fs/binfmt_elf_fdpic.c

   1 /* binfmt_elf_fdpic.c: FDPIC ELF binary format
   2  *
   3  * Copyright (C) 2003, 2004, 2006 Red Hat, Inc. All Rights Reserved.
   4  * Written by David Howells (dhowells@redhat.com)
   5  * Derived from binfmt_elf.c
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public License
   9  * as published by the Free Software Foundation; either version
  10  * 2 of the License, or (at your option) any later version.
  11  */
  12
  13 #include <linux/module.h>
  14
  15 #include <linux/fs.h>
  16 #include <linux/stat.h>
  17 #include <linux/sched.h>
  18 #include <linux/mm.h>
  19 #include <linux/mman.h>
  20 #include <linux/errno.h>
  21 #include <linux/signal.h>
  22 #include <linux/binfmts.h>
  23 #include <linux/string.h>
  24 #include <linux/file.h>
  25 #include <linux/fcntl.h>
  26 #include <linux/slab.h>
  27 #include <linux/pagemap.h>
  28 #include <linux/highmem.h>
  29 #include <linux/highuid.h>
  30 #include <linux/personality.h>
  31 #include <linux/ptrace.h>
  32 #include <linux/init.h>
  33 #include <linux/smp_lock.h>
  34 #include <linux/elf.h>
  35 #include <linux/elf-fdpic.h>
  36 #include <linux/elfcore.h>
  37 #include <linux/vs_base.h>
  38 #include <linux/vs_cvirt.h>
  39
  40 #include <asm/uaccess.h>
  41 #include <asm/param.h>
  42 #include <asm/pgalloc.h>
  43
  44 typedef char *elf_caddr_t;
  45 #ifndef elf_addr_t
  46 #define elf_addr_t unsigned long
  47 #endif
  48
  49 #if 0
  50 #define kdebug(fmt, ...) printk("FDPIC "fmt"\n" ,##__VA_ARGS__ )
  51 #else
  52 #define kdebug(fmt, ...) do {} while(0)
  53 #endif
  54
  55 #if 0
  56 #define kdcore(fmt, ...) printk("FDPIC "fmt"\n" ,##__VA_ARGS__ )
  57 #else
  58 #define kdcore(fmt, ...) do {} while(0)
  59 #endif
  60
  61 MODULE_LICENSE("GPL");
  62
  63 static int load_elf_fdpic_binary(struct linux_binprm *, struct pt_regs *);
  64 static int elf_fdpic_fetch_phdrs(struct elf_fdpic_params *, struct file *);
  65 static int elf_fdpic_map_file(struct elf_fdpic_params *, struct file *,
  66                               struct mm_struct *, const char *);
  67
  68 static int create_elf_fdpic_tables(struct linux_binprm *, struct mm_struct *,
  69                                    struct elf_fdpic_params *,
  70                                    struct elf_fdpic_params *);
  71
  72 #ifndef CONFIG_MMU
  73 static int elf_fdpic_transfer_args_to_stack(struct linux_binprm *,
  74                                             unsigned long *);
  75 static int elf_fdpic_map_file_constdisp_on_uclinux(struct elf_fdpic_params *,
  76                                                    struct file *,
  77                                                    struct mm_struct *);
  78 #endif
  79
  80 static int elf_fdpic_map_file_by_direct_mmap(struct elf_fdpic_params *,
  81                                              struct file *, struct mm_struct *);
  82
  83 #if defined(USE_ELF_CORE_DUMP) && defined(CONFIG_ELF_CORE)
  84 static int elf_fdpic_core_dump(long, struct pt_regs *, struct file *);
  85 #endif
  86
  87 static struct linux_binfmt elf_fdpic_format = {
  88         .module         = THIS_MODULE,
  89         .load_binary    = load_elf_fdpic_binary,
  90 #if defined(USE_ELF_CORE_DUMP) && defined(CONFIG_ELF_CORE)
  91         .core_dump      = elf_fdpic_core_dump,
  92 #endif
  93         .min_coredump   = ELF_EXEC_PAGESIZE,
  94 };
  95
  96 static int __init init_elf_fdpic_binfmt(void)
  97 {
  98         return register_binfmt(&elf_fdpic_format);
  99 }
 100
 101 static void __exit exit_elf_fdpic_binfmt(void)
 102 {
 103         unregister_binfmt(&elf_fdpic_format);
 104 }
 105
 106 core_initcall(init_elf_fdpic_binfmt);
 107 module_exit(exit_elf_fdpic_binfmt);
 108
 109 static int is_elf_fdpic(struct elfhdr *hdr, struct file *file)
 110 {
 111         if (memcmp(hdr->e_ident, ELFMAG, SELFMAG) != 0)
 112                 return 0;
 113         if (hdr->e_type != ET_EXEC && hdr->e_type != ET_DYN)
 114                 return 0;
 115         if (!elf_check_arch(hdr) || !elf_check_fdpic(hdr))
 116                 return 0;
 117         if (!file->f_op || !file->f_op->mmap)
 118                 return 0;
 119         return 1;
 120 }
 121
 122 /*****************************************************************************/
 123 /*
 124  * read the program headers table into memory
 125  */
 126 static int elf_fdpic_fetch_phdrs(struct elf_fdpic_params *params,
 127                                  struct file *file)
 128 {
 129         struct elf32_phdr *phdr;
 130         unsigned long size;
 131         int retval, loop;
 132
 133         if (params->hdr.e_phentsize != sizeof(struct elf_phdr))
 134                 return -ENOMEM;
 135         if (params->hdr.e_phnum > 65536U / sizeof(struct elf_phdr))
 136                 return -ENOMEM;
 137
 138         size = params->hdr.e_phnum * sizeof(struct elf_phdr);
 139         params->phdrs = kmalloc(size, GFP_KERNEL);
 140         if (!params->phdrs)
 141                 return -ENOMEM;
 142
 143         retval = kernel_read(file, params->hdr.e_phoff,
 144                              (char *) params->phdrs, size);
 145         if (retval < 0)
 146                 return retval;
 147
 148         /* determine stack size for this binary */
 149         phdr = params->phdrs;
 150         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 151                 if (phdr->p_type != PT_GNU_STACK)
 152                         continue;
 153
 154                 if (phdr->p_flags & PF_X)
 155                         params->flags |= ELF_FDPIC_FLAG_EXEC_STACK;
 156                 else
 157                         params->flags |= ELF_FDPIC_FLAG_NOEXEC_STACK;
 158
 159                 params->stack_size = phdr->p_memsz;
 160                 break;
 161         }
 162
 163         return 0;
 164 }
 165
 166 /*****************************************************************************/
 167 /*
 168  * load an fdpic binary into various bits of memory
 169  */
 170 static int load_elf_fdpic_binary(struct linux_binprm *bprm,
 171                                  struct pt_regs *regs)
 172 {
 173         struct elf_fdpic_params exec_params, interp_params;
 174         struct elf_phdr *phdr;
 175         unsigned long stack_size, entryaddr;
 176 #ifndef CONFIG_MMU
 177         unsigned long fullsize;
 178 #endif
 179 #ifdef ELF_FDPIC_PLAT_INIT
 180         unsigned long dynaddr;
 181 #endif
 182         struct file *interpreter = NULL; /* to shut gcc up */
 183         char *interpreter_name = NULL;
 184         int executable_stack;
 185         int retval, i;
 186
 187         memset(&exec_params, 0, sizeof(exec_params));
 188         memset(&interp_params, 0, sizeof(interp_params));
 189
 190         exec_params.hdr = *(struct elfhdr *) bprm->buf;
 191         exec_params.flags = ELF_FDPIC_FLAG_PRESENT | ELF_FDPIC_FLAG_EXECUTABLE;
 192
 193         /* check that this is a binary we know how to deal with */
 194         retval = -ENOEXEC;
 195         if (!is_elf_fdpic(&exec_params.hdr, bprm->file))
 196                 goto error;
 197
 198         /* read the program header table */
 199         retval = elf_fdpic_fetch_phdrs(&exec_params, bprm->file);
 200         if (retval < 0)
 201                 goto error;
 202
 203         /* scan for a program header that specifies an interpreter */
 204         phdr = exec_params.phdrs;
 205
 206         for (i = 0; i < exec_params.hdr.e_phnum; i++, phdr++) {
 207                 switch (phdr->p_type) {
 208                 case PT_INTERP:
 209                         retval = -ENOMEM;
 210                         if (phdr->p_filesz > PATH_MAX)
 211                                 goto error;
 212                         retval = -ENOENT;
 213                         if (phdr->p_filesz < 2)
 214                                 goto error;
 215
 216                         /* read the name of the interpreter into memory */
 217                         interpreter_name = kmalloc(phdr->p_filesz, GFP_KERNEL);
 218                         if (!interpreter_name)
 219                                 goto error;
 220
 221                         retval = kernel_read(bprm->file,
 222                                              phdr->p_offset,
 223                                              interpreter_name,
 224                                              phdr->p_filesz);
 225                         if (retval < 0)
 226                                 goto error;
 227
 228                         retval = -ENOENT;
 229                         if (interpreter_name[phdr->p_filesz - 1] != '\0')
 230                                 goto error;
 231
 232                         kdebug("Using ELF interpreter %s", interpreter_name);
 233
 234                         /* replace the program with the interpreter */
 235                         interpreter = open_exec(interpreter_name);
 236                         retval = PTR_ERR(interpreter);
 237                         if (IS_ERR(interpreter)) {
 238                                 interpreter = NULL;
 239                                 goto error;
 240                         }
 241
 242                         retval = kernel_read(interpreter, 0, bprm->buf,
 243                                              BINPRM_BUF_SIZE);
 244                         if (retval < 0)
 245                                 goto error;
 246
 247                         interp_params.hdr = *((struct elfhdr *) bprm->buf);
 248                         break;
 249
 250                 case PT_LOAD:
 251 #ifdef CONFIG_MMU
 252                         if (exec_params.load_addr == 0)
 253                                 exec_params.load_addr = phdr->p_vaddr;
 254 #endif
 255                         break;
 256                 }
 257
 258         }
 259
 260         if (elf_check_const_displacement(&exec_params.hdr))
 261                 exec_params.flags |= ELF_FDPIC_FLAG_CONSTDISP;
 262
 263         /* perform insanity checks on the interpreter */
 264         if (interpreter_name) {
 265                 retval = -ELIBBAD;
 266                 if (!is_elf_fdpic(&interp_params.hdr, interpreter))
 267                         goto error;
 268
 269                 interp_params.flags = ELF_FDPIC_FLAG_PRESENT;
 270
 271                 /* read the interpreter's program header table */
 272                 retval = elf_fdpic_fetch_phdrs(&interp_params, interpreter);
 273                 if (retval < 0)
 274                         goto error;
 275         }
 276
 277         stack_size = exec_params.stack_size;
 278         if (stack_size < interp_params.stack_size)
 279                 stack_size = interp_params.stack_size;
 280
 281         if (exec_params.flags & ELF_FDPIC_FLAG_EXEC_STACK)
 282                 executable_stack = EXSTACK_ENABLE_X;
 283         else if (exec_params.flags & ELF_FDPIC_FLAG_NOEXEC_STACK)
 284                 executable_stack = EXSTACK_DISABLE_X;
 285         else if (interp_params.flags & ELF_FDPIC_FLAG_EXEC_STACK)
 286                 executable_stack = EXSTACK_ENABLE_X;
 287         else if (interp_params.flags & ELF_FDPIC_FLAG_NOEXEC_STACK)
 288                 executable_stack = EXSTACK_DISABLE_X;
 289         else
 290                 executable_stack = EXSTACK_DEFAULT;
 291
 292         retval = -ENOEXEC;
 293         if (stack_size == 0)
 294                 goto error;
 295
 296         if (elf_check_const_displacement(&interp_params.hdr))
 297                 interp_params.flags |= ELF_FDPIC_FLAG_CONSTDISP;
 298
 299         /* flush all traces of the currently running executable */
 300         retval = flush_old_exec(bprm);
 301         if (retval)
 302                 goto error;
 303
 304         /* there's now no turning back... the old userspace image is dead,
 305          * defunct, deceased, etc. after this point we have to exit via
 306          * error_kill */
 307         set_personality(PER_LINUX_FDPIC);
 308         set_binfmt(&elf_fdpic_format);
 309
 310         current->mm->start_code = 0;
 311         current->mm->end_code = 0;
 312         current->mm->start_stack = 0;
 313         current->mm->start_data = 0;
 314         current->mm->end_data = 0;
 315         current->mm->context.exec_fdpic_loadmap = 0;
 316         current->mm->context.interp_fdpic_loadmap = 0;
 317
 318         current->flags &= ~PF_FORKNOEXEC;
 319
 320 #ifdef CONFIG_MMU
 321         elf_fdpic_arch_lay_out_mm(&exec_params,
 322                                   &interp_params,
 323                                   &current->mm->start_stack,
 324                                   &current->mm->start_brk);
 325
 326         retval = setup_arg_pages(bprm, current->mm->start_stack,
 327                                  executable_stack);
 328         if (retval < 0) {
 329                 send_sig(SIGKILL, current, 0);
 330                 goto error_kill;
 331         }
 332 #endif
 333
 334         /* load the executable and interpreter into memory */
 335         retval = elf_fdpic_map_file(&exec_params, bprm->file, current->mm,
 336                                     "executable");
 337         if (retval < 0)
 338                 goto error_kill;
 339
 340         if (interpreter_name) {
 341                 retval = elf_fdpic_map_file(&interp_params, interpreter,
 342                                             current->mm, "interpreter");
 343                 if (retval < 0) {
 344                         printk(KERN_ERR "Unable to load interpreter\n");
 345                         goto error_kill;
 346                 }
 347
 348                 allow_write_access(interpreter);
 349                 fput(interpreter);
 350                 interpreter = NULL;
 351         }
 352
 353 #ifdef CONFIG_MMU
 354         if (!current->mm->start_brk)
 355                 current->mm->start_brk = current->mm->end_data;
 356
 357         current->mm->brk = current->mm->start_brk =
 358                 PAGE_ALIGN(current->mm->start_brk);
 359
 360 #else
 361         /* create a stack and brk area big enough for everyone
 362          * - the brk heap starts at the bottom and works up
 363          * - the stack starts at the top and works down
 364          */
 365         stack_size = (stack_size + PAGE_SIZE - 1) & PAGE_MASK;
 366         if (stack_size < PAGE_SIZE * 2)
 367                 stack_size = PAGE_SIZE * 2;
 368
 369         down_write(&current->mm->mmap_sem);
 370         current->mm->start_brk = do_mmap(NULL, 0, stack_size,
 371                                          PROT_READ | PROT_WRITE | PROT_EXEC,
 372                                          MAP_PRIVATE | MAP_ANON | MAP_GROWSDOWN,
 373                                          0);
 374
 375         if (IS_ERR_VALUE(current->mm->start_brk)) {
 376                 up_write(&current->mm->mmap_sem);
 377                 retval = current->mm->start_brk;
 378                 current->mm->start_brk = 0;
 379                 goto error_kill;
 380         }
 381
 382         /* expand the stack mapping to use up the entire allocation granule */
 383         fullsize = ksize((char *) current->mm->start_brk);
 384         if (!IS_ERR_VALUE(do_mremap(current->mm->start_brk, stack_size,
 385                                     fullsize, 0, 0)))
 386                 stack_size = fullsize;
 387         up_write(&current->mm->mmap_sem);
 388
 389         current->mm->brk = current->mm->start_brk;
 390         current->mm->context.end_brk = current->mm->start_brk;
 391         current->mm->context.end_brk +=
 392                 (stack_size > PAGE_SIZE) ? (stack_size - PAGE_SIZE) : 0;
 393         current->mm->start_stack = current->mm->start_brk + stack_size;
 394 #endif
 395
 396         compute_creds(bprm);
 397         current->flags &= ~PF_FORKNOEXEC;
 398         if (create_elf_fdpic_tables(bprm, current->mm,
 399                                     &exec_params, &interp_params) < 0)
 400                 goto error_kill;
 401
 402         kdebug("- start_code  %lx", current->mm->start_code);
 403         kdebug("- end_code    %lx", current->mm->end_code);
 404         kdebug("- start_data  %lx", current->mm->start_data);
 405         kdebug("- end_data    %lx", current->mm->end_data);
 406         kdebug("- start_brk   %lx", current->mm->start_brk);
 407         kdebug("- brk         %lx", current->mm->brk);
 408         kdebug("- start_stack %lx", current->mm->start_stack);
 409
 410 #ifdef ELF_FDPIC_PLAT_INIT
 411         /*
 412          * The ABI may specify that certain registers be set up in special
 413          * ways (on i386 %edx is the address of a DT_FINI function, for
 414          * example.  This macro performs whatever initialization to
 415          * the regs structure is required.
 416          */
 417         dynaddr = interp_params.dynamic_addr ?: exec_params.dynamic_addr;
 418         ELF_FDPIC_PLAT_INIT(regs, exec_params.map_addr, interp_params.map_addr,
 419                             dynaddr);
 420 #endif
 421
 422         /* everything is now ready... get the userspace context ready to roll */
 423         entryaddr = interp_params.entry_addr ?: exec_params.entry_addr;
 424         start_thread(regs, entryaddr, current->mm->start_stack);
 425
 426         retval = 0;
 427
 428 error:
 429         if (interpreter) {
 430                 allow_write_access(interpreter);
 431                 fput(interpreter);
 432         }
 433         kfree(interpreter_name);
 434         kfree(exec_params.phdrs);
 435         kfree(exec_params.loadmap);
 436         kfree(interp_params.phdrs);
 437         kfree(interp_params.loadmap);
 438         return retval;
 439
 440         /* unrecoverable error - kill the process */
 441 error_kill:
 442         send_sig(SIGSEGV, current, 0);
 443         goto error;
 444
 445 }
 446
 447 /*****************************************************************************/
 448 /*
 449  * present useful information to the program
 450  */
 451 static int create_elf_fdpic_tables(struct linux_binprm *bprm,
 452                                    struct mm_struct *mm,
 453                                    struct elf_fdpic_params *exec_params,
 454                                    struct elf_fdpic_params *interp_params)
 455 {
 456         unsigned long sp, csp, nitems;
 457         elf_caddr_t __user *argv, *envp;
 458         size_t platform_len = 0, len;
 459         char *k_platform;
 460         char __user *u_platform, *p;
 461         long hwcap;
 462         int loop;
 463
 464         /* we're going to shovel a whole load of stuff onto the stack */
 465 #ifdef CONFIG_MMU
 466         sp = bprm->p;
 467 #else
 468         sp = mm->start_stack;
 469
 470         /* stack the program arguments and environment */
 471         if (elf_fdpic_transfer_args_to_stack(bprm, &sp) < 0)
 472                 return -EFAULT;
 473 #endif
 474
 475         /* get hold of platform and hardware capabilities masks for the machine
 476          * we are running on.  In some cases (Sparc), this info is impossible
 477          * to get, in others (i386) it is merely difficult.
 478          */
 479         hwcap = ELF_HWCAP;
 480         k_platform = ELF_PLATFORM;
 481         u_platform = NULL;
 482
 483         if (k_platform) {
 484                 platform_len = strlen(k_platform) + 1;
 485                 sp -= platform_len;
 486                 u_platform = (char __user *) sp;
 487                 if (__copy_to_user(u_platform, k_platform, platform_len) != 0)
 488                         return -EFAULT;
 489         }
 490
 491 #if defined(__i386__) && defined(CONFIG_SMP)
 492         /* in some cases (e.g. Hyper-Threading), we want to avoid L1 evictions
 493          * by the processes running on the same package. One thing we can do is
 494          * to shuffle the initial stack for them.
 495          *
 496          * the conditionals here are unneeded, but kept in to make the code
 497          * behaviour the same as pre change unless we have hyperthreaded
 498          * processors. This keeps Mr Marcelo Person happier but should be
 499          * removed for 2.5
 500          */
 501         if (smp_num_siblings > 1)
 502                 sp = sp - ((current->pid % 64) << 7);
 503 #endif
 504
 505         sp &= ~7UL;
 506
 507         /* stack the load map(s) */
 508         len = sizeof(struct elf32_fdpic_loadmap);
 509         len += sizeof(struct elf32_fdpic_loadseg) * exec_params->loadmap->nsegs;
 510         sp = (sp - len) & ~7UL;
 511         exec_params->map_addr = sp;
 512
 513         if (copy_to_user((void __user *) sp, exec_params->loadmap, len) != 0)
 514                 return -EFAULT;
 515
 516         current->mm->context.exec_fdpic_loadmap = (unsigned long) sp;
 517
 518         if (interp_params->loadmap) {
 519                 len = sizeof(struct elf32_fdpic_loadmap);
 520                 len += sizeof(struct elf32_fdpic_loadseg) *
 521                         interp_params->loadmap->nsegs;
 522                 sp = (sp - len) & ~7UL;
 523                 interp_params->map_addr = sp;
 524
 525                 if (copy_to_user((void __user *) sp, interp_params->loadmap,
 526                                  len) != 0)
 527                         return -EFAULT;
 528
 529                 current->mm->context.interp_fdpic_loadmap = (unsigned long) sp;
 530         }
 531
 532         /* force 16 byte _final_ alignment here for generality */
 533 #define DLINFO_ITEMS 13
 534
 535         nitems = 1 + DLINFO_ITEMS + (k_platform ? 1 : 0);
 536 #ifdef DLINFO_ARCH_ITEMS
 537         nitems += DLINFO_ARCH_ITEMS;
 538 #endif
 539
 540         csp = sp;
 541         sp -= nitems * 2 * sizeof(unsigned long);
 542         sp -= (bprm->envc + 1) * sizeof(char *);        /* envv[] */
 543         sp -= (bprm->argc + 1) * sizeof(char *);        /* argv[] */
 544         sp -= 1 * sizeof(unsigned long);                /* argc */
 545
 546         csp -= sp & 15UL;
 547         sp -= sp & 15UL;
 548
 549         /* put the ELF interpreter info on the stack */
 550 #define NEW_AUX_ENT(nr, id, val)                                        \
 551         do {                                                            \
 552                 struct { unsigned long _id, _val; } __user *ent;        \
 553                                                                         \
 554                 ent = (void __user *) csp;                              \
 555                 __put_user((id), &ent[nr]._id);                         \
 556                 __put_user((val), &ent[nr]._val);                       \
 557         } while (0)
 558
 559         csp -= 2 * sizeof(unsigned long);
 560         NEW_AUX_ENT(0, AT_NULL, 0);
 561         if (k_platform) {
 562                 csp -= 2 * sizeof(unsigned long);
 563                 NEW_AUX_ENT(0, AT_PLATFORM,
 564                             (elf_addr_t) (unsigned long) u_platform);
 565         }
 566
 567         csp -= DLINFO_ITEMS * 2 * sizeof(unsigned long);
 568         NEW_AUX_ENT( 0, AT_HWCAP,       hwcap);
 569         NEW_AUX_ENT( 1, AT_PAGESZ,      PAGE_SIZE);
 570         NEW_AUX_ENT( 2, AT_CLKTCK,      CLOCKS_PER_SEC);
 571         NEW_AUX_ENT( 3, AT_PHDR,        exec_params->ph_addr);
 572         NEW_AUX_ENT( 4, AT_PHENT,       sizeof(struct elf_phdr));
 573         NEW_AUX_ENT( 5, AT_PHNUM,       exec_params->hdr.e_phnum);
 574         NEW_AUX_ENT( 6, AT_BASE,        interp_params->elfhdr_addr);
 575         NEW_AUX_ENT( 7, AT_FLAGS,       0);
 576         NEW_AUX_ENT( 8, AT_ENTRY,       exec_params->entry_addr);
 577         NEW_AUX_ENT( 9, AT_UID,         (elf_addr_t) current->uid);
 578         NEW_AUX_ENT(10, AT_EUID,        (elf_addr_t) current->euid);
 579         NEW_AUX_ENT(11, AT_GID,         (elf_addr_t) current->gid);
 580         NEW_AUX_ENT(12, AT_EGID,        (elf_addr_t) current->egid);
 581
 582 #ifdef ARCH_DLINFO
 583         /* ARCH_DLINFO must come last so platform specific code can enforce
 584          * special alignment requirements on the AUXV if necessary (eg. PPC).
 585          */
 586         ARCH_DLINFO;
 587 #endif
 588 #undef NEW_AUX_ENT
 589
 590         /* allocate room for argv[] and envv[] */
 591         csp -= (bprm->envc + 1) * sizeof(elf_caddr_t);
 592         envp = (elf_caddr_t __user *) csp;
 593         csp -= (bprm->argc + 1) * sizeof(elf_caddr_t);
 594         argv = (elf_caddr_t __user *) csp;
 595
 596         /* stack argc */
 597         csp -= sizeof(unsigned long);
 598         __put_user(bprm->argc, (unsigned long __user *) csp);
 599
 600         BUG_ON(csp != sp);
 601
 602         /* fill in the argv[] array */
 603 #ifdef CONFIG_MMU
 604         current->mm->arg_start = bprm->p;
 605 #else
 606         current->mm->arg_start = current->mm->start_stack -
 607                 (MAX_ARG_PAGES * PAGE_SIZE - bprm->p);
 608 #endif
 609
 610         p = (char __user *) current->mm->arg_start;
 611         for (loop = bprm->argc; loop > 0; loop--) {
 612                 __put_user((elf_caddr_t) p, argv++);
 613                 len = strnlen_user(p, PAGE_SIZE * MAX_ARG_PAGES);
 614                 if (!len || len > PAGE_SIZE * MAX_ARG_PAGES)
 615                         return -EINVAL;
 616                 p += len;
 617         }
 618         __put_user(NULL, argv);
 619         current->mm->arg_end = (unsigned long) p;
 620
 621         /* fill in the envv[] array */
 622         current->mm->env_start = (unsigned long) p;
 623         for (loop = bprm->envc; loop > 0; loop--) {
 624                 __put_user((elf_caddr_t)(unsigned long) p, envp++);
 625                 len = strnlen_user(p, PAGE_SIZE * MAX_ARG_PAGES);
 626                 if (!len || len > PAGE_SIZE * MAX_ARG_PAGES)
 627                         return -EINVAL;
 628                 p += len;
 629         }
 630         __put_user(NULL, envp);
 631         current->mm->env_end = (unsigned long) p;
 632
 633         mm->start_stack = (unsigned long) sp;
 634         return 0;
 635 }
 636
 637 /*****************************************************************************/
 638 /*
 639  * transfer the program arguments and environment from the holding pages onto
 640  * the stack
 641  */
 642 #ifndef CONFIG_MMU
 643 static int elf_fdpic_transfer_args_to_stack(struct linux_binprm *bprm,
 644                                             unsigned long *_sp)
 645 {
 646         unsigned long index, stop, sp;
 647         char *src;
 648         int ret = 0;
 649
 650         stop = bprm->p >> PAGE_SHIFT;
 651         sp = *_sp;
 652
 653         for (index = MAX_ARG_PAGES - 1; index >= stop; index--) {
 654                 src = kmap(bprm->page[index]);
 655                 sp -= PAGE_SIZE;
 656                 if (copy_to_user((void *) sp, src, PAGE_SIZE) != 0)
 657                         ret = -EFAULT;
 658                 kunmap(bprm->page[index]);
 659                 if (ret < 0)
 660                         goto out;
 661         }
 662
 663         *_sp = (*_sp - (MAX_ARG_PAGES * PAGE_SIZE - bprm->p)) & ~15;
 664
 665 out:
 666         return ret;
 667 }
 668 #endif
 669
 670 /*****************************************************************************/
 671 /*
 672  * load the appropriate binary image (executable or interpreter) into memory
 673  * - we assume no MMU is available
 674  * - if no other PIC bits are set in params->hdr->e_flags
 675  *   - we assume that the LOADable segments in the binary are independently relocatable
 676  *   - we assume R/O executable segments are shareable
 677  * - else
 678  *   - we assume the loadable parts of the image to require fixed displacement
 679  *   - the image is not shareable
 680  */
 681 static int elf_fdpic_map_file(struct elf_fdpic_params *params,
 682                               struct file *file,
 683                               struct mm_struct *mm,
 684                               const char *what)
 685 {
 686         struct elf32_fdpic_loadmap *loadmap;
 687 #ifdef CONFIG_MMU
 688         struct elf32_fdpic_loadseg *mseg;
 689 #endif
 690         struct elf32_fdpic_loadseg *seg;
 691         struct elf32_phdr *phdr;
 692         unsigned long load_addr, stop;
 693         unsigned nloads, tmp;
 694         size_t size;
 695         int loop, ret;
 696
 697         /* allocate a load map table */
 698         nloads = 0;
 699         for (loop = 0; loop < params->hdr.e_phnum; loop++)
 700                 if (params->phdrs[loop].p_type == PT_LOAD)
 701                         nloads++;
 702
 703         if (nloads == 0)
 704                 return -ELIBBAD;
 705
 706         size = sizeof(*loadmap) + nloads * sizeof(*seg);
 707         loadmap = kmalloc(size, GFP_KERNEL);
 708         if (!loadmap)
 709                 return -ENOMEM;
 710
 711         params->loadmap = loadmap;
 712         memset(loadmap, 0, size);
 713
 714         loadmap->version = ELF32_FDPIC_LOADMAP_VERSION;
 715         loadmap->nsegs = nloads;
 716
 717         load_addr = params->load_addr;
 718         seg = loadmap->segs;
 719
 720         /* map the requested LOADs into the memory space */
 721         switch (params->flags & ELF_FDPIC_FLAG_ARRANGEMENT) {
 722         case ELF_FDPIC_FLAG_CONSTDISP:
 723         case ELF_FDPIC_FLAG_CONTIGUOUS:
 724 #ifndef CONFIG_MMU
 725                 ret = elf_fdpic_map_file_constdisp_on_uclinux(params, file, mm);
 726                 if (ret < 0)
 727                         return ret;
 728                 break;
 729 #endif
 730         default:
 731                 ret = elf_fdpic_map_file_by_direct_mmap(params, file, mm);
 732                 if (ret < 0)
 733                         return ret;
 734                 break;
 735         }
 736
 737         /* map the entry point */
 738         if (params->hdr.e_entry) {
 739                 seg = loadmap->segs;
 740                 for (loop = loadmap->nsegs; loop > 0; loop--, seg++) {
 741                         if (params->hdr.e_entry >= seg->p_vaddr &&
 742                             params->hdr.e_entry < seg->p_vaddr + seg->p_memsz) {
 743                                 params->entry_addr =
 744                                         (params->hdr.e_entry - seg->p_vaddr) +
 745                                         seg->addr;
 746                                 break;
 747                         }
 748                 }
 749         }
 750
 751         /* determine where the program header table has wound up if mapped */
 752         stop = params->hdr.e_phoff;
 753         stop += params->hdr.e_phnum * sizeof (struct elf_phdr);
 754         phdr = params->phdrs;
 755
 756         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 757                 if (phdr->p_type != PT_LOAD)
 758                         continue;
 759
 760                 if (phdr->p_offset > params->hdr.e_phoff ||
 761                     phdr->p_offset + phdr->p_filesz < stop)
 762                         continue;
 763
 764                 seg = loadmap->segs;
 765                 for (loop = loadmap->nsegs; loop > 0; loop--, seg++) {
 766                         if (phdr->p_vaddr >= seg->p_vaddr &&
 767                             phdr->p_vaddr + phdr->p_filesz <=
 768                             seg->p_vaddr + seg->p_memsz) {
 769                                 params->ph_addr =
 770                                         (phdr->p_vaddr - seg->p_vaddr) +
 771                                         seg->addr +
 772                                         params->hdr.e_phoff - phdr->p_offset;
 773                                 break;
 774                         }
 775                 }
 776                 break;
 777         }
 778
 779         /* determine where the dynamic section has wound up if there is one */
 780         phdr = params->phdrs;
 781         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 782                 if (phdr->p_type != PT_DYNAMIC)
 783                         continue;
 784
 785                 seg = loadmap->segs;
 786                 for (loop = loadmap->nsegs; loop > 0; loop--, seg++) {
 787                         if (phdr->p_vaddr >= seg->p_vaddr &&
 788                             phdr->p_vaddr + phdr->p_memsz <=
 789                             seg->p_vaddr + seg->p_memsz) {
 790                                 params->dynamic_addr =
 791                                         (phdr->p_vaddr - seg->p_vaddr) +
 792                                         seg->addr;
 793
 794                                 /* check the dynamic section contains at least
 795                                  * one item, and that the last item is a NULL
 796                                  * entry */
 797                                 if (phdr->p_memsz == 0 ||
 798                                     phdr->p_memsz % sizeof(Elf32_Dyn) != 0)
 799                                         goto dynamic_error;
 800
 801                                 tmp = phdr->p_memsz / sizeof(Elf32_Dyn);
 802                                 if (((Elf32_Dyn *)
 803                                      params->dynamic_addr)[tmp - 1].d_tag != 0)
 804                                         goto dynamic_error;
 805                                 break;
 806                         }
 807                 }
 808                 break;
 809         }
 810
 811         /* now elide adjacent segments in the load map on MMU linux
 812          * - on uClinux the holes between may actually be filled with system
 813          *   stuff or stuff from other processes
 814          */
 815 #ifdef CONFIG_MMU
 816         nloads = loadmap->nsegs;
 817         mseg = loadmap->segs;
 818         seg = mseg + 1;
 819         for (loop = 1; loop < nloads; loop++) {
 820                 /* see if we have a candidate for merging */
 821                 if (seg->p_vaddr - mseg->p_vaddr == seg->addr - mseg->addr) {
 822                         load_addr = PAGE_ALIGN(mseg->addr + mseg->p_memsz);
 823                         if (load_addr == (seg->addr & PAGE_MASK)) {
 824                                 mseg->p_memsz +=
 825                                         load_addr -
 826                                         (mseg->addr + mseg->p_memsz);
 827                                 mseg->p_memsz += seg->addr & ~PAGE_MASK;
 828                                 mseg->p_memsz += seg->p_memsz;
 829                                 loadmap->nsegs--;
 830                                 continue;
 831                         }
 832                 }
 833
 834                 mseg++;
 835                 if (mseg != seg)
 836                         *mseg = *seg;
 837         }
 838 #endif
 839
 840         kdebug("Mapped Object [%s]:", what);
 841         kdebug("- elfhdr   : %lx", params->elfhdr_addr);
 842         kdebug("- entry    : %lx", params->entry_addr);
 843         kdebug("- PHDR[]   : %lx", params->ph_addr);
 844         kdebug("- DYNAMIC[]: %lx", params->dynamic_addr);
 845         seg = loadmap->segs;
 846         for (loop = 0; loop < loadmap->nsegs; loop++, seg++)
 847                 kdebug("- LOAD[%d] : %08x-%08x [va=%x ms=%x]",
 848                        loop,
 849                        seg->addr, seg->addr + seg->p_memsz - 1,
 850                        seg->p_vaddr, seg->p_memsz);
 851
 852         return 0;
 853
 854 dynamic_error:
 855         printk("ELF FDPIC %s with invalid DYNAMIC section (inode=%lu)\n",
 856                what, file->f_dentry->d_inode->i_ino);
 857         return -ELIBBAD;
 858 }
 859
 860 /*****************************************************************************/
 861 /*
 862  * map a file with constant displacement under uClinux
 863  */
 864 #ifndef CONFIG_MMU
 865 static int elf_fdpic_map_file_constdisp_on_uclinux(
 866         struct elf_fdpic_params *params,
 867         struct file *file,
 868         struct mm_struct *mm)
 869 {
 870         struct elf32_fdpic_loadseg *seg;
 871         struct elf32_phdr *phdr;
 872         unsigned long load_addr, base = ULONG_MAX, top = 0, maddr = 0, mflags;
 873         loff_t fpos;
 874         int loop, ret;
 875
 876         load_addr = params->load_addr;
 877         seg = params->loadmap->segs;
 878
 879         /* determine the bounds of the contiguous overall allocation we must
 880          * make */
 881         phdr = params->phdrs;
 882         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 883                 if (params->phdrs[loop].p_type != PT_LOAD)
 884                         continue;
 885
 886                 if (base > phdr->p_vaddr)
 887                         base = phdr->p_vaddr;
 888                 if (top < phdr->p_vaddr + phdr->p_memsz)
 889                         top = phdr->p_vaddr + phdr->p_memsz;
 890         }
 891
 892         /* allocate one big anon block for everything */
 893         mflags = MAP_PRIVATE;
 894         if (params->flags & ELF_FDPIC_FLAG_EXECUTABLE)
 895                 mflags |= MAP_EXECUTABLE;
 896
 897         down_write(&mm->mmap_sem);
 898         maddr = do_mmap(NULL, load_addr, top - base,
 899                         PROT_READ | PROT_WRITE | PROT_EXEC, mflags, 0);
 900         up_write(&mm->mmap_sem);
 901         if (IS_ERR_VALUE(maddr))
 902                 return (int) maddr;
 903
 904         if (load_addr != 0)
 905                 load_addr += PAGE_ALIGN(top - base);
 906
 907         /* and then load the file segments into it */
 908         phdr = params->phdrs;
 909         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 910                 if (params->phdrs[loop].p_type != PT_LOAD)
 911                         continue;
 912
 913                 fpos = phdr->p_offset;
 914
 915                 seg->addr = maddr + (phdr->p_vaddr - base);
 916                 seg->p_vaddr = phdr->p_vaddr;
 917                 seg->p_memsz = phdr->p_memsz;
 918
 919                 ret = file->f_op->read(file, (void *) seg->addr,
 920                                        phdr->p_filesz, &fpos);
 921                 if (ret < 0)
 922                         return ret;
 923
 924                 /* map the ELF header address if in this segment */
 925                 if (phdr->p_offset == 0)
 926                         params->elfhdr_addr = seg->addr;
 927
 928                 /* clear any space allocated but not loaded */
 929                 if (phdr->p_filesz < phdr->p_memsz)
 930                         clear_user((void *) (seg->addr + phdr->p_filesz),
 931                                    phdr->p_memsz - phdr->p_filesz);
 932
 933                 if (mm) {
 934                         if (phdr->p_flags & PF_X) {
 935                                 mm->start_code = seg->addr;
 936                                 mm->end_code = seg->addr + phdr->p_memsz;
 937                         } else if (!mm->start_data) {
 938                                 mm->start_data = seg->addr;
 939 #ifndef CONFIG_MMU
 940                                 mm->end_data = seg->addr + phdr->p_memsz;
 941 #endif
 942                         }
 943
 944 #ifdef CONFIG_MMU
 945                         if (seg->addr + phdr->p_memsz > mm->end_data)
 946                                 mm->end_data = seg->addr + phdr->p_memsz;
 947 #endif
 948                 }
 949
 950                 seg++;
 951         }
 952
 953         return 0;
 954 }
 955 #endif
 956
 957 /*****************************************************************************/
 958 /*
 959  * map a binary by direct mmap() of the individual PT_LOAD segments
 960  */
 961 static int elf_fdpic_map_file_by_direct_mmap(struct elf_fdpic_params *params,
 962                                              struct file *file,
 963                                              struct mm_struct *mm)
 964 {
 965         struct elf32_fdpic_loadseg *seg;
 966         struct elf32_phdr *phdr;
 967         unsigned long load_addr, delta_vaddr;
 968         int loop, dvset;
 969
 970         load_addr = params->load_addr;
 971         delta_vaddr = 0;
 972         dvset = 0;
 973
 974         seg = params->loadmap->segs;
 975
 976         /* deal with each load segment separately */
 977         phdr = params->phdrs;
 978         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 979                 unsigned long maddr, disp, excess, excess1;
 980                 int prot = 0, flags;
 981
 982                 if (phdr->p_type != PT_LOAD)
 983                         continue;
 984
 985                 kdebug("[LOAD] va=%lx of=%lx fs=%lx ms=%lx",
 986                        (unsigned long) phdr->p_vaddr,
 987                        (unsigned long) phdr->p_offset,
 988                        (unsigned long) phdr->p_filesz,
 989                        (unsigned long) phdr->p_memsz);
 990
 991                 /* determine the mapping parameters */
 992                 if (phdr->p_flags & PF_R) prot |= PROT_READ;
 993                 if (phdr->p_flags & PF_W) prot |= PROT_WRITE;
 994                 if (phdr->p_flags & PF_X) prot |= PROT_EXEC;
 995
 996                 flags = MAP_PRIVATE | MAP_DENYWRITE;
 997                 if (params->flags & ELF_FDPIC_FLAG_EXECUTABLE)
 998                         flags |= MAP_EXECUTABLE;
 999
1000                 maddr = 0;
1001
1002                 switch (params->flags & ELF_FDPIC_FLAG_ARRANGEMENT) {
1003                 case ELF_FDPIC_FLAG_INDEPENDENT:
1004                         /* PT_LOADs are independently locatable */
1005                         break;
1006
1007                 case ELF_FDPIC_FLAG_HONOURVADDR:
1008                         /* the specified virtual address must be honoured */
1009                         maddr = phdr->p_vaddr;
1010                         flags |= MAP_FIXED;
1011                         break;
1012
1013                 case ELF_FDPIC_FLAG_CONSTDISP:
1014                         /* constant displacement
1015                          * - can be mapped anywhere, but must be mapped as a
1016                          *   unit
1017                          */
1018                         if (!dvset) {
1019                                 maddr = load_addr;
1020                                 delta_vaddr = phdr->p_vaddr;
1021                                 dvset = 1;
1022                         } else {
1023                                 maddr = load_addr + phdr->p_vaddr - delta_vaddr;
1024                                 flags |= MAP_FIXED;
1025                         }
1026                         break;
1027
1028                 case ELF_FDPIC_FLAG_CONTIGUOUS:
1029                         /* contiguity handled later */
1030                         break;
1031
1032                 default:
1033                         BUG();
1034                 }
1035
1036                 maddr &= PAGE_MASK;
1037
1038                 /* create the mapping */
1039                 disp = phdr->p_vaddr & ~PAGE_MASK;
1040                 down_write(&mm->mmap_sem);
1041                 maddr = do_mmap(file, maddr, phdr->p_memsz + disp, prot, flags,
1042                                 phdr->p_offset - disp);
1043                 up_write(&mm->mmap_sem);
1044
1045                 kdebug("mmap[%d] <file> sz=%lx pr=%x fl=%x of=%lx --> %08lx",
1046                        loop, phdr->p_memsz + disp, prot, flags,
1047                        phdr->p_offset - disp, maddr);
1048
1049                 if (IS_ERR_VALUE(maddr))
1050                         return (int) maddr;
1051
1052                 if ((params->flags & ELF_FDPIC_FLAG_ARRANGEMENT) ==
1053                     ELF_FDPIC_FLAG_CONTIGUOUS)
1054                         load_addr += PAGE_ALIGN(phdr->p_memsz + disp);
1055
1056                 seg->addr = maddr + disp;
1057                 seg->p_vaddr = phdr->p_vaddr;
1058                 seg->p_memsz = phdr->p_memsz;
1059
1060                 /* map the ELF header address if in this segment */
1061                 if (phdr->p_offset == 0)
1062                         params->elfhdr_addr = seg->addr;
1063
1064                 /* clear the bit between beginning of mapping and beginning of
1065                  * PT_LOAD */
1066                 if (prot & PROT_WRITE && disp > 0) {
1067                         kdebug("clear[%d] ad=%lx sz=%lx", loop, maddr, disp);
1068                         clear_user((void __user *) maddr, disp);
1069                         maddr += disp;
1070                 }
1071
1072                 /* clear any space allocated but not loaded
1073                  * - on uClinux we can just clear the lot
1074                  * - on MMU linux we'll get a SIGBUS beyond the last page
1075                  *   extant in the file
1076                  */
1077                 excess = phdr->p_memsz - phdr->p_filesz;
1078                 excess1 = PAGE_SIZE - ((maddr + phdr->p_filesz) & ~PAGE_MASK);
1079
1080 #ifdef CONFIG_MMU
1081                 if (excess > excess1) {
1082                         unsigned long xaddr = maddr + phdr->p_filesz + excess1;
1083                         unsigned long xmaddr;
1084
1085                         flags |= MAP_FIXED | MAP_ANONYMOUS;
1086                         down_write(&mm->mmap_sem);
1087                         xmaddr = do_mmap(NULL, xaddr, excess - excess1,
1088                                          prot, flags, 0);
1089                         up_write(&mm->mmap_sem);
1090
1091                         kdebug("mmap[%d] <anon>"
1092                                " ad=%lx sz=%lx pr=%x fl=%x of=0 --> %08lx",
1093                                loop, xaddr, excess - excess1, prot, flags,
1094                                xmaddr);
1095
1096                         if (xmaddr != xaddr)
1097                                 return -ENOMEM;
1098                 }
1099
1100                 if (prot & PROT_WRITE && excess1 > 0) {
1101                         kdebug("clear[%d] ad=%lx sz=%lx",
1102                                loop, maddr + phdr->p_filesz, excess1);
1103                         clear_user((void __user *) maddr + phdr->p_filesz,
1104                                    excess1);
1105                 }
1106
1107 #else
1108                 if (excess > 0) {
1109                         kdebug("clear[%d] ad=%lx sz=%lx",
1110                                loop, maddr + phdr->p_filesz, excess);
1111                         clear_user((void *) maddr + phdr->p_filesz, excess);
1112                 }
1113 #endif
1114
1115                 if (mm) {
1116                         if (phdr->p_flags & PF_X) {
1117                                 mm->start_code = maddr;
1118                                 mm->end_code = maddr + phdr->p_memsz;
1119                         } else if (!mm->start_data) {
1120                                 mm->start_data = maddr;
1121                                 mm->end_data = maddr + phdr->p_memsz;
1122                         }
1123                 }
1124
1125                 seg++;
1126         }
1127
1128         return 0;
1129 }
1130
1131 /*****************************************************************************/
1132 /*
1133  * ELF-FDPIC core dumper
1134  *
1135  * Modelled on fs/exec.c:aout_core_dump()
1136  * Jeremy Fitzhardinge <jeremy@sw.oz.au>
1137  *
1138  * Modelled on fs/binfmt_elf.c core dumper
1139  */
1140 #if defined(USE_ELF_CORE_DUMP) && defined(CONFIG_ELF_CORE)
1141
1142 /*
1143  * These are the only things you should do on a core-file: use only these
1144  * functions to write out all the necessary info.
1145  */
1146 static int dump_write(struct file *file, const void *addr, int nr)
1147 {
1148         return file->f_op->write(file, addr, nr, &file->f_pos) == nr;
1149 }
1150
1151 static int dump_seek(struct file *file, loff_t off)
1152 {
1153         if (file->f_op->llseek) {
1154                 if (file->f_op->llseek(file, off, SEEK_SET) != off)
1155                         return 0;
1156         } else {
1157                 file->f_pos = off;
1158         }
1159         return 1;
1160 }
1161
1162 /*
1163  * Decide whether a segment is worth dumping; default is yes to be
1164  * sure (missing info is worse than too much; etc).
1165  * Personally I'd include everything, and use the coredump limit...
1166  *
1167  * I think we should skip something. But I am not sure how. H.J.
1168  */
1169 static int maydump(struct vm_area_struct *vma)
1170 {
1171         /* Do not dump I/O mapped devices or special mappings */
1172         if (vma->vm_flags & (VM_IO | VM_RESERVED)) {
1173                 kdcore("%08lx: %08lx: no (IO)", vma->vm_start, vma->vm_flags);
1174                 return 0;
1175         }
1176
1177         /* If we may not read the contents, don't allow us to dump
1178          * them either. "dump_write()" can't handle it anyway.
1179          */
1180         if (!(vma->vm_flags & VM_READ)) {
1181                 kdcore("%08lx: %08lx: no (!read)", vma->vm_start, vma->vm_flags);
1182                 return 0;
1183         }
1184
1185         /* Dump shared memory only if mapped from an anonymous file. */
1186         if (vma->vm_flags & VM_SHARED) {
1187                 if (vma->vm_file->f_dentry->d_inode->i_nlink == 0) {
1188                         kdcore("%08lx: %08lx: no (share)", vma->vm_start, vma->vm_flags);
1189                         return 1;
1190                 }
1191
1192                 kdcore("%08lx: %08lx: no (share)", vma->vm_start, vma->vm_flags);
1193                 return 0;
1194         }
1195
1196 #ifdef CONFIG_MMU
1197         /* If it hasn't been written to, don't write it out */
1198         if (!vma->anon_vma) {
1199                 kdcore("%08lx: %08lx: no (!anon)", vma->vm_start, vma->vm_flags);
1200                 return 0;
1201         }
1202 #endif
1203
1204         kdcore("%08lx: %08lx: yes", vma->vm_start, vma->vm_flags);
1205         return 1;
1206 }
1207
1208 /* An ELF note in memory */
1209 struct memelfnote
1210 {
1211         const char *name;
1212         int type;
1213         unsigned int datasz;
1214         void *data;
1215 };
1216
1217 static int notesize(struct memelfnote *en)
1218 {
1219         int sz;
1220
1221         sz = sizeof(struct elf_note);
1222         sz += roundup(strlen(en->name) + 1, 4);
1223         sz += roundup(en->datasz, 4);
1224
1225         return sz;
1226 }
1227
1228 /* #define DEBUG */
1229
1230 #define DUMP_WRITE(addr, nr)    \
1231         do { if (!dump_write(file, (addr), (nr))) return 0; } while(0)
1232 #define DUMP_SEEK(off)  \
1233         do { if (!dump_seek(file, (off))) return 0; } while(0)
1234
1235 static int writenote(struct memelfnote *men, struct file *file)
1236 {
1237         struct elf_note en;
1238
1239         en.n_namesz = strlen(men->name) + 1;
1240         en.n_descsz = men->datasz;
1241         en.n_type = men->type;
1242
1243         DUMP_WRITE(&en, sizeof(en));
1244         DUMP_WRITE(men->name, en.n_namesz);
1245         /* XXX - cast from long long to long to avoid need for libgcc.a */
1246         DUMP_SEEK(roundup((unsigned long)file->f_pos, 4));      /* XXX */
1247         DUMP_WRITE(men->data, men->datasz);
1248         DUMP_SEEK(roundup((unsigned long)file->f_pos, 4));      /* XXX */
1249
1250         return 1;
1251 }
1252 #undef DUMP_WRITE
1253 #undef DUMP_SEEK
1254
1255 #define DUMP_WRITE(addr, nr)    \
1256         if ((size += (nr)) > limit || !dump_write(file, (addr), (nr))) \
1257                 goto end_coredump;
1258 #define DUMP_SEEK(off)  \
1259         if (!dump_seek(file, (off))) \
1260                 goto end_coredump;
1261
1262 static inline void fill_elf_fdpic_header(struct elfhdr *elf, int segs)
1263 {
1264         memcpy(elf->e_ident, ELFMAG, SELFMAG);
1265         elf->e_ident[EI_CLASS] = ELF_CLASS;
1266         elf->e_ident[EI_DATA] = ELF_DATA;
1267         elf->e_ident[EI_VERSION] = EV_CURRENT;
1268         elf->e_ident[EI_OSABI] = ELF_OSABI;
1269         memset(elf->e_ident+EI_PAD, 0, EI_NIDENT-EI_PAD);
1270
1271         elf->e_type = ET_CORE;
1272         elf->e_machine = ELF_ARCH;
1273         elf->e_version = EV_CURRENT;
1274         elf->e_entry = 0;
1275         elf->e_phoff = sizeof(struct elfhdr);
1276         elf->e_shoff = 0;
1277         elf->e_flags = ELF_FDPIC_CORE_EFLAGS;
1278         elf->e_ehsize = sizeof(struct elfhdr);
1279         elf->e_phentsize = sizeof(struct elf_phdr);
1280         elf->e_phnum = segs;
1281         elf->e_shentsize = 0;
1282         elf->e_shnum = 0;
1283         elf->e_shstrndx = 0;
1284         return;
1285 }
1286
1287 static inline void fill_elf_note_phdr(struct elf_phdr *phdr, int sz, loff_t offset)
1288 {
1289         phdr->p_type = PT_NOTE;
1290         phdr->p_offset = offset;
1291         phdr->p_vaddr = 0;
1292         phdr->p_paddr = 0;
1293         phdr->p_filesz = sz;
1294         phdr->p_memsz = 0;
1295         phdr->p_flags = 0;
1296         phdr->p_align = 0;
1297         return;
1298 }
1299
1300 static inline void fill_note(struct memelfnote *note, const char *name, int type,
1301                 unsigned int sz, void *data)
1302 {
1303         note->name = name;
1304         note->type = type;
1305         note->datasz = sz;
1306         note->data = data;
1307         return;
1308 }
1309
1310 /*
1311  * fill up all the fields in prstatus from the given task struct, except
1312  * registers which need to be filled up seperately.
1313  */
1314 static void fill_prstatus(struct elf_prstatus *prstatus,
1315                           struct task_struct *p, long signr)
1316 {
1317         prstatus->pr_info.si_signo = prstatus->pr_cursig = signr;
1318         prstatus->pr_sigpend = p->pending.signal.sig[0];
1319         prstatus->pr_sighold = p->blocked.sig[0];
1320         prstatus->pr_pid = p->pid;
1321         prstatus->pr_ppid = p->parent->pid;
1322         prstatus->pr_pgrp = process_group(p);
1323         prstatus->pr_sid = p->signal->session;
1324         if (thread_group_leader(p)) {
1325                 /*
1326                  * This is the record for the group leader.  Add in the
1327                  * cumulative times of previous dead threads.  This total
1328                  * won't include the time of each live thread whose state
1329                  * is included in the core dump.  The final total reported
1330                  * to our parent process when it calls wait4 will include
1331                  * those sums as well as the little bit more time it takes
1332                  * this and each other thread to finish dying after the
1333                  * core dump synchronization phase.
1334                  */
1335                 cputime_to_timeval(cputime_add(p->utime, p->signal->utime),
1336                                    &prstatus->pr_utime);
1337                 cputime_to_timeval(cputime_add(p->stime, p->signal->stime),
1338                                    &prstatus->pr_stime);
1339         } else {
1340                 cputime_to_timeval(p->utime, &prstatus->pr_utime);
1341                 cputime_to_timeval(p->stime, &prstatus->pr_stime);
1342         }
1343         cputime_to_timeval(p->signal->cutime, &prstatus->pr_cutime);
1344         cputime_to_timeval(p->signal->cstime, &prstatus->pr_cstime);
1345
1346         prstatus->pr_exec_fdpic_loadmap = p->mm->context.exec_fdpic_loadmap;
1347         prstatus->pr_interp_fdpic_loadmap = p->mm->context.interp_fdpic_loadmap;
1348 }
1349
1350 static int fill_psinfo(struct elf_prpsinfo *psinfo, struct task_struct *p,
1351                        struct mm_struct *mm)
1352 {
1353         unsigned int i, len;
1354
1355         /* first copy the parameters from user space */
1356         memset(psinfo, 0, sizeof(struct elf_prpsinfo));
1357
1358         len = mm->arg_end - mm->arg_start;
1359         if (len >= ELF_PRARGSZ)
1360                 len = ELF_PRARGSZ - 1;
1361         if (copy_from_user(&psinfo->pr_psargs,
1362                            (const char __user *) mm->arg_start, len))
1363                 return -EFAULT;
1364         for (i = 0; i < len; i++)
1365                 if (psinfo->pr_psargs[i] == 0)
1366                         psinfo->pr_psargs[i] = ' ';
1367         psinfo->pr_psargs[len] = 0;
1368
1369         psinfo->pr_pid = p->pid;
1370         psinfo->pr_ppid = p->parent->pid;
1371         psinfo->pr_pgrp = process_group(p);
1372         psinfo->pr_sid = p->signal->session;
1373
1374         i = p->state ? ffz(~p->state) + 1 : 0;
1375         psinfo->pr_state = i;
1376         psinfo->pr_sname = (i > 5) ? '.' : "RSDTZW"[i];
1377         psinfo->pr_zomb = psinfo->pr_sname == 'Z';
1378         psinfo->pr_nice = task_nice(p);
1379         psinfo->pr_flag = p->flags;
1380         SET_UID(psinfo->pr_uid, p->uid);
1381         SET_GID(psinfo->pr_gid, p->gid);
1382         strncpy(psinfo->pr_fname, p->comm, sizeof(psinfo->pr_fname));
1383
1384         return 0;
1385 }
1386
1387 /* Here is the structure in which status of each thread is captured. */
1388 struct elf_thread_status
1389 {
1390         struct list_head list;
1391         struct elf_prstatus prstatus;   /* NT_PRSTATUS */
1392         elf_fpregset_t fpu;             /* NT_PRFPREG */
1393         struct task_struct *thread;
1394 #ifdef ELF_CORE_COPY_XFPREGS
1395         elf_fpxregset_t xfpu;           /* NT_PRXFPREG */
1396 #endif
1397         struct memelfnote notes[3];
1398         int num_notes;
1399 };
1400
1401 /*
1402  * In order to add the specific thread information for the elf file format,
1403  * we need to keep a linked list of every thread's pr_status and then create
1404  * a single section for them in the final core file.
1405  */
1406 static int elf_dump_thread_status(long signr, struct elf_thread_status *t)
1407 {
1408         struct task_struct *p = t->thread;
1409         int sz = 0;
1410
1411         t->num_notes = 0;
1412
1413         fill_prstatus(&t->prstatus, p, signr);
1414         elf_core_copy_task_regs(p, &t->prstatus.pr_reg);
1415
1416         fill_note(&t->notes[0], "CORE", NT_PRSTATUS, sizeof(t->prstatus),
1417                   &t->prstatus);
1418         t->num_notes++;
1419         sz += notesize(&t->notes[0]);
1420
1421         t->prstatus.pr_fpvalid = elf_core_copy_task_fpregs(p, NULL, &t->fpu);
1422         if (t->prstatus.pr_fpvalid) {
1423                 fill_note(&t->notes[1], "CORE", NT_PRFPREG, sizeof(t->fpu),
1424                           &t->fpu);
1425                 t->num_notes++;
1426                 sz += notesize(&t->notes[1]);
1427         }
1428
1429 #ifdef ELF_CORE_COPY_XFPREGS
1430         if (elf_core_copy_task_xfpregs(p, &t->xfpu)) {
1431                 fill_note(&t->notes[2], "LINUX", NT_PRXFPREG, sizeof(t->xfpu),
1432                           &t->xfpu);
1433                 t->num_notes++;
1434                 sz += notesize(&t->notes[2]);
1435         }
1436 #endif
1437         return sz;
1438 }
1439
1440 /*
1441  * dump the segments for an MMU process
1442  */
1443 #ifdef CONFIG_MMU
1444 static int elf_fdpic_dump_segments(struct file *file, struct mm_struct *mm,
1445                                    size_t *size, unsigned long *limit)
1446 {
1447         struct vm_area_struct *vma;
1448
1449         for (vma = current->mm->mmap; vma; vma = vma->vm_next) {
1450                 unsigned long addr;
1451
1452                 if (!maydump(vma))
1453                         continue;
1454
1455                 for (addr = vma->vm_start;
1456                      addr < vma->vm_end;
1457                      addr += PAGE_SIZE
1458                      ) {
1459                         struct vm_area_struct *vma;
1460                         struct page *page;
1461
1462                         if (get_user_pages(current, current->mm, addr, 1, 0, 1,
1463                                            &page, &vma) <= 0) {
1464                                 DUMP_SEEK(file->f_pos + PAGE_SIZE);
1465                         }
1466                         else if (page == ZERO_PAGE(addr)) {
1467                                 DUMP_SEEK(file->f_pos + PAGE_SIZE);
1468                                 page_cache_release(page);
1469                         }
1470                         else {
1471                                 void *kaddr;
1472
1473                                 flush_cache_page(vma, addr, page_to_pfn(page));
1474                                 kaddr = kmap(page);
1475                                 if ((*size += PAGE_SIZE) > *limit ||
1476                                     !dump_write(file, kaddr, PAGE_SIZE)
1477                                     ) {
1478                                         kunmap(page);
1479                                         page_cache_release(page);
1480                                         return -EIO;
1481                                 }
1482                                 kunmap(page);
1483                                 page_cache_release(page);
1484                         }
1485                 }
1486         }
1487
1488         return 0;
1489
1490 end_coredump:
1491         return -EFBIG;
1492 }
1493 #endif
1494
1495 /*
1496  * dump the segments for a NOMMU process
1497  */
1498 #ifndef CONFIG_MMU
1499 static int elf_fdpic_dump_segments(struct file *file, struct mm_struct *mm,
1500                                    size_t *size, unsigned long *limit)
1501 {
1502         struct vm_list_struct *vml;
1503
1504         for (vml = current->mm->context.vmlist; vml; vml = vml->next) {
1505         struct vm_area_struct *vma = vml->vma;
1506
1507                 if (!maydump(vma))
1508                         continue;
1509
1510                 if ((*size += PAGE_SIZE) > *limit)
1511                         return -EFBIG;
1512
1513                 if (!dump_write(file, (void *) vma->vm_start,
1514                                 vma->vm_end - vma->vm_start))
1515                         return -EIO;
1516         }
1517
1518         return 0;
1519 }
1520 #endif
1521
1522 /*
1523  * Actual dumper
1524  *
1525  * This is a two-pass process; first we find the offsets of the bits,
1526  * and then they are actually written out.  If we run out of core limit
1527  * we just truncate.
1528  */
1529 static int elf_fdpic_core_dump(long signr, struct pt_regs *regs,
1530                                struct file *file)
1531 {
1532 #define NUM_NOTES       6
1533         int has_dumped = 0;
1534         mm_segment_t fs;
1535         int segs;
1536         size_t size = 0;
1537         int i;
1538         struct vm_area_struct *vma;
1539         struct elfhdr *elf = NULL;
1540         loff_t offset = 0, dataoff;
1541         unsigned long limit = current->signal->rlim[RLIMIT_CORE].rlim_cur;
1542         int numnote;
1543         struct memelfnote *notes = NULL;
1544         struct elf_prstatus *prstatus = NULL;   /* NT_PRSTATUS */
1545         struct elf_prpsinfo *psinfo = NULL;     /* NT_PRPSINFO */
1546         struct task_struct *g, *p;
1547         LIST_HEAD(thread_list);
1548         struct list_head *t;
1549         elf_fpregset_t *fpu = NULL;
1550 #ifdef ELF_CORE_COPY_XFPREGS
1551         elf_fpxregset_t *xfpu = NULL;
1552 #endif
1553         int thread_status_size = 0;
1554 #ifndef CONFIG_MMU
1555         struct vm_list_struct *vml;
1556 #endif
1557         elf_addr_t *auxv;
1558
1559         /*
1560          * We no longer stop all VM operations.
1561          *
1562          * This is because those proceses that could possibly change map_count
1563          * or the mmap / vma pages are now blocked in do_exit on current
1564          * finishing this core dump.
1565          *
1566          * Only ptrace can touch these memory addresses, but it doesn't change
1567          * the map_count or the pages allocated. So no possibility of crashing
1568          * exists while dumping the mm->vm_next areas to the core file.
1569          */
1570
1571         /* alloc memory for large data structures: too large to be on stack */
1572         elf = kmalloc(sizeof(*elf), GFP_KERNEL);
1573         if (!elf)
1574                 goto cleanup;
1575         prstatus = kzalloc(sizeof(*prstatus), GFP_KERNEL);
1576         if (!prstatus)
1577                 goto cleanup;
1578         psinfo = kmalloc(sizeof(*psinfo), GFP_KERNEL);
1579         if (!psinfo)
1580                 goto cleanup;
1581         notes = kmalloc(NUM_NOTES * sizeof(struct memelfnote), GFP_KERNEL);
1582         if (!notes)
1583                 goto cleanup;
1584         fpu = kmalloc(sizeof(*fpu), GFP_KERNEL);
1585         if (!fpu)
1586                 goto cleanup;
1587 #ifdef ELF_CORE_COPY_XFPREGS
1588         xfpu = kmalloc(sizeof(*xfpu), GFP_KERNEL);
1589         if (!xfpu)
1590                 goto cleanup;
1591 #endif
1592
1593         if (signr) {
1594                 struct elf_thread_status *tmp;
1595                 read_lock(&tasklist_lock);
1596                 do_each_thread(g,p)
1597                         if (current->mm == p->mm && current != p) {
1598                                 tmp = kzalloc(sizeof(*tmp), GFP_ATOMIC);
1599                                 if (!tmp) {
1600                                         read_unlock(&tasklist_lock);
1601                                         goto cleanup;
1602                                 }
1603                                 INIT_LIST_HEAD(&tmp->list);
1604                                 tmp->thread = p;
1605                                 list_add(&tmp->list, &thread_list);
1606                         }
1607                 while_each_thread(g,p);
1608                 read_unlock(&tasklist_lock);
1609                 list_for_each(t, &thread_list) {
1610                         struct elf_thread_status *tmp;
1611                         int sz;
1612
1613                         tmp = list_entry(t, struct elf_thread_status, list);
1614                         sz = elf_dump_thread_status(signr, tmp);
1615                         thread_status_size += sz;
1616                 }
1617         }
1618
1619         /* now collect the dump for the current */
1620         fill_prstatus(prstatus, current, signr);
1621         elf_core_copy_regs(&prstatus->pr_reg, regs);
1622
1623 #ifdef CONFIG_MMU
1624         segs = current->mm->map_count;
1625 #else
1626         segs = 0;
1627         for (vml = current->mm->context.vmlist; vml; vml = vml->next)
1628             segs++;
1629 #endif
1630 #ifdef ELF_CORE_EXTRA_PHDRS
1631         segs += ELF_CORE_EXTRA_PHDRS;
1632 #endif
1633
1634         /* Set up header */
1635         fill_elf_fdpic_header(elf, segs + 1);   /* including notes section */
1636
1637         has_dumped = 1;
1638         current->flags |= PF_DUMPCORE;
1639
1640         /*
1641          * Set up the notes in similar form to SVR4 core dumps made
1642          * with info from their /proc.
1643          */
1644
1645         fill_note(notes + 0, "CORE", NT_PRSTATUS, sizeof(*prstatus), prstatus);
1646         fill_psinfo(psinfo, current->group_leader, current->mm);
1647         fill_note(notes + 1, "CORE", NT_PRPSINFO, sizeof(*psinfo), psinfo);
1648
1649         numnote = 2;
1650
1651         auxv = (elf_addr_t *) current->mm->saved_auxv;
1652
1653         i = 0;
1654         do
1655                 i += 2;
1656         while (auxv[i - 2] != AT_NULL);
1657         fill_note(&notes[numnote++], "CORE", NT_AUXV,
1658                   i * sizeof(elf_addr_t), auxv);
1659
1660         /* Try to dump the FPU. */
1661         if ((prstatus->pr_fpvalid =
1662              elf_core_copy_task_fpregs(current, regs, fpu)))
1663                 fill_note(notes + numnote++,
1664                           "CORE", NT_PRFPREG, sizeof(*fpu), fpu);
1665 #ifdef ELF_CORE_COPY_XFPREGS
1666         if (elf_core_copy_task_xfpregs(current, xfpu))
1667                 fill_note(notes + numnote++,
1668                           "LINUX", NT_PRXFPREG, sizeof(*xfpu), xfpu);
1669 #endif
1670
1671         fs = get_fs();
1672         set_fs(KERNEL_DS);
1673
1674         DUMP_WRITE(elf, sizeof(*elf));
1675         offset += sizeof(*elf);                         /* Elf header */
1676         offset += (segs+1) * sizeof(struct elf_phdr);   /* Program headers */
1677
1678         /* Write notes phdr entry */
1679         {
1680                 struct elf_phdr phdr;
1681                 int sz = 0;
1682
1683                 for (i = 0; i < numnote; i++)
1684                         sz += notesize(notes + i);
1685
1686                 sz += thread_status_size;
1687
1688                 fill_elf_note_phdr(&phdr, sz, offset);
1689                 offset += sz;
1690                 DUMP_WRITE(&phdr, sizeof(phdr));
1691         }
1692
1693         /* Page-align dumped data */
1694         dataoff = offset = roundup(offset, ELF_EXEC_PAGESIZE);
1695
1696         /* write program headers for segments dump */
1697         for (
1698 #ifdef CONFIG_MMU
1699                 vma = current->mm->mmap; vma; vma = vma->vm_next
1700 #else
1701                         vml = current->mm->context.vmlist; vml; vml = vml->next
1702 #endif
1703              ) {
1704                 struct elf_phdr phdr;
1705                 size_t sz;
1706
1707 #ifndef CONFIG_MMU
1708                 vma = vml->vma;
1709 #endif
1710
1711                 sz = vma->vm_end - vma->vm_start;
1712
1713                 phdr.p_type = PT_LOAD;
1714                 phdr.p_offset = offset;
1715                 phdr.p_vaddr = vma->vm_start;
1716                 phdr.p_paddr = 0;
1717                 phdr.p_filesz = maydump(vma) ? sz : 0;
1718                 phdr.p_memsz = sz;
1719                 offset += phdr.p_filesz;
1720                 phdr.p_flags = vma->vm_flags & VM_READ ? PF_R : 0;
1721                 if (vma->vm_flags & VM_WRITE)
1722                         phdr.p_flags |= PF_W;
1723                 if (vma->vm_flags & VM_EXEC)
1724                         phdr.p_flags |= PF_X;
1725                 phdr.p_align = ELF_EXEC_PAGESIZE;
1726
1727                 DUMP_WRITE(&phdr, sizeof(phdr));
1728         }
1729
1730 #ifdef ELF_CORE_WRITE_EXTRA_PHDRS
1731         ELF_CORE_WRITE_EXTRA_PHDRS;
1732 #endif
1733
1734         /* write out the notes section */
1735         for (i = 0; i < numnote; i++)
1736                 if (!writenote(notes + i, file))
1737                         goto end_coredump;
1738
1739         /* write out the thread status notes section */
1740         list_for_each(t, &thread_list) {
1741                 struct elf_thread_status *tmp =
1742                                 list_entry(t, struct elf_thread_status, list);
1743
1744                 for (i = 0; i < tmp->num_notes; i++)
1745                         if (!writenote(&tmp->notes[i], file))
1746                                 goto end_coredump;
1747         }
1748
1749         DUMP_SEEK(dataoff);
1750
1751         if (elf_fdpic_dump_segments(file, current->mm, &size, &limit) < 0)
1752                 goto end_coredump;
1753
1754 #ifdef ELF_CORE_WRITE_EXTRA_DATA
1755         ELF_CORE_WRITE_EXTRA_DATA;
1756 #endif
1757
1758         if (file->f_pos != offset) {
1759                 /* Sanity check */
1760                 printk(KERN_WARNING
1761                        "elf_core_dump: file->f_pos (%lld) != offset (%lld)\n",
1762                        file->f_pos, offset);
1763         }
1764
1765 end_coredump:
1766         set_fs(fs);
1767
1768 cleanup:
1769         while (!list_empty(&thread_list)) {
1770                 struct list_head *tmp = thread_list.next;
1771                 list_del(tmp);
1772                 kfree(list_entry(tmp, struct elf_thread_status, list));
1773         }
1774
1775         kfree(elf);
1776         kfree(prstatus);
1777         kfree(psinfo);
1778         kfree(notes);
1779         kfree(fpu);
1780 #ifdef ELF_CORE_COPY_XFPREGS
1781         kfree(xfpu);
1782 #endif
1783         return has_dumped;
1784 #undef NUM_NOTES
1785 }
1786
1787 #endif          /* USE_ELF_CORE_DUMP */