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