Merge commit 'v2.6.29-rc1' into perfcounters/core
[linux-2.6.git] / 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_counter.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
57 #include <asm/uaccess.h>
58 #include <asm/mmu_context.h>
59 #include <asm/tlb.h>
60 #include "internal.h"
61
62 int core_uses_pid;
63 char core_pattern[CORENAME_MAX_SIZE] = "core";
64 int suid_dumpable = 0;
65
66 /* The maximal length of core_pattern is also specified in sysctl.c */
67
68 static LIST_HEAD(formats);
69 static DEFINE_RWLOCK(binfmt_lock);
70
71 int register_binfmt(struct linux_binfmt * fmt)
72 {
73         if (!fmt)
74                 return -EINVAL;
75         write_lock(&binfmt_lock);
76         list_add(&fmt->lh, &formats);
77         write_unlock(&binfmt_lock);
78         return 0;       
79 }
80
81 EXPORT_SYMBOL(register_binfmt);
82
83 void unregister_binfmt(struct linux_binfmt * fmt)
84 {
85         write_lock(&binfmt_lock);
86         list_del(&fmt->lh);
87         write_unlock(&binfmt_lock);
88 }
89
90 EXPORT_SYMBOL(unregister_binfmt);
91
92 static inline void put_binfmt(struct linux_binfmt * fmt)
93 {
94         module_put(fmt->module);
95 }
96
97 /*
98  * Note that a shared library must be both readable and executable due to
99  * security reasons.
100  *
101  * Also note that we take the address to load from from the file itself.
102  */
103 asmlinkage long sys_uselib(const char __user * library)
104 {
105         struct file *file;
106         struct nameidata nd;
107         char *tmp = getname(library);
108         int error = PTR_ERR(tmp);
109
110         if (!IS_ERR(tmp)) {
111                 error = path_lookup_open(AT_FDCWD, tmp,
112                                          LOOKUP_FOLLOW, &nd,
113                                          FMODE_READ|FMODE_EXEC);
114                 putname(tmp);
115         }
116         if (error)
117                 goto out;
118
119         error = -EINVAL;
120         if (!S_ISREG(nd.path.dentry->d_inode->i_mode))
121                 goto exit;
122
123         error = -EACCES;
124         if (nd.path.mnt->mnt_flags & MNT_NOEXEC)
125                 goto exit;
126
127         error = inode_permission(nd.path.dentry->d_inode,
128                                  MAY_READ | MAY_EXEC | MAY_OPEN);
129         if (error)
130                 goto exit;
131
132         file = nameidata_to_filp(&nd, O_RDONLY|O_LARGEFILE);
133         error = PTR_ERR(file);
134         if (IS_ERR(file))
135                 goto out;
136
137         fsnotify_open(file->f_path.dentry);
138
139         error = -ENOEXEC;
140         if(file->f_op) {
141                 struct linux_binfmt * fmt;
142
143                 read_lock(&binfmt_lock);
144                 list_for_each_entry(fmt, &formats, lh) {
145                         if (!fmt->load_shlib)
146                                 continue;
147                         if (!try_module_get(fmt->module))
148                                 continue;
149                         read_unlock(&binfmt_lock);
150                         error = fmt->load_shlib(file);
151                         read_lock(&binfmt_lock);
152                         put_binfmt(fmt);
153                         if (error != -ENOEXEC)
154                                 break;
155                 }
156                 read_unlock(&binfmt_lock);
157         }
158         fput(file);
159 out:
160         return error;
161 exit:
162         release_open_intent(&nd);
163         path_put(&nd.path);
164         goto out;
165 }
166
167 #ifdef CONFIG_MMU
168
169 static struct page *get_arg_page(struct linux_binprm *bprm, unsigned long pos,
170                 int write)
171 {
172         struct page *page;
173         int ret;
174
175 #ifdef CONFIG_STACK_GROWSUP
176         if (write) {
177                 ret = expand_stack_downwards(bprm->vma, pos);
178                 if (ret < 0)
179                         return NULL;
180         }
181 #endif
182         ret = get_user_pages(current, bprm->mm, pos,
183                         1, write, 1, &page, NULL);
184         if (ret <= 0)
185                 return NULL;
186
187         if (write) {
188                 unsigned long size = bprm->vma->vm_end - bprm->vma->vm_start;
189                 struct rlimit *rlim;
190
191                 /*
192                  * We've historically supported up to 32 pages (ARG_MAX)
193                  * of argument strings even with small stacks
194                  */
195                 if (size <= ARG_MAX)
196                         return page;
197
198                 /*
199                  * Limit to 1/4-th the stack size for the argv+env strings.
200                  * This ensures that:
201                  *  - the remaining binfmt code will not run out of stack space,
202                  *  - the program will have a reasonable amount of stack left
203                  *    to work from.
204                  */
205                 rlim = current->signal->rlim;
206                 if (size > rlim[RLIMIT_STACK].rlim_cur / 4) {
207                         put_page(page);
208                         return NULL;
209                 }
210         }
211
212         return page;
213 }
214
215 static void put_arg_page(struct page *page)
216 {
217         put_page(page);
218 }
219
220 static void free_arg_page(struct linux_binprm *bprm, int i)
221 {
222 }
223
224 static void free_arg_pages(struct linux_binprm *bprm)
225 {
226 }
227
228 static void flush_arg_page(struct linux_binprm *bprm, unsigned long pos,
229                 struct page *page)
230 {
231         flush_cache_page(bprm->vma, pos, page_to_pfn(page));
232 }
233
234 static int __bprm_mm_init(struct linux_binprm *bprm)
235 {
236         int err;
237         struct vm_area_struct *vma = NULL;
238         struct mm_struct *mm = bprm->mm;
239
240         bprm->vma = vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
241         if (!vma)
242                 return -ENOMEM;
243
244         down_write(&mm->mmap_sem);
245         vma->vm_mm = mm;
246
247         /*
248          * Place the stack at the largest stack address the architecture
249          * supports. Later, we'll move this to an appropriate place. We don't
250          * use STACK_TOP because that can depend on attributes which aren't
251          * configured yet.
252          */
253         vma->vm_end = STACK_TOP_MAX;
254         vma->vm_start = vma->vm_end - PAGE_SIZE;
255         vma->vm_flags = VM_STACK_FLAGS;
256         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
257         err = insert_vm_struct(mm, vma);
258         if (err)
259                 goto err;
260
261         mm->stack_vm = mm->total_vm = 1;
262         up_write(&mm->mmap_sem);
263         bprm->p = vma->vm_end - sizeof(void *);
264         return 0;
265 err:
266         up_write(&mm->mmap_sem);
267         bprm->vma = NULL;
268         kmem_cache_free(vm_area_cachep, vma);
269         return err;
270 }
271
272 static bool valid_arg_len(struct linux_binprm *bprm, long len)
273 {
274         return len <= MAX_ARG_STRLEN;
275 }
276
277 #else
278
279 static struct page *get_arg_page(struct linux_binprm *bprm, unsigned long pos,
280                 int write)
281 {
282         struct page *page;
283
284         page = bprm->page[pos / PAGE_SIZE];
285         if (!page && write) {
286                 page = alloc_page(GFP_HIGHUSER|__GFP_ZERO);
287                 if (!page)
288                         return NULL;
289                 bprm->page[pos / PAGE_SIZE] = page;
290         }
291
292         return page;
293 }
294
295 static void put_arg_page(struct page *page)
296 {
297 }
298
299 static void free_arg_page(struct linux_binprm *bprm, int i)
300 {
301         if (bprm->page[i]) {
302                 __free_page(bprm->page[i]);
303                 bprm->page[i] = NULL;
304         }
305 }
306
307 static void free_arg_pages(struct linux_binprm *bprm)
308 {
309         int i;
310
311         for (i = 0; i < MAX_ARG_PAGES; i++)
312                 free_arg_page(bprm, i);
313 }
314
315 static void flush_arg_page(struct linux_binprm *bprm, unsigned long pos,
316                 struct page *page)
317 {
318 }
319
320 static int __bprm_mm_init(struct linux_binprm *bprm)
321 {
322         bprm->p = PAGE_SIZE * MAX_ARG_PAGES - sizeof(void *);
323         return 0;
324 }
325
326 static bool valid_arg_len(struct linux_binprm *bprm, long len)
327 {
328         return len <= bprm->p;
329 }
330
331 #endif /* CONFIG_MMU */
332
333 /*
334  * Create a new mm_struct and populate it with a temporary stack
335  * vm_area_struct.  We don't have enough context at this point to set the stack
336  * flags, permissions, and offset, so we use temporary values.  We'll update
337  * them later in setup_arg_pages().
338  */
339 int bprm_mm_init(struct linux_binprm *bprm)
340 {
341         int err;
342         struct mm_struct *mm = NULL;
343
344         bprm->mm = mm = mm_alloc();
345         err = -ENOMEM;
346         if (!mm)
347                 goto err;
348
349         err = init_new_context(current, mm);
350         if (err)
351                 goto err;
352
353         err = __bprm_mm_init(bprm);
354         if (err)
355                 goto err;
356
357         return 0;
358
359 err:
360         if (mm) {
361                 bprm->mm = NULL;
362                 mmdrop(mm);
363         }
364
365         return err;
366 }
367
368 /*
369  * count() counts the number of strings in array ARGV.
370  */
371 static int count(char __user * __user * argv, int max)
372 {
373         int i = 0;
374
375         if (argv != NULL) {
376                 for (;;) {
377                         char __user * p;
378
379                         if (get_user(p, argv))
380                                 return -EFAULT;
381                         if (!p)
382                                 break;
383                         argv++;
384                         if (i++ >= max)
385                                 return -E2BIG;
386                         cond_resched();
387                 }
388         }
389         return i;
390 }
391
392 /*
393  * 'copy_strings()' copies argument/environment strings from the old
394  * processes's memory to the new process's stack.  The call to get_user_pages()
395  * ensures the destination page is created and not swapped out.
396  */
397 static int copy_strings(int argc, char __user * __user * argv,
398                         struct linux_binprm *bprm)
399 {
400         struct page *kmapped_page = NULL;
401         char *kaddr = NULL;
402         unsigned long kpos = 0;
403         int ret;
404
405         while (argc-- > 0) {
406                 char __user *str;
407                 int len;
408                 unsigned long pos;
409
410                 if (get_user(str, argv+argc) ||
411                                 !(len = strnlen_user(str, MAX_ARG_STRLEN))) {
412                         ret = -EFAULT;
413                         goto out;
414                 }
415
416                 if (!valid_arg_len(bprm, len)) {
417                         ret = -E2BIG;
418                         goto out;
419                 }
420
421                 /* We're going to work our way backwords. */
422                 pos = bprm->p;
423                 str += len;
424                 bprm->p -= len;
425
426                 while (len > 0) {
427                         int offset, bytes_to_copy;
428
429                         offset = pos % PAGE_SIZE;
430                         if (offset == 0)
431                                 offset = PAGE_SIZE;
432
433                         bytes_to_copy = offset;
434                         if (bytes_to_copy > len)
435                                 bytes_to_copy = len;
436
437                         offset -= bytes_to_copy;
438                         pos -= bytes_to_copy;
439                         str -= bytes_to_copy;
440                         len -= bytes_to_copy;
441
442                         if (!kmapped_page || kpos != (pos & PAGE_MASK)) {
443                                 struct page *page;
444
445                                 page = get_arg_page(bprm, pos, 1);
446                                 if (!page) {
447                                         ret = -E2BIG;
448                                         goto out;
449                                 }
450
451                                 if (kmapped_page) {
452                                         flush_kernel_dcache_page(kmapped_page);
453                                         kunmap(kmapped_page);
454                                         put_arg_page(kmapped_page);
455                                 }
456                                 kmapped_page = page;
457                                 kaddr = kmap(kmapped_page);
458                                 kpos = pos & PAGE_MASK;
459                                 flush_arg_page(bprm, kpos, kmapped_page);
460                         }
461                         if (copy_from_user(kaddr+offset, str, bytes_to_copy)) {
462                                 ret = -EFAULT;
463                                 goto out;
464                         }
465                 }
466         }
467         ret = 0;
468 out:
469         if (kmapped_page) {
470                 flush_kernel_dcache_page(kmapped_page);
471                 kunmap(kmapped_page);
472                 put_arg_page(kmapped_page);
473         }
474         return ret;
475 }
476
477 /*
478  * Like copy_strings, but get argv and its values from kernel memory.
479  */
480 int copy_strings_kernel(int argc,char ** argv, struct linux_binprm *bprm)
481 {
482         int r;
483         mm_segment_t oldfs = get_fs();
484         set_fs(KERNEL_DS);
485         r = copy_strings(argc, (char __user * __user *)argv, bprm);
486         set_fs(oldfs);
487         return r;
488 }
489 EXPORT_SYMBOL(copy_strings_kernel);
490
491 #ifdef CONFIG_MMU
492
493 /*
494  * During bprm_mm_init(), we create a temporary stack at STACK_TOP_MAX.  Once
495  * the binfmt code determines where the new stack should reside, we shift it to
496  * its final location.  The process proceeds as follows:
497  *
498  * 1) Use shift to calculate the new vma endpoints.
499  * 2) Extend vma to cover both the old and new ranges.  This ensures the
500  *    arguments passed to subsequent functions are consistent.
501  * 3) Move vma's page tables to the new range.
502  * 4) Free up any cleared pgd range.
503  * 5) Shrink the vma to cover only the new range.
504  */
505 static int shift_arg_pages(struct vm_area_struct *vma, unsigned long shift)
506 {
507         struct mm_struct *mm = vma->vm_mm;
508         unsigned long old_start = vma->vm_start;
509         unsigned long old_end = vma->vm_end;
510         unsigned long length = old_end - old_start;
511         unsigned long new_start = old_start - shift;
512         unsigned long new_end = old_end - shift;
513         struct mmu_gather *tlb;
514
515         BUG_ON(new_start > new_end);
516
517         /*
518          * ensure there are no vmas between where we want to go
519          * and where we are
520          */
521         if (vma != find_vma(mm, new_start))
522                 return -EFAULT;
523
524         /*
525          * cover the whole range: [new_start, old_end)
526          */
527         vma_adjust(vma, new_start, old_end, vma->vm_pgoff, NULL);
528
529         /*
530          * move the page tables downwards, on failure we rely on
531          * process cleanup to remove whatever mess we made.
532          */
533         if (length != move_page_tables(vma, old_start,
534                                        vma, new_start, length))
535                 return -ENOMEM;
536
537         lru_add_drain();
538         tlb = tlb_gather_mmu(mm, 0);
539         if (new_end > old_start) {
540                 /*
541                  * when the old and new regions overlap clear from new_end.
542                  */
543                 free_pgd_range(tlb, new_end, old_end, new_end,
544                         vma->vm_next ? vma->vm_next->vm_start : 0);
545         } else {
546                 /*
547                  * otherwise, clean from old_start; this is done to not touch
548                  * the address space in [new_end, old_start) some architectures
549                  * have constraints on va-space that make this illegal (IA64) -
550                  * for the others its just a little faster.
551                  */
552                 free_pgd_range(tlb, old_start, old_end, new_end,
553                         vma->vm_next ? vma->vm_next->vm_start : 0);
554         }
555         tlb_finish_mmu(tlb, new_end, old_end);
556
557         /*
558          * shrink the vma to just the new range.
559          */
560         vma_adjust(vma, new_start, new_end, vma->vm_pgoff, NULL);
561
562         return 0;
563 }
564
565 #define EXTRA_STACK_VM_PAGES    20      /* random */
566
567 /*
568  * Finalizes the stack vm_area_struct. The flags and permissions are updated,
569  * the stack is optionally relocated, and some extra space is added.
570  */
571 int setup_arg_pages(struct linux_binprm *bprm,
572                     unsigned long stack_top,
573                     int executable_stack)
574 {
575         unsigned long ret;
576         unsigned long stack_shift;
577         struct mm_struct *mm = current->mm;
578         struct vm_area_struct *vma = bprm->vma;
579         struct vm_area_struct *prev = NULL;
580         unsigned long vm_flags;
581         unsigned long stack_base;
582
583 #ifdef CONFIG_STACK_GROWSUP
584         /* Limit stack size to 1GB */
585         stack_base = current->signal->rlim[RLIMIT_STACK].rlim_max;
586         if (stack_base > (1 << 30))
587                 stack_base = 1 << 30;
588
589         /* Make sure we didn't let the argument array grow too large. */
590         if (vma->vm_end - vma->vm_start > stack_base)
591                 return -ENOMEM;
592
593         stack_base = PAGE_ALIGN(stack_top - stack_base);
594
595         stack_shift = vma->vm_start - stack_base;
596         mm->arg_start = bprm->p - stack_shift;
597         bprm->p = vma->vm_end - stack_shift;
598 #else
599         stack_top = arch_align_stack(stack_top);
600         stack_top = PAGE_ALIGN(stack_top);
601         stack_shift = vma->vm_end - stack_top;
602
603         bprm->p -= stack_shift;
604         mm->arg_start = bprm->p;
605 #endif
606
607         if (bprm->loader)
608                 bprm->loader -= stack_shift;
609         bprm->exec -= stack_shift;
610
611         down_write(&mm->mmap_sem);
612         vm_flags = VM_STACK_FLAGS;
613
614         /*
615          * Adjust stack execute permissions; explicitly enable for
616          * EXSTACK_ENABLE_X, disable for EXSTACK_DISABLE_X and leave alone
617          * (arch default) otherwise.
618          */
619         if (unlikely(executable_stack == EXSTACK_ENABLE_X))
620                 vm_flags |= VM_EXEC;
621         else if (executable_stack == EXSTACK_DISABLE_X)
622                 vm_flags &= ~VM_EXEC;
623         vm_flags |= mm->def_flags;
624
625         ret = mprotect_fixup(vma, &prev, vma->vm_start, vma->vm_end,
626                         vm_flags);
627         if (ret)
628                 goto out_unlock;
629         BUG_ON(prev != vma);
630
631         /* Move stack pages down in memory. */
632         if (stack_shift) {
633                 ret = shift_arg_pages(vma, stack_shift);
634                 if (ret) {
635                         up_write(&mm->mmap_sem);
636                         return ret;
637                 }
638         }
639
640 #ifdef CONFIG_STACK_GROWSUP
641         stack_base = vma->vm_end + EXTRA_STACK_VM_PAGES * PAGE_SIZE;
642 #else
643         stack_base = vma->vm_start - EXTRA_STACK_VM_PAGES * PAGE_SIZE;
644 #endif
645         ret = expand_stack(vma, stack_base);
646         if (ret)
647                 ret = -EFAULT;
648
649 out_unlock:
650         up_write(&mm->mmap_sem);
651         return 0;
652 }
653 EXPORT_SYMBOL(setup_arg_pages);
654
655 #endif /* CONFIG_MMU */
656
657 struct file *open_exec(const char *name)
658 {
659         struct nameidata nd;
660         struct file *file;
661         int err;
662
663         err = path_lookup_open(AT_FDCWD, name, LOOKUP_FOLLOW, &nd,
664                                 FMODE_READ|FMODE_EXEC);
665         if (err)
666                 goto out;
667
668         err = -EACCES;
669         if (!S_ISREG(nd.path.dentry->d_inode->i_mode))
670                 goto out_path_put;
671
672         if (nd.path.mnt->mnt_flags & MNT_NOEXEC)
673                 goto out_path_put;
674
675         err = inode_permission(nd.path.dentry->d_inode, MAY_EXEC | MAY_OPEN);
676         if (err)
677                 goto out_path_put;
678
679         file = nameidata_to_filp(&nd, O_RDONLY|O_LARGEFILE);
680         if (IS_ERR(file))
681                 return file;
682
683         fsnotify_open(file->f_path.dentry);
684
685         err = deny_write_access(file);
686         if (err) {
687                 fput(file);
688                 goto out;
689         }
690
691         return file;
692
693  out_path_put:
694         release_open_intent(&nd);
695         path_put(&nd.path);
696  out:
697         return ERR_PTR(err);
698 }
699 EXPORT_SYMBOL(open_exec);
700
701 int kernel_read(struct file *file, unsigned long 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         mm_release(tsk, old_mm);
727
728         if (old_mm) {
729                 /*
730                  * Make sure that if there is a core dump in progress
731                  * for the old mm, we get out and die instead of going
732                  * through with the exec.  We must hold mmap_sem around
733                  * checking core_state and changing tsk->mm.
734                  */
735                 down_read(&old_mm->mmap_sem);
736                 if (unlikely(old_mm->core_state)) {
737                         up_read(&old_mm->mmap_sem);
738                         return -EINTR;
739                 }
740         }
741         task_lock(tsk);
742         active_mm = tsk->active_mm;
743         tsk->mm = mm;
744         tsk->active_mm = mm;
745         activate_mm(active_mm, mm);
746         task_unlock(tsk);
747         arch_pick_mmap_layout(mm);
748         if (old_mm) {
749                 up_read(&old_mm->mmap_sem);
750                 BUG_ON(active_mm != old_mm);
751                 mm_update_next_owner(old_mm);
752                 mmput(old_mm);
753                 return 0;
754         }
755         mmdrop(active_mm);
756         return 0;
757 }
758
759 /*
760  * This function makes sure the current process has its own signal table,
761  * so that flush_signal_handlers can later reset the handlers without
762  * disturbing other processes.  (Other processes might share the signal
763  * table via the CLONE_SIGHAND option to clone().)
764  */
765 static int de_thread(struct task_struct *tsk)
766 {
767         struct signal_struct *sig = tsk->signal;
768         struct sighand_struct *oldsighand = tsk->sighand;
769         spinlock_t *lock = &oldsighand->siglock;
770         int count;
771
772         if (thread_group_empty(tsk))
773                 goto no_thread_group;
774
775         /*
776          * Kill all other threads in the thread group.
777          */
778         spin_lock_irq(lock);
779         if (signal_group_exit(sig)) {
780                 /*
781                  * Another group action in progress, just
782                  * return so that the signal is processed.
783                  */
784                 spin_unlock_irq(lock);
785                 return -EAGAIN;
786         }
787         sig->group_exit_task = tsk;
788         zap_other_threads(tsk);
789
790         /* Account for the thread group leader hanging around: */
791         count = thread_group_leader(tsk) ? 1 : 2;
792         sig->notify_count = count;
793         while (atomic_read(&sig->count) > count) {
794                 __set_current_state(TASK_UNINTERRUPTIBLE);
795                 spin_unlock_irq(lock);
796                 schedule();
797                 spin_lock_irq(lock);
798         }
799         spin_unlock_irq(lock);
800
801         /*
802          * At this point all other threads have exited, all we have to
803          * do is to wait for the thread group leader to become inactive,
804          * and to assume its PID:
805          */
806         if (!thread_group_leader(tsk)) {
807                 struct task_struct *leader = tsk->group_leader;
808
809                 sig->notify_count = -1; /* for exit_notify() */
810                 for (;;) {
811                         write_lock_irq(&tasklist_lock);
812                         if (likely(leader->exit_state))
813                                 break;
814                         __set_current_state(TASK_UNINTERRUPTIBLE);
815                         write_unlock_irq(&tasklist_lock);
816                         schedule();
817                 }
818
819                 /*
820                  * The only record we have of the real-time age of a
821                  * process, regardless of execs it's done, is start_time.
822                  * All the past CPU time is accumulated in signal_struct
823                  * from sister threads now dead.  But in this non-leader
824                  * exec, nothing survives from the original leader thread,
825                  * whose birth marks the true age of this process now.
826                  * When we take on its identity by switching to its PID, we
827                  * also take its birthdate (always earlier than our own).
828                  */
829                 tsk->start_time = leader->start_time;
830
831                 BUG_ON(!same_thread_group(leader, tsk));
832                 BUG_ON(has_group_leader_pid(tsk));
833                 /*
834                  * An exec() starts a new thread group with the
835                  * TGID of the previous thread group. Rehash the
836                  * two threads with a switched PID, and release
837                  * the former thread group leader:
838                  */
839
840                 /* Become a process group leader with the old leader's pid.
841                  * The old leader becomes a thread of the this thread group.
842                  * Note: The old leader also uses this pid until release_task
843                  *       is called.  Odd but simple and correct.
844                  */
845                 detach_pid(tsk, PIDTYPE_PID);
846                 tsk->pid = leader->pid;
847                 attach_pid(tsk, PIDTYPE_PID,  task_pid(leader));
848                 transfer_pid(leader, tsk, PIDTYPE_PGID);
849                 transfer_pid(leader, tsk, PIDTYPE_SID);
850                 list_replace_rcu(&leader->tasks, &tsk->tasks);
851
852                 tsk->group_leader = tsk;
853                 leader->group_leader = tsk;
854
855                 tsk->exit_signal = SIGCHLD;
856
857                 BUG_ON(leader->exit_state != EXIT_ZOMBIE);
858                 leader->exit_state = EXIT_DEAD;
859                 write_unlock_irq(&tasklist_lock);
860
861                 release_task(leader);
862         }
863
864         sig->group_exit_task = NULL;
865         sig->notify_count = 0;
866
867 no_thread_group:
868         exit_itimers(sig);
869         flush_itimer_signals();
870
871         if (atomic_read(&oldsighand->count) != 1) {
872                 struct sighand_struct *newsighand;
873                 /*
874                  * This ->sighand is shared with the CLONE_SIGHAND
875                  * but not CLONE_THREAD task, switch to the new one.
876                  */
877                 newsighand = kmem_cache_alloc(sighand_cachep, GFP_KERNEL);
878                 if (!newsighand)
879                         return -ENOMEM;
880
881                 atomic_set(&newsighand->count, 1);
882                 memcpy(newsighand->action, oldsighand->action,
883                        sizeof(newsighand->action));
884
885                 write_lock_irq(&tasklist_lock);
886                 spin_lock(&oldsighand->siglock);
887                 rcu_assign_pointer(tsk->sighand, newsighand);
888                 spin_unlock(&oldsighand->siglock);
889                 write_unlock_irq(&tasklist_lock);
890
891                 __cleanup_sighand(oldsighand);
892         }
893
894         BUG_ON(!thread_group_leader(tsk));
895         return 0;
896 }
897
898 /*
899  * These functions flushes out all traces of the currently running executable
900  * so that a new one can be started
901  */
902 static void flush_old_files(struct files_struct * files)
903 {
904         long j = -1;
905         struct fdtable *fdt;
906
907         spin_lock(&files->file_lock);
908         for (;;) {
909                 unsigned long set, i;
910
911                 j++;
912                 i = j * __NFDBITS;
913                 fdt = files_fdtable(files);
914                 if (i >= fdt->max_fds)
915                         break;
916                 set = fdt->close_on_exec->fds_bits[j];
917                 if (!set)
918                         continue;
919                 fdt->close_on_exec->fds_bits[j] = 0;
920                 spin_unlock(&files->file_lock);
921                 for ( ; set ; i++,set >>= 1) {
922                         if (set & 1) {
923                                 sys_close(i);
924                         }
925                 }
926                 spin_lock(&files->file_lock);
927
928         }
929         spin_unlock(&files->file_lock);
930 }
931
932 char *get_task_comm(char *buf, struct task_struct *tsk)
933 {
934         /* buf must be at least sizeof(tsk->comm) in size */
935         task_lock(tsk);
936         strncpy(buf, tsk->comm, sizeof(tsk->comm));
937         task_unlock(tsk);
938         return buf;
939 }
940
941 void set_task_comm(struct task_struct *tsk, char *buf)
942 {
943         task_lock(tsk);
944         strlcpy(tsk->comm, buf, sizeof(tsk->comm));
945         task_unlock(tsk);
946 }
947
948 int flush_old_exec(struct linux_binprm * bprm)
949 {
950         char * name;
951         int i, ch, retval;
952         char tcomm[sizeof(current->comm)];
953
954         /*
955          * Make sure we have a private signal table and that
956          * we are unassociated from the previous thread group.
957          */
958         retval = de_thread(current);
959         if (retval)
960                 goto out;
961
962         set_mm_exe_file(bprm->mm, bprm->file);
963
964         /*
965          * Release all of the old mmap stuff
966          */
967         retval = exec_mmap(bprm->mm);
968         if (retval)
969                 goto out;
970
971         bprm->mm = NULL;                /* We're using it now */
972
973         /* This is the point of no return */
974         current->sas_ss_sp = current->sas_ss_size = 0;
975
976         if (current_euid() == current_uid() && current_egid() == current_gid())
977                 set_dumpable(current->mm, 1);
978         else
979                 set_dumpable(current->mm, suid_dumpable);
980
981         name = bprm->filename;
982
983         /* Copies the binary name from after last slash */
984         for (i=0; (ch = *(name++)) != '\0';) {
985                 if (ch == '/')
986                         i = 0; /* overwrite what we wrote */
987                 else
988                         if (i < (sizeof(tcomm) - 1))
989                                 tcomm[i++] = ch;
990         }
991         tcomm[i] = '\0';
992         set_task_comm(current, tcomm);
993
994         current->flags &= ~PF_RANDOMIZE;
995         flush_thread();
996
997         /* Set the new mm task size. We have to do that late because it may
998          * depend on TIF_32BIT which is only updated in flush_thread() on
999          * some architectures like powerpc
1000          */
1001         current->mm->task_size = TASK_SIZE;
1002
1003         /* install the new credentials */
1004         if (bprm->cred->uid != current_euid() ||
1005             bprm->cred->gid != current_egid()) {
1006                 current->pdeath_signal = 0;
1007         } else if (file_permission(bprm->file, MAY_READ) ||
1008                    bprm->interp_flags & BINPRM_FLAGS_ENFORCE_NONDUMP) {
1009                 set_dumpable(current->mm, suid_dumpable);
1010         }
1011
1012         current->personality &= ~bprm->per_clear;
1013
1014         /*
1015          * Flush performance counters when crossing a
1016          * security domain:
1017          */
1018         if (!get_dumpable(current->mm))
1019                 perf_counter_exit_task(current);
1020
1021         /* An exec changes our domain. We are no longer part of the thread
1022            group */
1023
1024         current->self_exec_id++;
1025                         
1026         flush_signal_handlers(current, 0);
1027         flush_old_files(current->files);
1028
1029         return 0;
1030
1031 out:
1032         return retval;
1033 }
1034
1035 EXPORT_SYMBOL(flush_old_exec);
1036
1037 /*
1038  * install the new credentials for this executable
1039  */
1040 void install_exec_creds(struct linux_binprm *bprm)
1041 {
1042         security_bprm_committing_creds(bprm);
1043
1044         commit_creds(bprm->cred);
1045         bprm->cred = NULL;
1046
1047         /* cred_exec_mutex must be held at least to this point to prevent
1048          * ptrace_attach() from altering our determination of the task's
1049          * credentials; any time after this it may be unlocked */
1050
1051         security_bprm_committed_creds(bprm);
1052 }
1053 EXPORT_SYMBOL(install_exec_creds);
1054
1055 /*
1056  * determine how safe it is to execute the proposed program
1057  * - the caller must hold current->cred_exec_mutex to protect against
1058  *   PTRACE_ATTACH
1059  */
1060 void check_unsafe_exec(struct linux_binprm *bprm)
1061 {
1062         struct task_struct *p = current;
1063
1064         bprm->unsafe = tracehook_unsafe_exec(p);
1065
1066         if (atomic_read(&p->fs->count) > 1 ||
1067             atomic_read(&p->files->count) > 1 ||
1068             atomic_read(&p->sighand->count) > 1)
1069                 bprm->unsafe |= LSM_UNSAFE_SHARE;
1070 }
1071
1072 /* 
1073  * Fill the binprm structure from the inode. 
1074  * Check permissions, then read the first 128 (BINPRM_BUF_SIZE) bytes
1075  *
1076  * This may be called multiple times for binary chains (scripts for example).
1077  */
1078 int prepare_binprm(struct linux_binprm *bprm)
1079 {
1080         umode_t mode;
1081         struct inode * inode = bprm->file->f_path.dentry->d_inode;
1082         int retval;
1083
1084         mode = inode->i_mode;
1085         if (bprm->file->f_op == NULL)
1086                 return -EACCES;
1087
1088         /* clear any previous set[ug]id data from a previous binary */
1089         bprm->cred->euid = current_euid();
1090         bprm->cred->egid = current_egid();
1091
1092         if (!(bprm->file->f_path.mnt->mnt_flags & MNT_NOSUID)) {
1093                 /* Set-uid? */
1094                 if (mode & S_ISUID) {
1095                         bprm->per_clear |= PER_CLEAR_ON_SETID;
1096                         bprm->cred->euid = inode->i_uid;
1097                 }
1098
1099                 /* Set-gid? */
1100                 /*
1101                  * If setgid is set but no group execute bit then this
1102                  * is a candidate for mandatory locking, not a setgid
1103                  * executable.
1104                  */
1105                 if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP)) {
1106                         bprm->per_clear |= PER_CLEAR_ON_SETID;
1107                         bprm->cred->egid = inode->i_gid;
1108                 }
1109         }
1110
1111         /* fill in binprm security blob */
1112         retval = security_bprm_set_creds(bprm);
1113         if (retval)
1114                 return retval;
1115         bprm->cred_prepared = 1;
1116
1117         memset(bprm->buf, 0, BINPRM_BUF_SIZE);
1118         return kernel_read(bprm->file, 0, bprm->buf, BINPRM_BUF_SIZE);
1119 }
1120
1121 EXPORT_SYMBOL(prepare_binprm);
1122
1123 /*
1124  * Arguments are '\0' separated strings found at the location bprm->p
1125  * points to; chop off the first by relocating brpm->p to right after
1126  * the first '\0' encountered.
1127  */
1128 int remove_arg_zero(struct linux_binprm *bprm)
1129 {
1130         int ret = 0;
1131         unsigned long offset;
1132         char *kaddr;
1133         struct page *page;
1134
1135         if (!bprm->argc)
1136                 return 0;
1137
1138         do {
1139                 offset = bprm->p & ~PAGE_MASK;
1140                 page = get_arg_page(bprm, bprm->p, 0);
1141                 if (!page) {
1142                         ret = -EFAULT;
1143                         goto out;
1144                 }
1145                 kaddr = kmap_atomic(page, KM_USER0);
1146
1147                 for (; offset < PAGE_SIZE && kaddr[offset];
1148                                 offset++, bprm->p++)
1149                         ;
1150
1151                 kunmap_atomic(kaddr, KM_USER0);
1152                 put_arg_page(page);
1153
1154                 if (offset == PAGE_SIZE)
1155                         free_arg_page(bprm, (bprm->p >> PAGE_SHIFT) - 1);
1156         } while (offset == PAGE_SIZE);
1157
1158         bprm->p++;
1159         bprm->argc--;
1160         ret = 0;
1161
1162 out:
1163         return ret;
1164 }
1165 EXPORT_SYMBOL(remove_arg_zero);
1166
1167 /*
1168  * cycle the list of binary formats handler, until one recognizes the image
1169  */
1170 int search_binary_handler(struct linux_binprm *bprm,struct pt_regs *regs)
1171 {
1172         unsigned int depth = bprm->recursion_depth;
1173         int try,retval;
1174         struct linux_binfmt *fmt;
1175
1176         retval = security_bprm_check(bprm);
1177         if (retval)
1178                 return retval;
1179
1180         /* kernel module loader fixup */
1181         /* so we don't try to load run modprobe in kernel space. */
1182         set_fs(USER_DS);
1183
1184         retval = audit_bprm(bprm);
1185         if (retval)
1186                 return retval;
1187
1188         retval = -ENOENT;
1189         for (try=0; try<2; try++) {
1190                 read_lock(&binfmt_lock);
1191                 list_for_each_entry(fmt, &formats, lh) {
1192                         int (*fn)(struct linux_binprm *, struct pt_regs *) = fmt->load_binary;
1193                         if (!fn)
1194                                 continue;
1195                         if (!try_module_get(fmt->module))
1196                                 continue;
1197                         read_unlock(&binfmt_lock);
1198                         retval = fn(bprm, regs);
1199                         /*
1200                          * Restore the depth counter to its starting value
1201                          * in this call, so we don't have to rely on every
1202                          * load_binary function to restore it on return.
1203                          */
1204                         bprm->recursion_depth = depth;
1205                         if (retval >= 0) {
1206                                 if (depth == 0)
1207                                         tracehook_report_exec(fmt, bprm, regs);
1208                                 put_binfmt(fmt);
1209                                 allow_write_access(bprm->file);
1210                                 if (bprm->file)
1211                                         fput(bprm->file);
1212                                 bprm->file = NULL;
1213                                 current->did_exec = 1;
1214                                 proc_exec_connector(current);
1215                                 return retval;
1216                         }
1217                         read_lock(&binfmt_lock);
1218                         put_binfmt(fmt);
1219                         if (retval != -ENOEXEC || bprm->mm == NULL)
1220                                 break;
1221                         if (!bprm->file) {
1222                                 read_unlock(&binfmt_lock);
1223                                 return retval;
1224                         }
1225                 }
1226                 read_unlock(&binfmt_lock);
1227                 if (retval != -ENOEXEC || bprm->mm == NULL) {
1228                         break;
1229 #ifdef CONFIG_MODULES
1230                 } else {
1231 #define printable(c) (((c)=='\t') || ((c)=='\n') || (0x20<=(c) && (c)<=0x7e))
1232                         if (printable(bprm->buf[0]) &&
1233                             printable(bprm->buf[1]) &&
1234                             printable(bprm->buf[2]) &&
1235                             printable(bprm->buf[3]))
1236                                 break; /* -ENOEXEC */
1237                         request_module("binfmt-%04x", *(unsigned short *)(&bprm->buf[2]));
1238 #endif
1239                 }
1240         }
1241         return retval;
1242 }
1243
1244 EXPORT_SYMBOL(search_binary_handler);
1245
1246 void free_bprm(struct linux_binprm *bprm)
1247 {
1248         free_arg_pages(bprm);
1249         if (bprm->cred)
1250                 abort_creds(bprm->cred);
1251         kfree(bprm);
1252 }
1253
1254 /*
1255  * sys_execve() executes a new program.
1256  */
1257 int do_execve(char * filename,
1258         char __user *__user *argv,
1259         char __user *__user *envp,
1260         struct pt_regs * regs)
1261 {
1262         struct linux_binprm *bprm;
1263         struct file *file;
1264         struct files_struct *displaced;
1265         int retval;
1266
1267         retval = unshare_files(&displaced);
1268         if (retval)
1269                 goto out_ret;
1270
1271         retval = -ENOMEM;
1272         bprm = kzalloc(sizeof(*bprm), GFP_KERNEL);
1273         if (!bprm)
1274                 goto out_files;
1275
1276         retval = mutex_lock_interruptible(&current->cred_exec_mutex);
1277         if (retval < 0)
1278                 goto out_free;
1279
1280         retval = -ENOMEM;
1281         bprm->cred = prepare_exec_creds();
1282         if (!bprm->cred)
1283                 goto out_unlock;
1284         check_unsafe_exec(bprm);
1285
1286         file = open_exec(filename);
1287         retval = PTR_ERR(file);
1288         if (IS_ERR(file))
1289                 goto out_unlock;
1290
1291         sched_exec();
1292
1293         bprm->file = file;
1294         bprm->filename = filename;
1295         bprm->interp = filename;
1296
1297         retval = bprm_mm_init(bprm);
1298         if (retval)
1299                 goto out_file;
1300
1301         bprm->argc = count(argv, MAX_ARG_STRINGS);
1302         if ((retval = bprm->argc) < 0)
1303                 goto out;
1304
1305         bprm->envc = count(envp, MAX_ARG_STRINGS);
1306         if ((retval = bprm->envc) < 0)
1307                 goto out;
1308
1309         retval = prepare_binprm(bprm);
1310         if (retval < 0)
1311                 goto out;
1312
1313         retval = copy_strings_kernel(1, &bprm->filename, bprm);
1314         if (retval < 0)
1315                 goto out;
1316
1317         bprm->exec = bprm->p;
1318         retval = copy_strings(bprm->envc, envp, bprm);
1319         if (retval < 0)
1320                 goto out;
1321
1322         retval = copy_strings(bprm->argc, argv, bprm);
1323         if (retval < 0)
1324                 goto out;
1325
1326         current->flags &= ~PF_KTHREAD;
1327         retval = search_binary_handler(bprm,regs);
1328         if (retval < 0)
1329                 goto out;
1330
1331         /* execve succeeded */
1332         mutex_unlock(&current->cred_exec_mutex);
1333         acct_update_integrals(current);
1334         free_bprm(bprm);
1335         if (displaced)
1336                 put_files_struct(displaced);
1337         return retval;
1338
1339 out:
1340         if (bprm->mm)
1341                 mmput (bprm->mm);
1342
1343 out_file:
1344         if (bprm->file) {
1345                 allow_write_access(bprm->file);
1346                 fput(bprm->file);
1347         }
1348
1349 out_unlock:
1350         mutex_unlock(&current->cred_exec_mutex);
1351
1352 out_free:
1353         free_bprm(bprm);
1354
1355 out_files:
1356         if (displaced)
1357                 reset_files_struct(displaced);
1358 out_ret:
1359         return retval;
1360 }
1361
1362 int set_binfmt(struct linux_binfmt *new)
1363 {
1364         struct linux_binfmt *old = current->binfmt;
1365
1366         if (new) {
1367                 if (!try_module_get(new->module))
1368                         return -1;
1369         }
1370         current->binfmt = new;
1371         if (old)
1372                 module_put(old->module);
1373         return 0;
1374 }
1375
1376 EXPORT_SYMBOL(set_binfmt);
1377
1378 /* format_corename will inspect the pattern parameter, and output a
1379  * name into corename, which must have space for at least
1380  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
1381  */
1382 static int format_corename(char *corename, long signr)
1383 {
1384         const struct cred *cred = current_cred();
1385         const char *pat_ptr = core_pattern;
1386         int ispipe = (*pat_ptr == '|');
1387         char *out_ptr = corename;
1388         char *const out_end = corename + CORENAME_MAX_SIZE;
1389         int rc;
1390         int pid_in_pattern = 0;
1391
1392         /* Repeat as long as we have more pattern to process and more output
1393            space */
1394         while (*pat_ptr) {
1395                 if (*pat_ptr != '%') {
1396                         if (out_ptr == out_end)
1397                                 goto out;
1398                         *out_ptr++ = *pat_ptr++;
1399                 } else {
1400                         switch (*++pat_ptr) {
1401                         case 0:
1402                                 goto out;
1403                         /* Double percent, output one percent */
1404                         case '%':
1405                                 if (out_ptr == out_end)
1406                                         goto out;
1407                                 *out_ptr++ = '%';
1408                                 break;
1409                         /* pid */
1410                         case 'p':
1411                                 pid_in_pattern = 1;
1412                                 rc = snprintf(out_ptr, out_end - out_ptr,
1413                                               "%d", task_tgid_vnr(current));
1414                                 if (rc > out_end - out_ptr)
1415                                         goto out;
1416                                 out_ptr += rc;
1417                                 break;
1418                         /* uid */
1419                         case 'u':
1420                                 rc = snprintf(out_ptr, out_end - out_ptr,
1421                                               "%d", cred->uid);
1422                                 if (rc > out_end - out_ptr)
1423                                         goto out;
1424                                 out_ptr += rc;
1425                                 break;
1426                         /* gid */
1427                         case 'g':
1428                                 rc = snprintf(out_ptr, out_end - out_ptr,
1429                                               "%d", cred->gid);
1430                                 if (rc > out_end - out_ptr)
1431                                         goto out;
1432                                 out_ptr += rc;
1433                                 break;
1434                         /* signal that caused the coredump */
1435                         case 's':
1436                                 rc = snprintf(out_ptr, out_end - out_ptr,
1437                                               "%ld", signr);
1438                                 if (rc > out_end - out_ptr)
1439                                         goto out;
1440                                 out_ptr += rc;
1441                                 break;
1442                         /* UNIX time of coredump */
1443                         case 't': {
1444                                 struct timeval tv;
1445                                 do_gettimeofday(&tv);
1446                                 rc = snprintf(out_ptr, out_end - out_ptr,
1447                                               "%lu", tv.tv_sec);
1448                                 if (rc > out_end - out_ptr)
1449                                         goto out;
1450                                 out_ptr += rc;
1451                                 break;
1452                         }
1453                         /* hostname */
1454                         case 'h':
1455                                 down_read(&uts_sem);
1456                                 rc = snprintf(out_ptr, out_end - out_ptr,
1457                                               "%s", utsname()->nodename);
1458                                 up_read(&uts_sem);
1459                                 if (rc > out_end - out_ptr)
1460                                         goto out;
1461                                 out_ptr += rc;
1462                                 break;
1463                         /* executable */
1464                         case 'e':
1465                                 rc = snprintf(out_ptr, out_end - out_ptr,
1466                                               "%s", current->comm);
1467                                 if (rc > out_end - out_ptr)
1468                                         goto out;
1469                                 out_ptr += rc;
1470                                 break;
1471                         /* core limit size */
1472                         case 'c':
1473                                 rc = snprintf(out_ptr, out_end - out_ptr,
1474                                               "%lu", current->signal->rlim[RLIMIT_CORE].rlim_cur);
1475                                 if (rc > out_end - out_ptr)
1476                                         goto out;
1477                                 out_ptr += rc;
1478                                 break;
1479                         default:
1480                                 break;
1481                         }
1482                         ++pat_ptr;
1483                 }
1484         }
1485         /* Backward compatibility with core_uses_pid:
1486          *
1487          * If core_pattern does not include a %p (as is the default)
1488          * and core_uses_pid is set, then .%pid will be appended to
1489          * the filename. Do not do this for piped commands. */
1490         if (!ispipe && !pid_in_pattern && core_uses_pid) {
1491                 rc = snprintf(out_ptr, out_end - out_ptr,
1492                               ".%d", task_tgid_vnr(current));
1493                 if (rc > out_end - out_ptr)
1494                         goto out;
1495                 out_ptr += rc;
1496         }
1497 out:
1498         *out_ptr = 0;
1499         return ispipe;
1500 }
1501
1502 static int zap_process(struct task_struct *start)
1503 {
1504         struct task_struct *t;
1505         int nr = 0;
1506
1507         start->signal->flags = SIGNAL_GROUP_EXIT;
1508         start->signal->group_stop_count = 0;
1509
1510         t = start;
1511         do {
1512                 if (t != current && t->mm) {
1513                         sigaddset(&t->pending.signal, SIGKILL);
1514                         signal_wake_up(t, 1);
1515                         nr++;
1516                 }
1517         } while_each_thread(start, t);
1518
1519         return nr;
1520 }
1521
1522 static inline int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
1523                                 struct core_state *core_state, int exit_code)
1524 {
1525         struct task_struct *g, *p;
1526         unsigned long flags;
1527         int nr = -EAGAIN;
1528
1529         spin_lock_irq(&tsk->sighand->siglock);
1530         if (!signal_group_exit(tsk->signal)) {
1531                 mm->core_state = core_state;
1532                 tsk->signal->group_exit_code = exit_code;
1533                 nr = zap_process(tsk);
1534         }
1535         spin_unlock_irq(&tsk->sighand->siglock);
1536         if (unlikely(nr < 0))
1537                 return nr;
1538
1539         if (atomic_read(&mm->mm_users) == nr + 1)
1540                 goto done;
1541         /*
1542          * We should find and kill all tasks which use this mm, and we should
1543          * count them correctly into ->nr_threads. We don't take tasklist
1544          * lock, but this is safe wrt:
1545          *
1546          * fork:
1547          *      None of sub-threads can fork after zap_process(leader). All
1548          *      processes which were created before this point should be
1549          *      visible to zap_threads() because copy_process() adds the new
1550          *      process to the tail of init_task.tasks list, and lock/unlock
1551          *      of ->siglock provides a memory barrier.
1552          *
1553          * do_exit:
1554          *      The caller holds mm->mmap_sem. This means that the task which
1555          *      uses this mm can't pass exit_mm(), so it can't exit or clear
1556          *      its ->mm.
1557          *
1558          * de_thread:
1559          *      It does list_replace_rcu(&leader->tasks, &current->tasks),
1560          *      we must see either old or new leader, this does not matter.
1561          *      However, it can change p->sighand, so lock_task_sighand(p)
1562          *      must be used. Since p->mm != NULL and we hold ->mmap_sem
1563          *      it can't fail.
1564          *
1565          *      Note also that "g" can be the old leader with ->mm == NULL
1566          *      and already unhashed and thus removed from ->thread_group.
1567          *      This is OK, __unhash_process()->list_del_rcu() does not
1568          *      clear the ->next pointer, we will find the new leader via
1569          *      next_thread().
1570          */
1571         rcu_read_lock();
1572         for_each_process(g) {
1573                 if (g == tsk->group_leader)
1574                         continue;
1575                 if (g->flags & PF_KTHREAD)
1576                         continue;
1577                 p = g;
1578                 do {
1579                         if (p->mm) {
1580                                 if (unlikely(p->mm == mm)) {
1581                                         lock_task_sighand(p, &flags);
1582                                         nr += zap_process(p);
1583                                         unlock_task_sighand(p, &flags);
1584                                 }
1585                                 break;
1586                         }
1587                 } while_each_thread(g, p);
1588         }
1589         rcu_read_unlock();
1590 done:
1591         atomic_set(&core_state->nr_threads, nr);
1592         return nr;
1593 }
1594
1595 static int coredump_wait(int exit_code, struct core_state *core_state)
1596 {
1597         struct task_struct *tsk = current;
1598         struct mm_struct *mm = tsk->mm;
1599         struct completion *vfork_done;
1600         int core_waiters;
1601
1602         init_completion(&core_state->startup);
1603         core_state->dumper.task = tsk;
1604         core_state->dumper.next = NULL;
1605         core_waiters = zap_threads(tsk, mm, core_state, exit_code);
1606         up_write(&mm->mmap_sem);
1607
1608         if (unlikely(core_waiters < 0))
1609                 goto fail;
1610
1611         /*
1612          * Make sure nobody is waiting for us to release the VM,
1613          * otherwise we can deadlock when we wait on each other
1614          */
1615         vfork_done = tsk->vfork_done;
1616         if (vfork_done) {
1617                 tsk->vfork_done = NULL;
1618                 complete(vfork_done);
1619         }
1620
1621         if (core_waiters)
1622                 wait_for_completion(&core_state->startup);
1623 fail:
1624         return core_waiters;
1625 }
1626
1627 static void coredump_finish(struct mm_struct *mm)
1628 {
1629         struct core_thread *curr, *next;
1630         struct task_struct *task;
1631
1632         next = mm->core_state->dumper.next;
1633         while ((curr = next) != NULL) {
1634                 next = curr->next;
1635                 task = curr->task;
1636                 /*
1637                  * see exit_mm(), curr->task must not see
1638                  * ->task == NULL before we read ->next.
1639                  */
1640                 smp_mb();
1641                 curr->task = NULL;
1642                 wake_up_process(task);
1643         }
1644
1645         mm->core_state = NULL;
1646 }
1647
1648 /*
1649  * set_dumpable converts traditional three-value dumpable to two flags and
1650  * stores them into mm->flags.  It modifies lower two bits of mm->flags, but
1651  * these bits are not changed atomically.  So get_dumpable can observe the
1652  * intermediate state.  To avoid doing unexpected behavior, get get_dumpable
1653  * return either old dumpable or new one by paying attention to the order of
1654  * modifying the bits.
1655  *
1656  * dumpable |   mm->flags (binary)
1657  * old  new | initial interim  final
1658  * ---------+-----------------------
1659  *  0    1  |   00      01      01
1660  *  0    2  |   00      10(*)   11
1661  *  1    0  |   01      00      00
1662  *  1    2  |   01      11      11
1663  *  2    0  |   11      10(*)   00
1664  *  2    1  |   11      11      01
1665  *
1666  * (*) get_dumpable regards interim value of 10 as 11.
1667  */
1668 void set_dumpable(struct mm_struct *mm, int value)
1669 {
1670         switch (value) {
1671         case 0:
1672                 clear_bit(MMF_DUMPABLE, &mm->flags);
1673                 smp_wmb();
1674                 clear_bit(MMF_DUMP_SECURELY, &mm->flags);
1675                 break;
1676         case 1:
1677                 set_bit(MMF_DUMPABLE, &mm->flags);
1678                 smp_wmb();
1679                 clear_bit(MMF_DUMP_SECURELY, &mm->flags);
1680                 break;
1681         case 2:
1682                 set_bit(MMF_DUMP_SECURELY, &mm->flags);
1683                 smp_wmb();
1684                 set_bit(MMF_DUMPABLE, &mm->flags);
1685                 break;
1686         }
1687 }
1688
1689 int get_dumpable(struct mm_struct *mm)
1690 {
1691         int ret;
1692
1693         ret = mm->flags & 0x3;
1694         return (ret >= 2) ? 2 : ret;
1695 }
1696
1697 void do_coredump(long signr, int exit_code, struct pt_regs *regs)
1698 {
1699         struct core_state core_state;
1700         char corename[CORENAME_MAX_SIZE + 1];
1701         struct mm_struct *mm = current->mm;
1702         struct linux_binfmt * binfmt;
1703         struct inode * inode;
1704         struct file * file;
1705         const struct cred *old_cred;
1706         struct cred *cred;
1707         int retval = 0;
1708         int flag = 0;
1709         int ispipe = 0;
1710         unsigned long core_limit = current->signal->rlim[RLIMIT_CORE].rlim_cur;
1711         char **helper_argv = NULL;
1712         int helper_argc = 0;
1713         char *delimit;
1714
1715         audit_core_dumps(signr);
1716
1717         binfmt = current->binfmt;
1718         if (!binfmt || !binfmt->core_dump)
1719                 goto fail;
1720
1721         cred = prepare_creds();
1722         if (!cred) {
1723                 retval = -ENOMEM;
1724                 goto fail;
1725         }
1726
1727         down_write(&mm->mmap_sem);
1728         /*
1729          * If another thread got here first, or we are not dumpable, bail out.
1730          */
1731         if (mm->core_state || !get_dumpable(mm)) {
1732                 up_write(&mm->mmap_sem);
1733                 put_cred(cred);
1734                 goto fail;
1735         }
1736
1737         /*
1738          *      We cannot trust fsuid as being the "true" uid of the
1739          *      process nor do we know its entire history. We only know it
1740          *      was tainted so we dump it as root in mode 2.
1741          */
1742         if (get_dumpable(mm) == 2) {    /* Setuid core dump mode */
1743                 flag = O_EXCL;          /* Stop rewrite attacks */
1744                 cred->fsuid = 0;        /* Dump root private */
1745         }
1746
1747         retval = coredump_wait(exit_code, &core_state);
1748         if (retval < 0) {
1749                 put_cred(cred);
1750                 goto fail;
1751         }
1752
1753         old_cred = override_creds(cred);
1754
1755         /*
1756          * Clear any false indication of pending signals that might
1757          * be seen by the filesystem code called to write the core file.
1758          */
1759         clear_thread_flag(TIF_SIGPENDING);
1760
1761         /*
1762          * lock_kernel() because format_corename() is controlled by sysctl, which
1763          * uses lock_kernel()
1764          */
1765         lock_kernel();
1766         ispipe = format_corename(corename, signr);
1767         unlock_kernel();
1768         /*
1769          * Don't bother to check the RLIMIT_CORE value if core_pattern points
1770          * to a pipe.  Since we're not writing directly to the filesystem
1771          * RLIMIT_CORE doesn't really apply, as no actual core file will be
1772          * created unless the pipe reader choses to write out the core file
1773          * at which point file size limits and permissions will be imposed
1774          * as it does with any other process
1775          */
1776         if ((!ispipe) && (core_limit < binfmt->min_coredump))
1777                 goto fail_unlock;
1778
1779         if (ispipe) {
1780                 helper_argv = argv_split(GFP_KERNEL, corename+1, &helper_argc);
1781                 if (!helper_argv) {
1782                         printk(KERN_WARNING "%s failed to allocate memory\n",
1783                                __func__);
1784                         goto fail_unlock;
1785                 }
1786                 /* Terminate the string before the first option */
1787                 delimit = strchr(corename, ' ');
1788                 if (delimit)
1789                         *delimit = '\0';
1790                 delimit = strrchr(helper_argv[0], '/');
1791                 if (delimit)
1792                         delimit++;
1793                 else
1794                         delimit = helper_argv[0];
1795                 if (!strcmp(delimit, current->comm)) {
1796                         printk(KERN_NOTICE "Recursive core dump detected, "
1797                                         "aborting\n");
1798                         goto fail_unlock;
1799                 }
1800
1801                 core_limit = RLIM_INFINITY;
1802
1803                 /* SIGPIPE can happen, but it's just never processed */
1804                 if (call_usermodehelper_pipe(corename+1, helper_argv, NULL,
1805                                 &file)) {
1806                         printk(KERN_INFO "Core dump to %s pipe failed\n",
1807                                corename);
1808                         goto fail_unlock;
1809                 }
1810         } else
1811                 file = filp_open(corename,
1812                                  O_CREAT | 2 | O_NOFOLLOW | O_LARGEFILE | flag,
1813                                  0600);
1814         if (IS_ERR(file))
1815                 goto fail_unlock;
1816         inode = file->f_path.dentry->d_inode;
1817         if (inode->i_nlink > 1)
1818                 goto close_fail;        /* multiple links - don't dump */
1819         if (!ispipe && d_unhashed(file->f_path.dentry))
1820                 goto close_fail;
1821
1822         /* AK: actually i see no reason to not allow this for named pipes etc.,
1823            but keep the previous behaviour for now. */
1824         if (!ispipe && !S_ISREG(inode->i_mode))
1825                 goto close_fail;
1826         /*
1827          * Dont allow local users get cute and trick others to coredump
1828          * into their pre-created files:
1829          */
1830         if (inode->i_uid != current_fsuid())
1831                 goto close_fail;
1832         if (!file->f_op)
1833                 goto close_fail;
1834         if (!file->f_op->write)
1835                 goto close_fail;
1836         if (!ispipe && do_truncate(file->f_path.dentry, 0, 0, file) != 0)
1837                 goto close_fail;
1838
1839         retval = binfmt->core_dump(signr, regs, file, core_limit);
1840
1841         if (retval)
1842                 current->signal->group_exit_code |= 0x80;
1843 close_fail:
1844         filp_close(file, NULL);
1845 fail_unlock:
1846         if (helper_argv)
1847                 argv_free(helper_argv);
1848
1849         revert_creds(old_cred);
1850         put_cred(cred);
1851         coredump_finish(mm);
1852 fail:
1853         return;
1854 }