fs/exec.c

   1 /*
   2  *  linux/fs/exec.c
   3  *
   4  *  Copyright (C) 1991, 1992  Linus Torvalds
   5  */
   6
   7 /*
   8  * #!-checking implemented by tytso.
   9  */
  10 /*
  11  * Demand-loading implemented 01.12.91 - no need to read anything but
  12  * the header into memory. The inode of the executable is put into
  13  * "current->executable", and page faults do the actual loading. Clean.
  14  *
  15  * Once more I can proudly say that linux stood up to being changed: it
  16  * was less than 2 hours work to get demand-loading completely implemented.
  17  *
  18  * Demand loading changed July 1993 by Eric Youngdale.   Use mmap instead,
  19  * current->executable is only used by the procfs.  This allows a dispatch
  20  * table to check for several different types  of binary formats.  We keep
  21  * trying until we recognize the file or we run out of supported binary
  22  * formats.
  23  */
  24
  25 #include <linux/slab.h>
  26 #include <linux/file.h>
  27 #include <linux/fdtable.h>
  28 #include <linux/mm.h>
  29 #include <linux/vmacache.h>
  30 #include <linux/stat.h>
  31 #include <linux/fcntl.h>
  32 #include <linux/swap.h>
  33 #include <linux/string.h>
  34 #include <linux/init.h>
  35 #include <linux/sched/mm.h>
  36 #include <linux/sched/coredump.h>
  37 #include <linux/sched/signal.h>
  38 #include <linux/sched/numa_balancing.h>
  39 #include <linux/sched/task.h>
  40 #include <linux/pagemap.h>
  41 #include <linux/perf_event.h>
  42 #include <linux/highmem.h>
  43 #include <linux/spinlock.h>
  44 #include <linux/key.h>
  45 #include <linux/personality.h>
  46 #include <linux/binfmts.h>
  47 #include <linux/utsname.h>
  48 #include <linux/pid_namespace.h>
  49 #include <linux/module.h>
  50 #include <linux/namei.h>
  51 #include <linux/mount.h>
  52 #include <linux/security.h>
  53 #include <linux/syscalls.h>
  54 #include <linux/tsacct_kern.h>
  55 #include <linux/cn_proc.h>
  56 #include <linux/audit.h>
  57 #include <linux/tracehook.h>
  58 #include <linux/kmod.h>
  59 #include <linux/fsnotify.h>
  60 #include <linux/fs_struct.h>
  61 #include <linux/pipe_fs_i.h>
  62 #include <linux/oom.h>
  63 #include <linux/compat.h>
  64 #include <linux/vmalloc.h>
  65
  66 #include <linux/uaccess.h>
  67 #include <asm/mmu_context.h>
  68 #include <asm/tlb.h>
  69
  70 #include <trace/events/task.h>
  71 #include "internal.h"
  72
  73 #include <trace/events/sched.h>
  74
  75 int suid_dumpable = 0;
  76
  77 static LIST_HEAD(formats);
  78 static DEFINE_RWLOCK(binfmt_lock);
  79
  80 void __register_binfmt(struct linux_binfmt * fmt, int insert)
  81 {
  82         BUG_ON(!fmt);
  83         if (WARN_ON(!fmt->load_binary))
  84                 return;
  85         write_lock(&binfmt_lock);
  86         insert ? list_add(&fmt->lh, &formats) :
  87                  list_add_tail(&fmt->lh, &formats);
  88         write_unlock(&binfmt_lock);
  89 }
  90
  91 EXPORT_SYMBOL(__register_binfmt);
  92
  93 void unregister_binfmt(struct linux_binfmt * fmt)
  94 {
  95         write_lock(&binfmt_lock);
  96         list_del(&fmt->lh);
  97         write_unlock(&binfmt_lock);
  98 }
  99
 100 EXPORT_SYMBOL(unregister_binfmt);
 101
 102 static inline void put_binfmt(struct linux_binfmt * fmt)
 103 {
 104         module_put(fmt->module);
 105 }
 106
 107 bool path_noexec(const struct path *path)
 108 {
 109         return (path->mnt->mnt_flags & MNT_NOEXEC) ||
 110                (path->mnt->mnt_sb->s_iflags & SB_I_NOEXEC);
 111 }
 112
 113 #ifdef CONFIG_USELIB
 114 /*
 115  * Note that a shared library must be both readable and executable due to
 116  * security reasons.
 117  *
 118  * Also note that we take the address to load from from the file itself.
 119  */
 120 SYSCALL_DEFINE1(uselib, const char __user *, library)
 121 {
 122         struct linux_binfmt *fmt;
 123         struct file *file;
 124         struct filename *tmp = getname(library);
 125         int error = PTR_ERR(tmp);
 126         static const struct open_flags uselib_flags = {
 127                 .open_flag = O_LARGEFILE | O_RDONLY | __FMODE_EXEC,
 128                 .acc_mode = MAY_READ | MAY_EXEC,
 129                 .intent = LOOKUP_OPEN,
 130                 .lookup_flags = LOOKUP_FOLLOW,
 131         };
 132
 133         if (IS_ERR(tmp))
 134                 goto out;
 135
 136         file = do_filp_open(AT_FDCWD, tmp, &uselib_flags);
 137         putname(tmp);
 138         error = PTR_ERR(file);
 139         if (IS_ERR(file))
 140                 goto out;
 141
 142         error = -EINVAL;
 143         if (!S_ISREG(file_inode(file)->i_mode))
 144                 goto exit;
 145
 146         error = -EACCES;
 147         if (path_noexec(&file->f_path))
 148                 goto exit;
 149
 150         fsnotify_open(file);
 151
 152         error = -ENOEXEC;
 153
 154         read_lock(&binfmt_lock);
 155         list_for_each_entry(fmt, &formats, lh) {
 156                 if (!fmt->load_shlib)
 157                         continue;
 158                 if (!try_module_get(fmt->module))
 159                         continue;
 160                 read_unlock(&binfmt_lock);
 161                 error = fmt->load_shlib(file);
 162                 read_lock(&binfmt_lock);
 163                 put_binfmt(fmt);
 164                 if (error != -ENOEXEC)
 165                         break;
 166         }
 167         read_unlock(&binfmt_lock);
 168 exit:
 169         fput(file);
 170 out:
 171         return error;
 172 }
 173 #endif /* #ifdef CONFIG_USELIB */
 174
 175 #ifdef CONFIG_MMU
 176 /*
 177  * The nascent bprm->mm is not visible until exec_mmap() but it can
 178  * use a lot of memory, account these pages in current->mm temporary
 179  * for oom_badness()->get_mm_rss(). Once exec succeeds or fails, we
 180  * change the counter back via acct_arg_size(0).
 181  */
 182 static void acct_arg_size(struct linux_binprm *bprm, unsigned long pages)
 183 {
 184         struct mm_struct *mm = current->mm;
 185         long diff = (long)(pages - bprm->vma_pages);
 186
 187         if (!mm || !diff)
 188                 return;
 189
 190         bprm->vma_pages = pages;
 191         add_mm_counter(mm, MM_ANONPAGES, diff);
 192 }
 193
 194 static struct page *get_arg_page(struct linux_binprm *bprm, unsigned long pos,
 195                 int write)
 196 {
 197         struct page *page;
 198         int ret;
 199         unsigned int gup_flags = FOLL_FORCE;
 200
 201 #ifdef CONFIG_STACK_GROWSUP
 202         if (write) {
 203                 ret = expand_downwards(bprm->vma, pos);
 204                 if (ret < 0)
 205                         return NULL;
 206         }
 207 #endif
 208
 209         if (write)
 210                 gup_flags |= FOLL_WRITE;
 211
 212         /*
 213          * We are doing an exec().  'current' is the process
 214          * doing the exec and bprm->mm is the new process's mm.
 215          */
 216         ret = get_user_pages_remote(current, bprm->mm, pos, 1, gup_flags,
 217                         &page, NULL, NULL);
 218         if (ret <= 0)
 219                 return NULL;
 220
 221         if (write) {
 222                 unsigned long size = bprm->vma->vm_end - bprm->vma->vm_start;
 223                 unsigned long ptr_size, limit;
 224
 225                 /*
 226                  * Since the stack will hold pointers to the strings, we
 227                  * must account for them as well.
 228                  *
 229                  * The size calculation is the entire vma while each arg page is
 230                  * built, so each time we get here it's calculating how far it
 231                  * is currently (rather than each call being just the newly
 232                  * added size from the arg page).  As a result, we need to
 233                  * always add the entire size of the pointers, so that on the
 234                  * last call to get_arg_page() we'll actually have the entire
 235                  * correct size.
 236                  */
 237                 ptr_size = (bprm->argc + bprm->envc) * sizeof(void *);
 238                 if (ptr_size > ULONG_MAX - size)
 239                         goto fail;
 240                 size += ptr_size;
 241
 242                 acct_arg_size(bprm, size / PAGE_SIZE);
 243
 244                 /*
 245                  * We've historically supported up to 32 pages (ARG_MAX)
 246                  * of argument strings even with small stacks
 247                  */
 248                 if (size <= ARG_MAX)
 249                         return page;
 250
 251                 /*
 252                  * Limit to 1/4 of the max stack size or 3/4 of _STK_LIM
 253                  * (whichever is smaller) for the argv+env strings.
 254                  * This ensures that:
 255                  *  - the remaining binfmt code will not run out of stack space,
 256                  *  - the program will have a reasonable amount of stack left
 257                  *    to work from.
 258                  */
 259                 limit = _STK_LIM / 4 * 3;
 260                 limit = min(limit, rlimit(RLIMIT_STACK) / 4);
 261                 if (size > limit)
 262                         goto fail;
 263         }
 264
 265         return page;
 266
 267 fail:
 268         put_page(page);
 269         return NULL;
 270 }
 271
 272 static void put_arg_page(struct page *page)
 273 {
 274         put_page(page);
 275 }
 276
 277 static void free_arg_pages(struct linux_binprm *bprm)
 278 {
 279 }
 280
 281 static void flush_arg_page(struct linux_binprm *bprm, unsigned long pos,
 282                 struct page *page)
 283 {
 284         flush_cache_page(bprm->vma, pos, page_to_pfn(page));
 285 }
 286
 287 static int __bprm_mm_init(struct linux_binprm *bprm)
 288 {
 289         int err;
 290         struct vm_area_struct *vma = NULL;
 291         struct mm_struct *mm = bprm->mm;
 292
 293         bprm->vma = vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
 294         if (!vma)
 295                 return -ENOMEM;
 296
 297         if (down_write_killable(&mm->mmap_sem)) {
 298                 err = -EINTR;
 299                 goto err_free;
 300         }
 301         vma->vm_mm = mm;
 302
 303         /*
 304          * Place the stack at the largest stack address the architecture
 305          * supports. Later, we'll move this to an appropriate place. We don't
 306          * use STACK_TOP because that can depend on attributes which aren't
 307          * configured yet.
 308          */
 309         BUILD_BUG_ON(VM_STACK_FLAGS & VM_STACK_INCOMPLETE_SETUP);
 310         vma->vm_end = STACK_TOP_MAX;
 311         vma->vm_start = vma->vm_end - PAGE_SIZE;
 312         vma->vm_flags = VM_SOFTDIRTY | VM_STACK_FLAGS | VM_STACK_INCOMPLETE_SETUP;
 313         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
 314         INIT_LIST_HEAD(&vma->anon_vma_chain);
 315
 316         err = insert_vm_struct(mm, vma);
 317         if (err)
 318                 goto err;
 319
 320         mm->stack_vm = mm->total_vm = 1;
 321         arch_bprm_mm_init(mm, vma);
 322         up_write(&mm->mmap_sem);
 323         bprm->p = vma->vm_end - sizeof(void *);
 324         return 0;
 325 err:
 326         up_write(&mm->mmap_sem);
 327 err_free:
 328         bprm->vma = NULL;
 329         kmem_cache_free(vm_area_cachep, vma);
 330         return err;
 331 }
 332
 333 static bool valid_arg_len(struct linux_binprm *bprm, long len)
 334 {
 335         return len <= MAX_ARG_STRLEN;
 336 }
 337
 338 #else
 339
 340 static inline void acct_arg_size(struct linux_binprm *bprm, unsigned long pages)
 341 {
 342 }
 343
 344 static struct page *get_arg_page(struct linux_binprm *bprm, unsigned long pos,
 345                 int write)
 346 {
 347         struct page *page;
 348
 349         page = bprm->page[pos / PAGE_SIZE];
 350         if (!page && write) {
 351                 page = alloc_page(GFP_HIGHUSER|__GFP_ZERO);
 352                 if (!page)
 353                         return NULL;
 354                 bprm->page[pos / PAGE_SIZE] = page;
 355         }
 356
 357         return page;
 358 }
 359
 360 static void put_arg_page(struct page *page)
 361 {
 362 }
 363
 364 static void free_arg_page(struct linux_binprm *bprm, int i)
 365 {
 366         if (bprm->page[i]) {
 367                 __free_page(bprm->page[i]);
 368                 bprm->page[i] = NULL;
 369         }
 370 }
 371
 372 static void free_arg_pages(struct linux_binprm *bprm)
 373 {
 374         int i;
 375
 376         for (i = 0; i < MAX_ARG_PAGES; i++)
 377                 free_arg_page(bprm, i);
 378 }
 379
 380 static void flush_arg_page(struct linux_binprm *bprm, unsigned long pos,
 381                 struct page *page)
 382 {
 383 }
 384
 385 static int __bprm_mm_init(struct linux_binprm *bprm)
 386 {
 387         bprm->p = PAGE_SIZE * MAX_ARG_PAGES - sizeof(void *);
 388         return 0;
 389 }
 390
 391 static bool valid_arg_len(struct linux_binprm *bprm, long len)
 392 {
 393         return len <= bprm->p;
 394 }
 395
 396 #endif /* CONFIG_MMU */
 397
 398 /*
 399  * Create a new mm_struct and populate it with a temporary stack
 400  * vm_area_struct.  We don't have enough context at this point to set the stack
 401  * flags, permissions, and offset, so we use temporary values.  We'll update
 402  * them later in setup_arg_pages().
 403  */
 404 static int bprm_mm_init(struct linux_binprm *bprm)
 405 {
 406         int err;
 407         struct mm_struct *mm = NULL;
 408
 409         bprm->mm = mm = mm_alloc();
 410         err = -ENOMEM;
 411         if (!mm)
 412                 goto err;
 413
 414         err = __bprm_mm_init(bprm);
 415         if (err)
 416                 goto err;
 417
 418         return 0;
 419
 420 err:
 421         if (mm) {
 422                 bprm->mm = NULL;
 423                 mmdrop(mm);
 424         }
 425
 426         return err;
 427 }
 428
 429 struct user_arg_ptr {
 430 #ifdef CONFIG_COMPAT
 431         bool is_compat;
 432 #endif
 433         union {
 434                 const char __user *const __user *native;
 435 #ifdef CONFIG_COMPAT
 436                 const compat_uptr_t __user *compat;
 437 #endif
 438         } ptr;
 439 };
 440
 441 static const char __user *get_user_arg_ptr(struct user_arg_ptr argv, int nr)
 442 {
 443         const char __user *native;
 444
 445 #ifdef CONFIG_COMPAT
 446         if (unlikely(argv.is_compat)) {
 447                 compat_uptr_t compat;
 448
 449                 if (get_user(compat, argv.ptr.compat + nr))
 450                         return ERR_PTR(-EFAULT);
 451
 452                 return compat_ptr(compat);
 453         }
 454 #endif
 455
 456         if (get_user(native, argv.ptr.native + nr))
 457                 return ERR_PTR(-EFAULT);
 458
 459         return native;
 460 }
 461
 462 /*
 463  * count() counts the number of strings in array ARGV.
 464  */
 465 static int count(struct user_arg_ptr argv, int max)
 466 {
 467         int i = 0;
 468
 469         if (argv.ptr.native != NULL) {
 470                 for (;;) {
 471                         const char __user *p = get_user_arg_ptr(argv, i);
 472
 473                         if (!p)
 474                                 break;
 475
 476                         if (IS_ERR(p))
 477                                 return -EFAULT;
 478
 479                         if (i >= max)
 480                                 return -E2BIG;
 481                         ++i;
 482
 483                         if (fatal_signal_pending(current))
 484                                 return -ERESTARTNOHAND;
 485                         cond_resched();
 486                 }
 487         }
 488         return i;
 489 }
 490
 491 /*
 492  * 'copy_strings()' copies argument/environment strings from the old
 493  * processes's memory to the new process's stack.  The call to get_user_pages()
 494  * ensures the destination page is created and not swapped out.
 495  */
 496 static int copy_strings(int argc, struct user_arg_ptr argv,
 497                         struct linux_binprm *bprm)
 498 {
 499         struct page *kmapped_page = NULL;
 500         char *kaddr = NULL;
 501         unsigned long kpos = 0;
 502         int ret;
 503
 504         while (argc-- > 0) {
 505                 const char __user *str;
 506                 int len;
 507                 unsigned long pos;
 508
 509                 ret = -EFAULT;
 510                 str = get_user_arg_ptr(argv, argc);
 511                 if (IS_ERR(str))
 512                         goto out;
 513
 514                 len = strnlen_user(str, MAX_ARG_STRLEN);
 515                 if (!len)
 516                         goto out;
 517
 518                 ret = -E2BIG;
 519                 if (!valid_arg_len(bprm, len))
 520                         goto out;
 521
 522                 /* We're going to work our way backwords. */
 523                 pos = bprm->p;
 524                 str += len;
 525                 bprm->p -= len;
 526
 527                 while (len > 0) {
 528                         int offset, bytes_to_copy;
 529
 530                         if (fatal_signal_pending(current)) {
 531                                 ret = -ERESTARTNOHAND;
 532                                 goto out;
 533                         }
 534                         cond_resched();
 535
 536                         offset = pos % PAGE_SIZE;
 537                         if (offset == 0)
 538                                 offset = PAGE_SIZE;
 539
 540                         bytes_to_copy = offset;
 541                         if (bytes_to_copy > len)
 542                                 bytes_to_copy = len;
 543
 544                         offset -= bytes_to_copy;
 545                         pos -= bytes_to_copy;
 546                         str -= bytes_to_copy;
 547                         len -= bytes_to_copy;
 548
 549                         if (!kmapped_page || kpos != (pos & PAGE_MASK)) {
 550                                 struct page *page;
 551
 552                                 page = get_arg_page(bprm, pos, 1);
 553                                 if (!page) {
 554                                         ret = -E2BIG;
 555                                         goto out;
 556                                 }
 557
 558                                 if (kmapped_page) {
 559                                         flush_kernel_dcache_page(kmapped_page);
 560                                         kunmap(kmapped_page);
 561                                         put_arg_page(kmapped_page);
 562                                 }
 563                                 kmapped_page = page;
 564                                 kaddr = kmap(kmapped_page);
 565                                 kpos = pos & PAGE_MASK;
 566                                 flush_arg_page(bprm, kpos, kmapped_page);
 567                         }
 568                         if (copy_from_user(kaddr+offset, str, bytes_to_copy)) {
 569                                 ret = -EFAULT;
 570                                 goto out;
 571                         }
 572                 }
 573         }
 574         ret = 0;
 575 out:
 576         if (kmapped_page) {
 577                 flush_kernel_dcache_page(kmapped_page);
 578                 kunmap(kmapped_page);
 579                 put_arg_page(kmapped_page);
 580         }
 581         return ret;
 582 }
 583
 584 /*
 585  * Like copy_strings, but get argv and its values from kernel memory.
 586  */
 587 int copy_strings_kernel(int argc, const char *const *__argv,
 588                         struct linux_binprm *bprm)
 589 {
 590         int r;
 591         mm_segment_t oldfs = get_fs();
 592         struct user_arg_ptr argv = {
 593                 .ptr.native = (const char __user *const  __user *)__argv,
 594         };
 595
 596         set_fs(KERNEL_DS);
 597         r = copy_strings(argc, argv, bprm);
 598         set_fs(oldfs);
 599
 600         return r;
 601 }
 602 EXPORT_SYMBOL(copy_strings_kernel);
 603
 604 #ifdef CONFIG_MMU
 605
 606 /*
 607  * During bprm_mm_init(), we create a temporary stack at STACK_TOP_MAX.  Once
 608  * the binfmt code determines where the new stack should reside, we shift it to
 609  * its final location.  The process proceeds as follows:
 610  *
 611  * 1) Use shift to calculate the new vma endpoints.
 612  * 2) Extend vma to cover both the old and new ranges.  This ensures the
 613  *    arguments passed to subsequent functions are consistent.
 614  * 3) Move vma's page tables to the new range.
 615  * 4) Free up any cleared pgd range.
 616  * 5) Shrink the vma to cover only the new range.
 617  */
 618 static int shift_arg_pages(struct vm_area_struct *vma, unsigned long shift)
 619 {
 620         struct mm_struct *mm = vma->vm_mm;
 621         unsigned long old_start = vma->vm_start;
 622         unsigned long old_end = vma->vm_end;
 623         unsigned long length = old_end - old_start;
 624         unsigned long new_start = old_start - shift;
 625         unsigned long new_end = old_end - shift;
 626         struct mmu_gather tlb;
 627
 628         BUG_ON(new_start > new_end);
 629
 630         /*
 631          * ensure there are no vmas between where we want to go
 632          * and where we are
 633          */
 634         if (vma != find_vma(mm, new_start))
 635                 return -EFAULT;
 636
 637         /*
 638          * cover the whole range: [new_start, old_end)
 639          */
 640         if (vma_adjust(vma, new_start, old_end, vma->vm_pgoff, NULL))
 641                 return -ENOMEM;
 642
 643         /*
 644          * move the page tables downwards, on failure we rely on
 645          * process cleanup to remove whatever mess we made.
 646          */
 647         if (length != move_page_tables(vma, old_start,
 648                                        vma, new_start, length, false))
 649                 return -ENOMEM;
 650
 651         lru_add_drain();
 652         tlb_gather_mmu(&tlb, mm, old_start, old_end);
 653         if (new_end > old_start) {
 654                 /*
 655                  * when the old and new regions overlap clear from new_end.
 656                  */
 657                 free_pgd_range(&tlb, new_end, old_end, new_end,
 658                         vma->vm_next ? vma->vm_next->vm_start : USER_PGTABLES_CEILING);
 659         } else {
 660                 /*
 661                  * otherwise, clean from old_start; this is done to not touch
 662                  * the address space in [new_end, old_start) some architectures
 663                  * have constraints on va-space that make this illegal (IA64) -
 664                  * for the others its just a little faster.
 665                  */
 666                 free_pgd_range(&tlb, old_start, old_end, new_end,
 667                         vma->vm_next ? vma->vm_next->vm_start : USER_PGTABLES_CEILING);
 668         }
 669         tlb_finish_mmu(&tlb, old_start, old_end);
 670
 671         /*
 672          * Shrink the vma to just the new range.  Always succeeds.
 673          */
 674         vma_adjust(vma, new_start, new_end, vma->vm_pgoff, NULL);
 675
 676         return 0;
 677 }
 678
 679 /*
 680  * Finalizes the stack vm_area_struct. The flags and permissions are updated,
 681  * the stack is optionally relocated, and some extra space is added.
 682  */
 683 int setup_arg_pages(struct linux_binprm *bprm,
 684                     unsigned long stack_top,
 685                     int executable_stack)
 686 {
 687         unsigned long ret;
 688         unsigned long stack_shift;
 689         struct mm_struct *mm = current->mm;
 690         struct vm_area_struct *vma = bprm->vma;
 691         struct vm_area_struct *prev = NULL;
 692         unsigned long vm_flags;
 693         unsigned long stack_base;
 694         unsigned long stack_size;
 695         unsigned long stack_expand;
 696         unsigned long rlim_stack;
 697
 698 #ifdef CONFIG_STACK_GROWSUP
 699         /* Limit stack size */
 700         stack_base = rlimit_max(RLIMIT_STACK);
 701         if (stack_base > STACK_SIZE_MAX)
 702                 stack_base = STACK_SIZE_MAX;
 703
 704         /* Add space for stack randomization. */
 705         stack_base += (STACK_RND_MASK << PAGE_SHIFT);
 706
 707         /* Make sure we didn't let the argument array grow too large. */
 708         if (vma->vm_end - vma->vm_start > stack_base)
 709                 return -ENOMEM;
 710
 711         stack_base = PAGE_ALIGN(stack_top - stack_base);
 712
 713         stack_shift = vma->vm_start - stack_base;
 714         mm->arg_start = bprm->p - stack_shift;
 715         bprm->p = vma->vm_end - stack_shift;
 716 #else
 717         stack_top = arch_align_stack(stack_top);
 718         stack_top = PAGE_ALIGN(stack_top);
 719
 720         if (unlikely(stack_top < mmap_min_addr) ||
 721             unlikely(vma->vm_end - vma->vm_start >= stack_top - mmap_min_addr))
 722                 return -ENOMEM;
 723
 724         stack_shift = vma->vm_end - stack_top;
 725
 726         bprm->p -= stack_shift;
 727         mm->arg_start = bprm->p;
 728 #endif
 729
 730         if (bprm->loader)
 731                 bprm->loader -= stack_shift;
 732         bprm->exec -= stack_shift;
 733
 734         if (down_write_killable(&mm->mmap_sem))
 735                 return -EINTR;
 736
 737         vm_flags = VM_STACK_FLAGS;
 738
 739         /*
 740          * Adjust stack execute permissions; explicitly enable for
 741          * EXSTACK_ENABLE_X, disable for EXSTACK_DISABLE_X and leave alone
 742          * (arch default) otherwise.
 743          */
 744         if (unlikely(executable_stack == EXSTACK_ENABLE_X))
 745                 vm_flags |= VM_EXEC;
 746         else if (executable_stack == EXSTACK_DISABLE_X)
 747                 vm_flags &= ~VM_EXEC;
 748         vm_flags |= mm->def_flags;
 749         vm_flags |= VM_STACK_INCOMPLETE_SETUP;
 750
 751         ret = mprotect_fixup(vma, &prev, vma->vm_start, vma->vm_end,
 752                         vm_flags);
 753         if (ret)
 754                 goto out_unlock;
 755         BUG_ON(prev != vma);
 756
 757         /* Move stack pages down in memory. */
 758         if (stack_shift) {
 759                 ret = shift_arg_pages(vma, stack_shift);
 760                 if (ret)
 761                         goto out_unlock;
 762         }
 763
 764         /* mprotect_fixup is overkill to remove the temporary stack flags */
 765         vma->vm_flags &= ~VM_STACK_INCOMPLETE_SETUP;
 766
 767         stack_expand = 131072UL; /* randomly 32*4k (or 2*64k) pages */
 768         stack_size = vma->vm_end - vma->vm_start;
 769         /*
 770          * Align this down to a page boundary as expand_stack
 771          * will align it up.
 772          */
 773         rlim_stack = rlimit(RLIMIT_STACK) & PAGE_MASK;
 774 #ifdef CONFIG_STACK_GROWSUP
 775         if (stack_size + stack_expand > rlim_stack)
 776                 stack_base = vma->vm_start + rlim_stack;
 777         else
 778                 stack_base = vma->vm_end + stack_expand;
 779 #else
 780         if (stack_size + stack_expand > rlim_stack)
 781                 stack_base = vma->vm_end - rlim_stack;
 782         else
 783                 stack_base = vma->vm_start - stack_expand;
 784 #endif
 785         current->mm->start_stack = bprm->p;
 786         ret = expand_stack(vma, stack_base);
 787         if (ret)
 788                 ret = -EFAULT;
 789
 790 out_unlock:
 791         up_write(&mm->mmap_sem);
 792         return ret;
 793 }
 794 EXPORT_SYMBOL(setup_arg_pages);
 795
 796 #else
 797
 798 /*
 799  * Transfer the program arguments and environment from the holding pages
 800  * onto the stack. The provided stack pointer is adjusted accordingly.
 801  */
 802 int transfer_args_to_stack(struct linux_binprm *bprm,
 803                            unsigned long *sp_location)
 804 {
 805         unsigned long index, stop, sp;
 806         int ret = 0;
 807
 808         stop = bprm->p >> PAGE_SHIFT;
 809         sp = *sp_location;
 810
 811         for (index = MAX_ARG_PAGES - 1; index >= stop; index--) {
 812                 unsigned int offset = index == stop ? bprm->p & ~PAGE_MASK : 0;
 813                 char *src = kmap(bprm->page[index]) + offset;
 814                 sp -= PAGE_SIZE - offset;
 815                 if (copy_to_user((void *) sp, src, PAGE_SIZE - offset) != 0)
 816                         ret = -EFAULT;
 817                 kunmap(bprm->page[index]);
 818                 if (ret)
 819                         goto out;
 820         }
 821
 822         *sp_location = sp;
 823
 824 out:
 825         return ret;
 826 }
 827 EXPORT_SYMBOL(transfer_args_to_stack);
 828
 829 #endif /* CONFIG_MMU */
 830
 831 static struct file *do_open_execat(int fd, struct filename *name, int flags)
 832 {
 833         struct file *file;
 834         int err;
 835         struct open_flags open_exec_flags = {
 836                 .open_flag = O_LARGEFILE | O_RDONLY | __FMODE_EXEC,
 837                 .acc_mode = MAY_EXEC,
 838                 .intent = LOOKUP_OPEN,
 839                 .lookup_flags = LOOKUP_FOLLOW,
 840         };
 841
 842         if ((flags & ~(AT_SYMLINK_NOFOLLOW | AT_EMPTY_PATH)) != 0)
 843                 return ERR_PTR(-EINVAL);
 844         if (flags & AT_SYMLINK_NOFOLLOW)
 845                 open_exec_flags.lookup_flags &= ~LOOKUP_FOLLOW;
 846         if (flags & AT_EMPTY_PATH)
 847                 open_exec_flags.lookup_flags |= LOOKUP_EMPTY;
 848
 849         file = do_filp_open(fd, name, &open_exec_flags);
 850         if (IS_ERR(file))
 851                 goto out;
 852
 853         err = -EACCES;
 854         if (!S_ISREG(file_inode(file)->i_mode))
 855                 goto exit;
 856
 857         if (path_noexec(&file->f_path))
 858                 goto exit;
 859
 860         err = deny_write_access(file);
 861         if (err)
 862                 goto exit;
 863
 864         if (name->name[0] != '\0')
 865                 fsnotify_open(file);
 866
 867 out:
 868         return file;
 869
 870 exit:
 871         fput(file);
 872         return ERR_PTR(err);
 873 }
 874
 875 struct file *open_exec(const char *name)
 876 {
 877         struct filename *filename = getname_kernel(name);
 878         struct file *f = ERR_CAST(filename);
 879
 880         if (!IS_ERR(filename)) {
 881                 f = do_open_execat(AT_FDCWD, filename, 0);
 882                 putname(filename);
 883         }
 884         return f;
 885 }
 886 EXPORT_SYMBOL(open_exec);
 887
 888 int kernel_read_file(struct file *file, void **buf, loff_t *size,
 889                      loff_t max_size, enum kernel_read_file_id id)
 890 {
 891         loff_t i_size, pos;
 892         ssize_t bytes = 0;
 893         int ret;
 894
 895         if (!S_ISREG(file_inode(file)->i_mode) || max_size < 0)
 896                 return -EINVAL;
 897
 898         ret = security_kernel_read_file(file, id);
 899         if (ret)
 900                 return ret;
 901
 902         ret = deny_write_access(file);
 903         if (ret)
 904                 return ret;
 905
 906         i_size = i_size_read(file_inode(file));
 907         if (max_size > 0 && i_size > max_size) {
 908                 ret = -EFBIG;
 909                 goto out;
 910         }
 911         if (i_size <= 0) {
 912                 ret = -EINVAL;
 913                 goto out;
 914         }
 915
 916         if (id != READING_FIRMWARE_PREALLOC_BUFFER)
 917                 *buf = vmalloc(i_size);
 918         if (!*buf) {
 919                 ret = -ENOMEM;
 920                 goto out;
 921         }
 922
 923         pos = 0;
 924         while (pos < i_size) {
 925                 bytes = kernel_read(file, *buf + pos, i_size - pos, &pos);
 926                 if (bytes < 0) {
 927                         ret = bytes;
 928                         goto out_free;
 929                 }
 930
 931                 if (bytes == 0)
 932                         break;
 933         }
 934
 935         if (pos != i_size) {
 936                 ret = -EIO;
 937                 goto out_free;
 938         }
 939
 940         ret = security_kernel_post_read_file(file, *buf, i_size, id);
 941         if (!ret)
 942                 *size = pos;
 943
 944 out_free:
 945         if (ret < 0) {
 946                 if (id != READING_FIRMWARE_PREALLOC_BUFFER) {
 947                         vfree(*buf);
 948                         *buf = NULL;
 949                 }
 950         }
 951
 952 out:
 953         allow_write_access(file);
 954         return ret;
 955 }
 956 EXPORT_SYMBOL_GPL(kernel_read_file);
 957
 958 int kernel_read_file_from_path(const char *path, void **buf, loff_t *size,
 959                                loff_t max_size, enum kernel_read_file_id id)
 960 {
 961         struct file *file;
 962         int ret;
 963
 964         if (!path || !*path)
 965                 return -EINVAL;
 966
 967         file = filp_open(path, O_RDONLY, 0);
 968         if (IS_ERR(file))
 969                 return PTR_ERR(file);
 970
 971         ret = kernel_read_file(file, buf, size, max_size, id);
 972         fput(file);
 973         return ret;
 974 }
 975 EXPORT_SYMBOL_GPL(kernel_read_file_from_path);
 976
 977 int kernel_read_file_from_fd(int fd, void **buf, loff_t *size, loff_t max_size,
 978                              enum kernel_read_file_id id)
 979 {
 980         struct fd f = fdget(fd);
 981         int ret = -EBADF;
 982
 983         if (!f.file || !(f.file->f_mode & FMODE_READ))
 984                 goto out;
 985
 986         ret = kernel_read_file(f.file, buf, size, max_size, id);
 987 out:
 988         fdput(f);
 989         return ret;
 990 }
 991 EXPORT_SYMBOL_GPL(kernel_read_file_from_fd);
 992
 993 ssize_t read_code(struct file *file, unsigned long addr, loff_t pos, size_t len)
 994 {
 995         ssize_t res = vfs_read(file, (void __user *)addr, len, &pos);
 996         if (res > 0)
 997                 flush_icache_range(addr, addr + len);
 998         return res;
 999 }
1000 EXPORT_SYMBOL(read_code);
1001
1002 static int exec_mmap(struct mm_struct *mm)
1003 {
1004         struct task_struct *tsk;
1005         struct mm_struct *old_mm, *active_mm;
1006
1007         /* Notify parent that we're no longer interested in the old VM */
1008         tsk = current;
1009         old_mm = current->mm;
1010         exec_mm_release(tsk, old_mm);
1011
1012         if (old_mm) {
1013                 sync_mm_rss(old_mm);
1014                 /*
1015                  * Make sure that if there is a core dump in progress
1016                  * for the old mm, we get out and die instead of going
1017                  * through with the exec.  We must hold mmap_sem around
1018                  * checking core_state and changing tsk->mm.
1019                  */
1020                 down_read(&old_mm->mmap_sem);
1021                 if (unlikely(old_mm->core_state)) {
1022                         up_read(&old_mm->mmap_sem);
1023                         return -EINTR;
1024                 }
1025         }
1026         task_lock(tsk);
1027
1028         local_irq_disable();
1029         active_mm = tsk->active_mm;
1030         tsk->active_mm = mm;
1031         tsk->mm = mm;
1032         /*
1033          * This prevents preemption while active_mm is being loaded and
1034          * it and mm are being updated, which could cause problems for
1035          * lazy tlb mm refcounting when these are updated by context
1036          * switches. Not all architectures can handle irqs off over
1037          * activate_mm yet.
1038          */
1039         if (!IS_ENABLED(CONFIG_ARCH_WANT_IRQS_OFF_ACTIVATE_MM))
1040                 local_irq_enable();
1041         activate_mm(active_mm, mm);
1042         if (IS_ENABLED(CONFIG_ARCH_WANT_IRQS_OFF_ACTIVATE_MM))
1043                 local_irq_enable();
1044         tsk->mm->vmacache_seqnum = 0;
1045         vmacache_flush(tsk);
1046         task_unlock(tsk);
1047         if (old_mm) {
1048                 up_read(&old_mm->mmap_sem);
1049                 BUG_ON(active_mm != old_mm);
1050                 setmax_mm_hiwater_rss(&tsk->signal->maxrss, old_mm);
1051                 mm_update_next_owner(old_mm);
1052                 mmput(old_mm);
1053                 return 0;
1054         }
1055         mmdrop(active_mm);
1056         return 0;
1057 }
1058
1059 /*
1060  * This function makes sure the current process has its own signal table,
1061  * so that flush_signal_handlers can later reset the handlers without
1062  * disturbing other processes.  (Other processes might share the signal
1063  * table via the CLONE_SIGHAND option to clone().)
1064  */
1065 static int de_thread(struct task_struct *tsk)
1066 {
1067         struct signal_struct *sig = tsk->signal;
1068         struct sighand_struct *oldsighand = tsk->sighand;
1069         spinlock_t *lock = &oldsighand->siglock;
1070
1071         if (thread_group_empty(tsk))
1072                 goto no_thread_group;
1073
1074         /*
1075          * Kill all other threads in the thread group.
1076          */
1077         spin_lock_irq(lock);
1078         if (signal_group_exit(sig)) {
1079                 /*
1080                  * Another group action in progress, just
1081                  * return so that the signal is processed.
1082                  */
1083                 spin_unlock_irq(lock);
1084                 return -EAGAIN;
1085         }
1086
1087         sig->group_exit_task = tsk;
1088         sig->notify_count = zap_other_threads(tsk);
1089         if (!thread_group_leader(tsk))
1090                 sig->notify_count--;
1091
1092         while (sig->notify_count) {
1093                 __set_current_state(TASK_KILLABLE);
1094                 spin_unlock_irq(lock);
1095                 schedule();
1096                 if (unlikely(__fatal_signal_pending(tsk)))
1097                         goto killed;
1098                 spin_lock_irq(lock);
1099         }
1100         spin_unlock_irq(lock);
1101
1102         /*
1103          * At this point all other threads have exited, all we have to
1104          * do is to wait for the thread group leader to become inactive,
1105          * and to assume its PID:
1106          */
1107         if (!thread_group_leader(tsk)) {
1108                 struct task_struct *leader = tsk->group_leader;
1109
1110                 for (;;) {
1111                         cgroup_threadgroup_change_begin(tsk);
1112                         write_lock_irq(&tasklist_lock);
1113                         /*
1114                          * Do this under tasklist_lock to ensure that
1115                          * exit_notify() can't miss ->group_exit_task
1116                          */
1117                         sig->notify_count = -1;
1118                         if (likely(leader->exit_state))
1119                                 break;
1120                         __set_current_state(TASK_KILLABLE);
1121                         write_unlock_irq(&tasklist_lock);
1122                         cgroup_threadgroup_change_end(tsk);
1123                         schedule();
1124                         if (unlikely(__fatal_signal_pending(tsk)))
1125                                 goto killed;
1126                 }
1127
1128                 /*
1129                  * The only record we have of the real-time age of a
1130                  * process, regardless of execs it's done, is start_time.
1131                  * All the past CPU time is accumulated in signal_struct
1132                  * from sister threads now dead.  But in this non-leader
1133                  * exec, nothing survives from the original leader thread,
1134                  * whose birth marks the true age of this process now.
1135                  * When we take on its identity by switching to its PID, we
1136                  * also take its birthdate (always earlier than our own).
1137                  */
1138                 tsk->start_time = leader->start_time;
1139                 tsk->real_start_time = leader->real_start_time;
1140
1141                 BUG_ON(!same_thread_group(leader, tsk));
1142                 BUG_ON(has_group_leader_pid(tsk));
1143                 /*
1144                  * An exec() starts a new thread group with the
1145                  * TGID of the previous thread group. Rehash the
1146                  * two threads with a switched PID, and release
1147                  * the former thread group leader:
1148                  */
1149
1150                 /* Become a process group leader with the old leader's pid.
1151                  * The old leader becomes a thread of the this thread group.
1152                  * Note: The old leader also uses this pid until release_task
1153                  *       is called.  Odd but simple and correct.
1154                  */
1155                 tsk->pid = leader->pid;
1156                 change_pid(tsk, PIDTYPE_PID, task_pid(leader));
1157                 transfer_pid(leader, tsk, PIDTYPE_PGID);
1158                 transfer_pid(leader, tsk, PIDTYPE_SID);
1159
1160                 list_replace_rcu(&leader->tasks, &tsk->tasks);
1161                 list_replace_init(&leader->sibling, &tsk->sibling);
1162
1163                 tsk->group_leader = tsk;
1164                 leader->group_leader = tsk;
1165
1166                 tsk->exit_signal = SIGCHLD;
1167                 leader->exit_signal = -1;
1168
1169                 BUG_ON(leader->exit_state != EXIT_ZOMBIE);
1170                 leader->exit_state = EXIT_DEAD;
1171
1172                 /*
1173                  * We are going to release_task()->ptrace_unlink() silently,
1174                  * the tracer can sleep in do_wait(). EXIT_DEAD guarantees
1175                  * the tracer wont't block again waiting for this thread.
1176                  */
1177                 if (unlikely(leader->ptrace))
1178                         __wake_up_parent(leader, leader->parent);
1179                 write_unlock_irq(&tasklist_lock);
1180                 cgroup_threadgroup_change_end(tsk);
1181
1182                 release_task(leader);
1183         }
1184
1185         sig->group_exit_task = NULL;
1186         sig->notify_count = 0;
1187
1188 no_thread_group:
1189         /* we have changed execution domain */
1190         tsk->exit_signal = SIGCHLD;
1191
1192 #ifdef CONFIG_POSIX_TIMERS
1193         exit_itimers(sig);
1194         flush_itimer_signals();
1195 #endif
1196
1197         if (atomic_read(&oldsighand->count) != 1) {
1198                 struct sighand_struct *newsighand;
1199                 /*
1200                  * This ->sighand is shared with the CLONE_SIGHAND
1201                  * but not CLONE_THREAD task, switch to the new one.
1202                  */
1203                 newsighand = kmem_cache_alloc(sighand_cachep, GFP_KERNEL);
1204                 if (!newsighand)
1205                         return -ENOMEM;
1206
1207                 atomic_set(&newsighand->count, 1);
1208                 memcpy(newsighand->action, oldsighand->action,
1209                        sizeof(newsighand->action));
1210
1211                 write_lock_irq(&tasklist_lock);
1212                 spin_lock(&oldsighand->siglock);
1213                 rcu_assign_pointer(tsk->sighand, newsighand);
1214                 spin_unlock(&oldsighand->siglock);
1215                 write_unlock_irq(&tasklist_lock);
1216
1217                 __cleanup_sighand(oldsighand);
1218         }
1219
1220         BUG_ON(!thread_group_leader(tsk));
1221         return 0;
1222
1223 killed:
1224         /* protects against exit_notify() and __exit_signal() */
1225         read_lock(&tasklist_lock);
1226         sig->group_exit_task = NULL;
1227         sig->notify_count = 0;
1228         read_unlock(&tasklist_lock);
1229         return -EAGAIN;
1230 }
1231
1232 char *__get_task_comm(char *buf, size_t buf_size, struct task_struct *tsk)
1233 {
1234         task_lock(tsk);
1235         strncpy(buf, tsk->comm, buf_size);
1236         task_unlock(tsk);
1237         return buf;
1238 }
1239 EXPORT_SYMBOL_GPL(__get_task_comm);
1240
1241 /*
1242  * These functions flushes out all traces of the currently running executable
1243  * so that a new one can be started
1244  */
1245
1246 void __set_task_comm(struct task_struct *tsk, const char *buf, bool exec)
1247 {
1248         task_lock(tsk);
1249         trace_task_rename(tsk, buf);
1250         strlcpy(tsk->comm, buf, sizeof(tsk->comm));
1251         task_unlock(tsk);
1252         perf_event_comm(tsk, exec);
1253 }
1254
1255 /*
1256  * Calling this is the point of no return. None of the failures will be
1257  * seen by userspace since either the process is already taking a fatal
1258  * signal (via de_thread() or coredump), or will have SEGV raised
1259  * (after exec_mmap()) by search_binary_handlers (see below).
1260  */
1261 int flush_old_exec(struct linux_binprm * bprm)
1262 {
1263         int retval;
1264
1265         /*
1266          * Make sure we have a private signal table and that
1267          * we are unassociated from the previous thread group.
1268          */
1269         retval = de_thread(current);
1270         if (retval)
1271                 goto out;
1272
1273         /*
1274          * Must be called _before_ exec_mmap() as bprm->mm is
1275          * not visibile until then. This also enables the update
1276          * to be lockless.
1277          */
1278         set_mm_exe_file(bprm->mm, bprm->file);
1279
1280         would_dump(bprm, bprm->file);
1281
1282         /*
1283          * Release all of the old mmap stuff
1284          */
1285         acct_arg_size(bprm, 0);
1286         retval = exec_mmap(bprm->mm);
1287         if (retval)
1288                 goto out;
1289
1290         /*
1291          * After clearing bprm->mm (to mark that current is using the
1292          * prepared mm now), we have nothing left of the original
1293          * process. If anything from here on returns an error, the check
1294          * in search_binary_handler() will SEGV current.
1295          */
1296         bprm->mm = NULL;
1297
1298         set_fs(USER_DS);
1299         current->flags &= ~(PF_RANDOMIZE | PF_FORKNOEXEC | PF_KTHREAD |
1300                                         PF_NOFREEZE | PF_NO_SETAFFINITY);
1301         flush_thread();
1302         current->personality &= ~bprm->per_clear;
1303
1304         /*
1305          * We have to apply CLOEXEC before we change whether the process is
1306          * dumpable (in setup_new_exec) to avoid a race with a process in userspace
1307          * trying to access the should-be-closed file descriptors of a process
1308          * undergoing exec(2).
1309          */
1310         do_close_on_exec(current->files);
1311         return 0;
1312
1313 out:
1314         return retval;
1315 }
1316 EXPORT_SYMBOL(flush_old_exec);
1317
1318 void would_dump(struct linux_binprm *bprm, struct file *file)
1319 {
1320         struct inode *inode = file_inode(file);
1321         if (inode_permission(inode, MAY_READ) < 0) {
1322                 struct user_namespace *old, *user_ns;
1323                 bprm->interp_flags |= BINPRM_FLAGS_ENFORCE_NONDUMP;
1324
1325                 /* Ensure mm->user_ns contains the executable */
1326                 user_ns = old = bprm->mm->user_ns;
1327                 while ((user_ns != &init_user_ns) &&
1328                        !privileged_wrt_inode_uidgid(user_ns, inode))
1329                         user_ns = user_ns->parent;
1330
1331                 if (old != user_ns) {
1332                         bprm->mm->user_ns = get_user_ns(user_ns);
1333                         put_user_ns(old);
1334                 }
1335         }
1336 }
1337 EXPORT_SYMBOL(would_dump);
1338
1339 void setup_new_exec(struct linux_binprm * bprm)
1340 {
1341         /*
1342          * Once here, prepare_binrpm() will not be called any more, so
1343          * the final state of setuid/setgid/fscaps can be merged into the
1344          * secureexec flag.
1345          */
1346         bprm->secureexec |= bprm->cap_elevated;
1347
1348         if (bprm->secureexec) {
1349                 /* Make sure parent cannot signal privileged process. */
1350                 current->pdeath_signal = 0;
1351
1352                 /*
1353                  * For secureexec, reset the stack limit to sane default to
1354                  * avoid bad behavior from the prior rlimits. This has to
1355                  * happen before arch_pick_mmap_layout(), which examines
1356                  * RLIMIT_STACK, but after the point of no return to avoid
1357                  * needing to clean up the change on failure.
1358                  */
1359                 if (current->signal->rlim[RLIMIT_STACK].rlim_cur > _STK_LIM)
1360                         current->signal->rlim[RLIMIT_STACK].rlim_cur = _STK_LIM;
1361         }
1362
1363         arch_pick_mmap_layout(current->mm);
1364
1365         current->sas_ss_sp = current->sas_ss_size = 0;
1366
1367         /*
1368          * Figure out dumpability. Note that this checking only of current
1369          * is wrong, but userspace depends on it. This should be testing
1370          * bprm->secureexec instead.
1371          */
1372         if (bprm->interp_flags & BINPRM_FLAGS_ENFORCE_NONDUMP ||
1373             !(uid_eq(current_euid(), current_uid()) &&
1374               gid_eq(current_egid(), current_gid())))
1375                 set_dumpable(current->mm, suid_dumpable);
1376         else
1377                 set_dumpable(current->mm, SUID_DUMP_USER);
1378
1379         arch_setup_new_exec();
1380         perf_event_exec();
1381         __set_task_comm(current, kbasename(bprm->filename), true);
1382
1383         /* Set the new mm task size. We have to do that late because it may
1384          * depend on TIF_32BIT which is only updated in flush_thread() on
1385          * some architectures like powerpc
1386          */
1387         current->mm->task_size = TASK_SIZE;
1388
1389         /* An exec changes our domain. We are no longer part of the thread
1390            group */
1391         WRITE_ONCE(current->self_exec_id, current->self_exec_id + 1);
1392         flush_signal_handlers(current, 0);
1393 }
1394 EXPORT_SYMBOL(setup_new_exec);
1395
1396 /*
1397  * Prepare credentials and lock ->cred_guard_mutex.
1398  * install_exec_creds() commits the new creds and drops the lock.
1399  * Or, if exec fails before, free_bprm() should release ->cred and
1400  * and unlock.
1401  */
1402 int prepare_bprm_creds(struct linux_binprm *bprm)
1403 {
1404         if (mutex_lock_interruptible(&current->signal->cred_guard_mutex))
1405                 return -ERESTARTNOINTR;
1406
1407         bprm->cred = prepare_exec_creds();
1408         if (likely(bprm->cred))
1409                 return 0;
1410
1411         mutex_unlock(&current->signal->cred_guard_mutex);
1412         return -ENOMEM;
1413 }
1414
1415 static void free_bprm(struct linux_binprm *bprm)
1416 {
1417         free_arg_pages(bprm);
1418         if (bprm->cred) {
1419                 mutex_unlock(&current->signal->cred_guard_mutex);
1420                 abort_creds(bprm->cred);
1421         }
1422         if (bprm->file) {
1423                 allow_write_access(bprm->file);
1424                 fput(bprm->file);
1425         }
1426         /* If a binfmt changed the interp, free it. */
1427         if (bprm->interp != bprm->filename)
1428                 kfree(bprm->interp);
1429         kfree(bprm);
1430 }
1431
1432 int bprm_change_interp(const char *interp, struct linux_binprm *bprm)
1433 {
1434         /* If a binfmt changed the interp, free it first. */
1435         if (bprm->interp != bprm->filename)
1436                 kfree(bprm->interp);
1437         bprm->interp = kstrdup(interp, GFP_KERNEL);
1438         if (!bprm->interp)
1439                 return -ENOMEM;
1440         return 0;
1441 }
1442 EXPORT_SYMBOL(bprm_change_interp);
1443
1444 /*
1445  * install the new credentials for this executable
1446  */
1447 void install_exec_creds(struct linux_binprm *bprm)
1448 {
1449         security_bprm_committing_creds(bprm);
1450
1451         commit_creds(bprm->cred);
1452         bprm->cred = NULL;
1453
1454         /*
1455          * Disable monitoring for regular users
1456          * when executing setuid binaries. Must
1457          * wait until new credentials are committed
1458          * by commit_creds() above
1459          */
1460         if (get_dumpable(current->mm) != SUID_DUMP_USER)
1461                 perf_event_exit_task(current);
1462         /*
1463          * cred_guard_mutex must be held at least to this point to prevent
1464          * ptrace_attach() from altering our determination of the task's
1465          * credentials; any time after this it may be unlocked.
1466          */
1467         security_bprm_committed_creds(bprm);
1468         mutex_unlock(&current->signal->cred_guard_mutex);
1469 }
1470 EXPORT_SYMBOL(install_exec_creds);
1471
1472 /*
1473  * determine how safe it is to execute the proposed program
1474  * - the caller must hold ->cred_guard_mutex to protect against
1475  *   PTRACE_ATTACH or seccomp thread-sync
1476  */
1477 static void check_unsafe_exec(struct linux_binprm *bprm)
1478 {
1479         struct task_struct *p = current, *t;
1480         unsigned n_fs;
1481
1482         if (p->ptrace)
1483                 bprm->unsafe |= LSM_UNSAFE_PTRACE;
1484
1485         /*
1486          * This isn't strictly necessary, but it makes it harder for LSMs to
1487          * mess up.
1488          */
1489         if (task_no_new_privs(current))
1490                 bprm->unsafe |= LSM_UNSAFE_NO_NEW_PRIVS;
1491
1492         t = p;
1493         n_fs = 1;
1494         spin_lock(&p->fs->lock);
1495         rcu_read_lock();
1496         while_each_thread(p, t) {
1497                 if (t->fs == p->fs)
1498                         n_fs++;
1499         }
1500         rcu_read_unlock();
1501
1502         if (p->fs->users > n_fs)
1503                 bprm->unsafe |= LSM_UNSAFE_SHARE;
1504         else
1505                 p->fs->in_exec = 1;
1506         spin_unlock(&p->fs->lock);
1507 }
1508
1509 static void bprm_fill_uid(struct linux_binprm *bprm)
1510 {
1511         struct inode *inode;
1512         unsigned int mode;
1513         kuid_t uid;
1514         kgid_t gid;
1515
1516         /*
1517          * Since this can be called multiple times (via prepare_binprm),
1518          * we must clear any previous work done when setting set[ug]id
1519          * bits from any earlier bprm->file uses (for example when run
1520          * first for a setuid script then again for its interpreter).
1521          */
1522         bprm->cred->euid = current_euid();
1523         bprm->cred->egid = current_egid();
1524
1525         if (!mnt_may_suid(bprm->file->f_path.mnt))
1526                 return;
1527
1528         if (task_no_new_privs(current))
1529                 return;
1530
1531         inode = bprm->file->f_path.dentry->d_inode;
1532         mode = READ_ONCE(inode->i_mode);
1533         if (!(mode & (S_ISUID|S_ISGID)))
1534                 return;
1535
1536         /* Be careful if suid/sgid is set */
1537         inode_lock(inode);
1538
1539         /* reload atomically mode/uid/gid now that lock held */
1540         mode = inode->i_mode;
1541         uid = inode->i_uid;
1542         gid = inode->i_gid;
1543         inode_unlock(inode);
1544
1545         /* We ignore suid/sgid if there are no mappings for them in the ns */
1546         if (!kuid_has_mapping(bprm->cred->user_ns, uid) ||
1547                  !kgid_has_mapping(bprm->cred->user_ns, gid))
1548                 return;
1549
1550         if (mode & S_ISUID) {
1551                 bprm->per_clear |= PER_CLEAR_ON_SETID;
1552                 bprm->cred->euid = uid;
1553         }
1554
1555         if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP)) {
1556                 bprm->per_clear |= PER_CLEAR_ON_SETID;
1557                 bprm->cred->egid = gid;
1558         }
1559 }
1560
1561 /*
1562  * Fill the binprm structure from the inode.
1563  * Check permissions, then read the first 128 (BINPRM_BUF_SIZE) bytes
1564  *
1565  * This may be called multiple times for binary chains (scripts for example).
1566  */
1567 int prepare_binprm(struct linux_binprm *bprm)
1568 {
1569         int retval;
1570         loff_t pos = 0;
1571
1572         bprm_fill_uid(bprm);
1573
1574         /* fill in binprm security blob */
1575         retval = security_bprm_set_creds(bprm);
1576         if (retval)
1577                 return retval;
1578         bprm->called_set_creds = 1;
1579
1580         memset(bprm->buf, 0, BINPRM_BUF_SIZE);
1581         return kernel_read(bprm->file, bprm->buf, BINPRM_BUF_SIZE, &pos);
1582 }
1583
1584 EXPORT_SYMBOL(prepare_binprm);
1585
1586 /*
1587  * Arguments are '\0' separated strings found at the location bprm->p
1588  * points to; chop off the first by relocating brpm->p to right after
1589  * the first '\0' encountered.
1590  */
1591 int remove_arg_zero(struct linux_binprm *bprm)
1592 {
1593         int ret = 0;
1594         unsigned long offset;
1595         char *kaddr;
1596         struct page *page;
1597
1598         if (!bprm->argc)
1599                 return 0;
1600
1601         do {
1602                 offset = bprm->p & ~PAGE_MASK;
1603                 page = get_arg_page(bprm, bprm->p, 0);
1604                 if (!page) {
1605                         ret = -EFAULT;
1606                         goto out;
1607                 }
1608                 kaddr = kmap_atomic(page);
1609
1610                 for (; offset < PAGE_SIZE && kaddr[offset];
1611                                 offset++, bprm->p++)
1612                         ;
1613
1614                 kunmap_atomic(kaddr);
1615                 put_arg_page(page);
1616         } while (offset == PAGE_SIZE);
1617
1618         bprm->p++;
1619         bprm->argc--;
1620         ret = 0;
1621
1622 out:
1623         return ret;
1624 }
1625 EXPORT_SYMBOL(remove_arg_zero);
1626
1627 #define printable(c) (((c)=='\t') || ((c)=='\n') || (0x20<=(c) && (c)<=0x7e))
1628 /*
1629  * cycle the list of binary formats handler, until one recognizes the image
1630  */
1631 int search_binary_handler(struct linux_binprm *bprm)
1632 {
1633         bool need_retry = IS_ENABLED(CONFIG_MODULES);
1634         struct linux_binfmt *fmt;
1635         int retval;
1636
1637         /* This allows 4 levels of binfmt rewrites before failing hard. */
1638         if (bprm->recursion_depth > 5)
1639                 return -ELOOP;
1640
1641         retval = security_bprm_check(bprm);
1642         if (retval)
1643                 return retval;
1644
1645         retval = -ENOENT;
1646  retry:
1647         read_lock(&binfmt_lock);
1648         list_for_each_entry(fmt, &formats, lh) {
1649                 if (!try_module_get(fmt->module))
1650                         continue;
1651                 read_unlock(&binfmt_lock);
1652                 bprm->recursion_depth++;
1653                 retval = fmt->load_binary(bprm);
1654                 read_lock(&binfmt_lock);
1655                 put_binfmt(fmt);
1656                 bprm->recursion_depth--;
1657                 if (retval < 0 && !bprm->mm) {
1658                         /* we got to flush_old_exec() and failed after it */
1659                         read_unlock(&binfmt_lock);
1660                         force_sigsegv(SIGSEGV, current);
1661                         return retval;
1662                 }
1663                 if (retval != -ENOEXEC || !bprm->file) {
1664                         read_unlock(&binfmt_lock);
1665                         return retval;
1666                 }
1667         }
1668         read_unlock(&binfmt_lock);
1669
1670         if (need_retry) {
1671                 if (printable(bprm->buf[0]) && printable(bprm->buf[1]) &&
1672                     printable(bprm->buf[2]) && printable(bprm->buf[3]))
1673                         return retval;
1674                 if (request_module("binfmt-%04x", *(ushort *)(bprm->buf + 2)) < 0)
1675                         return retval;
1676                 need_retry = false;
1677                 goto retry;
1678         }
1679
1680         return retval;
1681 }
1682 EXPORT_SYMBOL(search_binary_handler);
1683
1684 static int exec_binprm(struct linux_binprm *bprm)
1685 {
1686         pid_t old_pid, old_vpid;
1687         int ret;
1688
1689         /* Need to fetch pid before load_binary changes it */
1690         old_pid = current->pid;
1691         rcu_read_lock();
1692         old_vpid = task_pid_nr_ns(current, task_active_pid_ns(current->parent));
1693         rcu_read_unlock();
1694
1695         ret = search_binary_handler(bprm);
1696         if (ret >= 0) {
1697                 audit_bprm(bprm);
1698                 trace_sched_process_exec(current, old_pid, bprm);
1699                 ptrace_event(PTRACE_EVENT_EXEC, old_vpid);
1700                 proc_exec_connector(current);
1701         }
1702
1703         return ret;
1704 }
1705
1706 /*
1707  * sys_execve() executes a new program.
1708  */
1709 static int do_execveat_common(int fd, struct filename *filename,
1710                               struct user_arg_ptr argv,
1711                               struct user_arg_ptr envp,
1712                               int flags)
1713 {
1714         char *pathbuf = NULL;
1715         struct linux_binprm *bprm;
1716         struct file *file;
1717         struct files_struct *displaced;
1718         int retval;
1719
1720         if (IS_ERR(filename))
1721                 return PTR_ERR(filename);
1722
1723         /*
1724          * We move the actual failure in case of RLIMIT_NPROC excess from
1725          * set*uid() to execve() because too many poorly written programs
1726          * don't check setuid() return code.  Here we additionally recheck
1727          * whether NPROC limit is still exceeded.
1728          */
1729         if ((current->flags & PF_NPROC_EXCEEDED) &&
1730             atomic_read(&current_user()->processes) > rlimit(RLIMIT_NPROC)) {
1731                 retval = -EAGAIN;
1732                 goto out_ret;
1733         }
1734
1735         /* We're below the limit (still or again), so we don't want to make
1736          * further execve() calls fail. */
1737         current->flags &= ~PF_NPROC_EXCEEDED;
1738
1739         retval = unshare_files(&displaced);
1740         if (retval)
1741                 goto out_ret;
1742
1743         retval = -ENOMEM;
1744         bprm = kzalloc(sizeof(*bprm), GFP_KERNEL);
1745         if (!bprm)
1746                 goto out_files;
1747
1748         retval = prepare_bprm_creds(bprm);
1749         if (retval)
1750                 goto out_free;
1751
1752         check_unsafe_exec(bprm);
1753         current->in_execve = 1;
1754
1755         file = do_open_execat(fd, filename, flags);
1756         retval = PTR_ERR(file);
1757         if (IS_ERR(file))
1758                 goto out_unmark;
1759
1760         sched_exec();
1761
1762         bprm->file = file;
1763         if (fd == AT_FDCWD || filename->name[0] == '/') {
1764                 bprm->filename = filename->name;
1765         } else {
1766                 if (filename->name[0] == '\0')
1767                         pathbuf = kasprintf(GFP_KERNEL, "/dev/fd/%d", fd);
1768                 else
1769                         pathbuf = kasprintf(GFP_KERNEL, "/dev/fd/%d/%s",
1770                                             fd, filename->name);
1771                 if (!pathbuf) {
1772                         retval = -ENOMEM;
1773                         goto out_unmark;
1774                 }
1775                 /*
1776                  * Record that a name derived from an O_CLOEXEC fd will be
1777                  * inaccessible after exec. Relies on having exclusive access to
1778                  * current->files (due to unshare_files above).
1779                  */
1780                 if (close_on_exec(fd, rcu_dereference_raw(current->files->fdt)))
1781                         bprm->interp_flags |= BINPRM_FLAGS_PATH_INACCESSIBLE;
1782                 bprm->filename = pathbuf;
1783         }
1784         bprm->interp = bprm->filename;
1785
1786         retval = bprm_mm_init(bprm);
1787         if (retval)
1788                 goto out_unmark;
1789
1790         bprm->argc = count(argv, MAX_ARG_STRINGS);
1791         if ((retval = bprm->argc) < 0)
1792                 goto out;
1793
1794         bprm->envc = count(envp, MAX_ARG_STRINGS);
1795         if ((retval = bprm->envc) < 0)
1796                 goto out;
1797
1798         retval = prepare_binprm(bprm);
1799         if (retval < 0)
1800                 goto out;
1801
1802         retval = copy_strings_kernel(1, &bprm->filename, bprm);
1803         if (retval < 0)
1804                 goto out;
1805
1806         bprm->exec = bprm->p;
1807         retval = copy_strings(bprm->envc, envp, bprm);
1808         if (retval < 0)
1809                 goto out;
1810
1811         retval = copy_strings(bprm->argc, argv, bprm);
1812         if (retval < 0)
1813                 goto out;
1814
1815         retval = exec_binprm(bprm);
1816         if (retval < 0)
1817                 goto out;
1818
1819         /* execve succeeded */
1820         current->fs->in_exec = 0;
1821         current->in_execve = 0;
1822         membarrier_execve(current);
1823         acct_update_integrals(current);
1824         task_numa_free(current, false);
1825         free_bprm(bprm);
1826         kfree(pathbuf);
1827         putname(filename);
1828         if (displaced)
1829                 put_files_struct(displaced);
1830         return retval;
1831
1832 out:
1833         if (bprm->mm) {
1834                 acct_arg_size(bprm, 0);
1835                 mmput(bprm->mm);
1836         }
1837
1838 out_unmark:
1839         current->fs->in_exec = 0;
1840         current->in_execve = 0;
1841
1842 out_free:
1843         free_bprm(bprm);
1844         kfree(pathbuf);
1845
1846 out_files:
1847         if (displaced)
1848                 reset_files_struct(displaced);
1849 out_ret:
1850         putname(filename);
1851         return retval;
1852 }
1853
1854 int do_execve(struct filename *filename,
1855         const char __user *const __user *__argv,
1856         const char __user *const __user *__envp)
1857 {
1858         struct user_arg_ptr argv = { .ptr.native = __argv };
1859         struct user_arg_ptr envp = { .ptr.native = __envp };
1860         return do_execveat_common(AT_FDCWD, filename, argv, envp, 0);
1861 }
1862
1863 int do_execveat(int fd, struct filename *filename,
1864                 const char __user *const __user *__argv,
1865                 const char __user *const __user *__envp,
1866                 int flags)
1867 {
1868         struct user_arg_ptr argv = { .ptr.native = __argv };
1869         struct user_arg_ptr envp = { .ptr.native = __envp };
1870
1871         return do_execveat_common(fd, filename, argv, envp, flags);
1872 }
1873
1874 #ifdef CONFIG_COMPAT
1875 static int compat_do_execve(struct filename *filename,
1876         const compat_uptr_t __user *__argv,
1877         const compat_uptr_t __user *__envp)
1878 {
1879         struct user_arg_ptr argv = {
1880                 .is_compat = true,
1881                 .ptr.compat = __argv,
1882         };
1883         struct user_arg_ptr envp = {
1884                 .is_compat = true,
1885                 .ptr.compat = __envp,
1886         };
1887         return do_execveat_common(AT_FDCWD, filename, argv, envp, 0);
1888 }
1889
1890 static int compat_do_execveat(int fd, struct filename *filename,
1891                               const compat_uptr_t __user *__argv,
1892                               const compat_uptr_t __user *__envp,
1893                               int flags)
1894 {
1895         struct user_arg_ptr argv = {
1896                 .is_compat = true,
1897                 .ptr.compat = __argv,
1898         };
1899         struct user_arg_ptr envp = {
1900                 .is_compat = true,
1901                 .ptr.compat = __envp,
1902         };
1903         return do_execveat_common(fd, filename, argv, envp, flags);
1904 }
1905 #endif
1906
1907 void set_binfmt(struct linux_binfmt *new)
1908 {
1909         struct mm_struct *mm = current->mm;
1910
1911         if (mm->binfmt)
1912                 module_put(mm->binfmt->module);
1913
1914         mm->binfmt = new;
1915         if (new)
1916                 __module_get(new->module);
1917 }
1918 EXPORT_SYMBOL(set_binfmt);
1919
1920 /*
1921  * set_dumpable stores three-value SUID_DUMP_* into mm->flags.
1922  */
1923 void set_dumpable(struct mm_struct *mm, int value)
1924 {
1925         unsigned long old, new;
1926
1927         if (WARN_ON((unsigned)value > SUID_DUMP_ROOT))
1928                 return;
1929
1930         do {
1931                 old = ACCESS_ONCE(mm->flags);
1932                 new = (old & ~MMF_DUMPABLE_MASK) | value;
1933         } while (cmpxchg(&mm->flags, old, new) != old);
1934 }
1935
1936 SYSCALL_DEFINE3(execve,
1937                 const char __user *, filename,
1938                 const char __user *const __user *, argv,
1939                 const char __user *const __user *, envp)
1940 {
1941         return do_execve(getname(filename), argv, envp);
1942 }
1943
1944 SYSCALL_DEFINE5(execveat,
1945                 int, fd, const char __user *, filename,
1946                 const char __user *const __user *, argv,
1947                 const char __user *const __user *, envp,
1948                 int, flags)
1949 {
1950         int lookup_flags = (flags & AT_EMPTY_PATH) ? LOOKUP_EMPTY : 0;
1951
1952         return do_execveat(fd,
1953                            getname_flags(filename, lookup_flags, NULL),
1954                            argv, envp, flags);
1955 }
1956
1957 #ifdef CONFIG_COMPAT
1958 COMPAT_SYSCALL_DEFINE3(execve, const char __user *, filename,
1959         const compat_uptr_t __user *, argv,
1960         const compat_uptr_t __user *, envp)
1961 {
1962         return compat_do_execve(getname(filename), argv, envp);
1963 }
1964
1965 COMPAT_SYSCALL_DEFINE5(execveat, int, fd,
1966                        const char __user *, filename,
1967                        const compat_uptr_t __user *, argv,
1968                        const compat_uptr_t __user *, envp,
1969                        int,  flags)
1970 {
1971         int lookup_flags = (flags & AT_EMPTY_PATH) ? LOOKUP_EMPTY : 0;
1972
1973         return compat_do_execveat(fd,
1974                                   getname_flags(filename, lookup_flags, NULL),
1975                                   argv, envp, flags);
1976 }
1977 #endif