LXR linux/mm/mprotect.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  mm/mprotect.c
   4 *
   5 *  (C) Copyright 1994 Linus Torvalds
   6 *  (C) Copyright 2002 Christoph Hellwig
   7 *
   8 *  Address space accounting code       <alan@lxorguk.ukuu.org.uk>
   9 *  (C) Copyright 2002 Red Hat Inc, All Rights Reserved
  10 */
  11
  12#include <linux/mm.h>
  13#include <linux/hugetlb.h>
  14#include <linux/shm.h>
  15#include <linux/mman.h>
  16#include <linux/fs.h>
  17#include <linux/highmem.h>
  18#include <linux/security.h>
  19#include <linux/mempolicy.h>
  20#include <linux/personality.h>
  21#include <linux/syscalls.h>
  22#include <linux/swap.h>
  23#include <linux/swapops.h>
  24#include <linux/mmu_notifier.h>
  25#include <linux/migrate.h>
  26#include <linux/perf_event.h>
  27#include <linux/pkeys.h>
  28#include <linux/ksm.h>
  29#include <linux/uaccess.h>
  30#include <linux/mm_inline.h>
  31#include <asm/pgtable.h>
  32#include <asm/cacheflush.h>
  33#include <asm/mmu_context.h>
  34#include <asm/tlbflush.h>
  35
  36#include "internal.h"
  37
  38static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  39                unsigned long addr, unsigned long end, pgprot_t newprot,
  40                int dirty_accountable, int prot_numa)
  41{
  42        struct mm_struct *mm = vma->vm_mm;
  43        pte_t *pte, oldpte;
  44        spinlock_t *ptl;
  45        unsigned long pages = 0;
  46        int target_node = NUMA_NO_NODE;
  47
  48        /*
  49         * Can be called with only the mmap_sem for reading by
  50         * prot_numa so we must check the pmd isn't constantly
  51         * changing from under us from pmd_none to pmd_trans_huge
  52         * and/or the other way around.
  53         */
  54        if (pmd_trans_unstable(pmd))
  55                return 0;
  56
  57        /*
  58         * The pmd points to a regular pte so the pmd can't change
  59         * from under us even if the mmap_sem is only hold for
  60         * reading.
  61         */
  62        pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
  63
  64        /* Get target node for single threaded private VMAs */
  65        if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
  66            atomic_read(&vma->vm_mm->mm_users) == 1)
  67                target_node = numa_node_id();
  68
  69        flush_tlb_batched_pending(vma->vm_mm);
  70        arch_enter_lazy_mmu_mode();
  71        do {
  72                oldpte = *pte;
  73                if (pte_present(oldpte)) {
  74                        pte_t ptent;
  75                        bool preserve_write = prot_numa && pte_write(oldpte);
  76
  77                        /*
  78                         * Avoid trapping faults against the zero or KSM
  79                         * pages. See similar comment in change_huge_pmd.
  80                         */
  81                        if (prot_numa) {
  82                                struct page *page;
  83
  84                                page = vm_normal_page(vma, addr, oldpte);
  85                                if (!page || PageKsm(page))
  86                                        continue;
  87
  88                                /* Also skip shared copy-on-write pages */
  89                                if (is_cow_mapping(vma->vm_flags) &&
  90                                    page_mapcount(page) != 1)
  91                                        continue;
  92
  93                                /*
  94                                 * While migration can move some dirty pages,
  95                                 * it cannot move them all from MIGRATE_ASYNC
  96                                 * context.
  97                                 */
  98                                if (page_is_file_cache(page) && PageDirty(page))
  99                                        continue;
 100
 101                                /* Avoid TLB flush if possible */
 102                                if (pte_protnone(oldpte))
 103                                        continue;
 104
 105                                /*
 106                                 * Don't mess with PTEs if page is already on the node
 107                                 * a single-threaded process is running on.
 108                                 */
 109                                if (target_node == page_to_nid(page))
 110                                        continue;
 111                        }
 112
 113                        ptent = ptep_modify_prot_start(mm, addr, pte);
 114                        ptent = pte_modify(ptent, newprot);
 115                        if (preserve_write)
 116                                ptent = pte_mk_savedwrite(ptent);
 117
 118                        /* Avoid taking write faults for known dirty pages */
 119                        if (dirty_accountable && pte_dirty(ptent) &&
 120                                        (pte_soft_dirty(ptent) ||
 121                                         !(vma->vm_flags & VM_SOFTDIRTY))) {
 122                                ptent = pte_mkwrite(ptent);
 123                        }
 124                        ptep_modify_prot_commit(mm, addr, pte, ptent);
 125                        pages++;
 126                } else if (IS_ENABLED(CONFIG_MIGRATION)) {
 127                        swp_entry_t entry = pte_to_swp_entry(oldpte);
 128
 129                        if (is_write_migration_entry(entry)) {
 130                                pte_t newpte;
 131                                /*
 132                                 * A protection check is difficult so
 133                                 * just be safe and disable write
 134                                 */
 135                                make_migration_entry_read(&entry);
 136                                newpte = swp_entry_to_pte(entry);
 137                                if (pte_swp_soft_dirty(oldpte))
 138                                        newpte = pte_swp_mksoft_dirty(newpte);
 139                                set_pte_at(mm, addr, pte, newpte);
 140
 141                                pages++;
 142                        }
 143
 144                        if (is_write_device_private_entry(entry)) {
 145                                pte_t newpte;
 146
 147                                /*
 148                                 * We do not preserve soft-dirtiness. See
 149                                 * copy_one_pte() for explanation.
 150                                 */
 151                                make_device_private_entry_read(&entry);
 152                                newpte = swp_entry_to_pte(entry);
 153                                set_pte_at(mm, addr, pte, newpte);
 154
 155                                pages++;
 156                        }
 157                }
 158        } while (pte++, addr += PAGE_SIZE, addr != end);
 159        arch_leave_lazy_mmu_mode();
 160        pte_unmap_unlock(pte - 1, ptl);
 161
 162        return pages;
 163}
 164
 165static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 166                pud_t *pud, unsigned long addr, unsigned long end,
 167                pgprot_t newprot, int dirty_accountable, int prot_numa)
 168{
 169        pmd_t *pmd;
 170        struct mm_struct *mm = vma->vm_mm;
 171        unsigned long next;
 172        unsigned long pages = 0;
 173        unsigned long nr_huge_updates = 0;
 174        unsigned long mni_start = 0;
 175
 176        pmd = pmd_offset(pud, addr);
 177        do {
 178                unsigned long this_pages;
 179
 180                next = pmd_addr_end(addr, end);
 181                if (!is_swap_pmd(*pmd) && !pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)
 182                                && pmd_none_or_clear_bad(pmd))
 183                        goto next;
 184
 185                /* invoke the mmu notifier if the pmd is populated */
 186                if (!mni_start) {
 187                        mni_start = addr;
 188                        mmu_notifier_invalidate_range_start(mm, mni_start, end);
 189                }
 190
 191                if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 192                        if (next - addr != HPAGE_PMD_SIZE) {
 193                                __split_huge_pmd(vma, pmd, addr, false, NULL);
 194                        } else {
 195                                int nr_ptes = change_huge_pmd(vma, pmd, addr,
 196                                                newprot, prot_numa);
 197
 198                                if (nr_ptes) {
 199                                        if (nr_ptes == HPAGE_PMD_NR) {
 200                                                pages += HPAGE_PMD_NR;
 201                                                nr_huge_updates++;
 202                                        }
 203
 204                                        /* huge pmd was handled */
 205                                        goto next;
 206                                }
 207                        }
 208                        /* fall through, the trans huge pmd just split */
 209                }
 210                this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 211                                 dirty_accountable, prot_numa);
 212                pages += this_pages;
 213next:
 214                cond_resched();
 215        } while (pmd++, addr = next, addr != end);
 216
 217        if (mni_start)
 218                mmu_notifier_invalidate_range_end(mm, mni_start, end);
 219
 220        if (nr_huge_updates)
 221                count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 222        return pages;
 223}
 224
 225static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 226                p4d_t *p4d, unsigned long addr, unsigned long end,
 227                pgprot_t newprot, int dirty_accountable, int prot_numa)
 228{
 229        pud_t *pud;
 230        unsigned long next;
 231        unsigned long pages = 0;
 232
 233        pud = pud_offset(p4d, addr);
 234        do {
 235                next = pud_addr_end(addr, end);
 236                if (pud_none_or_clear_bad(pud))
 237                        continue;
 238                pages += change_pmd_range(vma, pud, addr, next, newprot,
 239                                 dirty_accountable, prot_numa);
 240        } while (pud++, addr = next, addr != end);
 241
 242        return pages;
 243}
 244
 245static inline unsigned long change_p4d_range(struct vm_area_struct *vma,
 246                pgd_t *pgd, unsigned long addr, unsigned long end,
 247                pgprot_t newprot, int dirty_accountable, int prot_numa)
 248{
 249        p4d_t *p4d;
 250        unsigned long next;
 251        unsigned long pages = 0;
 252
 253        p4d = p4d_offset(pgd, addr);
 254        do {
 255                next = p4d_addr_end(addr, end);
 256                if (p4d_none_or_clear_bad(p4d))
 257                        continue;
 258                pages += change_pud_range(vma, p4d, addr, next, newprot,
 259                                 dirty_accountable, prot_numa);
 260        } while (p4d++, addr = next, addr != end);
 261
 262        return pages;
 263}
 264
 265static unsigned long change_protection_range(struct vm_area_struct *vma,
 266                unsigned long addr, unsigned long end, pgprot_t newprot,
 267                int dirty_accountable, int prot_numa)
 268{
 269        struct mm_struct *mm = vma->vm_mm;
 270        pgd_t *pgd;
 271        unsigned long next;
 272        unsigned long start = addr;
 273        unsigned long pages = 0;
 274
 275        BUG_ON(addr >= end);
 276        pgd = pgd_offset(mm, addr);
 277        flush_cache_range(vma, addr, end);
 278        inc_tlb_flush_pending(mm);
 279        do {
 280                next = pgd_addr_end(addr, end);
 281                if (pgd_none_or_clear_bad(pgd))
 282                        continue;
 283                pages += change_p4d_range(vma, pgd, addr, next, newprot,
 284                                 dirty_accountable, prot_numa);
 285        } while (pgd++, addr = next, addr != end);
 286
 287        /* Only flush the TLB if we actually modified any entries: */
 288        if (pages)
 289                flush_tlb_range(vma, start, end);
 290        dec_tlb_flush_pending(mm);
 291
 292        return pages;
 293}
 294
 295unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 296                       unsigned long end, pgprot_t newprot,
 297                       int dirty_accountable, int prot_numa)
 298{
 299        unsigned long pages;
 300
 301        if (is_vm_hugetlb_page(vma))
 302                pages = hugetlb_change_protection(vma, start, end, newprot);
 303        else
 304                pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 305
 306        return pages;
 307}
 308
 309int
 310mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 311        unsigned long start, unsigned long end, unsigned long newflags)
 312{
 313        struct mm_struct *mm = vma->vm_mm;
 314        unsigned long oldflags = vma->vm_flags;
 315        long nrpages = (end - start) >> PAGE_SHIFT;
 316        unsigned long charged = 0;
 317        pgoff_t pgoff;
 318        int error;
 319        int dirty_accountable = 0;
 320
 321        if (newflags == oldflags) {
 322                *pprev = vma;
 323                return 0;
 324        }
 325
 326        /*
 327         * If we make a private mapping writable we increase our commit;
 328         * but (without finer accounting) cannot reduce our commit if we
 329         * make it unwritable again. hugetlb mapping were accounted for
 330         * even if read-only so there is no need to account for them here
 331         */
 332        if (newflags & VM_WRITE) {
 333                /* Check space limits when area turns into data. */
 334                if (!may_expand_vm(mm, newflags, nrpages) &&
 335                                may_expand_vm(mm, oldflags, nrpages))
 336                        return -ENOMEM;
 337                if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 338                                                VM_SHARED|VM_NORESERVE))) {
 339                        charged = nrpages;
 340                        if (security_vm_enough_memory_mm(mm, charged))
 341                                return -ENOMEM;
 342                        newflags |= VM_ACCOUNT;
 343                }
 344        }
 345
 346        /*
 347         * First try to merge with previous and/or next vma.
 348         */
 349        pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 350        *pprev = vma_merge(mm, *pprev, start, end, newflags,
 351                           vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 352                           vma->vm_userfaultfd_ctx);
 353        if (*pprev) {
 354                vma = *pprev;
 355                VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 356                goto success;
 357        }
 358
 359        *pprev = vma;
 360
 361        if (start != vma->vm_start) {
 362                error = split_vma(mm, vma, start, 1);
 363                if (error)
 364                        goto fail;
 365        }
 366
 367        if (end != vma->vm_end) {
 368                error = split_vma(mm, vma, end, 0);
 369                if (error)
 370                        goto fail;
 371        }
 372
 373success:
 374        /*
 375         * vm_flags and vm_page_prot are protected by the mmap_sem
 376         * held in write mode.
 377         */
 378        vma->vm_flags = newflags;
 379        dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
 380        vma_set_page_prot(vma);
 381
 382        change_protection(vma, start, end, vma->vm_page_prot,
 383                          dirty_accountable, 0);
 384
 385        /*
 386         * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 387         * fault on access.
 388         */
 389        if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 390                        (newflags & VM_WRITE)) {
 391                populate_vma_page_range(vma, start, end, NULL);
 392        }
 393
 394        vm_stat_account(mm, oldflags, -nrpages);
 395        vm_stat_account(mm, newflags, nrpages);
 396        perf_event_mmap(vma);
 397        return 0;
 398
 399fail:
 400        vm_unacct_memory(charged);
 401        return error;
 402}
 403
 404/*
 405 * pkey==-1 when doing a legacy mprotect()
 406 */
 407static int do_mprotect_pkey(unsigned long start, size_t len,
 408                unsigned long prot, int pkey)
 409{
 410        unsigned long nstart, end, tmp, reqprot;
 411        struct vm_area_struct *vma, *prev;
 412        int error = -EINVAL;
 413        const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 414        const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 415                                (prot & PROT_READ);
 416
 417        prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 418        if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 419                return -EINVAL;
 420
 421        if (start & ~PAGE_MASK)
 422                return -EINVAL;
 423        if (!len)
 424                return 0;
 425        len = PAGE_ALIGN(len);
 426        end = start + len;
 427        if (end <= start)
 428                return -ENOMEM;
 429        if (!arch_validate_prot(prot, start))
 430                return -EINVAL;
 431
 432        reqprot = prot;
 433
 434        if (down_write_killable(&current->mm->mmap_sem))
 435                return -EINTR;
 436
 437        /*
 438         * If userspace did not allocate the pkey, do not let
 439         * them use it here.
 440         */
 441        error = -EINVAL;
 442        if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 443                goto out;
 444
 445        vma = find_vma(current->mm, start);
 446        error = -ENOMEM;
 447        if (!vma)
 448                goto out;
 449        prev = vma->vm_prev;
 450        if (unlikely(grows & PROT_GROWSDOWN)) {
 451                if (vma->vm_start >= end)
 452                        goto out;
 453                start = vma->vm_start;
 454                error = -EINVAL;
 455                if (!(vma->vm_flags & VM_GROWSDOWN))
 456                        goto out;
 457        } else {
 458                if (vma->vm_start > start)
 459                        goto out;
 460                if (unlikely(grows & PROT_GROWSUP)) {
 461                        end = vma->vm_end;
 462                        error = -EINVAL;
 463                        if (!(vma->vm_flags & VM_GROWSUP))
 464                                goto out;
 465                }
 466        }
 467        if (start > vma->vm_start)
 468                prev = vma;
 469
 470        for (nstart = start ; ; ) {
 471                unsigned long mask_off_old_flags;
 472                unsigned long newflags;
 473                int new_vma_pkey;
 474
 475                /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 476
 477                /* Does the application expect PROT_READ to imply PROT_EXEC */
 478                if (rier && (vma->vm_flags & VM_MAYEXEC))
 479                        prot |= PROT_EXEC;
 480
 481                /*
 482                 * Each mprotect() call explicitly passes r/w/x permissions.
 483                 * If a permission is not passed to mprotect(), it must be
 484                 * cleared from the VMA.
 485                 */
 486                mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 487                                        VM_FLAGS_CLEAR;
 488
 489                new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 490                newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 491                newflags |= (vma->vm_flags & ~mask_off_old_flags);
 492
 493                /* newflags >> 4 shift VM_MAY% in place of VM_% */
 494                if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 495                        error = -EACCES;
 496                        goto out;
 497                }
 498
 499                error = security_file_mprotect(vma, reqprot, prot);
 500                if (error)
 501                        goto out;
 502
 503                tmp = vma->vm_end;
 504                if (tmp > end)
 505                        tmp = end;
 506                error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 507                if (error)
 508                        goto out;
 509                nstart = tmp;
 510
 511                if (nstart < prev->vm_end)
 512                        nstart = prev->vm_end;
 513                if (nstart >= end)
 514                        goto out;
 515
 516                vma = prev->vm_next;
 517                if (!vma || vma->vm_start != nstart) {
 518                        error = -ENOMEM;
 519                        goto out;
 520                }
 521                prot = reqprot;
 522        }
 523out:
 524        up_write(&current->mm->mmap_sem);
 525        return error;
 526}
 527
 528SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 529                unsigned long, prot)
 530{
 531        return do_mprotect_pkey(start, len, prot, -1);
 532}
 533
 534#ifdef CONFIG_ARCH_HAS_PKEYS
 535
 536SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 537                unsigned long, prot, int, pkey)
 538{
 539        return do_mprotect_pkey(start, len, prot, pkey);
 540}
 541
 542SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 543{
 544        int pkey;
 545        int ret;
 546
 547        /* No flags supported yet. */
 548        if (flags)
 549                return -EINVAL;
 550        /* check for unsupported init values */
 551        if (init_val & ~PKEY_ACCESS_MASK)
 552                return -EINVAL;
 553
 554        down_write(&current->mm->mmap_sem);
 555        pkey = mm_pkey_alloc(current->mm);
 556
 557        ret = -ENOSPC;
 558        if (pkey == -1)
 559                goto out;
 560
 561        ret = arch_set_user_pkey_access(current, pkey, init_val);
 562        if (ret) {
 563                mm_pkey_free(current->mm, pkey);
 564                goto out;
 565        }
 566        ret = pkey;
 567out:
 568        up_write(&current->mm->mmap_sem);
 569        return ret;
 570}
 571
 572SYSCALL_DEFINE1(pkey_free, int, pkey)
 573{
 574        int ret;
 575
 576        down_write(&current->mm->mmap_sem);
 577        ret = mm_pkey_free(current->mm, pkey);
 578        up_write(&current->mm->mmap_sem);
 579
 580        /*
 581         * We could provie warnings or errors if any VMA still
 582         * has the pkey set here.
 583         */
 584        return ret;
 585}
 586
 587#endif /* CONFIG_ARCH_HAS_PKEYS */
 588