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