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/stat.h>
  30 #include <linux/fcntl.h>
  31 #include <linux/smp_lock.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/proc_fs.h>
  47 #include <linux/mount.h>
  48 #include <linux/security.h>
  49 #include <linux/syscalls.h>
  50 #include <linux/tsacct_kern.h>
  51 #include <linux/cn_proc.h>
  52 #include <linux/audit.h>
  53 #include <linux/tracehook.h>
  54 #include <linux/kmod.h>
  55 #include <linux/fsnotify.h>
  56 #include <linux/fs_struct.h>
  57 #include <linux/pipe_fs_i.h>
  58
  59 #include <asm/uaccess.h>
  60 #include <asm/mmu_context.h>
  61 #include <asm/tlb.h>
  62 #include "internal.h"
  63
  64 int core_uses_pid;
  65 char core_pattern[CORENAME_MAX_SIZE] = "core";
  66 unsigned int core_pipe_limit;
  67 int suid_dumpable = 0;
  68
  69 /* The maximal length of core_pattern is also specified in sysctl.c */
  70
  71 static LIST_HEAD(formats);
  72 static DEFINE_RWLOCK(binfmt_lock);
  73
  74 int __register_binfmt(struct linux_binfmt * fmt, int insert)
  75 {
  76         if (!fmt)
  77                 return -EINVAL;
  78         write_lock(&binfmt_lock);
  79         insert ? list_add(&fmt->lh, &formats) :
  80                  list_add_tail(&fmt->lh, &formats);
  81         write_unlock(&binfmt_lock);
  82         return 0;
  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 /*
 102  * Note that a shared library must be both readable and executable due to
 103  * security reasons.
 104  *
 105  * Also note that we take the address to load from from the file itself.
 106  */
 107 SYSCALL_DEFINE1(uselib, const char __user *, library)
 108 {
 109         struct file *file;
 110         char *tmp = getname(library);
 111         int error = PTR_ERR(tmp);
 112
 113         if (IS_ERR(tmp))
 114                 goto out;
 115
 116         file = do_filp_open(AT_FDCWD, tmp,
 117                                 O_LARGEFILE | O_RDONLY | FMODE_EXEC, 0,
 118                                 MAY_READ | MAY_EXEC | MAY_OPEN);
 119         putname(tmp);
 120         error = PTR_ERR(file);
 121         if (IS_ERR(file))
 122                 goto out;
 123
 124         error = -EINVAL;
 125         if (!S_ISREG(file->f_path.dentry->d_inode->i_mode))
 126                 goto exit;
 127
 128         error = -EACCES;
 129         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC)
 130                 goto exit;
 131
 132         fsnotify_open(file->f_path.dentry);
 133
 134         error = -ENOEXEC;
 135         if(file->f_op) {
 136                 struct linux_binfmt * fmt;
 137
 138                 read_lock(&binfmt_lock);
 139                 list_for_each_entry(fmt, &formats, lh) {
 140                         if (!fmt->load_shlib)
 141                                 continue;
 142                         if (!try_module_get(fmt->module))
 143                                 continue;
 144                         read_unlock(&binfmt_lock);
 145                         error = fmt->load_shlib(file);
 146                         read_lock(&binfmt_lock);
 147                         put_binfmt(fmt);
 148                         if (error != -ENOEXEC)
 149                                 break;
 150                 }
 151                 read_unlock(&binfmt_lock);
 152         }
 153 exit:
 154         fput(file);
 155 out:
 156         return error;
 157 }
 158
 159 #ifdef CONFIG_MMU
 160
 161 static struct page *get_arg_page(struct linux_binprm *bprm, unsigned long pos,
 162                 int write)
 163 {
 164         struct page *page;
 165         int ret;
 166
 167 #ifdef CONFIG_STACK_GROWSUP
 168         if (write) {
 169                 ret = expand_stack_downwards(bprm->vma, pos);
 170                 if (ret < 0)
 171                         return NULL;
 172         }
 173 #endif
 174         ret = get_user_pages(current, bprm->mm, pos,
 175                         1, write, 1, &page, NULL);
 176         if (ret <= 0)
 177                 return NULL;
 178
 179         if (write) {
 180                 unsigned long size = bprm->vma->vm_end - bprm->vma->vm_start;
 181                 struct rlimit *rlim;
 182
 183                 /*
 184                  * We've historically supported up to 32 pages (ARG_MAX)
 185                  * of argument strings even with small stacks
 186                  */
 187                 if (size <= ARG_MAX)
 188                         return page;
 189
 190                 /*
 191                  * Limit to 1/4-th the stack size for the argv+env strings.
 192                  * This ensures that:
 193                  *  - the remaining binfmt code will not run out of stack space,
 194                  *  - the program will have a reasonable amount of stack left
 195                  *    to work from.
 196                  */
 197                 rlim = current->signal->rlim;
 198                 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur) / 4) {
 199                         put_page(page);
 200                         return NULL;
 201                 }
 202         }
 203
 204         return page;
 205 }
 206
 207 static void put_arg_page(struct page *page)
 208 {
 209         put_page(page);
 210 }
 211
 212 static void free_arg_page(struct linux_binprm *bprm, int i)
 213 {
 214 }
 215
 216 static void free_arg_pages(struct linux_binprm *bprm)
 217 {
 218 }
 219
 220 static void flush_arg_page(struct linux_binprm *bprm, unsigned long pos,
 221                 struct page *page)
 222 {
 223         flush_cache_page(bprm->vma, pos, page_to_pfn(page));
 224 }
 225
 226 static int __bprm_mm_init(struct linux_binprm *bprm)
 227 {
 228         int err;
 229         struct vm_area_struct *vma = NULL;
 230         struct mm_struct *mm = bprm->mm;
 231
 232         bprm->vma = vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
 233         if (!vma)
 234                 return -ENOMEM;
 235
 236         down_write(&mm->mmap_sem);
 237         vma->vm_mm = mm;
 238
 239         /*
 240          * Place the stack at the largest stack address the architecture
 241          * supports. Later, we'll move this to an appropriate place. We don't
 242          * use STACK_TOP because that can depend on attributes which aren't
 243          * configured yet.
 244          */
 245         BUG_ON(VM_STACK_FLAGS & VM_STACK_INCOMPLETE_SETUP);
 246         vma->vm_end = STACK_TOP_MAX;
 247         vma->vm_start = vma->vm_end - PAGE_SIZE;
 248         vma->vm_flags = VM_STACK_FLAGS | VM_STACK_INCOMPLETE_SETUP;
 249         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
 250         INIT_LIST_HEAD(&vma->anon_vma_chain);
 251         err = insert_vm_struct(mm, vma);
 252         if (err)
 253                 goto err;
 254
 255         mm->stack_vm = mm->total_vm = 1;
 256         up_write(&mm->mmap_sem);
 257         bprm->p = vma->vm_end - sizeof(void *);
 258         return 0;
 259 err:
 260         up_write(&mm->mmap_sem);
 261         bprm->vma = NULL;
 262         kmem_cache_free(vm_area_cachep, vma);
 263         return err;
 264 }
 265
 266 static bool valid_arg_len(struct linux_binprm *bprm, long len)
 267 {
 268         return len <= MAX_ARG_STRLEN;
 269 }
 270
 271 #else
 272
 273 static struct page *get_arg_page(struct linux_binprm *bprm, unsigned long pos,
 274                 int write)
 275 {
 276         struct page *page;
 277
 278         page = bprm->page[pos / PAGE_SIZE];
 279         if (!page && write) {
 280                 page = alloc_page(GFP_HIGHUSER|__GFP_ZERO);
 281                 if (!page)
 282                         return NULL;
 283                 bprm->page[pos / PAGE_SIZE] = page;
 284         }
 285
 286         return page;
 287 }
 288
 289 static void put_arg_page(struct page *page)
 290 {
 291 }
 292
 293 static void free_arg_page(struct linux_binprm *bprm, int i)
 294 {
 295         if (bprm->page[i]) {
 296                 __free_page(bprm->page[i]);
 297                 bprm->page[i] = NULL;
 298         }
 299 }
 300
 301 static void free_arg_pages(struct linux_binprm *bprm)
 302 {
 303         int i;
 304
 305         for (i = 0; i < MAX_ARG_PAGES; i++)
 306                 free_arg_page(bprm, i);
 307 }
 308
 309 static void flush_arg_page(struct linux_binprm *bprm, unsigned long pos,
 310                 struct page *page)
 311 {
 312 }
 313
 314 static int __bprm_mm_init(struct linux_binprm *bprm)
 315 {
 316         bprm->p = PAGE_SIZE * MAX_ARG_PAGES - sizeof(void *);
 317         return 0;
 318 }
 319
 320 static bool valid_arg_len(struct linux_binprm *bprm, long len)
 321 {
 322         return len <= bprm->p;
 323 }
 324
 325 #endif /* CONFIG_MMU */
 326
 327 /*
 328  * Create a new mm_struct and populate it with a temporary stack
 329  * vm_area_struct.  We don't have enough context at this point to set the stack
 330  * flags, permissions, and offset, so we use temporary values.  We'll update
 331  * them later in setup_arg_pages().
 332  */
 333 int bprm_mm_init(struct linux_binprm *bprm)
 334 {
 335         int err;
 336         struct mm_struct *mm = NULL;
 337
 338         bprm->mm = mm = mm_alloc();
 339         err = -ENOMEM;
 340         if (!mm)
 341                 goto err;
 342
 343         err = init_new_context(current, mm);
 344         if (err)
 345                 goto err;
 346
 347         err = __bprm_mm_init(bprm);
 348         if (err)
 349                 goto err;
 350
 351         return 0;
 352
 353 err:
 354         if (mm) {
 355                 bprm->mm = NULL;
 356                 mmdrop(mm);
 357         }
 358
 359         return err;
 360 }
 361
 362 /*
 363  * count() counts the number of strings in array ARGV.
 364  */
 365 static int count(char __user * __user * argv, int max)
 366 {
 367         int i = 0;
 368
 369         if (argv != NULL) {
 370                 for (;;) {
 371                         char __user * p;
 372
 373                         if (get_user(p, argv))
 374                                 return -EFAULT;
 375                         if (!p)
 376                                 break;
 377                         argv++;
 378                         if (i++ >= max)
 379                                 return -E2BIG;
 380                         cond_resched();
 381                 }
 382         }
 383         return i;
 384 }
 385
 386 /*
 387  * 'copy_strings()' copies argument/environment strings from the old
 388  * processes's memory to the new process's stack.  The call to get_user_pages()
 389  * ensures the destination page is created and not swapped out.
 390  */
 391 static int copy_strings(int argc, char __user * __user * argv,
 392                         struct linux_binprm *bprm)
 393 {
 394         struct page *kmapped_page = NULL;
 395         char *kaddr = NULL;
 396         unsigned long kpos = 0;
 397         int ret;
 398
 399         while (argc-- > 0) {
 400                 char __user *str;
 401                 int len;
 402                 unsigned long pos;
 403
 404                 if (get_user(str, argv+argc) ||
 405                                 !(len = strnlen_user(str, MAX_ARG_STRLEN))) {
 406                         ret = -EFAULT;
 407                         goto out;
 408                 }
 409
 410                 if (!valid_arg_len(bprm, len)) {
 411                         ret = -E2BIG;
 412                         goto out;
 413                 }
 414
 415                 /* We're going to work our way backwords. */
 416                 pos = bprm->p;
 417                 str += len;
 418                 bprm->p -= len;
 419
 420                 while (len > 0) {
 421                         int offset, bytes_to_copy;
 422
 423                         offset = pos % PAGE_SIZE;
 424                         if (offset == 0)
 425                                 offset = PAGE_SIZE;
 426
 427                         bytes_to_copy = offset;
 428                         if (bytes_to_copy > len)
 429                                 bytes_to_copy = len;
 430
 431                         offset -= bytes_to_copy;
 432                         pos -= bytes_to_copy;
 433                         str -= bytes_to_copy;
 434                         len -= bytes_to_copy;
 435
 436                         if (!kmapped_page || kpos != (pos & PAGE_MASK)) {
 437                                 struct page *page;
 438
 439                                 page = get_arg_page(bprm, pos, 1);
 440                                 if (!page) {
 441                                         ret = -E2BIG;
 442                                         goto out;
 443                                 }
 444
 445                                 if (kmapped_page) {
 446                                         flush_kernel_dcache_page(kmapped_page);
 447                                         kunmap(kmapped_page);
 448                                         put_arg_page(kmapped_page);
 449                                 }
 450                                 kmapped_page = page;
 451                                 kaddr = kmap(kmapped_page);
 452                                 kpos = pos & PAGE_MASK;
 453                                 flush_arg_page(bprm, kpos, kmapped_page);
 454                         }
 455                         if (copy_from_user(kaddr+offset, str, bytes_to_copy)) {
 456                                 ret = -EFAULT;
 457                                 goto out;
 458                         }
 459                 }
 460         }
 461         ret = 0;
 462 out:
 463         if (kmapped_page) {
 464                 flush_kernel_dcache_page(kmapped_page);
 465                 kunmap(kmapped_page);
 466                 put_arg_page(kmapped_page);
 467         }
 468         return ret;
 469 }
 470
 471 /*
 472  * Like copy_strings, but get argv and its values from kernel memory.
 473  */
 474 int copy_strings_kernel(int argc,char ** argv, struct linux_binprm *bprm)
 475 {
 476         int r;
 477         mm_segment_t oldfs = get_fs();
 478         set_fs(KERNEL_DS);
 479         r = copy_strings(argc, (char __user * __user *)argv, bprm);
 480         set_fs(oldfs);
 481         return r;
 482 }
 483 EXPORT_SYMBOL(copy_strings_kernel);
 484
 485 #ifdef CONFIG_MMU
 486
 487 /*
 488  * During bprm_mm_init(), we create a temporary stack at STACK_TOP_MAX.  Once
 489  * the binfmt code determines where the new stack should reside, we shift it to
 490  * its final location.  The process proceeds as follows:
 491  *
 492  * 1) Use shift to calculate the new vma endpoints.
 493  * 2) Extend vma to cover both the old and new ranges.  This ensures the
 494  *    arguments passed to subsequent functions are consistent.
 495  * 3) Move vma's page tables to the new range.
 496  * 4) Free up any cleared pgd range.
 497  * 5) Shrink the vma to cover only the new range.
 498  */
 499 static int shift_arg_pages(struct vm_area_struct *vma, unsigned long shift)
 500 {
 501         struct mm_struct *mm = vma->vm_mm;
 502         unsigned long old_start = vma->vm_start;
 503         unsigned long old_end = vma->vm_end;
 504         unsigned long length = old_end - old_start;
 505         unsigned long new_start = old_start - shift;
 506         unsigned long new_end = old_end - shift;
 507         struct mmu_gather *tlb;
 508
 509         BUG_ON(new_start > new_end);
 510
 511         /*
 512          * ensure there are no vmas between where we want to go
 513          * and where we are
 514          */
 515         if (vma != find_vma(mm, new_start))
 516                 return -EFAULT;
 517
 518         /*
 519          * cover the whole range: [new_start, old_end)
 520          */
 521         if (vma_adjust(vma, new_start, old_end, vma->vm_pgoff, NULL))
 522                 return -ENOMEM;
 523
 524         /*
 525          * move the page tables downwards, on failure we rely on
 526          * process cleanup to remove whatever mess we made.
 527          */
 528         if (length != move_page_tables(vma, old_start,
 529                                        vma, new_start, length))
 530                 return -ENOMEM;
 531
 532         lru_add_drain();
 533         tlb = tlb_gather_mmu(mm, 0);
 534         if (new_end > old_start) {
 535                 /*
 536                  * when the old and new regions overlap clear from new_end.
 537                  */
 538                 free_pgd_range(tlb, new_end, old_end, new_end,
 539                         vma->vm_next ? vma->vm_next->vm_start : 0);
 540         } else {
 541                 /*
 542                  * otherwise, clean from old_start; this is done to not touch
 543                  * the address space in [new_end, old_start) some architectures
 544                  * have constraints on va-space that make this illegal (IA64) -
 545                  * for the others its just a little faster.
 546                  */
 547                 free_pgd_range(tlb, old_start, old_end, new_end,
 548                         vma->vm_next ? vma->vm_next->vm_start : 0);
 549         }
 550         tlb_finish_mmu(tlb, new_end, old_end);
 551
 552         /*
 553          * Shrink the vma to just the new range.  Always succeeds.
 554          */
 555         vma_adjust(vma, new_start, new_end, vma->vm_pgoff, NULL);
 556
 557         return 0;
 558 }
 559
 560 /*
 561  * Finalizes the stack vm_area_struct. The flags and permissions are updated,
 562  * the stack is optionally relocated, and some extra space is added.
 563  */
 564 int setup_arg_pages(struct linux_binprm *bprm,
 565                     unsigned long stack_top,
 566                     int executable_stack)
 567 {
 568         unsigned long ret;
 569         unsigned long stack_shift;
 570         struct mm_struct *mm = current->mm;
 571         struct vm_area_struct *vma = bprm->vma;
 572         struct vm_area_struct *prev = NULL;
 573         unsigned long vm_flags;
 574         unsigned long stack_base;
 575         unsigned long stack_size;
 576         unsigned long stack_expand;
 577         unsigned long rlim_stack;
 578
 579 #ifdef CONFIG_STACK_GROWSUP
 580         /* Limit stack size to 1GB */
 581         stack_base = rlimit_max(RLIMIT_STACK);
 582         if (stack_base > (1 << 30))
 583                 stack_base = 1 << 30;
 584
 585         /* Make sure we didn't let the argument array grow too large. */
 586         if (vma->vm_end - vma->vm_start > stack_base)
 587                 return -ENOMEM;
 588
 589         stack_base = PAGE_ALIGN(stack_top - stack_base);
 590
 591         stack_shift = vma->vm_start - stack_base;
 592         mm->arg_start = bprm->p - stack_shift;
 593         bprm->p = vma->vm_end - stack_shift;
 594 #else
 595         stack_top = arch_align_stack(stack_top);
 596         stack_top = PAGE_ALIGN(stack_top);
 597         stack_shift = vma->vm_end - stack_top;
 598
 599         bprm->p -= stack_shift;
 600         mm->arg_start = bprm->p;
 601 #endif
 602
 603         if (bprm->loader)
 604                 bprm->loader -= stack_shift;
 605         bprm->exec -= stack_shift;
 606
 607         down_write(&mm->mmap_sem);
 608         vm_flags = VM_STACK_FLAGS;
 609
 610         /*
 611          * Adjust stack execute permissions; explicitly enable for
 612          * EXSTACK_ENABLE_X, disable for EXSTACK_DISABLE_X and leave alone
 613          * (arch default) otherwise.
 614          */
 615         if (unlikely(executable_stack == EXSTACK_ENABLE_X))
 616                 vm_flags |= VM_EXEC;
 617         else if (executable_stack == EXSTACK_DISABLE_X)
 618                 vm_flags &= ~VM_EXEC;
 619         vm_flags |= mm->def_flags;
 620         vm_flags |= VM_STACK_INCOMPLETE_SETUP;
 621
 622         ret = mprotect_fixup(vma, &prev, vma->vm_start, vma->vm_end,
 623                         vm_flags);
 624         if (ret)
 625                 goto out_unlock;
 626         BUG_ON(prev != vma);
 627
 628         /* Move stack pages down in memory. */
 629         if (stack_shift) {
 630                 ret = shift_arg_pages(vma, stack_shift);
 631                 if (ret)
 632                         goto out_unlock;
 633         }
 634
 635         /* mprotect_fixup is overkill to remove the temporary stack flags */
 636         vma->vm_flags &= ~VM_STACK_INCOMPLETE_SETUP;
 637
 638         stack_expand = 131072UL; /* randomly 32*4k (or 2*64k) pages */
 639         stack_size = vma->vm_end - vma->vm_start;
 640         /*
 641          * Align this down to a page boundary as expand_stack
 642          * will align it up.
 643          */
 644         rlim_stack = rlimit(RLIMIT_STACK) & PAGE_MASK;
 645 #ifdef CONFIG_STACK_GROWSUP
 646         if (stack_size + stack_expand > rlim_stack)
 647                 stack_base = vma->vm_start + rlim_stack;
 648         else
 649                 stack_base = vma->vm_end + stack_expand;
 650 #else
 651         if (stack_size + stack_expand > rlim_stack)
 652                 stack_base = vma->vm_end - rlim_stack;
 653         else
 654                 stack_base = vma->vm_start - stack_expand;
 655 #endif
 656         ret = expand_stack(vma, stack_base);
 657         if (ret)
 658                 ret = -EFAULT;
 659
 660 out_unlock:
 661         up_write(&mm->mmap_sem);
 662         return ret;
 663 }
 664 EXPORT_SYMBOL(setup_arg_pages);
 665
 666 #endif /* CONFIG_MMU */
 667
 668 struct file *open_exec(const char *name)
 669 {
 670         struct file *file;
 671         int err;
 672
 673         file = do_filp_open(AT_FDCWD, name,
 674                                 O_LARGEFILE | O_RDONLY | FMODE_EXEC, 0,
 675                                 MAY_EXEC | MAY_OPEN);
 676         if (IS_ERR(file))
 677                 goto out;
 678
 679         err = -EACCES;
 680         if (!S_ISREG(file->f_path.dentry->d_inode->i_mode))
 681                 goto exit;
 682
 683         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC)
 684                 goto exit;
 685
 686         fsnotify_open(file->f_path.dentry);
 687
 688         err = deny_write_access(file);
 689         if (err)
 690                 goto exit;
 691
 692 out:
 693         return file;
 694
 695 exit:
 696         fput(file);
 697         return ERR_PTR(err);
 698 }
 699 EXPORT_SYMBOL(open_exec);
 700
 701 int kernel_read(struct file *file, loff_t offset,
 702                 char *addr, unsigned long count)
 703 {
 704         mm_segment_t old_fs;
 705         loff_t pos = offset;
 706         int result;
 707
 708         old_fs = get_fs();
 709         set_fs(get_ds());
 710         /* The cast to a user pointer is valid due to the set_fs() */
 711         result = vfs_read(file, (void __user *)addr, count, &pos);
 712         set_fs(old_fs);
 713         return result;
 714 }
 715
 716 EXPORT_SYMBOL(kernel_read);
 717
 718 static int exec_mmap(struct mm_struct *mm)
 719 {
 720         struct task_struct *tsk;
 721         struct mm_struct * old_mm, *active_mm;
 722
 723         /* Notify parent that we're no longer interested in the old VM */
 724         tsk = current;
 725         old_mm = current->mm;
 726         sync_mm_rss(tsk, old_mm);
 727         mm_release(tsk, old_mm);
 728
 729         if (old_mm) {
 730                 /*
 731                  * Make sure that if there is a core dump in progress
 732                  * for the old mm, we get out and die instead of going
 733                  * through with the exec.  We must hold mmap_sem around
 734                  * checking core_state and changing tsk->mm.
 735                  */
 736                 down_read(&old_mm->mmap_sem);
 737                 if (unlikely(old_mm->core_state)) {
 738                         up_read(&old_mm->mmap_sem);
 739                         return -EINTR;
 740                 }
 741         }
 742         task_lock(tsk);
 743         active_mm = tsk->active_mm;
 744         tsk->mm = mm;
 745         tsk->active_mm = mm;
 746         activate_mm(active_mm, mm);
 747         task_unlock(tsk);
 748         arch_pick_mmap_layout(mm);
 749         if (old_mm) {
 750                 up_read(&old_mm->mmap_sem);
 751                 BUG_ON(active_mm != old_mm);
 752                 mm_update_next_owner(old_mm);
 753                 mmput(old_mm);
 754                 return 0;
 755         }
 756         mmdrop(active_mm);
 757         return 0;
 758 }
 759
 760 /*
 761  * This function makes sure the current process has its own signal table,
 762  * so that flush_signal_handlers can later reset the handlers without
 763  * disturbing other processes.  (Other processes might share the signal
 764  * table via the CLONE_SIGHAND option to clone().)
 765  */
 766 static int de_thread(struct task_struct *tsk)
 767 {
 768         struct signal_struct *sig = tsk->signal;
 769         struct sighand_struct *oldsighand = tsk->sighand;
 770         spinlock_t *lock = &oldsighand->siglock;
 771         int count;
 772
 773         if (thread_group_empty(tsk))
 774                 goto no_thread_group;
 775
 776         /*
 777          * Kill all other threads in the thread group.
 778          */
 779         spin_lock_irq(lock);
 780         if (signal_group_exit(sig)) {
 781                 /*
 782                  * Another group action in progress, just
 783                  * return so that the signal is processed.
 784                  */
 785                 spin_unlock_irq(lock);
 786                 return -EAGAIN;
 787         }
 788         sig->group_exit_task = tsk;
 789         zap_other_threads(tsk);
 790
 791         /* Account for the thread group leader hanging around: */
 792         count = thread_group_leader(tsk) ? 1 : 2;
 793         sig->notify_count = count;
 794         while (atomic_read(&sig->count) > count) {
 795                 __set_current_state(TASK_UNINTERRUPTIBLE);
 796                 spin_unlock_irq(lock);
 797                 schedule();
 798                 spin_lock_irq(lock);
 799         }
 800         spin_unlock_irq(lock);
 801
 802         /*
 803          * At this point all other threads have exited, all we have to
 804          * do is to wait for the thread group leader to become inactive,
 805          * and to assume its PID:
 806          */
 807         if (!thread_group_leader(tsk)) {
 808                 struct task_struct *leader = tsk->group_leader;
 809
 810                 sig->notify_count = -1; /* for exit_notify() */
 811                 for (;;) {
 812                         write_lock_irq(&tasklist_lock);
 813                         if (likely(leader->exit_state))
 814                                 break;
 815                         __set_current_state(TASK_UNINTERRUPTIBLE);
 816                         write_unlock_irq(&tasklist_lock);
 817                         schedule();
 818                 }
 819
 820                 /*
 821                  * The only record we have of the real-time age of a
 822                  * process, regardless of execs it's done, is start_time.
 823                  * All the past CPU time is accumulated in signal_struct
 824                  * from sister threads now dead.  But in this non-leader
 825                  * exec, nothing survives from the original leader thread,
 826                  * whose birth marks the true age of this process now.
 827                  * When we take on its identity by switching to its PID, we
 828                  * also take its birthdate (always earlier than our own).
 829                  */
 830                 tsk->start_time = leader->start_time;
 831
 832                 BUG_ON(!same_thread_group(leader, tsk));
 833                 BUG_ON(has_group_leader_pid(tsk));
 834                 /*
 835                  * An exec() starts a new thread group with the
 836                  * TGID of the previous thread group. Rehash the
 837                  * two threads with a switched PID, and release
 838                  * the former thread group leader:
 839                  */
 840
 841                 /* Become a process group leader with the old leader's pid.
 842                  * The old leader becomes a thread of the this thread group.
 843                  * Note: The old leader also uses this pid until release_task
 844                  *       is called.  Odd but simple and correct.
 845                  */
 846                 detach_pid(tsk, PIDTYPE_PID);
 847                 tsk->pid = leader->pid;
 848                 attach_pid(tsk, PIDTYPE_PID,  task_pid(leader));
 849                 transfer_pid(leader, tsk, PIDTYPE_PGID);
 850                 transfer_pid(leader, tsk, PIDTYPE_SID);
 851
 852                 list_replace_rcu(&leader->tasks, &tsk->tasks);
 853                 list_replace_init(&leader->sibling, &tsk->sibling);
 854
 855                 tsk->group_leader = tsk;
 856                 leader->group_leader = tsk;
 857
 858                 tsk->exit_signal = SIGCHLD;
 859
 860                 BUG_ON(leader->exit_state != EXIT_ZOMBIE);
 861                 leader->exit_state = EXIT_DEAD;
 862                 write_unlock_irq(&tasklist_lock);
 863
 864                 release_task(leader);
 865         }
 866
 867         sig->group_exit_task = NULL;
 868         sig->notify_count = 0;
 869
 870 no_thread_group:
 871         if (current->mm)
 872                 setmax_mm_hiwater_rss(&sig->maxrss, current->mm);
 873
 874         exit_itimers(sig);
 875         flush_itimer_signals();
 876
 877         if (atomic_read(&oldsighand->count) != 1) {
 878                 struct sighand_struct *newsighand;
 879                 /*
 880                  * This ->sighand is shared with the CLONE_SIGHAND
 881                  * but not CLONE_THREAD task, switch to the new one.
 882                  */
 883                 newsighand = kmem_cache_alloc(sighand_cachep, GFP_KERNEL);
 884                 if (!newsighand)
 885                         return -ENOMEM;
 886
 887                 atomic_set(&newsighand->count, 1);
 888                 memcpy(newsighand->action, oldsighand->action,
 889                        sizeof(newsighand->action));
 890
 891                 write_lock_irq(&tasklist_lock);
 892                 spin_lock(&oldsighand->siglock);
 893                 rcu_assign_pointer(tsk->sighand, newsighand);
 894                 spin_unlock(&oldsighand->siglock);
 895                 write_unlock_irq(&tasklist_lock);
 896
 897                 __cleanup_sighand(oldsighand);
 898         }
 899
 900         BUG_ON(!thread_group_leader(tsk));
 901         return 0;
 902 }
 903
 904 /*
 905  * These functions flushes out all traces of the currently running executable
 906  * so that a new one can be started
 907  */
 908 static void flush_old_files(struct files_struct * files)
 909 {
 910         long j = -1;
 911         struct fdtable *fdt;
 912
 913         spin_lock(&files->file_lock);
 914         for (;;) {
 915                 unsigned long set, i;
 916
 917                 j++;
 918                 i = j * __NFDBITS;
 919                 fdt = files_fdtable(files);
 920                 if (i >= fdt->max_fds)
 921                         break;
 922                 set = fdt->close_on_exec->fds_bits[j];
 923                 if (!set)
 924                         continue;
 925                 fdt->close_on_exec->fds_bits[j] = 0;
 926                 spin_unlock(&files->file_lock);
 927                 for ( ; set ; i++,set >>= 1) {
 928                         if (set & 1) {
 929                                 sys_close(i);
 930                         }
 931                 }
 932                 spin_lock(&files->file_lock);
 933
 934         }
 935         spin_unlock(&files->file_lock);
 936 }
 937
 938 char *get_task_comm(char *buf, struct task_struct *tsk)
 939 {
 940         /* buf must be at least sizeof(tsk->comm) in size */
 941         task_lock(tsk);
 942         strncpy(buf, tsk->comm, sizeof(tsk->comm));
 943         task_unlock(tsk);
 944         return buf;
 945 }
 946
 947 void set_task_comm(struct task_struct *tsk, char *buf)
 948 {
 949         task_lock(tsk);
 950
 951         /*
 952          * Threads may access current->comm without holding
 953          * the task lock, so write the string carefully.
 954          * Readers without a lock may see incomplete new
 955          * names but are safe from non-terminating string reads.
 956          */
 957         memset(tsk->comm, 0, TASK_COMM_LEN);
 958         wmb();
 959         strlcpy(tsk->comm, buf, sizeof(tsk->comm));
 960         task_unlock(tsk);
 961         perf_event_comm(tsk);
 962 }
 963
 964 int flush_old_exec(struct linux_binprm * bprm)
 965 {
 966         int retval;
 967
 968         /*
 969          * Make sure we have a private signal table and that
 970          * we are unassociated from the previous thread group.
 971          */
 972         retval = de_thread(current);
 973         if (retval)
 974                 goto out;
 975
 976         set_mm_exe_file(bprm->mm, bprm->file);
 977
 978         /*
 979          * Release all of the old mmap stuff
 980          */
 981         retval = exec_mmap(bprm->mm);
 982         if (retval)
 983                 goto out;
 984
 985         bprm->mm = NULL;                /* We're using it now */
 986
 987         current->flags &= ~PF_RANDOMIZE;
 988         flush_thread();
 989         current->personality &= ~bprm->per_clear;
 990
 991         return 0;
 992
 993 out:
 994         return retval;
 995 }
 996 EXPORT_SYMBOL(flush_old_exec);
 997
 998 void setup_new_exec(struct linux_binprm * bprm)
 999 {
1000         int i, ch;
1001         char * name;
1002         char tcomm[sizeof(current->comm)];
1003
1004         arch_pick_mmap_layout(current->mm);
1005
1006         /* This is the point of no return */
1007         current->sas_ss_sp = current->sas_ss_size = 0;
1008
1009         if (current_euid() == current_uid() && current_egid() == current_gid())
1010                 set_dumpable(current->mm, 1);
1011         else
1012                 set_dumpable(current->mm, suid_dumpable);
1013
1014         name = bprm->filename;
1015
1016         /* Copies the binary name from after last slash */
1017         for (i=0; (ch = *(name++)) != '\0';) {
1018                 if (ch == '/')
1019                         i = 0; /* overwrite what we wrote */
1020                 else
1021                         if (i < (sizeof(tcomm) - 1))
1022                                 tcomm[i++] = ch;
1023         }
1024         tcomm[i] = '\0';
1025         set_task_comm(current, tcomm);
1026
1027         /* Set the new mm task size. We have to do that late because it may
1028          * depend on TIF_32BIT which is only updated in flush_thread() on
1029          * some architectures like powerpc
1030          */
1031         current->mm->task_size = TASK_SIZE;
1032
1033         /* install the new credentials */
1034         if (bprm->cred->uid != current_euid() ||
1035             bprm->cred->gid != current_egid()) {
1036                 current->pdeath_signal = 0;
1037         } else if (file_permission(bprm->file, MAY_READ) ||
1038                    bprm->interp_flags & BINPRM_FLAGS_ENFORCE_NONDUMP) {
1039                 set_dumpable(current->mm, suid_dumpable);
1040         }
1041
1042         /*
1043          * Flush performance counters when crossing a
1044          * security domain:
1045          */
1046         if (!get_dumpable(current->mm))
1047                 perf_event_exit_task(current);
1048
1049         /* An exec changes our domain. We are no longer part of the thread
1050            group */
1051
1052         current->self_exec_id++;
1053
1054         flush_signal_handlers(current, 0);
1055         flush_old_files(current->files);
1056 }
1057 EXPORT_SYMBOL(setup_new_exec);
1058
1059 /*
1060  * Prepare credentials and lock ->cred_guard_mutex.
1061  * install_exec_creds() commits the new creds and drops the lock.
1062  * Or, if exec fails before, free_bprm() should release ->cred and
1063  * and unlock.
1064  */
1065 int prepare_bprm_creds(struct linux_binprm *bprm)
1066 {
1067         if (mutex_lock_interruptible(&current->cred_guard_mutex))
1068                 return -ERESTARTNOINTR;
1069
1070         bprm->cred = prepare_exec_creds();
1071         if (likely(bprm->cred))
1072                 return 0;
1073
1074         mutex_unlock(&current->cred_guard_mutex);
1075         return -ENOMEM;
1076 }
1077
1078 void free_bprm(struct linux_binprm *bprm)
1079 {
1080         free_arg_pages(bprm);
1081         if (bprm->cred) {
1082                 mutex_unlock(&current->cred_guard_mutex);
1083                 abort_creds(bprm->cred);
1084         }
1085         kfree(bprm);
1086 }
1087
1088 /*
1089  * install the new credentials for this executable
1090  */
1091 void install_exec_creds(struct linux_binprm *bprm)
1092 {
1093         security_bprm_committing_creds(bprm);
1094
1095         commit_creds(bprm->cred);
1096         bprm->cred = NULL;
1097         /*
1098          * cred_guard_mutex must be held at least to this point to prevent
1099          * ptrace_attach() from altering our determination of the task's
1100          * credentials; any time after this it may be unlocked.
1101          */
1102         security_bprm_committed_creds(bprm);
1103         mutex_unlock(&current->cred_guard_mutex);
1104 }
1105 EXPORT_SYMBOL(install_exec_creds);
1106
1107 /*
1108  * determine how safe it is to execute the proposed program
1109  * - the caller must hold current->cred_guard_mutex to protect against
1110  *   PTRACE_ATTACH
1111  */
1112 int check_unsafe_exec(struct linux_binprm *bprm)
1113 {
1114         struct task_struct *p = current, *t;
1115         unsigned n_fs;
1116         int res = 0;
1117
1118         bprm->unsafe = tracehook_unsafe_exec(p);
1119
1120         n_fs = 1;
1121         write_lock(&p->fs->lock);
1122         rcu_read_lock();
1123         for (t = next_thread(p); t != p; t = next_thread(t)) {
1124                 if (t->fs == p->fs)
1125                         n_fs++;
1126         }
1127         rcu_read_unlock();
1128
1129         if (p->fs->users > n_fs) {
1130                 bprm->unsafe |= LSM_UNSAFE_SHARE;
1131         } else {
1132                 res = -EAGAIN;
1133                 if (!p->fs->in_exec) {
1134                         p->fs->in_exec = 1;
1135                         res = 1;
1136                 }
1137         }
1138         write_unlock(&p->fs->lock);
1139
1140         return res;
1141 }
1142
1143 /*
1144  * Fill the binprm structure from the inode.
1145  * Check permissions, then read the first 128 (BINPRM_BUF_SIZE) bytes
1146  *
1147  * This may be called multiple times for binary chains (scripts for example).
1148  */
1149 int prepare_binprm(struct linux_binprm *bprm)
1150 {
1151         umode_t mode;
1152         struct inode * inode = bprm->file->f_path.dentry->d_inode;
1153         int retval;
1154
1155         mode = inode->i_mode;
1156         if (bprm->file->f_op == NULL)
1157                 return -EACCES;
1158
1159         /* clear any previous set[ug]id data from a previous binary */
1160         bprm->cred->euid = current_euid();
1161         bprm->cred->egid = current_egid();
1162
1163         if (!(bprm->file->f_path.mnt->mnt_flags & MNT_NOSUID)) {
1164                 /* Set-uid? */
1165                 if (mode & S_ISUID) {
1166                         bprm->per_clear |= PER_CLEAR_ON_SETID;
1167                         bprm->cred->euid = inode->i_uid;
1168                 }
1169
1170                 /* Set-gid? */
1171                 /*
1172                  * If setgid is set but no group execute bit then this
1173                  * is a candidate for mandatory locking, not a setgid
1174                  * executable.
1175                  */
1176                 if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP)) {
1177                         bprm->per_clear |= PER_CLEAR_ON_SETID;
1178                         bprm->cred->egid = inode->i_gid;
1179                 }
1180         }
1181
1182         /* fill in binprm security blob */
1183         retval = security_bprm_set_creds(bprm);
1184         if (retval)
1185                 return retval;
1186         bprm->cred_prepared = 1;
1187
1188         memset(bprm->buf, 0, BINPRM_BUF_SIZE);
1189         return kernel_read(bprm->file, 0, bprm->buf, BINPRM_BUF_SIZE);
1190 }
1191
1192 EXPORT_SYMBOL(prepare_binprm);
1193
1194 /*
1195  * Arguments are '\0' separated strings found at the location bprm->p
1196  * points to; chop off the first by relocating brpm->p to right after
1197  * the first '\0' encountered.
1198  */
1199 int remove_arg_zero(struct linux_binprm *bprm)
1200 {
1201         int ret = 0;
1202         unsigned long offset;
1203         char *kaddr;
1204         struct page *page;
1205
1206         if (!bprm->argc)
1207                 return 0;
1208
1209         do {
1210                 offset = bprm->p & ~PAGE_MASK;
1211                 page = get_arg_page(bprm, bprm->p, 0);
1212                 if (!page) {
1213                         ret = -EFAULT;
1214                         goto out;
1215                 }
1216                 kaddr = kmap_atomic(page, KM_USER0);
1217
1218                 for (; offset < PAGE_SIZE && kaddr[offset];
1219                                 offset++, bprm->p++)
1220                         ;
1221
1222                 kunmap_atomic(kaddr, KM_USER0);
1223                 put_arg_page(page);
1224
1225                 if (offset == PAGE_SIZE)
1226                         free_arg_page(bprm, (bprm->p >> PAGE_SHIFT) - 1);
1227         } while (offset == PAGE_SIZE);
1228
1229         bprm->p++;
1230         bprm->argc--;
1231         ret = 0;
1232
1233 out:
1234         return ret;
1235 }
1236 EXPORT_SYMBOL(remove_arg_zero);
1237
1238 /*
1239  * cycle the list of binary formats handler, until one recognizes the image
1240  */
1241 int search_binary_handler(struct linux_binprm *bprm,struct pt_regs *regs)
1242 {
1243         unsigned int depth = bprm->recursion_depth;
1244         int try,retval;
1245         struct linux_binfmt *fmt;
1246
1247         retval = security_bprm_check(bprm);
1248         if (retval)
1249                 return retval;
1250
1251         /* kernel module loader fixup */
1252         /* so we don't try to load run modprobe in kernel space. */
1253         set_fs(USER_DS);
1254
1255         retval = audit_bprm(bprm);
1256         if (retval)
1257                 return retval;
1258
1259         retval = -ENOENT;
1260         for (try=0; try<2; try++) {
1261                 read_lock(&binfmt_lock);
1262                 list_for_each_entry(fmt, &formats, lh) {
1263                         int (*fn)(struct linux_binprm *, struct pt_regs *) = fmt->load_binary;
1264                         if (!fn)
1265                                 continue;
1266                         if (!try_module_get(fmt->module))
1267                                 continue;
1268                         read_unlock(&binfmt_lock);
1269                         retval = fn(bprm, regs);
1270                         /*
1271                          * Restore the depth counter to its starting value
1272                          * in this call, so we don't have to rely on every
1273                          * load_binary function to restore it on return.
1274                          */
1275                         bprm->recursion_depth = depth;
1276                         if (retval >= 0) {
1277                                 if (depth == 0)
1278                                         tracehook_report_exec(fmt, bprm, regs);
1279                                 put_binfmt(fmt);
1280                                 allow_write_access(bprm->file);
1281                                 if (bprm->file)
1282                                         fput(bprm->file);
1283                                 bprm->file = NULL;
1284                                 current->did_exec = 1;
1285                                 proc_exec_connector(current);
1286                                 return retval;
1287                         }
1288                         read_lock(&binfmt_lock);
1289                         put_binfmt(fmt);
1290                         if (retval != -ENOEXEC || bprm->mm == NULL)
1291                                 break;
1292                         if (!bprm->file) {
1293                                 read_unlock(&binfmt_lock);
1294                                 return retval;
1295                         }
1296                 }
1297                 read_unlock(&binfmt_lock);
1298                 if (retval != -ENOEXEC || bprm->mm == NULL) {
1299                         break;
1300 #ifdef CONFIG_MODULES
1301                 } else {
1302 #define printable(c) (((c)=='\t') || ((c)=='\n') || (0x20<=(c) && (c)<=0x7e))
1303                         if (printable(bprm->buf[0]) &&
1304                             printable(bprm->buf[1]) &&
1305                             printable(bprm->buf[2]) &&
1306                             printable(bprm->buf[3]))
1307                                 break; /* -ENOEXEC */
1308                         request_module("binfmt-%04x", *(unsigned short *)(&bprm->buf[2]));
1309 #endif
1310                 }
1311         }
1312         return retval;
1313 }
1314
1315 EXPORT_SYMBOL(search_binary_handler);
1316
1317 /*
1318  * sys_execve() executes a new program.
1319  */
1320 int do_execve(char * filename,
1321         char __user *__user *argv,
1322         char __user *__user *envp,
1323         struct pt_regs * regs)
1324 {
1325         struct linux_binprm *bprm;
1326         struct file *file;
1327         struct files_struct *displaced;
1328         bool clear_in_exec;
1329         int retval;
1330
1331         retval = unshare_files(&displaced);
1332         if (retval)
1333                 goto out_ret;
1334
1335         retval = -ENOMEM;
1336         bprm = kzalloc(sizeof(*bprm), GFP_KERNEL);
1337         if (!bprm)
1338                 goto out_files;
1339
1340         retval = prepare_bprm_creds(bprm);
1341         if (retval)
1342                 goto out_free;
1343
1344         retval = check_unsafe_exec(bprm);
1345         if (retval < 0)
1346                 goto out_free;
1347         clear_in_exec = retval;
1348         current->in_execve = 1;
1349
1350         file = open_exec(filename);
1351         retval = PTR_ERR(file);
1352         if (IS_ERR(file))
1353                 goto out_unmark;
1354
1355         sched_exec();
1356
1357         bprm->file = file;
1358         bprm->filename = filename;
1359         bprm->interp = filename;
1360
1361         retval = bprm_mm_init(bprm);
1362         if (retval)
1363                 goto out_file;
1364
1365         bprm->argc = count(argv, MAX_ARG_STRINGS);
1366         if ((retval = bprm->argc) < 0)
1367                 goto out;
1368
1369         bprm->envc = count(envp, MAX_ARG_STRINGS);
1370         if ((retval = bprm->envc) < 0)
1371                 goto out;
1372
1373         retval = prepare_binprm(bprm);
1374         if (retval < 0)
1375                 goto out;
1376
1377         retval = copy_strings_kernel(1, &bprm->filename, bprm);
1378         if (retval < 0)
1379                 goto out;
1380
1381         bprm->exec = bprm->p;
1382         retval = copy_strings(bprm->envc, envp, bprm);
1383         if (retval < 0)
1384                 goto out;
1385
1386         retval = copy_strings(bprm->argc, argv, bprm);
1387         if (retval < 0)
1388                 goto out;
1389
1390         current->flags &= ~PF_KTHREAD;
1391         retval = search_binary_handler(bprm,regs);
1392         if (retval < 0)
1393                 goto out;
1394
1395         /* execve succeeded */
1396         current->fs->in_exec = 0;
1397         current->in_execve = 0;
1398         acct_update_integrals(current);
1399         free_bprm(bprm);
1400         if (displaced)
1401                 put_files_struct(displaced);
1402         return retval;
1403
1404 out:
1405         if (bprm->mm)
1406                 mmput (bprm->mm);
1407
1408 out_file:
1409         if (bprm->file) {
1410                 allow_write_access(bprm->file);
1411                 fput(bprm->file);
1412         }
1413
1414 out_unmark:
1415         if (clear_in_exec)
1416                 current->fs->in_exec = 0;
1417         current->in_execve = 0;
1418
1419 out_free:
1420         free_bprm(bprm);
1421
1422 out_files:
1423         if (displaced)
1424                 reset_files_struct(displaced);
1425 out_ret:
1426         return retval;
1427 }
1428
1429 void set_binfmt(struct linux_binfmt *new)
1430 {
1431         struct mm_struct *mm = current->mm;
1432
1433         if (mm->binfmt)
1434                 module_put(mm->binfmt->module);
1435
1436         mm->binfmt = new;
1437         if (new)
1438                 __module_get(new->module);
1439 }
1440
1441 EXPORT_SYMBOL(set_binfmt);
1442
1443 /* format_corename will inspect the pattern parameter, and output a
1444  * name into corename, which must have space for at least
1445  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
1446  */
1447 static int format_corename(char *corename, long signr)
1448 {
1449         const struct cred *cred = current_cred();
1450         const char *pat_ptr = core_pattern;
1451         int ispipe = (*pat_ptr == '|');
1452         char *out_ptr = corename;
1453         char *const out_end = corename + CORENAME_MAX_SIZE;
1454         int rc;
1455         int pid_in_pattern = 0;
1456
1457         /* Repeat as long as we have more pattern to process and more output
1458            space */
1459         while (*pat_ptr) {
1460                 if (*pat_ptr != '%') {
1461                         if (out_ptr == out_end)
1462                                 goto out;
1463                         *out_ptr++ = *pat_ptr++;
1464                 } else {
1465                         switch (*++pat_ptr) {
1466                         case 0:
1467                                 goto out;
1468                         /* Double percent, output one percent */
1469                         case '%':
1470                                 if (out_ptr == out_end)
1471                                         goto out;
1472                                 *out_ptr++ = '%';
1473                                 break;
1474                         /* pid */
1475                         case 'p':
1476                                 pid_in_pattern = 1;
1477                                 rc = snprintf(out_ptr, out_end - out_ptr,
1478                                               "%d", task_tgid_vnr(current));
1479                                 if (rc > out_end - out_ptr)
1480                                         goto out;
1481                                 out_ptr += rc;
1482                                 break;
1483                         /* uid */
1484                         case 'u':
1485                                 rc = snprintf(out_ptr, out_end - out_ptr,
1486                                               "%d", cred->uid);
1487                                 if (rc > out_end - out_ptr)
1488                                         goto out;
1489                                 out_ptr += rc;
1490                                 break;
1491                         /* gid */
1492                         case 'g':
1493                                 rc = snprintf(out_ptr, out_end - out_ptr,
1494                                               "%d", cred->gid);
1495                                 if (rc > out_end - out_ptr)
1496                                         goto out;
1497                                 out_ptr += rc;
1498                                 break;
1499                         /* signal that caused the coredump */
1500                         case 's':
1501                                 rc = snprintf(out_ptr, out_end - out_ptr,
1502                                               "%ld", signr);
1503                                 if (rc > out_end - out_ptr)
1504                                         goto out;
1505                                 out_ptr += rc;
1506                                 break;
1507                         /* UNIX time of coredump */
1508                         case 't': {
1509                                 struct timeval tv;
1510                                 do_gettimeofday(&tv);
1511                                 rc = snprintf(out_ptr, out_end - out_ptr,
1512                                               "%lu", tv.tv_sec);
1513                                 if (rc > out_end - out_ptr)
1514                                         goto out;
1515                                 out_ptr += rc;
1516                                 break;
1517                         }
1518                         /* hostname */
1519                         case 'h':
1520                                 down_read(&uts_sem);
1521                                 rc = snprintf(out_ptr, out_end - out_ptr,
1522                                               "%s", utsname()->nodename);
1523                                 up_read(&uts_sem);
1524                                 if (rc > out_end - out_ptr)
1525                                         goto out;
1526                                 out_ptr += rc;
1527                                 break;
1528                         /* executable */
1529                         case 'e':
1530                                 rc = snprintf(out_ptr, out_end - out_ptr,
1531                                               "%s", current->comm);
1532                                 if (rc > out_end - out_ptr)
1533                                         goto out;
1534                                 out_ptr += rc;
1535                                 break;
1536                         /* core limit size */
1537                         case 'c':
1538                                 rc = snprintf(out_ptr, out_end - out_ptr,
1539                                               "%lu", rlimit(RLIMIT_CORE));
1540                                 if (rc > out_end - out_ptr)
1541                                         goto out;
1542                                 out_ptr += rc;
1543                                 break;
1544                         default:
1545                                 break;
1546                         }
1547                         ++pat_ptr;
1548                 }
1549         }
1550         /* Backward compatibility with core_uses_pid:
1551          *
1552          * If core_pattern does not include a %p (as is the default)
1553          * and core_uses_pid is set, then .%pid will be appended to
1554          * the filename. Do not do this for piped commands. */
1555         if (!ispipe && !pid_in_pattern && core_uses_pid) {
1556                 rc = snprintf(out_ptr, out_end - out_ptr,
1557                               ".%d", task_tgid_vnr(current));
1558                 if (rc > out_end - out_ptr)
1559                         goto out;
1560                 out_ptr += rc;
1561         }
1562 out:
1563         *out_ptr = 0;
1564         return ispipe;
1565 }
1566
1567 static int zap_process(struct task_struct *start, int exit_code)
1568 {
1569         struct task_struct *t;
1570         int nr = 0;
1571
1572         start->signal->flags = SIGNAL_GROUP_EXIT;
1573         start->signal->group_exit_code = exit_code;
1574         start->signal->group_stop_count = 0;
1575
1576         t = start;
1577         do {
1578                 if (t != current && t->mm) {
1579                         sigaddset(&t->pending.signal, SIGKILL);
1580                         signal_wake_up(t, 1);
1581                         nr++;
1582                 }
1583         } while_each_thread(start, t);
1584
1585         return nr;
1586 }
1587
1588 static inline int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
1589                                 struct core_state *core_state, int exit_code)
1590 {
1591         struct task_struct *g, *p;
1592         unsigned long flags;
1593         int nr = -EAGAIN;
1594
1595         spin_lock_irq(&tsk->sighand->siglock);
1596         if (!signal_group_exit(tsk->signal)) {
1597                 mm->core_state = core_state;
1598                 nr = zap_process(tsk, exit_code);
1599         }
1600         spin_unlock_irq(&tsk->sighand->siglock);
1601         if (unlikely(nr < 0))
1602                 return nr;
1603
1604         if (atomic_read(&mm->mm_users) == nr + 1)
1605                 goto done;
1606         /*
1607          * We should find and kill all tasks which use this mm, and we should
1608          * count them correctly into ->nr_threads. We don't take tasklist
1609          * lock, but this is safe wrt:
1610          *
1611          * fork:
1612          *      None of sub-threads can fork after zap_process(leader). All
1613          *      processes which were created before this point should be
1614          *      visible to zap_threads() because copy_process() adds the new
1615          *      process to the tail of init_task.tasks list, and lock/unlock
1616          *      of ->siglock provides a memory barrier.
1617          *
1618          * do_exit:
1619          *      The caller holds mm->mmap_sem. This means that the task which
1620          *      uses this mm can't pass exit_mm(), so it can't exit or clear
1621          *      its ->mm.
1622          *
1623          * de_thread:
1624          *      It does list_replace_rcu(&leader->tasks, &current->tasks),
1625          *      we must see either old or new leader, this does not matter.
1626          *      However, it can change p->sighand, so lock_task_sighand(p)
1627          *      must be used. Since p->mm != NULL and we hold ->mmap_sem
1628          *      it can't fail.
1629          *
1630          *      Note also that "g" can be the old leader with ->mm == NULL
1631          *      and already unhashed and thus removed from ->thread_group.
1632          *      This is OK, __unhash_process()->list_del_rcu() does not
1633          *      clear the ->next pointer, we will find the new leader via
1634          *      next_thread().
1635          */
1636         rcu_read_lock();
1637         for_each_process(g) {
1638                 if (g == tsk->group_leader)
1639                         continue;
1640                 if (g->flags & PF_KTHREAD)
1641                         continue;
1642                 p = g;
1643                 do {
1644                         if (p->mm) {
1645                                 if (unlikely(p->mm == mm)) {
1646                                         lock_task_sighand(p, &flags);
1647                                         nr += zap_process(p, exit_code);
1648                                         unlock_task_sighand(p, &flags);
1649                                 }
1650                                 break;
1651                         }
1652                 } while_each_thread(g, p);
1653         }
1654         rcu_read_unlock();
1655 done:
1656         atomic_set(&core_state->nr_threads, nr);
1657         return nr;
1658 }
1659
1660 static int coredump_wait(int exit_code, struct core_state *core_state)
1661 {
1662         struct task_struct *tsk = current;
1663         struct mm_struct *mm = tsk->mm;
1664         struct completion *vfork_done;
1665         int core_waiters;
1666
1667         init_completion(&core_state->startup);
1668         core_state->dumper.task = tsk;
1669         core_state->dumper.next = NULL;
1670         core_waiters = zap_threads(tsk, mm, core_state, exit_code);
1671         up_write(&mm->mmap_sem);
1672
1673         if (unlikely(core_waiters < 0))
1674                 goto fail;
1675
1676         /*
1677          * Make sure nobody is waiting for us to release the VM,
1678          * otherwise we can deadlock when we wait on each other
1679          */
1680         vfork_done = tsk->vfork_done;
1681         if (vfork_done) {
1682                 tsk->vfork_done = NULL;
1683                 complete(vfork_done);
1684         }
1685
1686         if (core_waiters)
1687                 wait_for_completion(&core_state->startup);
1688 fail:
1689         return core_waiters;
1690 }
1691
1692 static void coredump_finish(struct mm_struct *mm)
1693 {
1694         struct core_thread *curr, *next;
1695         struct task_struct *task;
1696
1697         next = mm->core_state->dumper.next;
1698         while ((curr = next) != NULL) {
1699                 next = curr->next;
1700                 task = curr->task;
1701                 /*
1702                  * see exit_mm(), curr->task must not see
1703                  * ->task == NULL before we read ->next.
1704                  */
1705                 smp_mb();
1706                 curr->task = NULL;
1707                 wake_up_process(task);
1708         }
1709
1710         mm->core_state = NULL;
1711 }
1712
1713 /*
1714  * set_dumpable converts traditional three-value dumpable to two flags and
1715  * stores them into mm->flags.  It modifies lower two bits of mm->flags, but
1716  * these bits are not changed atomically.  So get_dumpable can observe the
1717  * intermediate state.  To avoid doing unexpected behavior, get get_dumpable
1718  * return either old dumpable or new one by paying attention to the order of
1719  * modifying the bits.
1720  *
1721  * dumpable |   mm->flags (binary)
1722  * old  new | initial interim  final
1723  * ---------+-----------------------
1724  *  0    1  |   00      01      01
1725  *  0    2  |   00      10(*)   11
1726  *  1    0  |   01      00      00
1727  *  1    2  |   01      11      11
1728  *  2    0  |   11      10(*)   00
1729  *  2    1  |   11      11      01
1730  *
1731  * (*) get_dumpable regards interim value of 10 as 11.
1732  */
1733 void set_dumpable(struct mm_struct *mm, int value)
1734 {
1735         switch (value) {
1736         case 0:
1737                 clear_bit(MMF_DUMPABLE, &mm->flags);
1738                 smp_wmb();
1739                 clear_bit(MMF_DUMP_SECURELY, &mm->flags);
1740                 break;
1741         case 1:
1742                 set_bit(MMF_DUMPABLE, &mm->flags);
1743                 smp_wmb();
1744                 clear_bit(MMF_DUMP_SECURELY, &mm->flags);
1745                 break;
1746         case 2:
1747                 set_bit(MMF_DUMP_SECURELY, &mm->flags);
1748                 smp_wmb();
1749                 set_bit(MMF_DUMPABLE, &mm->flags);
1750                 break;
1751         }
1752 }
1753
1754 static int __get_dumpable(unsigned long mm_flags)
1755 {
1756         int ret;
1757
1758         ret = mm_flags & MMF_DUMPABLE_MASK;
1759         return (ret >= 2) ? 2 : ret;
1760 }
1761
1762 int get_dumpable(struct mm_struct *mm)
1763 {
1764         return __get_dumpable(mm->flags);
1765 }
1766
1767 static void wait_for_dump_helpers(struct file *file)
1768 {
1769         struct pipe_inode_info *pipe;
1770
1771         pipe = file->f_path.dentry->d_inode->i_pipe;
1772
1773         pipe_lock(pipe);
1774         pipe->readers++;
1775         pipe->writers--;
1776
1777         while ((pipe->readers > 1) && (!signal_pending(current))) {
1778                 wake_up_interruptible_sync(&pipe->wait);
1779                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
1780                 pipe_wait(pipe);
1781         }
1782
1783         pipe->readers--;
1784         pipe->writers++;
1785         pipe_unlock(pipe);
1786
1787 }
1788
1789
1790 void do_coredump(long signr, int exit_code, struct pt_regs *regs)
1791 {
1792         struct core_state core_state;
1793         char corename[CORENAME_MAX_SIZE + 1];
1794         struct mm_struct *mm = current->mm;
1795         struct linux_binfmt * binfmt;
1796         struct inode * inode;
1797         const struct cred *old_cred;
1798         struct cred *cred;
1799         int retval = 0;
1800         int flag = 0;
1801         int ispipe = 0;
1802         char **helper_argv = NULL;
1803         int helper_argc = 0;
1804         int dump_count = 0;
1805         static atomic_t core_dump_count = ATOMIC_INIT(0);
1806         struct coredump_params cprm = {
1807                 .signr = signr,
1808                 .regs = regs,
1809                 .limit = rlimit(RLIMIT_CORE),
1810                 /*
1811                  * We must use the same mm->flags while dumping core to avoid
1812                  * inconsistency of bit flags, since this flag is not protected
1813                  * by any locks.
1814                  */
1815                 .mm_flags = mm->flags,
1816         };
1817
1818         audit_core_dumps(signr);
1819
1820         binfmt = mm->binfmt;
1821         if (!binfmt || !binfmt->core_dump)
1822                 goto fail;
1823
1824         cred = prepare_creds();
1825         if (!cred) {
1826                 retval = -ENOMEM;
1827                 goto fail;
1828         }
1829
1830         down_write(&mm->mmap_sem);
1831         /*
1832          * If another thread got here first, or we are not dumpable, bail out.
1833          */
1834         if (mm->core_state || !__get_dumpable(cprm.mm_flags)) {
1835                 up_write(&mm->mmap_sem);
1836                 put_cred(cred);
1837                 goto fail;
1838         }
1839
1840         /*
1841          *      We cannot trust fsuid as being the "true" uid of the
1842          *      process nor do we know its entire history. We only know it
1843          *      was tainted so we dump it as root in mode 2.
1844          */
1845         if (__get_dumpable(cprm.mm_flags) == 2) {
1846                 /* Setuid core dump mode */
1847                 flag = O_EXCL;          /* Stop rewrite attacks */
1848                 cred->fsuid = 0;        /* Dump root private */
1849         }
1850
1851         retval = coredump_wait(exit_code, &core_state);
1852         if (retval < 0) {
1853                 put_cred(cred);
1854                 goto fail;
1855         }
1856
1857         old_cred = override_creds(cred);
1858
1859         /*
1860          * Clear any false indication of pending signals that might
1861          * be seen by the filesystem code called to write the core file.
1862          */
1863         clear_thread_flag(TIF_SIGPENDING);
1864
1865         /*
1866          * lock_kernel() because format_corename() is controlled by sysctl, which
1867          * uses lock_kernel()
1868          */
1869         lock_kernel();
1870         ispipe = format_corename(corename, signr);
1871         unlock_kernel();
1872
1873         if ((!ispipe) && (cprm.limit < binfmt->min_coredump))
1874                 goto fail_unlock;
1875
1876         if (ispipe) {
1877                 if (cprm.limit == 0) {
1878                         /*
1879                          * Normally core limits are irrelevant to pipes, since
1880                          * we're not writing to the file system, but we use
1881                          * cprm.limit of 0 here as a speacial value. Any
1882                          * non-zero limit gets set to RLIM_INFINITY below, but
1883                          * a limit of 0 skips the dump.  This is a consistent
1884                          * way to catch recursive crashes.  We can still crash
1885                          * if the core_pattern binary sets RLIM_CORE =  !0
1886                          * but it runs as root, and can do lots of stupid things
1887                          * Note that we use task_tgid_vnr here to grab the pid
1888                          * of the process group leader.  That way we get the
1889                          * right pid if a thread in a multi-threaded
1890                          * core_pattern process dies.
1891                          */
1892                         printk(KERN_WARNING
1893                                 "Process %d(%s) has RLIMIT_CORE set to 0\n",
1894                                 task_tgid_vnr(current), current->comm);
1895                         printk(KERN_WARNING "Aborting core\n");
1896                         goto fail_unlock;
1897                 }
1898
1899                 dump_count = atomic_inc_return(&core_dump_count);
1900                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
1901                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
1902                                task_tgid_vnr(current), current->comm);
1903                         printk(KERN_WARNING "Skipping core dump\n");
1904                         goto fail_dropcount;
1905                 }
1906
1907                 helper_argv = argv_split(GFP_KERNEL, corename+1, &helper_argc);
1908                 if (!helper_argv) {
1909                         printk(KERN_WARNING "%s failed to allocate memory\n",
1910                                __func__);
1911                         goto fail_dropcount;
1912                 }
1913
1914                 cprm.limit = RLIM_INFINITY;
1915
1916                 /* SIGPIPE can happen, but it's just never processed */
1917                 if (call_usermodehelper_pipe(helper_argv[0], helper_argv, NULL,
1918                                 &cprm.file)) {
1919                         printk(KERN_INFO "Core dump to %s pipe failed\n",
1920                                corename);
1921                         goto fail_dropcount;
1922                 }
1923         } else
1924                 cprm.file = filp_open(corename,
1925                                  O_CREAT | 2 | O_NOFOLLOW | O_LARGEFILE | flag,
1926                                  0600);
1927         if (IS_ERR(cprm.file))
1928                 goto fail_dropcount;
1929         inode = cprm.file->f_path.dentry->d_inode;
1930         if (inode->i_nlink > 1)
1931                 goto close_fail;        /* multiple links - don't dump */
1932         if (!ispipe && d_unhashed(cprm.file->f_path.dentry))
1933                 goto close_fail;
1934
1935         /* AK: actually i see no reason to not allow this for named pipes etc.,
1936            but keep the previous behaviour for now. */
1937         if (!ispipe && !S_ISREG(inode->i_mode))
1938                 goto close_fail;
1939         /*
1940          * Dont allow local users get cute and trick others to coredump
1941          * into their pre-created files:
1942          * Note, this is not relevant for pipes
1943          */
1944         if (!ispipe && (inode->i_uid != current_fsuid()))
1945                 goto close_fail;
1946         if (!cprm.file->f_op)
1947                 goto close_fail;
1948         if (!cprm.file->f_op->write)
1949                 goto close_fail;
1950         if (!ispipe &&
1951             do_truncate(cprm.file->f_path.dentry, 0, 0, cprm.file) != 0)
1952                 goto close_fail;
1953
1954         retval = binfmt->core_dump(&cprm);
1955
1956         if (retval)
1957                 current->signal->group_exit_code |= 0x80;
1958 close_fail:
1959         if (ispipe && core_pipe_limit)
1960                 wait_for_dump_helpers(cprm.file);
1961         filp_close(cprm.file, NULL);
1962 fail_dropcount:
1963         if (dump_count)
1964                 atomic_dec(&core_dump_count);
1965 fail_unlock:
1966         if (helper_argv)
1967                 argv_free(helper_argv);
1968
1969         revert_creds(old_cred);
1970         put_cred(cred);
1971         coredump_finish(mm);
1972 fail:
1973         return;
1974 }