559db08SKAMEZAWA Hiroyuki add_mm_rss_vec(dst_mm, rss); 1204c36987e2SDaisuke Nishimura pte_unmap_unlock(orig_dst_pte, dst_ptl); 1205c74df32cSHugh Dickins cond_resched(); 1206570a335bSHugh Dickins 12079a5cc85cSAlistair Popple if (ret == -EIO) { 12089a5cc85cSAlistair Popple VM_WARN_ON_ONCE(!entry.val); 120970e806e4SPeter Xu if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) { 121070e806e4SPeter Xu ret = -ENOMEM; 121170e806e4SPeter Xu goto out; 121270e806e4SPeter Xu } 121370e806e4SPeter Xu entry.val = 0; 1214b756a3b5SAlistair Popple } else if (ret == -EBUSY) { 1215b756a3b5SAlistair Popple goto out; 12169a5cc85cSAlistair Popple } else if (ret == -EAGAIN) { 1217294de6d8SKefeng Wang prealloc = folio_prealloc(src_mm, src_vma, addr, false); 121870e806e4SPeter Xu if (!prealloc) 1219570a335bSHugh Dickins return -ENOMEM; 1220*f8d93776SDavid Hildenbrand } else if (ret < 0) { 12219a5cc85cSAlistair Popple VM_WARN_ON_ONCE(1); 12229a5cc85cSAlistair Popple } 12239a5cc85cSAlistair Popple 122470e806e4SPeter Xu /* We've captured and resolved the error. Reset, try again. */ 122570e806e4SPeter Xu ret = 0; 12269a5cc85cSAlistair Popple 12271da177e4SLinus Torvalds if (addr != end) 12281da177e4SLinus Torvalds goto again; 122970e806e4SPeter Xu out: 123070e806e4SPeter Xu if (unlikely(prealloc)) 1231edf50470SMatthew Wilcox (Oracle) folio_put(prealloc); 123270e806e4SPeter Xu return ret; 12331da177e4SLinus Torvalds } 12341da177e4SLinus Torvalds 1235c78f4636SPeter Xu static inline int 1236c78f4636SPeter Xu copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1237c78f4636SPeter Xu pud_t *dst_pud, pud_t *src_pud, unsigned long addr, 1238c78f4636SPeter Xu unsigned long end) 12391da177e4SLinus Torvalds { 1240c78f4636SPeter Xu struct mm_struct *dst_mm = dst_vma->vm_mm; 1241c78f4636SPeter Xu struct mm_struct *src_mm = src_vma->vm_mm; 12421da177e4SLinus Torvalds pmd_t *src_pmd, *dst_pmd; 12431da177e4SLinus Torvalds unsigned long next; 12441da177e4SLinus Torvalds 12451da177e4SLinus Torvalds dst_pmd = pmd_alloc(dst_mm, dst_pud, addr); 12461da177e4SLinus Torvalds if (!dst_pmd) 12471da177e4SLinus Torvalds return -ENOMEM; 12481da177e4SLinus Torvalds src_pmd = pmd_offset(src_pud, addr); 12491da177e4SLinus Torvalds do { 12501da177e4SLinus Torvalds next = pmd_addr_end(addr, end); 125184c3fc4eSZi Yan if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd) 125284c3fc4eSZi Yan || pmd_devmap(*src_pmd)) { 125371e3aac0SAndrea Arcangeli int err; 1254c78f4636SPeter Xu VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma); 12558f34f1eaSPeter Xu err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd, 12568f34f1eaSPeter Xu addr, dst_vma, src_vma); 125771e3aac0SAndrea Arcangeli if (err == -ENOMEM) 125871e3aac0SAndrea Arcangeli return -ENOMEM; 125971e3aac0SAndrea Arcangeli if (!err) 126071e3aac0SAndrea Arcangeli continue; 126171e3aac0SAndrea Arcangeli /* fall through */ 126271e3aac0SAndrea Arcangeli } 12631da177e4SLinus Torvalds if (pmd_none_or_clear_bad(src_pmd)) 12641da177e4SLinus Torvalds continue; 1265c78f4636SPeter Xu if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd, 1266c78f4636SPeter Xu addr, next)) 12671da177e4SLinus Torvalds return -ENOMEM; 12681da177e4SLinus Torvalds } while (dst_pmd++, src_pmd++, addr = next, addr != end); 12691da177e4SLinus Torvalds return 0; 12701da177e4SLinus Torvalds } 12711da177e4SLinus Torvalds 1272c78f4636SPeter Xu static inline int 1273c78f4636SPeter Xu copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1274c78f4636SPeter Xu p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr, 1275c78f4636SPeter Xu unsigned long end) 12761da177e4SLinus Torvalds { 1277c78f4636SPeter Xu struct mm_struct *dst_mm = dst_vma->vm_mm; 1278c78f4636SPeter Xu struct mm_struct *src_mm = src_vma->vm_mm; 12791da177e4SLinus Torvalds pud_t *src_pud, *dst_pud; 12801da177e4SLinus Torvalds unsigned long next; 12811da177e4SLinus Torvalds 1282c2febafcSKirill A. Shutemov dst_pud = pud_alloc(dst_mm, dst_p4d, addr); 12831da177e4SLinus Torvalds if (!dst_pud) 12841da177e4SLinus Torvalds return -ENOMEM; 1285c2febafcSKirill A. Shutemov src_pud = pud_offset(src_p4d, addr); 12861da177e4SLinus Torvalds do { 12871da177e4SLinus Torvalds next = pud_addr_end(addr, end); 1288a00cc7d9SMatthew Wilcox if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) { 1289a00cc7d9SMatthew Wilcox int err; 1290a00cc7d9SMatthew Wilcox 1291c78f4636SPeter Xu VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma); 1292a00cc7d9SMatthew Wilcox err = copy_huge_pud(dst_mm, src_mm, 1293c78f4636SPeter Xu dst_pud, src_pud, addr, src_vma); 1294a00cc7d9SMatthew Wilcox if (err == -ENOMEM) 1295a00cc7d9SMatthew Wilcox return -ENOMEM; 1296a00cc7d9SMatthew Wilcox if (!err) 1297a00cc7d9SMatthew Wilcox continue; 1298a00cc7d9SMatthew Wilcox /* fall through */ 1299a00cc7d9SMatthew Wilcox } 13001da177e4SLinus Torvalds if (pud_none_or_clear_bad(src_pud)) 13011da177e4SLinus Torvalds continue; 1302c78f4636SPeter Xu if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud, 1303c78f4636SPeter Xu addr, next)) 13041da177e4SLinus Torvalds return -ENOMEM; 13051da177e4SLinus Torvalds } while (dst_pud++, src_pud++, addr = next, addr != end); 13061da177e4SLinus Torvalds return 0; 13071da177e4SLinus Torvalds } 13081da177e4SLinus Torvalds 1309c78f4636SPeter Xu static inline int 1310c78f4636SPeter Xu copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma, 1311c78f4636SPeter Xu pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr, 1312c78f4636SPeter Xu unsigned long end) 1313c2febafcSKirill A. Shutemov { 1314c78f4636SPeter Xu struct mm_struct *dst_mm = dst_vma->vm_mm; 1315c2febafcSKirill A. Shutemov p4d_t *src_p4d, *dst_p4d; 1316c2febafcSKirill A. Shutemov unsigned long next; 1317c2febafcSKirill A. Shutemov 1318c2febafcSKirill A. Shutemov dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr); 1319c2febafcSKirill A. Shutemov if (!dst_p4d) 1320c2febafcSKirill A. Shutemov return -ENOMEM; 1321c2febafcSKirill A. Shutemov src_p4d = p4d_offset(src_pgd, addr); 1322c2febafcSKirill A. Shutemov do { 1323c2febafcSKirill A. Shutemov next = p4d_addr_end(addr, end); 1324c2febafcSKirill A. Shutemov if (p4d_none_or_clear_bad(src_p4d)) 1325c2febafcSKirill A. Shutemov continue; 1326c78f4636SPeter Xu if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d, 1327c78f4636SPeter Xu addr, next)) 1328c2febafcSKirill A. Shutemov return -ENOMEM; 1329c2febafcSKirill A. Shutemov } while (dst_p4d++, src_p4d++, addr = next, addr != end); 1330c2febafcSKirill A. Shutemov return 0; 1331c2febafcSKirill A. Shutemov } 1332c2febafcSKirill A. Shutemov 1333c56d1b62SPeter Xu /* 1334c56d1b62SPeter Xu * Return true if the vma needs to copy the pgtable during this fork(). Return 1335c56d1b62SPeter Xu * false when we can speed up fork() by allowing lazy page faults later until 1336c56d1b62SPeter Xu * when the child accesses the memory range. 1337c56d1b62SPeter Xu */ 1338bc70fbf2SPeter Xu static bool 1339c56d1b62SPeter Xu vma_needs_copy(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma) 1340c56d1b62SPeter Xu { 1341c56d1b62SPeter Xu /* 1342c56d1b62SPeter Xu * Always copy pgtables when dst_vma has uffd-wp enabled even if it's 1343c56d1b62SPeter Xu * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable 1344c56d1b62SPeter Xu * contains uffd-wp protection information, that's something we can't 1345c56d1b62SPeter Xu * retrieve from page cache, and skip copying will lose those info. 1346c56d1b62SPeter Xu */ 1347c56d1b62SPeter Xu if (userfaultfd_wp(dst_vma)) 1348c56d1b62SPeter Xu return true; 1349c56d1b62SPeter Xu 1350bcd51a3cSMike Kravetz if (src_vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) 1351c56d1b62SPeter Xu return true; 1352c56d1b62SPeter Xu 1353c56d1b62SPeter Xu if (src_vma->anon_vma) 1354c56d1b62SPeter Xu return true; 1355c56d1b62SPeter Xu 1356c56d1b62SPeter Xu /* 1357c56d1b62SPeter Xu * Don't copy ptes where a page fault will fill them correctly. Fork 1358c56d1b62SPeter Xu * becomes much lighter when there are big shared or private readonly 1359c56d1b62SPeter Xu * mappings. The tradeoff is that copy_page_range is more efficient 1360c56d1b62SPeter Xu * than faulting. 1361c56d1b62SPeter Xu */ 1362c56d1b62SPeter Xu return false; 1363c56d1b62SPeter Xu } 1364c56d1b62SPeter Xu 1365c78f4636SPeter Xu int 1366c78f4636SPeter Xu copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma) 13671da177e4SLinus Torvalds { 13681da177e4SLinus Torvalds pgd_t *src_pgd, *dst_pgd; 13691da177e4SLinus Torvalds unsigned long next; 1370c78f4636SPeter Xu unsigned long addr = src_vma->vm_start; 1371c78f4636SPeter Xu unsigned long end = src_vma->vm_end; 1372c78f4636SPeter Xu struct mm_struct *dst_mm = dst_vma->vm_mm; 1373c78f4636SPeter Xu struct mm_struct *src_mm = src_vma->vm_mm; 1374ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 13752ec74c3eSSagi Grimberg bool is_cow; 1376cddb8a5cSAndrea Arcangeli int ret; 13771da177e4SLinus Torvalds 1378c56d1b62SPeter Xu if (!vma_needs_copy(dst_vma, src_vma)) 1379d992895bSNick Piggin return 0; 1380d992895bSNick Piggin 1381c78f4636SPeter Xu if (is_vm_hugetlb_page(src_vma)) 1382bc70fbf2SPeter Xu return copy_hugetlb_page_range(dst_mm, src_mm, dst_vma, src_vma); 13831da177e4SLinus Torvalds 1384c78f4636SPeter Xu if (unlikely(src_vma->vm_flags & VM_PFNMAP)) { 13852ab64037Svenkatesh.pallipadi@intel.com /* 13862ab64037Svenkatesh.pallipadi@intel.com * We do not free on error cases below as remove_vma 13872ab64037Svenkatesh.pallipadi@intel.com * gets called on error from higher level routine 13882ab64037Svenkatesh.pallipadi@intel.com */ 1389c78f4636SPeter Xu ret = track_pfn_copy(src_vma); 13902ab64037Svenkatesh.pallipadi@intel.com if (ret) 13912ab64037Svenkatesh.pallipadi@intel.com return ret; 13922ab64037Svenkatesh.pallipadi@intel.com } 13932ab64037Svenkatesh.pallipadi@intel.com 1394cddb8a5cSAndrea Arcangeli /* 1395cddb8a5cSAndrea Arcangeli * We need to invalidate the secondary MMU mappings only when 1396cddb8a5cSAndrea Arcangeli * there could be a permission downgrade on the ptes of the 1397cddb8a5cSAndrea Arcangeli * parent mm. And a permission downgrade will only happen if 1398cddb8a5cSAndrea Arcangeli * is_cow_mapping() returns true. 1399cddb8a5cSAndrea Arcangeli */ 1400c78f4636SPeter Xu is_cow = is_cow_mapping(src_vma->vm_flags); 1401ac46d4f3SJérôme Glisse 1402ac46d4f3SJérôme Glisse if (is_cow) { 14037269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 14047d4a8be0SAlistair Popple 0, src_mm, addr, end); 1405ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 140657efa1feSJason Gunthorpe /* 140757efa1feSJason Gunthorpe * Disabling preemption is not needed for the write side, as 140857efa1feSJason Gunthorpe * the read side doesn't spin, but goes to the mmap_lock. 140957efa1feSJason Gunthorpe * 141057efa1feSJason Gunthorpe * Use the raw variant of the seqcount_t write API to avoid 141157efa1feSJason Gunthorpe * lockdep complaining about preemptibility. 141257efa1feSJason Gunthorpe */ 1413e727bfd5SSuren Baghdasaryan vma_assert_write_locked(src_vma); 141457efa1feSJason Gunthorpe raw_write_seqcount_begin(&src_mm->write_protect_seq); 1415ac46d4f3SJérôme Glisse } 1416cddb8a5cSAndrea Arcangeli 1417cddb8a5cSAndrea Arcangeli ret = 0; 14181da177e4SLinus Torvalds dst_pgd = pgd_offset(dst_mm, addr); 14191da177e4SLinus Torvalds src_pgd = pgd_offset(src_mm, addr); 14201da177e4SLinus Torvalds do { 14211da177e4SLinus Torvalds next = pgd_addr_end(addr, end); 14221da177e4SLinus Torvalds if (pgd_none_or_clear_bad(src_pgd)) 14231da177e4SLinus Torvalds continue; 1424c78f4636SPeter Xu if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd, 1425c78f4636SPeter Xu addr, next))) { 1426d155df53SMa Wupeng untrack_pfn_clear(dst_vma); 1427cddb8a5cSAndrea Arcangeli ret = -ENOMEM; 1428cddb8a5cSAndrea Arcangeli break; 1429cddb8a5cSAndrea Arcangeli } 14301da177e4SLinus Torvalds } while (dst_pgd++, src_pgd++, addr = next, addr != end); 1431cddb8a5cSAndrea Arcangeli 143257efa1feSJason Gunthorpe if (is_cow) { 143357efa1feSJason Gunthorpe raw_write_seqcount_end(&src_mm->write_protect_seq); 1434ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 143557efa1feSJason Gunthorpe } 1436cddb8a5cSAndrea Arcangeli return ret; 14371da177e4SLinus Torvalds } 14381da177e4SLinus Torvalds 14395abfd71dSPeter Xu /* Whether we should zap all COWed (private) pages too */ 14405abfd71dSPeter Xu static inline bool should_zap_cows(struct zap_details *details) 14415abfd71dSPeter Xu { 14425abfd71dSPeter Xu /* By default, zap all pages */ 14435abfd71dSPeter Xu if (!details) 14445abfd71dSPeter Xu return true; 14455abfd71dSPeter Xu 14465abfd71dSPeter Xu /* Or, we zap COWed pages only if the caller wants to */ 14472e148f1eSPeter Xu return details->even_cows; 14485abfd71dSPeter Xu } 14495abfd71dSPeter Xu 1450eabafaaaSKefeng Wang /* Decides whether we should zap this folio with the folio pointer specified */ 1451eabafaaaSKefeng Wang static inline bool should_zap_folio(struct zap_details *details, 1452eabafaaaSKefeng Wang struct folio *folio) 14533506659eSMatthew Wilcox (Oracle) { 1454eabafaaaSKefeng Wang /* If we can make a decision without *folio.. */ 14555abfd71dSPeter Xu if (should_zap_cows(details)) 1456254ab940SPeter Xu return true; 14573506659eSMatthew Wilcox (Oracle) 1458eabafaaaSKefeng Wang /* E.g. the caller passes NULL for the case of a zero folio */ 1459eabafaaaSKefeng Wang if (!folio) 1460254ab940SPeter Xu return true; 14615abfd71dSPeter Xu 1462eabafaaaSKefeng Wang /* Otherwise we should only zap non-anon folios */ 1463eabafaaaSKefeng Wang return !folio_test_anon(folio); 14643506659eSMatthew Wilcox (Oracle) } 14653506659eSMatthew Wilcox (Oracle) 1466999dad82SPeter Xu static inline bool zap_drop_file_uffd_wp(struct zap_details *details) 1467999dad82SPeter Xu { 1468999dad82SPeter Xu if (!details) 1469999dad82SPeter Xu return false; 1470999dad82SPeter Xu 1471999dad82SPeter Xu return details->zap_flags & ZAP_FLAG_DROP_MARKER; 1472999dad82SPeter Xu } 1473999dad82SPeter Xu 1474999dad82SPeter Xu /* 1475999dad82SPeter Xu * This function makes sure that we'll replace the none pte with an uffd-wp 1476999dad82SPeter Xu * swap special pte marker when necessary. Must be with the pgtable lock held. 1477999dad82SPeter Xu */ 1478999dad82SPeter Xu static inline void 1479999dad82SPeter Xu zap_install_uffd_wp_if_needed(struct vm_area_struct *vma, 1480999dad82SPeter Xu unsigned long addr, pte_t *pte, 1481999dad82SPeter Xu struct zap_details *details, pte_t pteval) 1482999dad82SPeter Xu { 14832bad466cSPeter Xu /* Zap on anonymous always means dropping everything */ 14842bad466cSPeter Xu if (vma_is_anonymous(vma)) 14852bad466cSPeter Xu return; 14862bad466cSPeter Xu 1487999dad82SPeter Xu if (zap_drop_file_uffd_wp(details)) 1488999dad82SPeter Xu return; 1489999dad82SPeter Xu 1490999dad82SPeter Xu pte_install_uffd_wp_if_needed(vma, addr, pte, pteval); 1491999dad82SPeter Xu } 1492999dad82SPeter Xu 149351c6f666SRobin Holt static unsigned long zap_pte_range(struct mmu_gather *tlb, 1494b5810039SNick Piggin struct vm_area_struct *vma, pmd_t *pmd, 14951da177e4SLinus Torvalds unsigned long addr, unsigned long end, 149697a89413SPeter Zijlstra struct zap_details *details) 14971da177e4SLinus Torvalds { 1498b5810039SNick Piggin struct mm_struct *mm = tlb->mm; 1499d16dfc55SPeter Zijlstra int force_flush = 0; 1500d559db08SKAMEZAWA Hiroyuki int rss[NR_MM_COUNTERS]; 150197a89413SPeter Zijlstra spinlock_t *ptl; 15025f1a1907SSteven Rostedt pte_t *start_pte; 150397a89413SPeter Zijlstra pte_t *pte; 15048a5f14a2SKirill A. Shutemov swp_entry_t entry; 1505d559db08SKAMEZAWA Hiroyuki 1506ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, PAGE_SIZE); 1507e303297eSPeter Zijlstra init_rss_vec(rss); 15083db82b93SHugh Dickins start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl); 15093db82b93SHugh Dickins if (!pte) 15103db82b93SHugh Dickins return addr; 15113db82b93SHugh Dickins 15123ea27719SMel Gorman flush_tlb_batched_pending(mm); 15136606c3e0SZachary Amsden arch_enter_lazy_mmu_mode(); 15141da177e4SLinus Torvalds do { 1515c33c7948SRyan Roberts pte_t ptent = ptep_get(pte); 1516eabafaaaSKefeng Wang struct folio *folio = NULL; 15178018db85SPeter Xu struct page *page; 15188018db85SPeter Xu 1519166f61b9STobin C Harding if (pte_none(ptent)) 15201da177e4SLinus Torvalds continue; 152151c6f666SRobin Holt 15227b167b68SMinchan Kim if (need_resched()) 15237b167b68SMinchan Kim break; 15247b167b68SMinchan Kim 15256f5e6b9eSHugh Dickins if (pte_present(ptent)) { 15265df397deSLinus Torvalds unsigned int delay_rmap; 15275df397deSLinus Torvalds 152825b2995aSChristoph Hellwig page = vm_normal_page(vma, addr, ptent); 1529eabafaaaSKefeng Wang if (page) 1530eabafaaaSKefeng Wang folio = page_folio(page); 1531eabafaaaSKefeng Wang 1532eabafaaaSKefeng Wang if (unlikely(!should_zap_folio(details, folio))) 15331da177e4SLinus Torvalds continue; 1534b5810039SNick Piggin ptent = ptep_get_and_clear_full(mm, addr, pte, 1535a600388dSZachary Amsden tlb->fullmm); 1536e5136e87SRick Edgecombe arch_check_zapped_pte(vma, ptent); 15371da177e4SLinus Torvalds tlb_remove_tlb_entry(tlb, pte, addr); 1538999dad82SPeter Xu zap_install_uffd_wp_if_needed(vma, addr, pte, details, 1539999dad82SPeter Xu ptent); 1540e2942062Sxu xin if (unlikely(!page)) { 15416080d19fSxu xin ksm_might_unmap_zero_page(mm, ptent); 15421da177e4SLinus Torvalds continue; 1543e2942062Sxu xin } 1544eca56ff9SJerome Marchand 15455df397deSLinus Torvalds delay_rmap = 0; 1546c4626503SDavid Hildenbrand if (!folio_test_anon(folio)) { 15471cf35d47SLinus Torvalds if (pte_dirty(ptent)) { 1548e4e3df29SDavid Hildenbrand folio_mark_dirty(folio); 15495df397deSLinus Torvalds if (tlb_delay_rmap(tlb)) { 15505df397deSLinus Torvalds delay_rmap = 1; 15515df397deSLinus Torvalds force_flush = 1; 15525df397deSLinus Torvalds } 15531cf35d47SLinus Torvalds } 15548788f678SYu Zhao if (pte_young(ptent) && likely(vma_has_recency(vma))) 1555c4626503SDavid Hildenbrand folio_mark_accessed(folio); 15566237bcd9SHugh Dickins } 1557a23f517bSKefeng Wang rss[mm_counter(folio)]--; 15585df397deSLinus Torvalds if (!delay_rmap) { 1559c4626503SDavid Hildenbrand folio_remove_rmap_pte(folio, page, vma); 15603dc14741SHugh Dickins if (unlikely(page_mapcount(page) < 0)) 15613dc14741SHugh Dickins print_bad_pte(vma, addr, ptent, page); 15625df397deSLinus Torvalds } 15635df397deSLinus Torvalds if (unlikely(__tlb_remove_page(tlb, page, delay_rmap))) { 15641cf35d47SLinus Torvalds force_flush = 1; 1565ce9ec37bSWill Deacon addr += PAGE_SIZE; 1566d16dfc55SPeter Zijlstra break; 15671cf35d47SLinus Torvalds } 15681da177e4SLinus Torvalds continue; 15691da177e4SLinus Torvalds } 15705042db43SJérôme Glisse 15715042db43SJérôme Glisse entry = pte_to_swp_entry(ptent); 1572b756a3b5SAlistair Popple if (is_device_private_entry(entry) || 1573b756a3b5SAlistair Popple is_device_exclusive_entry(entry)) { 15748018db85SPeter Xu page = pfn_swap_entry_to_page(entry); 1575c4626503SDavid Hildenbrand folio = page_folio(page); 1576eabafaaaSKefeng Wang if (unlikely(!should_zap_folio(details, folio))) 15775042db43SJérôme Glisse continue; 1578999dad82SPeter Xu /* 1579999dad82SPeter Xu * Both device private/exclusive mappings should only 1580999dad82SPeter Xu * work with anonymous page so far, so we don't need to 1581999dad82SPeter Xu * consider uffd-wp bit when zap. For more information, 1582999dad82SPeter Xu * see zap_install_uffd_wp_if_needed(). 1583999dad82SPeter Xu */ 1584999dad82SPeter Xu WARN_ON_ONCE(!vma_is_anonymous(vma)); 1585a23f517bSKefeng Wang rss[mm_counter(folio)]--; 1586b756a3b5SAlistair Popple if (is_device_private_entry(entry)) 1587c4626503SDavid Hildenbrand folio_remove_rmap_pte(folio, page, vma); 1588c4626503SDavid Hildenbrand folio_put(folio); 15898018db85SPeter Xu } else if (!non_swap_entry(entry)) { 15905abfd71dSPeter Xu /* Genuine swap entry, hence a private anon page */ 15915abfd71dSPeter Xu if (!should_zap_cows(details)) 15921da177e4SLinus Torvalds continue; 1593b084d435SKAMEZAWA Hiroyuki rss[MM_SWAPENTS]--; 15948018db85SPeter Xu if (unlikely(!free_swap_and_cache(entry))) 15958018db85SPeter Xu print_bad_pte(vma, addr, ptent, NULL); 15965abfd71dSPeter Xu } else if (is_migration_entry(entry)) { 1597eabafaaaSKefeng Wang folio = pfn_swap_entry_folio(entry); 1598eabafaaaSKefeng Wang if (!should_zap_folio(details, folio)) 15995abfd71dSPeter Xu continue; 1600a23f517bSKefeng Wang rss[mm_counter(folio)]--; 1601999dad82SPeter Xu } else if (pte_marker_entry_uffd_wp(entry)) { 16022bad466cSPeter Xu /* 16032bad466cSPeter Xu * For anon: always drop the marker; for file: only 16042bad466cSPeter Xu * drop the marker if explicitly requested. 16052bad466cSPeter Xu */ 16062bad466cSPeter Xu if (!vma_is_anonymous(vma) && 16072bad466cSPeter Xu !zap_drop_file_uffd_wp(details)) 1608999dad82SPeter Xu continue; 16099f186f9eSMiaohe Lin } else if (is_hwpoison_entry(entry) || 1610af19487fSAxel Rasmussen is_poisoned_swp_entry(entry)) { 16115abfd71dSPeter Xu if (!should_zap_cows(details)) 16125abfd71dSPeter Xu continue; 16135abfd71dSPeter Xu } else { 16145abfd71dSPeter Xu /* We should have covered all the swap entry types */ 1615727d16f1SAndrew Morton pr_alert("unrecognized swap entry 0x%lx\n", entry.val); 16165abfd71dSPeter Xu WARN_ON_ONCE(1); 16179f9f1acdSKonstantin Khlebnikov } 16189888a1caSZachary Amsden pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); 1619999dad82SPeter Xu zap_install_uffd_wp_if_needed(vma, addr, pte, details, ptent); 162097a89413SPeter Zijlstra } while (pte++, addr += PAGE_SIZE, addr != end); 1621ae859762SHugh Dickins 1622d559db08SKAMEZAWA Hiroyuki add_mm_rss_vec(mm, rss); 16236606c3e0SZachary Amsden arch_leave_lazy_mmu_mode(); 162451c6f666SRobin Holt 16251cf35d47SLinus Torvalds /* Do the actual TLB flush before dropping ptl */ 16265df397deSLinus Torvalds if (force_flush) { 16271cf35d47SLinus Torvalds tlb_flush_mmu_tlbonly(tlb); 16285df397deSLinus Torvalds tlb_flush_rmaps(tlb, vma); 16295df397deSLinus Torvalds } 16301cf35d47SLinus Torvalds pte_unmap_unlock(start_pte, ptl); 16311cf35d47SLinus Torvalds 16321cf35d47SLinus Torvalds /* 16331cf35d47SLinus Torvalds * If we forced a TLB flush (either due to running out of 16341cf35d47SLinus Torvalds * batch buffers or because we needed to flush dirty TLB 16351cf35d47SLinus Torvalds * entries before releasing the ptl), free the batched 16363db82b93SHugh Dickins * memory too. Come back again if we didn't do everything. 16371cf35d47SLinus Torvalds */ 16383db82b93SHugh Dickins if (force_flush) 1639fa0aafb8SPeter Zijlstra tlb_flush_mmu(tlb); 1640d16dfc55SPeter Zijlstra 164151c6f666SRobin Holt return addr; 16421da177e4SLinus Torvalds } 16431da177e4SLinus Torvalds 164451c6f666SRobin Holt static inline unsigned long zap_pmd_range(struct mmu_gather *tlb, 1645b5810039SNick Piggin struct vm_area_struct *vma, pud_t *pud, 16461da177e4SLinus Torvalds unsigned long addr, unsigned long end, 164797a89413SPeter Zijlstra struct zap_details *details) 16481da177e4SLinus Torvalds { 16491da177e4SLinus Torvalds pmd_t *pmd; 16501da177e4SLinus Torvalds unsigned long next; 16511da177e4SLinus Torvalds 16521da177e4SLinus Torvalds pmd = pmd_offset(pud, addr); 16531da177e4SLinus Torvalds do { 16541da177e4SLinus Torvalds next = pmd_addr_end(addr, end); 165584c3fc4eSZi Yan if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) { 165653406ed1SHugh Dickins if (next - addr != HPAGE_PMD_SIZE) 1657fd60775aSDavid Rientjes __split_huge_pmd(vma, pmd, addr, false, NULL); 16583db82b93SHugh Dickins else if (zap_huge_pmd(tlb, vma, pmd, addr)) { 16593db82b93SHugh Dickins addr = next; 16603db82b93SHugh Dickins continue; 16613db82b93SHugh Dickins } 166271e3aac0SAndrea Arcangeli /* fall through */ 16633506659eSMatthew Wilcox (Oracle) } else if (details && details->single_folio && 16643506659eSMatthew Wilcox (Oracle) folio_test_pmd_mappable(details->single_folio) && 166522061a1fSHugh Dickins next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) { 166622061a1fSHugh Dickins spinlock_t *ptl = pmd_lock(tlb->mm, pmd); 166722061a1fSHugh Dickins /* 166822061a1fSHugh Dickins * Take and drop THP pmd lock so that we cannot return 166922061a1fSHugh Dickins * prematurely, while zap_huge_pmd() has cleared *pmd, 167022061a1fSHugh Dickins * but not yet decremented compound_mapcount(). 167122061a1fSHugh Dickins */ 167222061a1fSHugh Dickins spin_unlock(ptl); 167371e3aac0SAndrea Arcangeli } 16743db82b93SHugh Dickins if (pmd_none(*pmd)) { 16753db82b93SHugh Dickins addr = next; 16763db82b93SHugh Dickins continue; 16773db82b93SHugh Dickins } 16783db82b93SHugh Dickins addr = zap_pte_range(tlb, vma, pmd, addr, next, details); 16793db82b93SHugh Dickins if (addr != next) 16803db82b93SHugh Dickins pmd--; 16813db82b93SHugh Dickins } while (pmd++, cond_resched(), addr != end); 168251c6f666SRobin Holt 168351c6f666SRobin Holt return addr; 16841da177e4SLinus Torvalds } 16851da177e4SLinus Torvalds 168651c6f666SRobin Holt static inline unsigned long zap_pud_range(struct mmu_gather *tlb, 1687c2febafcSKirill A. Shutemov struct vm_area_struct *vma, p4d_t *p4d, 16881da177e4SLinus Torvalds unsigned long addr, unsigned long end, 168997a89413SPeter Zijlstra struct zap_details *details) 16901da177e4SLinus Torvalds { 16911da177e4SLinus Torvalds pud_t *pud; 16921da177e4SLinus Torvalds unsigned long next; 16931da177e4SLinus Torvalds 1694c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 16951da177e4SLinus Torvalds do { 16961da177e4SLinus Torvalds next = pud_addr_end(addr, end); 1697a00cc7d9SMatthew Wilcox if (pud_trans_huge(*pud) || pud_devmap(*pud)) { 1698a00cc7d9SMatthew Wilcox if (next - addr != HPAGE_PUD_SIZE) { 169942fc5414SMichel Lespinasse mmap_assert_locked(tlb->mm); 1700a00cc7d9SMatthew Wilcox split_huge_pud(vma, pud, addr); 1701a00cc7d9SMatthew Wilcox } else if (zap_huge_pud(tlb, vma, pud, addr)) 1702a00cc7d9SMatthew Wilcox goto next; 1703a00cc7d9SMatthew Wilcox /* fall through */ 1704a00cc7d9SMatthew Wilcox } 170597a89413SPeter Zijlstra if (pud_none_or_clear_bad(pud)) 17061da177e4SLinus Torvalds continue; 170797a89413SPeter Zijlstra next = zap_pmd_range(tlb, vma, pud, addr, next, details); 1708a00cc7d9SMatthew Wilcox next: 1709a00cc7d9SMatthew Wilcox cond_resched(); 171097a89413SPeter Zijlstra } while (pud++, addr = next, addr != end); 171151c6f666SRobin Holt 171251c6f666SRobin Holt return addr; 17131da177e4SLinus Torvalds } 17141da177e4SLinus Torvalds 1715c2febafcSKirill A. Shutemov static inline unsigned long zap_p4d_range(struct mmu_gather *tlb, 1716c2febafcSKirill A. Shutemov struct vm_area_struct *vma, pgd_t *pgd, 1717c2febafcSKirill A. Shutemov unsigned long addr, unsigned long end, 1718c2febafcSKirill A. Shutemov struct zap_details *details) 1719c2febafcSKirill A. Shutemov { 1720c2febafcSKirill A. Shutemov p4d_t *p4d; 1721c2febafcSKirill A. Shutemov unsigned long next; 1722c2febafcSKirill A. Shutemov 1723c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 1724c2febafcSKirill A. Shutemov do { 1725c2febafcSKirill A. Shutemov next = p4d_addr_end(addr, end); 1726c2febafcSKirill A. Shutemov if (p4d_none_or_clear_bad(p4d)) 1727c2febafcSKirill A. Shutemov continue; 1728c2febafcSKirill A. Shutemov next = zap_pud_range(tlb, vma, p4d, addr, next, details); 1729c2febafcSKirill A. Shutemov } while (p4d++, addr = next, addr != end); 1730c2febafcSKirill A. Shutemov 1731c2febafcSKirill A. Shutemov return addr; 1732c2febafcSKirill A. Shutemov } 1733c2febafcSKirill A. Shutemov 1734aac45363SMichal Hocko void unmap_page_range(struct mmu_gather *tlb, 173551c6f666SRobin Holt struct vm_area_struct *vma, 17361da177e4SLinus Torvalds unsigned long addr, unsigned long end, 173797a89413SPeter Zijlstra struct zap_details *details) 17381da177e4SLinus Torvalds { 17391da177e4SLinus Torvalds pgd_t *pgd; 17401da177e4SLinus Torvalds unsigned long next; 17411da177e4SLinus Torvalds 17421da177e4SLinus Torvalds BUG_ON(addr >= end); 17431da177e4SLinus Torvalds tlb_start_vma(tlb, vma); 17441da177e4SLinus Torvalds pgd = pgd_offset(vma->vm_mm, addr); 17451da177e4SLinus Torvalds do { 17461da177e4SLinus Torvalds next = pgd_addr_end(addr, end); 174797a89413SPeter Zijlstra if (pgd_none_or_clear_bad(pgd)) 17481da177e4SLinus Torvalds continue; 1749c2febafcSKirill A. Shutemov next = zap_p4d_range(tlb, vma, pgd, addr, next, details); 175097a89413SPeter Zijlstra } while (pgd++, addr = next, addr != end); 17511da177e4SLinus Torvalds tlb_end_vma(tlb, vma); 17521da177e4SLinus Torvalds } 17531da177e4SLinus Torvalds 1754f5cc4eefSAl Viro 1755f5cc4eefSAl Viro static void unmap_single_vma(struct mmu_gather *tlb, 17561da177e4SLinus Torvalds struct vm_area_struct *vma, unsigned long start_addr, 17574f74d2c8SLinus Torvalds unsigned long end_addr, 175868f48381SSuren Baghdasaryan struct zap_details *details, bool mm_wr_locked) 17591da177e4SLinus Torvalds { 1760f5cc4eefSAl Viro unsigned long start = max(vma->vm_start, start_addr); 17611da177e4SLinus Torvalds unsigned long end; 17621da177e4SLinus Torvalds 17631da177e4SLinus Torvalds if (start >= vma->vm_end) 1764f5cc4eefSAl Viro return; 17651da177e4SLinus Torvalds end = min(vma->vm_end, end_addr); 17661da177e4SLinus Torvalds if (end <= vma->vm_start) 1767f5cc4eefSAl Viro return; 17681da177e4SLinus Torvalds 1769cbc91f71SSrikar Dronamraju if (vma->vm_file) 1770cbc91f71SSrikar Dronamraju uprobe_munmap(vma, start, end); 1771cbc91f71SSrikar Dronamraju 1772b3b9c293SKonstantin Khlebnikov if (unlikely(vma->vm_flags & VM_PFNMAP)) 177368f48381SSuren Baghdasaryan untrack_pfn(vma, 0, 0, mm_wr_locked); 17742ab64037Svenkatesh.pallipadi@intel.com 17758b2a1238SAl Viro if (start != end) { 177651c6f666SRobin Holt if (unlikely(is_vm_hugetlb_page(vma))) { 1777a137e1ccSAndi Kleen /* 1778a137e1ccSAndi Kleen * It is undesirable to test vma->vm_file as it 1779a137e1ccSAndi Kleen * should be non-null for valid hugetlb area. 1780a137e1ccSAndi Kleen * However, vm_file will be NULL in the error 17817aa6b4adSDavidlohr Bueso * cleanup path of mmap_region. When 1782a137e1ccSAndi Kleen * hugetlbfs ->mmap method fails, 17837aa6b4adSDavidlohr Bueso * mmap_region() nullifies vma->vm_file 1784a137e1ccSAndi Kleen * before calling this function to clean up. 1785a137e1ccSAndi Kleen * Since no pte has actually been setup, it is 1786a137e1ccSAndi Kleen * safe to do nothing in this case. 1787a137e1ccSAndi Kleen */ 178824669e58SAneesh Kumar K.V if (vma->vm_file) { 178905e90bd0SPeter Xu zap_flags_t zap_flags = details ? 179005e90bd0SPeter Xu details->zap_flags : 0; 17912820b0f0SRik van Riel __unmap_hugepage_range(tlb, vma, start, end, 179205e90bd0SPeter Xu NULL, zap_flags); 179324669e58SAneesh Kumar K.V } 179451c6f666SRobin Holt } else 1795038c7aa1SAl Viro unmap_page_range(tlb, vma, start, end, details); 179697a89413SPeter Zijlstra } 179751c6f666SRobin Holt } 17981da177e4SLinus Torvalds 1799f5cc4eefSAl Viro /** 1800f5cc4eefSAl Viro * unmap_vmas - unmap a range of memory covered by a list of vma's 1801f5cc4eefSAl Viro * @tlb: address of the caller's struct mmu_gather 18026e412203SYang Li * @mas: the maple state 1803f5cc4eefSAl Viro * @vma: the starting vma 1804f5cc4eefSAl Viro * @start_addr: virtual address at which to start unmapping 1805f5cc4eefSAl Viro * @end_addr: virtual address at which to end unmapping 18066e412203SYang Li * @tree_end: The maximum index to check 1807809ef83cSYang Li * @mm_wr_locked: lock flag 1808f5cc4eefSAl Viro * 1809f5cc4eefSAl Viro * Unmap all pages in the vma list. 1810f5cc4eefSAl Viro * 1811f5cc4eefSAl Viro * Only addresses between `start' and `end' will be unmapped. 1812f5cc4eefSAl Viro * 1813f5cc4eefSAl Viro * The VMA list must be sorted in ascending virtual address order. 1814f5cc4eefSAl Viro * 1815f5cc4eefSAl Viro * unmap_vmas() assumes that the caller will flush the whole unmapped address 1816f5cc4eefSAl Viro * range after unmap_vmas() returns. So the only responsibility here is to 1817f5cc4eefSAl Viro * ensure that any thus-far unmapped pages are flushed before unmap_vmas() 1818f5cc4eefSAl Viro * drops the lock and schedules. 1819f5cc4eefSAl Viro */ 1820fd892593SLiam R. Howlett void unmap_vmas(struct mmu_gather *tlb, struct ma_state *mas, 1821f5cc4eefSAl Viro struct vm_area_struct *vma, unsigned long start_addr, 1822fd892593SLiam R. Howlett unsigned long end_addr, unsigned long tree_end, 1823fd892593SLiam R. Howlett bool mm_wr_locked) 1824f5cc4eefSAl Viro { 1825ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 1826999dad82SPeter Xu struct zap_details details = { 182704ada095SMike Kravetz .zap_flags = ZAP_FLAG_DROP_MARKER | ZAP_FLAG_UNMAP, 1828999dad82SPeter Xu /* Careful - we need to zap private pages too! */ 1829999dad82SPeter Xu .even_cows = true, 1830999dad82SPeter Xu }; 1831f5cc4eefSAl Viro 18327d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma->vm_mm, 18336f4f13e8SJérôme Glisse start_addr, end_addr); 1834ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 1835763ecb03SLiam R. Howlett do { 18362820b0f0SRik van Riel unsigned long start = start_addr; 18372820b0f0SRik van Riel unsigned long end = end_addr; 18382820b0f0SRik van Riel hugetlb_zap_begin(vma, &start, &end); 18392820b0f0SRik van Riel unmap_single_vma(tlb, vma, start, end, &details, 184068f48381SSuren Baghdasaryan mm_wr_locked); 18412820b0f0SRik van Riel hugetlb_zap_end(vma, &details); 1842d2406291SPeng Zhang vma = mas_find(mas, tree_end - 1); 1843d2406291SPeng Zhang } while (vma && likely(!xa_is_zero(vma))); 1844ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 18451da177e4SLinus Torvalds } 18461da177e4SLinus Torvalds 18471da177e4SLinus Torvalds /** 1848f5cc4eefSAl Viro * zap_page_range_single - remove user pages in a given range 1849f5cc4eefSAl Viro * @vma: vm_area_struct holding the applicable pages 1850f5cc4eefSAl Viro * @address: starting address of pages to zap 1851f5cc4eefSAl Viro * @size: number of bytes to zap 18528a5f14a2SKirill A. Shutemov * @details: details of shared cache invalidation 1853f5cc4eefSAl Viro * 1854f5cc4eefSAl Viro * The range must fit into one VMA. 1855f5cc4eefSAl Viro */ 185621b85b09SMike Kravetz void zap_page_range_single(struct vm_area_struct *vma, unsigned long address, 1857f5cc4eefSAl Viro unsigned long size, struct zap_details *details) 1858f5cc4eefSAl Viro { 185921b85b09SMike Kravetz const unsigned long end = address + size; 1860ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 1861f5cc4eefSAl Viro struct mmu_gather tlb; 1862f5cc4eefSAl Viro 1863f5cc4eefSAl Viro lru_add_drain(); 18647d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 186521b85b09SMike Kravetz address, end); 18662820b0f0SRik van Riel hugetlb_zap_begin(vma, &range.start, &range.end); 1867a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 1868ac46d4f3SJérôme Glisse update_hiwater_rss(vma->vm_mm); 1869ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 187021b85b09SMike Kravetz /* 187121b85b09SMike Kravetz * unmap 'address-end' not 'range.start-range.end' as range 187221b85b09SMike Kravetz * could have been expanded for hugetlb pmd sharing. 187321b85b09SMike Kravetz */ 187468f48381SSuren Baghdasaryan unmap_single_vma(&tlb, vma, address, end, details, false); 1875ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 1876ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 18772820b0f0SRik van Riel hugetlb_zap_end(vma, details); 18781da177e4SLinus Torvalds } 18791da177e4SLinus Torvalds 1880c627f9ccSJack Steiner /** 1881c627f9ccSJack Steiner * zap_vma_ptes - remove ptes mapping the vma 1882c627f9ccSJack Steiner * @vma: vm_area_struct holding ptes to be zapped 1883c627f9ccSJack Steiner * @address: starting address of pages to zap 1884c627f9ccSJack Steiner * @size: number of bytes to zap 1885c627f9ccSJack Steiner * 1886c627f9ccSJack Steiner * This function only unmaps ptes assigned to VM_PFNMAP vmas. 1887c627f9ccSJack Steiner * 1888c627f9ccSJack Steiner * The entire address range must be fully contained within the vma. 1889c627f9ccSJack Steiner * 1890c627f9ccSJack Steiner */ 189127d036e3SLeon Romanovsky void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address, 1892c627f9ccSJack Steiner unsigned long size) 1893c627f9ccSJack Steiner { 189488a35912SMiaohe Lin if (!range_in_vma(vma, address, address + size) || 1895c627f9ccSJack Steiner !(vma->vm_flags & VM_PFNMAP)) 189627d036e3SLeon Romanovsky return; 189727d036e3SLeon Romanovsky 1898f5cc4eefSAl Viro zap_page_range_single(vma, address, size, NULL); 1899c627f9ccSJack Steiner } 1900c627f9ccSJack Steiner EXPORT_SYMBOL_GPL(zap_vma_ptes); 1901c627f9ccSJack Steiner 19028cd3984dSArjun Roy static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr) 1903c9cfcddfSLinus Torvalds { 1904c2febafcSKirill A. Shutemov pgd_t *pgd; 1905c2febafcSKirill A. Shutemov p4d_t *p4d; 1906c2febafcSKirill A. Shutemov pud_t *pud; 1907c2febafcSKirill A. Shutemov pmd_t *pmd; 1908c2febafcSKirill A. Shutemov 1909c2febafcSKirill A. Shutemov pgd = pgd_offset(mm, addr); 1910c2febafcSKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 1911c2febafcSKirill A. Shutemov if (!p4d) 1912c2febafcSKirill A. Shutemov return NULL; 1913c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 1914c2febafcSKirill A. Shutemov if (!pud) 1915c2febafcSKirill A. Shutemov return NULL; 1916c2febafcSKirill A. Shutemov pmd = pmd_alloc(mm, pud, addr); 1917c2febafcSKirill A. Shutemov if (!pmd) 1918c2febafcSKirill A. Shutemov return NULL; 1919c2febafcSKirill A. Shutemov 1920f66055abSAndrea Arcangeli VM_BUG_ON(pmd_trans_huge(*pmd)); 19218cd3984dSArjun Roy return pmd; 19228cd3984dSArjun Roy } 19238cd3984dSArjun Roy 19248cd3984dSArjun Roy pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr, 19258cd3984dSArjun Roy spinlock_t **ptl) 19268cd3984dSArjun Roy { 19278cd3984dSArjun Roy pmd_t *pmd = walk_to_pmd(mm, addr); 19288cd3984dSArjun Roy 19298cd3984dSArjun Roy if (!pmd) 19308cd3984dSArjun Roy return NULL; 1931c9cfcddfSLinus Torvalds return pte_alloc_map_lock(mm, pmd, addr, ptl); 1932c9cfcddfSLinus Torvalds } 1933c9cfcddfSLinus Torvalds 19348efd6f5bSArjun Roy static int validate_page_before_insert(struct page *page) 19358efd6f5bSArjun Roy { 1936f8b6187dSKefeng Wang struct folio *folio = page_folio(page); 1937f8b6187dSKefeng Wang 1938f8b6187dSKefeng Wang if (folio_test_anon(folio) || folio_test_slab(folio) || 1939f8b6187dSKefeng Wang page_has_type(page)) 19408efd6f5bSArjun Roy return -EINVAL; 1941f8b6187dSKefeng Wang flush_dcache_folio(folio); 19428efd6f5bSArjun Roy return 0; 19438efd6f5bSArjun Roy } 19448efd6f5bSArjun Roy 1945cea86fe2SHugh Dickins static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte, 19468efd6f5bSArjun Roy unsigned long addr, struct page *page, pgprot_t prot) 19478efd6f5bSArjun Roy { 1948ef37b2eaSDavid Hildenbrand struct folio *folio = page_folio(page); 1949ef37b2eaSDavid Hildenbrand 1950c33c7948SRyan Roberts if (!pte_none(ptep_get(pte))) 19518efd6f5bSArjun Roy return -EBUSY; 19528efd6f5bSArjun Roy /* Ok, finally just insert the thing.. */ 1953ef37b2eaSDavid Hildenbrand folio_get(folio); 19546b27cc6cSKefeng Wang inc_mm_counter(vma->vm_mm, mm_counter_file(folio)); 1955ef37b2eaSDavid Hildenbrand folio_add_file_rmap_pte(folio, page, vma); 1956cea86fe2SHugh Dickins set_pte_at(vma->vm_mm, addr, pte, mk_pte(page, prot)); 19578efd6f5bSArjun Roy return 0; 19588efd6f5bSArjun Roy } 19598efd6f5bSArjun Roy 19601da177e4SLinus Torvalds /* 1961238f58d8SLinus Torvalds * This is the old fallback for page remapping. 1962238f58d8SLinus Torvalds * 1963238f58d8SLinus Torvalds * For historical reasons, it only allows reserved pages. Only 1964238f58d8SLinus Torvalds * old drivers should use this, and they needed to mark their 1965238f58d8SLinus Torvalds * pages reserved for the old functions anyway. 1966238f58d8SLinus Torvalds */ 1967423bad60SNick Piggin static int insert_page(struct vm_area_struct *vma, unsigned long addr, 1968423bad60SNick Piggin struct page *page, pgprot_t prot) 1969238f58d8SLinus Torvalds { 1970238f58d8SLinus Torvalds int retval; 1971238f58d8SLinus Torvalds pte_t *pte; 1972238f58d8SLinus Torvalds spinlock_t *ptl; 1973238f58d8SLinus Torvalds 19748efd6f5bSArjun Roy retval = validate_page_before_insert(page); 19758efd6f5bSArjun Roy if (retval) 19765b4e655eSKAMEZAWA Hiroyuki goto out; 1977238f58d8SLinus Torvalds retval = -ENOMEM; 1978cea86fe2SHugh Dickins pte = get_locked_pte(vma->vm_mm, addr, &ptl); 1979238f58d8SLinus Torvalds if (!pte) 19805b4e655eSKAMEZAWA Hiroyuki goto out; 1981cea86fe2SHugh Dickins retval = insert_page_into_pte_locked(vma, pte, addr, page, prot); 1982238f58d8SLinus Torvalds pte_unmap_unlock(pte, ptl); 1983238f58d8SLinus Torvalds out: 1984238f58d8SLinus Torvalds return retval; 1985238f58d8SLinus Torvalds } 1986238f58d8SLinus Torvalds 1987cea86fe2SHugh Dickins static int insert_page_in_batch_locked(struct vm_area_struct *vma, pte_t *pte, 19888cd3984dSArjun Roy unsigned long addr, struct page *page, pgprot_t prot) 19898cd3984dSArjun Roy { 19908cd3984dSArjun Roy int err; 19918cd3984dSArjun Roy 19928cd3984dSArjun Roy if (!page_count(page)) 19938cd3984dSArjun Roy return -EINVAL; 19948cd3984dSArjun Roy err = validate_page_before_insert(page); 19957f70c2a6SArjun Roy if (err) 19967f70c2a6SArjun Roy return err; 1997cea86fe2SHugh Dickins return insert_page_into_pte_locked(vma, pte, addr, page, prot); 19988cd3984dSArjun Roy } 19998cd3984dSArjun Roy 20008cd3984dSArjun Roy /* insert_pages() amortizes the cost of spinlock operations 2001bb7dbaafSMatthew Wilcox (Oracle) * when inserting pages in a loop. 20028cd3984dSArjun Roy */ 20038cd3984dSArjun Roy static int insert_pages(struct vm_area_struct *vma, unsigned long addr, 20048cd3984dSArjun Roy struct page **pages, unsigned long *num, pgprot_t prot) 20058cd3984dSArjun Roy { 20068cd3984dSArjun Roy pmd_t *pmd = NULL; 20077f70c2a6SArjun Roy pte_t *start_pte, *pte; 20087f70c2a6SArjun Roy spinlock_t *pte_lock; 20098cd3984dSArjun Roy struct mm_struct *const mm = vma->vm_mm; 20108cd3984dSArjun Roy unsigned long curr_page_idx = 0; 20118cd3984dSArjun Roy unsigned long remaining_pages_total = *num; 20128cd3984dSArjun Roy unsigned long pages_to_write_in_pmd; 20138cd3984dSArjun Roy int ret; 20148cd3984dSArjun Roy more: 20158cd3984dSArjun Roy ret = -EFAULT; 20168cd3984dSArjun Roy pmd = walk_to_pmd(mm, addr); 20178cd3984dSArjun Roy if (!pmd) 20188cd3984dSArjun Roy goto out; 20198cd3984dSArjun Roy 20208cd3984dSArjun Roy pages_to_write_in_pmd = min_t(unsigned long, 20218cd3984dSArjun Roy remaining_pages_total, PTRS_PER_PTE - pte_index(addr)); 20228cd3984dSArjun Roy 20238cd3984dSArjun Roy /* Allocate the PTE if necessary; takes PMD lock once only. */ 20248cd3984dSArjun Roy ret = -ENOMEM; 20258cd3984dSArjun Roy if (pte_alloc(mm, pmd)) 20268cd3984dSArjun Roy goto out; 20278cd3984dSArjun Roy 20288cd3984dSArjun Roy while (pages_to_write_in_pmd) { 20298cd3984dSArjun Roy int pte_idx = 0; 20308cd3984dSArjun Roy const int batch_size = min_t(int, pages_to_write_in_pmd, 8); 20318cd3984dSArjun Roy 20327f70c2a6SArjun Roy start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock); 20333db82b93SHugh Dickins if (!start_pte) { 20343db82b93SHugh Dickins ret = -EFAULT; 20353db82b93SHugh Dickins goto out; 20363db82b93SHugh Dickins } 20377f70c2a6SArjun Roy for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) { 2038cea86fe2SHugh Dickins int err = insert_page_in_batch_locked(vma, pte, 20398cd3984dSArjun Roy addr, pages[curr_page_idx], prot); 20408cd3984dSArjun Roy if (unlikely(err)) { 20417f70c2a6SArjun Roy pte_unmap_unlock(start_pte, pte_lock); 20428cd3984dSArjun Roy ret = err; 20438cd3984dSArjun Roy remaining_pages_total -= pte_idx; 20448cd3984dSArjun Roy goto out; 20458cd3984dSArjun Roy } 20468cd3984dSArjun Roy addr += PAGE_SIZE; 20478cd3984dSArjun Roy ++curr_page_idx; 20488cd3984dSArjun Roy } 20497f70c2a6SArjun Roy pte_unmap_unlock(start_pte, pte_lock); 20508cd3984dSArjun Roy pages_to_write_in_pmd -= batch_size; 20518cd3984dSArjun Roy remaining_pages_total -= batch_size; 20528cd3984dSArjun Roy } 20538cd3984dSArjun Roy if (remaining_pages_total) 20548cd3984dSArjun Roy goto more; 20558cd3984dSArjun Roy ret = 0; 20568cd3984dSArjun Roy out: 20578cd3984dSArjun Roy *num = remaining_pages_total; 20588cd3984dSArjun Roy return ret; 20598cd3984dSArjun Roy } 20608cd3984dSArjun Roy 20618cd3984dSArjun Roy /** 20628cd3984dSArjun Roy * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock. 20638cd3984dSArjun Roy * @vma: user vma to map to 20648cd3984dSArjun Roy * @addr: target start user address of these pages 20658cd3984dSArjun Roy * @pages: source kernel pages 20668cd3984dSArjun Roy * @num: in: number of pages to map. out: number of pages that were *not* 20678cd3984dSArjun Roy * mapped. (0 means all pages were successfully mapped). 20688cd3984dSArjun Roy * 20698cd3984dSArjun Roy * Preferred over vm_insert_page() when inserting multiple pages. 20708cd3984dSArjun Roy * 20718cd3984dSArjun Roy * In case of error, we may have mapped a subset of the provided 20728cd3984dSArjun Roy * pages. It is the caller's responsibility to account for this case. 20738cd3984dSArjun Roy * 20748cd3984dSArjun Roy * The same restrictions apply as in vm_insert_page(). 20758cd3984dSArjun Roy */ 20768cd3984dSArjun Roy int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr, 20778cd3984dSArjun Roy struct page **pages, unsigned long *num) 20788cd3984dSArjun Roy { 20798cd3984dSArjun Roy const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1; 20808cd3984dSArjun Roy 20818cd3984dSArjun Roy if (addr < vma->vm_start || end_addr >= vma->vm_end) 20828cd3984dSArjun Roy return -EFAULT; 20838cd3984dSArjun Roy if (!(vma->vm_flags & VM_MIXEDMAP)) { 2084d8ed45c5SMichel Lespinasse BUG_ON(mmap_read_trylock(vma->vm_mm)); 20858cd3984dSArjun Roy BUG_ON(vma->vm_flags & VM_PFNMAP); 20861c71222eSSuren Baghdasaryan vm_flags_set(vma, VM_MIXEDMAP); 20878cd3984dSArjun Roy } 20888cd3984dSArjun Roy /* Defer page refcount checking till we're about to map that page. */ 20898cd3984dSArjun Roy return insert_pages(vma, addr, pages, num, vma->vm_page_prot); 20908cd3984dSArjun Roy } 20918cd3984dSArjun Roy EXPORT_SYMBOL(vm_insert_pages); 20928cd3984dSArjun Roy 2093bfa5bf6dSRolf Eike Beer /** 2094bfa5bf6dSRolf Eike Beer * vm_insert_page - insert single page into user vma 2095bfa5bf6dSRolf Eike Beer * @vma: user vma to map to 2096bfa5bf6dSRolf Eike Beer * @addr: target user address of this page 2097bfa5bf6dSRolf Eike Beer * @page: source kernel page 2098bfa5bf6dSRolf Eike Beer * 2099a145dd41SLinus Torvalds * This allows drivers to insert individual pages they've allocated 2100a145dd41SLinus Torvalds * into a user vma. 2101a145dd41SLinus Torvalds * 2102a145dd41SLinus Torvalds * The page has to be a nice clean _individual_ kernel allocation. 2103a145dd41SLinus Torvalds * If you allocate a compound page, you need to have marked it as 2104a145dd41SLinus Torvalds * such (__GFP_COMP), or manually just split the page up yourself 21058dfcc9baSNick Piggin * (see split_page()). 2106a145dd41SLinus Torvalds * 2107a145dd41SLinus Torvalds * NOTE! Traditionally this was done with "remap_pfn_range()" which 2108a145dd41SLinus Torvalds * took an arbitrary page protection parameter. This doesn't allow 2109a145dd41SLinus Torvalds * that. Your vma protection will have to be set up correctly, which 2110a145dd41SLinus Torvalds * means that if you want a shared writable mapping, you'd better 2111a145dd41SLinus Torvalds * ask for a shared writable mapping! 2112a145dd41SLinus Torvalds * 2113a145dd41SLinus Torvalds * The page does not need to be reserved. 21144b6e1e37SKonstantin Khlebnikov * 21154b6e1e37SKonstantin Khlebnikov * Usually this function is called from f_op->mmap() handler 2116c1e8d7c6SMichel Lespinasse * under mm->mmap_lock write-lock, so it can change vma->vm_flags. 21174b6e1e37SKonstantin Khlebnikov * Caller must set VM_MIXEDMAP on vma if it wants to call this 21184b6e1e37SKonstantin Khlebnikov * function from other places, for example from page-fault handler. 2119a862f68aSMike Rapoport * 2120a862f68aSMike Rapoport * Return: %0 on success, negative error code otherwise. 2121a145dd41SLinus Torvalds */ 2122423bad60SNick Piggin int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, 2123423bad60SNick Piggin struct page *page) 2124a145dd41SLinus Torvalds { 2125a145dd41SLinus Torvalds if (addr < vma->vm_start || addr >= vma->vm_end) 2126a145dd41SLinus Torvalds return -EFAULT; 2127a145dd41SLinus Torvalds if (!page_count(page)) 2128a145dd41SLinus Torvalds return -EINVAL; 21294b6e1e37SKonstantin Khlebnikov if (!(vma->vm_flags & VM_MIXEDMAP)) { 2130d8ed45c5SMichel Lespinasse BUG_ON(mmap_read_trylock(vma->vm_mm)); 21314b6e1e37SKonstantin Khlebnikov BUG_ON(vma->vm_flags & VM_PFNMAP); 21321c71222eSSuren Baghdasaryan vm_flags_set(vma, VM_MIXEDMAP); 21334b6e1e37SKonstantin Khlebnikov } 2134423bad60SNick Piggin return insert_page(vma, addr, page, vma->vm_page_prot); 2135a145dd41SLinus Torvalds } 2136e3c3374fSLinus Torvalds EXPORT_SYMBOL(vm_insert_page); 2137a145dd41SLinus Torvalds 2138a667d745SSouptick Joarder /* 2139a667d745SSouptick Joarder * __vm_map_pages - maps range of kernel pages into user vma 2140a667d745SSouptick Joarder * @vma: user vma to map to 2141a667d745SSouptick Joarder * @pages: pointer to array of source kernel pages 2142a667d745SSouptick Joarder * @num: number of pages in page array 2143a667d745SSouptick Joarder * @offset: user's requested vm_pgoff 2144a667d745SSouptick Joarder * 2145a667d745SSouptick Joarder * This allows drivers to map range of kernel pages into a user vma. 2146a667d745SSouptick Joarder * 2147a667d745SSouptick Joarder * Return: 0 on success and error code otherwise. 2148a667d745SSouptick Joarder */ 2149a667d745SSouptick Joarder static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages, 2150a667d745SSouptick Joarder unsigned long num, unsigned long offset) 2151a667d745SSouptick Joarder { 2152a667d745SSouptick Joarder unsigned long count = vma_pages(vma); 2153a667d745SSouptick Joarder unsigned long uaddr = vma->vm_start; 2154a667d745SSouptick Joarder int ret, i; 2155a667d745SSouptick Joarder 2156a667d745SSouptick Joarder /* Fail if the user requested offset is beyond the end of the object */ 215796756fcbSMiguel Ojeda if (offset >= num) 2158a667d745SSouptick Joarder return -ENXIO; 2159a667d745SSouptick Joarder 2160a667d745SSouptick Joarder /* Fail if the user requested size exceeds available object size */ 2161a667d745SSouptick Joarder if (count > num - offset) 2162a667d745SSouptick Joarder return -ENXIO; 2163a667d745SSouptick Joarder 2164a667d745SSouptick Joarder for (i = 0; i < count; i++) { 2165a667d745SSouptick Joarder ret = vm_insert_page(vma, uaddr, pages[offset + i]); 2166a667d745SSouptick Joarder if (ret < 0) 2167a667d745SSouptick Joarder return ret; 2168a667d745SSouptick Joarder uaddr += PAGE_SIZE; 2169a667d745SSouptick Joarder } 2170a667d745SSouptick Joarder 2171a667d745SSouptick Joarder return 0; 2172a667d745SSouptick Joarder } 2173a667d745SSouptick Joarder 2174a667d745SSouptick Joarder /** 2175a667d745SSouptick Joarder * vm_map_pages - maps range of kernel pages starts with non zero offset 2176a667d745SSouptick Joarder * @vma: user vma to map to 2177a667d745SSouptick Joarder * @pages: pointer to array of source kernel pages 2178a667d745SSouptick Joarder * @num: number of pages in page array 2179a667d745SSouptick Joarder * 2180a667d745SSouptick Joarder * Maps an object consisting of @num pages, catering for the user's 2181a667d745SSouptick Joarder * requested vm_pgoff 2182a667d745SSouptick Joarder * 2183a667d745SSouptick Joarder * If we fail to insert any page into the vma, the function will return 2184a667d745SSouptick Joarder * immediately leaving any previously inserted pages present. Callers 2185a667d745SSouptick Joarder * from the mmap handler may immediately return the error as their caller 2186a667d745SSouptick Joarder * will destroy the vma, removing any successfully inserted pages. Other 2187a667d745SSouptick Joarder * callers should make their own arrangements for calling unmap_region(). 2188a667d745SSouptick Joarder * 2189a667d745SSouptick Joarder * Context: Process context. Called by mmap handlers. 2190a667d745SSouptick Joarder * Return: 0 on success and error code otherwise. 2191a667d745SSouptick Joarder */ 2192a667d745SSouptick Joarder int vm_map_pages(struct vm_area_struct *vma, struct page **pages, 2193a667d745SSouptick Joarder unsigned long num) 2194a667d745SSouptick Joarder { 2195a667d745SSouptick Joarder return __vm_map_pages(vma, pages, num, vma->vm_pgoff); 2196a667d745SSouptick Joarder } 2197a667d745SSouptick Joarder EXPORT_SYMBOL(vm_map_pages); 2198a667d745SSouptick Joarder 2199a667d745SSouptick Joarder /** 2200a667d745SSouptick Joarder * vm_map_pages_zero - map range of kernel pages starts with zero offset 2201a667d745SSouptick Joarder * @vma: user vma to map to 2202a667d745SSouptick Joarder * @pages: pointer to array of source kernel pages 2203a667d745SSouptick Joarder * @num: number of pages in page array 2204a667d745SSouptick Joarder * 2205a667d745SSouptick Joarder * Similar to vm_map_pages(), except that it explicitly sets the offset 2206a667d745SSouptick Joarder * to 0. This function is intended for the drivers that did not consider 2207a667d745SSouptick Joarder * vm_pgoff. 2208a667d745SSouptick Joarder * 2209a667d745SSouptick Joarder * Context: Process context. Called by mmap handlers. 2210a667d745SSouptick Joarder * Return: 0 on success and error code otherwise. 2211a667d745SSouptick Joarder */ 2212a667d745SSouptick Joarder int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages, 2213a667d745SSouptick Joarder unsigned long num) 2214a667d745SSouptick Joarder { 2215a667d745SSouptick Joarder return __vm_map_pages(vma, pages, num, 0); 2216a667d745SSouptick Joarder } 2217a667d745SSouptick Joarder EXPORT_SYMBOL(vm_map_pages_zero); 2218a667d745SSouptick Joarder 22199b5a8e00SMatthew Wilcox static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr, 2220b2770da6SRoss Zwisler pfn_t pfn, pgprot_t prot, bool mkwrite) 2221423bad60SNick Piggin { 2222423bad60SNick Piggin struct mm_struct *mm = vma->vm_mm; 2223423bad60SNick Piggin pte_t *pte, entry; 2224423bad60SNick Piggin spinlock_t *ptl; 2225423bad60SNick Piggin 2226423bad60SNick Piggin pte = get_locked_pte(mm, addr, &ptl); 2227423bad60SNick Piggin if (!pte) 22289b5a8e00SMatthew Wilcox return VM_FAULT_OOM; 2229c33c7948SRyan Roberts entry = ptep_get(pte); 2230c33c7948SRyan Roberts if (!pte_none(entry)) { 2231b2770da6SRoss Zwisler if (mkwrite) { 2232b2770da6SRoss Zwisler /* 2233b2770da6SRoss Zwisler * For read faults on private mappings the PFN passed 2234b2770da6SRoss Zwisler * in may not match the PFN we have mapped if the 2235b2770da6SRoss Zwisler * mapped PFN is a writeable COW page. In the mkwrite 2236b2770da6SRoss Zwisler * case we are creating a writable PTE for a shared 2237f2c57d91SJan Kara * mapping and we expect the PFNs to match. If they 2238f2c57d91SJan Kara * don't match, we are likely racing with block 2239f2c57d91SJan Kara * allocation and mapping invalidation so just skip the 2240f2c57d91SJan Kara * update. 2241b2770da6SRoss Zwisler */ 2242c33c7948SRyan Roberts if (pte_pfn(entry) != pfn_t_to_pfn(pfn)) { 2243c33c7948SRyan Roberts WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry))); 2244423bad60SNick Piggin goto out_unlock; 2245f2c57d91SJan Kara } 2246c33c7948SRyan Roberts entry = pte_mkyoung(entry); 2247cae85cb8SJan Kara entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2248cae85cb8SJan Kara if (ptep_set_access_flags(vma, addr, pte, entry, 1)) 2249cae85cb8SJan Kara update_mmu_cache(vma, addr, pte); 2250cae85cb8SJan Kara } 2251b2770da6SRoss Zwisler goto out_unlock; 2252b2770da6SRoss Zwisler } 2253423bad60SNick Piggin 2254423bad60SNick Piggin /* Ok, finally just insert the thing.. */ 225501c8f1c4SDan Williams if (pfn_t_devmap(pfn)) 225601c8f1c4SDan Williams entry = pte_mkdevmap(pfn_t_pte(pfn, prot)); 225701c8f1c4SDan Williams else 225801c8f1c4SDan Williams entry = pte_mkspecial(pfn_t_pte(pfn, prot)); 2259b2770da6SRoss Zwisler 2260b2770da6SRoss Zwisler if (mkwrite) { 2261b2770da6SRoss Zwisler entry = pte_mkyoung(entry); 2262b2770da6SRoss Zwisler entry = maybe_mkwrite(pte_mkdirty(entry), vma); 2263b2770da6SRoss Zwisler } 2264b2770da6SRoss Zwisler 2265423bad60SNick Piggin set_pte_at(mm, addr, pte, entry); 22664b3073e1SRussell King update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */ 2267423bad60SNick Piggin 2268423bad60SNick Piggin out_unlock: 2269423bad60SNick Piggin pte_unmap_unlock(pte, ptl); 22709b5a8e00SMatthew Wilcox return VM_FAULT_NOPAGE; 2271423bad60SNick Piggin } 2272423bad60SNick Piggin 2273f5e6d1d5SMatthew Wilcox /** 2274f5e6d1d5SMatthew Wilcox * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot 2275f5e6d1d5SMatthew Wilcox * @vma: user vma to map to 2276f5e6d1d5SMatthew Wilcox * @addr: target user address of this page 2277f5e6d1d5SMatthew Wilcox * @pfn: source kernel pfn 2278f5e6d1d5SMatthew Wilcox * @pgprot: pgprot flags for the inserted page 2279f5e6d1d5SMatthew Wilcox * 2280a1a0aea5SRandy Dunlap * This is exactly like vmf_insert_pfn(), except that it allows drivers 2281f5e6d1d5SMatthew Wilcox * to override pgprot on a per-page basis. 2282f5e6d1d5SMatthew Wilcox * 2283f5e6d1d5SMatthew Wilcox * This only makes sense for IO mappings, and it makes no sense for 2284f5e6d1d5SMatthew Wilcox * COW mappings. In general, using multiple vmas is preferable; 2285ae2b01f3SMatthew Wilcox * vmf_insert_pfn_prot should only be used if using multiple VMAs is 2286f5e6d1d5SMatthew Wilcox * impractical. 2287f5e6d1d5SMatthew Wilcox * 228828d8b812SLorenzo Stoakes * pgprot typically only differs from @vma->vm_page_prot when drivers set 228928d8b812SLorenzo Stoakes * caching- and encryption bits different than those of @vma->vm_page_prot, 229028d8b812SLorenzo Stoakes * because the caching- or encryption mode may not be known at mmap() time. 229128d8b812SLorenzo Stoakes * 229228d8b812SLorenzo Stoakes * This is ok as long as @vma->vm_page_prot is not used by the core vm 229328d8b812SLorenzo Stoakes * to set caching and encryption bits for those vmas (except for COW pages). 229428d8b812SLorenzo Stoakes * This is ensured by core vm only modifying these page table entries using 229528d8b812SLorenzo Stoakes * functions that don't touch caching- or encryption bits, using pte_modify() 229628d8b812SLorenzo Stoakes * if needed. (See for example mprotect()). 229728d8b812SLorenzo Stoakes * 229828d8b812SLorenzo Stoakes * Also when new page-table entries are created, this is only done using the 229928d8b812SLorenzo Stoakes * fault() callback, and never using the value of vma->vm_page_prot, 230028d8b812SLorenzo Stoakes * except for page-table entries that point to anonymous pages as the result 230128d8b812SLorenzo Stoakes * of COW. 2302574c5b3dSThomas Hellstrom * 2303ae2b01f3SMatthew Wilcox * Context: Process context. May allocate using %GFP_KERNEL. 2304f5e6d1d5SMatthew Wilcox * Return: vm_fault_t value. 2305f5e6d1d5SMatthew Wilcox */ 2306f5e6d1d5SMatthew Wilcox vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr, 2307f5e6d1d5SMatthew Wilcox unsigned long pfn, pgprot_t pgprot) 2308f5e6d1d5SMatthew Wilcox { 23096d958546SMatthew Wilcox /* 23106d958546SMatthew Wilcox * Technically, architectures with pte_special can avoid all these 23116d958546SMatthew Wilcox * restrictions (same for remap_pfn_range). However we would like 23126d958546SMatthew Wilcox * consistency in testing and feature parity among all, so we should 23136d958546SMatthew Wilcox * try to keep these invariants in place for everybody. 23146d958546SMatthew Wilcox */ 23156d958546SMatthew Wilcox BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); 23166d958546SMatthew Wilcox BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == 23176d958546SMatthew Wilcox (VM_PFNMAP|VM_MIXEDMAP)); 23186d958546SMatthew Wilcox BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); 23196d958546SMatthew Wilcox BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); 23206d958546SMatthew Wilcox 23216d958546SMatthew Wilcox if (addr < vma->vm_start || addr >= vma->vm_end) 23226d958546SMatthew Wilcox return VM_FAULT_SIGBUS; 23236d958546SMatthew Wilcox 23246d958546SMatthew Wilcox if (!pfn_modify_allowed(pfn, pgprot)) 23256d958546SMatthew Wilcox return VM_FAULT_SIGBUS; 23266d958546SMatthew Wilcox 23276d958546SMatthew Wilcox track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)); 23286d958546SMatthew Wilcox 23299b5a8e00SMatthew Wilcox return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot, 23306d958546SMatthew Wilcox false); 2331f5e6d1d5SMatthew Wilcox } 2332f5e6d1d5SMatthew Wilcox EXPORT_SYMBOL(vmf_insert_pfn_prot); 2333e0dc0d8fSNick Piggin 2334ae2b01f3SMatthew Wilcox /** 2335ae2b01f3SMatthew Wilcox * vmf_insert_pfn - insert single pfn into user vma 2336ae2b01f3SMatthew Wilcox * @vma: user vma to map to 2337ae2b01f3SMatthew Wilcox * @addr: target user address of this page 2338ae2b01f3SMatthew Wilcox * @pfn: source kernel pfn 2339ae2b01f3SMatthew Wilcox * 2340ae2b01f3SMatthew Wilcox * Similar to vm_insert_page, this allows drivers to insert individual pages 2341ae2b01f3SMatthew Wilcox * they've allocated into a user vma. Same comments apply. 2342ae2b01f3SMatthew Wilcox * 2343ae2b01f3SMatthew Wilcox * This function should only be called from a vm_ops->fault handler, and 2344ae2b01f3SMatthew Wilcox * in that case the handler should return the result of this function. 2345ae2b01f3SMatthew Wilcox * 2346ae2b01f3SMatthew Wilcox * vma cannot be a COW mapping. 2347ae2b01f3SMatthew Wilcox * 2348ae2b01f3SMatthew Wilcox * As this is called only for pages that do not currently exist, we 2349ae2b01f3SMatthew Wilcox * do not need to flush old virtual caches or the TLB. 2350ae2b01f3SMatthew Wilcox * 2351ae2b01f3SMatthew Wilcox * Context: Process context. May allocate using %GFP_KERNEL. 2352ae2b01f3SMatthew Wilcox * Return: vm_fault_t value. 2353ae2b01f3SMatthew Wilcox */ 2354ae2b01f3SMatthew Wilcox vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr, 2355ae2b01f3SMatthew Wilcox unsigned long pfn) 2356ae2b01f3SMatthew Wilcox { 2357ae2b01f3SMatthew Wilcox return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot); 2358ae2b01f3SMatthew Wilcox } 2359ae2b01f3SMatthew Wilcox EXPORT_SYMBOL(vmf_insert_pfn); 2360ae2b01f3SMatthew Wilcox 2361785a3fabSDan Williams static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn) 2362785a3fabSDan Williams { 2363785a3fabSDan Williams /* these checks mirror the abort conditions in vm_normal_page */ 2364785a3fabSDan Williams if (vma->vm_flags & VM_MIXEDMAP) 2365785a3fabSDan Williams return true; 2366785a3fabSDan Williams if (pfn_t_devmap(pfn)) 2367785a3fabSDan Williams return true; 2368785a3fabSDan Williams if (pfn_t_special(pfn)) 2369785a3fabSDan Williams return true; 2370785a3fabSDan Williams if (is_zero_pfn(pfn_t_to_pfn(pfn))) 2371785a3fabSDan Williams return true; 2372785a3fabSDan Williams return false; 2373785a3fabSDan Williams } 2374785a3fabSDan Williams 237579f3aa5bSMatthew Wilcox static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma, 237628d8b812SLorenzo Stoakes unsigned long addr, pfn_t pfn, bool mkwrite) 2377423bad60SNick Piggin { 237828d8b812SLorenzo Stoakes pgprot_t pgprot = vma->vm_page_prot; 237979f3aa5bSMatthew Wilcox int err; 238087744ab3SDan Williams 2381785a3fabSDan Williams BUG_ON(!vm_mixed_ok(vma, pfn)); 2382423bad60SNick Piggin 2383423bad60SNick Piggin if (addr < vma->vm_start || addr >= vma->vm_end) 238479f3aa5bSMatthew Wilcox return VM_FAULT_SIGBUS; 2385308a047cSBorislav Petkov 2386308a047cSBorislav Petkov track_pfn_insert(vma, &pgprot, pfn); 2387423bad60SNick Piggin 238842e4089cSAndi Kleen if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot)) 238979f3aa5bSMatthew Wilcox return VM_FAULT_SIGBUS; 239042e4089cSAndi Kleen 2391423bad60SNick Piggin /* 2392423bad60SNick Piggin * If we don't have pte special, then we have to use the pfn_valid() 2393423bad60SNick Piggin * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must* 2394423bad60SNick Piggin * refcount the page if pfn_valid is true (hence insert_page rather 239562eede62SHugh Dickins * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP 239662eede62SHugh Dickins * without pte special, it would there be refcounted as a normal page. 2397423bad60SNick Piggin */ 239800b3a331SLaurent Dufour if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && 239900b3a331SLaurent Dufour !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) { 2400423bad60SNick Piggin struct page *page; 2401423bad60SNick Piggin 240203fc2da6SDan Williams /* 240303fc2da6SDan Williams * At this point we are committed to insert_page() 240403fc2da6SDan Williams * regardless of whether the caller specified flags that 240503fc2da6SDan Williams * result in pfn_t_has_page() == false. 240603fc2da6SDan Williams */ 240703fc2da6SDan Williams page = pfn_to_page(pfn_t_to_pfn(pfn)); 240879f3aa5bSMatthew Wilcox err = insert_page(vma, addr, page, pgprot); 240979f3aa5bSMatthew Wilcox } else { 24109b5a8e00SMatthew Wilcox return insert_pfn(vma, addr, pfn, pgprot, mkwrite); 2411423bad60SNick Piggin } 2412b2770da6SRoss Zwisler 24135d747637SMatthew Wilcox if (err == -ENOMEM) 24145d747637SMatthew Wilcox return VM_FAULT_OOM; 24155d747637SMatthew Wilcox if (err < 0 && err != -EBUSY) 24165d747637SMatthew Wilcox return VM_FAULT_SIGBUS; 24175d747637SMatthew Wilcox 24185d747637SMatthew Wilcox return VM_FAULT_NOPAGE; 2419423bad60SNick Piggin } 242079f3aa5bSMatthew Wilcox 242179f3aa5bSMatthew Wilcox vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr, 242279f3aa5bSMatthew Wilcox pfn_t pfn) 242379f3aa5bSMatthew Wilcox { 242428d8b812SLorenzo Stoakes return __vm_insert_mixed(vma, addr, pfn, false); 242579f3aa5bSMatthew Wilcox } 24265d747637SMatthew Wilcox EXPORT_SYMBOL(vmf_insert_mixed); 2427423bad60SNick Piggin 2428ab77dab4SSouptick Joarder /* 2429ab77dab4SSouptick Joarder * If the insertion of PTE failed because someone else already added a 2430ab77dab4SSouptick Joarder * different entry in the mean time, we treat that as success as we assume 2431ab77dab4SSouptick Joarder * the same entry was actually inserted. 2432ab77dab4SSouptick Joarder */ 2433ab77dab4SSouptick Joarder vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma, 2434ab77dab4SSouptick Joarder unsigned long addr, pfn_t pfn) 2435b2770da6SRoss Zwisler { 243628d8b812SLorenzo Stoakes return __vm_insert_mixed(vma, addr, pfn, true); 2437b2770da6SRoss Zwisler } 2438ab77dab4SSouptick Joarder EXPORT_SYMBOL(vmf_insert_mixed_mkwrite); 2439b2770da6SRoss Zwisler 2440a145dd41SLinus Torvalds /* 24411da177e4SLinus Torvalds * maps a range of physical memory into the requested pages. the old 24421da177e4SLinus Torvalds * mappings are removed. any references to nonexistent pages results 24431da177e4SLinus Torvalds * in null mappings (currently treated as "copy-on-access") 24441da177e4SLinus Torvalds */ 24451da177e4SLinus Torvalds static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd, 24461da177e4SLinus Torvalds unsigned long addr, unsigned long end, 24471da177e4SLinus Torvalds unsigned long pfn, pgprot_t prot) 24481da177e4SLinus Torvalds { 244990a3e375SMiaohe Lin pte_t *pte, *mapped_pte; 2450c74df32cSHugh Dickins spinlock_t *ptl; 245142e4089cSAndi Kleen int err = 0; 24521da177e4SLinus Torvalds 245390a3e375SMiaohe Lin mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); 24541da177e4SLinus Torvalds if (!pte) 24551da177e4SLinus Torvalds return -ENOMEM; 24566606c3e0SZachary Amsden arch_enter_lazy_mmu_mode(); 24571da177e4SLinus Torvalds do { 2458c33c7948SRyan Roberts BUG_ON(!pte_none(ptep_get(pte))); 245942e4089cSAndi Kleen if (!pfn_modify_allowed(pfn, prot)) { 246042e4089cSAndi Kleen err = -EACCES; 246142e4089cSAndi Kleen break; 246242e4089cSAndi Kleen } 24637e675137SNick Piggin set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot))); 24641da177e4SLinus Torvalds pfn++; 24651da177e4SLinus Torvalds } while (pte++, addr += PAGE_SIZE, addr != end); 24666606c3e0SZachary Amsden arch_leave_lazy_mmu_mode(); 246790a3e375SMiaohe Lin pte_unmap_unlock(mapped_pte, ptl); 246842e4089cSAndi Kleen return err; 24691da177e4SLinus Torvalds } 24701da177e4SLinus Torvalds 24711da177e4SLinus Torvalds static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud, 24721da177e4SLinus Torvalds unsigned long addr, unsigned long end, 24731da177e4SLinus Torvalds unsigned long pfn, pgprot_t prot) 24741da177e4SLinus Torvalds { 24751da177e4SLinus Torvalds pmd_t *pmd; 24761da177e4SLinus Torvalds unsigned long next; 247742e4089cSAndi Kleen int err; 24781da177e4SLinus Torvalds 24791da177e4SLinus Torvalds pfn -= addr >> PAGE_SHIFT; 24801da177e4SLinus Torvalds pmd = pmd_alloc(mm, pud, addr); 24811da177e4SLinus Torvalds if (!pmd) 24821da177e4SLinus Torvalds return -ENOMEM; 2483f66055abSAndrea Arcangeli VM_BUG_ON(pmd_trans_huge(*pmd)); 24841da177e4SLinus Torvalds do { 24851da177e4SLinus Torvalds next = pmd_addr_end(addr, end); 248642e4089cSAndi Kleen err = remap_pte_range(mm, pmd, addr, next, 248742e4089cSAndi Kleen pfn + (addr >> PAGE_SHIFT), prot); 248842e4089cSAndi Kleen if (err) 248942e4089cSAndi Kleen return err; 24901da177e4SLinus Torvalds } while (pmd++, addr = next, addr != end); 24911da177e4SLinus Torvalds return 0; 24921da177e4SLinus Torvalds } 24931da177e4SLinus Torvalds 2494c2febafcSKirill A. Shutemov static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d, 24951da177e4SLinus Torvalds unsigned long addr, unsigned long end, 24961da177e4SLinus Torvalds unsigned long pfn, pgprot_t prot) 24971da177e4SLinus Torvalds { 24981da177e4SLinus Torvalds pud_t *pud; 24991da177e4SLinus Torvalds unsigned long next; 250042e4089cSAndi Kleen int err; 25011da177e4SLinus Torvalds 25021da177e4SLinus Torvalds pfn -= addr >> PAGE_SHIFT; 2503c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 25041da177e4SLinus Torvalds if (!pud) 25051da177e4SLinus Torvalds return -ENOMEM; 25061da177e4SLinus Torvalds do { 25071da177e4SLinus Torvalds next = pud_addr_end(addr, end); 250842e4089cSAndi Kleen err = remap_pmd_range(mm, pud, addr, next, 250942e4089cSAndi Kleen pfn + (addr >> PAGE_SHIFT), prot); 251042e4089cSAndi Kleen if (err) 251142e4089cSAndi Kleen return err; 25121da177e4SLinus Torvalds } while (pud++, addr = next, addr != end); 25131da177e4SLinus Torvalds return 0; 25141da177e4SLinus Torvalds } 25151da177e4SLinus Torvalds 2516c2febafcSKirill A. Shutemov static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd, 2517c2febafcSKirill A. Shutemov unsigned long addr, unsigned long end, 2518c2febafcSKirill A. Shutemov unsigned long pfn, pgprot_t prot) 2519c2febafcSKirill A. Shutemov { 2520c2febafcSKirill A. Shutemov p4d_t *p4d; 2521c2febafcSKirill A. Shutemov unsigned long next; 252242e4089cSAndi Kleen int err; 2523c2febafcSKirill A. Shutemov 2524c2febafcSKirill A. Shutemov pfn -= addr >> PAGE_SHIFT; 2525c2febafcSKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 2526c2febafcSKirill A. Shutemov if (!p4d) 2527c2febafcSKirill A. Shutemov return -ENOMEM; 2528c2febafcSKirill A. Shutemov do { 2529c2febafcSKirill A. Shutemov next = p4d_addr_end(addr, end); 253042e4089cSAndi Kleen err = remap_pud_range(mm, p4d, addr, next, 253142e4089cSAndi Kleen pfn + (addr >> PAGE_SHIFT), prot); 253242e4089cSAndi Kleen if (err) 253342e4089cSAndi Kleen return err; 2534c2febafcSKirill A. Shutemov } while (p4d++, addr = next, addr != end); 2535c2febafcSKirill A. Shutemov return 0; 2536c2febafcSKirill A. Shutemov } 2537c2febafcSKirill A. Shutemov 253874ffa5a3SChristoph Hellwig /* 253974ffa5a3SChristoph Hellwig * Variant of remap_pfn_range that does not call track_pfn_remap. The caller 254074ffa5a3SChristoph Hellwig * must have pre-validated the caching bits of the pgprot_t. 2541bfa5bf6dSRolf Eike Beer */ 254274ffa5a3SChristoph Hellwig int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr, 25431da177e4SLinus Torvalds unsigned long pfn, unsigned long size, pgprot_t prot) 25441da177e4SLinus Torvalds { 25451da177e4SLinus Torvalds pgd_t *pgd; 25461da177e4SLinus Torvalds unsigned long next; 25472d15cab8SHugh Dickins unsigned long end = addr + PAGE_ALIGN(size); 25481da177e4SLinus Torvalds struct mm_struct *mm = vma->vm_mm; 25491da177e4SLinus Torvalds int err; 25501da177e4SLinus Torvalds 25510c4123e3SAlex Zhang if (WARN_ON_ONCE(!PAGE_ALIGNED(addr))) 25520c4123e3SAlex Zhang return -EINVAL; 25530c4123e3SAlex Zhang 25541da177e4SLinus Torvalds /* 25551da177e4SLinus Torvalds * Physically remapped pages are special. Tell the 25561da177e4SLinus Torvalds * rest of the world about it: 25571da177e4SLinus Torvalds * VM_IO tells people not to look at these pages 25581da177e4SLinus Torvalds * (accesses can have side effects). 25596aab341eSLinus Torvalds * VM_PFNMAP tells the core MM that the base pages are just 25606aab341eSLinus Torvalds * raw PFN mappings, and do not have a "struct page" associated 25616aab341eSLinus Torvalds * with them. 2562314e51b9SKonstantin Khlebnikov * VM_DONTEXPAND 2563314e51b9SKonstantin Khlebnikov * Disable vma merging and expanding with mremap(). 2564314e51b9SKonstantin Khlebnikov * VM_DONTDUMP 2565314e51b9SKonstantin Khlebnikov * Omit vma from core dump, even when VM_IO turned off. 2566fb155c16SLinus Torvalds * 2567fb155c16SLinus Torvalds * There's a horrible special case to handle copy-on-write 2568fb155c16SLinus Torvalds * behaviour that some programs depend on. We mark the "original" 2569fb155c16SLinus Torvalds * un-COW'ed pages by matching them up with "vma->vm_pgoff". 2570b3b9c293SKonstantin Khlebnikov * See vm_normal_page() for details. 25711da177e4SLinus Torvalds */ 2572b3b9c293SKonstantin Khlebnikov if (is_cow_mapping(vma->vm_flags)) { 2573b3b9c293SKonstantin Khlebnikov if (addr != vma->vm_start || end != vma->vm_end) 2574b3b9c293SKonstantin Khlebnikov return -EINVAL; 25756aab341eSLinus Torvalds vma->vm_pgoff = pfn; 2576b3b9c293SKonstantin Khlebnikov } 2577b3b9c293SKonstantin Khlebnikov 25781c71222eSSuren Baghdasaryan vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP); 25791da177e4SLinus Torvalds 25801da177e4SLinus Torvalds BUG_ON(addr >= end); 25811da177e4SLinus Torvalds pfn -= addr >> PAGE_SHIFT; 25821da177e4SLinus Torvalds pgd = pgd_offset(mm, addr); 25831da177e4SLinus Torvalds flush_cache_range(vma, addr, end); 25841da177e4SLinus Torvalds do { 25851da177e4SLinus Torvalds next = pgd_addr_end(addr, end); 2586c2febafcSKirill A. Shutemov err = remap_p4d_range(mm, pgd, addr, next, 25871da177e4SLinus Torvalds pfn + (addr >> PAGE_SHIFT), prot); 25881da177e4SLinus Torvalds if (err) 258974ffa5a3SChristoph Hellwig return err; 25901da177e4SLinus Torvalds } while (pgd++, addr = next, addr != end); 25912ab64037Svenkatesh.pallipadi@intel.com 259274ffa5a3SChristoph Hellwig return 0; 259374ffa5a3SChristoph Hellwig } 25942ab64037Svenkatesh.pallipadi@intel.com 259574ffa5a3SChristoph Hellwig /** 259674ffa5a3SChristoph Hellwig * remap_pfn_range - remap kernel memory to userspace 259774ffa5a3SChristoph Hellwig * @vma: user vma to map to 259874ffa5a3SChristoph Hellwig * @addr: target page aligned user address to start at 259974ffa5a3SChristoph Hellwig * @pfn: page frame number of kernel physical memory address 260074ffa5a3SChristoph Hellwig * @size: size of mapping area 260174ffa5a3SChristoph Hellwig * @prot: page protection flags for this mapping 260274ffa5a3SChristoph Hellwig * 260374ffa5a3SChristoph Hellwig * Note: this is only safe if the mm semaphore is held when called. 260474ffa5a3SChristoph Hellwig * 260574ffa5a3SChristoph Hellwig * Return: %0 on success, negative error code otherwise. 260674ffa5a3SChristoph Hellwig */ 260774ffa5a3SChristoph Hellwig int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr, 260874ffa5a3SChristoph Hellwig unsigned long pfn, unsigned long size, pgprot_t prot) 260974ffa5a3SChristoph Hellwig { 261074ffa5a3SChristoph Hellwig int err; 261174ffa5a3SChristoph Hellwig 261274ffa5a3SChristoph Hellwig err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size)); 261374ffa5a3SChristoph Hellwig if (err) 261474ffa5a3SChristoph Hellwig return -EINVAL; 261574ffa5a3SChristoph Hellwig 261674ffa5a3SChristoph Hellwig err = remap_pfn_range_notrack(vma, addr, pfn, size, prot); 261774ffa5a3SChristoph Hellwig if (err) 261868f48381SSuren Baghdasaryan untrack_pfn(vma, pfn, PAGE_ALIGN(size), true); 26191da177e4SLinus Torvalds return err; 26201da177e4SLinus Torvalds } 26211da177e4SLinus Torvalds EXPORT_SYMBOL(remap_pfn_range); 26221da177e4SLinus Torvalds 2623b4cbb197SLinus Torvalds /** 2624b4cbb197SLinus Torvalds * vm_iomap_memory - remap memory to userspace 2625b4cbb197SLinus Torvalds * @vma: user vma to map to 2626abd69b9eSWang Wenhu * @start: start of the physical memory to be mapped 2627b4cbb197SLinus Torvalds * @len: size of area 2628b4cbb197SLinus Torvalds * 2629b4cbb197SLinus Torvalds * This is a simplified io_remap_pfn_range() for common driver use. The 2630b4cbb197SLinus Torvalds * driver just needs to give us the physical memory range to be mapped, 2631b4cbb197SLinus Torvalds * we'll figure out the rest from the vma information. 2632b4cbb197SLinus Torvalds * 2633b4cbb197SLinus Torvalds * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get 2634b4cbb197SLinus Torvalds * whatever write-combining details or similar. 2635a862f68aSMike Rapoport * 2636a862f68aSMike Rapoport * Return: %0 on success, negative error code otherwise. 2637b4cbb197SLinus Torvalds */ 2638b4cbb197SLinus Torvalds int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len) 2639b4cbb197SLinus Torvalds { 2640b4cbb197SLinus Torvalds unsigned long vm_len, pfn, pages; 2641b4cbb197SLinus Torvalds 2642b4cbb197SLinus Torvalds /* Check that the physical memory area passed in looks valid */ 2643b4cbb197SLinus Torvalds if (start + len < start) 2644b4cbb197SLinus Torvalds return -EINVAL; 2645b4cbb197SLinus Torvalds /* 2646b4cbb197SLinus Torvalds * You *really* shouldn't map things that aren't page-aligned, 2647b4cbb197SLinus Torvalds * but we've historically allowed it because IO memory might 2648b4cbb197SLinus Torvalds * just have smaller alignment. 2649b4cbb197SLinus Torvalds */ 2650b4cbb197SLinus Torvalds len += start & ~PAGE_MASK; 2651b4cbb197SLinus Torvalds pfn = start >> PAGE_SHIFT; 2652b4cbb197SLinus Torvalds pages = (len + ~PAGE_MASK) >> PAGE_SHIFT; 2653b4cbb197SLinus Torvalds if (pfn + pages < pfn) 2654b4cbb197SLinus Torvalds return -EINVAL; 2655b4cbb197SLinus Torvalds 2656b4cbb197SLinus Torvalds /* We start the mapping 'vm_pgoff' pages into the area */ 2657b4cbb197SLinus Torvalds if (vma->vm_pgoff > pages) 2658b4cbb197SLinus Torvalds return -EINVAL; 2659b4cbb197SLinus Torvalds pfn += vma->vm_pgoff; 2660b4cbb197SLinus Torvalds pages -= vma->vm_pgoff; 2661b4cbb197SLinus Torvalds 2662b4cbb197SLinus Torvalds /* Can we fit all of the mapping? */ 2663b4cbb197SLinus Torvalds vm_len = vma->vm_end - vma->vm_start; 2664b4cbb197SLinus Torvalds if (vm_len >> PAGE_SHIFT > pages) 2665b4cbb197SLinus Torvalds return -EINVAL; 2666b4cbb197SLinus Torvalds 2667b4cbb197SLinus Torvalds /* Ok, let it rip */ 2668b4cbb197SLinus Torvalds return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot); 2669b4cbb197SLinus Torvalds } 2670b4cbb197SLinus Torvalds EXPORT_SYMBOL(vm_iomap_memory); 2671b4cbb197SLinus Torvalds 2672aee16b3cSJeremy Fitzhardinge static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd, 2673aee16b3cSJeremy Fitzhardinge unsigned long addr, unsigned long end, 2674e80d3909SJoerg Roedel pte_fn_t fn, void *data, bool create, 2675e80d3909SJoerg Roedel pgtbl_mod_mask *mask) 2676aee16b3cSJeremy Fitzhardinge { 26778abb50c7SMiaohe Lin pte_t *pte, *mapped_pte; 2678be1db475SDaniel Axtens int err = 0; 26793f649ab7SKees Cook spinlock_t *ptl; 2680aee16b3cSJeremy Fitzhardinge 2681be1db475SDaniel Axtens if (create) { 26828abb50c7SMiaohe Lin mapped_pte = pte = (mm == &init_mm) ? 2683e80d3909SJoerg Roedel pte_alloc_kernel_track(pmd, addr, mask) : 2684aee16b3cSJeremy Fitzhardinge pte_alloc_map_lock(mm, pmd, addr, &ptl); 2685aee16b3cSJeremy Fitzhardinge if (!pte) 2686aee16b3cSJeremy Fitzhardinge return -ENOMEM; 2687be1db475SDaniel Axtens } else { 26888abb50c7SMiaohe Lin mapped_pte = pte = (mm == &init_mm) ? 2689be1db475SDaniel Axtens pte_offset_kernel(pmd, addr) : 2690be1db475SDaniel Axtens pte_offset_map_lock(mm, pmd, addr, &ptl); 26913db82b93SHugh Dickins if (!pte) 26923db82b93SHugh Dickins return -EINVAL; 2693be1db475SDaniel Axtens } 2694aee16b3cSJeremy Fitzhardinge 269538e0edb1SJeremy Fitzhardinge arch_enter_lazy_mmu_mode(); 269638e0edb1SJeremy Fitzhardinge 2697eeb4a05fSChristoph Hellwig if (fn) { 2698aee16b3cSJeremy Fitzhardinge do { 2699c33c7948SRyan Roberts if (create || !pte_none(ptep_get(pte))) { 27008b1e0f81SAnshuman Khandual err = fn(pte++, addr, data); 2701aee16b3cSJeremy Fitzhardinge if (err) 2702aee16b3cSJeremy Fitzhardinge break; 2703be1db475SDaniel Axtens } 2704c36987e2SDaisuke Nishimura } while (addr += PAGE_SIZE, addr != end); 2705eeb4a05fSChristoph Hellwig } 2706e80d3909SJoerg Roedel *mask |= PGTBL_PTE_MODIFIED; 2707aee16b3cSJeremy Fitzhardinge 270838e0edb1SJeremy Fitzhardinge arch_leave_lazy_mmu_mode(); 270938e0edb1SJeremy Fitzhardinge 2710aee16b3cSJeremy Fitzhardinge if (mm != &init_mm) 27118abb50c7SMiaohe Lin pte_unmap_unlock(mapped_pte, ptl); 2712aee16b3cSJeremy Fitzhardinge return err; 2713aee16b3cSJeremy Fitzhardinge } 2714aee16b3cSJeremy Fitzhardinge 2715aee16b3cSJeremy Fitzhardinge static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud, 2716aee16b3cSJeremy Fitzhardinge unsigned long addr, unsigned long end, 2717e80d3909SJoerg Roedel pte_fn_t fn, void *data, bool create, 2718e80d3909SJoerg Roedel pgtbl_mod_mask *mask) 2719aee16b3cSJeremy Fitzhardinge { 2720aee16b3cSJeremy Fitzhardinge pmd_t *pmd; 2721aee16b3cSJeremy Fitzhardinge unsigned long next; 2722be1db475SDaniel Axtens int err = 0; 2723aee16b3cSJeremy Fitzhardinge 2724ceb86879SAndi Kleen BUG_ON(pud_huge(*pud)); 2725ceb86879SAndi Kleen 2726be1db475SDaniel Axtens if (create) { 2727e80d3909SJoerg Roedel pmd = pmd_alloc_track(mm, pud, addr, mask); 2728aee16b3cSJeremy Fitzhardinge if (!pmd) 2729aee16b3cSJeremy Fitzhardinge return -ENOMEM; 2730be1db475SDaniel Axtens } else { 2731be1db475SDaniel Axtens pmd = pmd_offset(pud, addr); 2732be1db475SDaniel Axtens } 2733aee16b3cSJeremy Fitzhardinge do { 2734aee16b3cSJeremy Fitzhardinge next = pmd_addr_end(addr, end); 27350c95cba4SNicholas Piggin if (pmd_none(*pmd) && !create) 27360c95cba4SNicholas Piggin continue; 27370c95cba4SNicholas Piggin if (WARN_ON_ONCE(pmd_leaf(*pmd))) 27380c95cba4SNicholas Piggin return -EINVAL; 27390c95cba4SNicholas Piggin if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) { 27400c95cba4SNicholas Piggin if (!create) 27410c95cba4SNicholas Piggin continue; 27420c95cba4SNicholas Piggin pmd_clear_bad(pmd); 27430c95cba4SNicholas Piggin } 27440c95cba4SNicholas Piggin err = apply_to_pte_range(mm, pmd, addr, next, 27450c95cba4SNicholas Piggin fn, data, create, mask); 2746aee16b3cSJeremy Fitzhardinge if (err) 2747aee16b3cSJeremy Fitzhardinge break; 2748aee16b3cSJeremy Fitzhardinge } while (pmd++, addr = next, addr != end); 27490c95cba4SNicholas Piggin 2750aee16b3cSJeremy Fitzhardinge return err; 2751aee16b3cSJeremy Fitzhardinge } 2752aee16b3cSJeremy Fitzhardinge 2753c2febafcSKirill A. Shutemov static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d, 2754aee16b3cSJeremy Fitzhardinge unsigned long addr, unsigned long end, 2755e80d3909SJoerg Roedel pte_fn_t fn, void *data, bool create, 2756e80d3909SJoerg Roedel pgtbl_mod_mask *mask) 2757aee16b3cSJeremy Fitzhardinge { 2758aee16b3cSJeremy Fitzhardinge pud_t *pud; 2759aee16b3cSJeremy Fitzhardinge unsigned long next; 2760be1db475SDaniel Axtens int err = 0; 2761aee16b3cSJeremy Fitzhardinge 2762be1db475SDaniel Axtens if (create) { 2763e80d3909SJoerg Roedel pud = pud_alloc_track(mm, p4d, addr, mask); 2764aee16b3cSJeremy Fitzhardinge if (!pud) 2765aee16b3cSJeremy Fitzhardinge return -ENOMEM; 2766be1db475SDaniel Axtens } else { 2767be1db475SDaniel Axtens pud = pud_offset(p4d, addr); 2768be1db475SDaniel Axtens } 2769aee16b3cSJeremy Fitzhardinge do { 2770aee16b3cSJeremy Fitzhardinge next = pud_addr_end(addr, end); 27710c95cba4SNicholas Piggin if (pud_none(*pud) && !create) 27720c95cba4SNicholas Piggin continue; 27730c95cba4SNicholas Piggin if (WARN_ON_ONCE(pud_leaf(*pud))) 27740c95cba4SNicholas Piggin return -EINVAL; 27750c95cba4SNicholas Piggin if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) { 27760c95cba4SNicholas Piggin if (!create) 27770c95cba4SNicholas Piggin continue; 27780c95cba4SNicholas Piggin pud_clear_bad(pud); 27790c95cba4SNicholas Piggin } 27800c95cba4SNicholas Piggin err = apply_to_pmd_range(mm, pud, addr, next, 27810c95cba4SNicholas Piggin fn, data, create, mask); 2782aee16b3cSJeremy Fitzhardinge if (err) 2783aee16b3cSJeremy Fitzhardinge break; 2784aee16b3cSJeremy Fitzhardinge } while (pud++, addr = next, addr != end); 27850c95cba4SNicholas Piggin 2786aee16b3cSJeremy Fitzhardinge return err; 2787aee16b3cSJeremy Fitzhardinge } 2788aee16b3cSJeremy Fitzhardinge 2789c2febafcSKirill A. Shutemov static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd, 2790c2febafcSKirill A. Shutemov unsigned long addr, unsigned long end, 2791e80d3909SJoerg Roedel pte_fn_t fn, void *data, bool create, 2792e80d3909SJoerg Roedel pgtbl_mod_mask *mask) 2793c2febafcSKirill A. Shutemov { 2794c2febafcSKirill A. Shutemov p4d_t *p4d; 2795c2febafcSKirill A. Shutemov unsigned long next; 2796be1db475SDaniel Axtens int err = 0; 2797c2febafcSKirill A. Shutemov 2798be1db475SDaniel Axtens if (create) { 2799e80d3909SJoerg Roedel p4d = p4d_alloc_track(mm, pgd, addr, mask); 2800c2febafcSKirill A. Shutemov if (!p4d) 2801c2febafcSKirill A. Shutemov return -ENOMEM; 2802be1db475SDaniel Axtens } else { 2803be1db475SDaniel Axtens p4d = p4d_offset(pgd, addr); 2804be1db475SDaniel Axtens } 2805c2febafcSKirill A. Shutemov do { 2806c2febafcSKirill A. Shutemov next = p4d_addr_end(addr, end); 28070c95cba4SNicholas Piggin if (p4d_none(*p4d) && !create) 28080c95cba4SNicholas Piggin continue; 28090c95cba4SNicholas Piggin if (WARN_ON_ONCE(p4d_leaf(*p4d))) 28100c95cba4SNicholas Piggin return -EINVAL; 28110c95cba4SNicholas Piggin if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) { 28120c95cba4SNicholas Piggin if (!create) 28130c95cba4SNicholas Piggin continue; 28140c95cba4SNicholas Piggin p4d_clear_bad(p4d); 28150c95cba4SNicholas Piggin } 28160c95cba4SNicholas Piggin err = apply_to_pud_range(mm, p4d, addr, next, 28170c95cba4SNicholas Piggin fn, data, create, mask); 2818c2febafcSKirill A. Shutemov if (err) 2819c2febafcSKirill A. Shutemov break; 2820c2febafcSKirill A. Shutemov } while (p4d++, addr = next, addr != end); 28210c95cba4SNicholas Piggin 2822c2febafcSKirill A. Shutemov return err; 2823c2febafcSKirill A. Shutemov } 2824c2febafcSKirill A. Shutemov 2825be1db475SDaniel Axtens static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr, 2826be1db475SDaniel Axtens unsigned long size, pte_fn_t fn, 2827be1db475SDaniel Axtens void *data, bool create) 2828be1db475SDaniel Axtens { 2829be1db475SDaniel Axtens pgd_t *pgd; 2830e80d3909SJoerg Roedel unsigned long start = addr, next; 2831be1db475SDaniel Axtens unsigned long end = addr + size; 2832e80d3909SJoerg Roedel pgtbl_mod_mask mask = 0; 2833be1db475SDaniel Axtens int err = 0; 2834be1db475SDaniel Axtens 2835be1db475SDaniel Axtens if (WARN_ON(addr >= end)) 2836be1db475SDaniel Axtens return -EINVAL; 2837be1db475SDaniel Axtens 2838be1db475SDaniel Axtens pgd = pgd_offset(mm, addr); 2839be1db475SDaniel Axtens do { 2840be1db475SDaniel Axtens next = pgd_addr_end(addr, end); 28410c95cba4SNicholas Piggin if (pgd_none(*pgd) && !create) 2842be1db475SDaniel Axtens continue; 28430c95cba4SNicholas Piggin if (WARN_ON_ONCE(pgd_leaf(*pgd))) 28440c95cba4SNicholas Piggin return -EINVAL; 28450c95cba4SNicholas Piggin if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) { 28460c95cba4SNicholas Piggin if (!create) 28470c95cba4SNicholas Piggin continue; 28480c95cba4SNicholas Piggin pgd_clear_bad(pgd); 28490c95cba4SNicholas Piggin } 28500c95cba4SNicholas Piggin err = apply_to_p4d_range(mm, pgd, addr, next, 28510c95cba4SNicholas Piggin fn, data, create, &mask); 2852be1db475SDaniel Axtens if (err) 2853be1db475SDaniel Axtens break; 2854be1db475SDaniel Axtens } while (pgd++, addr = next, addr != end); 2855be1db475SDaniel Axtens 2856e80d3909SJoerg Roedel if (mask & ARCH_PAGE_TABLE_SYNC_MASK) 2857e80d3909SJoerg Roedel arch_sync_kernel_mappings(start, start + size); 2858e80d3909SJoerg Roedel 2859be1db475SDaniel Axtens return err; 2860be1db475SDaniel Axtens } 2861be1db475SDaniel Axtens 2862aee16b3cSJeremy Fitzhardinge /* 2863aee16b3cSJeremy Fitzhardinge * Scan a region of virtual memory, filling in page tables as necessary 2864aee16b3cSJeremy Fitzhardinge * and calling a provided function on each leaf page table. 2865aee16b3cSJeremy Fitzhardinge */ 2866aee16b3cSJeremy Fitzhardinge int apply_to_page_range(struct mm_struct *mm, unsigned long addr, 2867aee16b3cSJeremy Fitzhardinge unsigned long size, pte_fn_t fn, void *data) 2868aee16b3cSJeremy Fitzhardinge { 2869be1db475SDaniel Axtens return __apply_to_page_range(mm, addr, size, fn, data, true); 2870aee16b3cSJeremy Fitzhardinge } 2871aee16b3cSJeremy Fitzhardinge EXPORT_SYMBOL_GPL(apply_to_page_range); 2872aee16b3cSJeremy Fitzhardinge 28731da177e4SLinus Torvalds /* 2874be1db475SDaniel Axtens * Scan a region of virtual memory, calling a provided function on 2875be1db475SDaniel Axtens * each leaf page table where it exists. 2876be1db475SDaniel Axtens * 2877be1db475SDaniel Axtens * Unlike apply_to_page_range, this does _not_ fill in page tables 2878be1db475SDaniel Axtens * where they are absent. 2879be1db475SDaniel Axtens */ 2880be1db475SDaniel Axtens int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr, 2881be1db475SDaniel Axtens unsigned long size, pte_fn_t fn, void *data) 2882be1db475SDaniel Axtens { 2883be1db475SDaniel Axtens return __apply_to_page_range(mm, addr, size, fn, data, false); 2884be1db475SDaniel Axtens } 2885be1db475SDaniel Axtens EXPORT_SYMBOL_GPL(apply_to_existing_page_range); 2886be1db475SDaniel Axtens 2887be1db475SDaniel Axtens /* 28889b4bdd2fSKirill A. Shutemov * handle_pte_fault chooses page fault handler according to an entry which was 28899b4bdd2fSKirill A. Shutemov * read non-atomically. Before making any commitment, on those architectures 28909b4bdd2fSKirill A. Shutemov * or configurations (e.g. i386 with PAE) which might give a mix of unmatched 28919b4bdd2fSKirill A. Shutemov * parts, do_swap_page must check under lock before unmapping the pte and 28929b4bdd2fSKirill A. Shutemov * proceeding (but do_wp_page is only called after already making such a check; 2893a335b2e1SRyota Ozaki * and do_anonymous_page can safely check later on). 28948f4e2101SHugh Dickins */ 28952ca99358SPeter Xu static inline int pte_unmap_same(struct vm_fault *vmf) 28968f4e2101SHugh Dickins { 28978f4e2101SHugh Dickins int same = 1; 2898923717cbSThomas Gleixner #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION) 28998f4e2101SHugh Dickins if (sizeof(pte_t) > sizeof(unsigned long)) { 2900c7ad0880SHugh Dickins spin_lock(vmf->ptl); 2901c33c7948SRyan Roberts same = pte_same(ptep_get(vmf->pte), vmf->orig_pte); 2902c7ad0880SHugh Dickins spin_unlock(vmf->ptl); 29038f4e2101SHugh Dickins } 29048f4e2101SHugh Dickins #endif 29052ca99358SPeter Xu pte_unmap(vmf->pte); 29062ca99358SPeter Xu vmf->pte = NULL; 29078f4e2101SHugh Dickins return same; 29088f4e2101SHugh Dickins } 29098f4e2101SHugh Dickins 2910a873dfe1STony Luck /* 2911a873dfe1STony Luck * Return: 2912a873dfe1STony Luck * 0: copied succeeded 2913a873dfe1STony Luck * -EHWPOISON: copy failed due to hwpoison in source page 2914a873dfe1STony Luck * -EAGAIN: copied failed (some other reason) 2915a873dfe1STony Luck */ 2916a873dfe1STony Luck static inline int __wp_page_copy_user(struct page *dst, struct page *src, 291783d116c5SJia He struct vm_fault *vmf) 29186aab341eSLinus Torvalds { 2919a873dfe1STony Luck int ret; 292083d116c5SJia He void *kaddr; 292183d116c5SJia He void __user *uaddr; 292283d116c5SJia He struct vm_area_struct *vma = vmf->vma; 292383d116c5SJia He struct mm_struct *mm = vma->vm_mm; 292483d116c5SJia He unsigned long addr = vmf->address; 292583d116c5SJia He 292683d116c5SJia He if (likely(src)) { 2927d302c239STony Luck if (copy_mc_user_highpage(dst, src, addr, vma)) { 2928d302c239STony Luck memory_failure_queue(page_to_pfn(src), 0); 2929a873dfe1STony Luck return -EHWPOISON; 2930d302c239STony Luck } 2931a873dfe1STony Luck return 0; 293283d116c5SJia He } 293383d116c5SJia He 29346aab341eSLinus Torvalds /* 29356aab341eSLinus Torvalds * If the source page was a PFN mapping, we don't have 29366aab341eSLinus Torvalds * a "struct page" for it. We do a best-effort copy by 29376aab341eSLinus Torvalds * just copying from the original user address. If that 29386aab341eSLinus Torvalds * fails, we just zero-fill it. Live with it. 29396aab341eSLinus Torvalds */ 294024d2613aSFabio M. De Francesco kaddr = kmap_local_page(dst); 294124d2613aSFabio M. De Francesco pagefault_disable(); 294283d116c5SJia He uaddr = (void __user *)(addr & PAGE_MASK); 294383d116c5SJia He 294483d116c5SJia He /* 294583d116c5SJia He * On architectures with software "accessed" bits, we would 294683d116c5SJia He * take a double page fault, so mark it accessed here. 294783d116c5SJia He */ 29483db82b93SHugh Dickins vmf->pte = NULL; 2949e1fd09e3SYu Zhao if (!arch_has_hw_pte_young() && !pte_young(vmf->orig_pte)) { 295083d116c5SJia He pte_t entry; 295183d116c5SJia He 295283d116c5SJia He vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl); 2953c33c7948SRyan Roberts if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) { 295483d116c5SJia He /* 295583d116c5SJia He * Other thread has already handled the fault 29567df67697SBibo Mao * and update local tlb only 295783d116c5SJia He */ 2958a92cbb82SHugh Dickins if (vmf->pte) 29597df67697SBibo Mao update_mmu_tlb(vma, addr, vmf->pte); 2960a873dfe1STony Luck ret = -EAGAIN; 296183d116c5SJia He goto pte_unlock; 296283d116c5SJia He } 296383d116c5SJia He 296483d116c5SJia He entry = pte_mkyoung(vmf->orig_pte); 296583d116c5SJia He if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0)) 29665003a2bdSMatthew Wilcox (Oracle) update_mmu_cache_range(vmf, vma, addr, vmf->pte, 1); 296783d116c5SJia He } 29685d2a2dbbSLinus Torvalds 29695d2a2dbbSLinus Torvalds /* 29705d2a2dbbSLinus Torvalds * This really shouldn't fail, because the page is there 29715d2a2dbbSLinus Torvalds * in the page tables. But it might just be unreadable, 29725d2a2dbbSLinus Torvalds * in which case we just give up and fill the result with 29735d2a2dbbSLinus Torvalds * zeroes. 29745d2a2dbbSLinus Torvalds */ 297583d116c5SJia He if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) { 29763db82b93SHugh Dickins if (vmf->pte) 2977c3e5ea6eSKirill A. Shutemov goto warn; 2978c3e5ea6eSKirill A. Shutemov 2979c3e5ea6eSKirill A. Shutemov /* Re-validate under PTL if the page is still mapped */ 2980c3e5ea6eSKirill A. Shutemov vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl); 2981c33c7948SRyan Roberts if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) { 29827df67697SBibo Mao /* The PTE changed under us, update local tlb */ 2983a92cbb82SHugh Dickins if (vmf->pte) 29847df67697SBibo Mao update_mmu_tlb(vma, addr, vmf->pte); 2985a873dfe1STony Luck ret = -EAGAIN; 2986c3e5ea6eSKirill A. Shutemov goto pte_unlock; 2987c3e5ea6eSKirill A. Shutemov } 2988c3e5ea6eSKirill A. Shutemov 2989c3e5ea6eSKirill A. Shutemov /* 2990985ba004SEthon Paul * The same page can be mapped back since last copy attempt. 2991c3e5ea6eSKirill A. Shutemov * Try to copy again under PTL. 2992c3e5ea6eSKirill A. Shutemov */ 2993c3e5ea6eSKirill A. Shutemov if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) { 299483d116c5SJia He /* 299583d116c5SJia He * Give a warn in case there can be some obscure 299683d116c5SJia He * use-case 299783d116c5SJia He */ 2998c3e5ea6eSKirill A. Shutemov warn: 299983d116c5SJia He WARN_ON_ONCE(1); 30003ecb01dfSJan Beulich clear_page(kaddr); 300183d116c5SJia He } 3002c3e5ea6eSKirill A. Shutemov } 300383d116c5SJia He 3004a873dfe1STony Luck ret = 0; 300583d116c5SJia He 300683d116c5SJia He pte_unlock: 30073db82b93SHugh Dickins if (vmf->pte) 300883d116c5SJia He pte_unmap_unlock(vmf->pte, vmf->ptl); 300924d2613aSFabio M. De Francesco pagefault_enable(); 301024d2613aSFabio M. De Francesco kunmap_local(kaddr); 3011c4ec7b0dSDmitriy Monakhov flush_dcache_page(dst); 301283d116c5SJia He 301383d116c5SJia He return ret; 30146aab341eSLinus Torvalds } 30156aab341eSLinus Torvalds 3016c20cd45eSMichal Hocko static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma) 3017c20cd45eSMichal Hocko { 3018c20cd45eSMichal Hocko struct file *vm_file = vma->vm_file; 3019c20cd45eSMichal Hocko 3020c20cd45eSMichal Hocko if (vm_file) 3021c20cd45eSMichal Hocko return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO; 3022c20cd45eSMichal Hocko 3023c20cd45eSMichal Hocko /* 3024c20cd45eSMichal Hocko * Special mappings (e.g. VDSO) do not have any file so fake 3025c20cd45eSMichal Hocko * a default GFP_KERNEL for them. 3026c20cd45eSMichal Hocko */ 3027c20cd45eSMichal Hocko return GFP_KERNEL; 3028c20cd45eSMichal Hocko } 3029c20cd45eSMichal Hocko 30301da177e4SLinus Torvalds /* 3031fb09a464SKirill A. Shutemov * Notify the address space that the page is about to become writable so that 3032fb09a464SKirill A. Shutemov * it can prohibit this or wait for the page to get into an appropriate state. 3033fb09a464SKirill A. Shutemov * 3034fb09a464SKirill A. Shutemov * We do this without the lock held, so that it can sleep if it needs to. 3035fb09a464SKirill A. Shutemov */ 303686aa6998SSidhartha Kumar static vm_fault_t do_page_mkwrite(struct vm_fault *vmf, struct folio *folio) 3037fb09a464SKirill A. Shutemov { 30382b740303SSouptick Joarder vm_fault_t ret; 303938b8cb7fSJan Kara unsigned int old_flags = vmf->flags; 3040fb09a464SKirill A. Shutemov 304138b8cb7fSJan Kara vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE; 3042fb09a464SKirill A. Shutemov 3043dc617f29SDarrick J. Wong if (vmf->vma->vm_file && 3044dc617f29SDarrick J. Wong IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host)) 3045dc617f29SDarrick J. Wong return VM_FAULT_SIGBUS; 3046dc617f29SDarrick J. Wong 304711bac800SDave Jiang ret = vmf->vma->vm_ops->page_mkwrite(vmf); 304838b8cb7fSJan Kara /* Restore original flags so that caller is not surprised */ 304938b8cb7fSJan Kara vmf->flags = old_flags; 3050fb09a464SKirill A. Shutemov if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) 3051fb09a464SKirill A. Shutemov return ret; 3052fb09a464SKirill A. Shutemov if (unlikely(!(ret & VM_FAULT_LOCKED))) { 30533d243659SSidhartha Kumar folio_lock(folio); 30543d243659SSidhartha Kumar if (!folio->mapping) { 30553d243659SSidhartha Kumar folio_unlock(folio); 3056fb09a464SKirill A. Shutemov return 0; /* retry */ 3057fb09a464SKirill A. Shutemov } 3058fb09a464SKirill A. Shutemov ret |= VM_FAULT_LOCKED; 3059fb09a464SKirill A. Shutemov } else 30603d243659SSidhartha Kumar VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 3061fb09a464SKirill A. Shutemov return ret; 3062fb09a464SKirill A. Shutemov } 3063fb09a464SKirill A. Shutemov 3064fb09a464SKirill A. Shutemov /* 306597ba0c2bSJan Kara * Handle dirtying of a page in shared file mapping on a write fault. 30664e047f89SShachar Raindel * 306797ba0c2bSJan Kara * The function expects the page to be locked and unlocks it. 30684e047f89SShachar Raindel */ 306989b15332SJohannes Weiner static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf) 30704e047f89SShachar Raindel { 307189b15332SJohannes Weiner struct vm_area_struct *vma = vmf->vma; 30724e047f89SShachar Raindel struct address_space *mapping; 307315b4919aSZhangPeng struct folio *folio = page_folio(vmf->page); 307497ba0c2bSJan Kara bool dirtied; 307597ba0c2bSJan Kara bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite; 30764e047f89SShachar Raindel 307715b4919aSZhangPeng dirtied = folio_mark_dirty(folio); 307815b4919aSZhangPeng VM_BUG_ON_FOLIO(folio_test_anon(folio), folio); 307997ba0c2bSJan Kara /* 308015b4919aSZhangPeng * Take a local copy of the address_space - folio.mapping may be zeroed 308115b4919aSZhangPeng * by truncate after folio_unlock(). The address_space itself remains 308215b4919aSZhangPeng * pinned by vma->vm_file's reference. We rely on folio_unlock()'s 308397ba0c2bSJan Kara * release semantics to prevent the compiler from undoing this copying. 308497ba0c2bSJan Kara */ 308515b4919aSZhangPeng mapping = folio_raw_mapping(folio); 308615b4919aSZhangPeng folio_unlock(folio); 30874e047f89SShachar Raindel 30884e047f89SShachar Raindel if (!page_mkwrite) 30894e047f89SShachar Raindel file_update_time(vma->vm_file); 309089b15332SJohannes Weiner 309189b15332SJohannes Weiner /* 309289b15332SJohannes Weiner * Throttle page dirtying rate down to writeback speed. 309389b15332SJohannes Weiner * 309489b15332SJohannes Weiner * mapping may be NULL here because some device drivers do not 309589b15332SJohannes Weiner * set page.mapping but still dirty their pages 309689b15332SJohannes Weiner * 3097c1e8d7c6SMichel Lespinasse * Drop the mmap_lock before waiting on IO, if we can. The file 309889b15332SJohannes Weiner * is pinning the mapping, as per above. 309989b15332SJohannes Weiner */ 310089b15332SJohannes Weiner if ((dirtied || page_mkwrite) && mapping) { 310189b15332SJohannes Weiner struct file *fpin; 310289b15332SJohannes Weiner 310389b15332SJohannes Weiner fpin = maybe_unlock_mmap_for_io(vmf, NULL); 310489b15332SJohannes Weiner balance_dirty_pages_ratelimited(mapping); 310589b15332SJohannes Weiner if (fpin) { 310689b15332SJohannes Weiner fput(fpin); 3107d9272525SPeter Xu return VM_FAULT_COMPLETED; 310889b15332SJohannes Weiner } 310989b15332SJohannes Weiner } 311089b15332SJohannes Weiner 311189b15332SJohannes Weiner return 0; 31124e047f89SShachar Raindel } 31134e047f89SShachar Raindel 311497ba0c2bSJan Kara /* 31154e047f89SShachar Raindel * Handle write page faults for pages that can be reused in the current vma 31164e047f89SShachar Raindel * 31174e047f89SShachar Raindel * This can happen either due to the mapping being with the VM_SHARED flag, 31184e047f89SShachar Raindel * or due to us being the last reference standing to the page. In either 31194e047f89SShachar Raindel * case, all we need to do here is to mark the page as writable and update 31204e047f89SShachar Raindel * any related book-keeping. 31214e047f89SShachar Raindel */ 3122a86bc96bSKefeng Wang static inline void wp_page_reuse(struct vm_fault *vmf, struct folio *folio) 312382b0f8c3SJan Kara __releases(vmf->ptl) 31244e047f89SShachar Raindel { 312582b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 31264e047f89SShachar Raindel pte_t entry; 31276c287605SDavid Hildenbrand 3128c89357e2SDavid Hildenbrand VM_BUG_ON(!(vmf->flags & FAULT_FLAG_WRITE)); 31296c287605SDavid Hildenbrand 3130c2c3b514SKefeng Wang if (folio) { 3131c2c3b514SKefeng Wang VM_BUG_ON(folio_test_anon(folio) && 3132c2c3b514SKefeng Wang !PageAnonExclusive(vmf->page)); 31334e047f89SShachar Raindel /* 3134c2c3b514SKefeng Wang * Clear the folio's cpupid information as the existing 31354e047f89SShachar Raindel * information potentially belongs to a now completely 31364e047f89SShachar Raindel * unrelated process. 31374e047f89SShachar Raindel */ 3138c2c3b514SKefeng Wang folio_xchg_last_cpupid(folio, (1 << LAST_CPUPID_SHIFT) - 1); 3139c2c3b514SKefeng Wang } 31404e047f89SShachar Raindel 31412994302bSJan Kara flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte)); 31422994302bSJan Kara entry = pte_mkyoung(vmf->orig_pte); 31434e047f89SShachar Raindel entry = maybe_mkwrite(pte_mkdirty(entry), vma); 314482b0f8c3SJan Kara if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1)) 31455003a2bdSMatthew Wilcox (Oracle) update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1); 314682b0f8c3SJan Kara pte_unmap_unlock(vmf->pte, vmf->ptl); 3147798a6b87SPeter Xu count_vm_event(PGREUSE); 31484e047f89SShachar Raindel } 31494e047f89SShachar Raindel 31504ed43798SMatthew Wilcox (Oracle) /* 31514ed43798SMatthew Wilcox (Oracle) * We could add a bitflag somewhere, but for now, we know that all 31524ed43798SMatthew Wilcox (Oracle) * vm_ops that have a ->map_pages have been audited and don't need 31534ed43798SMatthew Wilcox (Oracle) * the mmap_lock to be held. 31544ed43798SMatthew Wilcox (Oracle) */ 31554ed43798SMatthew Wilcox (Oracle) static inline vm_fault_t vmf_can_call_fault(const struct vm_fault *vmf) 31564ed43798SMatthew Wilcox (Oracle) { 31574ed43798SMatthew Wilcox (Oracle) struct vm_area_struct *vma = vmf->vma; 31584ed43798SMatthew Wilcox (Oracle) 31594ed43798SMatthew Wilcox (Oracle) if (vma->vm_ops->map_pages || !(vmf->flags & FAULT_FLAG_VMA_LOCK)) 31604ed43798SMatthew Wilcox (Oracle) return 0; 31614ed43798SMatthew Wilcox (Oracle) vma_end_read(vma); 31624ed43798SMatthew Wilcox (Oracle) return VM_FAULT_RETRY; 31634ed43798SMatthew Wilcox (Oracle) } 31644ed43798SMatthew Wilcox (Oracle) 3165164b06f2SMatthew Wilcox (Oracle) static vm_fault_t vmf_anon_prepare(struct vm_fault *vmf) 3166164b06f2SMatthew Wilcox (Oracle) { 3167164b06f2SMatthew Wilcox (Oracle) struct vm_area_struct *vma = vmf->vma; 3168164b06f2SMatthew Wilcox (Oracle) 3169164b06f2SMatthew Wilcox (Oracle) if (likely(vma->anon_vma)) 3170164b06f2SMatthew Wilcox (Oracle) return 0; 3171164b06f2SMatthew Wilcox (Oracle) if (vmf->flags & FAULT_FLAG_VMA_LOCK) { 3172164b06f2SMatthew Wilcox (Oracle) vma_end_read(vma); 3173164b06f2SMatthew Wilcox (Oracle) return VM_FAULT_RETRY; 3174164b06f2SMatthew Wilcox (Oracle) } 3175164b06f2SMatthew Wilcox (Oracle) if (__anon_vma_prepare(vma)) 3176164b06f2SMatthew Wilcox (Oracle) return VM_FAULT_OOM; 3177164b06f2SMatthew Wilcox (Oracle) return 0; 3178164b06f2SMatthew Wilcox (Oracle) } 3179164b06f2SMatthew Wilcox (Oracle) 31804e047f89SShachar Raindel /* 3181c89357e2SDavid Hildenbrand * Handle the case of a page which we actually need to copy to a new page, 3182c89357e2SDavid Hildenbrand * either due to COW or unsharing. 31832f38ab2cSShachar Raindel * 3184c1e8d7c6SMichel Lespinasse * Called with mmap_lock locked and the old page referenced, but 31852f38ab2cSShachar Raindel * without the ptl held. 31862f38ab2cSShachar Raindel * 31872f38ab2cSShachar Raindel * High level logic flow: 31882f38ab2cSShachar Raindel * 31892f38ab2cSShachar Raindel * - Allocate a page, copy the content of the old page to the new one. 31902f38ab2cSShachar Raindel * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc. 31912f38ab2cSShachar Raindel * - Take the PTL. If the pte changed, bail out and release the allocated page 31922f38ab2cSShachar Raindel * - If the pte is still the way we remember it, update the page table and all 31932f38ab2cSShachar Raindel * relevant references. This includes dropping the reference the page-table 31942f38ab2cSShachar Raindel * held to the old page, as well as updating the rmap. 31952f38ab2cSShachar Raindel * - In any case, unlock the PTL and drop the reference we took to the old page. 31962f38ab2cSShachar Raindel */ 31972b740303SSouptick Joarder static vm_fault_t wp_page_copy(struct vm_fault *vmf) 31982f38ab2cSShachar Raindel { 3199c89357e2SDavid Hildenbrand const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; 320082b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 3201bae473a4SKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 320228d41a48SMatthew Wilcox (Oracle) struct folio *old_folio = NULL; 320328d41a48SMatthew Wilcox (Oracle) struct folio *new_folio = NULL; 32042f38ab2cSShachar Raindel pte_t entry; 32052f38ab2cSShachar Raindel int page_copied = 0; 3206ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3207164b06f2SMatthew Wilcox (Oracle) vm_fault_t ret; 3208cf503cc6SKefeng Wang bool pfn_is_zero; 32092f38ab2cSShachar Raindel 3210662ce1dcSYang Yang delayacct_wpcopy_start(); 3211662ce1dcSYang Yang 321228d41a48SMatthew Wilcox (Oracle) if (vmf->page) 321328d41a48SMatthew Wilcox (Oracle) old_folio = page_folio(vmf->page); 3214164b06f2SMatthew Wilcox (Oracle) ret = vmf_anon_prepare(vmf); 3215164b06f2SMatthew Wilcox (Oracle) if (unlikely(ret)) 3216164b06f2SMatthew Wilcox (Oracle) goto out; 32172f38ab2cSShachar Raindel 3218cf503cc6SKefeng Wang pfn_is_zero = is_zero_pfn(pte_pfn(vmf->orig_pte)); 3219cf503cc6SKefeng Wang new_folio = folio_prealloc(mm, vma, vmf->address, pfn_is_zero); 32206bc56a4dSMatthew Wilcox (Oracle) if (!new_folio) 32212f38ab2cSShachar Raindel goto oom; 3222cf503cc6SKefeng Wang 3223cf503cc6SKefeng Wang if (!pfn_is_zero) { 3224164b06f2SMatthew Wilcox (Oracle) int err; 322583d116c5SJia He 3226164b06f2SMatthew Wilcox (Oracle) err = __wp_page_copy_user(&new_folio->page, vmf->page, vmf); 3227164b06f2SMatthew Wilcox (Oracle) if (err) { 322883d116c5SJia He /* 322983d116c5SJia He * COW failed, if the fault was solved by other, 323083d116c5SJia He * it's fine. If not, userspace would re-fault on 323183d116c5SJia He * the same address and we will handle the fault 323283d116c5SJia He * from the second attempt. 3233a873dfe1STony Luck * The -EHWPOISON case will not be retried. 323483d116c5SJia He */ 323528d41a48SMatthew Wilcox (Oracle) folio_put(new_folio); 323628d41a48SMatthew Wilcox (Oracle) if (old_folio) 323728d41a48SMatthew Wilcox (Oracle) folio_put(old_folio); 3238662ce1dcSYang Yang 3239662ce1dcSYang Yang delayacct_wpcopy_end(); 3240164b06f2SMatthew Wilcox (Oracle) return err == -EHWPOISON ? VM_FAULT_HWPOISON : 0; 324183d116c5SJia He } 324228d41a48SMatthew Wilcox (Oracle) kmsan_copy_page_meta(&new_folio->page, vmf->page); 32432f38ab2cSShachar Raindel } 32442f38ab2cSShachar Raindel 324528d41a48SMatthew Wilcox (Oracle) __folio_mark_uptodate(new_folio); 3246eb3c24f3SMel Gorman 32477d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, 32486f4f13e8SJérôme Glisse vmf->address & PAGE_MASK, 3249ac46d4f3SJérôme Glisse (vmf->address & PAGE_MASK) + PAGE_SIZE); 3250ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 32512f38ab2cSShachar Raindel 32522f38ab2cSShachar Raindel /* 32532f38ab2cSShachar Raindel * Re-check the pte - we dropped the lock 32542f38ab2cSShachar Raindel */ 325582b0f8c3SJan Kara vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl); 3256c33c7948SRyan Roberts if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte))) { 325728d41a48SMatthew Wilcox (Oracle) if (old_folio) { 325828d41a48SMatthew Wilcox (Oracle) if (!folio_test_anon(old_folio)) { 32596b27cc6cSKefeng Wang dec_mm_counter(mm, mm_counter_file(old_folio)); 3260f1a79412SShakeel Butt inc_mm_counter(mm, MM_ANONPAGES); 32612f38ab2cSShachar Raindel } 32622f38ab2cSShachar Raindel } else { 32636080d19fSxu xin ksm_might_unmap_zero_page(mm, vmf->orig_pte); 3264f1a79412SShakeel Butt inc_mm_counter(mm, MM_ANONPAGES); 32652f38ab2cSShachar Raindel } 32662994302bSJan Kara flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte)); 326728d41a48SMatthew Wilcox (Oracle) entry = mk_pte(&new_folio->page, vma->vm_page_prot); 326850c25ee9SThomas Bogendoerfer entry = pte_sw_mkyoung(entry); 3269c89357e2SDavid Hildenbrand if (unlikely(unshare)) { 3270c89357e2SDavid Hildenbrand if (pte_soft_dirty(vmf->orig_pte)) 3271c89357e2SDavid Hildenbrand entry = pte_mksoft_dirty(entry); 3272c89357e2SDavid Hildenbrand if (pte_uffd_wp(vmf->orig_pte)) 3273c89357e2SDavid Hildenbrand entry = pte_mkuffd_wp(entry); 3274c89357e2SDavid Hildenbrand } else { 32752f38ab2cSShachar Raindel entry = maybe_mkwrite(pte_mkdirty(entry), vma); 3276c89357e2SDavid Hildenbrand } 3277111fe718SNicholas Piggin 32782f38ab2cSShachar Raindel /* 32792f38ab2cSShachar Raindel * Clear the pte entry and flush it first, before updating the 3280111fe718SNicholas Piggin * pte with the new entry, to keep TLBs on different CPUs in 3281111fe718SNicholas Piggin * sync. This code used to set the new PTE then flush TLBs, but 3282111fe718SNicholas Piggin * that left a window where the new PTE could be loaded into 3283111fe718SNicholas Piggin * some TLBs while the old PTE remains in others. 32842f38ab2cSShachar Raindel */ 3285ec8832d0SAlistair Popple ptep_clear_flush(vma, vmf->address, vmf->pte); 328628d41a48SMatthew Wilcox (Oracle) folio_add_new_anon_rmap(new_folio, vma, vmf->address); 328728d41a48SMatthew Wilcox (Oracle) folio_add_lru_vma(new_folio, vma); 32882f38ab2cSShachar Raindel /* 32892f38ab2cSShachar Raindel * We call the notify macro here because, when using secondary 32902f38ab2cSShachar Raindel * mmu page tables (such as kvm shadow page tables), we want the 32912f38ab2cSShachar Raindel * new page to be mapped directly into the secondary page table. 32922f38ab2cSShachar Raindel */ 3293c89357e2SDavid Hildenbrand BUG_ON(unshare && pte_write(entry)); 329482b0f8c3SJan Kara set_pte_at_notify(mm, vmf->address, vmf->pte, entry); 32955003a2bdSMatthew Wilcox (Oracle) update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1); 329628d41a48SMatthew Wilcox (Oracle) if (old_folio) { 32972f38ab2cSShachar Raindel /* 32982f38ab2cSShachar Raindel * Only after switching the pte to the new page may 32992f38ab2cSShachar Raindel * we remove the mapcount here. Otherwise another 33002f38ab2cSShachar Raindel * process may come and find the rmap count decremented 33012f38ab2cSShachar Raindel * before the pte is switched to the new page, and 33022f38ab2cSShachar Raindel * "reuse" the old page writing into it while our pte 33032f38ab2cSShachar Raindel * here still points into it and can be read by other 33042f38ab2cSShachar Raindel * threads. 33052f38ab2cSShachar Raindel * 33062f38ab2cSShachar Raindel * The critical issue is to order this 3307c4626503SDavid Hildenbrand * folio_remove_rmap_pte() with the ptp_clear_flush 3308c4626503SDavid Hildenbrand * above. Those stores are ordered by (if nothing else,) 33092f38ab2cSShachar Raindel * the barrier present in the atomic_add_negative 3310c4626503SDavid Hildenbrand * in folio_remove_rmap_pte(); 33112f38ab2cSShachar Raindel * 33122f38ab2cSShachar Raindel * Then the TLB flush in ptep_clear_flush ensures that 33132f38ab2cSShachar Raindel * no process can access the old page before the 33142f38ab2cSShachar Raindel * decremented mapcount is visible. And the old page 33152f38ab2cSShachar Raindel * cannot be reused until after the decremented 33162f38ab2cSShachar Raindel * mapcount is visible. So transitively, TLBs to 33172f38ab2cSShachar Raindel * old page will be flushed before it can be reused. 33182f38ab2cSShachar Raindel */ 3319c4626503SDavid Hildenbrand folio_remove_rmap_pte(old_folio, vmf->page, vma); 33202f38ab2cSShachar Raindel } 33212f38ab2cSShachar Raindel 33222f38ab2cSShachar Raindel /* Free the old page.. */ 332328d41a48SMatthew Wilcox (Oracle) new_folio = old_folio; 33242f38ab2cSShachar Raindel page_copied = 1; 33253db82b93SHugh Dickins pte_unmap_unlock(vmf->pte, vmf->ptl); 33263db82b93SHugh Dickins } else if (vmf->pte) { 33277df67697SBibo Mao update_mmu_tlb(vma, vmf->address, vmf->pte); 33283db82b93SHugh Dickins pte_unmap_unlock(vmf->pte, vmf->ptl); 33292f38ab2cSShachar Raindel } 33302f38ab2cSShachar Raindel 3331ec8832d0SAlistair Popple mmu_notifier_invalidate_range_end(&range); 33323db82b93SHugh Dickins 33333db82b93SHugh Dickins if (new_folio) 33343db82b93SHugh Dickins folio_put(new_folio); 333528d41a48SMatthew Wilcox (Oracle) if (old_folio) { 3336f4c4a3f4SHuang Ying if (page_copied) 333728d41a48SMatthew Wilcox (Oracle) free_swap_cache(&old_folio->page); 333828d41a48SMatthew Wilcox (Oracle) folio_put(old_folio); 33392f38ab2cSShachar Raindel } 3340662ce1dcSYang Yang 3341662ce1dcSYang Yang delayacct_wpcopy_end(); 3342cb8d8633SDavid Hildenbrand return 0; 33432f38ab2cSShachar Raindel oom: 3344164b06f2SMatthew Wilcox (Oracle) ret = VM_FAULT_OOM; 3345164b06f2SMatthew Wilcox (Oracle) out: 334628d41a48SMatthew Wilcox (Oracle) if (old_folio) 334728d41a48SMatthew Wilcox (Oracle) folio_put(old_folio); 3348662ce1dcSYang Yang 3349662ce1dcSYang Yang delayacct_wpcopy_end(); 3350164b06f2SMatthew Wilcox (Oracle) return ret; 33512f38ab2cSShachar Raindel } 33522f38ab2cSShachar Raindel 335366a6197cSJan Kara /** 335466a6197cSJan Kara * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE 335566a6197cSJan Kara * writeable once the page is prepared 335666a6197cSJan Kara * 335766a6197cSJan Kara * @vmf: structure describing the fault 3358a86bc96bSKefeng Wang * @folio: the folio of vmf->page 335966a6197cSJan Kara * 336066a6197cSJan Kara * This function handles all that is needed to finish a write page fault in a 336166a6197cSJan Kara * shared mapping due to PTE being read-only once the mapped page is prepared. 3362a862f68aSMike Rapoport * It handles locking of PTE and modifying it. 336366a6197cSJan Kara * 336466a6197cSJan Kara * The function expects the page to be locked or other protection against 336566a6197cSJan Kara * concurrent faults / writeback (such as DAX radix tree locks). 3366a862f68aSMike Rapoport * 33672797e79fSLiu Xiang * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before 3368a862f68aSMike Rapoport * we acquired PTE lock. 336966a6197cSJan Kara */ 3370a86bc96bSKefeng Wang static vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf, struct folio *folio) 337166a6197cSJan Kara { 337266a6197cSJan Kara WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED)); 337366a6197cSJan Kara vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address, 337466a6197cSJan Kara &vmf->ptl); 33753db82b93SHugh Dickins if (!vmf->pte) 33763db82b93SHugh Dickins return VM_FAULT_NOPAGE; 337766a6197cSJan Kara /* 337866a6197cSJan Kara * We might have raced with another page fault while we released the 337966a6197cSJan Kara * pte_offset_map_lock. 338066a6197cSJan Kara */ 3381c33c7948SRyan Roberts if (!pte_same(ptep_get(vmf->pte), vmf->orig_pte)) { 33827df67697SBibo Mao update_mmu_tlb(vmf->vma, vmf->address, vmf->pte); 338366a6197cSJan Kara pte_unmap_unlock(vmf->pte, vmf->ptl); 3384a19e2553SJan Kara return VM_FAULT_NOPAGE; 338566a6197cSJan Kara } 3386a86bc96bSKefeng Wang wp_page_reuse(vmf, folio); 3387a19e2553SJan Kara return 0; 338866a6197cSJan Kara } 338966a6197cSJan Kara 3390dd906184SBoaz Harrosh /* 3391dd906184SBoaz Harrosh * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED 3392dd906184SBoaz Harrosh * mapping 3393dd906184SBoaz Harrosh */ 33942b740303SSouptick Joarder static vm_fault_t wp_pfn_shared(struct vm_fault *vmf) 3395dd906184SBoaz Harrosh { 339682b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 3397bae473a4SKirill A. Shutemov 3398dd906184SBoaz Harrosh if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) { 33992b740303SSouptick Joarder vm_fault_t ret; 3400dd906184SBoaz Harrosh 340182b0f8c3SJan Kara pte_unmap_unlock(vmf->pte, vmf->ptl); 34024a68fef1SMatthew Wilcox (Oracle) ret = vmf_can_call_fault(vmf); 34034a68fef1SMatthew Wilcox (Oracle) if (ret) 34044a68fef1SMatthew Wilcox (Oracle) return ret; 3405063e60d8SMatthew Wilcox (Oracle) 3406fe82221fSJan Kara vmf->flags |= FAULT_FLAG_MKWRITE; 340711bac800SDave Jiang ret = vma->vm_ops->pfn_mkwrite(vmf); 34082f89dc12SJan Kara if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)) 3409dd906184SBoaz Harrosh return ret; 3410a86bc96bSKefeng Wang return finish_mkwrite_fault(vmf, NULL); 3411dd906184SBoaz Harrosh } 3412a86bc96bSKefeng Wang wp_page_reuse(vmf, NULL); 3413cb8d8633SDavid Hildenbrand return 0; 3414dd906184SBoaz Harrosh } 3415dd906184SBoaz Harrosh 34165a97858bSSidhartha Kumar static vm_fault_t wp_page_shared(struct vm_fault *vmf, struct folio *folio) 341782b0f8c3SJan Kara __releases(vmf->ptl) 341893e478d4SShachar Raindel { 341982b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 3420cb8d8633SDavid Hildenbrand vm_fault_t ret = 0; 342193e478d4SShachar Raindel 34225a97858bSSidhartha Kumar folio_get(folio); 342393e478d4SShachar Raindel 342493e478d4SShachar Raindel if (vma->vm_ops && vma->vm_ops->page_mkwrite) { 34252b740303SSouptick Joarder vm_fault_t tmp; 342693e478d4SShachar Raindel 342782b0f8c3SJan Kara pte_unmap_unlock(vmf->pte, vmf->ptl); 34284a68fef1SMatthew Wilcox (Oracle) tmp = vmf_can_call_fault(vmf); 34294a68fef1SMatthew Wilcox (Oracle) if (tmp) { 3430063e60d8SMatthew Wilcox (Oracle) folio_put(folio); 34314a68fef1SMatthew Wilcox (Oracle) return tmp; 3432063e60d8SMatthew Wilcox (Oracle) } 3433063e60d8SMatthew Wilcox (Oracle) 343486aa6998SSidhartha Kumar tmp = do_page_mkwrite(vmf, folio); 343593e478d4SShachar Raindel if (unlikely(!tmp || (tmp & 343693e478d4SShachar Raindel (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 34375a97858bSSidhartha Kumar folio_put(folio); 343893e478d4SShachar Raindel return tmp; 343993e478d4SShachar Raindel } 3440a86bc96bSKefeng Wang tmp = finish_mkwrite_fault(vmf, folio); 3441a19e2553SJan Kara if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) { 34425a97858bSSidhartha Kumar folio_unlock(folio); 34435a97858bSSidhartha Kumar folio_put(folio); 344466a6197cSJan Kara return tmp; 344593e478d4SShachar Raindel } 344666a6197cSJan Kara } else { 3447a86bc96bSKefeng Wang wp_page_reuse(vmf, folio); 34485a97858bSSidhartha Kumar folio_lock(folio); 344993e478d4SShachar Raindel } 345089b15332SJohannes Weiner ret |= fault_dirty_shared_page(vmf); 34515a97858bSSidhartha Kumar folio_put(folio); 345293e478d4SShachar Raindel 345389b15332SJohannes Weiner return ret; 345493e478d4SShachar Raindel } 345593e478d4SShachar Raindel 3456dec078ccSDavid Hildenbrand static bool wp_can_reuse_anon_folio(struct folio *folio, 3457dec078ccSDavid Hildenbrand struct vm_area_struct *vma) 3458dec078ccSDavid Hildenbrand { 3459dec078ccSDavid Hildenbrand /* 3460dec078ccSDavid Hildenbrand * We have to verify under folio lock: these early checks are 3461dec078ccSDavid Hildenbrand * just an optimization to avoid locking the folio and freeing 3462dec078ccSDavid Hildenbrand * the swapcache if there is little hope that we can reuse. 3463dec078ccSDavid Hildenbrand * 3464dec078ccSDavid Hildenbrand * KSM doesn't necessarily raise the folio refcount. 3465dec078ccSDavid Hildenbrand */ 3466dec078ccSDavid Hildenbrand if (folio_test_ksm(folio) || folio_ref_count(folio) > 3) 3467dec078ccSDavid Hildenbrand return false; 3468dec078ccSDavid Hildenbrand if (!folio_test_lru(folio)) 3469dec078ccSDavid Hildenbrand /* 3470dec078ccSDavid Hildenbrand * We cannot easily detect+handle references from 3471dec078ccSDavid Hildenbrand * remote LRU caches or references to LRU folios. 3472dec078ccSDavid Hildenbrand */ 3473dec078ccSDavid Hildenbrand lru_add_drain(); 3474dec078ccSDavid Hildenbrand if (folio_ref_count(folio) > 1 + folio_test_swapcache(folio)) 3475dec078ccSDavid Hildenbrand return false; 3476dec078ccSDavid Hildenbrand if (!folio_trylock(folio)) 3477dec078ccSDavid Hildenbrand return false; 3478dec078ccSDavid Hildenbrand if (folio_test_swapcache(folio)) 3479dec078ccSDavid Hildenbrand folio_free_swap(folio); 3480dec078ccSDavid Hildenbrand if (folio_test_ksm(folio) || folio_ref_count(folio) != 1) { 3481dec078ccSDavid Hildenbrand folio_unlock(folio); 3482dec078ccSDavid Hildenbrand return false; 3483dec078ccSDavid Hildenbrand } 3484dec078ccSDavid Hildenbrand /* 3485dec078ccSDavid Hildenbrand * Ok, we've got the only folio reference from our mapping 3486dec078ccSDavid Hildenbrand * and the folio is locked, it's dark out, and we're wearing 3487dec078ccSDavid Hildenbrand * sunglasses. Hit it. 3488dec078ccSDavid Hildenbrand */ 3489dec078ccSDavid Hildenbrand folio_move_anon_rmap(folio, vma); 3490dec078ccSDavid Hildenbrand folio_unlock(folio); 3491dec078ccSDavid Hildenbrand return true; 3492dec078ccSDavid Hildenbrand } 3493dec078ccSDavid Hildenbrand 34942f38ab2cSShachar Raindel /* 3495c89357e2SDavid Hildenbrand * This routine handles present pages, when 3496c89357e2SDavid Hildenbrand * * users try to write to a shared page (FAULT_FLAG_WRITE) 3497c89357e2SDavid Hildenbrand * * GUP wants to take a R/O pin on a possibly shared anonymous page 3498c89357e2SDavid Hildenbrand * (FAULT_FLAG_UNSHARE) 3499c89357e2SDavid Hildenbrand * 3500c89357e2SDavid Hildenbrand * It is done by copying the page to a new address and decrementing the 3501c89357e2SDavid Hildenbrand * shared-page counter for the old page. 35021da177e4SLinus Torvalds * 35031da177e4SLinus Torvalds * Note that this routine assumes that the protection checks have been 35041da177e4SLinus Torvalds * done by the caller (the low-level page fault routine in most cases). 3505c89357e2SDavid Hildenbrand * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've 3506c89357e2SDavid Hildenbrand * done any necessary COW. 35071da177e4SLinus Torvalds * 3508c89357e2SDavid Hildenbrand * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even 3509c89357e2SDavid Hildenbrand * though the page will change only once the write actually happens. This 3510c89357e2SDavid Hildenbrand * avoids a few races, and potentially makes it more efficient. 35111da177e4SLinus Torvalds * 3512c1e8d7c6SMichel Lespinasse * We enter with non-exclusive mmap_lock (to exclude vma changes, 35138f4e2101SHugh Dickins * but allow concurrent faults), with pte both mapped and locked. 3514c1e8d7c6SMichel Lespinasse * We return with mmap_lock still held, but pte unmapped and unlocked. 35151da177e4SLinus Torvalds */ 35162b740303SSouptick Joarder static vm_fault_t do_wp_page(struct vm_fault *vmf) 351782b0f8c3SJan Kara __releases(vmf->ptl) 35181da177e4SLinus Torvalds { 3519c89357e2SDavid Hildenbrand const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; 352082b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 3521b9086fdeSDavid Hildenbrand struct folio *folio = NULL; 3522d61ea1cbSPeter Xu pte_t pte; 35231da177e4SLinus Torvalds 3524c89357e2SDavid Hildenbrand if (likely(!unshare)) { 3525c33c7948SRyan Roberts if (userfaultfd_pte_wp(vma, ptep_get(vmf->pte))) { 3526d61ea1cbSPeter Xu if (!userfaultfd_wp_async(vma)) { 3527529b930bSAndrea Arcangeli pte_unmap_unlock(vmf->pte, vmf->ptl); 3528529b930bSAndrea Arcangeli return handle_userfault(vmf, VM_UFFD_WP); 3529529b930bSAndrea Arcangeli } 3530529b930bSAndrea Arcangeli 35316ce64428SNadav Amit /* 3532d61ea1cbSPeter Xu * Nothing needed (cache flush, TLB invalidations, 3533d61ea1cbSPeter Xu * etc.) because we're only removing the uffd-wp bit, 3534d61ea1cbSPeter Xu * which is completely invisible to the user. 3535d61ea1cbSPeter Xu */ 3536d61ea1cbSPeter Xu pte = pte_clear_uffd_wp(ptep_get(vmf->pte)); 3537d61ea1cbSPeter Xu 3538d61ea1cbSPeter Xu set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte); 3539d61ea1cbSPeter Xu /* 3540d61ea1cbSPeter Xu * Update this to be prepared for following up CoW 3541d61ea1cbSPeter Xu * handling 3542d61ea1cbSPeter Xu */ 3543d61ea1cbSPeter Xu vmf->orig_pte = pte; 3544d61ea1cbSPeter Xu } 3545d61ea1cbSPeter Xu 3546d61ea1cbSPeter Xu /* 35476ce64428SNadav Amit * Userfaultfd write-protect can defer flushes. Ensure the TLB 35486ce64428SNadav Amit * is flushed in this case before copying. 35496ce64428SNadav Amit */ 35506ce64428SNadav Amit if (unlikely(userfaultfd_wp(vmf->vma) && 35516ce64428SNadav Amit mm_tlb_flush_pending(vmf->vma->vm_mm))) 35526ce64428SNadav Amit flush_tlb_page(vmf->vma, vmf->address); 3553c89357e2SDavid Hildenbrand } 35546ce64428SNadav Amit 3555a41b70d6SJan Kara vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte); 3556c89357e2SDavid Hildenbrand 35575a97858bSSidhartha Kumar if (vmf->page) 35585a97858bSSidhartha Kumar folio = page_folio(vmf->page); 35595a97858bSSidhartha Kumar 3560251b97f5SPeter Zijlstra /* 3561b9086fdeSDavid Hildenbrand * Shared mapping: we are guaranteed to have VM_WRITE and 3562b9086fdeSDavid Hildenbrand * FAULT_FLAG_WRITE set at this point. 3563b9086fdeSDavid Hildenbrand */ 3564b9086fdeSDavid Hildenbrand if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) { 3565b9086fdeSDavid Hildenbrand /* 356664e45507SPeter Feiner * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a 356764e45507SPeter Feiner * VM_PFNMAP VMA. 3568251b97f5SPeter Zijlstra * 3569251b97f5SPeter Zijlstra * We should not cow pages in a shared writeable mapping. 3570dd906184SBoaz Harrosh * Just mark the pages writable and/or call ops->pfn_mkwrite. 3571251b97f5SPeter Zijlstra */ 3572b9086fdeSDavid Hildenbrand if (!vmf->page) 35732994302bSJan Kara return wp_pfn_shared(vmf); 35745a97858bSSidhartha Kumar return wp_page_shared(vmf, folio); 3575251b97f5SPeter Zijlstra } 35761da177e4SLinus Torvalds 3577b9086fdeSDavid Hildenbrand /* 3578b9086fdeSDavid Hildenbrand * Private mapping: create an exclusive anonymous page copy if reuse 3579b9086fdeSDavid Hildenbrand * is impossible. We might miss VM_WRITE for FOLL_FORCE handling. 358053a05ad9SDavid Hildenbrand * 3581dec078ccSDavid Hildenbrand * If we encounter a page that is marked exclusive, we must reuse 3582dec078ccSDavid Hildenbrand * the page without further checks. 358353a05ad9SDavid Hildenbrand */ 3584dec078ccSDavid Hildenbrand if (folio && folio_test_anon(folio) && 3585dec078ccSDavid Hildenbrand (PageAnonExclusive(vmf->page) || wp_can_reuse_anon_folio(folio, vma))) { 3586dec078ccSDavid Hildenbrand if (!PageAnonExclusive(vmf->page)) 35875ca43289SDavid Hildenbrand SetPageAnonExclusive(vmf->page); 3588c89357e2SDavid Hildenbrand if (unlikely(unshare)) { 3589c89357e2SDavid Hildenbrand pte_unmap_unlock(vmf->pte, vmf->ptl); 3590c89357e2SDavid Hildenbrand return 0; 3591c89357e2SDavid Hildenbrand } 3592a86bc96bSKefeng Wang wp_page_reuse(vmf, folio); 3593cb8d8633SDavid Hildenbrand return 0; 35941da177e4SLinus Torvalds } 35951da177e4SLinus Torvalds /* 35961da177e4SLinus Torvalds * Ok, we need to copy. Oh, well.. 35971da177e4SLinus Torvalds */ 3598b9086fdeSDavid Hildenbrand if (folio) 3599b9086fdeSDavid Hildenbrand folio_get(folio); 360028766805SShachar Raindel 360182b0f8c3SJan Kara pte_unmap_unlock(vmf->pte, vmf->ptl); 360294bfe85bSYang Yang #ifdef CONFIG_KSM 3603b9086fdeSDavid Hildenbrand if (folio && folio_test_ksm(folio)) 360494bfe85bSYang Yang count_vm_event(COW_KSM); 360594bfe85bSYang Yang #endif 3606a41b70d6SJan Kara return wp_page_copy(vmf); 36071da177e4SLinus Torvalds } 36081da177e4SLinus Torvalds 360997a89413SPeter Zijlstra static void unmap_mapping_range_vma(struct vm_area_struct *vma, 36101da177e4SLinus Torvalds unsigned long start_addr, unsigned long end_addr, 36111da177e4SLinus Torvalds struct zap_details *details) 36121da177e4SLinus Torvalds { 3613f5cc4eefSAl Viro zap_page_range_single(vma, start_addr, end_addr - start_addr, details); 36141da177e4SLinus Torvalds } 36151da177e4SLinus Torvalds 3616f808c13fSDavidlohr Bueso static inline void unmap_mapping_range_tree(struct rb_root_cached *root, 3617232a6a1cSPeter Xu pgoff_t first_index, 3618232a6a1cSPeter Xu pgoff_t last_index, 36191da177e4SLinus Torvalds struct zap_details *details) 36201da177e4SLinus Torvalds { 36211da177e4SLinus Torvalds struct vm_area_struct *vma; 36221da177e4SLinus Torvalds pgoff_t vba, vea, zba, zea; 36231da177e4SLinus Torvalds 3624232a6a1cSPeter Xu vma_interval_tree_foreach(vma, root, first_index, last_index) { 36251da177e4SLinus Torvalds vba = vma->vm_pgoff; 3626d6e93217SLibin vea = vba + vma_pages(vma) - 1; 3627f9871da9SMiaohe Lin zba = max(first_index, vba); 3628f9871da9SMiaohe Lin zea = min(last_index, vea); 36291da177e4SLinus Torvalds 363097a89413SPeter Zijlstra unmap_mapping_range_vma(vma, 36311da177e4SLinus Torvalds ((zba - vba) << PAGE_SHIFT) + vma->vm_start, 36321da177e4SLinus Torvalds ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start, 363397a89413SPeter Zijlstra details); 36341da177e4SLinus Torvalds } 36351da177e4SLinus Torvalds } 36361da177e4SLinus Torvalds 36371da177e4SLinus Torvalds /** 36383506659eSMatthew Wilcox (Oracle) * unmap_mapping_folio() - Unmap single folio from processes. 36393506659eSMatthew Wilcox (Oracle) * @folio: The locked folio to be unmapped. 364022061a1fSHugh Dickins * 36413506659eSMatthew Wilcox (Oracle) * Unmap this folio from any userspace process which still has it mmaped. 364222061a1fSHugh Dickins * Typically, for efficiency, the range of nearby pages has already been 364322061a1fSHugh Dickins * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once 36443506659eSMatthew Wilcox (Oracle) * truncation or invalidation holds the lock on a folio, it may find that 36453506659eSMatthew Wilcox (Oracle) * the page has been remapped again: and then uses unmap_mapping_folio() 364622061a1fSHugh Dickins * to unmap it finally. 364722061a1fSHugh Dickins */ 36483506659eSMatthew Wilcox (Oracle) void unmap_mapping_folio(struct folio *folio) 364922061a1fSHugh Dickins { 36503506659eSMatthew Wilcox (Oracle) struct address_space *mapping = folio->mapping; 365122061a1fSHugh Dickins struct zap_details details = { }; 3652232a6a1cSPeter Xu pgoff_t first_index; 3653232a6a1cSPeter Xu pgoff_t last_index; 365422061a1fSHugh Dickins 36553506659eSMatthew Wilcox (Oracle) VM_BUG_ON(!folio_test_locked(folio)); 365622061a1fSHugh Dickins 36573506659eSMatthew Wilcox (Oracle) first_index = folio->index; 365887b11f86SSidhartha Kumar last_index = folio_next_index(folio) - 1; 3659232a6a1cSPeter Xu 36602e148f1eSPeter Xu details.even_cows = false; 36613506659eSMatthew Wilcox (Oracle) details.single_folio = folio; 3662999dad82SPeter Xu details.zap_flags = ZAP_FLAG_DROP_MARKER; 366322061a1fSHugh Dickins 36642c865995SHugh Dickins i_mmap_lock_read(mapping); 366522061a1fSHugh Dickins if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))) 3666232a6a1cSPeter Xu unmap_mapping_range_tree(&mapping->i_mmap, first_index, 3667232a6a1cSPeter Xu last_index, &details); 36682c865995SHugh Dickins i_mmap_unlock_read(mapping); 366922061a1fSHugh Dickins } 367022061a1fSHugh Dickins 367122061a1fSHugh Dickins /** 3672977fbdcdSMatthew Wilcox * unmap_mapping_pages() - Unmap pages from processes. 3673977fbdcdSMatthew Wilcox * @mapping: The address space containing pages to be unmapped. 3674977fbdcdSMatthew Wilcox * @start: Index of first page to be unmapped. 3675977fbdcdSMatthew Wilcox * @nr: Number of pages to be unmapped. 0 to unmap to end of file. 3676977fbdcdSMatthew Wilcox * @even_cows: Whether to unmap even private COWed pages. 3677977fbdcdSMatthew Wilcox * 3678977fbdcdSMatthew Wilcox * Unmap the pages in this address space from any userspace process which 3679977fbdcdSMatthew Wilcox * has them mmaped. Generally, you want to remove COWed pages as well when 3680977fbdcdSMatthew Wilcox * a file is being truncated, but not when invalidating pages from the page 3681977fbdcdSMatthew Wilcox * cache. 3682977fbdcdSMatthew Wilcox */ 3683977fbdcdSMatthew Wilcox void unmap_mapping_pages(struct address_space *mapping, pgoff_t start, 3684977fbdcdSMatthew Wilcox pgoff_t nr, bool even_cows) 3685977fbdcdSMatthew Wilcox { 3686977fbdcdSMatthew Wilcox struct zap_details details = { }; 3687232a6a1cSPeter Xu pgoff_t first_index = start; 3688232a6a1cSPeter Xu pgoff_t last_index = start + nr - 1; 3689977fbdcdSMatthew Wilcox 36902e148f1eSPeter Xu details.even_cows = even_cows; 3691232a6a1cSPeter Xu if (last_index < first_index) 3692232a6a1cSPeter Xu last_index = ULONG_MAX; 3693977fbdcdSMatthew Wilcox 36942c865995SHugh Dickins i_mmap_lock_read(mapping); 3695977fbdcdSMatthew Wilcox if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))) 3696232a6a1cSPeter Xu unmap_mapping_range_tree(&mapping->i_mmap, first_index, 3697232a6a1cSPeter Xu last_index, &details); 36982c865995SHugh Dickins i_mmap_unlock_read(mapping); 3699977fbdcdSMatthew Wilcox } 37006e0e99d5SDavid Howells EXPORT_SYMBOL_GPL(unmap_mapping_pages); 3701977fbdcdSMatthew Wilcox 3702977fbdcdSMatthew Wilcox /** 37038a5f14a2SKirill A. Shutemov * unmap_mapping_range - unmap the portion of all mmaps in the specified 3704977fbdcdSMatthew Wilcox * address_space corresponding to the specified byte range in the underlying 37058a5f14a2SKirill A. Shutemov * file. 37068a5f14a2SKirill A. Shutemov * 37073d41088fSMartin Waitz * @mapping: the address space containing mmaps to be unmapped. 37081da177e4SLinus Torvalds * @holebegin: byte in first page to unmap, relative to the start of 37091da177e4SLinus Torvalds * the underlying file. This will be rounded down to a PAGE_SIZE 371025d9e2d1Snpiggin@suse.de * boundary. Note that this is different from truncate_pagecache(), which 37111da177e4SLinus Torvalds * must keep the partial page. In contrast, we must get rid of 37121da177e4SLinus Torvalds * partial pages. 37131da177e4SLinus Torvalds * @holelen: size of prospective hole in bytes. This will be rounded 37141da177e4SLinus Torvalds * up to a PAGE_SIZE boundary. A holelen of zero truncates to the 37151da177e4SLinus Torvalds * end of the file. 37161da177e4SLinus Torvalds * @even_cows: 1 when truncating a file, unmap even private COWed pages; 37171da177e4SLinus Torvalds * but 0 when invalidating pagecache, don't throw away private data. 37181da177e4SLinus Torvalds */ 37191da177e4SLinus Torvalds void unmap_mapping_range(struct address_space *mapping, 37201da177e4SLinus Torvalds loff_t const holebegin, loff_t const holelen, int even_cows) 37211da177e4SLinus Torvalds { 37229eab0421SJiajun Xie pgoff_t hba = (pgoff_t)(holebegin) >> PAGE_SHIFT; 37239eab0421SJiajun Xie pgoff_t hlen = ((pgoff_t)(holelen) + PAGE_SIZE - 1) >> PAGE_SHIFT; 37241da177e4SLinus Torvalds 37251da177e4SLinus Torvalds /* Check for overflow. */ 37261da177e4SLinus Torvalds if (sizeof(holelen) > sizeof(hlen)) { 37271da177e4SLinus Torvalds long long holeend = 37281da177e4SLinus Torvalds (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; 37291da177e4SLinus Torvalds if (holeend & ~(long long)ULONG_MAX) 37301da177e4SLinus Torvalds hlen = ULONG_MAX - hba + 1; 37311da177e4SLinus Torvalds } 37321da177e4SLinus Torvalds 3733977fbdcdSMatthew Wilcox unmap_mapping_pages(mapping, hba, hlen, even_cows); 37341da177e4SLinus Torvalds } 37351da177e4SLinus Torvalds EXPORT_SYMBOL(unmap_mapping_range); 37361da177e4SLinus Torvalds 37371da177e4SLinus Torvalds /* 3738b756a3b5SAlistair Popple * Restore a potential device exclusive pte to a working pte entry 3739b756a3b5SAlistair Popple */ 3740b756a3b5SAlistair Popple static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf) 3741b756a3b5SAlistair Popple { 374219672a9eSMatthew Wilcox (Oracle) struct folio *folio = page_folio(vmf->page); 3743b756a3b5SAlistair Popple struct vm_area_struct *vma = vmf->vma; 3744b756a3b5SAlistair Popple struct mmu_notifier_range range; 3745fdc724d6SSuren Baghdasaryan vm_fault_t ret; 3746b756a3b5SAlistair Popple 37477c7b9629SAlistair Popple /* 37487c7b9629SAlistair Popple * We need a reference to lock the folio because we don't hold 37497c7b9629SAlistair Popple * the PTL so a racing thread can remove the device-exclusive 37507c7b9629SAlistair Popple * entry and unmap it. If the folio is free the entry must 37517c7b9629SAlistair Popple * have been removed already. If it happens to have already 37527c7b9629SAlistair Popple * been re-allocated after being freed all we do is lock and 37537c7b9629SAlistair Popple * unlock it. 37547c7b9629SAlistair Popple */ 37557c7b9629SAlistair Popple if (!folio_try_get(folio)) 37567c7b9629SAlistair Popple return 0; 37577c7b9629SAlistair Popple 3758fdc724d6SSuren Baghdasaryan ret = folio_lock_or_retry(folio, vmf); 3759fdc724d6SSuren Baghdasaryan if (ret) { 37607c7b9629SAlistair Popple folio_put(folio); 3761fdc724d6SSuren Baghdasaryan return ret; 37627c7b9629SAlistair Popple } 37637d4a8be0SAlistair Popple mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, 3764b756a3b5SAlistair Popple vma->vm_mm, vmf->address & PAGE_MASK, 3765b756a3b5SAlistair Popple (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL); 3766b756a3b5SAlistair Popple mmu_notifier_invalidate_range_start(&range); 3767b756a3b5SAlistair Popple 3768b756a3b5SAlistair Popple vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 3769b756a3b5SAlistair Popple &vmf->ptl); 3770c33c7948SRyan Roberts if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte))) 377119672a9eSMatthew Wilcox (Oracle) restore_exclusive_pte(vma, vmf->page, vmf->address, vmf->pte); 3772b756a3b5SAlistair Popple 37733db82b93SHugh Dickins if (vmf->pte) 3774b756a3b5SAlistair Popple pte_unmap_unlock(vmf->pte, vmf->ptl); 377519672a9eSMatthew Wilcox (Oracle) folio_unlock(folio); 37767c7b9629SAlistair Popple folio_put(folio); 3777b756a3b5SAlistair Popple 3778b756a3b5SAlistair Popple mmu_notifier_invalidate_range_end(&range); 3779b756a3b5SAlistair Popple return 0; 3780b756a3b5SAlistair Popple } 3781b756a3b5SAlistair Popple 3782a160e537SMatthew Wilcox (Oracle) static inline bool should_try_to_free_swap(struct folio *folio, 3783c145e0b4SDavid Hildenbrand struct vm_area_struct *vma, 3784c145e0b4SDavid Hildenbrand unsigned int fault_flags) 3785c145e0b4SDavid Hildenbrand { 3786a160e537SMatthew Wilcox (Oracle) if (!folio_test_swapcache(folio)) 3787c145e0b4SDavid Hildenbrand return false; 37889202d527SMatthew Wilcox (Oracle) if (mem_cgroup_swap_full(folio) || (vma->vm_flags & VM_LOCKED) || 3789a160e537SMatthew Wilcox (Oracle) folio_test_mlocked(folio)) 3790c145e0b4SDavid Hildenbrand return true; 3791c145e0b4SDavid Hildenbrand /* 3792c145e0b4SDavid Hildenbrand * If we want to map a page that's in the swapcache writable, we 3793c145e0b4SDavid Hildenbrand * have to detect via the refcount if we're really the exclusive 3794c145e0b4SDavid Hildenbrand * user. Try freeing the swapcache to get rid of the swapcache 3795c145e0b4SDavid Hildenbrand * reference only in case it's likely that we'll be the exlusive user. 3796c145e0b4SDavid Hildenbrand */ 3797a160e537SMatthew Wilcox (Oracle) return (fault_flags & FAULT_FLAG_WRITE) && !folio_test_ksm(folio) && 3798a160e537SMatthew Wilcox (Oracle) folio_ref_count(folio) == 2; 3799c145e0b4SDavid Hildenbrand } 3800c145e0b4SDavid Hildenbrand 38019c28a205SPeter Xu static vm_fault_t pte_marker_clear(struct vm_fault *vmf) 38029c28a205SPeter Xu { 38039c28a205SPeter Xu vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, 38049c28a205SPeter Xu vmf->address, &vmf->ptl); 38053db82b93SHugh Dickins if (!vmf->pte) 38063db82b93SHugh Dickins return 0; 38079c28a205SPeter Xu /* 38089c28a205SPeter Xu * Be careful so that we will only recover a special uffd-wp pte into a 38099c28a205SPeter Xu * none pte. Otherwise it means the pte could have changed, so retry. 38107e3ce3f8SPeter Xu * 38117e3ce3f8SPeter Xu * This should also cover the case where e.g. the pte changed 3812af19487fSAxel Rasmussen * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED. 38137e3ce3f8SPeter Xu * So is_pte_marker() check is not enough to safely drop the pte. 38149c28a205SPeter Xu */ 3815c33c7948SRyan Roberts if (pte_same(vmf->orig_pte, ptep_get(vmf->pte))) 38169c28a205SPeter Xu pte_clear(vmf->vma->vm_mm, vmf->address, vmf->pte); 38179c28a205SPeter Xu pte_unmap_unlock(vmf->pte, vmf->ptl); 38189c28a205SPeter Xu return 0; 38199c28a205SPeter Xu } 38209c28a205SPeter Xu 38212bad466cSPeter Xu static vm_fault_t do_pte_missing(struct vm_fault *vmf) 38222bad466cSPeter Xu { 38232bad466cSPeter Xu if (vma_is_anonymous(vmf->vma)) 38242bad466cSPeter Xu return do_anonymous_page(vmf); 38252bad466cSPeter Xu else 38262bad466cSPeter Xu return do_fault(vmf); 38272bad466cSPeter Xu } 38282bad466cSPeter Xu 38299c28a205SPeter Xu /* 38309c28a205SPeter Xu * This is actually a page-missing access, but with uffd-wp special pte 38319c28a205SPeter Xu * installed. It means this pte was wr-protected before being unmapped. 38329c28a205SPeter Xu */ 38339c28a205SPeter Xu static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf) 38349c28a205SPeter Xu { 38359c28a205SPeter Xu /* 38369c28a205SPeter Xu * Just in case there're leftover special ptes even after the region 38377a079ba2SPeter Xu * got unregistered - we can simply clear them. 38389c28a205SPeter Xu */ 38392bad466cSPeter Xu if (unlikely(!userfaultfd_wp(vmf->vma))) 38409c28a205SPeter Xu return pte_marker_clear(vmf); 38419c28a205SPeter Xu 38422bad466cSPeter Xu return do_pte_missing(vmf); 38439c28a205SPeter Xu } 38449c28a205SPeter Xu 38455c041f5dSPeter Xu static vm_fault_t handle_pte_marker(struct vm_fault *vmf) 38465c041f5dSPeter Xu { 38475c041f5dSPeter Xu swp_entry_t entry = pte_to_swp_entry(vmf->orig_pte); 38485c041f5dSPeter Xu unsigned long marker = pte_marker_get(entry); 38495c041f5dSPeter Xu 38505c041f5dSPeter Xu /* 3851ca92ea3dSPeter Xu * PTE markers should never be empty. If anything weird happened, 3852ca92ea3dSPeter Xu * the best thing to do is to kill the process along with its mm. 38535c041f5dSPeter Xu */ 3854ca92ea3dSPeter Xu if (WARN_ON_ONCE(!marker)) 38555c041f5dSPeter Xu return VM_FAULT_SIGBUS; 38565c041f5dSPeter Xu 385715520a3fSPeter Xu /* Higher priority than uffd-wp when data corrupted */ 3858af19487fSAxel Rasmussen if (marker & PTE_MARKER_POISONED) 3859af19487fSAxel Rasmussen return VM_FAULT_HWPOISON; 386015520a3fSPeter Xu 38619c28a205SPeter Xu if (pte_marker_entry_uffd_wp(entry)) 38629c28a205SPeter Xu return pte_marker_handle_uffd_wp(vmf); 38639c28a205SPeter Xu 38649c28a205SPeter Xu /* This is an unknown pte marker */ 38659c28a205SPeter Xu return VM_FAULT_SIGBUS; 38665c041f5dSPeter Xu } 38675c041f5dSPeter Xu 3868b756a3b5SAlistair Popple /* 3869c1e8d7c6SMichel Lespinasse * We enter with non-exclusive mmap_lock (to exclude vma changes, 38708f4e2101SHugh Dickins * but allow concurrent faults), and pte mapped but not yet locked. 38719a95f3cfSPaul Cassella * We return with pte unmapped and unlocked. 38729a95f3cfSPaul Cassella * 3873c1e8d7c6SMichel Lespinasse * We return with the mmap_lock locked or unlocked in the same cases 38749a95f3cfSPaul Cassella * as does filemap_fault(). 38751da177e4SLinus Torvalds */ 38762b740303SSouptick Joarder vm_fault_t do_swap_page(struct vm_fault *vmf) 38771da177e4SLinus Torvalds { 387882b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 3879d4f9565aSMatthew Wilcox (Oracle) struct folio *swapcache, *folio = NULL; 3880d4f9565aSMatthew Wilcox (Oracle) struct page *page; 38812799e775SMiaohe Lin struct swap_info_struct *si = NULL; 388214f9135dSDavid Hildenbrand rmap_t rmap_flags = RMAP_NONE; 388313ddaf26SKairui Song bool need_clear_cache = false; 38841493a191SDavid Hildenbrand bool exclusive = false; 388565500d23SHugh Dickins swp_entry_t entry; 38861da177e4SLinus Torvalds pte_t pte; 38872b740303SSouptick Joarder vm_fault_t ret = 0; 3888aae466b0SJoonsoo Kim void *shadow = NULL; 38891da177e4SLinus Torvalds 38902ca99358SPeter Xu if (!pte_unmap_same(vmf)) 38918f4e2101SHugh Dickins goto out; 389265500d23SHugh Dickins 38932994302bSJan Kara entry = pte_to_swp_entry(vmf->orig_pte); 3894d1737fdbSAndi Kleen if (unlikely(non_swap_entry(entry))) { 38950697212aSChristoph Lameter if (is_migration_entry(entry)) { 389682b0f8c3SJan Kara migration_entry_wait(vma->vm_mm, vmf->pmd, 389782b0f8c3SJan Kara vmf->address); 3898b756a3b5SAlistair Popple } else if (is_device_exclusive_entry(entry)) { 3899b756a3b5SAlistair Popple vmf->page = pfn_swap_entry_to_page(entry); 3900b756a3b5SAlistair Popple ret = remove_device_exclusive_entry(vmf); 39015042db43SJérôme Glisse } else if (is_device_private_entry(entry)) { 39021235ccd0SSuren Baghdasaryan if (vmf->flags & FAULT_FLAG_VMA_LOCK) { 39031235ccd0SSuren Baghdasaryan /* 39041235ccd0SSuren Baghdasaryan * migrate_to_ram is not yet ready to operate 39051235ccd0SSuren Baghdasaryan * under VMA lock. 39061235ccd0SSuren Baghdasaryan */ 39071235ccd0SSuren Baghdasaryan vma_end_read(vma); 39081235ccd0SSuren Baghdasaryan ret = VM_FAULT_RETRY; 39091235ccd0SSuren Baghdasaryan goto out; 39101235ccd0SSuren Baghdasaryan } 39111235ccd0SSuren Baghdasaryan 3912af5cdaf8SAlistair Popple vmf->page = pfn_swap_entry_to_page(entry); 391316ce101dSAlistair Popple vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 391416ce101dSAlistair Popple vmf->address, &vmf->ptl); 39153db82b93SHugh Dickins if (unlikely(!vmf->pte || 3916c33c7948SRyan Roberts !pte_same(ptep_get(vmf->pte), 3917c33c7948SRyan Roberts vmf->orig_pte))) 39183b65f437SRyan Roberts goto unlock; 391916ce101dSAlistair Popple 392016ce101dSAlistair Popple /* 392116ce101dSAlistair Popple * Get a page reference while we know the page can't be 392216ce101dSAlistair Popple * freed. 392316ce101dSAlistair Popple */ 392416ce101dSAlistair Popple get_page(vmf->page); 392516ce101dSAlistair Popple pte_unmap_unlock(vmf->pte, vmf->ptl); 39264a955bedSAlistair Popple ret = vmf->page->pgmap->ops->migrate_to_ram(vmf); 392716ce101dSAlistair Popple put_page(vmf->page); 3928d1737fdbSAndi Kleen } else if (is_hwpoison_entry(entry)) { 3929d1737fdbSAndi Kleen ret = VM_FAULT_HWPOISON; 39305c041f5dSPeter Xu } else if (is_pte_marker_entry(entry)) { 39315c041f5dSPeter Xu ret = handle_pte_marker(vmf); 3932d1737fdbSAndi Kleen } else { 39332994302bSJan Kara print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL); 3934d99be1a8SHugh Dickins ret = VM_FAULT_SIGBUS; 3935d1737fdbSAndi Kleen } 39360697212aSChristoph Lameter goto out; 39370697212aSChristoph Lameter } 39380bcac06fSMinchan Kim 39392799e775SMiaohe Lin /* Prevent swapoff from happening to us. */ 39402799e775SMiaohe Lin si = get_swap_device(entry); 39412799e775SMiaohe Lin if (unlikely(!si)) 39422799e775SMiaohe Lin goto out; 39430bcac06fSMinchan Kim 39445a423081SMatthew Wilcox (Oracle) folio = swap_cache_get_folio(entry, vma, vmf->address); 39455a423081SMatthew Wilcox (Oracle) if (folio) 39465a423081SMatthew Wilcox (Oracle) page = folio_file_page(folio, swp_offset(entry)); 3947d4f9565aSMatthew Wilcox (Oracle) swapcache = folio; 3948f8020772SMinchan Kim 3949d4f9565aSMatthew Wilcox (Oracle) if (!folio) { 3950a449bf58SQian Cai if (data_race(si->flags & SWP_SYNCHRONOUS_IO) && 3951eb085574SHuang Ying __swap_count(entry) == 1) { 395213ddaf26SKairui Song /* 395313ddaf26SKairui Song * Prevent parallel swapin from proceeding with 395413ddaf26SKairui Song * the cache flag. Otherwise, another thread may 395513ddaf26SKairui Song * finish swapin first, free the entry, and swapout 395613ddaf26SKairui Song * reusing the same entry. It's undetectable as 395713ddaf26SKairui Song * pte_same() returns true due to entry reuse. 395813ddaf26SKairui Song */ 395913ddaf26SKairui Song if (swapcache_prepare(entry)) { 396013ddaf26SKairui Song /* Relax a bit to prevent rapid repeated page faults */ 396113ddaf26SKairui Song schedule_timeout_uninterruptible(1); 396213ddaf26SKairui Song goto out; 396313ddaf26SKairui Song } 396413ddaf26SKairui Song need_clear_cache = true; 396513ddaf26SKairui Song 39660bcac06fSMinchan Kim /* skip swapcache */ 396763ad4addSMatthew Wilcox (Oracle) folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, 396863ad4addSMatthew Wilcox (Oracle) vma, vmf->address, false); 396963ad4addSMatthew Wilcox (Oracle) page = &folio->page; 397063ad4addSMatthew Wilcox (Oracle) if (folio) { 397163ad4addSMatthew Wilcox (Oracle) __folio_set_locked(folio); 397263ad4addSMatthew Wilcox (Oracle) __folio_set_swapbacked(folio); 39734c6355b2SJohannes Weiner 397465995918SMatthew Wilcox (Oracle) if (mem_cgroup_swapin_charge_folio(folio, 397563ad4addSMatthew Wilcox (Oracle) vma->vm_mm, GFP_KERNEL, 397663ad4addSMatthew Wilcox (Oracle) entry)) { 3977545b1b07SMichal Hocko ret = VM_FAULT_OOM; 39784c6355b2SJohannes Weiner goto out_page; 3979545b1b07SMichal Hocko } 39800add0c77SShakeel Butt mem_cgroup_swapin_uncharge_swap(entry); 39814c6355b2SJohannes Weiner 3982aae466b0SJoonsoo Kim shadow = get_shadow_from_swap_cache(entry); 3983aae466b0SJoonsoo Kim if (shadow) 398463ad4addSMatthew Wilcox (Oracle) workingset_refault(folio, shadow); 39850076f029SJoonsoo Kim 398663ad4addSMatthew Wilcox (Oracle) folio_add_lru(folio); 39870add0c77SShakeel Butt 3988c9bdf768SMatthew Wilcox (Oracle) /* To provide entry to swap_read_folio() */ 39893d2c9087SDavid Hildenbrand folio->swap = entry; 3990c9bdf768SMatthew Wilcox (Oracle) swap_read_folio(folio, true, NULL); 399163ad4addSMatthew Wilcox (Oracle) folio->private = NULL; 39920bcac06fSMinchan Kim } 3993aa8d22a1SMinchan Kim } else { 3994e9e9b7ecSMinchan Kim page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, 3995e9e9b7ecSMinchan Kim vmf); 399663ad4addSMatthew Wilcox (Oracle) if (page) 399763ad4addSMatthew Wilcox (Oracle) folio = page_folio(page); 3998d4f9565aSMatthew Wilcox (Oracle) swapcache = folio; 39990bcac06fSMinchan Kim } 40000bcac06fSMinchan Kim 4001d4f9565aSMatthew Wilcox (Oracle) if (!folio) { 40021da177e4SLinus Torvalds /* 40038f4e2101SHugh Dickins * Back out if somebody else faulted in this pte 40048f4e2101SHugh Dickins * while we released the pte lock. 40051da177e4SLinus Torvalds */ 400682b0f8c3SJan Kara vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 400782b0f8c3SJan Kara vmf->address, &vmf->ptl); 4008c33c7948SRyan Roberts if (likely(vmf->pte && 4009c33c7948SRyan Roberts pte_same(ptep_get(vmf->pte), vmf->orig_pte))) 40101da177e4SLinus Torvalds ret = VM_FAULT_OOM; 401165500d23SHugh Dickins goto unlock; 40121da177e4SLinus Torvalds } 40131da177e4SLinus Torvalds 40141da177e4SLinus Torvalds /* Had to read the page from swap area: Major fault */ 40151da177e4SLinus Torvalds ret = VM_FAULT_MAJOR; 4016f8891e5eSChristoph Lameter count_vm_event(PGMAJFAULT); 40172262185cSRoman Gushchin count_memcg_event_mm(vma->vm_mm, PGMAJFAULT); 4018d1737fdbSAndi Kleen } else if (PageHWPoison(page)) { 401971f72525SWu Fengguang /* 402071f72525SWu Fengguang * hwpoisoned dirty swapcache pages are kept for killing 402171f72525SWu Fengguang * owner processes (which may be unknown at hwpoison time) 402271f72525SWu Fengguang */ 4023d1737fdbSAndi Kleen ret = VM_FAULT_HWPOISON; 40244779cb31SAndi Kleen goto out_release; 40251da177e4SLinus Torvalds } 40261da177e4SLinus Torvalds 4027fdc724d6SSuren Baghdasaryan ret |= folio_lock_or_retry(folio, vmf); 4028fdc724d6SSuren Baghdasaryan if (ret & VM_FAULT_RETRY) 4029d065bd81SMichel Lespinasse goto out_release; 40301da177e4SLinus Torvalds 403184d60fddSDavid Hildenbrand if (swapcache) { 40324969c119SAndrea Arcangeli /* 40333b344157SMatthew Wilcox (Oracle) * Make sure folio_free_swap() or swapoff did not release the 403484d60fddSDavid Hildenbrand * swapcache from under us. The page pin, and pte_same test 403584d60fddSDavid Hildenbrand * below, are not enough to exclude that. Even if it is still 403684d60fddSDavid Hildenbrand * swapcache, we need to check that the page's swap has not 403784d60fddSDavid Hildenbrand * changed. 40384969c119SAndrea Arcangeli */ 403963ad4addSMatthew Wilcox (Oracle) if (unlikely(!folio_test_swapcache(folio) || 4040cfeed8ffSDavid Hildenbrand page_swap_entry(page).val != entry.val)) 40414969c119SAndrea Arcangeli goto out_page; 40424969c119SAndrea Arcangeli 404384d60fddSDavid Hildenbrand /* 404484d60fddSDavid Hildenbrand * KSM sometimes has to copy on read faults, for example, if 404584d60fddSDavid Hildenbrand * page->index of !PageKSM() pages would be nonlinear inside the 404684d60fddSDavid Hildenbrand * anon VMA -- PageKSM() is lost on actual swapout. 404784d60fddSDavid Hildenbrand */ 404896db66d9SMatthew Wilcox (Oracle) folio = ksm_might_need_to_copy(folio, vma, vmf->address); 404996db66d9SMatthew Wilcox (Oracle) if (unlikely(!folio)) { 40505ad64688SHugh Dickins ret = VM_FAULT_OOM; 405196db66d9SMatthew Wilcox (Oracle) folio = swapcache; 40524969c119SAndrea Arcangeli goto out_page; 405396db66d9SMatthew Wilcox (Oracle) } else if (unlikely(folio == ERR_PTR(-EHWPOISON))) { 40546b970599SKefeng Wang ret = VM_FAULT_HWPOISON; 405596db66d9SMatthew Wilcox (Oracle) folio = swapcache; 40566b970599SKefeng Wang goto out_page; 40574969c119SAndrea Arcangeli } 405896db66d9SMatthew Wilcox (Oracle) if (folio != swapcache) 405996db66d9SMatthew Wilcox (Oracle) page = folio_page(folio, 0); 4060c145e0b4SDavid Hildenbrand 4061c145e0b4SDavid Hildenbrand /* 4062c145e0b4SDavid Hildenbrand * If we want to map a page that's in the swapcache writable, we 4063c145e0b4SDavid Hildenbrand * have to detect via the refcount if we're really the exclusive 4064c145e0b4SDavid Hildenbrand * owner. Try removing the extra reference from the local LRU 40651fec6890SMatthew Wilcox (Oracle) * caches if required. 4066c145e0b4SDavid Hildenbrand */ 4067d4f9565aSMatthew Wilcox (Oracle) if ((vmf->flags & FAULT_FLAG_WRITE) && folio == swapcache && 406863ad4addSMatthew Wilcox (Oracle) !folio_test_ksm(folio) && !folio_test_lru(folio)) 4069c145e0b4SDavid Hildenbrand lru_add_drain(); 407084d60fddSDavid Hildenbrand } 40715ad64688SHugh Dickins 40724231f842SKefeng Wang folio_throttle_swaprate(folio, GFP_KERNEL); 4073073e587eSKAMEZAWA Hiroyuki 40741da177e4SLinus Torvalds /* 40758f4e2101SHugh Dickins * Back out if somebody else already faulted in this pte. 40761da177e4SLinus Torvalds */ 407782b0f8c3SJan Kara vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, 407882b0f8c3SJan Kara &vmf->ptl); 4079c33c7948SRyan Roberts if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) 4080b8107480SKirill Korotaev goto out_nomap; 4081b8107480SKirill Korotaev 408263ad4addSMatthew Wilcox (Oracle) if (unlikely(!folio_test_uptodate(folio))) { 4083b8107480SKirill Korotaev ret = VM_FAULT_SIGBUS; 4084b8107480SKirill Korotaev goto out_nomap; 40851da177e4SLinus Torvalds } 40861da177e4SLinus Torvalds 40878c7c6e34SKAMEZAWA Hiroyuki /* 408878fbe906SDavid Hildenbrand * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte 408978fbe906SDavid Hildenbrand * must never point at an anonymous page in the swapcache that is 409078fbe906SDavid Hildenbrand * PG_anon_exclusive. Sanity check that this holds and especially, that 409178fbe906SDavid Hildenbrand * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity 409278fbe906SDavid Hildenbrand * check after taking the PT lock and making sure that nobody 409378fbe906SDavid Hildenbrand * concurrently faulted in this page and set PG_anon_exclusive. 409478fbe906SDavid Hildenbrand */ 409563ad4addSMatthew Wilcox (Oracle) BUG_ON(!folio_test_anon(folio) && folio_test_mappedtodisk(folio)); 409663ad4addSMatthew Wilcox (Oracle) BUG_ON(folio_test_anon(folio) && PageAnonExclusive(page)); 409778fbe906SDavid Hildenbrand 409878fbe906SDavid Hildenbrand /* 40991493a191SDavid Hildenbrand * Check under PT lock (to protect against concurrent fork() sharing 41001493a191SDavid Hildenbrand * the swap entry concurrently) for certainly exclusive pages. 41011493a191SDavid Hildenbrand */ 410263ad4addSMatthew Wilcox (Oracle) if (!folio_test_ksm(folio)) { 41031493a191SDavid Hildenbrand exclusive = pte_swp_exclusive(vmf->orig_pte); 4104d4f9565aSMatthew Wilcox (Oracle) if (folio != swapcache) { 41051493a191SDavid Hildenbrand /* 41061493a191SDavid Hildenbrand * We have a fresh page that is not exposed to the 41071493a191SDavid Hildenbrand * swapcache -> certainly exclusive. 41081493a191SDavid Hildenbrand */ 41091493a191SDavid Hildenbrand exclusive = true; 411063ad4addSMatthew Wilcox (Oracle) } else if (exclusive && folio_test_writeback(folio) && 4111eacde327SMiaohe Lin data_race(si->flags & SWP_STABLE_WRITES)) { 41121493a191SDavid Hildenbrand /* 41131493a191SDavid Hildenbrand * This is tricky: not all swap backends support 41141493a191SDavid Hildenbrand * concurrent page modifications while under writeback. 41151493a191SDavid Hildenbrand * 41161493a191SDavid Hildenbrand * So if we stumble over such a page in the swapcache 41171493a191SDavid Hildenbrand * we must not set the page exclusive, otherwise we can 41181493a191SDavid Hildenbrand * map it writable without further checks and modify it 41191493a191SDavid Hildenbrand * while still under writeback. 41201493a191SDavid Hildenbrand * 41211493a191SDavid Hildenbrand * For these problematic swap backends, simply drop the 41221493a191SDavid Hildenbrand * exclusive marker: this is perfectly fine as we start 41231493a191SDavid Hildenbrand * writeback only if we fully unmapped the page and 41241493a191SDavid Hildenbrand * there are no unexpected references on the page after 41251493a191SDavid Hildenbrand * unmapping succeeded. After fully unmapped, no 41261493a191SDavid Hildenbrand * further GUP references (FOLL_GET and FOLL_PIN) can 41271493a191SDavid Hildenbrand * appear, so dropping the exclusive marker and mapping 41281493a191SDavid Hildenbrand * it only R/O is fine. 41291493a191SDavid Hildenbrand */ 41301493a191SDavid Hildenbrand exclusive = false; 41311493a191SDavid Hildenbrand } 41321493a191SDavid Hildenbrand } 41331493a191SDavid Hildenbrand 41341493a191SDavid Hildenbrand /* 41356dca4ac6SPeter Collingbourne * Some architectures may have to restore extra metadata to the page 41366dca4ac6SPeter Collingbourne * when reading from swap. This metadata may be indexed by swap entry 41376dca4ac6SPeter Collingbourne * so this must be called before swap_free(). 41386dca4ac6SPeter Collingbourne */ 41396dca4ac6SPeter Collingbourne arch_swap_restore(entry, folio); 41406dca4ac6SPeter Collingbourne 41416dca4ac6SPeter Collingbourne /* 4142c145e0b4SDavid Hildenbrand * Remove the swap entry and conditionally try to free up the swapcache. 4143c145e0b4SDavid Hildenbrand * We're already holding a reference on the page but haven't mapped it 4144c145e0b4SDavid Hildenbrand * yet. 41458c7c6e34SKAMEZAWA Hiroyuki */ 4146c145e0b4SDavid Hildenbrand swap_free(entry); 4147a160e537SMatthew Wilcox (Oracle) if (should_try_to_free_swap(folio, vma, vmf->flags)) 4148a160e537SMatthew Wilcox (Oracle) folio_free_swap(folio); 41491da177e4SLinus Torvalds 4150f1a79412SShakeel Butt inc_mm_counter(vma->vm_mm, MM_ANONPAGES); 4151f1a79412SShakeel Butt dec_mm_counter(vma->vm_mm, MM_SWAPENTS); 41521da177e4SLinus Torvalds pte = mk_pte(page, vma->vm_page_prot); 4153c145e0b4SDavid Hildenbrand 4154c145e0b4SDavid Hildenbrand /* 41551493a191SDavid Hildenbrand * Same logic as in do_wp_page(); however, optimize for pages that are 41561493a191SDavid Hildenbrand * certainly not shared either because we just allocated them without 41571493a191SDavid Hildenbrand * exposing them to the swapcache or because the swap entry indicates 41581493a191SDavid Hildenbrand * exclusivity. 4159c145e0b4SDavid Hildenbrand */ 416063ad4addSMatthew Wilcox (Oracle) if (!folio_test_ksm(folio) && 416163ad4addSMatthew Wilcox (Oracle) (exclusive || folio_ref_count(folio) == 1)) { 41626c287605SDavid Hildenbrand if (vmf->flags & FAULT_FLAG_WRITE) { 41631da177e4SLinus Torvalds pte = maybe_mkwrite(pte_mkdirty(pte), vma); 416482b0f8c3SJan Kara vmf->flags &= ~FAULT_FLAG_WRITE; 41656c287605SDavid Hildenbrand } 416614f9135dSDavid Hildenbrand rmap_flags |= RMAP_EXCLUSIVE; 41671da177e4SLinus Torvalds } 41681da177e4SLinus Torvalds flush_icache_page(vma, page); 41692994302bSJan Kara if (pte_swp_soft_dirty(vmf->orig_pte)) 4170179ef71cSCyrill Gorcunov pte = pte_mksoft_dirty(pte); 4171f1eb1bacSPeter Xu if (pte_swp_uffd_wp(vmf->orig_pte)) 4172f45ec5ffSPeter Xu pte = pte_mkuffd_wp(pte); 41732994302bSJan Kara vmf->orig_pte = pte; 41740bcac06fSMinchan Kim 41750bcac06fSMinchan Kim /* ksm created a completely new copy */ 4176d4f9565aSMatthew Wilcox (Oracle) if (unlikely(folio != swapcache && swapcache)) { 41772853b66bSMatthew Wilcox (Oracle) folio_add_new_anon_rmap(folio, vma, vmf->address); 417863ad4addSMatthew Wilcox (Oracle) folio_add_lru_vma(folio, vma); 41790bcac06fSMinchan Kim } else { 4180b832a354SDavid Hildenbrand folio_add_anon_rmap_pte(folio, page, vma, vmf->address, 4181b832a354SDavid Hildenbrand rmap_flags); 418200501b53SJohannes Weiner } 41831da177e4SLinus Torvalds 418463ad4addSMatthew Wilcox (Oracle) VM_BUG_ON(!folio_test_anon(folio) || 418563ad4addSMatthew Wilcox (Oracle) (pte_write(pte) && !PageAnonExclusive(page))); 41861eba86c0SPasha Tatashin set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte); 41871eba86c0SPasha Tatashin arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte); 41881eba86c0SPasha Tatashin 418963ad4addSMatthew Wilcox (Oracle) folio_unlock(folio); 4190d4f9565aSMatthew Wilcox (Oracle) if (folio != swapcache && swapcache) { 41914969c119SAndrea Arcangeli /* 41924969c119SAndrea Arcangeli * Hold the lock to avoid the swap entry to be reused 41934969c119SAndrea Arcangeli * until we take the PT lock for the pte_same() check 41944969c119SAndrea Arcangeli * (to avoid false positives from pte_same). For 41954969c119SAndrea Arcangeli * further safety release the lock after the swap_free 41964969c119SAndrea Arcangeli * so that the swap count won't change under a 41974969c119SAndrea Arcangeli * parallel locked swapcache. 41984969c119SAndrea Arcangeli */ 4199d4f9565aSMatthew Wilcox (Oracle) folio_unlock(swapcache); 4200d4f9565aSMatthew Wilcox (Oracle) folio_put(swapcache); 42014969c119SAndrea Arcangeli } 4202c475a8abSHugh Dickins 420382b0f8c3SJan Kara if (vmf->flags & FAULT_FLAG_WRITE) { 42042994302bSJan Kara ret |= do_wp_page(vmf); 420561469f1dSHugh Dickins if (ret & VM_FAULT_ERROR) 420661469f1dSHugh Dickins ret &= VM_FAULT_ERROR; 42071da177e4SLinus Torvalds goto out; 42081da177e4SLinus Torvalds } 42091da177e4SLinus Torvalds 42101da177e4SLinus Torvalds /* No need to invalidate - it was non-present before */ 42115003a2bdSMatthew Wilcox (Oracle) update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1); 421265500d23SHugh Dickins unlock: 42133db82b93SHugh Dickins if (vmf->pte) 421482b0f8c3SJan Kara pte_unmap_unlock(vmf->pte, vmf->ptl); 42151da177e4SLinus Torvalds out: 421613ddaf26SKairui Song /* Clear the swap cache pin for direct swapin after PTL unlock */ 421713ddaf26SKairui Song if (need_clear_cache) 421813ddaf26SKairui Song swapcache_clear(si, entry); 42192799e775SMiaohe Lin if (si) 42202799e775SMiaohe Lin put_swap_device(si); 42211da177e4SLinus Torvalds return ret; 4222b8107480SKirill Korotaev out_nomap: 42233db82b93SHugh Dickins if (vmf->pte) 422482b0f8c3SJan Kara pte_unmap_unlock(vmf->pte, vmf->ptl); 4225bc43f75cSJohannes Weiner out_page: 422663ad4addSMatthew Wilcox (Oracle) folio_unlock(folio); 42274779cb31SAndi Kleen out_release: 422863ad4addSMatthew Wilcox (Oracle) folio_put(folio); 4229d4f9565aSMatthew Wilcox (Oracle) if (folio != swapcache && swapcache) { 4230d4f9565aSMatthew Wilcox (Oracle) folio_unlock(swapcache); 4231d4f9565aSMatthew Wilcox (Oracle) folio_put(swapcache); 42324969c119SAndrea Arcangeli } 423313ddaf26SKairui Song if (need_clear_cache) 423413ddaf26SKairui Song swapcache_clear(si, entry); 42352799e775SMiaohe Lin if (si) 42362799e775SMiaohe Lin put_swap_device(si); 423765500d23SHugh Dickins return ret; 42381da177e4SLinus Torvalds } 42391da177e4SLinus Torvalds 424019eaf449SRyan Roberts static bool pte_range_none(pte_t *pte, int nr_pages) 424119eaf449SRyan Roberts { 424219eaf449SRyan Roberts int i; 424319eaf449SRyan Roberts 424419eaf449SRyan Roberts for (i = 0; i < nr_pages; i++) { 424519eaf449SRyan Roberts if (!pte_none(ptep_get_lockless(pte + i))) 424619eaf449SRyan Roberts return false; 424719eaf449SRyan Roberts } 424819eaf449SRyan Roberts 424919eaf449SRyan Roberts return true; 425019eaf449SRyan Roberts } 425119eaf449SRyan Roberts 425219eaf449SRyan Roberts static struct folio *alloc_anon_folio(struct vm_fault *vmf) 425319eaf449SRyan Roberts { 425419eaf449SRyan Roberts struct vm_area_struct *vma = vmf->vma; 4255085ff35eSKefeng Wang #ifdef CONFIG_TRANSPARENT_HUGEPAGE 425619eaf449SRyan Roberts unsigned long orders; 425719eaf449SRyan Roberts struct folio *folio; 425819eaf449SRyan Roberts unsigned long addr; 425919eaf449SRyan Roberts pte_t *pte; 426019eaf449SRyan Roberts gfp_t gfp; 426119eaf449SRyan Roberts int order; 426219eaf449SRyan Roberts 426319eaf449SRyan Roberts /* 426419eaf449SRyan Roberts * If uffd is active for the vma we need per-page fault fidelity to 426519eaf449SRyan Roberts * maintain the uffd semantics. 426619eaf449SRyan Roberts */ 426719eaf449SRyan Roberts if (unlikely(userfaultfd_armed(vma))) 426819eaf449SRyan Roberts goto fallback; 426919eaf449SRyan Roberts 427019eaf449SRyan Roberts /* 427119eaf449SRyan Roberts * Get a list of all the (large) orders below PMD_ORDER that are enabled 427219eaf449SRyan Roberts * for this vma. Then filter out the orders that can't be allocated over 427319eaf449SRyan Roberts * the faulting address and still be fully contained in the vma. 427419eaf449SRyan Roberts */ 427519eaf449SRyan Roberts orders = thp_vma_allowable_orders(vma, vma->vm_flags, false, true, true, 427619eaf449SRyan Roberts BIT(PMD_ORDER) - 1); 427719eaf449SRyan Roberts orders = thp_vma_suitable_orders(vma, vmf->address, orders); 427819eaf449SRyan Roberts 427919eaf449SRyan Roberts if (!orders) 428019eaf449SRyan Roberts goto fallback; 428119eaf449SRyan Roberts 428219eaf449SRyan Roberts pte = pte_offset_map(vmf->pmd, vmf->address & PMD_MASK); 428319eaf449SRyan Roberts if (!pte) 428419eaf449SRyan Roberts return ERR_PTR(-EAGAIN); 428519eaf449SRyan Roberts 428619eaf449SRyan Roberts /* 428719eaf449SRyan Roberts * Find the highest order where the aligned range is completely 428819eaf449SRyan Roberts * pte_none(). Note that all remaining orders will be completely 428919eaf449SRyan Roberts * pte_none(). 429019eaf449SRyan Roberts */ 429119eaf449SRyan Roberts order = highest_order(orders); 429219eaf449SRyan Roberts while (orders) { 429319eaf449SRyan Roberts addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order); 429419eaf449SRyan Roberts if (pte_range_none(pte + pte_index(addr), 1 << order)) 429519eaf449SRyan Roberts break; 429619eaf449SRyan Roberts order = next_order(&orders, order); 429719eaf449SRyan Roberts } 429819eaf449SRyan Roberts 429919eaf449SRyan Roberts pte_unmap(pte); 430019eaf449SRyan Roberts 430119eaf449SRyan Roberts /* Try allocating the highest of the remaining orders. */ 430219eaf449SRyan Roberts gfp = vma_thp_gfp_mask(vma); 430319eaf449SRyan Roberts while (orders) { 430419eaf449SRyan Roberts addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order); 430519eaf449SRyan Roberts folio = vma_alloc_folio(gfp, order, vma, addr, true); 430619eaf449SRyan Roberts if (folio) { 4307085ff35eSKefeng Wang if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) { 4308085ff35eSKefeng Wang folio_put(folio); 4309085ff35eSKefeng Wang goto next; 4310085ff35eSKefeng Wang } 4311085ff35eSKefeng Wang folio_throttle_swaprate(folio, gfp); 431219eaf449SRyan Roberts clear_huge_page(&folio->page, vmf->address, 1 << order); 431319eaf449SRyan Roberts return folio; 431419eaf449SRyan Roberts } 4315085ff35eSKefeng Wang next: 431619eaf449SRyan Roberts order = next_order(&orders, order); 431719eaf449SRyan Roberts } 431819eaf449SRyan Roberts 431919eaf449SRyan Roberts fallback: 432019eaf449SRyan Roberts #endif 4321085ff35eSKefeng Wang return folio_prealloc(vma->vm_mm, vma, vmf->address, true); 432219eaf449SRyan Roberts } 432319eaf449SRyan Roberts 43241da177e4SLinus Torvalds /* 4325c1e8d7c6SMichel Lespinasse * We enter with non-exclusive mmap_lock (to exclude vma changes, 43268f4e2101SHugh Dickins * but allow concurrent faults), and pte mapped but not yet locked. 4327c1e8d7c6SMichel Lespinasse * We return with mmap_lock still held, but pte unmapped and unlocked. 43281da177e4SLinus Torvalds */ 43292b740303SSouptick Joarder static vm_fault_t do_anonymous_page(struct vm_fault *vmf) 43301da177e4SLinus Torvalds { 43312bad466cSPeter Xu bool uffd_wp = vmf_orig_pte_uffd_wp(vmf); 433282b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 433319eaf449SRyan Roberts unsigned long addr = vmf->address; 43346bc56a4dSMatthew Wilcox (Oracle) struct folio *folio; 43352b740303SSouptick Joarder vm_fault_t ret = 0; 433619eaf449SRyan Roberts int nr_pages = 1; 43371da177e4SLinus Torvalds pte_t entry; 433819eaf449SRyan Roberts int i; 43391da177e4SLinus Torvalds 43406b7339f4SKirill A. Shutemov /* File mapping without ->vm_ops ? */ 43416b7339f4SKirill A. Shutemov if (vma->vm_flags & VM_SHARED) 43426b7339f4SKirill A. Shutemov return VM_FAULT_SIGBUS; 43436b7339f4SKirill A. Shutemov 43447267ec00SKirill A. Shutemov /* 43453db82b93SHugh Dickins * Use pte_alloc() instead of pte_alloc_map(), so that OOM can 43463db82b93SHugh Dickins * be distinguished from a transient failure of pte_offset_map(). 43477267ec00SKirill A. Shutemov */ 43484cf58924SJoel Fernandes (Google) if (pte_alloc(vma->vm_mm, vmf->pmd)) 43497267ec00SKirill A. Shutemov return VM_FAULT_OOM; 43507267ec00SKirill A. Shutemov 435111ac5524SLinus Torvalds /* Use the zero-page for reads */ 435282b0f8c3SJan Kara if (!(vmf->flags & FAULT_FLAG_WRITE) && 4353bae473a4SKirill A. Shutemov !mm_forbids_zeropage(vma->vm_mm)) { 435482b0f8c3SJan Kara entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address), 435562eede62SHugh Dickins vma->vm_page_prot)); 435682b0f8c3SJan Kara vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 435782b0f8c3SJan Kara vmf->address, &vmf->ptl); 43583db82b93SHugh Dickins if (!vmf->pte) 43593db82b93SHugh Dickins goto unlock; 43602bad466cSPeter Xu if (vmf_pte_changed(vmf)) { 43617df67697SBibo Mao update_mmu_tlb(vma, vmf->address, vmf->pte); 4362a13ea5b7SHugh Dickins goto unlock; 43637df67697SBibo Mao } 43646b31d595SMichal Hocko ret = check_stable_address_space(vma->vm_mm); 43656b31d595SMichal Hocko if (ret) 43666b31d595SMichal Hocko goto unlock; 43676b251fc9SAndrea Arcangeli /* Deliver the page fault to userland, check inside PT lock */ 43686b251fc9SAndrea Arcangeli if (userfaultfd_missing(vma)) { 436982b0f8c3SJan Kara pte_unmap_unlock(vmf->pte, vmf->ptl); 437082b0f8c3SJan Kara return handle_userfault(vmf, VM_UFFD_MISSING); 43716b251fc9SAndrea Arcangeli } 4372a13ea5b7SHugh Dickins goto setpte; 4373a13ea5b7SHugh Dickins } 4374a13ea5b7SHugh Dickins 43751da177e4SLinus Torvalds /* Allocate our own private page. */ 43761da177e4SLinus Torvalds if (unlikely(anon_vma_prepare(vma))) 437765500d23SHugh Dickins goto oom; 437819eaf449SRyan Roberts /* Returns NULL on OOM or ERR_PTR(-EAGAIN) if we must retry the fault */ 437919eaf449SRyan Roberts folio = alloc_anon_folio(vmf); 438019eaf449SRyan Roberts if (IS_ERR(folio)) 438119eaf449SRyan Roberts return 0; 43826bc56a4dSMatthew Wilcox (Oracle) if (!folio) 438365500d23SHugh Dickins goto oom; 4384eb3c24f3SMel Gorman 438519eaf449SRyan Roberts nr_pages = folio_nr_pages(folio); 438619eaf449SRyan Roberts addr = ALIGN_DOWN(vmf->address, nr_pages * PAGE_SIZE); 438719eaf449SRyan Roberts 438852f37629SMinchan Kim /* 4389cb3184deSMatthew Wilcox (Oracle) * The memory barrier inside __folio_mark_uptodate makes sure that 4390f4f5329dSWei Yang * preceding stores to the page contents become visible before 439152f37629SMinchan Kim * the set_pte_at() write. 439252f37629SMinchan Kim */ 4393cb3184deSMatthew Wilcox (Oracle) __folio_mark_uptodate(folio); 43941da177e4SLinus Torvalds 4395cb3184deSMatthew Wilcox (Oracle) entry = mk_pte(&folio->page, vma->vm_page_prot); 439650c25ee9SThomas Bogendoerfer entry = pte_sw_mkyoung(entry); 43971ac0cb5dSHugh Dickins if (vma->vm_flags & VM_WRITE) 4398161e393cSRick Edgecombe entry = pte_mkwrite(pte_mkdirty(entry), vma); 43998f4e2101SHugh Dickins 440019eaf449SRyan Roberts vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl); 44013db82b93SHugh Dickins if (!vmf->pte) 44023db82b93SHugh Dickins goto release; 440319eaf449SRyan Roberts if (nr_pages == 1 && vmf_pte_changed(vmf)) { 440419eaf449SRyan Roberts update_mmu_tlb(vma, addr, vmf->pte); 440519eaf449SRyan Roberts goto release; 440619eaf449SRyan Roberts } else if (nr_pages > 1 && !pte_range_none(vmf->pte, nr_pages)) { 440719eaf449SRyan Roberts for (i = 0; i < nr_pages; i++) 440819eaf449SRyan Roberts update_mmu_tlb(vma, addr + PAGE_SIZE * i, vmf->pte + i); 44098f4e2101SHugh Dickins goto release; 44107df67697SBibo Mao } 44119ba69294SHugh Dickins 44126b31d595SMichal Hocko ret = check_stable_address_space(vma->vm_mm); 44136b31d595SMichal Hocko if (ret) 44146b31d595SMichal Hocko goto release; 44156b31d595SMichal Hocko 44166b251fc9SAndrea Arcangeli /* Deliver the page fault to userland, check inside PT lock */ 44176b251fc9SAndrea Arcangeli if (userfaultfd_missing(vma)) { 441882b0f8c3SJan Kara pte_unmap_unlock(vmf->pte, vmf->ptl); 4419cb3184deSMatthew Wilcox (Oracle) folio_put(folio); 442082b0f8c3SJan Kara return handle_userfault(vmf, VM_UFFD_MISSING); 44216b251fc9SAndrea Arcangeli } 44226b251fc9SAndrea Arcangeli 442319eaf449SRyan Roberts folio_ref_add(folio, nr_pages - 1); 442419eaf449SRyan Roberts add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr_pages); 442519eaf449SRyan Roberts folio_add_new_anon_rmap(folio, vma, addr); 4426cb3184deSMatthew Wilcox (Oracle) folio_add_lru_vma(folio, vma); 4427a13ea5b7SHugh Dickins setpte: 44282bad466cSPeter Xu if (uffd_wp) 44292bad466cSPeter Xu entry = pte_mkuffd_wp(entry); 443019eaf449SRyan Roberts set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr_pages); 44311da177e4SLinus Torvalds 44321da177e4SLinus Torvalds /* No need to invalidate - it was non-present before */ 443319eaf449SRyan Roberts update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr_pages); 443465500d23SHugh Dickins unlock: 44353db82b93SHugh Dickins if (vmf->pte) 443682b0f8c3SJan Kara pte_unmap_unlock(vmf->pte, vmf->ptl); 44376b31d595SMichal Hocko return ret; 44388f4e2101SHugh Dickins release: 4439cb3184deSMatthew Wilcox (Oracle) folio_put(folio); 44408f4e2101SHugh Dickins goto unlock; 444165500d23SHugh Dickins oom: 44421da177e4SLinus Torvalds return VM_FAULT_OOM; 44431da177e4SLinus Torvalds } 44441da177e4SLinus Torvalds 44459a95f3cfSPaul Cassella /* 4446c1e8d7c6SMichel Lespinasse * The mmap_lock must have been held on entry, and may have been 44479a95f3cfSPaul Cassella * released depending on flags and vma->vm_ops->fault() return value. 44489a95f3cfSPaul Cassella * See filemap_fault() and __lock_page_retry(). 44499a95f3cfSPaul Cassella */ 44502b740303SSouptick Joarder static vm_fault_t __do_fault(struct vm_fault *vmf) 44517eae74afSKirill A. Shutemov { 445282b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 445301d1e0e6SMatthew Wilcox (Oracle) struct folio *folio; 44542b740303SSouptick Joarder vm_fault_t ret; 44557eae74afSKirill A. Shutemov 445663f3655fSMichal Hocko /* 445763f3655fSMichal Hocko * Preallocate pte before we take page_lock because this might lead to 445863f3655fSMichal Hocko * deadlocks for memcg reclaim which waits for pages under writeback: 445963f3655fSMichal Hocko * lock_page(A) 446063f3655fSMichal Hocko * SetPageWriteback(A) 446163f3655fSMichal Hocko * unlock_page(A) 446263f3655fSMichal Hocko * lock_page(B) 446363f3655fSMichal Hocko * lock_page(B) 4464d383807aSYanfei Xu * pte_alloc_one 446563f3655fSMichal Hocko * shrink_page_list 446663f3655fSMichal Hocko * wait_on_page_writeback(A) 446763f3655fSMichal Hocko * SetPageWriteback(B) 446863f3655fSMichal Hocko * unlock_page(B) 446963f3655fSMichal Hocko * # flush A, B to clear the writeback 447063f3655fSMichal Hocko */ 447163f3655fSMichal Hocko if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) { 4472a7069ee3SYanfei Xu vmf->prealloc_pte = pte_alloc_one(vma->vm_mm); 447363f3655fSMichal Hocko if (!vmf->prealloc_pte) 447463f3655fSMichal Hocko return VM_FAULT_OOM; 447563f3655fSMichal Hocko } 447663f3655fSMichal Hocko 447711bac800SDave Jiang ret = vma->vm_ops->fault(vmf); 44783917048dSJan Kara if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY | 4479b1aa812bSJan Kara VM_FAULT_DONE_COW))) 44807eae74afSKirill A. Shutemov return ret; 44817eae74afSKirill A. Shutemov 448201d1e0e6SMatthew Wilcox (Oracle) folio = page_folio(vmf->page); 4483667240e0SJan Kara if (unlikely(PageHWPoison(vmf->page))) { 4484e53ac737SRik van Riel vm_fault_t poisonret = VM_FAULT_HWPOISON; 4485e53ac737SRik van Riel if (ret & VM_FAULT_LOCKED) { 448601d1e0e6SMatthew Wilcox (Oracle) if (page_mapped(vmf->page)) 448701d1e0e6SMatthew Wilcox (Oracle) unmap_mapping_folio(folio); 448801d1e0e6SMatthew Wilcox (Oracle) /* Retry if a clean folio was removed from the cache. */ 448901d1e0e6SMatthew Wilcox (Oracle) if (mapping_evict_folio(folio->mapping, folio)) 44903149c79fSRik van Riel poisonret = VM_FAULT_NOPAGE; 449101d1e0e6SMatthew Wilcox (Oracle) folio_unlock(folio); 4492e53ac737SRik van Riel } 449301d1e0e6SMatthew Wilcox (Oracle) folio_put(folio); 4494936ca80dSJan Kara vmf->page = NULL; 4495e53ac737SRik van Riel return poisonret; 44967eae74afSKirill A. Shutemov } 44977eae74afSKirill A. Shutemov 44987eae74afSKirill A. Shutemov if (unlikely(!(ret & VM_FAULT_LOCKED))) 449901d1e0e6SMatthew Wilcox (Oracle) folio_lock(folio); 45007eae74afSKirill A. Shutemov else 450101d1e0e6SMatthew Wilcox (Oracle) VM_BUG_ON_PAGE(!folio_test_locked(folio), vmf->page); 45027eae74afSKirill A. Shutemov 45037eae74afSKirill A. Shutemov return ret; 45047eae74afSKirill A. Shutemov } 45057eae74afSKirill A. Shutemov 4506396bcc52SMatthew Wilcox (Oracle) #ifdef CONFIG_TRANSPARENT_HUGEPAGE 450782b0f8c3SJan Kara static void deposit_prealloc_pte(struct vm_fault *vmf) 4508953c66c2SAneesh Kumar K.V { 450982b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 4510953c66c2SAneesh Kumar K.V 451182b0f8c3SJan Kara pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte); 4512953c66c2SAneesh Kumar K.V /* 4513953c66c2SAneesh Kumar K.V * We are going to consume the prealloc table, 4514953c66c2SAneesh Kumar K.V * count that as nr_ptes. 4515953c66c2SAneesh Kumar K.V */ 4516c4812909SKirill A. Shutemov mm_inc_nr_ptes(vma->vm_mm); 45177f2b6ce8STobin C Harding vmf->prealloc_pte = NULL; 4518953c66c2SAneesh Kumar K.V } 4519953c66c2SAneesh Kumar K.V 4520f9ce0be7SKirill A. Shutemov vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page) 452110102459SKirill A. Shutemov { 4522ef37b2eaSDavid Hildenbrand struct folio *folio = page_folio(page); 452382b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 452482b0f8c3SJan Kara bool write = vmf->flags & FAULT_FLAG_WRITE; 452582b0f8c3SJan Kara unsigned long haddr = vmf->address & HPAGE_PMD_MASK; 452610102459SKirill A. Shutemov pmd_t entry; 4527d01ac3c3SMatthew Wilcox (Oracle) vm_fault_t ret = VM_FAULT_FALLBACK; 452810102459SKirill A. Shutemov 45293485b883SRyan Roberts if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER)) 4530d01ac3c3SMatthew Wilcox (Oracle) return ret; 453110102459SKirill A. Shutemov 4532ef37b2eaSDavid Hildenbrand if (page != &folio->page || folio_order(folio) != HPAGE_PMD_ORDER) 4533d01ac3c3SMatthew Wilcox (Oracle) return ret; 453410102459SKirill A. Shutemov 4535953c66c2SAneesh Kumar K.V /* 4536eac96c3eSYang Shi * Just backoff if any subpage of a THP is corrupted otherwise 4537eac96c3eSYang Shi * the corrupted page may mapped by PMD silently to escape the 4538eac96c3eSYang Shi * check. This kind of THP just can be PTE mapped. Access to 4539eac96c3eSYang Shi * the corrupted subpage should trigger SIGBUS as expected. 4540eac96c3eSYang Shi */ 4541ef37b2eaSDavid Hildenbrand if (unlikely(folio_test_has_hwpoisoned(folio))) 4542eac96c3eSYang Shi return ret; 4543eac96c3eSYang Shi 4544eac96c3eSYang Shi /* 4545f0953a1bSIngo Molnar * Archs like ppc64 need additional space to store information 4546953c66c2SAneesh Kumar K.V * related to pte entry. Use the preallocated table for that. 4547953c66c2SAneesh Kumar K.V */ 454882b0f8c3SJan Kara if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) { 45494cf58924SJoel Fernandes (Google) vmf->prealloc_pte = pte_alloc_one(vma->vm_mm); 455082b0f8c3SJan Kara if (!vmf->prealloc_pte) 4551953c66c2SAneesh Kumar K.V return VM_FAULT_OOM; 4552953c66c2SAneesh Kumar K.V } 4553953c66c2SAneesh Kumar K.V 455482b0f8c3SJan Kara vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); 455582b0f8c3SJan Kara if (unlikely(!pmd_none(*vmf->pmd))) 455610102459SKirill A. Shutemov goto out; 455710102459SKirill A. Shutemov 45589f1f5b60SMatthew Wilcox (Oracle) flush_icache_pages(vma, page, HPAGE_PMD_NR); 455910102459SKirill A. Shutemov 456010102459SKirill A. Shutemov entry = mk_huge_pmd(page, vma->vm_page_prot); 456110102459SKirill A. Shutemov if (write) 4562f55e1014SLinus Torvalds entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 456310102459SKirill A. Shutemov 45646b27cc6cSKefeng Wang add_mm_counter(vma->vm_mm, mm_counter_file(folio), HPAGE_PMD_NR); 4565ef37b2eaSDavid Hildenbrand folio_add_file_rmap_pmd(folio, page, vma); 4566cea86fe2SHugh Dickins 4567953c66c2SAneesh Kumar K.V /* 4568953c66c2SAneesh Kumar K.V * deposit and withdraw with pmd lock held 4569953c66c2SAneesh Kumar K.V */ 4570953c66c2SAneesh Kumar K.V if (arch_needs_pgtable_deposit()) 457182b0f8c3SJan Kara deposit_prealloc_pte(vmf); 457210102459SKirill A. Shutemov 457382b0f8c3SJan Kara set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); 457410102459SKirill A. Shutemov 457582b0f8c3SJan Kara update_mmu_cache_pmd(vma, haddr, vmf->pmd); 457610102459SKirill A. Shutemov 457710102459SKirill A. Shutemov /* fault is handled */ 457810102459SKirill A. Shutemov ret = 0; 457995ecedcdSKirill A. Shutemov count_vm_event(THP_FILE_MAPPED); 458010102459SKirill A. Shutemov out: 458182b0f8c3SJan Kara spin_unlock(vmf->ptl); 458210102459SKirill A. Shutemov return ret; 458310102459SKirill A. Shutemov } 458410102459SKirill A. Shutemov #else 4585f9ce0be7SKirill A. Shutemov vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page) 458610102459SKirill A. Shutemov { 4587f9ce0be7SKirill A. Shutemov return VM_FAULT_FALLBACK; 458810102459SKirill A. Shutemov } 458910102459SKirill A. Shutemov #endif 459010102459SKirill A. Shutemov 45913bd786f7SYin Fengwei /** 45923bd786f7SYin Fengwei * set_pte_range - Set a range of PTEs to point to pages in a folio. 45933bd786f7SYin Fengwei * @vmf: Fault decription. 45943bd786f7SYin Fengwei * @folio: The folio that contains @page. 45953bd786f7SYin Fengwei * @page: The first page to create a PTE for. 45963bd786f7SYin Fengwei * @nr: The number of PTEs to create. 45973bd786f7SYin Fengwei * @addr: The first address to create a PTE for. 45983bd786f7SYin Fengwei */ 45993bd786f7SYin Fengwei void set_pte_range(struct vm_fault *vmf, struct folio *folio, 46003bd786f7SYin Fengwei struct page *page, unsigned int nr, unsigned long addr) 46013bb97794SKirill A. Shutemov { 460282b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 46032bad466cSPeter Xu bool uffd_wp = vmf_orig_pte_uffd_wp(vmf); 460482b0f8c3SJan Kara bool write = vmf->flags & FAULT_FLAG_WRITE; 46053bd786f7SYin Fengwei bool prefault = in_range(vmf->address, addr, nr * PAGE_SIZE); 46063bb97794SKirill A. Shutemov pte_t entry; 46077267ec00SKirill A. Shutemov 46083bd786f7SYin Fengwei flush_icache_pages(vma, page, nr); 46093bb97794SKirill A. Shutemov entry = mk_pte(page, vma->vm_page_prot); 461046bdb427SWill Deacon 461146bdb427SWill Deacon if (prefault && arch_wants_old_prefaulted_pte()) 461246bdb427SWill Deacon entry = pte_mkold(entry); 461350c25ee9SThomas Bogendoerfer else 461450c25ee9SThomas Bogendoerfer entry = pte_sw_mkyoung(entry); 461546bdb427SWill Deacon 46163bb97794SKirill A. Shutemov if (write) 46173bb97794SKirill A. Shutemov entry = maybe_mkwrite(pte_mkdirty(entry), vma); 46189c28a205SPeter Xu if (unlikely(uffd_wp)) 4619f1eb1bacSPeter Xu entry = pte_mkuffd_wp(entry); 4620bae473a4SKirill A. Shutemov /* copy-on-write page */ 4621bae473a4SKirill A. Shutemov if (write && !(vma->vm_flags & VM_SHARED)) { 46223bd786f7SYin Fengwei add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr); 46233bd786f7SYin Fengwei VM_BUG_ON_FOLIO(nr != 1, folio); 46243bd786f7SYin Fengwei folio_add_new_anon_rmap(folio, vma, addr); 46253bd786f7SYin Fengwei folio_add_lru_vma(folio, vma); 46263bb97794SKirill A. Shutemov } else { 46276b27cc6cSKefeng Wang add_mm_counter(vma->vm_mm, mm_counter_file(folio), nr); 462868f03208SDavid Hildenbrand folio_add_file_rmap_ptes(folio, page, nr, vma); 46293bb97794SKirill A. Shutemov } 46303bd786f7SYin Fengwei set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr); 46313bd786f7SYin Fengwei 46323bd786f7SYin Fengwei /* no need to invalidate: a not-present page won't be cached */ 46333bd786f7SYin Fengwei update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr); 46343bb97794SKirill A. Shutemov } 46353bb97794SKirill A. Shutemov 4636f46f2adeSPeter Xu static bool vmf_pte_changed(struct vm_fault *vmf) 4637f46f2adeSPeter Xu { 4638f46f2adeSPeter Xu if (vmf->flags & FAULT_FLAG_ORIG_PTE_VALID) 4639c33c7948SRyan Roberts return !pte_same(ptep_get(vmf->pte), vmf->orig_pte); 4640f46f2adeSPeter Xu 4641c33c7948SRyan Roberts return !pte_none(ptep_get(vmf->pte)); 4642f46f2adeSPeter Xu } 4643f46f2adeSPeter Xu 46449118c0cbSJan Kara /** 46459118c0cbSJan Kara * finish_fault - finish page fault once we have prepared the page to fault 46469118c0cbSJan Kara * 46479118c0cbSJan Kara * @vmf: structure describing the fault 46489118c0cbSJan Kara * 46499118c0cbSJan Kara * This function handles all that is needed to finish a page fault once the 46509118c0cbSJan Kara * page to fault in is prepared. It handles locking of PTEs, inserts PTE for 46519118c0cbSJan Kara * given page, adds reverse page mapping, handles memcg charges and LRU 4652a862f68aSMike Rapoport * addition. 46539118c0cbSJan Kara * 46549118c0cbSJan Kara * The function expects the page to be locked and on success it consumes a 46559118c0cbSJan Kara * reference of a page being mapped (for the PTE which maps it). 4656a862f68aSMike Rapoport * 4657a862f68aSMike Rapoport * Return: %0 on success, %VM_FAULT_ code in case of error. 46589118c0cbSJan Kara */ 46592b740303SSouptick Joarder vm_fault_t finish_fault(struct vm_fault *vmf) 46609118c0cbSJan Kara { 4661f9ce0be7SKirill A. Shutemov struct vm_area_struct *vma = vmf->vma; 46629118c0cbSJan Kara struct page *page; 4663f9ce0be7SKirill A. Shutemov vm_fault_t ret; 46649118c0cbSJan Kara 46659118c0cbSJan Kara /* Did we COW the page? */ 4666f9ce0be7SKirill A. Shutemov if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 46679118c0cbSJan Kara page = vmf->cow_page; 46689118c0cbSJan Kara else 46699118c0cbSJan Kara page = vmf->page; 46706b31d595SMichal Hocko 46716b31d595SMichal Hocko /* 46726b31d595SMichal Hocko * check even for read faults because we might have lost our CoWed 46736b31d595SMichal Hocko * page 46746b31d595SMichal Hocko */ 4675f9ce0be7SKirill A. Shutemov if (!(vma->vm_flags & VM_SHARED)) { 4676f9ce0be7SKirill A. Shutemov ret = check_stable_address_space(vma->vm_mm); 4677f9ce0be7SKirill A. Shutemov if (ret) 4678f9ce0be7SKirill A. Shutemov return ret; 4679f9ce0be7SKirill A. Shutemov } 4680f9ce0be7SKirill A. Shutemov 4681f9ce0be7SKirill A. Shutemov if (pmd_none(*vmf->pmd)) { 4682f9ce0be7SKirill A. Shutemov if (PageTransCompound(page)) { 4683f9ce0be7SKirill A. Shutemov ret = do_set_pmd(vmf, page); 4684f9ce0be7SKirill A. Shutemov if (ret != VM_FAULT_FALLBACK) 4685f9ce0be7SKirill A. Shutemov return ret; 4686f9ce0be7SKirill A. Shutemov } 4687f9ce0be7SKirill A. Shutemov 468803c4f204SQi Zheng if (vmf->prealloc_pte) 468903c4f204SQi Zheng pmd_install(vma->vm_mm, vmf->pmd, &vmf->prealloc_pte); 469003c4f204SQi Zheng else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) 4691f9ce0be7SKirill A. Shutemov return VM_FAULT_OOM; 4692f9ce0be7SKirill A. Shutemov } 4693f9ce0be7SKirill A. Shutemov 4694f9ce0be7SKirill A. Shutemov vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 4695f9ce0be7SKirill A. Shutemov vmf->address, &vmf->ptl); 46963db82b93SHugh Dickins if (!vmf->pte) 46973db82b93SHugh Dickins return VM_FAULT_NOPAGE; 4698f9ce0be7SKirill A. Shutemov 469970427f6eSSergei Antonov /* Re-check under ptl */ 470070427f6eSSergei Antonov if (likely(!vmf_pte_changed(vmf))) { 47013bd786f7SYin Fengwei struct folio *folio = page_folio(page); 470270427f6eSSergei Antonov 47033bd786f7SYin Fengwei set_pte_range(vmf, folio, page, 1, vmf->address); 470470427f6eSSergei Antonov ret = 0; 470570427f6eSSergei Antonov } else { 4706f9ce0be7SKirill A. Shutemov update_mmu_tlb(vma, vmf->address, vmf->pte); 470770427f6eSSergei Antonov ret = VM_FAULT_NOPAGE; 470870427f6eSSergei Antonov } 470970427f6eSSergei Antonov 47109118c0cbSJan Kara pte_unmap_unlock(vmf->pte, vmf->ptl); 47119118c0cbSJan Kara return ret; 47129118c0cbSJan Kara } 47139118c0cbSJan Kara 471453d36a56SLorenzo Stoakes static unsigned long fault_around_pages __read_mostly = 471553d36a56SLorenzo Stoakes 65536 >> PAGE_SHIFT; 4716a9b0f861SKirill A. Shutemov 47171592eef0SKirill A. Shutemov #ifdef CONFIG_DEBUG_FS 4718a9b0f861SKirill A. Shutemov static int fault_around_bytes_get(void *data, u64 *val) 47191592eef0SKirill A. Shutemov { 472053d36a56SLorenzo Stoakes *val = fault_around_pages << PAGE_SHIFT; 47211592eef0SKirill A. Shutemov return 0; 47221592eef0SKirill A. Shutemov } 47231592eef0SKirill A. Shutemov 4724b4903d6eSAndrey Ryabinin /* 4725da391d64SWilliam Kucharski * fault_around_bytes must be rounded down to the nearest page order as it's 4726da391d64SWilliam Kucharski * what do_fault_around() expects to see. 4727b4903d6eSAndrey Ryabinin */ 4728a9b0f861SKirill A. Shutemov static int fault_around_bytes_set(void *data, u64 val) 47291592eef0SKirill A. Shutemov { 4730a9b0f861SKirill A. Shutemov if (val / PAGE_SIZE > PTRS_PER_PTE) 47311592eef0SKirill A. Shutemov return -EINVAL; 473253d36a56SLorenzo Stoakes 473353d36a56SLorenzo Stoakes /* 473453d36a56SLorenzo Stoakes * The minimum value is 1 page, however this results in no fault-around 473553d36a56SLorenzo Stoakes * at all. See should_fault_around(). 473653d36a56SLorenzo Stoakes */ 473753d36a56SLorenzo Stoakes fault_around_pages = max(rounddown_pow_of_two(val) >> PAGE_SHIFT, 1UL); 473853d36a56SLorenzo Stoakes 47391592eef0SKirill A. Shutemov return 0; 47401592eef0SKirill A. Shutemov } 47410a1345f8SYevgen Pronenko DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops, 4742a9b0f861SKirill A. Shutemov fault_around_bytes_get, fault_around_bytes_set, "%llu\n"); 47431592eef0SKirill A. Shutemov 47441592eef0SKirill A. Shutemov static int __init fault_around_debugfs(void) 47451592eef0SKirill A. Shutemov { 4746d9f7979cSGreg Kroah-Hartman debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL, 4747a9b0f861SKirill A. Shutemov &fault_around_bytes_fops); 47481592eef0SKirill A. Shutemov return 0; 47491592eef0SKirill A. Shutemov } 47501592eef0SKirill A. Shutemov late_initcall(fault_around_debugfs); 47511592eef0SKirill A. Shutemov #endif 47528c6e50b0SKirill A. Shutemov 47531fdb412bSKirill A. Shutemov /* 47541fdb412bSKirill A. Shutemov * do_fault_around() tries to map few pages around the fault address. The hope 47551fdb412bSKirill A. Shutemov * is that the pages will be needed soon and this will lower the number of 47561fdb412bSKirill A. Shutemov * faults to handle. 47571fdb412bSKirill A. Shutemov * 47581fdb412bSKirill A. Shutemov * It uses vm_ops->map_pages() to map the pages, which skips the page if it's 47591fdb412bSKirill A. Shutemov * not ready to be mapped: not up-to-date, locked, etc. 47601fdb412bSKirill A. Shutemov * 47619042599eSLorenzo Stoakes * This function doesn't cross VMA or page table boundaries, in order to call 47629042599eSLorenzo Stoakes * map_pages() and acquire a PTE lock only once. 47631fdb412bSKirill A. Shutemov * 476453d36a56SLorenzo Stoakes * fault_around_pages defines how many pages we'll try to map. 4765da391d64SWilliam Kucharski * do_fault_around() expects it to be set to a power of two less than or equal 4766da391d64SWilliam Kucharski * to PTRS_PER_PTE. 47671fdb412bSKirill A. Shutemov * 4768da391d64SWilliam Kucharski * The virtual address of the area that we map is naturally aligned to 476953d36a56SLorenzo Stoakes * fault_around_pages * PAGE_SIZE rounded down to the machine page size 4770da391d64SWilliam Kucharski * (and therefore to page order). This way it's easier to guarantee 4771da391d64SWilliam Kucharski * that we don't cross page table boundaries. 47721fdb412bSKirill A. Shutemov */ 47732b740303SSouptick Joarder static vm_fault_t do_fault_around(struct vm_fault *vmf) 47748c6e50b0SKirill A. Shutemov { 477553d36a56SLorenzo Stoakes pgoff_t nr_pages = READ_ONCE(fault_around_pages); 47769042599eSLorenzo Stoakes pgoff_t pte_off = pte_index(vmf->address); 47779042599eSLorenzo Stoakes /* The page offset of vmf->address within the VMA. */ 47789042599eSLorenzo Stoakes pgoff_t vma_off = vmf->pgoff - vmf->vma->vm_pgoff; 47799042599eSLorenzo Stoakes pgoff_t from_pte, to_pte; 478058ef47efSMatthew Wilcox (Oracle) vm_fault_t ret; 47818c6e50b0SKirill A. Shutemov 47829042599eSLorenzo Stoakes /* The PTE offset of the start address, clamped to the VMA. */ 47839042599eSLorenzo Stoakes from_pte = max(ALIGN_DOWN(pte_off, nr_pages), 47849042599eSLorenzo Stoakes pte_off - min(pte_off, vma_off)); 4785aecd6f44SKirill A. Shutemov 47869042599eSLorenzo Stoakes /* The PTE offset of the end address, clamped to the VMA and PTE. */ 47879042599eSLorenzo Stoakes to_pte = min3(from_pte + nr_pages, (pgoff_t)PTRS_PER_PTE, 47889042599eSLorenzo Stoakes pte_off + vma_pages(vmf->vma) - vma_off) - 1; 47898c6e50b0SKirill A. Shutemov 479082b0f8c3SJan Kara if (pmd_none(*vmf->pmd)) { 47914cf58924SJoel Fernandes (Google) vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm); 479282b0f8c3SJan Kara if (!vmf->prealloc_pte) 4793f9ce0be7SKirill A. Shutemov return VM_FAULT_OOM; 47948c6e50b0SKirill A. Shutemov } 47958c6e50b0SKirill A. Shutemov 479658ef47efSMatthew Wilcox (Oracle) rcu_read_lock(); 479758ef47efSMatthew Wilcox (Oracle) ret = vmf->vma->vm_ops->map_pages(vmf, 47989042599eSLorenzo Stoakes vmf->pgoff + from_pte - pte_off, 47999042599eSLorenzo Stoakes vmf->pgoff + to_pte - pte_off); 480058ef47efSMatthew Wilcox (Oracle) rcu_read_unlock(); 480158ef47efSMatthew Wilcox (Oracle) 480258ef47efSMatthew Wilcox (Oracle) return ret; 48037267ec00SKirill A. Shutemov } 48047267ec00SKirill A. Shutemov 48059c28a205SPeter Xu /* Return true if we should do read fault-around, false otherwise */ 48069c28a205SPeter Xu static inline bool should_fault_around(struct vm_fault *vmf) 48079c28a205SPeter Xu { 48089c28a205SPeter Xu /* No ->map_pages? No way to fault around... */ 48099c28a205SPeter Xu if (!vmf->vma->vm_ops->map_pages) 48109c28a205SPeter Xu return false; 48119c28a205SPeter Xu 48129c28a205SPeter Xu if (uffd_disable_fault_around(vmf->vma)) 48139c28a205SPeter Xu return false; 48149c28a205SPeter Xu 481553d36a56SLorenzo Stoakes /* A single page implies no faulting 'around' at all. */ 481653d36a56SLorenzo Stoakes return fault_around_pages > 1; 48179c28a205SPeter Xu } 48189c28a205SPeter Xu 48192b740303SSouptick Joarder static vm_fault_t do_read_fault(struct vm_fault *vmf) 4820e655fb29SKirill A. Shutemov { 48212b740303SSouptick Joarder vm_fault_t ret = 0; 482222d1e68fSSidhartha Kumar struct folio *folio; 48238c6e50b0SKirill A. Shutemov 48248c6e50b0SKirill A. Shutemov /* 48258c6e50b0SKirill A. Shutemov * Let's call ->map_pages() first and use ->fault() as fallback 48268c6e50b0SKirill A. Shutemov * if page by the offset is not ready to be mapped (cold cache or 48278c6e50b0SKirill A. Shutemov * something). 48288c6e50b0SKirill A. Shutemov */ 48299c28a205SPeter Xu if (should_fault_around(vmf)) { 48300721ec8bSJan Kara ret = do_fault_around(vmf); 48317267ec00SKirill A. Shutemov if (ret) 48327267ec00SKirill A. Shutemov return ret; 48338c6e50b0SKirill A. Shutemov } 4834e655fb29SKirill A. Shutemov 483512214ebaSMatthew Wilcox (Oracle) ret = vmf_can_call_fault(vmf); 483612214ebaSMatthew Wilcox (Oracle) if (ret) 483712214ebaSMatthew Wilcox (Oracle) return ret; 4838f5617ffeSMatthew Wilcox (Oracle) 4839936ca80dSJan Kara ret = __do_fault(vmf); 4840e655fb29SKirill A. Shutemov if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4841e655fb29SKirill A. Shutemov return ret; 4842e655fb29SKirill A. Shutemov 48439118c0cbSJan Kara ret |= finish_fault(vmf); 484422d1e68fSSidhartha Kumar folio = page_folio(vmf->page); 484522d1e68fSSidhartha Kumar folio_unlock(folio); 48467267ec00SKirill A. Shutemov if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 484722d1e68fSSidhartha Kumar folio_put(folio); 4848e655fb29SKirill A. Shutemov return ret; 4849e655fb29SKirill A. Shutemov } 4850e655fb29SKirill A. Shutemov 48512b740303SSouptick Joarder static vm_fault_t do_cow_fault(struct vm_fault *vmf) 4852ec47c3b9SKirill A. Shutemov { 485382b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 4854e4621e70SKefeng Wang struct folio *folio; 48552b740303SSouptick Joarder vm_fault_t ret; 4856ec47c3b9SKirill A. Shutemov 48574de8c93aSMatthew Wilcox (Oracle) ret = vmf_can_call_fault(vmf); 48584de8c93aSMatthew Wilcox (Oracle) if (!ret) 48594de8c93aSMatthew Wilcox (Oracle) ret = vmf_anon_prepare(vmf); 48604de8c93aSMatthew Wilcox (Oracle) if (ret) 48614de8c93aSMatthew Wilcox (Oracle) return ret; 4862ec47c3b9SKirill A. Shutemov 4863e4621e70SKefeng Wang folio = folio_prealloc(vma->vm_mm, vma, vmf->address, false); 4864e4621e70SKefeng Wang if (!folio) 4865ec47c3b9SKirill A. Shutemov return VM_FAULT_OOM; 4866ec47c3b9SKirill A. Shutemov 4867e4621e70SKefeng Wang vmf->cow_page = &folio->page; 4868ec47c3b9SKirill A. Shutemov 4869936ca80dSJan Kara ret = __do_fault(vmf); 4870ec47c3b9SKirill A. Shutemov if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4871ec47c3b9SKirill A. Shutemov goto uncharge_out; 48723917048dSJan Kara if (ret & VM_FAULT_DONE_COW) 48733917048dSJan Kara return ret; 4874ec47c3b9SKirill A. Shutemov 4875936ca80dSJan Kara copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma); 4876e4621e70SKefeng Wang __folio_mark_uptodate(folio); 4877ec47c3b9SKirill A. Shutemov 48789118c0cbSJan Kara ret |= finish_fault(vmf); 4879936ca80dSJan Kara unlock_page(vmf->page); 4880936ca80dSJan Kara put_page(vmf->page); 48817267ec00SKirill A. Shutemov if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 48827267ec00SKirill A. Shutemov goto uncharge_out; 4883ec47c3b9SKirill A. Shutemov return ret; 4884ec47c3b9SKirill A. Shutemov uncharge_out: 4885e4621e70SKefeng Wang folio_put(folio); 4886ec47c3b9SKirill A. Shutemov return ret; 4887ec47c3b9SKirill A. Shutemov } 4888ec47c3b9SKirill A. Shutemov 48892b740303SSouptick Joarder static vm_fault_t do_shared_fault(struct vm_fault *vmf) 48901da177e4SLinus Torvalds { 489182b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 48922b740303SSouptick Joarder vm_fault_t ret, tmp; 48936f609b7eSSidhartha Kumar struct folio *folio; 48941d65f86dSKAMEZAWA Hiroyuki 48954ed43798SMatthew Wilcox (Oracle) ret = vmf_can_call_fault(vmf); 48964ed43798SMatthew Wilcox (Oracle) if (ret) 48974ed43798SMatthew Wilcox (Oracle) return ret; 489854cb8821SNick Piggin 4899936ca80dSJan Kara ret = __do_fault(vmf); 49007eae74afSKirill A. Shutemov if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) 4901f0c6d4d2SKirill A. Shutemov return ret; 49021da177e4SLinus Torvalds 49036f609b7eSSidhartha Kumar folio = page_folio(vmf->page); 49046f609b7eSSidhartha Kumar 49051da177e4SLinus Torvalds /* 4906f0c6d4d2SKirill A. Shutemov * Check if the backing address space wants to know that the page is 4907f0c6d4d2SKirill A. Shutemov * about to become writable 49081da177e4SLinus Torvalds */ 4909fb09a464SKirill A. Shutemov if (vma->vm_ops->page_mkwrite) { 49106f609b7eSSidhartha Kumar folio_unlock(folio); 491186aa6998SSidhartha Kumar tmp = do_page_mkwrite(vmf, folio); 4912fb09a464SKirill A. Shutemov if (unlikely(!tmp || 4913fb09a464SKirill A. Shutemov (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { 49146f609b7eSSidhartha Kumar folio_put(folio); 4915f0c6d4d2SKirill A. Shutemov return tmp; 491669676147SMark Fasheh } 4917d0217ac0SNick Piggin } 4918fb09a464SKirill A. Shutemov 49199118c0cbSJan Kara ret |= finish_fault(vmf); 49207267ec00SKirill A. Shutemov if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | 49217267ec00SKirill A. Shutemov VM_FAULT_RETRY))) { 49226f609b7eSSidhartha Kumar folio_unlock(folio); 49236f609b7eSSidhartha Kumar folio_put(folio); 4924f0c6d4d2SKirill A. Shutemov return ret; 49259637a5efSDavid Howells } 4926d00806b1SNick Piggin 492789b15332SJohannes Weiner ret |= fault_dirty_shared_page(vmf); 4928b827e496SNick Piggin return ret; 492954cb8821SNick Piggin } 4930d00806b1SNick Piggin 49319a95f3cfSPaul Cassella /* 4932c1e8d7c6SMichel Lespinasse * We enter with non-exclusive mmap_lock (to exclude vma changes, 49339a95f3cfSPaul Cassella * but allow concurrent faults). 4934c1e8d7c6SMichel Lespinasse * The mmap_lock may have been released depending on flags and our 49359138e47eSMatthew Wilcox (Oracle) * return value. See filemap_fault() and __folio_lock_or_retry(). 4936c1e8d7c6SMichel Lespinasse * If mmap_lock is released, vma may become invalid (for example 4937fc8efd2dSJan Stancek * by other thread calling munmap()). 49389a95f3cfSPaul Cassella */ 49392b740303SSouptick Joarder static vm_fault_t do_fault(struct vm_fault *vmf) 494054cb8821SNick Piggin { 494182b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 4942fc8efd2dSJan Stancek struct mm_struct *vm_mm = vma->vm_mm; 49432b740303SSouptick Joarder vm_fault_t ret; 494454cb8821SNick Piggin 4945ff09d7ecSAneesh Kumar K.V /* 4946ff09d7ecSAneesh Kumar K.V * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND 4947ff09d7ecSAneesh Kumar K.V */ 4948ff09d7ecSAneesh Kumar K.V if (!vma->vm_ops->fault) { 49493db82b93SHugh Dickins vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, 49503db82b93SHugh Dickins vmf->address, &vmf->ptl); 49513db82b93SHugh Dickins if (unlikely(!vmf->pte)) 4952b0b9b3dfSHugh Dickins ret = VM_FAULT_SIGBUS; 4953ff09d7ecSAneesh Kumar K.V else { 4954ff09d7ecSAneesh Kumar K.V /* 4955ff09d7ecSAneesh Kumar K.V * Make sure this is not a temporary clearing of pte 4956ff09d7ecSAneesh Kumar K.V * by holding ptl and checking again. A R/M/W update 4957ff09d7ecSAneesh Kumar K.V * of pte involves: take ptl, clearing the pte so that 4958ff09d7ecSAneesh Kumar K.V * we don't have concurrent modification by hardware 4959ff09d7ecSAneesh Kumar K.V * followed by an update. 4960ff09d7ecSAneesh Kumar K.V */ 4961c33c7948SRyan Roberts if (unlikely(pte_none(ptep_get(vmf->pte)))) 4962ff09d7ecSAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 4963ff09d7ecSAneesh Kumar K.V else 4964ff09d7ecSAneesh Kumar K.V ret = VM_FAULT_NOPAGE; 4965ff09d7ecSAneesh Kumar K.V 4966ff09d7ecSAneesh Kumar K.V pte_unmap_unlock(vmf->pte, vmf->ptl); 4967ff09d7ecSAneesh Kumar K.V } 4968ff09d7ecSAneesh Kumar K.V } else if (!(vmf->flags & FAULT_FLAG_WRITE)) 4969b0b9b3dfSHugh Dickins ret = do_read_fault(vmf); 4970b0b9b3dfSHugh Dickins else if (!(vma->vm_flags & VM_SHARED)) 4971b0b9b3dfSHugh Dickins ret = do_cow_fault(vmf); 4972b0b9b3dfSHugh Dickins else 4973b0b9b3dfSHugh Dickins ret = do_shared_fault(vmf); 4974b0b9b3dfSHugh Dickins 4975b0b9b3dfSHugh Dickins /* preallocated pagetable is unused: free it */ 4976b0b9b3dfSHugh Dickins if (vmf->prealloc_pte) { 4977fc8efd2dSJan Stancek pte_free(vm_mm, vmf->prealloc_pte); 49787f2b6ce8STobin C Harding vmf->prealloc_pte = NULL; 4979b0b9b3dfSHugh Dickins } 4980b0b9b3dfSHugh Dickins return ret; 498154cb8821SNick Piggin } 498254cb8821SNick Piggin 4983cda6d936SKefeng Wang int numa_migrate_prep(struct folio *folio, struct vm_area_struct *vma, 4984f4c0d836SYang Shi unsigned long addr, int page_nid, int *flags) 49859532fec1SMel Gorman { 4986cda6d936SKefeng Wang folio_get(folio); 49879532fec1SMel Gorman 4988fc137c0dSRaghavendra K T /* Record the current PID acceesing VMA */ 4989fc137c0dSRaghavendra K T vma_set_access_pid_bit(vma); 4990fc137c0dSRaghavendra K T 49919532fec1SMel Gorman count_vm_numa_event(NUMA_HINT_FAULTS); 499204bb2f94SRik van Riel if (page_nid == numa_node_id()) { 49939532fec1SMel Gorman count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); 499404bb2f94SRik van Riel *flags |= TNF_FAULT_LOCAL; 499504bb2f94SRik van Riel } 49969532fec1SMel Gorman 499775c70128SKefeng Wang return mpol_misplaced(folio, vma, addr); 49989532fec1SMel Gorman } 49999532fec1SMel Gorman 50002b740303SSouptick Joarder static vm_fault_t do_numa_page(struct vm_fault *vmf) 5001d10e63f2SMel Gorman { 500282b0f8c3SJan Kara struct vm_area_struct *vma = vmf->vma; 50036695cf68SKefeng Wang struct folio *folio = NULL; 50046695cf68SKefeng Wang int nid = NUMA_NO_NODE; 50056a56ccbcSDavid Hildenbrand bool writable = false; 500690572890SPeter Zijlstra int last_cpupid; 5007cbee9f88SPeter Zijlstra int target_nid; 500804a86453SAneesh Kumar K.V pte_t pte, old_pte; 50096688cc05SPeter Zijlstra int flags = 0; 5010d10e63f2SMel Gorman 5011d10e63f2SMel Gorman /* 5012d10e63f2SMel Gorman * The "pte" at this point cannot be used safely without 5013d10e63f2SMel Gorman * validation through pte_unmap_same(). It's of NUMA type but 5014d10e63f2SMel Gorman * the pfn may be screwed if the read is non atomic. 5015d10e63f2SMel Gorman */ 501682b0f8c3SJan Kara spin_lock(vmf->ptl); 5017c33c7948SRyan Roberts if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) { 501882b0f8c3SJan Kara pte_unmap_unlock(vmf->pte, vmf->ptl); 50194daae3b4SMel Gorman goto out; 50204daae3b4SMel Gorman } 50214daae3b4SMel Gorman 5022b99a342dSHuang Ying /* Get the normal PTE */ 5023b99a342dSHuang Ying old_pte = ptep_get(vmf->pte); 502404a86453SAneesh Kumar K.V pte = pte_modify(old_pte, vma->vm_page_prot); 5025d10e63f2SMel Gorman 50266a56ccbcSDavid Hildenbrand /* 50276a56ccbcSDavid Hildenbrand * Detect now whether the PTE could be writable; this information 50286a56ccbcSDavid Hildenbrand * is only valid while holding the PT lock. 50296a56ccbcSDavid Hildenbrand */ 50306a56ccbcSDavid Hildenbrand writable = pte_write(pte); 50316a56ccbcSDavid Hildenbrand if (!writable && vma_wants_manual_pte_write_upgrade(vma) && 50326a56ccbcSDavid Hildenbrand can_change_pte_writable(vma, vmf->address, pte)) 50336a56ccbcSDavid Hildenbrand writable = true; 50346a56ccbcSDavid Hildenbrand 50356695cf68SKefeng Wang folio = vm_normal_folio(vma, vmf->address, pte); 50366695cf68SKefeng Wang if (!folio || folio_is_zone_device(folio)) 5037b99a342dSHuang Ying goto out_map; 5038d10e63f2SMel Gorman 5039e81c4802SKirill A. Shutemov /* TODO: handle PTE-mapped THP */ 50406695cf68SKefeng Wang if (folio_test_large(folio)) 5041b99a342dSHuang Ying goto out_map; 5042e81c4802SKirill A. Shutemov 50436688cc05SPeter Zijlstra /* 5044bea66fbdSMel Gorman * Avoid grouping on RO pages in general. RO pages shouldn't hurt as 5045bea66fbdSMel Gorman * much anyway since they can be in shared cache state. This misses 5046bea66fbdSMel Gorman * the case where a mapping is writable but the process never writes 5047bea66fbdSMel Gorman * to it but pte_write gets cleared during protection updates and 5048bea66fbdSMel Gorman * pte_dirty has unpredictable behaviour between PTE scan updates, 5049bea66fbdSMel Gorman * background writeback, dirty balancing and application behaviour. 50506688cc05SPeter Zijlstra */ 50516a56ccbcSDavid Hildenbrand if (!writable) 50526688cc05SPeter Zijlstra flags |= TNF_NO_GROUP; 50536688cc05SPeter Zijlstra 5054dabe1d99SRik van Riel /* 50556695cf68SKefeng Wang * Flag if the folio is shared between multiple address spaces. This 5056dabe1d99SRik van Riel * is later used when determining whether to group tasks together 5057dabe1d99SRik van Riel */ 50586695cf68SKefeng Wang if (folio_estimated_sharers(folio) > 1 && (vma->vm_flags & VM_SHARED)) 5059dabe1d99SRik van Riel flags |= TNF_SHARED; 5060dabe1d99SRik van Riel 50616695cf68SKefeng Wang nid = folio_nid(folio); 506233024536SHuang Ying /* 506333024536SHuang Ying * For memory tiering mode, cpupid of slow memory page is used 506433024536SHuang Ying * to record page access time. So use default value. 506533024536SHuang Ying */ 506633024536SHuang Ying if ((sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) && 50676695cf68SKefeng Wang !node_is_toptier(nid)) 506833024536SHuang Ying last_cpupid = (-1 & LAST_CPUPID_MASK); 506933024536SHuang Ying else 507067b33e3fSKefeng Wang last_cpupid = folio_last_cpupid(folio); 5071cda6d936SKefeng Wang target_nid = numa_migrate_prep(folio, vma, vmf->address, nid, &flags); 507298fa15f3SAnshuman Khandual if (target_nid == NUMA_NO_NODE) { 50736695cf68SKefeng Wang folio_put(folio); 5074b99a342dSHuang Ying goto out_map; 50754daae3b4SMel Gorman } 5076b99a342dSHuang Ying pte_unmap_unlock(vmf->pte, vmf->ptl); 50776a56ccbcSDavid Hildenbrand writable = false; 50784daae3b4SMel Gorman 50794daae3b4SMel Gorman /* Migrate to the requested node */ 50806695cf68SKefeng Wang if (migrate_misplaced_folio(folio, vma, target_nid)) { 50816695cf68SKefeng Wang nid = target_nid; 50826688cc05SPeter Zijlstra flags |= TNF_MIGRATED; 5083b99a342dSHuang Ying } else { 5084074c2381SMel Gorman flags |= TNF_MIGRATE_FAIL; 5085c7ad0880SHugh Dickins vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, 5086c7ad0880SHugh Dickins vmf->address, &vmf->ptl); 5087c7ad0880SHugh Dickins if (unlikely(!vmf->pte)) 5088c7ad0880SHugh Dickins goto out; 5089c33c7948SRyan Roberts if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) { 5090b99a342dSHuang Ying pte_unmap_unlock(vmf->pte, vmf->ptl); 5091b99a342dSHuang Ying goto out; 5092b99a342dSHuang Ying } 5093b99a342dSHuang Ying goto out_map; 5094b99a342dSHuang Ying } 50954daae3b4SMel Gorman 50964daae3b4SMel Gorman out: 50976695cf68SKefeng Wang if (nid != NUMA_NO_NODE) 50986695cf68SKefeng Wang task_numa_fault(last_cpupid, nid, 1, flags); 5099d10e63f2SMel Gorman return 0; 5100b99a342dSHuang Ying out_map: 5101b99a342dSHuang Ying /* 5102b99a342dSHuang Ying * Make it present again, depending on how arch implements 5103b99a342dSHuang Ying * non-accessible ptes, some can allow access by kernel mode. 5104b99a342dSHuang Ying */ 5105b99a342dSHuang Ying old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte); 5106b99a342dSHuang Ying pte = pte_modify(old_pte, vma->vm_page_prot); 5107b99a342dSHuang Ying pte = pte_mkyoung(pte); 51086a56ccbcSDavid Hildenbrand if (writable) 5109161e393cSRick Edgecombe pte = pte_mkwrite(pte, vma); 5110b99a342dSHuang Ying ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte); 51115003a2bdSMatthew Wilcox (Oracle) update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1); 5112b99a342dSHuang Ying pte_unmap_unlock(vmf->pte, vmf->ptl); 5113b99a342dSHuang Ying goto out; 5114d10e63f2SMel Gorman } 5115d10e63f2SMel Gorman 51162b740303SSouptick Joarder static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf) 5117b96375f7SMatthew Wilcox { 51188f5fd0e1SMatthew Wilcox (Oracle) struct vm_area_struct *vma = vmf->vma; 51198f5fd0e1SMatthew Wilcox (Oracle) if (vma_is_anonymous(vma)) 512082b0f8c3SJan Kara return do_huge_pmd_anonymous_page(vmf); 512140d49a3cSMatthew Wilcox (Oracle) if (vma->vm_ops->huge_fault) 51221d024e7aSMatthew Wilcox (Oracle) return vma->vm_ops->huge_fault(vmf, PMD_ORDER); 5123b96375f7SMatthew Wilcox return VM_FAULT_FALLBACK; 5124b96375f7SMatthew Wilcox } 5125b96375f7SMatthew Wilcox 5126183f24aaSGeert Uytterhoeven /* `inline' is required to avoid gcc 4.1.2 build error */ 51275db4f15cSYang Shi static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf) 5128b96375f7SMatthew Wilcox { 51298f5fd0e1SMatthew Wilcox (Oracle) struct vm_area_struct *vma = vmf->vma; 5130c89357e2SDavid Hildenbrand const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; 5131aea06577SDavid Hildenbrand vm_fault_t ret; 5132c89357e2SDavid Hildenbrand 51338f5fd0e1SMatthew Wilcox (Oracle) if (vma_is_anonymous(vma)) { 5134c89357e2SDavid Hildenbrand if (likely(!unshare) && 5135d61ea1cbSPeter Xu userfaultfd_huge_pmd_wp(vma, vmf->orig_pmd)) { 5136d61ea1cbSPeter Xu if (userfaultfd_wp_async(vmf->vma)) 5137d61ea1cbSPeter Xu goto split; 5138529b930bSAndrea Arcangeli return handle_userfault(vmf, VM_UFFD_WP); 5139d61ea1cbSPeter Xu } 51405db4f15cSYang Shi return do_huge_pmd_wp_page(vmf); 5141529b930bSAndrea Arcangeli } 5142af9e4d5fSKirill A. Shutemov 51438f5fd0e1SMatthew Wilcox (Oracle) if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) { 51448f5fd0e1SMatthew Wilcox (Oracle) if (vma->vm_ops->huge_fault) { 51451d024e7aSMatthew Wilcox (Oracle) ret = vma->vm_ops->huge_fault(vmf, PMD_ORDER); 5146327e9fd4SThomas Hellstrom (VMware) if (!(ret & VM_FAULT_FALLBACK)) 5147327e9fd4SThomas Hellstrom (VMware) return ret; 5148327e9fd4SThomas Hellstrom (VMware) } 5149aea06577SDavid Hildenbrand } 5150327e9fd4SThomas Hellstrom (VMware) 5151d61ea1cbSPeter Xu split: 5152327e9fd4SThomas Hellstrom (VMware) /* COW or write-notify handled on pte level: split pmd. */ 51538f5fd0e1SMatthew Wilcox (Oracle) __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL); 5154af9e4d5fSKirill A. Shutemov 5155b96375f7SMatthew Wilcox return VM_FAULT_FALLBACK; 5156b96375f7SMatthew Wilcox } 5157b96375f7SMatthew Wilcox 51582b740303SSouptick Joarder static vm_fault_t create_huge_pud(struct vm_fault *vmf) 5159a00cc7d9SMatthew Wilcox { 5160327e9fd4SThomas Hellstrom (VMware) #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \ 5161327e9fd4SThomas Hellstrom (VMware) defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) 5162c4fd825eSMatthew Wilcox (Oracle) struct vm_area_struct *vma = vmf->vma; 5163a00cc7d9SMatthew Wilcox /* No support for anonymous transparent PUD pages yet */ 5164c4fd825eSMatthew Wilcox (Oracle) if (vma_is_anonymous(vma)) 516514c99d65SGowans, James return VM_FAULT_FALLBACK; 516640d49a3cSMatthew Wilcox (Oracle) if (vma->vm_ops->huge_fault) 51671d024e7aSMatthew Wilcox (Oracle) return vma->vm_ops->huge_fault(vmf, PUD_ORDER); 516814c99d65SGowans, James #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 516914c99d65SGowans, James return VM_FAULT_FALLBACK; 517014c99d65SGowans, James } 517114c99d65SGowans, James 517214c99d65SGowans, James static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud) 517314c99d65SGowans, James { 517414c99d65SGowans, James #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \ 517514c99d65SGowans, James defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) 5176c4fd825eSMatthew Wilcox (Oracle) struct vm_area_struct *vma = vmf->vma; 5177aea06577SDavid Hildenbrand vm_fault_t ret; 5178aea06577SDavid Hildenbrand 517914c99d65SGowans, James /* No support for anonymous transparent PUD pages yet */ 5180c4fd825eSMatthew Wilcox (Oracle) if (vma_is_anonymous(vma)) 5181327e9fd4SThomas Hellstrom (VMware) goto split; 5182c4fd825eSMatthew Wilcox (Oracle) if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) { 5183c4fd825eSMatthew Wilcox (Oracle) if (vma->vm_ops->huge_fault) { 51841d024e7aSMatthew Wilcox (Oracle) ret = vma->vm_ops->huge_fault(vmf, PUD_ORDER); 5185327e9fd4SThomas Hellstrom (VMware) if (!(ret & VM_FAULT_FALLBACK)) 5186327e9fd4SThomas Hellstrom (VMware) return ret; 5187327e9fd4SThomas Hellstrom (VMware) } 5188aea06577SDavid Hildenbrand } 5189327e9fd4SThomas Hellstrom (VMware) split: 5190327e9fd4SThomas Hellstrom (VMware) /* COW or write-notify not handled on PUD level: split pud.*/ 5191c4fd825eSMatthew Wilcox (Oracle) __split_huge_pud(vma, vmf->pud, vmf->address); 519214c99d65SGowans, James #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ 5193a00cc7d9SMatthew Wilcox return VM_FAULT_FALLBACK; 5194a00cc7d9SMatthew Wilcox } 5195a00cc7d9SMatthew Wilcox 51961da177e4SLinus Torvalds /* 51971da177e4SLinus Torvalds * These routines also need to handle stuff like marking pages dirty 51981da177e4SLinus Torvalds * and/or accessed for architectures that don't do it in hardware (most 51991da177e4SLinus Torvalds * RISC architectures). The early dirtying is also good on the i386. 52001da177e4SLinus Torvalds * 52011da177e4SLinus Torvalds * There is also a hook called "update_mmu_cache()" that architectures 52021da177e4SLinus Torvalds * with external mmu caches can use to update those (ie the Sparc or 52031da177e4SLinus Torvalds * PowerPC hashed page tables that act as extended TLBs). 52041da177e4SLinus Torvalds * 5205c1e8d7c6SMichel Lespinasse * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow 52067267ec00SKirill A. Shutemov * concurrent faults). 52079a95f3cfSPaul Cassella * 5208c1e8d7c6SMichel Lespinasse * The mmap_lock may have been released depending on flags and our return value. 52099138e47eSMatthew Wilcox (Oracle) * See filemap_fault() and __folio_lock_or_retry(). 52101da177e4SLinus Torvalds */ 52112b740303SSouptick Joarder static vm_fault_t handle_pte_fault(struct vm_fault *vmf) 52121da177e4SLinus Torvalds { 52131da177e4SLinus Torvalds pte_t entry; 52141da177e4SLinus Torvalds 521582b0f8c3SJan Kara if (unlikely(pmd_none(*vmf->pmd))) { 52167267ec00SKirill A. Shutemov /* 52177267ec00SKirill A. Shutemov * Leave __pte_alloc() until later: because vm_ops->fault may 52187267ec00SKirill A. Shutemov * want to allocate huge page, and if we expose page table 52197267ec00SKirill A. Shutemov * for an instant, it will be difficult to retract from 52207267ec00SKirill A. Shutemov * concurrent faults and from rmap lookups. 52217267ec00SKirill A. Shutemov */ 522282b0f8c3SJan Kara vmf->pte = NULL; 5223f46f2adeSPeter Xu vmf->flags &= ~FAULT_FLAG_ORIG_PTE_VALID; 52247267ec00SKirill A. Shutemov } else { 5225f9ce0be7SKirill A. Shutemov /* 52267267ec00SKirill A. Shutemov * A regular pmd is established and it can't morph into a huge 5227c7ad0880SHugh Dickins * pmd by anon khugepaged, since that takes mmap_lock in write 5228c7ad0880SHugh Dickins * mode; but shmem or file collapse to THP could still morph 5229c7ad0880SHugh Dickins * it into a huge pmd: just retry later if so. 52307267ec00SKirill A. Shutemov */ 5231c7ad0880SHugh Dickins vmf->pte = pte_offset_map_nolock(vmf->vma->vm_mm, vmf->pmd, 5232c7ad0880SHugh Dickins vmf->address, &vmf->ptl); 5233c7ad0880SHugh Dickins if (unlikely(!vmf->pte)) 5234c7ad0880SHugh Dickins return 0; 523526e1a0c3SHugh Dickins vmf->orig_pte = ptep_get_lockless(vmf->pte); 5236f46f2adeSPeter Xu vmf->flags |= FAULT_FLAG_ORIG_PTE_VALID; 52377267ec00SKirill A. Shutemov 52382994302bSJan Kara if (pte_none(vmf->orig_pte)) { 523982b0f8c3SJan Kara pte_unmap(vmf->pte); 524082b0f8c3SJan Kara vmf->pte = NULL; 52417267ec00SKirill A. Shutemov } 52427267ec00SKirill A. Shutemov } 52437267ec00SKirill A. Shutemov 52442bad466cSPeter Xu if (!vmf->pte) 52452bad466cSPeter Xu return do_pte_missing(vmf); 52467267ec00SKirill A. Shutemov 52472994302bSJan Kara if (!pte_present(vmf->orig_pte)) 52482994302bSJan Kara return do_swap_page(vmf); 52491da177e4SLinus Torvalds 52502994302bSJan Kara if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma)) 52512994302bSJan Kara return do_numa_page(vmf); 5252d10e63f2SMel Gorman 525382b0f8c3SJan Kara spin_lock(vmf->ptl); 52542994302bSJan Kara entry = vmf->orig_pte; 5255c33c7948SRyan Roberts if (unlikely(!pte_same(ptep_get(vmf->pte), entry))) { 52567df67697SBibo Mao update_mmu_tlb(vmf->vma, vmf->address, vmf->pte); 52578f4e2101SHugh Dickins goto unlock; 52587df67697SBibo Mao } 5259c89357e2SDavid Hildenbrand if (vmf->flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 5260f6f37321SLinus Torvalds if (!pte_write(entry)) 52612994302bSJan Kara return do_wp_page(vmf); 5262c89357e2SDavid Hildenbrand else if (likely(vmf->flags & FAULT_FLAG_WRITE)) 52631da177e4SLinus Torvalds entry = pte_mkdirty(entry); 52641da177e4SLinus Torvalds } 52651da177e4SLinus Torvalds entry = pte_mkyoung(entry); 526682b0f8c3SJan Kara if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry, 526782b0f8c3SJan Kara vmf->flags & FAULT_FLAG_WRITE)) { 52685003a2bdSMatthew Wilcox (Oracle) update_mmu_cache_range(vmf, vmf->vma, vmf->address, 52695003a2bdSMatthew Wilcox (Oracle) vmf->pte, 1); 52701a44e149SAndrea Arcangeli } else { 5271b7333b58SYang Shi /* Skip spurious TLB flush for retried page fault */ 5272b7333b58SYang Shi if (vmf->flags & FAULT_FLAG_TRIED) 5273b7333b58SYang Shi goto unlock; 52741a44e149SAndrea Arcangeli /* 52751a44e149SAndrea Arcangeli * This is needed only for protection faults but the arch code 52761a44e149SAndrea Arcangeli * is not yet telling us if this is a protection fault or not. 52771a44e149SAndrea Arcangeli * This still avoids useless tlb flushes for .text page faults 52781a44e149SAndrea Arcangeli * with threads. 52791a44e149SAndrea Arcangeli */ 528082b0f8c3SJan Kara if (vmf->flags & FAULT_FLAG_WRITE) 528199c29133SGerald Schaefer flush_tlb_fix_spurious_fault(vmf->vma, vmf->address, 528299c29133SGerald Schaefer vmf->pte); 52831a44e149SAndrea Arcangeli } 52848f4e2101SHugh Dickins unlock: 528582b0f8c3SJan Kara pte_unmap_unlock(vmf->pte, vmf->ptl); 528683c54070SNick Piggin return 0; 52871da177e4SLinus Torvalds } 52881da177e4SLinus Torvalds 52891da177e4SLinus Torvalds /* 52904ec31152SMatthew Wilcox (Oracle) * On entry, we hold either the VMA lock or the mmap_lock 52914ec31152SMatthew Wilcox (Oracle) * (FAULT_FLAG_VMA_LOCK tells you which). If VM_FAULT_RETRY is set in 52924ec31152SMatthew Wilcox (Oracle) * the result, the mmap_lock is not held on exit. See filemap_fault() 52934ec31152SMatthew Wilcox (Oracle) * and __folio_lock_or_retry(). 52941da177e4SLinus Torvalds */ 52952b740303SSouptick Joarder static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma, 52962b740303SSouptick Joarder unsigned long address, unsigned int flags) 52971da177e4SLinus Torvalds { 529882b0f8c3SJan Kara struct vm_fault vmf = { 5299bae473a4SKirill A. Shutemov .vma = vma, 53001a29d85eSJan Kara .address = address & PAGE_MASK, 5301824ddc60SNadav Amit .real_address = address, 5302bae473a4SKirill A. Shutemov .flags = flags, 53030721ec8bSJan Kara .pgoff = linear_page_index(vma, address), 5304667240e0SJan Kara .gfp_mask = __get_fault_gfp_mask(vma), 5305bae473a4SKirill A. Shutemov }; 5306dcddffd4SKirill A. Shutemov struct mm_struct *mm = vma->vm_mm; 53077da4e2cbSYang Shi unsigned long vm_flags = vma->vm_flags; 53081da177e4SLinus Torvalds pgd_t *pgd; 5309c2febafcSKirill A. Shutemov p4d_t *p4d; 53102b740303SSouptick Joarder vm_fault_t ret; 53111da177e4SLinus Torvalds 53121da177e4SLinus Torvalds pgd = pgd_offset(mm, address); 5313c2febafcSKirill A. Shutemov p4d = p4d_alloc(mm, pgd, address); 5314c2febafcSKirill A. Shutemov if (!p4d) 5315c2febafcSKirill A. Shutemov return VM_FAULT_OOM; 5316a00cc7d9SMatthew Wilcox 5317c2febafcSKirill A. Shutemov vmf.pud = pud_alloc(mm, p4d, address); 5318a00cc7d9SMatthew Wilcox if (!vmf.pud) 5319c74df32cSHugh Dickins return VM_FAULT_OOM; 5320625110b5SThomas Hellstrom retry_pud: 53217da4e2cbSYang Shi if (pud_none(*vmf.pud) && 53223485b883SRyan Roberts thp_vma_allowable_order(vma, vm_flags, false, true, true, PUD_ORDER)) { 5323a00cc7d9SMatthew Wilcox ret = create_huge_pud(&vmf); 5324a00cc7d9SMatthew Wilcox if (!(ret & VM_FAULT_FALLBACK)) 5325a00cc7d9SMatthew Wilcox return ret; 5326a00cc7d9SMatthew Wilcox } else { 5327a00cc7d9SMatthew Wilcox pud_t orig_pud = *vmf.pud; 5328a00cc7d9SMatthew Wilcox 5329a00cc7d9SMatthew Wilcox barrier(); 5330a00cc7d9SMatthew Wilcox if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) { 5331a00cc7d9SMatthew Wilcox 5332c89357e2SDavid Hildenbrand /* 5333c89357e2SDavid Hildenbrand * TODO once we support anonymous PUDs: NUMA case and 5334c89357e2SDavid Hildenbrand * FAULT_FLAG_UNSHARE handling. 5335c89357e2SDavid Hildenbrand */ 5336c89357e2SDavid Hildenbrand if ((flags & FAULT_FLAG_WRITE) && !pud_write(orig_pud)) { 5337a00cc7d9SMatthew Wilcox ret = wp_huge_pud(&vmf, orig_pud); 5338a00cc7d9SMatthew Wilcox if (!(ret & VM_FAULT_FALLBACK)) 5339a00cc7d9SMatthew Wilcox return ret; 5340a00cc7d9SMatthew Wilcox } else { 5341a00cc7d9SMatthew Wilcox huge_pud_set_accessed(&vmf, orig_pud); 5342a00cc7d9SMatthew Wilcox return 0; 5343a00cc7d9SMatthew Wilcox } 5344a00cc7d9SMatthew Wilcox } 5345a00cc7d9SMatthew Wilcox } 5346a00cc7d9SMatthew Wilcox 5347a00cc7d9SMatthew Wilcox vmf.pmd = pmd_alloc(mm, vmf.pud, address); 534882b0f8c3SJan Kara if (!vmf.pmd) 5349c74df32cSHugh Dickins return VM_FAULT_OOM; 5350625110b5SThomas Hellstrom 5351625110b5SThomas Hellstrom /* Huge pud page fault raced with pmd_alloc? */ 5352625110b5SThomas Hellstrom if (pud_trans_unstable(vmf.pud)) 5353625110b5SThomas Hellstrom goto retry_pud; 5354625110b5SThomas Hellstrom 53557da4e2cbSYang Shi if (pmd_none(*vmf.pmd) && 53563485b883SRyan Roberts thp_vma_allowable_order(vma, vm_flags, false, true, true, PMD_ORDER)) { 5357a2d58167SDave Jiang ret = create_huge_pmd(&vmf); 5358c0292554SKirill A. Shutemov if (!(ret & VM_FAULT_FALLBACK)) 5359c0292554SKirill A. Shutemov return ret; 536071e3aac0SAndrea Arcangeli } else { 536126e1a0c3SHugh Dickins vmf.orig_pmd = pmdp_get_lockless(vmf.pmd); 53621f1d06c3SDavid Rientjes 53635db4f15cSYang Shi if (unlikely(is_swap_pmd(vmf.orig_pmd))) { 536484c3fc4eSZi Yan VM_BUG_ON(thp_migration_supported() && 53655db4f15cSYang Shi !is_pmd_migration_entry(vmf.orig_pmd)); 53665db4f15cSYang Shi if (is_pmd_migration_entry(vmf.orig_pmd)) 536784c3fc4eSZi Yan pmd_migration_entry_wait(mm, vmf.pmd); 536884c3fc4eSZi Yan return 0; 536984c3fc4eSZi Yan } 53705db4f15cSYang Shi if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) { 53715db4f15cSYang Shi if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma)) 53725db4f15cSYang Shi return do_huge_pmd_numa_page(&vmf); 5373d10e63f2SMel Gorman 5374c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 5375c89357e2SDavid Hildenbrand !pmd_write(vmf.orig_pmd)) { 53765db4f15cSYang Shi ret = wp_huge_pmd(&vmf); 53779845cbbdSKirill A. Shutemov if (!(ret & VM_FAULT_FALLBACK)) 53781f1d06c3SDavid Rientjes return ret; 5379a1dd450bSWill Deacon } else { 53805db4f15cSYang Shi huge_pmd_set_accessed(&vmf); 538171e3aac0SAndrea Arcangeli return 0; 538271e3aac0SAndrea Arcangeli } 538371e3aac0SAndrea Arcangeli } 53849845cbbdSKirill A. Shutemov } 538571e3aac0SAndrea Arcangeli 538682b0f8c3SJan Kara return handle_pte_fault(&vmf); 53871da177e4SLinus Torvalds } 53881da177e4SLinus Torvalds 5389bce617edSPeter Xu /** 5390f0953a1bSIngo Molnar * mm_account_fault - Do page fault accounting 5391809ef83cSYang Li * @mm: mm from which memcg should be extracted. It can be NULL. 5392bce617edSPeter Xu * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting 5393bce617edSPeter Xu * of perf event counters, but we'll still do the per-task accounting to 5394bce617edSPeter Xu * the task who triggered this page fault. 5395bce617edSPeter Xu * @address: the faulted address. 5396bce617edSPeter Xu * @flags: the fault flags. 5397bce617edSPeter Xu * @ret: the fault retcode. 5398bce617edSPeter Xu * 5399f0953a1bSIngo Molnar * This will take care of most of the page fault accounting. Meanwhile, it 5400bce617edSPeter Xu * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter 5401f0953a1bSIngo Molnar * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should 5402bce617edSPeter Xu * still be in per-arch page fault handlers at the entry of page fault. 5403bce617edSPeter Xu */ 540453156443SSuren Baghdasaryan static inline void mm_account_fault(struct mm_struct *mm, struct pt_regs *regs, 5405bce617edSPeter Xu unsigned long address, unsigned int flags, 5406bce617edSPeter Xu vm_fault_t ret) 5407bce617edSPeter Xu { 5408bce617edSPeter Xu bool major; 5409bce617edSPeter Xu 541053156443SSuren Baghdasaryan /* Incomplete faults will be accounted upon completion. */ 541153156443SSuren Baghdasaryan if (ret & VM_FAULT_RETRY) 541253156443SSuren Baghdasaryan return; 541353156443SSuren Baghdasaryan 5414bce617edSPeter Xu /* 541553156443SSuren Baghdasaryan * To preserve the behavior of older kernels, PGFAULT counters record 541653156443SSuren Baghdasaryan * both successful and failed faults, as opposed to perf counters, 541753156443SSuren Baghdasaryan * which ignore failed cases. 5418bce617edSPeter Xu */ 541953156443SSuren Baghdasaryan count_vm_event(PGFAULT); 542053156443SSuren Baghdasaryan count_memcg_event_mm(mm, PGFAULT); 542153156443SSuren Baghdasaryan 542253156443SSuren Baghdasaryan /* 542353156443SSuren Baghdasaryan * Do not account for unsuccessful faults (e.g. when the address wasn't 542453156443SSuren Baghdasaryan * valid). That includes arch_vma_access_permitted() failing before 542553156443SSuren Baghdasaryan * reaching here. So this is not a "this many hardware page faults" 542653156443SSuren Baghdasaryan * counter. We should use the hw profiling for that. 542753156443SSuren Baghdasaryan */ 542853156443SSuren Baghdasaryan if (ret & VM_FAULT_ERROR) 5429bce617edSPeter Xu return; 5430bce617edSPeter Xu 5431bce617edSPeter Xu /* 5432bce617edSPeter Xu * We define the fault as a major fault when the final successful fault 5433bce617edSPeter Xu * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't 5434bce617edSPeter Xu * handle it immediately previously). 5435bce617edSPeter Xu */ 5436bce617edSPeter Xu major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED); 5437bce617edSPeter Xu 5438a2beb5f1SPeter Xu if (major) 5439a2beb5f1SPeter Xu current->maj_flt++; 5440a2beb5f1SPeter Xu else 5441a2beb5f1SPeter Xu current->min_flt++; 5442a2beb5f1SPeter Xu 5443bce617edSPeter Xu /* 5444a2beb5f1SPeter Xu * If the fault is done for GUP, regs will be NULL. We only do the 5445a2beb5f1SPeter Xu * accounting for the per thread fault counters who triggered the 5446a2beb5f1SPeter Xu * fault, and we skip the perf event updates. 5447bce617edSPeter Xu */ 5448bce617edSPeter Xu if (!regs) 5449bce617edSPeter Xu return; 5450bce617edSPeter Xu 5451a2beb5f1SPeter Xu if (major) 5452bce617edSPeter Xu perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address); 5453a2beb5f1SPeter Xu else 5454bce617edSPeter Xu perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address); 5455bce617edSPeter Xu } 5456bce617edSPeter Xu 5457ec1c86b2SYu Zhao #ifdef CONFIG_LRU_GEN 5458ec1c86b2SYu Zhao static void lru_gen_enter_fault(struct vm_area_struct *vma) 5459ec1c86b2SYu Zhao { 54608788f678SYu Zhao /* the LRU algorithm only applies to accesses with recency */ 54618788f678SYu Zhao current->in_lru_fault = vma_has_recency(vma); 5462ec1c86b2SYu Zhao } 5463ec1c86b2SYu Zhao 5464ec1c86b2SYu Zhao static void lru_gen_exit_fault(void) 5465ec1c86b2SYu Zhao { 5466ec1c86b2SYu Zhao current->in_lru_fault = false; 5467ec1c86b2SYu Zhao } 5468ec1c86b2SYu Zhao #else 5469ec1c86b2SYu Zhao static void lru_gen_enter_fault(struct vm_area_struct *vma) 5470ec1c86b2SYu Zhao { 5471ec1c86b2SYu Zhao } 5472ec1c86b2SYu Zhao 5473ec1c86b2SYu Zhao static void lru_gen_exit_fault(void) 5474ec1c86b2SYu Zhao { 5475ec1c86b2SYu Zhao } 5476ec1c86b2SYu Zhao #endif /* CONFIG_LRU_GEN */ 5477ec1c86b2SYu Zhao 5478cdc5021cSDavid Hildenbrand static vm_fault_t sanitize_fault_flags(struct vm_area_struct *vma, 5479cdc5021cSDavid Hildenbrand unsigned int *flags) 5480cdc5021cSDavid Hildenbrand { 5481cdc5021cSDavid Hildenbrand if (unlikely(*flags & FAULT_FLAG_UNSHARE)) { 5482cdc5021cSDavid Hildenbrand if (WARN_ON_ONCE(*flags & FAULT_FLAG_WRITE)) 5483cdc5021cSDavid Hildenbrand return VM_FAULT_SIGSEGV; 5484cdc5021cSDavid Hildenbrand /* 5485cdc5021cSDavid Hildenbrand * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's 5486cdc5021cSDavid Hildenbrand * just treat it like an ordinary read-fault otherwise. 5487cdc5021cSDavid Hildenbrand */ 5488cdc5021cSDavid Hildenbrand if (!is_cow_mapping(vma->vm_flags)) 5489cdc5021cSDavid Hildenbrand *flags &= ~FAULT_FLAG_UNSHARE; 549079881fedSDavid Hildenbrand } else if (*flags & FAULT_FLAG_WRITE) { 549179881fedSDavid Hildenbrand /* Write faults on read-only mappings are impossible ... */ 549279881fedSDavid Hildenbrand if (WARN_ON_ONCE(!(vma->vm_flags & VM_MAYWRITE))) 549379881fedSDavid Hildenbrand return VM_FAULT_SIGSEGV; 549479881fedSDavid Hildenbrand /* ... and FOLL_FORCE only applies to COW mappings. */ 549579881fedSDavid Hildenbrand if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE) && 549679881fedSDavid Hildenbrand !is_cow_mapping(vma->vm_flags))) 549779881fedSDavid Hildenbrand return VM_FAULT_SIGSEGV; 5498cdc5021cSDavid Hildenbrand } 54994089eef0SSuren Baghdasaryan #ifdef CONFIG_PER_VMA_LOCK 55004089eef0SSuren Baghdasaryan /* 55014089eef0SSuren Baghdasaryan * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of 55024089eef0SSuren Baghdasaryan * the assumption that lock is dropped on VM_FAULT_RETRY. 55034089eef0SSuren Baghdasaryan */ 55044089eef0SSuren Baghdasaryan if (WARN_ON_ONCE((*flags & 55054089eef0SSuren Baghdasaryan (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)) == 55064089eef0SSuren Baghdasaryan (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT))) 55074089eef0SSuren Baghdasaryan return VM_FAULT_SIGSEGV; 55084089eef0SSuren Baghdasaryan #endif 55094089eef0SSuren Baghdasaryan 5510cdc5021cSDavid Hildenbrand return 0; 5511cdc5021cSDavid Hildenbrand } 5512cdc5021cSDavid Hildenbrand 55139a95f3cfSPaul Cassella /* 55149a95f3cfSPaul Cassella * By the time we get here, we already hold the mm semaphore 55159a95f3cfSPaul Cassella * 5516c1e8d7c6SMichel Lespinasse * The mmap_lock may have been released depending on flags and our 55179138e47eSMatthew Wilcox (Oracle) * return value. See filemap_fault() and __folio_lock_or_retry(). 55189a95f3cfSPaul Cassella */ 55192b740303SSouptick Joarder vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address, 5520bce617edSPeter Xu unsigned int flags, struct pt_regs *regs) 5521519e5247SJohannes Weiner { 552253156443SSuren Baghdasaryan /* If the fault handler drops the mmap_lock, vma may be freed */ 552353156443SSuren Baghdasaryan struct mm_struct *mm = vma->vm_mm; 55242b740303SSouptick Joarder vm_fault_t ret; 5525519e5247SJohannes Weiner 5526519e5247SJohannes Weiner __set_current_state(TASK_RUNNING); 5527519e5247SJohannes Weiner 5528cdc5021cSDavid Hildenbrand ret = sanitize_fault_flags(vma, &flags); 5529cdc5021cSDavid Hildenbrand if (ret) 553053156443SSuren Baghdasaryan goto out; 5531cdc5021cSDavid Hildenbrand 5532de0c799bSLaurent Dufour if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE, 5533de0c799bSLaurent Dufour flags & FAULT_FLAG_INSTRUCTION, 553453156443SSuren Baghdasaryan flags & FAULT_FLAG_REMOTE)) { 553553156443SSuren Baghdasaryan ret = VM_FAULT_SIGSEGV; 553653156443SSuren Baghdasaryan goto out; 553753156443SSuren Baghdasaryan } 5538de0c799bSLaurent Dufour 5539519e5247SJohannes Weiner /* 5540519e5247SJohannes Weiner * Enable the memcg OOM handling for faults triggered in user 5541519e5247SJohannes Weiner * space. Kernel faults are handled more gracefully. 5542519e5247SJohannes Weiner */ 5543519e5247SJohannes Weiner if (flags & FAULT_FLAG_USER) 554429ef680aSMichal Hocko mem_cgroup_enter_user_fault(); 5545519e5247SJohannes Weiner 5546ec1c86b2SYu Zhao lru_gen_enter_fault(vma); 5547ec1c86b2SYu Zhao 5548bae473a4SKirill A. Shutemov if (unlikely(is_vm_hugetlb_page(vma))) 5549bae473a4SKirill A. Shutemov ret = hugetlb_fault(vma->vm_mm, vma, address, flags); 5550bae473a4SKirill A. Shutemov else 5551dcddffd4SKirill A. Shutemov ret = __handle_mm_fault(vma, address, flags); 5552519e5247SJohannes Weiner 5553ec1c86b2SYu Zhao lru_gen_exit_fault(); 5554ec1c86b2SYu Zhao 555549426420SJohannes Weiner if (flags & FAULT_FLAG_USER) { 555629ef680aSMichal Hocko mem_cgroup_exit_user_fault(); 555749426420SJohannes Weiner /* 555849426420SJohannes Weiner * The task may have entered a memcg OOM situation but 555949426420SJohannes Weiner * if the allocation error was handled gracefully (no 556049426420SJohannes Weiner * VM_FAULT_OOM), there is no need to kill anything. 556149426420SJohannes Weiner * Just clean up the OOM state peacefully. 556249426420SJohannes Weiner */ 556349426420SJohannes Weiner if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM)) 556449426420SJohannes Weiner mem_cgroup_oom_synchronize(false); 556549426420SJohannes Weiner } 556653156443SSuren Baghdasaryan out: 556753156443SSuren Baghdasaryan mm_account_fault(mm, regs, address, flags, ret); 5568bce617edSPeter Xu 5569519e5247SJohannes Weiner return ret; 5570519e5247SJohannes Weiner } 5571e1d6d01aSJesse Barnes EXPORT_SYMBOL_GPL(handle_mm_fault); 5572519e5247SJohannes Weiner 5573c2508ec5SLinus Torvalds #ifdef CONFIG_LOCK_MM_AND_FIND_VMA 5574c2508ec5SLinus Torvalds #include <linux/extable.h> 5575c2508ec5SLinus Torvalds 5576c2508ec5SLinus Torvalds static inline bool get_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs) 5577c2508ec5SLinus Torvalds { 55784542057eSLinus Torvalds if (likely(mmap_read_trylock(mm))) 5579c2508ec5SLinus Torvalds return true; 5580c2508ec5SLinus Torvalds 5581c2508ec5SLinus Torvalds if (regs && !user_mode(regs)) { 55828fa50708SJiaxun Yang unsigned long ip = exception_ip(regs); 5583c2508ec5SLinus Torvalds if (!search_exception_tables(ip)) 5584c2508ec5SLinus Torvalds return false; 5585c2508ec5SLinus Torvalds } 5586c2508ec5SLinus Torvalds 5587eda00472SLinus Torvalds return !mmap_read_lock_killable(mm); 5588c2508ec5SLinus Torvalds } 5589c2508ec5SLinus Torvalds 5590c2508ec5SLinus Torvalds static inline bool mmap_upgrade_trylock(struct mm_struct *mm) 5591c2508ec5SLinus Torvalds { 5592c2508ec5SLinus Torvalds /* 5593c2508ec5SLinus Torvalds * We don't have this operation yet. 5594c2508ec5SLinus Torvalds * 5595c2508ec5SLinus Torvalds * It should be easy enough to do: it's basically a 5596c2508ec5SLinus Torvalds * atomic_long_try_cmpxchg_acquire() 5597c2508ec5SLinus Torvalds * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but 5598c2508ec5SLinus Torvalds * it also needs the proper lockdep magic etc. 5599c2508ec5SLinus Torvalds */ 5600c2508ec5SLinus Torvalds return false; 5601c2508ec5SLinus Torvalds } 5602c2508ec5SLinus Torvalds 5603c2508ec5SLinus Torvalds static inline bool upgrade_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs) 5604c2508ec5SLinus Torvalds { 5605c2508ec5SLinus Torvalds mmap_read_unlock(mm); 5606c2508ec5SLinus Torvalds if (regs && !user_mode(regs)) { 56078fa50708SJiaxun Yang unsigned long ip = exception_ip(regs); 5608c2508ec5SLinus Torvalds if (!search_exception_tables(ip)) 5609c2508ec5SLinus Torvalds return false; 5610c2508ec5SLinus Torvalds } 5611eda00472SLinus Torvalds return !mmap_write_lock_killable(mm); 5612c2508ec5SLinus Torvalds } 5613c2508ec5SLinus Torvalds 5614c2508ec5SLinus Torvalds /* 5615c2508ec5SLinus Torvalds * Helper for page fault handling. 5616c2508ec5SLinus Torvalds * 5617c2508ec5SLinus Torvalds * This is kind of equivalend to "mmap_read_lock()" followed 5618c2508ec5SLinus Torvalds * by "find_extend_vma()", except it's a lot more careful about 5619c2508ec5SLinus Torvalds * the locking (and will drop the lock on failure). 5620c2508ec5SLinus Torvalds * 5621c2508ec5SLinus Torvalds * For example, if we have a kernel bug that causes a page 5622c2508ec5SLinus Torvalds * fault, we don't want to just use mmap_read_lock() to get 5623c2508ec5SLinus Torvalds * the mm lock, because that would deadlock if the bug were 5624c2508ec5SLinus Torvalds * to happen while we're holding the mm lock for writing. 5625c2508ec5SLinus Torvalds * 5626c2508ec5SLinus Torvalds * So this checks the exception tables on kernel faults in 5627c2508ec5SLinus Torvalds * order to only do this all for instructions that are actually 5628c2508ec5SLinus Torvalds * expected to fault. 5629c2508ec5SLinus Torvalds * 5630c2508ec5SLinus Torvalds * We can also actually take the mm lock for writing if we 5631c2508ec5SLinus Torvalds * need to extend the vma, which helps the VM layer a lot. 5632c2508ec5SLinus Torvalds */ 5633c2508ec5SLinus Torvalds struct vm_area_struct *lock_mm_and_find_vma(struct mm_struct *mm, 5634c2508ec5SLinus Torvalds unsigned long addr, struct pt_regs *regs) 5635c2508ec5SLinus Torvalds { 5636c2508ec5SLinus Torvalds struct vm_area_struct *vma; 5637c2508ec5SLinus Torvalds 5638c2508ec5SLinus Torvalds if (!get_mmap_lock_carefully(mm, regs)) 5639c2508ec5SLinus Torvalds return NULL; 5640c2508ec5SLinus Torvalds 5641c2508ec5SLinus Torvalds vma = find_vma(mm, addr); 5642c2508ec5SLinus Torvalds if (likely(vma && (vma->vm_start <= addr))) 5643c2508ec5SLinus Torvalds return vma; 5644c2508ec5SLinus Torvalds 5645c2508ec5SLinus Torvalds /* 5646c2508ec5SLinus Torvalds * Well, dang. We might still be successful, but only 5647c2508ec5SLinus Torvalds * if we can extend a vma to do so. 5648c2508ec5SLinus Torvalds */ 5649c2508ec5SLinus Torvalds if (!vma || !(vma->vm_flags & VM_GROWSDOWN)) { 5650c2508ec5SLinus Torvalds mmap_read_unlock(mm); 5651c2508ec5SLinus Torvalds return NULL; 5652c2508ec5SLinus Torvalds } 5653c2508ec5SLinus Torvalds 5654c2508ec5SLinus Torvalds /* 5655c2508ec5SLinus Torvalds * We can try to upgrade the mmap lock atomically, 5656c2508ec5SLinus Torvalds * in which case we can continue to use the vma 5657c2508ec5SLinus Torvalds * we already looked up. 5658c2508ec5SLinus Torvalds * 5659c2508ec5SLinus Torvalds * Otherwise we'll have to drop the mmap lock and 5660c2508ec5SLinus Torvalds * re-take it, and also look up the vma again, 5661c2508ec5SLinus Torvalds * re-checking it. 5662c2508ec5SLinus Torvalds */ 5663c2508ec5SLinus Torvalds if (!mmap_upgrade_trylock(mm)) { 5664c2508ec5SLinus Torvalds if (!upgrade_mmap_lock_carefully(mm, regs)) 5665c2508ec5SLinus Torvalds return NULL; 5666c2508ec5SLinus Torvalds 5667c2508ec5SLinus Torvalds vma = find_vma(mm, addr); 5668c2508ec5SLinus Torvalds if (!vma) 5669c2508ec5SLinus Torvalds goto fail; 5670c2508ec5SLinus Torvalds if (vma->vm_start <= addr) 5671c2508ec5SLinus Torvalds goto success; 5672c2508ec5SLinus Torvalds if (!(vma->vm_flags & VM_GROWSDOWN)) 5673c2508ec5SLinus Torvalds goto fail; 5674c2508ec5SLinus Torvalds } 5675c2508ec5SLinus Torvalds 56768d7071afSLinus Torvalds if (expand_stack_locked(vma, addr)) 5677c2508ec5SLinus Torvalds goto fail; 5678c2508ec5SLinus Torvalds 5679c2508ec5SLinus Torvalds success: 5680c2508ec5SLinus Torvalds mmap_write_downgrade(mm); 5681c2508ec5SLinus Torvalds return vma; 5682c2508ec5SLinus Torvalds 5683c2508ec5SLinus Torvalds fail: 5684c2508ec5SLinus Torvalds mmap_write_unlock(mm); 5685c2508ec5SLinus Torvalds return NULL; 5686c2508ec5SLinus Torvalds } 5687c2508ec5SLinus Torvalds #endif 5688c2508ec5SLinus Torvalds 568950ee3253SSuren Baghdasaryan #ifdef CONFIG_PER_VMA_LOCK 569050ee3253SSuren Baghdasaryan /* 569150ee3253SSuren Baghdasaryan * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be 569250ee3253SSuren Baghdasaryan * stable and not isolated. If the VMA is not found or is being modified the 569350ee3253SSuren Baghdasaryan * function returns NULL. 569450ee3253SSuren Baghdasaryan */ 569550ee3253SSuren Baghdasaryan struct vm_area_struct *lock_vma_under_rcu(struct mm_struct *mm, 569650ee3253SSuren Baghdasaryan unsigned long address) 569750ee3253SSuren Baghdasaryan { 569850ee3253SSuren Baghdasaryan MA_STATE(mas, &mm->mm_mt, address, address); 569950ee3253SSuren Baghdasaryan struct vm_area_struct *vma; 570050ee3253SSuren Baghdasaryan 570150ee3253SSuren Baghdasaryan rcu_read_lock(); 570250ee3253SSuren Baghdasaryan retry: 570350ee3253SSuren Baghdasaryan vma = mas_walk(&mas); 570450ee3253SSuren Baghdasaryan if (!vma) 570550ee3253SSuren Baghdasaryan goto inval; 570650ee3253SSuren Baghdasaryan 570750ee3253SSuren Baghdasaryan if (!vma_start_read(vma)) 570850ee3253SSuren Baghdasaryan goto inval; 570950ee3253SSuren Baghdasaryan 5710444eeb17SSuren Baghdasaryan /* 5711657b5146SJann Horn * find_mergeable_anon_vma uses adjacent vmas which are not locked. 5712657b5146SJann Horn * This check must happen after vma_start_read(); otherwise, a 5713657b5146SJann Horn * concurrent mremap() with MREMAP_DONTUNMAP could dissociate the VMA 5714657b5146SJann Horn * from its anon_vma. 5715657b5146SJann Horn */ 571629a22b9eSSuren Baghdasaryan if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) 5717657b5146SJann Horn goto inval_end_read; 5718444eeb17SSuren Baghdasaryan 571950ee3253SSuren Baghdasaryan /* Check since vm_start/vm_end might change before we lock the VMA */ 5720657b5146SJann Horn if (unlikely(address < vma->vm_start || address >= vma->vm_end)) 5721657b5146SJann Horn goto inval_end_read; 572250ee3253SSuren Baghdasaryan 572350ee3253SSuren Baghdasaryan /* Check if the VMA got isolated after we found it */ 572450ee3253SSuren Baghdasaryan if (vma->detached) { 572550ee3253SSuren Baghdasaryan vma_end_read(vma); 572652f23865SSuren Baghdasaryan count_vm_vma_lock_event(VMA_LOCK_MISS); 572750ee3253SSuren Baghdasaryan /* The area was replaced with another one */ 572850ee3253SSuren Baghdasaryan goto retry; 572950ee3253SSuren Baghdasaryan } 573050ee3253SSuren Baghdasaryan 573150ee3253SSuren Baghdasaryan rcu_read_unlock(); 573250ee3253SSuren Baghdasaryan return vma; 5733657b5146SJann Horn 5734657b5146SJann Horn inval_end_read: 5735657b5146SJann Horn vma_end_read(vma); 573650ee3253SSuren Baghdasaryan inval: 573750ee3253SSuren Baghdasaryan rcu_read_unlock(); 573852f23865SSuren Baghdasaryan count_vm_vma_lock_event(VMA_LOCK_ABORT); 573950ee3253SSuren Baghdasaryan return NULL; 574050ee3253SSuren Baghdasaryan } 574150ee3253SSuren Baghdasaryan #endif /* CONFIG_PER_VMA_LOCK */ 574250ee3253SSuren Baghdasaryan 574390eceff1SKirill A. Shutemov #ifndef __PAGETABLE_P4D_FOLDED 574490eceff1SKirill A. Shutemov /* 574590eceff1SKirill A. Shutemov * Allocate p4d page table. 574690eceff1SKirill A. Shutemov * We've already handled the fast-path in-line. 574790eceff1SKirill A. Shutemov */ 574890eceff1SKirill A. Shutemov int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) 574990eceff1SKirill A. Shutemov { 575090eceff1SKirill A. Shutemov p4d_t *new = p4d_alloc_one(mm, address); 575190eceff1SKirill A. Shutemov if (!new) 575290eceff1SKirill A. Shutemov return -ENOMEM; 575390eceff1SKirill A. Shutemov 575490eceff1SKirill A. Shutemov spin_lock(&mm->page_table_lock); 5755ed33b5a6SQi Zheng if (pgd_present(*pgd)) { /* Another has populated it */ 575690eceff1SKirill A. Shutemov p4d_free(mm, new); 5757ed33b5a6SQi Zheng } else { 5758ed33b5a6SQi Zheng smp_wmb(); /* See comment in pmd_install() */ 575990eceff1SKirill A. Shutemov pgd_populate(mm, pgd, new); 5760ed33b5a6SQi Zheng } 576190eceff1SKirill A. Shutemov spin_unlock(&mm->page_table_lock); 576290eceff1SKirill A. Shutemov return 0; 576390eceff1SKirill A. Shutemov } 576490eceff1SKirill A. Shutemov #endif /* __PAGETABLE_P4D_FOLDED */ 576590eceff1SKirill A. Shutemov 57661da177e4SLinus Torvalds #ifndef __PAGETABLE_PUD_FOLDED 57671da177e4SLinus Torvalds /* 57681da177e4SLinus Torvalds * Allocate page upper directory. 5769872fec16SHugh Dickins * We've already handled the fast-path in-line. 57701da177e4SLinus Torvalds */ 5771c2febafcSKirill A. Shutemov int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address) 57721da177e4SLinus Torvalds { 5773c74df32cSHugh Dickins pud_t *new = pud_alloc_one(mm, address); 5774c74df32cSHugh Dickins if (!new) 57751bb3630eSHugh Dickins return -ENOMEM; 57761da177e4SLinus Torvalds 5777872fec16SHugh Dickins spin_lock(&mm->page_table_lock); 5778b4e98d9aSKirill A. Shutemov if (!p4d_present(*p4d)) { 5779b4e98d9aSKirill A. Shutemov mm_inc_nr_puds(mm); 5780ed33b5a6SQi Zheng smp_wmb(); /* See comment in pmd_install() */ 5781c2febafcSKirill A. Shutemov p4d_populate(mm, p4d, new); 5782b4e98d9aSKirill A. Shutemov } else /* Another has populated it */ 5783c2febafcSKirill A. Shutemov pud_free(mm, new); 5784872fec16SHugh Dickins spin_unlock(&mm->page_table_lock); 57851bb3630eSHugh Dickins return 0; 57861da177e4SLinus Torvalds } 57871da177e4SLinus Torvalds #endif /* __PAGETABLE_PUD_FOLDED */ 57881da177e4SLinus Torvalds 57891da177e4SLinus Torvalds #ifndef __PAGETABLE_PMD_FOLDED 57901da177e4SLinus Torvalds /* 57911da177e4SLinus Torvalds * Allocate page middle directory. 5792872fec16SHugh Dickins * We've already handled the fast-path in-line. 57931da177e4SLinus Torvalds */ 57941bb3630eSHugh Dickins int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) 57951da177e4SLinus Torvalds { 5796a00cc7d9SMatthew Wilcox spinlock_t *ptl; 5797c74df32cSHugh Dickins pmd_t *new = pmd_alloc_one(mm, address); 5798c74df32cSHugh Dickins if (!new) 57991bb3630eSHugh Dickins return -ENOMEM; 58001da177e4SLinus Torvalds 5801a00cc7d9SMatthew Wilcox ptl = pud_lock(mm, pud); 5802dc6c9a35SKirill A. Shutemov if (!pud_present(*pud)) { 5803dc6c9a35SKirill A. Shutemov mm_inc_nr_pmds(mm); 5804ed33b5a6SQi Zheng smp_wmb(); /* See comment in pmd_install() */ 58051da177e4SLinus Torvalds pud_populate(mm, pud, new); 5806ed33b5a6SQi Zheng } else { /* Another has populated it */ 58075e541973SBenjamin Herrenschmidt pmd_free(mm, new); 5808ed33b5a6SQi Zheng } 5809a00cc7d9SMatthew Wilcox spin_unlock(ptl); 58101bb3630eSHugh Dickins return 0; 58111da177e4SLinus Torvalds } 58121da177e4SLinus Torvalds #endif /* __PAGETABLE_PMD_FOLDED */ 58131da177e4SLinus Torvalds 58143b6748e2SJohannes Weiner /** 58159fd6dad1SPaolo Bonzini * follow_pte - look up PTE at a user virtual address 58169fd6dad1SPaolo Bonzini * @mm: the mm_struct of the target address space 58179fd6dad1SPaolo Bonzini * @address: user virtual address 58189fd6dad1SPaolo Bonzini * @ptepp: location to store found PTE 58199fd6dad1SPaolo Bonzini * @ptlp: location to store the lock for the PTE 58209fd6dad1SPaolo Bonzini * 58219fd6dad1SPaolo Bonzini * On a successful return, the pointer to the PTE is stored in @ptepp; 58229fd6dad1SPaolo Bonzini * the corresponding lock is taken and its location is stored in @ptlp. 58239fd6dad1SPaolo Bonzini * The contents of the PTE are only stable until @ptlp is released; 58249fd6dad1SPaolo Bonzini * any further use, if any, must be protected against invalidation 58259fd6dad1SPaolo Bonzini * with MMU notifiers. 58269fd6dad1SPaolo Bonzini * 58279fd6dad1SPaolo Bonzini * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore 58289fd6dad1SPaolo Bonzini * should be taken for read. 58299fd6dad1SPaolo Bonzini * 58309fd6dad1SPaolo Bonzini * KVM uses this function. While it is arguably less bad than ``follow_pfn``, 58319fd6dad1SPaolo Bonzini * it is not a good general-purpose API. 58329fd6dad1SPaolo Bonzini * 58339fd6dad1SPaolo Bonzini * Return: zero on success, -ve otherwise. 58349fd6dad1SPaolo Bonzini */ 58359fd6dad1SPaolo Bonzini int follow_pte(struct mm_struct *mm, unsigned long address, 58369fd6dad1SPaolo Bonzini pte_t **ptepp, spinlock_t **ptlp) 58379fd6dad1SPaolo Bonzini { 58380e5e64c0SMuchun Song pgd_t *pgd; 58390e5e64c0SMuchun Song p4d_t *p4d; 58400e5e64c0SMuchun Song pud_t *pud; 58410e5e64c0SMuchun Song pmd_t *pmd; 58420e5e64c0SMuchun Song pte_t *ptep; 58430e5e64c0SMuchun Song 58440e5e64c0SMuchun Song pgd = pgd_offset(mm, address); 58450e5e64c0SMuchun Song if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) 58460e5e64c0SMuchun Song goto out; 58470e5e64c0SMuchun Song 58480e5e64c0SMuchun Song p4d = p4d_offset(pgd, address); 58490e5e64c0SMuchun Song if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d))) 58500e5e64c0SMuchun Song goto out; 58510e5e64c0SMuchun Song 58520e5e64c0SMuchun Song pud = pud_offset(p4d, address); 58530e5e64c0SMuchun Song if (pud_none(*pud) || unlikely(pud_bad(*pud))) 58540e5e64c0SMuchun Song goto out; 58550e5e64c0SMuchun Song 58560e5e64c0SMuchun Song pmd = pmd_offset(pud, address); 58570e5e64c0SMuchun Song VM_BUG_ON(pmd_trans_huge(*pmd)); 58580e5e64c0SMuchun Song 58590e5e64c0SMuchun Song ptep = pte_offset_map_lock(mm, pmd, address, ptlp); 58603db82b93SHugh Dickins if (!ptep) 58613db82b93SHugh Dickins goto out; 5862c33c7948SRyan Roberts if (!pte_present(ptep_get(ptep))) 58630e5e64c0SMuchun Song goto unlock; 58640e5e64c0SMuchun Song *ptepp = ptep; 58650e5e64c0SMuchun Song return 0; 58660e5e64c0SMuchun Song unlock: 58670e5e64c0SMuchun Song pte_unmap_unlock(ptep, *ptlp); 58680e5e64c0SMuchun Song out: 58690e5e64c0SMuchun Song return -EINVAL; 58709fd6dad1SPaolo Bonzini } 58719fd6dad1SPaolo Bonzini EXPORT_SYMBOL_GPL(follow_pte); 58729fd6dad1SPaolo Bonzini 58739fd6dad1SPaolo Bonzini /** 58743b6748e2SJohannes Weiner * follow_pfn - look up PFN at a user virtual address 58753b6748e2SJohannes Weiner * @vma: memory mapping 58763b6748e2SJohannes Weiner * @address: user virtual address 58773b6748e2SJohannes Weiner * @pfn: location to store found PFN 58783b6748e2SJohannes Weiner * 58793b6748e2SJohannes Weiner * Only IO mappings and raw PFN mappings are allowed. 58803b6748e2SJohannes Weiner * 58819fd6dad1SPaolo Bonzini * This function does not allow the caller to read the permissions 58829fd6dad1SPaolo Bonzini * of the PTE. Do not use it. 58839fd6dad1SPaolo Bonzini * 5884a862f68aSMike Rapoport * Return: zero and the pfn at @pfn on success, -ve otherwise. 58853b6748e2SJohannes Weiner */ 58863b6748e2SJohannes Weiner int follow_pfn(struct vm_area_struct *vma, unsigned long address, 58873b6748e2SJohannes Weiner unsigned long *pfn) 58883b6748e2SJohannes Weiner { 58893b6748e2SJohannes Weiner int ret = -EINVAL; 58903b6748e2SJohannes Weiner spinlock_t *ptl; 58913b6748e2SJohannes Weiner pte_t *ptep; 58923b6748e2SJohannes Weiner 58933b6748e2SJohannes Weiner if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 58943b6748e2SJohannes Weiner return ret; 58953b6748e2SJohannes Weiner 58969fd6dad1SPaolo Bonzini ret = follow_pte(vma->vm_mm, address, &ptep, &ptl); 58973b6748e2SJohannes Weiner if (ret) 58983b6748e2SJohannes Weiner return ret; 5899c33c7948SRyan Roberts *pfn = pte_pfn(ptep_get(ptep)); 59003b6748e2SJohannes Weiner pte_unmap_unlock(ptep, ptl); 59013b6748e2SJohannes Weiner return 0; 59023b6748e2SJohannes Weiner } 59033b6748e2SJohannes Weiner EXPORT_SYMBOL(follow_pfn); 59043b6748e2SJohannes Weiner 590528b2ee20SRik van Riel #ifdef CONFIG_HAVE_IOREMAP_PROT 5906d87fe660Svenkatesh.pallipadi@intel.com int follow_phys(struct vm_area_struct *vma, 590728b2ee20SRik van Riel unsigned long address, unsigned int flags, 5908d87fe660Svenkatesh.pallipadi@intel.com unsigned long *prot, resource_size_t *phys) 590928b2ee20SRik van Riel { 591003668a4dSJohannes Weiner int ret = -EINVAL; 591128b2ee20SRik van Riel pte_t *ptep, pte; 591228b2ee20SRik van Riel spinlock_t *ptl; 591328b2ee20SRik van Riel 5914d87fe660Svenkatesh.pallipadi@intel.com if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 5915d87fe660Svenkatesh.pallipadi@intel.com goto out; 591628b2ee20SRik van Riel 59179fd6dad1SPaolo Bonzini if (follow_pte(vma->vm_mm, address, &ptep, &ptl)) 5918d87fe660Svenkatesh.pallipadi@intel.com goto out; 5919c33c7948SRyan Roberts pte = ptep_get(ptep); 592003668a4dSJohannes Weiner 5921f6f37321SLinus Torvalds if ((flags & FOLL_WRITE) && !pte_write(pte)) 592228b2ee20SRik van Riel goto unlock; 592328b2ee20SRik van Riel 592428b2ee20SRik van Riel *prot = pgprot_val(pte_pgprot(pte)); 592503668a4dSJohannes Weiner *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; 592628b2ee20SRik van Riel 592703668a4dSJohannes Weiner ret = 0; 592828b2ee20SRik van Riel unlock: 592928b2ee20SRik van Riel pte_unmap_unlock(ptep, ptl); 593028b2ee20SRik van Riel out: 5931d87fe660Svenkatesh.pallipadi@intel.com return ret; 593228b2ee20SRik van Riel } 593328b2ee20SRik van Riel 593496667f8aSDaniel Vetter /** 593596667f8aSDaniel Vetter * generic_access_phys - generic implementation for iomem mmap access 593696667f8aSDaniel Vetter * @vma: the vma to access 5937f0953a1bSIngo Molnar * @addr: userspace address, not relative offset within @vma 593896667f8aSDaniel Vetter * @buf: buffer to read/write 593996667f8aSDaniel Vetter * @len: length of transfer 594096667f8aSDaniel Vetter * @write: set to FOLL_WRITE when writing, otherwise reading 594196667f8aSDaniel Vetter * 594296667f8aSDaniel Vetter * This is a generic implementation for &vm_operations_struct.access for an 594396667f8aSDaniel Vetter * iomem mapping. This callback is used by access_process_vm() when the @vma is 594496667f8aSDaniel Vetter * not page based. 594596667f8aSDaniel Vetter */ 594628b2ee20SRik van Riel int generic_access_phys(struct vm_area_struct *vma, unsigned long addr, 594728b2ee20SRik van Riel void *buf, int len, int write) 594828b2ee20SRik van Riel { 594928b2ee20SRik van Riel resource_size_t phys_addr; 595028b2ee20SRik van Riel unsigned long prot = 0; 59512bc7273bSKOSAKI Motohiro void __iomem *maddr; 595296667f8aSDaniel Vetter pte_t *ptep, pte; 595396667f8aSDaniel Vetter spinlock_t *ptl; 595496667f8aSDaniel Vetter int offset = offset_in_page(addr); 595596667f8aSDaniel Vetter int ret = -EINVAL; 595628b2ee20SRik van Riel 595796667f8aSDaniel Vetter if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) 595896667f8aSDaniel Vetter return -EINVAL; 595996667f8aSDaniel Vetter 596096667f8aSDaniel Vetter retry: 5961e913a8cdSLinus Torvalds if (follow_pte(vma->vm_mm, addr, &ptep, &ptl)) 596296667f8aSDaniel Vetter return -EINVAL; 5963c33c7948SRyan Roberts pte = ptep_get(ptep); 596496667f8aSDaniel Vetter pte_unmap_unlock(ptep, ptl); 596596667f8aSDaniel Vetter 596696667f8aSDaniel Vetter prot = pgprot_val(pte_pgprot(pte)); 596796667f8aSDaniel Vetter phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; 596896667f8aSDaniel Vetter 596996667f8aSDaniel Vetter if ((write & FOLL_WRITE) && !pte_write(pte)) 597028b2ee20SRik van Riel return -EINVAL; 597128b2ee20SRik van Riel 59729cb12d7bSGrazvydas Ignotas maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot); 597324eee1e4Sjie@chenjie6@huwei.com if (!maddr) 597424eee1e4Sjie@chenjie6@huwei.com return -ENOMEM; 597524eee1e4Sjie@chenjie6@huwei.com 5976e913a8cdSLinus Torvalds if (follow_pte(vma->vm_mm, addr, &ptep, &ptl)) 597796667f8aSDaniel Vetter goto out_unmap; 597896667f8aSDaniel Vetter 5979c33c7948SRyan Roberts if (!pte_same(pte, ptep_get(ptep))) { 598096667f8aSDaniel Vetter pte_unmap_unlock(ptep, ptl); 598196667f8aSDaniel Vetter iounmap(maddr); 598296667f8aSDaniel Vetter 598396667f8aSDaniel Vetter goto retry; 598496667f8aSDaniel Vetter } 598596667f8aSDaniel Vetter 598628b2ee20SRik van Riel if (write) 598728b2ee20SRik van Riel memcpy_toio(maddr + offset, buf, len); 598828b2ee20SRik van Riel else 598928b2ee20SRik van Riel memcpy_fromio(buf, maddr + offset, len); 599096667f8aSDaniel Vetter ret = len; 599196667f8aSDaniel Vetter pte_unmap_unlock(ptep, ptl); 599296667f8aSDaniel Vetter out_unmap: 599328b2ee20SRik van Riel iounmap(maddr); 599428b2ee20SRik van Riel 599596667f8aSDaniel Vetter return ret; 599628b2ee20SRik van Riel } 59975a73633eSUwe Kleine-König EXPORT_SYMBOL_GPL(generic_access_phys); 599828b2ee20SRik van Riel #endif 599928b2ee20SRik van Riel 60000ec76a11SDavid Howells /* 6001d3f5ffcaSJohn Hubbard * Access another process' address space as given in mm. 60020ec76a11SDavid Howells */ 6003c43cfa42SLorenzo Stoakes static int __access_remote_vm(struct mm_struct *mm, unsigned long addr, 6004c43cfa42SLorenzo Stoakes void *buf, int len, unsigned int gup_flags) 60050ec76a11SDavid Howells { 60060ec76a11SDavid Howells void *old_buf = buf; 6007442486ecSLorenzo Stoakes int write = gup_flags & FOLL_WRITE; 60080ec76a11SDavid Howells 6009d8ed45c5SMichel Lespinasse if (mmap_read_lock_killable(mm)) 60101e426fe2SKonstantin Khlebnikov return 0; 60111e426fe2SKonstantin Khlebnikov 601222883973SKirill A. Shutemov /* Untag the address before looking up the VMA */ 601322883973SKirill A. Shutemov addr = untagged_addr_remote(mm, addr); 601422883973SKirill A. Shutemov 6015eee9c708SLinus Torvalds /* Avoid triggering the temporary warning in __get_user_pages */ 6016eee9c708SLinus Torvalds if (!vma_lookup(mm, addr) && !expand_stack(mm, addr)) 6017eee9c708SLinus Torvalds return 0; 6018eee9c708SLinus Torvalds 6019183ff22bSSimon Arlott /* ignore errors, just check how much was successfully transferred */ 60200ec76a11SDavid Howells while (len) { 6021ca5e8632SLorenzo Stoakes int bytes, offset; 60220ec76a11SDavid Howells void *maddr; 6023ca5e8632SLorenzo Stoakes struct vm_area_struct *vma = NULL; 6024ca5e8632SLorenzo Stoakes struct page *page = get_user_page_vma_remote(mm, addr, 6025ca5e8632SLorenzo Stoakes gup_flags, &vma); 60260ec76a11SDavid Howells 60276a1960b8SLorenzo Stoakes if (IS_ERR(page)) { 60288d7071afSLinus Torvalds /* We might need to expand the stack to access it */ 60298d7071afSLinus Torvalds vma = vma_lookup(mm, addr); 60308d7071afSLinus Torvalds if (!vma) { 60318d7071afSLinus Torvalds vma = expand_stack(mm, addr); 60329471f1f2SLinus Torvalds 60339471f1f2SLinus Torvalds /* mmap_lock was dropped on failure */ 60348d7071afSLinus Torvalds if (!vma) 60359471f1f2SLinus Torvalds return buf - old_buf; 60369471f1f2SLinus Torvalds 60379471f1f2SLinus Torvalds /* Try again if stack expansion worked */ 60389471f1f2SLinus Torvalds continue; 60398d7071afSLinus Torvalds } 60408d7071afSLinus Torvalds 604128b2ee20SRik van Riel /* 604228b2ee20SRik van Riel * Check if this is a VM_IO | VM_PFNMAP VMA, which 604328b2ee20SRik van Riel * we can access using slightly different code. 604428b2ee20SRik van Riel */ 60459471f1f2SLinus Torvalds bytes = 0; 60469471f1f2SLinus Torvalds #ifdef CONFIG_HAVE_IOREMAP_PROT 604728b2ee20SRik van Riel if (vma->vm_ops && vma->vm_ops->access) 60489471f1f2SLinus Torvalds bytes = vma->vm_ops->access(vma, addr, buf, 604928b2ee20SRik van Riel len, write); 6050dbffcd03SRik van Riel #endif 60519471f1f2SLinus Torvalds if (bytes <= 0) 60529471f1f2SLinus Torvalds break; 605328b2ee20SRik van Riel } else { 60540ec76a11SDavid Howells bytes = len; 60550ec76a11SDavid Howells offset = addr & (PAGE_SIZE-1); 60560ec76a11SDavid Howells if (bytes > PAGE_SIZE-offset) 60570ec76a11SDavid Howells bytes = PAGE_SIZE-offset; 60580ec76a11SDavid Howells 6059f7ef5fe7SFabio M. De Francesco maddr = kmap_local_page(page); 60600ec76a11SDavid Howells if (write) { 60610ec76a11SDavid Howells copy_to_user_page(vma, page, addr, 60620ec76a11SDavid Howells maddr + offset, buf, bytes); 60630ec76a11SDavid Howells set_page_dirty_lock(page); 60640ec76a11SDavid Howells } else { 60650ec76a11SDavid Howells copy_from_user_page(vma, page, addr, 60660ec76a11SDavid Howells buf, maddr + offset, bytes); 60670ec76a11SDavid Howells } 6068f7ef5fe7SFabio M. De Francesco unmap_and_put_page(page, maddr); 606928b2ee20SRik van Riel } 60700ec76a11SDavid Howells len -= bytes; 60710ec76a11SDavid Howells buf += bytes; 60720ec76a11SDavid Howells addr += bytes; 60730ec76a11SDavid Howells } 6074d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 60750ec76a11SDavid Howells 60760ec76a11SDavid Howells return buf - old_buf; 60770ec76a11SDavid Howells } 607803252919SAndi Kleen 60795ddd36b9SStephen Wilson /** 6080ae91dbfcSRandy Dunlap * access_remote_vm - access another process' address space 60815ddd36b9SStephen Wilson * @mm: the mm_struct of the target address space 60825ddd36b9SStephen Wilson * @addr: start address to access 60835ddd36b9SStephen Wilson * @buf: source or destination buffer 60845ddd36b9SStephen Wilson * @len: number of bytes to transfer 60856347e8d5SLorenzo Stoakes * @gup_flags: flags modifying lookup behaviour 60865ddd36b9SStephen Wilson * 60875ddd36b9SStephen Wilson * The caller must hold a reference on @mm. 6088a862f68aSMike Rapoport * 6089a862f68aSMike Rapoport * Return: number of bytes copied from source to destination. 60905ddd36b9SStephen Wilson */ 60915ddd36b9SStephen Wilson int access_remote_vm(struct mm_struct *mm, unsigned long addr, 60926347e8d5SLorenzo Stoakes void *buf, int len, unsigned int gup_flags) 60935ddd36b9SStephen Wilson { 6094d3f5ffcaSJohn Hubbard return __access_remote_vm(mm, addr, buf, len, gup_flags); 60955ddd36b9SStephen Wilson } 60965ddd36b9SStephen Wilson 609703252919SAndi Kleen /* 6098206cb636SStephen Wilson * Access another process' address space. 6099206cb636SStephen Wilson * Source/target buffer must be kernel space, 6100206cb636SStephen Wilson * Do not walk the page table directly, use get_user_pages 6101206cb636SStephen Wilson */ 6102206cb636SStephen Wilson int access_process_vm(struct task_struct *tsk, unsigned long addr, 6103f307ab6dSLorenzo Stoakes void *buf, int len, unsigned int gup_flags) 6104206cb636SStephen Wilson { 6105206cb636SStephen Wilson struct mm_struct *mm; 6106206cb636SStephen Wilson int ret; 6107206cb636SStephen Wilson 6108206cb636SStephen Wilson mm = get_task_mm(tsk); 6109206cb636SStephen Wilson if (!mm) 6110206cb636SStephen Wilson return 0; 6111206cb636SStephen Wilson 6112d3f5ffcaSJohn Hubbard ret = __access_remote_vm(mm, addr, buf, len, gup_flags); 6113442486ecSLorenzo Stoakes 6114206cb636SStephen Wilson mmput(mm); 6115206cb636SStephen Wilson 6116206cb636SStephen Wilson return ret; 6117206cb636SStephen Wilson } 6118fcd35857SCatalin Marinas EXPORT_SYMBOL_GPL(access_process_vm); 6119206cb636SStephen Wilson 612003252919SAndi Kleen /* 612103252919SAndi Kleen * Print the name of a VMA. 612203252919SAndi Kleen */ 612303252919SAndi Kleen void print_vma_addr(char *prefix, unsigned long ip) 612403252919SAndi Kleen { 612503252919SAndi Kleen struct mm_struct *mm = current->mm; 612603252919SAndi Kleen struct vm_area_struct *vma; 612703252919SAndi Kleen 6128e8bff74aSIngo Molnar /* 61290a7f682dSMichal Hocko * we might be running from an atomic context so we cannot sleep 6130e8bff74aSIngo Molnar */ 6131d8ed45c5SMichel Lespinasse if (!mmap_read_trylock(mm)) 6132e8bff74aSIngo Molnar return; 6133e8bff74aSIngo Molnar 613403252919SAndi Kleen vma = find_vma(mm, ip); 613503252919SAndi Kleen if (vma && vma->vm_file) { 613603252919SAndi Kleen struct file *f = vma->vm_file; 61370a7f682dSMichal Hocko char *buf = (char *)__get_free_page(GFP_NOWAIT); 613803252919SAndi Kleen if (buf) { 61392fbc57c5SAndy Shevchenko char *p; 614003252919SAndi Kleen 61419bf39ab2SMiklos Szeredi p = file_path(f, buf, PAGE_SIZE); 614203252919SAndi Kleen if (IS_ERR(p)) 614303252919SAndi Kleen p = "?"; 61442fbc57c5SAndy Shevchenko printk("%s%s[%lx+%lx]", prefix, kbasename(p), 614503252919SAndi Kleen vma->vm_start, 614603252919SAndi Kleen vma->vm_end - vma->vm_start); 614703252919SAndi Kleen free_page((unsigned long)buf); 614803252919SAndi Kleen } 614903252919SAndi Kleen } 6150d8ed45c5SMichel Lespinasse mmap_read_unlock(mm); 615103252919SAndi Kleen } 61523ee1afa3SNick Piggin 6153662bbcb2SMichael S. Tsirkin #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP) 61549ec23531SDavid Hildenbrand void __might_fault(const char *file, int line) 61553ee1afa3SNick Piggin { 61569ec23531SDavid Hildenbrand if (pagefault_disabled()) 6157662bbcb2SMichael S. Tsirkin return; 615842a38756SThomas Gleixner __might_sleep(file, line); 61599ec23531SDavid Hildenbrand #if defined(CONFIG_DEBUG_ATOMIC_SLEEP) 6160662bbcb2SMichael S. Tsirkin if (current->mm) 6161da1c55f1SMichel Lespinasse might_lock_read(¤t->mm->mmap_lock); 61629ec23531SDavid Hildenbrand #endif 61633ee1afa3SNick Piggin } 61649ec23531SDavid Hildenbrand EXPORT_SYMBOL(__might_fault); 61653ee1afa3SNick Piggin #endif 616647ad8475SAndrea Arcangeli 616747ad8475SAndrea Arcangeli #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) 6168c6ddfb6cSHuang Ying /* 6169c6ddfb6cSHuang Ying * Process all subpages of the specified huge page with the specified 6170c6ddfb6cSHuang Ying * operation. The target subpage will be processed last to keep its 6171c6ddfb6cSHuang Ying * cache lines hot. 6172c6ddfb6cSHuang Ying */ 61731cb9dc4bSLiu Shixin static inline int process_huge_page( 6174c6ddfb6cSHuang Ying unsigned long addr_hint, unsigned int pages_per_huge_page, 61751cb9dc4bSLiu Shixin int (*process_subpage)(unsigned long addr, int idx, void *arg), 6176c6ddfb6cSHuang Ying void *arg) 6177c6ddfb6cSHuang Ying { 61781cb9dc4bSLiu Shixin int i, n, base, l, ret; 6179c6ddfb6cSHuang Ying unsigned long addr = addr_hint & 6180c6ddfb6cSHuang Ying ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 6181c6ddfb6cSHuang Ying 6182c6ddfb6cSHuang Ying /* Process target subpage last to keep its cache lines hot */ 6183c6ddfb6cSHuang Ying might_sleep(); 6184c6ddfb6cSHuang Ying n = (addr_hint - addr) / PAGE_SIZE; 6185c6ddfb6cSHuang Ying if (2 * n <= pages_per_huge_page) { 6186c6ddfb6cSHuang Ying /* If target subpage in first half of huge page */ 6187c6ddfb6cSHuang Ying base = 0; 6188c6ddfb6cSHuang Ying l = n; 6189c6ddfb6cSHuang Ying /* Process subpages at the end of huge page */ 6190c6ddfb6cSHuang Ying for (i = pages_per_huge_page - 1; i >= 2 * n; i--) { 6191c6ddfb6cSHuang Ying cond_resched(); 61921cb9dc4bSLiu Shixin ret = process_subpage(addr + i * PAGE_SIZE, i, arg); 61931cb9dc4bSLiu Shixin if (ret) 61941cb9dc4bSLiu Shixin return ret; 6195c6ddfb6cSHuang Ying } 6196c6ddfb6cSHuang Ying } else { 6197c6ddfb6cSHuang Ying /* If target subpage in second half of huge page */ 6198c6ddfb6cSHuang Ying base = pages_per_huge_page - 2 * (pages_per_huge_page - n); 6199c6ddfb6cSHuang Ying l = pages_per_huge_page - n; 6200c6ddfb6cSHuang Ying /* Process subpages at the begin of huge page */ 6201c6ddfb6cSHuang Ying for (i = 0; i < base; i++) { 6202c6ddfb6cSHuang Ying cond_resched(); 62031cb9dc4bSLiu Shixin ret = process_subpage(addr + i * PAGE_SIZE, i, arg); 62041cb9dc4bSLiu Shixin if (ret) 62051cb9dc4bSLiu Shixin return ret; 6206c6ddfb6cSHuang Ying } 6207c6ddfb6cSHuang Ying } 6208c6ddfb6cSHuang Ying /* 6209c6ddfb6cSHuang Ying * Process remaining subpages in left-right-left-right pattern 6210c6ddfb6cSHuang Ying * towards the target subpage 6211c6ddfb6cSHuang Ying */ 6212c6ddfb6cSHuang Ying for (i = 0; i < l; i++) { 6213c6ddfb6cSHuang Ying int left_idx = base + i; 6214c6ddfb6cSHuang Ying int right_idx = base + 2 * l - 1 - i; 6215c6ddfb6cSHuang Ying 6216c6ddfb6cSHuang Ying cond_resched(); 62171cb9dc4bSLiu Shixin ret = process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg); 62181cb9dc4bSLiu Shixin if (ret) 62191cb9dc4bSLiu Shixin return ret; 6220c6ddfb6cSHuang Ying cond_resched(); 62211cb9dc4bSLiu Shixin ret = process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg); 62221cb9dc4bSLiu Shixin if (ret) 62231cb9dc4bSLiu Shixin return ret; 6224c6ddfb6cSHuang Ying } 62251cb9dc4bSLiu Shixin return 0; 6226c6ddfb6cSHuang Ying } 6227c6ddfb6cSHuang Ying 622847ad8475SAndrea Arcangeli static void clear_gigantic_page(struct page *page, 622947ad8475SAndrea Arcangeli unsigned long addr, 623047ad8475SAndrea Arcangeli unsigned int pages_per_huge_page) 623147ad8475SAndrea Arcangeli { 623247ad8475SAndrea Arcangeli int i; 623314455eabSCheng Li struct page *p; 623447ad8475SAndrea Arcangeli 623547ad8475SAndrea Arcangeli might_sleep(); 623614455eabSCheng Li for (i = 0; i < pages_per_huge_page; i++) { 623714455eabSCheng Li p = nth_page(page, i); 623847ad8475SAndrea Arcangeli cond_resched(); 623947ad8475SAndrea Arcangeli clear_user_highpage(p, addr + i * PAGE_SIZE); 624047ad8475SAndrea Arcangeli } 624147ad8475SAndrea Arcangeli } 6242c6ddfb6cSHuang Ying 62431cb9dc4bSLiu Shixin static int clear_subpage(unsigned long addr, int idx, void *arg) 6244c6ddfb6cSHuang Ying { 6245c6ddfb6cSHuang Ying struct page *page = arg; 6246c6ddfb6cSHuang Ying 624721fff064SKefeng Wang clear_user_highpage(nth_page(page, idx), addr); 62481cb9dc4bSLiu Shixin return 0; 6249c6ddfb6cSHuang Ying } 6250c6ddfb6cSHuang Ying 625147ad8475SAndrea Arcangeli void clear_huge_page(struct page *page, 6252c79b57e4SHuang Ying unsigned long addr_hint, unsigned int pages_per_huge_page) 625347ad8475SAndrea Arcangeli { 6254c79b57e4SHuang Ying unsigned long addr = addr_hint & 6255c79b57e4SHuang Ying ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 625647ad8475SAndrea Arcangeli 625747ad8475SAndrea Arcangeli if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { 625847ad8475SAndrea Arcangeli clear_gigantic_page(page, addr, pages_per_huge_page); 625947ad8475SAndrea Arcangeli return; 626047ad8475SAndrea Arcangeli } 626147ad8475SAndrea Arcangeli 6262c6ddfb6cSHuang Ying process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page); 626347ad8475SAndrea Arcangeli } 626447ad8475SAndrea Arcangeli 62651cb9dc4bSLiu Shixin static int copy_user_gigantic_page(struct folio *dst, struct folio *src, 626647ad8475SAndrea Arcangeli unsigned long addr, 626747ad8475SAndrea Arcangeli struct vm_area_struct *vma, 626847ad8475SAndrea Arcangeli unsigned int pages_per_huge_page) 626947ad8475SAndrea Arcangeli { 627047ad8475SAndrea Arcangeli int i; 6271c0e8150eSZhangPeng struct page *dst_page; 6272c0e8150eSZhangPeng struct page *src_page; 627347ad8475SAndrea Arcangeli 627414455eabSCheng Li for (i = 0; i < pages_per_huge_page; i++) { 6275c0e8150eSZhangPeng dst_page = folio_page(dst, i); 6276c0e8150eSZhangPeng src_page = folio_page(src, i); 627714455eabSCheng Li 627847ad8475SAndrea Arcangeli cond_resched(); 62791cb9dc4bSLiu Shixin if (copy_mc_user_highpage(dst_page, src_page, 62801cb9dc4bSLiu Shixin addr + i*PAGE_SIZE, vma)) { 62811cb9dc4bSLiu Shixin memory_failure_queue(page_to_pfn(src_page), 0); 62821cb9dc4bSLiu Shixin return -EHWPOISON; 628347ad8475SAndrea Arcangeli } 628447ad8475SAndrea Arcangeli } 62851cb9dc4bSLiu Shixin return 0; 62861cb9dc4bSLiu Shixin } 628747ad8475SAndrea Arcangeli 6288c9f4cd71SHuang Ying struct copy_subpage_arg { 6289c9f4cd71SHuang Ying struct page *dst; 6290c9f4cd71SHuang Ying struct page *src; 6291c9f4cd71SHuang Ying struct vm_area_struct *vma; 6292c9f4cd71SHuang Ying }; 6293c9f4cd71SHuang Ying 62941cb9dc4bSLiu Shixin static int copy_subpage(unsigned long addr, int idx, void *arg) 6295c9f4cd71SHuang Ying { 6296c9f4cd71SHuang Ying struct copy_subpage_arg *copy_arg = arg; 629721fff064SKefeng Wang struct page *dst = nth_page(copy_arg->dst, idx); 629821fff064SKefeng Wang struct page *src = nth_page(copy_arg->src, idx); 6299c9f4cd71SHuang Ying 630021fff064SKefeng Wang if (copy_mc_user_highpage(dst, src, addr, copy_arg->vma)) { 630121fff064SKefeng Wang memory_failure_queue(page_to_pfn(src), 0); 63021cb9dc4bSLiu Shixin return -EHWPOISON; 63031cb9dc4bSLiu Shixin } 63041cb9dc4bSLiu Shixin return 0; 6305c9f4cd71SHuang Ying } 6306c9f4cd71SHuang Ying 63071cb9dc4bSLiu Shixin int copy_user_large_folio(struct folio *dst, struct folio *src, 6308c0e8150eSZhangPeng unsigned long addr_hint, struct vm_area_struct *vma) 630947ad8475SAndrea Arcangeli { 6310c0e8150eSZhangPeng unsigned int pages_per_huge_page = folio_nr_pages(dst); 6311c9f4cd71SHuang Ying unsigned long addr = addr_hint & 6312c9f4cd71SHuang Ying ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1); 6313c9f4cd71SHuang Ying struct copy_subpage_arg arg = { 6314c0e8150eSZhangPeng .dst = &dst->page, 6315c0e8150eSZhangPeng .src = &src->page, 6316c9f4cd71SHuang Ying .vma = vma, 6317c9f4cd71SHuang Ying }; 631847ad8475SAndrea Arcangeli 63191cb9dc4bSLiu Shixin if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) 63201cb9dc4bSLiu Shixin return copy_user_gigantic_page(dst, src, addr, vma, 632147ad8475SAndrea Arcangeli pages_per_huge_page); 632247ad8475SAndrea Arcangeli 63231cb9dc4bSLiu Shixin return process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg); 632447ad8475SAndrea Arcangeli } 6325fa4d75c1SMike Kravetz 6326e87340caSZhangPeng long copy_folio_from_user(struct folio *dst_folio, 6327fa4d75c1SMike Kravetz const void __user *usr_src, 6328810a56b9SMike Kravetz bool allow_pagefault) 6329fa4d75c1SMike Kravetz { 6330e87340caSZhangPeng void *kaddr; 6331fa4d75c1SMike Kravetz unsigned long i, rc = 0; 6332e87340caSZhangPeng unsigned int nr_pages = folio_nr_pages(dst_folio); 6333e87340caSZhangPeng unsigned long ret_val = nr_pages * PAGE_SIZE; 633414455eabSCheng Li struct page *subpage; 6335fa4d75c1SMike Kravetz 6336e87340caSZhangPeng for (i = 0; i < nr_pages; i++) { 6337e87340caSZhangPeng subpage = folio_page(dst_folio, i); 6338e87340caSZhangPeng kaddr = kmap_local_page(subpage); 63390d508c1fSZhangPeng if (!allow_pagefault) 63400d508c1fSZhangPeng pagefault_disable(); 6341e87340caSZhangPeng rc = copy_from_user(kaddr, usr_src + i * PAGE_SIZE, PAGE_SIZE); 63420d508c1fSZhangPeng if (!allow_pagefault) 63430d508c1fSZhangPeng pagefault_enable(); 6344e87340caSZhangPeng kunmap_local(kaddr); 6345fa4d75c1SMike Kravetz 6346fa4d75c1SMike Kravetz ret_val -= (PAGE_SIZE - rc); 6347fa4d75c1SMike Kravetz if (rc) 6348fa4d75c1SMike Kravetz break; 6349fa4d75c1SMike Kravetz 6350e763243cSMuchun Song flush_dcache_page(subpage); 6351e763243cSMuchun Song 6352fa4d75c1SMike Kravetz cond_resched(); 6353fa4d75c1SMike Kravetz } 6354fa4d75c1SMike Kravetz return ret_val; 6355fa4d75c1SMike Kravetz } 635647ad8475SAndrea Arcangeli #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */ 635749076ec2SKirill A. Shutemov 635840b64acdSOlof Johansson #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS 6359b35f1819SKirill A. Shutemov 6360b35f1819SKirill A. Shutemov static struct kmem_cache *page_ptl_cachep; 6361b35f1819SKirill A. Shutemov 6362b35f1819SKirill A. Shutemov void __init ptlock_cache_init(void) 6363b35f1819SKirill A. Shutemov { 6364b35f1819SKirill A. Shutemov page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0, 6365b35f1819SKirill A. Shutemov SLAB_PANIC, NULL); 6366b35f1819SKirill A. Shutemov } 6367b35f1819SKirill A. Shutemov 6368f5ecca06SVishal Moola (Oracle) bool ptlock_alloc(struct ptdesc *ptdesc) 636949076ec2SKirill A. Shutemov { 637049076ec2SKirill A. Shutemov spinlock_t *ptl; 637149076ec2SKirill A. Shutemov 6372b35f1819SKirill A. Shutemov ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL); 637349076ec2SKirill A. Shutemov if (!ptl) 637449076ec2SKirill A. Shutemov return false; 6375f5ecca06SVishal Moola (Oracle) ptdesc->ptl = ptl; 637649076ec2SKirill A. Shutemov return true; 637749076ec2SKirill A. Shutemov } 637849076ec2SKirill A. Shutemov 63796ed1b8a0SVishal Moola (Oracle) void ptlock_free(struct ptdesc *ptdesc) 638049076ec2SKirill A. Shutemov { 63816ed1b8a0SVishal Moola (Oracle) kmem_cache_free(page_ptl_cachep, ptdesc->ptl); 638249076ec2SKirill A. Shutemov } 638349076ec2SKirill A. Shutemov #endif 6384