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 #include <linux/freezer.h>
  66
  67 #include <linux/uaccess.h>
  68 #include <asm/mmu_context.h>
  69 #include <asm/tlb.h>
  70
  71 #include <trace/events/task.h>
  72 #include "internal.h"
  73
  74 #include <trace/events/sched.h>
  75
  76 int suid_dumpable = 0;
  77
  78 static LIST_HEAD(formats);
  79 static DEFINE_RWLOCK(binfmt_lock);
  80
  81 void __register_binfmt(struct linux_binfmt * fmt, int insert)
  82 {
  83         BUG_ON(!fmt);
  84         if (WARN_ON(!fmt->load_binary))
  85                 return;
  86         write_lock(&binfmt_lock);
  87         insert ? list_add(&fmt->lh, &formats) :
  88                  list_add_tail(&fmt->lh, &formats);
  89         write_unlock(&binfmt_lock);
  90 }
  91
  92 EXPORT_SYMBOL(__register_binfmt);
  93
  94 void unregister_binfmt(struct linux_binfmt * fmt)
  95 {
  96         write_lock(&binfmt_lock);
  97         list_del(&fmt->lh);
  98         write_unlock(&binfmt_lock);
  99 }
 100
 101 EXPORT_SYMBOL(unregister_binfmt);
 102
 103 static inline void put_binfmt(struct linux_binfmt * fmt)
 104 {
 105         module_put(fmt->module);
 106 }
 107
 108 bool path_noexec(const struct path *path)
 109 {
 110         return (path->mnt->mnt_flags & MNT_NOEXEC) ||
 111                (path->mnt->mnt_sb->s_iflags & SB_I_NOEXEC);
 112 }
 113
 114 #ifdef CONFIG_USELIB
 115 /*
 116  * Note that a shared library must be both readable and executable due to
 117  * security reasons.
 118  *
 119  * Also note that we take the address to load from from the file itself.
 120  */
 121 SYSCALL_DEFINE1(uselib, const char __user *, library)
 122 {
 123         struct linux_binfmt *fmt;
 124         struct file *file;
 125         struct filename *tmp = getname(library);
 126         int error = PTR_ERR(tmp);
 127         static const struct open_flags uselib_flags = {
 128                 .open_flag = O_LARGEFILE | O_RDONLY | __FMODE_EXEC,
 129                 .acc_mode = MAY_READ | MAY_EXEC,
 130                 .intent = LOOKUP_OPEN,
 131                 .lookup_flags = LOOKUP_FOLLOW,
 132         };
 133
 134         if (IS_ERR(tmp))
 135                 goto out;
 136
 137         file = do_filp_open(AT_FDCWD, tmp, &uselib_flags);
 138         putname(tmp);
 139         error = PTR_ERR(file);
 140         if (IS_ERR(file))
 141                 goto out;
 142
 143         error = -EINVAL;
 144         if (!S_ISREG(file_inode(file)->i_mode))
 145                 goto exit;
 146
 147         error = -EACCES;
 148         if (path_noexec(&file->f_path))
 149                 goto exit;
 150
 151         fsnotify_open(file);
 152
 153         error = -ENOEXEC;
 154
 155         read_lock(&binfmt_lock);
 156         list_for_each_entry(fmt, &formats, lh) {
 157                 if (!fmt->load_shlib)
 158                         continue;
 159                 if (!try_module_get(fmt->module))
 160                         continue;
 161                 read_unlock(&binfmt_lock);
 162                 error = fmt->load_shlib(file);
 163                 read_lock(&binfmt_lock);
 164                 put_binfmt(fmt);
 165                 if (error != -ENOEXEC)
 166                         break;
 167         }
 168         read_unlock(&binfmt_lock);
 169 exit:
 170         fput(file);
 171 out:
 172         return error;
 173 }
 174 #endif /* #ifdef CONFIG_USELIB */
 175
 176 #ifdef CONFIG_MMU
 177 /*
 178  * The nascent bprm->mm is not visible until exec_mmap() but it can
 179  * use a lot of memory, account these pages in current->mm temporary
 180  * for oom_badness()->get_mm_rss(). Once exec succeeds or fails, we
 181  * change the counter back via acct_arg_size(0).
 182  */
 183 static void acct_arg_size(struct linux_binprm *bprm, unsigned long pages)
 184 {
 185         struct mm_struct *mm = current->mm;
 186         long diff = (long)(pages - bprm->vma_pages);
 187
 188         if (!mm || !diff)
 189                 return;
 190
 191         bprm->vma_pages = pages;
 192         add_mm_counter(mm, MM_ANONPAGES, diff);
 193 }
 194
 195 static struct page *get_arg_page(struct linux_binprm *bprm, unsigned long pos,
 196                 int write)
 197 {
 198         struct page *page;
 199         int ret;
 200         unsigned int gup_flags = FOLL_FORCE;
 201
 202 #ifdef CONFIG_STACK_GROWSUP
 203         if (write) {
 204                 ret = expand_downwards(bprm->vma, pos);
 205                 if (ret < 0)
 206                         return NULL;
 207         }
 208 #endif
 209
 210         if (write)
 211                 gup_flags |= FOLL_WRITE;
 212
 213         /*
 214          * We are doing an exec().  'current' is the process
 215          * doing the exec and bprm->mm is the new process's mm.
 216          */
 217         ret = get_user_pages_remote(current, bprm->mm, pos, 1, gup_flags,
 218                         &page, NULL, NULL);
 219         if (ret <= 0)
 220                 return NULL;
 221
 222         if (write) {
 223                 unsigned long size = bprm->vma->vm_end - bprm->vma->vm_start;
 224                 unsigned long ptr_size, limit;
 225
 226                 /*
 227                  * Since the stack will hold pointers to the strings, we
 228                  * must account for them as well.
 229                  *
 230                  * The size calculation is the entire vma while each arg page is
 231                  * built, so each time we get here it's calculating how far it
 232                  * is currently (rather than each call being just the newly
 233                  * added size from the arg page).  As a result, we need to
 234                  * always add the entire size of the pointers, so that on the
 235                  * last call to get_arg_page() we'll actually have the entire
 236                  * correct size.
 237                  */
 238                 ptr_size = (bprm->argc + bprm->envc) * sizeof(void *);
 239                 if (ptr_size > ULONG_MAX - size)
 240                         goto fail;
 241                 size += ptr_size;
 242
 243                 acct_arg_size(bprm, size / PAGE_SIZE);
 244
 245                 /*
 246                  * We've historically supported up to 32 pages (ARG_MAX)
 247                  * of argument strings even with small stacks
 248                  */
 249                 if (size <= ARG_MAX)
 250                         return page;
 251
 252                 /*
 253                  * Limit to 1/4 of the max stack size or 3/4 of _STK_LIM
 254                  * (whichever is smaller) for the argv+env strings.
 255                  * This ensures that:
 256                  *  - the remaining binfmt code will not run out of stack space,
 257                  *  - the program will have a reasonable amount of stack left
 258                  *    to work from.
 259                  */
 260                 limit = _STK_LIM / 4 * 3;
 261                 limit = min(limit, bprm->rlim_stack.rlim_cur / 4);
 262                 if (size > limit)
 263                         goto fail;
 264         }
 265
 266         return page;
 267
 268 fail:
 269         put_page(page);
 270         return NULL;
 271 }
 272
 273 static void put_arg_page(struct page *page)
 274 {
 275         put_page(page);
 276 }
 277
 278 static void free_arg_pages(struct linux_binprm *bprm)
 279 {
 280 }
 281
 282 static void flush_arg_page(struct linux_binprm *bprm, unsigned long pos,
 283                 struct page *page)
 284 {
 285         flush_cache_page(bprm->vma, pos, page_to_pfn(page));
 286 }
 287
 288 static int __bprm_mm_init(struct linux_binprm *bprm)
 289 {
 290         int err;
 291         struct vm_area_struct *vma = NULL;
 292         struct mm_struct *mm = bprm->mm;
 293
 294         bprm->vma = vma = vm_area_alloc(mm);
 295         if (!vma)
 296                 return -ENOMEM;
 297         vma_set_anonymous(vma);
 298
 299         if (down_write_killable(&mm->mmap_sem)) {
 300                 err = -EINTR;
 301                 goto err_free;
 302         }
 303
 304         /*
 305          * Place the stack at the largest stack address the architecture
 306          * supports. Later, we'll move this to an appropriate place. We don't
 307          * use STACK_TOP because that can depend on attributes which aren't
 308          * configured yet.
 309          */
 310         BUILD_BUG_ON(VM_STACK_FLAGS & VM_STACK_INCOMPLETE_SETUP);
 311         vma->vm_end = STACK_TOP_MAX;
 312         vma->vm_start = vma->vm_end - PAGE_SIZE;
 313         vma->vm_flags = VM_SOFTDIRTY | VM_STACK_FLAGS | VM_STACK_INCOMPLETE_SETUP;
 314         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
 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         vm_area_free(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         /* Save current stack limit for all calculations made during exec. */
 415         task_lock(current->group_leader);
 416         bprm->rlim_stack = current->signal->rlim[RLIMIT_STACK];
 417         task_unlock(current->group_leader);
 418
 419         err = __bprm_mm_init(bprm);
 420         if (err)
 421                 goto err;
 422
 423         return 0;
 424
 425 err:
 426         if (mm) {
 427                 bprm->mm = NULL;
 428                 mmdrop(mm);
 429         }
 430
 431         return err;
 432 }
 433
 434 struct user_arg_ptr {
 435 #ifdef CONFIG_COMPAT
 436         bool is_compat;
 437 #endif
 438         union {
 439                 const char __user *const __user *native;
 440 #ifdef CONFIG_COMPAT
 441                 const compat_uptr_t __user *compat;
 442 #endif
 443         } ptr;
 444 };
 445
 446 static const char __user *get_user_arg_ptr(struct user_arg_ptr argv, int nr)
 447 {
 448         const char __user *native;
 449
 450 #ifdef CONFIG_COMPAT
 451         if (unlikely(argv.is_compat)) {
 452                 compat_uptr_t compat;
 453
 454                 if (get_user(compat, argv.ptr.compat + nr))
 455                         return ERR_PTR(-EFAULT);
 456
 457                 return compat_ptr(compat);
 458         }
 459 #endif
 460
 461         if (get_user(native, argv.ptr.native + nr))
 462                 return ERR_PTR(-EFAULT);
 463
 464         return native;
 465 }
 466
 467 /*
 468  * count() counts the number of strings in array ARGV.
 469  */
 470 static int count(struct user_arg_ptr argv, int max)
 471 {
 472         int i = 0;
 473
 474         if (argv.ptr.native != NULL) {
 475                 for (;;) {
 476                         const char __user *p = get_user_arg_ptr(argv, i);
 477
 478                         if (!p)
 479                                 break;
 480
 481                         if (IS_ERR(p))
 482                                 return -EFAULT;
 483
 484                         if (i >= max)
 485                                 return -E2BIG;
 486                         ++i;
 487
 488                         if (fatal_signal_pending(current))
 489                                 return -ERESTARTNOHAND;
 490                         cond_resched();
 491                 }
 492         }
 493         return i;
 494 }
 495
 496 /*
 497  * 'copy_strings()' copies argument/environment strings from the old
 498  * processes's memory to the new process's stack.  The call to get_user_pages()
 499  * ensures the destination page is created and not swapped out.
 500  */
 501 static int copy_strings(int argc, struct user_arg_ptr argv,
 502                         struct linux_binprm *bprm)
 503 {
 504         struct page *kmapped_page = NULL;
 505         char *kaddr = NULL;
 506         unsigned long kpos = 0;
 507         int ret;
 508
 509         while (argc-- > 0) {
 510                 const char __user *str;
 511                 int len;
 512                 unsigned long pos;
 513
 514                 ret = -EFAULT;
 515                 str = get_user_arg_ptr(argv, argc);
 516                 if (IS_ERR(str))
 517                         goto out;
 518
 519                 len = strnlen_user(str, MAX_ARG_STRLEN);
 520                 if (!len)
 521                         goto out;
 522
 523                 ret = -E2BIG;
 524                 if (!valid_arg_len(bprm, len))
 525                         goto out;
 526
 527                 /* We're going to work our way backwords. */
 528                 pos = bprm->p;
 529                 str += len;
 530                 bprm->p -= len;
 531
 532                 while (len > 0) {
 533                         int offset, bytes_to_copy;
 534
 535                         if (fatal_signal_pending(current)) {
 536                                 ret = -ERESTARTNOHAND;
 537                                 goto out;
 538                         }
 539                         cond_resched();
 540
 541                         offset = pos % PAGE_SIZE;
 542                         if (offset == 0)
 543                                 offset = PAGE_SIZE;
 544
 545                         bytes_to_copy = offset;
 546                         if (bytes_to_copy > len)
 547                                 bytes_to_copy = len;
 548
 549                         offset -= bytes_to_copy;
 550                         pos -= bytes_to_copy;
 551                         str -= bytes_to_copy;
 552                         len -= bytes_to_copy;
 553
 554                         if (!kmapped_page || kpos != (pos & PAGE_MASK)) {
 555                                 struct page *page;
 556
 557                                 page = get_arg_page(bprm, pos, 1);
 558                                 if (!page) {
 559                                         ret = -E2BIG;
 560                                         goto out;
 561                                 }
 562
 563                                 if (kmapped_page) {
 564                                         flush_kernel_dcache_page(kmapped_page);
 565                                         kunmap(kmapped_page);
 566                                         put_arg_page(kmapped_page);
 567                                 }
 568                                 kmapped_page = page;
 569                                 kaddr = kmap(kmapped_page);
 570                                 kpos = pos & PAGE_MASK;
 571                                 flush_arg_page(bprm, kpos, kmapped_page);
 572                         }
 573                         if (copy_from_user(kaddr+offset, str, bytes_to_copy)) {
 574                                 ret = -EFAULT;
 575                                 goto out;
 576                         }
 577                 }
 578         }
 579         ret = 0;
 580 out:
 581         if (kmapped_page) {
 582                 flush_kernel_dcache_page(kmapped_page);
 583                 kunmap(kmapped_page);
 584                 put_arg_page(kmapped_page);
 585         }
 586         return ret;
 587 }
 588
 589 /*
 590  * Like copy_strings, but get argv and its values from kernel memory.
 591  */
 592 int copy_strings_kernel(int argc, const char *const *__argv,
 593                         struct linux_binprm *bprm)
 594 {
 595         int r;
 596         mm_segment_t oldfs = get_fs();
 597         struct user_arg_ptr argv = {
 598                 .ptr.native = (const char __user *const  __user *)__argv,
 599         };
 600
 601         set_fs(KERNEL_DS);
 602         r = copy_strings(argc, argv, bprm);
 603         set_fs(oldfs);
 604
 605         return r;
 606 }
 607 EXPORT_SYMBOL(copy_strings_kernel);
 608
 609 #ifdef CONFIG_MMU
 610
 611 /*
 612  * During bprm_mm_init(), we create a temporary stack at STACK_TOP_MAX.  Once
 613  * the binfmt code determines where the new stack should reside, we shift it to
 614  * its final location.  The process proceeds as follows:
 615  *
 616  * 1) Use shift to calculate the new vma endpoints.
 617  * 2) Extend vma to cover both the old and new ranges.  This ensures the
 618  *    arguments passed to subsequent functions are consistent.
 619  * 3) Move vma's page tables to the new range.
 620  * 4) Free up any cleared pgd range.
 621  * 5) Shrink the vma to cover only the new range.
 622  */
 623 static int shift_arg_pages(struct vm_area_struct *vma, unsigned long shift)
 624 {
 625         struct mm_struct *mm = vma->vm_mm;
 626         unsigned long old_start = vma->vm_start;
 627         unsigned long old_end = vma->vm_end;
 628         unsigned long length = old_end - old_start;
 629         unsigned long new_start = old_start - shift;
 630         unsigned long new_end = old_end - shift;
 631         struct mmu_gather tlb;
 632
 633         BUG_ON(new_start > new_end);
 634
 635         /*
 636          * ensure there are no vmas between where we want to go
 637          * and where we are
 638          */
 639         if (vma != find_vma(mm, new_start))
 640                 return -EFAULT;
 641
 642         /*
 643          * cover the whole range: [new_start, old_end)
 644          */
 645         if (vma_adjust(vma, new_start, old_end, vma->vm_pgoff, NULL))
 646                 return -ENOMEM;
 647
 648         /*
 649          * move the page tables downwards, on failure we rely on
 650          * process cleanup to remove whatever mess we made.
 651          */
 652         if (length != move_page_tables(vma, old_start,
 653                                        vma, new_start, length, false))
 654                 return -ENOMEM;
 655
 656         lru_add_drain();
 657         tlb_gather_mmu(&tlb, mm, old_start, old_end);
 658         if (new_end > old_start) {
 659                 /*
 660                  * when the old and new regions overlap clear from new_end.
 661                  */
 662                 free_pgd_range(&tlb, new_end, old_end, new_end,
 663                         vma->vm_next ? vma->vm_next->vm_start : USER_PGTABLES_CEILING);
 664         } else {
 665                 /*
 666                  * otherwise, clean from old_start; this is done to not touch
 667                  * the address space in [new_end, old_start) some architectures
 668                  * have constraints on va-space that make this illegal (IA64) -
 669                  * for the others its just a little faster.
 670                  */
 671                 free_pgd_range(&tlb, old_start, old_end, new_end,
 672                         vma->vm_next ? vma->vm_next->vm_start : USER_PGTABLES_CEILING);
 673         }
 674         tlb_finish_mmu(&tlb, old_start, old_end);
 675
 676         /*
 677          * Shrink the vma to just the new range.  Always succeeds.
 678          */
 679         vma_adjust(vma, new_start, new_end, vma->vm_pgoff, NULL);
 680
 681         return 0;
 682 }
 683
 684 /*
 685  * Finalizes the stack vm_area_struct. The flags and permissions are updated,
 686  * the stack is optionally relocated, and some extra space is added.
 687  */
 688 int setup_arg_pages(struct linux_binprm *bprm,
 689                     unsigned long stack_top,
 690                     int executable_stack)
 691 {
 692         unsigned long ret;
 693         unsigned long stack_shift;
 694         struct mm_struct *mm = current->mm;
 695         struct vm_area_struct *vma = bprm->vma;
 696         struct vm_area_struct *prev = NULL;
 697         unsigned long vm_flags;
 698         unsigned long stack_base;
 699         unsigned long stack_size;
 700         unsigned long stack_expand;
 701         unsigned long rlim_stack;
 702
 703 #ifdef CONFIG_STACK_GROWSUP
 704         /* Limit stack size */
 705         stack_base = bprm->rlim_stack.rlim_max;
 706         if (stack_base > STACK_SIZE_MAX)
 707                 stack_base = STACK_SIZE_MAX;
 708
 709         /* Add space for stack randomization. */
 710         stack_base += (STACK_RND_MASK << PAGE_SHIFT);
 711
 712         /* Make sure we didn't let the argument array grow too large. */
 713         if (vma->vm_end - vma->vm_start > stack_base)
 714                 return -ENOMEM;
 715
 716         stack_base = PAGE_ALIGN(stack_top - stack_base);
 717
 718         stack_shift = vma->vm_start - stack_base;
 719         mm->arg_start = bprm->p - stack_shift;
 720         bprm->p = vma->vm_end - stack_shift;
 721 #else
 722         stack_top = arch_align_stack(stack_top);
 723         stack_top = PAGE_ALIGN(stack_top);
 724
 725         if (unlikely(stack_top < mmap_min_addr) ||
 726             unlikely(vma->vm_end - vma->vm_start >= stack_top - mmap_min_addr))
 727                 return -ENOMEM;
 728
 729         stack_shift = vma->vm_end - stack_top;
 730
 731         bprm->p -= stack_shift;
 732         mm->arg_start = bprm->p;
 733 #endif
 734
 735         if (bprm->loader)
 736                 bprm->loader -= stack_shift;
 737         bprm->exec -= stack_shift;
 738
 739         if (down_write_killable(&mm->mmap_sem))
 740                 return -EINTR;
 741
 742         vm_flags = VM_STACK_FLAGS;
 743
 744         /*
 745          * Adjust stack execute permissions; explicitly enable for
 746          * EXSTACK_ENABLE_X, disable for EXSTACK_DISABLE_X and leave alone
 747          * (arch default) otherwise.
 748          */
 749         if (unlikely(executable_stack == EXSTACK_ENABLE_X))
 750                 vm_flags |= VM_EXEC;
 751         else if (executable_stack == EXSTACK_DISABLE_X)
 752                 vm_flags &= ~VM_EXEC;
 753         vm_flags |= mm->def_flags;
 754         vm_flags |= VM_STACK_INCOMPLETE_SETUP;
 755
 756         ret = mprotect_fixup(vma, &prev, vma->vm_start, vma->vm_end,
 757                         vm_flags);
 758         if (ret)
 759                 goto out_unlock;
 760         BUG_ON(prev != vma);
 761
 762         /* Move stack pages down in memory. */
 763         if (stack_shift) {
 764                 ret = shift_arg_pages(vma, stack_shift);
 765                 if (ret)
 766                         goto out_unlock;
 767         }
 768
 769         /* mprotect_fixup is overkill to remove the temporary stack flags */
 770         vma->vm_flags &= ~VM_STACK_INCOMPLETE_SETUP;
 771
 772         stack_expand = 131072UL; /* randomly 32*4k (or 2*64k) pages */
 773         stack_size = vma->vm_end - vma->vm_start;
 774         /*
 775          * Align this down to a page boundary as expand_stack
 776          * will align it up.
 777          */
 778         rlim_stack = bprm->rlim_stack.rlim_cur & PAGE_MASK;
 779 #ifdef CONFIG_STACK_GROWSUP
 780         if (stack_size + stack_expand > rlim_stack)
 781                 stack_base = vma->vm_start + rlim_stack;
 782         else
 783                 stack_base = vma->vm_end + stack_expand;
 784 #else
 785         if (stack_size + stack_expand > rlim_stack)
 786                 stack_base = vma->vm_end - rlim_stack;
 787         else
 788                 stack_base = vma->vm_start - stack_expand;
 789 #endif
 790         current->mm->start_stack = bprm->p;
 791         ret = expand_stack(vma, stack_base);
 792         if (ret)
 793                 ret = -EFAULT;
 794
 795 out_unlock:
 796         up_write(&mm->mmap_sem);
 797         return ret;
 798 }
 799 EXPORT_SYMBOL(setup_arg_pages);
 800
 801 #else
 802
 803 /*
 804  * Transfer the program arguments and environment from the holding pages
 805  * onto the stack. The provided stack pointer is adjusted accordingly.
 806  */
 807 int transfer_args_to_stack(struct linux_binprm *bprm,
 808                            unsigned long *sp_location)
 809 {
 810         unsigned long index, stop, sp;
 811         int ret = 0;
 812
 813         stop = bprm->p >> PAGE_SHIFT;
 814         sp = *sp_location;
 815
 816         for (index = MAX_ARG_PAGES - 1; index >= stop; index--) {
 817                 unsigned int offset = index == stop ? bprm->p & ~PAGE_MASK : 0;
 818                 char *src = kmap(bprm->page[index]) + offset;
 819                 sp -= PAGE_SIZE - offset;
 820                 if (copy_to_user((void *) sp, src, PAGE_SIZE - offset) != 0)
 821                         ret = -EFAULT;
 822                 kunmap(bprm->page[index]);
 823                 if (ret)
 824                         goto out;
 825         }
 826
 827         *sp_location = sp;
 828
 829 out:
 830         return ret;
 831 }
 832 EXPORT_SYMBOL(transfer_args_to_stack);
 833
 834 #endif /* CONFIG_MMU */
 835
 836 static struct file *do_open_execat(int fd, struct filename *name, int flags)
 837 {
 838         struct file *file;
 839         int err;
 840         struct open_flags open_exec_flags = {
 841                 .open_flag = O_LARGEFILE | O_RDONLY | __FMODE_EXEC,
 842                 .acc_mode = MAY_EXEC,
 843                 .intent = LOOKUP_OPEN,
 844                 .lookup_flags = LOOKUP_FOLLOW,
 845         };
 846
 847         if ((flags & ~(AT_SYMLINK_NOFOLLOW | AT_EMPTY_PATH)) != 0)
 848                 return ERR_PTR(-EINVAL);
 849         if (flags & AT_SYMLINK_NOFOLLOW)
 850                 open_exec_flags.lookup_flags &= ~LOOKUP_FOLLOW;
 851         if (flags & AT_EMPTY_PATH)
 852                 open_exec_flags.lookup_flags |= LOOKUP_EMPTY;
 853
 854         file = do_filp_open(fd, name, &open_exec_flags);
 855         if (IS_ERR(file))
 856                 goto out;
 857
 858         err = -EACCES;
 859         if (!S_ISREG(file_inode(file)->i_mode))
 860                 goto exit;
 861
 862         if (path_noexec(&file->f_path))
 863                 goto exit;
 864
 865         err = deny_write_access(file);
 866         if (err)
 867                 goto exit;
 868
 869         if (name->name[0] != '\0')
 870                 fsnotify_open(file);
 871
 872 out:
 873         return file;
 874
 875 exit:
 876         fput(file);
 877         return ERR_PTR(err);
 878 }
 879
 880 struct file *open_exec(const char *name)
 881 {
 882         struct filename *filename = getname_kernel(name);
 883         struct file *f = ERR_CAST(filename);
 884
 885         if (!IS_ERR(filename)) {
 886                 f = do_open_execat(AT_FDCWD, filename, 0);
 887                 putname(filename);
 888         }
 889         return f;
 890 }
 891 EXPORT_SYMBOL(open_exec);
 892
 893 int kernel_read_file(struct file *file, void **buf, loff_t *size,
 894                      loff_t max_size, enum kernel_read_file_id id)
 895 {
 896         loff_t i_size, pos;
 897         ssize_t bytes = 0;
 898         int ret;
 899
 900         if (!S_ISREG(file_inode(file)->i_mode) || max_size < 0)
 901                 return -EINVAL;
 902
 903         ret = deny_write_access(file);
 904         if (ret)
 905                 return ret;
 906
 907         ret = security_kernel_read_file(file, id);
 908         if (ret)
 909                 goto out;
 910
 911         i_size = i_size_read(file_inode(file));
 912         if (i_size <= 0) {
 913                 ret = -EINVAL;
 914                 goto out;
 915         }
 916         if (i_size > SIZE_MAX || (max_size > 0 && i_size > max_size)) {
 917                 ret = -EFBIG;
 918                 goto out;
 919         }
 920
 921         if (id != READING_FIRMWARE_PREALLOC_BUFFER)
 922                 *buf = vmalloc(i_size);
 923         if (!*buf) {
 924                 ret = -ENOMEM;
 925                 goto out;
 926         }
 927
 928         pos = 0;
 929         while (pos < i_size) {
 930                 bytes = kernel_read(file, *buf + pos, i_size - pos, &pos);
 931                 if (bytes < 0) {
 932                         ret = bytes;
 933                         goto out;
 934                 }
 935
 936                 if (bytes == 0)
 937                         break;
 938         }
 939
 940         if (pos != i_size) {
 941                 ret = -EIO;
 942                 goto out_free;
 943         }
 944
 945         ret = security_kernel_post_read_file(file, *buf, i_size, id);
 946         if (!ret)
 947                 *size = pos;
 948
 949 out_free:
 950         if (ret < 0) {
 951                 if (id != READING_FIRMWARE_PREALLOC_BUFFER) {
 952                         vfree(*buf);
 953                         *buf = NULL;
 954                 }
 955         }
 956
 957 out:
 958         allow_write_access(file);
 959         return ret;
 960 }
 961 EXPORT_SYMBOL_GPL(kernel_read_file);
 962
 963 int kernel_read_file_from_path(const char *path, void **buf, loff_t *size,
 964                                loff_t max_size, enum kernel_read_file_id id)
 965 {
 966         struct file *file;
 967         int ret;
 968
 969         if (!path || !*path)
 970                 return -EINVAL;
 971
 972         file = filp_open(path, O_RDONLY, 0);
 973         if (IS_ERR(file))
 974                 return PTR_ERR(file);
 975
 976         ret = kernel_read_file(file, buf, size, max_size, id);
 977         fput(file);
 978         return ret;
 979 }
 980 EXPORT_SYMBOL_GPL(kernel_read_file_from_path);
 981
 982 int kernel_read_file_from_fd(int fd, void **buf, loff_t *size, loff_t max_size,
 983                              enum kernel_read_file_id id)
 984 {
 985         struct fd f = fdget(fd);
 986         int ret = -EBADF;
 987
 988         if (!f.file)
 989                 goto out;
 990
 991         ret = kernel_read_file(f.file, buf, size, max_size, id);
 992 out:
 993         fdput(f);
 994         return ret;
 995 }
 996 EXPORT_SYMBOL_GPL(kernel_read_file_from_fd);
 997
 998 ssize_t read_code(struct file *file, unsigned long addr, loff_t pos, size_t len)
 999 {
1000         ssize_t res = vfs_read(file, (void __user *)addr, len, &pos);
1001         if (res > 0)
1002                 flush_icache_range(addr, addr + len);
1003         return res;
1004 }
1005 EXPORT_SYMBOL(read_code);
1006
1007 static int exec_mmap(struct mm_struct *mm)
1008 {
1009         struct task_struct *tsk;
1010         struct mm_struct *old_mm, *active_mm;
1011
1012         /* Notify parent that we're no longer interested in the old VM */
1013         tsk = current;
1014         old_mm = current->mm;
1015         mm_release(tsk, old_mm);
1016
1017         if (old_mm) {
1018                 sync_mm_rss(old_mm);
1019                 /*
1020                  * Make sure that if there is a core dump in progress
1021                  * for the old mm, we get out and die instead of going
1022                  * through with the exec.  We must hold mmap_sem around
1023                  * checking core_state and changing tsk->mm.
1024                  */
1025                 down_read(&old_mm->mmap_sem);
1026                 if (unlikely(old_mm->core_state)) {
1027                         up_read(&old_mm->mmap_sem);
1028                         return -EINTR;
1029                 }
1030         }
1031         task_lock(tsk);
1032         active_mm = tsk->active_mm;
1033         tsk->mm = mm;
1034         tsk->active_mm = mm;
1035         activate_mm(active_mm, mm);
1036         tsk->mm->vmacache_seqnum = 0;
1037         vmacache_flush(tsk);
1038         task_unlock(tsk);
1039         if (old_mm) {
1040                 up_read(&old_mm->mmap_sem);
1041                 BUG_ON(active_mm != old_mm);
1042                 setmax_mm_hiwater_rss(&tsk->signal->maxrss, old_mm);
1043                 mm_update_next_owner(old_mm);
1044                 mmput(old_mm);
1045                 return 0;
1046         }
1047         mmdrop(active_mm);
1048         return 0;
1049 }
1050
1051 /*
1052  * This function makes sure the current process has its own signal table,
1053  * so that flush_signal_handlers can later reset the handlers without
1054  * disturbing other processes.  (Other processes might share the signal
1055  * table via the CLONE_SIGHAND option to clone().)
1056  */
1057 static int de_thread(struct task_struct *tsk)
1058 {
1059         struct signal_struct *sig = tsk->signal;
1060         struct sighand_struct *oldsighand = tsk->sighand;
1061         spinlock_t *lock = &oldsighand->siglock;
1062
1063         if (thread_group_empty(tsk))
1064                 goto no_thread_group;
1065
1066         /*
1067          * Kill all other threads in the thread group.
1068          */
1069         spin_lock_irq(lock);
1070         if (signal_group_exit(sig)) {
1071                 /*
1072                  * Another group action in progress, just
1073                  * return so that the signal is processed.
1074                  */
1075                 spin_unlock_irq(lock);
1076                 return -EAGAIN;
1077         }
1078
1079         sig->group_exit_task = tsk;
1080         sig->notify_count = zap_other_threads(tsk);
1081         if (!thread_group_leader(tsk))
1082                 sig->notify_count--;
1083
1084         while (sig->notify_count) {
1085                 __set_current_state(TASK_KILLABLE);
1086                 spin_unlock_irq(lock);
1087                 freezable_schedule();
1088                 if (unlikely(__fatal_signal_pending(tsk)))
1089                         goto killed;
1090                 spin_lock_irq(lock);
1091         }
1092         spin_unlock_irq(lock);
1093
1094         /*
1095          * At this point all other threads have exited, all we have to
1096          * do is to wait for the thread group leader to become inactive,
1097          * and to assume its PID:
1098          */
1099         if (!thread_group_leader(tsk)) {
1100                 struct task_struct *leader = tsk->group_leader;
1101
1102                 for (;;) {
1103                         cgroup_threadgroup_change_begin(tsk);
1104                         write_lock_irq(&tasklist_lock);
1105                         /*
1106                          * Do this under tasklist_lock to ensure that
1107                          * exit_notify() can't miss ->group_exit_task
1108                          */
1109                         sig->notify_count = -1;
1110                         if (likely(leader->exit_state))
1111                                 break;
1112                         __set_current_state(TASK_KILLABLE);
1113                         write_unlock_irq(&tasklist_lock);
1114                         cgroup_threadgroup_change_end(tsk);
1115                         freezable_schedule();
1116                         if (unlikely(__fatal_signal_pending(tsk)))
1117                                 goto killed;
1118                 }
1119
1120                 /*
1121                  * The only record we have of the real-time age of a
1122                  * process, regardless of execs it's done, is start_time.
1123                  * All the past CPU time is accumulated in signal_struct
1124                  * from sister threads now dead.  But in this non-leader
1125                  * exec, nothing survives from the original leader thread,
1126                  * whose birth marks the true age of this process now.
1127                  * When we take on its identity by switching to its PID, we
1128                  * also take its birthdate (always earlier than our own).
1129                  */
1130                 tsk->start_time = leader->start_time;
1131                 tsk->real_start_time = leader->real_start_time;
1132
1133                 BUG_ON(!same_thread_group(leader, tsk));
1134                 BUG_ON(has_group_leader_pid(tsk));
1135                 /*
1136                  * An exec() starts a new thread group with the
1137                  * TGID of the previous thread group. Rehash the
1138                  * two threads with a switched PID, and release
1139                  * the former thread group leader:
1140                  */
1141
1142                 /* Become a process group leader with the old leader's pid.
1143                  * The old leader becomes a thread of the this thread group.
1144                  * Note: The old leader also uses this pid until release_task
1145                  *       is called.  Odd but simple and correct.
1146                  */
1147                 tsk->pid = leader->pid;
1148                 change_pid(tsk, PIDTYPE_PID, task_pid(leader));
1149                 transfer_pid(leader, tsk, PIDTYPE_TGID);
1150                 transfer_pid(leader, tsk, PIDTYPE_PGID);
1151                 transfer_pid(leader, tsk, PIDTYPE_SID);
1152
1153                 list_replace_rcu(&leader->tasks, &tsk->tasks);
1154                 list_replace_init(&leader->sibling, &tsk->sibling);
1155
1156                 tsk->group_leader = tsk;
1157                 leader->group_leader = tsk;
1158
1159                 tsk->exit_signal = SIGCHLD;
1160                 leader->exit_signal = -1;
1161
1162                 BUG_ON(leader->exit_state != EXIT_ZOMBIE);
1163                 leader->exit_state = EXIT_DEAD;
1164
1165                 /*
1166                  * We are going to release_task()->ptrace_unlink() silently,
1167                  * the tracer can sleep in do_wait(). EXIT_DEAD guarantees
1168                  * the tracer wont't block again waiting for this thread.
1169                  */
1170                 if (unlikely(leader->ptrace))
1171                         __wake_up_parent(leader, leader->parent);
1172                 write_unlock_irq(&tasklist_lock);
1173                 cgroup_threadgroup_change_end(tsk);
1174
1175                 release_task(leader);
1176         }
1177
1178         sig->group_exit_task = NULL;
1179         sig->notify_count = 0;
1180
1181 no_thread_group:
1182         /* we have changed execution domain */
1183         tsk->exit_signal = SIGCHLD;
1184
1185 #ifdef CONFIG_POSIX_TIMERS
1186         exit_itimers(sig);
1187         flush_itimer_signals();
1188 #endif
1189
1190         if (atomic_read(&oldsighand->count) != 1) {
1191                 struct sighand_struct *newsighand;
1192                 /*
1193                  * This ->sighand is shared with the CLONE_SIGHAND
1194                  * but not CLONE_THREAD task, switch to the new one.
1195                  */
1196                 newsighand = kmem_cache_alloc(sighand_cachep, GFP_KERNEL);
1197                 if (!newsighand)
1198                         return -ENOMEM;
1199
1200                 atomic_set(&newsighand->count, 1);
1201                 memcpy(newsighand->action, oldsighand->action,
1202                        sizeof(newsighand->action));
1203
1204                 write_lock_irq(&tasklist_lock);
1205                 spin_lock(&oldsighand->siglock);
1206                 rcu_assign_pointer(tsk->sighand, newsighand);
1207                 spin_unlock(&oldsighand->siglock);
1208                 write_unlock_irq(&tasklist_lock);
1209
1210                 __cleanup_sighand(oldsighand);
1211         }
1212
1213         BUG_ON(!thread_group_leader(tsk));
1214         return 0;
1215
1216 killed:
1217         /* protects against exit_notify() and __exit_signal() */
1218         read_lock(&tasklist_lock);
1219         sig->group_exit_task = NULL;
1220         sig->notify_count = 0;
1221         read_unlock(&tasklist_lock);
1222         return -EAGAIN;
1223 }
1224
1225 char *__get_task_comm(char *buf, size_t buf_size, struct task_struct *tsk)
1226 {
1227         task_lock(tsk);
1228         strncpy(buf, tsk->comm, buf_size);
1229         task_unlock(tsk);
1230         return buf;
1231 }
1232 EXPORT_SYMBOL_GPL(__get_task_comm);
1233
1234 /*
1235  * These functions flushes out all traces of the currently running executable
1236  * so that a new one can be started
1237  */
1238
1239 void __set_task_comm(struct task_struct *tsk, const char *buf, bool exec)
1240 {
1241         task_lock(tsk);
1242         trace_task_rename(tsk, buf);
1243         strlcpy(tsk->comm, buf, sizeof(tsk->comm));
1244         task_unlock(tsk);
1245         perf_event_comm(tsk, exec);
1246 }
1247
1248 /*
1249  * Calling this is the point of no return. None of the failures will be
1250  * seen by userspace since either the process is already taking a fatal
1251  * signal (via de_thread() or coredump), or will have SEGV raised
1252  * (after exec_mmap()) by search_binary_handlers (see below).
1253  */
1254 int flush_old_exec(struct linux_binprm * bprm)
1255 {
1256         int retval;
1257
1258         /*
1259          * Make sure we have a private signal table and that
1260          * we are unassociated from the previous thread group.
1261          */
1262         retval = de_thread(current);
1263         if (retval)
1264                 goto out;
1265
1266         /*
1267          * Must be called _before_ exec_mmap() as bprm->mm is
1268          * not visibile until then. This also enables the update
1269          * to be lockless.
1270          */
1271         set_mm_exe_file(bprm->mm, bprm->file);
1272
1273         /*
1274          * Release all of the old mmap stuff
1275          */
1276         acct_arg_size(bprm, 0);
1277         retval = exec_mmap(bprm->mm);
1278         if (retval)
1279                 goto out;
1280
1281         /*
1282          * After clearing bprm->mm (to mark that current is using the
1283          * prepared mm now), we have nothing left of the original
1284          * process. If anything from here on returns an error, the check
1285          * in search_binary_handler() will SEGV current.
1286          */
1287         bprm->mm = NULL;
1288
1289         set_fs(USER_DS);
1290         current->flags &= ~(PF_RANDOMIZE | PF_FORKNOEXEC | PF_KTHREAD |
1291                                         PF_NOFREEZE | PF_NO_SETAFFINITY);
1292         flush_thread();
1293         current->personality &= ~bprm->per_clear;
1294
1295         /*
1296          * We have to apply CLOEXEC before we change whether the process is
1297          * dumpable (in setup_new_exec) to avoid a race with a process in userspace
1298          * trying to access the should-be-closed file descriptors of a process
1299          * undergoing exec(2).
1300          */
1301         do_close_on_exec(current->files);
1302         return 0;
1303
1304 out:
1305         return retval;
1306 }
1307 EXPORT_SYMBOL(flush_old_exec);
1308
1309 void would_dump(struct linux_binprm *bprm, struct file *file)
1310 {
1311         struct inode *inode = file_inode(file);
1312         if (inode_permission(inode, MAY_READ) < 0) {
1313                 struct user_namespace *old, *user_ns;
1314                 bprm->interp_flags |= BINPRM_FLAGS_ENFORCE_NONDUMP;
1315
1316                 /* Ensure mm->user_ns contains the executable */
1317                 user_ns = old = bprm->mm->user_ns;
1318                 while ((user_ns != &init_user_ns) &&
1319                        !privileged_wrt_inode_uidgid(user_ns, inode))
1320                         user_ns = user_ns->parent;
1321
1322                 if (old != user_ns) {
1323                         bprm->mm->user_ns = get_user_ns(user_ns);
1324                         put_user_ns(old);
1325                 }
1326         }
1327 }
1328 EXPORT_SYMBOL(would_dump);
1329
1330 void setup_new_exec(struct linux_binprm * bprm)
1331 {
1332         /*
1333          * Once here, prepare_binrpm() will not be called any more, so
1334          * the final state of setuid/setgid/fscaps can be merged into the
1335          * secureexec flag.
1336          */
1337         bprm->secureexec |= bprm->cap_elevated;
1338
1339         if (bprm->secureexec) {
1340                 /* Make sure parent cannot signal privileged process. */
1341                 current->pdeath_signal = 0;
1342
1343                 /*
1344                  * For secureexec, reset the stack limit to sane default to
1345                  * avoid bad behavior from the prior rlimits. This has to
1346                  * happen before arch_pick_mmap_layout(), which examines
1347                  * RLIMIT_STACK, but after the point of no return to avoid
1348                  * needing to clean up the change on failure.
1349                  */
1350                 if (bprm->rlim_stack.rlim_cur > _STK_LIM)
1351                         bprm->rlim_stack.rlim_cur = _STK_LIM;
1352         }
1353
1354         arch_pick_mmap_layout(current->mm, &bprm->rlim_stack);
1355
1356         current->sas_ss_sp = current->sas_ss_size = 0;
1357
1358         /*
1359          * Figure out dumpability. Note that this checking only of current
1360          * is wrong, but userspace depends on it. This should be testing
1361          * bprm->secureexec instead.
1362          */
1363         if (bprm->interp_flags & BINPRM_FLAGS_ENFORCE_NONDUMP ||
1364             !(uid_eq(current_euid(), current_uid()) &&
1365               gid_eq(current_egid(), current_gid())))
1366                 set_dumpable(current->mm, suid_dumpable);
1367         else
1368                 set_dumpable(current->mm, SUID_DUMP_USER);
1369
1370         arch_setup_new_exec();
1371         perf_event_exec();
1372         __set_task_comm(current, kbasename(bprm->filename), true);
1373
1374         /* Set the new mm task size. We have to do that late because it may
1375          * depend on TIF_32BIT which is only updated in flush_thread() on
1376          * some architectures like powerpc
1377          */
1378         current->mm->task_size = TASK_SIZE;
1379
1380         /* An exec changes our domain. We are no longer part of the thread
1381            group */
1382         current->self_exec_id++;
1383         flush_signal_handlers(current, 0);
1384 }
1385 EXPORT_SYMBOL(setup_new_exec);
1386
1387 /* Runs immediately before start_thread() takes over. */
1388 void finalize_exec(struct linux_binprm *bprm)
1389 {
1390         /* Store any stack rlimit changes before starting thread. */
1391         task_lock(current->group_leader);
1392         current->signal->rlim[RLIMIT_STACK] = bprm->rlim_stack;
1393         task_unlock(current->group_leader);
1394 }
1395 EXPORT_SYMBOL(finalize_exec);
1396
1397 /*
1398  * Prepare credentials and lock ->cred_guard_mutex.
1399  * install_exec_creds() commits the new creds and drops the lock.
1400  * Or, if exec fails before, free_bprm() should release ->cred and
1401  * and unlock.
1402  */
1403 int prepare_bprm_creds(struct linux_binprm *bprm)
1404 {
1405         if (mutex_lock_interruptible(&current->signal->cred_guard_mutex))
1406                 return -ERESTARTNOINTR;
1407
1408         bprm->cred = prepare_exec_creds();
1409         if (likely(bprm->cred))
1410                 return 0;
1411
1412         mutex_unlock(&current->signal->cred_guard_mutex);
1413         return -ENOMEM;
1414 }
1415
1416 static void free_bprm(struct linux_binprm *bprm)
1417 {
1418         free_arg_pages(bprm);
1419         if (bprm->cred) {
1420                 mutex_unlock(&current->signal->cred_guard_mutex);
1421                 abort_creds(bprm->cred);
1422         }
1423         if (bprm->file) {
1424                 allow_write_access(bprm->file);
1425                 fput(bprm->file);
1426         }
1427         /* If a binfmt changed the interp, free it. */
1428         if (bprm->interp != bprm->filename)
1429                 kfree(bprm->interp);
1430         kfree(bprm);
1431 }
1432
1433 int bprm_change_interp(const char *interp, struct linux_binprm *bprm)
1434 {
1435         /* If a binfmt changed the interp, free it first. */
1436         if (bprm->interp != bprm->filename)
1437                 kfree(bprm->interp);
1438         bprm->interp = kstrdup(interp, GFP_KERNEL);
1439         if (!bprm->interp)
1440                 return -ENOMEM;
1441         return 0;
1442 }
1443 EXPORT_SYMBOL(bprm_change_interp);
1444
1445 /*
1446  * install the new credentials for this executable
1447  */
1448 void install_exec_creds(struct linux_binprm *bprm)
1449 {
1450         security_bprm_committing_creds(bprm);
1451
1452         commit_creds(bprm->cred);
1453         bprm->cred = NULL;
1454
1455         /*
1456          * Disable monitoring for regular users
1457          * when executing setuid binaries. Must
1458          * wait until new credentials are committed
1459          * by commit_creds() above
1460          */
1461         if (get_dumpable(current->mm) != SUID_DUMP_USER)
1462                 perf_event_exit_task(current);
1463         /*
1464          * cred_guard_mutex must be held at least to this point to prevent
1465          * ptrace_attach() from altering our determination of the task's
1466          * credentials; any time after this it may be unlocked.
1467          */
1468         security_bprm_committed_creds(bprm);
1469         mutex_unlock(&current->signal->cred_guard_mutex);
1470 }
1471 EXPORT_SYMBOL(install_exec_creds);
1472
1473 /*
1474  * determine how safe it is to execute the proposed program
1475  * - the caller must hold ->cred_guard_mutex to protect against
1476  *   PTRACE_ATTACH or seccomp thread-sync
1477  */
1478 static void check_unsafe_exec(struct linux_binprm *bprm)
1479 {
1480         struct task_struct *p = current, *t;
1481         unsigned n_fs;
1482
1483         if (p->ptrace)
1484                 bprm->unsafe |= LSM_UNSAFE_PTRACE;
1485
1486         /*
1487          * This isn't strictly necessary, but it makes it harder for LSMs to
1488          * mess up.
1489          */
1490         if (task_no_new_privs(current))
1491                 bprm->unsafe |= LSM_UNSAFE_NO_NEW_PRIVS;
1492
1493         t = p;
1494         n_fs = 1;
1495         spin_lock(&p->fs->lock);
1496         rcu_read_lock();
1497         while_each_thread(p, t) {
1498                 if (t->fs == p->fs)
1499                         n_fs++;
1500         }
1501         rcu_read_unlock();
1502
1503         if (p->fs->users > n_fs)
1504                 bprm->unsafe |= LSM_UNSAFE_SHARE;
1505         else
1506                 p->fs->in_exec = 1;
1507         spin_unlock(&p->fs->lock);
1508 }
1509
1510 static void bprm_fill_uid(struct linux_binprm *bprm)
1511 {
1512         struct inode *inode;
1513         unsigned int mode;
1514         kuid_t uid;
1515         kgid_t gid;
1516
1517         /*
1518          * Since this can be called multiple times (via prepare_binprm),
1519          * we must clear any previous work done when setting set[ug]id
1520          * bits from any earlier bprm->file uses (for example when run
1521          * first for a setuid script then again for its interpreter).
1522          */
1523         bprm->cred->euid = current_euid();
1524         bprm->cred->egid = current_egid();
1525
1526         if (!mnt_may_suid(bprm->file->f_path.mnt))
1527                 return;
1528
1529         if (task_no_new_privs(current))
1530                 return;
1531
1532         inode = bprm->file->f_path.dentry->d_inode;
1533         mode = READ_ONCE(inode->i_mode);
1534         if (!(mode & (S_ISUID|S_ISGID)))
1535                 return;
1536
1537         /* Be careful if suid/sgid is set */
1538         inode_lock(inode);
1539
1540         /* reload atomically mode/uid/gid now that lock held */
1541         mode = inode->i_mode;
1542         uid = inode->i_uid;
1543         gid = inode->i_gid;
1544         inode_unlock(inode);
1545
1546         /* We ignore suid/sgid if there are no mappings for them in the ns */
1547         if (!kuid_has_mapping(bprm->cred->user_ns, uid) ||
1548                  !kgid_has_mapping(bprm->cred->user_ns, gid))
1549                 return;
1550
1551         if (mode & S_ISUID) {
1552                 bprm->per_clear |= PER_CLEAR_ON_SETID;
1553                 bprm->cred->euid = uid;
1554         }
1555
1556         if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP)) {
1557                 bprm->per_clear |= PER_CLEAR_ON_SETID;
1558                 bprm->cred->egid = gid;
1559         }
1560 }
1561
1562 /*
1563  * Fill the binprm structure from the inode.
1564  * Check permissions, then read the first 128 (BINPRM_BUF_SIZE) bytes
1565  *
1566  * This may be called multiple times for binary chains (scripts for example).
1567  */
1568 int prepare_binprm(struct linux_binprm *bprm)
1569 {
1570         int retval;
1571         loff_t pos = 0;
1572
1573         bprm_fill_uid(bprm);
1574
1575         /* fill in binprm security blob */
1576         retval = security_bprm_set_creds(bprm);
1577         if (retval)
1578                 return retval;
1579         bprm->called_set_creds = 1;
1580
1581         memset(bprm->buf, 0, BINPRM_BUF_SIZE);
1582         return kernel_read(bprm->file, bprm->buf, BINPRM_BUF_SIZE, &pos);
1583 }
1584
1585 EXPORT_SYMBOL(prepare_binprm);
1586
1587 /*
1588  * Arguments are '\0' separated strings found at the location bprm->p
1589  * points to; chop off the first by relocating brpm->p to right after
1590  * the first '\0' encountered.
1591  */
1592 int remove_arg_zero(struct linux_binprm *bprm)
1593 {
1594         int ret = 0;
1595         unsigned long offset;
1596         char *kaddr;
1597         struct page *page;
1598
1599         if (!bprm->argc)
1600                 return 0;
1601
1602         do {
1603                 offset = bprm->p & ~PAGE_MASK;
1604                 page = get_arg_page(bprm, bprm->p, 0);
1605                 if (!page) {
1606                         ret = -EFAULT;
1607                         goto out;
1608                 }
1609                 kaddr = kmap_atomic(page);
1610
1611                 for (; offset < PAGE_SIZE && kaddr[offset];
1612                                 offset++, bprm->p++)
1613                         ;
1614
1615                 kunmap_atomic(kaddr);
1616                 put_arg_page(page);
1617         } while (offset == PAGE_SIZE);
1618
1619         bprm->p++;
1620         bprm->argc--;
1621         ret = 0;
1622
1623 out:
1624         return ret;
1625 }
1626 EXPORT_SYMBOL(remove_arg_zero);
1627
1628 #define printable(c) (((c)=='\t') || ((c)=='\n') || (0x20<=(c) && (c)<=0x7e))
1629 /*
1630  * cycle the list of binary formats handler, until one recognizes the image
1631  */
1632 int search_binary_handler(struct linux_binprm *bprm)
1633 {
1634         bool need_retry = IS_ENABLED(CONFIG_MODULES);
1635         struct linux_binfmt *fmt;
1636         int retval;
1637
1638         /* This allows 4 levels of binfmt rewrites before failing hard. */
1639         if (bprm->recursion_depth > 5)
1640                 return -ELOOP;
1641
1642         retval = security_bprm_check(bprm);
1643         if (retval)
1644                 return retval;
1645
1646         retval = -ENOENT;
1647  retry:
1648         read_lock(&binfmt_lock);
1649         list_for_each_entry(fmt, &formats, lh) {
1650                 if (!try_module_get(fmt->module))
1651                         continue;
1652                 read_unlock(&binfmt_lock);
1653                 bprm->recursion_depth++;
1654                 retval = fmt->load_binary(bprm);
1655                 read_lock(&binfmt_lock);
1656                 put_binfmt(fmt);
1657                 bprm->recursion_depth--;
1658                 if (retval < 0 && !bprm->mm) {
1659                         /* we got to flush_old_exec() and failed after it */
1660                         read_unlock(&binfmt_lock);
1661                         force_sigsegv(SIGSEGV, current);
1662                         return retval;
1663                 }
1664                 if (retval != -ENOEXEC || !bprm->file) {
1665                         read_unlock(&binfmt_lock);
1666                         return retval;
1667                 }
1668         }
1669         read_unlock(&binfmt_lock);
1670
1671         if (need_retry) {
1672                 if (printable(bprm->buf[0]) && printable(bprm->buf[1]) &&
1673                     printable(bprm->buf[2]) && printable(bprm->buf[3]))
1674                         return retval;
1675                 if (request_module("binfmt-%04x", *(ushort *)(bprm->buf + 2)) < 0)
1676                         return retval;
1677                 need_retry = false;
1678                 goto retry;
1679         }
1680
1681         return retval;
1682 }
1683 EXPORT_SYMBOL(search_binary_handler);
1684
1685 static int exec_binprm(struct linux_binprm *bprm)
1686 {
1687         pid_t old_pid, old_vpid;
1688         int ret;
1689
1690         /* Need to fetch pid before load_binary changes it */
1691         old_pid = current->pid;
1692         rcu_read_lock();
1693         old_vpid = task_pid_nr_ns(current, task_active_pid_ns(current->parent));
1694         rcu_read_unlock();
1695
1696         ret = search_binary_handler(bprm);
1697         if (ret >= 0) {
1698                 audit_bprm(bprm);
1699                 trace_sched_process_exec(current, old_pid, bprm);
1700                 ptrace_event(PTRACE_EVENT_EXEC, old_vpid);
1701                 proc_exec_connector(current);
1702         }
1703
1704         return ret;
1705 }
1706
1707 /*
1708  * sys_execve() executes a new program.
1709  */
1710 static int __do_execve_file(int fd, struct filename *filename,
1711                             struct user_arg_ptr argv,
1712                             struct user_arg_ptr envp,
1713                             int flags, struct file *file)
1714 {
1715         char *pathbuf = NULL;
1716         struct linux_binprm *bprm;
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         if (!file)
1756                 file = do_open_execat(fd, filename, flags);
1757         retval = PTR_ERR(file);
1758         if (IS_ERR(file))
1759                 goto out_unmark;
1760
1761         sched_exec();
1762
1763         bprm->file = file;
1764         if (!filename) {
1765                 bprm->filename = "none";
1766         } else if (fd == AT_FDCWD || filename->name[0] == '/') {
1767                 bprm->filename = filename->name;
1768         } else {
1769                 if (filename->name[0] == '\0')
1770                         pathbuf = kasprintf(GFP_KERNEL, "/dev/fd/%d", fd);
1771                 else
1772                         pathbuf = kasprintf(GFP_KERNEL, "/dev/fd/%d/%s",
1773                                             fd, filename->name);
1774                 if (!pathbuf) {
1775                         retval = -ENOMEM;
1776                         goto out_unmark;
1777                 }
1778                 /*
1779                  * Record that a name derived from an O_CLOEXEC fd will be
1780                  * inaccessible after exec. Relies on having exclusive access to
1781                  * current->files (due to unshare_files above).
1782                  */
1783                 if (close_on_exec(fd, rcu_dereference_raw(current->files->fdt)))
1784                         bprm->interp_flags |= BINPRM_FLAGS_PATH_INACCESSIBLE;
1785                 bprm->filename = pathbuf;
1786         }
1787         bprm->interp = bprm->filename;
1788
1789         retval = bprm_mm_init(bprm);
1790         if (retval)
1791                 goto out_unmark;
1792
1793         bprm->argc = count(argv, MAX_ARG_STRINGS);
1794         if ((retval = bprm->argc) < 0)
1795                 goto out;
1796
1797         bprm->envc = count(envp, MAX_ARG_STRINGS);
1798         if ((retval = bprm->envc) < 0)
1799                 goto out;
1800
1801         retval = prepare_binprm(bprm);
1802         if (retval < 0)
1803                 goto out;
1804
1805         retval = copy_strings_kernel(1, &bprm->filename, bprm);
1806         if (retval < 0)
1807                 goto out;
1808
1809         bprm->exec = bprm->p;
1810         retval = copy_strings(bprm->envc, envp, bprm);
1811         if (retval < 0)
1812                 goto out;
1813
1814         retval = copy_strings(bprm->argc, argv, bprm);
1815         if (retval < 0)
1816                 goto out;
1817
1818         would_dump(bprm, bprm->file);
1819
1820         retval = exec_binprm(bprm);
1821         if (retval < 0)
1822                 goto out;
1823
1824         /* execve succeeded */
1825         current->fs->in_exec = 0;
1826         current->in_execve = 0;
1827         membarrier_execve(current);
1828         rseq_execve(current);
1829         acct_update_integrals(current);
1830         task_numa_free(current);
1831         free_bprm(bprm);
1832         kfree(pathbuf);
1833         if (filename)
1834                 putname(filename);
1835         if (displaced)
1836                 put_files_struct(displaced);
1837         return retval;
1838
1839 out:
1840         if (bprm->mm) {
1841                 acct_arg_size(bprm, 0);
1842                 mmput(bprm->mm);
1843         }
1844
1845 out_unmark:
1846         current->fs->in_exec = 0;
1847         current->in_execve = 0;
1848
1849 out_free:
1850         free_bprm(bprm);
1851         kfree(pathbuf);
1852
1853 out_files:
1854         if (displaced)
1855                 reset_files_struct(displaced);
1856 out_ret:
1857         if (filename)
1858                 putname(filename);
1859         return retval;
1860 }
1861
1862 static int do_execveat_common(int fd, struct filename *filename,
1863                               struct user_arg_ptr argv,
1864                               struct user_arg_ptr envp,
1865                               int flags)
1866 {
1867         return __do_execve_file(fd, filename, argv, envp, flags, NULL);
1868 }
1869
1870 int do_execve_file(struct file *file, void *__argv, void *__envp)
1871 {
1872         struct user_arg_ptr argv = { .ptr.native = __argv };
1873         struct user_arg_ptr envp = { .ptr.native = __envp };
1874
1875         return __do_execve_file(AT_FDCWD, NULL, argv, envp, 0, file);
1876 }
1877
1878 int do_execve(struct filename *filename,
1879         const char __user *const __user *__argv,
1880         const char __user *const __user *__envp)
1881 {
1882         struct user_arg_ptr argv = { .ptr.native = __argv };
1883         struct user_arg_ptr envp = { .ptr.native = __envp };
1884         return do_execveat_common(AT_FDCWD, filename, argv, envp, 0);
1885 }
1886
1887 int do_execveat(int fd, struct filename *filename,
1888                 const char __user *const __user *__argv,
1889                 const char __user *const __user *__envp,
1890                 int flags)
1891 {
1892         struct user_arg_ptr argv = { .ptr.native = __argv };
1893         struct user_arg_ptr envp = { .ptr.native = __envp };
1894
1895         return do_execveat_common(fd, filename, argv, envp, flags);
1896 }
1897
1898 #ifdef CONFIG_COMPAT
1899 static int compat_do_execve(struct filename *filename,
1900         const compat_uptr_t __user *__argv,
1901         const compat_uptr_t __user *__envp)
1902 {
1903         struct user_arg_ptr argv = {
1904                 .is_compat = true,
1905                 .ptr.compat = __argv,
1906         };
1907         struct user_arg_ptr envp = {
1908                 .is_compat = true,
1909                 .ptr.compat = __envp,
1910         };
1911         return do_execveat_common(AT_FDCWD, filename, argv, envp, 0);
1912 }
1913
1914 static int compat_do_execveat(int fd, struct filename *filename,
1915                               const compat_uptr_t __user *__argv,
1916                               const compat_uptr_t __user *__envp,
1917                               int flags)
1918 {
1919         struct user_arg_ptr argv = {
1920                 .is_compat = true,
1921                 .ptr.compat = __argv,
1922         };
1923         struct user_arg_ptr envp = {
1924                 .is_compat = true,
1925                 .ptr.compat = __envp,
1926         };
1927         return do_execveat_common(fd, filename, argv, envp, flags);
1928 }
1929 #endif
1930
1931 void set_binfmt(struct linux_binfmt *new)
1932 {
1933         struct mm_struct *mm = current->mm;
1934
1935         if (mm->binfmt)
1936                 module_put(mm->binfmt->module);
1937
1938         mm->binfmt = new;
1939         if (new)
1940                 __module_get(new->module);
1941 }
1942 EXPORT_SYMBOL(set_binfmt);
1943
1944 /*
1945  * set_dumpable stores three-value SUID_DUMP_* into mm->flags.
1946  */
1947 void set_dumpable(struct mm_struct *mm, int value)
1948 {
1949         unsigned long old, new;
1950
1951         if (WARN_ON((unsigned)value > SUID_DUMP_ROOT))
1952                 return;
1953
1954         do {
1955                 old = READ_ONCE(mm->flags);
1956                 new = (old & ~MMF_DUMPABLE_MASK) | value;
1957         } while (cmpxchg(&mm->flags, old, new) != old);
1958 }
1959
1960 SYSCALL_DEFINE3(execve,
1961                 const char __user *, filename,
1962                 const char __user *const __user *, argv,
1963                 const char __user *const __user *, envp)
1964 {
1965         return do_execve(getname(filename), argv, envp);
1966 }
1967
1968 SYSCALL_DEFINE5(execveat,
1969                 int, fd, const char __user *, filename,
1970                 const char __user *const __user *, argv,
1971                 const char __user *const __user *, envp,
1972                 int, flags)
1973 {
1974         int lookup_flags = (flags & AT_EMPTY_PATH) ? LOOKUP_EMPTY : 0;
1975
1976         return do_execveat(fd,
1977                            getname_flags(filename, lookup_flags, NULL),
1978                            argv, envp, flags);
1979 }
1980
1981 #ifdef CONFIG_COMPAT
1982 COMPAT_SYSCALL_DEFINE3(execve, const char __user *, filename,
1983         const compat_uptr_t __user *, argv,
1984         const compat_uptr_t __user *, envp)
1985 {
1986         return compat_do_execve(getname(filename), argv, envp);
1987 }
1988
1989 COMPAT_SYSCALL_DEFINE5(execveat, int, fd,
1990                        const char __user *, filename,
1991                        const compat_uptr_t __user *, argv,
1992                        const compat_uptr_t __user *, envp,
1993                        int,  flags)
1994 {
1995         int lookup_flags = (flags & AT_EMPTY_PATH) ? LOOKUP_EMPTY : 0;
1996
1997         return compat_do_execveat(fd,
1998                                   getname_flags(filename, lookup_flags, NULL),
1999                                   argv, envp, flags);
2000 }
2001 #endif