git.oblomov.eu Git - linux-2.6/blob - 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/security.h>
  29 #include <linux/highmem.h>
  30 #include <linux/highuid.h>
  31 #include <linux/personality.h>
  32 #include <linux/ptrace.h>
  33 #include <linux/init.h>
  34 #include <linux/elf.h>
  35 #include <linux/elf-fdpic.h>
  36 #include <linux/elfcore.h>
  37
  38 #include <asm/uaccess.h>
  39 #include <asm/param.h>
  40 #include <asm/pgalloc.h>
  41
  42 typedef char *elf_caddr_t;
  43
  44 #if 0
  45 #define kdebug(fmt, ...) printk("FDPIC "fmt"\n" ,##__VA_ARGS__ )
  46 #else
  47 #define kdebug(fmt, ...) do {} while(0)
  48 #endif
  49
  50 #if 0
  51 #define kdcore(fmt, ...) printk("FDPIC "fmt"\n" ,##__VA_ARGS__ )
  52 #else
  53 #define kdcore(fmt, ...) do {} while(0)
  54 #endif
  55
  56 MODULE_LICENSE("GPL");
  57
  58 static int load_elf_fdpic_binary(struct linux_binprm *, struct pt_regs *);
  59 static int elf_fdpic_fetch_phdrs(struct elf_fdpic_params *, struct file *);
  60 static int elf_fdpic_map_file(struct elf_fdpic_params *, struct file *,
  61                               struct mm_struct *, const char *);
  62
  63 static int create_elf_fdpic_tables(struct linux_binprm *, struct mm_struct *,
  64                                    struct elf_fdpic_params *,
  65                                    struct elf_fdpic_params *);
  66
  67 #ifndef CONFIG_MMU
  68 static int elf_fdpic_transfer_args_to_stack(struct linux_binprm *,
  69                                             unsigned long *);
  70 static int elf_fdpic_map_file_constdisp_on_uclinux(struct elf_fdpic_params *,
  71                                                    struct file *,
  72                                                    struct mm_struct *);
  73 #endif
  74
  75 static int elf_fdpic_map_file_by_direct_mmap(struct elf_fdpic_params *,
  76                                              struct file *, struct mm_struct *);
  77
  78 #if defined(USE_ELF_CORE_DUMP) && defined(CONFIG_ELF_CORE)
  79 static int elf_fdpic_core_dump(long, struct pt_regs *, struct file *, unsigned long limit);
  80 #endif
  81
  82 static struct linux_binfmt elf_fdpic_format = {
  83         .module         = THIS_MODULE,
  84         .load_binary    = load_elf_fdpic_binary,
  85 #if defined(USE_ELF_CORE_DUMP) && defined(CONFIG_ELF_CORE)
  86         .core_dump      = elf_fdpic_core_dump,
  87 #endif
  88         .min_coredump   = ELF_EXEC_PAGESIZE,
  89 };
  90
  91 static int __init init_elf_fdpic_binfmt(void)
  92 {
  93         return register_binfmt(&elf_fdpic_format);
  94 }
  95
  96 static void __exit exit_elf_fdpic_binfmt(void)
  97 {
  98         unregister_binfmt(&elf_fdpic_format);
  99 }
 100
 101 core_initcall(init_elf_fdpic_binfmt);
 102 module_exit(exit_elf_fdpic_binfmt);
 103
 104 static int is_elf_fdpic(struct elfhdr *hdr, struct file *file)
 105 {
 106         if (memcmp(hdr->e_ident, ELFMAG, SELFMAG) != 0)
 107                 return 0;
 108         if (hdr->e_type != ET_EXEC && hdr->e_type != ET_DYN)
 109                 return 0;
 110         if (!elf_check_arch(hdr) || !elf_check_fdpic(hdr))
 111                 return 0;
 112         if (!file->f_op || !file->f_op->mmap)
 113                 return 0;
 114         return 1;
 115 }
 116
 117 /*****************************************************************************/
 118 /*
 119  * read the program headers table into memory
 120  */
 121 static int elf_fdpic_fetch_phdrs(struct elf_fdpic_params *params,
 122                                  struct file *file)
 123 {
 124         struct elf32_phdr *phdr;
 125         unsigned long size;
 126         int retval, loop;
 127
 128         if (params->hdr.e_phentsize != sizeof(struct elf_phdr))
 129                 return -ENOMEM;
 130         if (params->hdr.e_phnum > 65536U / sizeof(struct elf_phdr))
 131                 return -ENOMEM;
 132
 133         size = params->hdr.e_phnum * sizeof(struct elf_phdr);
 134         params->phdrs = kmalloc(size, GFP_KERNEL);
 135         if (!params->phdrs)
 136                 return -ENOMEM;
 137
 138         retval = kernel_read(file, params->hdr.e_phoff,
 139                              (char *) params->phdrs, size);
 140         if (unlikely(retval != size))
 141                 return retval < 0 ? retval : -ENOEXEC;
 142
 143         /* determine stack size for this binary */
 144         phdr = params->phdrs;
 145         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 146                 if (phdr->p_type != PT_GNU_STACK)
 147                         continue;
 148
 149                 if (phdr->p_flags & PF_X)
 150                         params->flags |= ELF_FDPIC_FLAG_EXEC_STACK;
 151                 else
 152                         params->flags |= ELF_FDPIC_FLAG_NOEXEC_STACK;
 153
 154                 params->stack_size = phdr->p_memsz;
 155                 break;
 156         }
 157
 158         return 0;
 159 }
 160
 161 /*****************************************************************************/
 162 /*
 163  * load an fdpic binary into various bits of memory
 164  */
 165 static int load_elf_fdpic_binary(struct linux_binprm *bprm,
 166                                  struct pt_regs *regs)
 167 {
 168         struct elf_fdpic_params exec_params, interp_params;
 169         struct elf_phdr *phdr;
 170         unsigned long stack_size, entryaddr;
 171 #ifdef ELF_FDPIC_PLAT_INIT
 172         unsigned long dynaddr;
 173 #endif
 174         struct file *interpreter = NULL; /* to shut gcc up */
 175         char *interpreter_name = NULL;
 176         int executable_stack;
 177         int retval, i;
 178
 179         kdebug("____ LOAD %d ____", current->pid);
 180
 181         memset(&exec_params, 0, sizeof(exec_params));
 182         memset(&interp_params, 0, sizeof(interp_params));
 183
 184         exec_params.hdr = *(struct elfhdr *) bprm->buf;
 185         exec_params.flags = ELF_FDPIC_FLAG_PRESENT | ELF_FDPIC_FLAG_EXECUTABLE;
 186
 187         /* check that this is a binary we know how to deal with */
 188         retval = -ENOEXEC;
 189         if (!is_elf_fdpic(&exec_params.hdr, bprm->file))
 190                 goto error;
 191
 192         /* read the program header table */
 193         retval = elf_fdpic_fetch_phdrs(&exec_params, bprm->file);
 194         if (retval < 0)
 195                 goto error;
 196
 197         /* scan for a program header that specifies an interpreter */
 198         phdr = exec_params.phdrs;
 199
 200         for (i = 0; i < exec_params.hdr.e_phnum; i++, phdr++) {
 201                 switch (phdr->p_type) {
 202                 case PT_INTERP:
 203                         retval = -ENOMEM;
 204                         if (phdr->p_filesz > PATH_MAX)
 205                                 goto error;
 206                         retval = -ENOENT;
 207                         if (phdr->p_filesz < 2)
 208                                 goto error;
 209
 210                         /* read the name of the interpreter into memory */
 211                         interpreter_name = kmalloc(phdr->p_filesz, GFP_KERNEL);
 212                         if (!interpreter_name)
 213                                 goto error;
 214
 215                         retval = kernel_read(bprm->file,
 216                                              phdr->p_offset,
 217                                              interpreter_name,
 218                                              phdr->p_filesz);
 219                         if (unlikely(retval != phdr->p_filesz)) {
 220                                 if (retval >= 0)
 221                                         retval = -ENOEXEC;
 222                                 goto error;
 223                         }
 224
 225                         retval = -ENOENT;
 226                         if (interpreter_name[phdr->p_filesz - 1] != '\0')
 227                                 goto error;
 228
 229                         kdebug("Using ELF interpreter %s", interpreter_name);
 230
 231                         /* replace the program with the interpreter */
 232                         interpreter = open_exec(interpreter_name);
 233                         retval = PTR_ERR(interpreter);
 234                         if (IS_ERR(interpreter)) {
 235                                 interpreter = NULL;
 236                                 goto error;
 237                         }
 238
 239                         /*
 240                          * If the binary is not readable then enforce
 241                          * mm->dumpable = 0 regardless of the interpreter's
 242                          * permissions.
 243                          */
 244                         if (file_permission(interpreter, MAY_READ) < 0)
 245                                 bprm->interp_flags |= BINPRM_FLAGS_ENFORCE_NONDUMP;
 246
 247                         retval = kernel_read(interpreter, 0, bprm->buf,
 248                                              BINPRM_BUF_SIZE);
 249                         if (unlikely(retval != BINPRM_BUF_SIZE)) {
 250                                 if (retval >= 0)
 251                                         retval = -ENOEXEC;
 252                                 goto error;
 253                         }
 254
 255                         interp_params.hdr = *((struct elfhdr *) bprm->buf);
 256                         break;
 257
 258                 case PT_LOAD:
 259 #ifdef CONFIG_MMU
 260                         if (exec_params.load_addr == 0)
 261                                 exec_params.load_addr = phdr->p_vaddr;
 262 #endif
 263                         break;
 264                 }
 265
 266         }
 267
 268         if (elf_check_const_displacement(&exec_params.hdr))
 269                 exec_params.flags |= ELF_FDPIC_FLAG_CONSTDISP;
 270
 271         /* perform insanity checks on the interpreter */
 272         if (interpreter_name) {
 273                 retval = -ELIBBAD;
 274                 if (!is_elf_fdpic(&interp_params.hdr, interpreter))
 275                         goto error;
 276
 277                 interp_params.flags = ELF_FDPIC_FLAG_PRESENT;
 278
 279                 /* read the interpreter's program header table */
 280                 retval = elf_fdpic_fetch_phdrs(&interp_params, interpreter);
 281                 if (retval < 0)
 282                         goto error;
 283         }
 284
 285         stack_size = exec_params.stack_size;
 286         if (stack_size < interp_params.stack_size)
 287                 stack_size = interp_params.stack_size;
 288
 289         if (exec_params.flags & ELF_FDPIC_FLAG_EXEC_STACK)
 290                 executable_stack = EXSTACK_ENABLE_X;
 291         else if (exec_params.flags & ELF_FDPIC_FLAG_NOEXEC_STACK)
 292                 executable_stack = EXSTACK_DISABLE_X;
 293         else if (interp_params.flags & ELF_FDPIC_FLAG_EXEC_STACK)
 294                 executable_stack = EXSTACK_ENABLE_X;
 295         else if (interp_params.flags & ELF_FDPIC_FLAG_NOEXEC_STACK)
 296                 executable_stack = EXSTACK_DISABLE_X;
 297         else
 298                 executable_stack = EXSTACK_DEFAULT;
 299
 300         retval = -ENOEXEC;
 301         if (stack_size == 0)
 302                 goto error;
 303
 304         if (elf_check_const_displacement(&interp_params.hdr))
 305                 interp_params.flags |= ELF_FDPIC_FLAG_CONSTDISP;
 306
 307         /* flush all traces of the currently running executable */
 308         retval = flush_old_exec(bprm);
 309         if (retval)
 310                 goto error;
 311
 312         /* there's now no turning back... the old userspace image is dead,
 313          * defunct, deceased, etc. after this point we have to exit via
 314          * error_kill */
 315         set_personality(PER_LINUX_FDPIC);
 316         set_binfmt(&elf_fdpic_format);
 317
 318         current->mm->start_code = 0;
 319         current->mm->end_code = 0;
 320         current->mm->start_stack = 0;
 321         current->mm->start_data = 0;
 322         current->mm->end_data = 0;
 323         current->mm->context.exec_fdpic_loadmap = 0;
 324         current->mm->context.interp_fdpic_loadmap = 0;
 325
 326         current->flags &= ~PF_FORKNOEXEC;
 327
 328 #ifdef CONFIG_MMU
 329         elf_fdpic_arch_lay_out_mm(&exec_params,
 330                                   &interp_params,
 331                                   &current->mm->start_stack,
 332                                   &current->mm->start_brk);
 333
 334         retval = setup_arg_pages(bprm, current->mm->start_stack,
 335                                  executable_stack);
 336         if (retval < 0) {
 337                 send_sig(SIGKILL, current, 0);
 338                 goto error_kill;
 339         }
 340 #endif
 341
 342         /* load the executable and interpreter into memory */
 343         retval = elf_fdpic_map_file(&exec_params, bprm->file, current->mm,
 344                                     "executable");
 345         if (retval < 0)
 346                 goto error_kill;
 347
 348         if (interpreter_name) {
 349                 retval = elf_fdpic_map_file(&interp_params, interpreter,
 350                                             current->mm, "interpreter");
 351                 if (retval < 0) {
 352                         printk(KERN_ERR "Unable to load interpreter\n");
 353                         goto error_kill;
 354                 }
 355
 356                 allow_write_access(interpreter);
 357                 fput(interpreter);
 358                 interpreter = NULL;
 359         }
 360
 361 #ifdef CONFIG_MMU
 362         if (!current->mm->start_brk)
 363                 current->mm->start_brk = current->mm->end_data;
 364
 365         current->mm->brk = current->mm->start_brk =
 366                 PAGE_ALIGN(current->mm->start_brk);
 367
 368 #else
 369         /* create a stack and brk area big enough for everyone
 370          * - the brk heap starts at the bottom and works up
 371          * - the stack starts at the top and works down
 372          */
 373         stack_size = (stack_size + PAGE_SIZE - 1) & PAGE_MASK;
 374         if (stack_size < PAGE_SIZE * 2)
 375                 stack_size = PAGE_SIZE * 2;
 376
 377         down_write(&current->mm->mmap_sem);
 378         current->mm->start_brk = do_mmap(NULL, 0, stack_size,
 379                                          PROT_READ | PROT_WRITE | PROT_EXEC,
 380                                          MAP_PRIVATE | MAP_ANONYMOUS | MAP_GROWSDOWN,
 381                                          0);
 382
 383         if (IS_ERR_VALUE(current->mm->start_brk)) {
 384                 up_write(&current->mm->mmap_sem);
 385                 retval = current->mm->start_brk;
 386                 current->mm->start_brk = 0;
 387                 goto error_kill;
 388         }
 389
 390         up_write(&current->mm->mmap_sem);
 391
 392         current->mm->brk = current->mm->start_brk;
 393         current->mm->context.end_brk = current->mm->start_brk;
 394         current->mm->context.end_brk +=
 395                 (stack_size > PAGE_SIZE) ? (stack_size - PAGE_SIZE) : 0;
 396         current->mm->start_stack = current->mm->start_brk + stack_size;
 397 #endif
 398
 399         install_exec_creds(bprm);
 400         current->flags &= ~PF_FORKNOEXEC;
 401         if (create_elf_fdpic_tables(bprm, current->mm,
 402                                     &exec_params, &interp_params) < 0)
 403                 goto error_kill;
 404
 405         kdebug("- start_code  %lx", current->mm->start_code);
 406         kdebug("- end_code    %lx", current->mm->end_code);
 407         kdebug("- start_data  %lx", current->mm->start_data);
 408         kdebug("- end_data    %lx", current->mm->end_data);
 409         kdebug("- start_brk   %lx", current->mm->start_brk);
 410         kdebug("- brk         %lx", current->mm->brk);
 411         kdebug("- start_stack %lx", current->mm->start_stack);
 412
 413 #ifdef ELF_FDPIC_PLAT_INIT
 414         /*
 415          * The ABI may specify that certain registers be set up in special
 416          * ways (on i386 %edx is the address of a DT_FINI function, for
 417          * example.  This macro performs whatever initialization to
 418          * the regs structure is required.
 419          */
 420         dynaddr = interp_params.dynamic_addr ?: exec_params.dynamic_addr;
 421         ELF_FDPIC_PLAT_INIT(regs, exec_params.map_addr, interp_params.map_addr,
 422                             dynaddr);
 423 #endif
 424
 425         /* everything is now ready... get the userspace context ready to roll */
 426         entryaddr = interp_params.entry_addr ?: exec_params.entry_addr;
 427         start_thread(regs, entryaddr, current->mm->start_stack);
 428
 429         retval = 0;
 430
 431 error:
 432         if (interpreter) {
 433                 allow_write_access(interpreter);
 434                 fput(interpreter);
 435         }
 436         kfree(interpreter_name);
 437         kfree(exec_params.phdrs);
 438         kfree(exec_params.loadmap);
 439         kfree(interp_params.phdrs);
 440         kfree(interp_params.loadmap);
 441         return retval;
 442
 443         /* unrecoverable error - kill the process */
 444 error_kill:
 445         send_sig(SIGSEGV, current, 0);
 446         goto error;
 447
 448 }
 449
 450 /*****************************************************************************/
 451
 452 #ifndef ELF_BASE_PLATFORM
 453 /*
 454  * AT_BASE_PLATFORM indicates the "real" hardware/microarchitecture.
 455  * If the arch defines ELF_BASE_PLATFORM (in asm/elf.h), the value
 456  * will be copied to the user stack in the same manner as AT_PLATFORM.
 457  */
 458 #define ELF_BASE_PLATFORM NULL
 459 #endif
 460
 461 /*
 462  * present useful information to the program by shovelling it onto the new
 463  * process's stack
 464  */
 465 static int create_elf_fdpic_tables(struct linux_binprm *bprm,
 466                                    struct mm_struct *mm,
 467                                    struct elf_fdpic_params *exec_params,
 468                                    struct elf_fdpic_params *interp_params)
 469 {
 470         const struct cred *cred = current_cred();
 471         unsigned long sp, csp, nitems;
 472         elf_caddr_t __user *argv, *envp;
 473         size_t platform_len = 0, len;
 474         char *k_platform, *k_base_platform;
 475         char __user *u_platform, *u_base_platform, *p;
 476         long hwcap;
 477         int loop;
 478         int nr; /* reset for each csp adjustment */
 479
 480 #ifdef CONFIG_MMU
 481         /* In some cases (e.g. Hyper-Threading), we want to avoid L1 evictions
 482          * by the processes running on the same package. One thing we can do is
 483          * to shuffle the initial stack for them, so we give the architecture
 484          * an opportunity to do so here.
 485          */
 486         sp = arch_align_stack(bprm->p);
 487 #else
 488         sp = mm->start_stack;
 489
 490         /* stack the program arguments and environment */
 491         if (elf_fdpic_transfer_args_to_stack(bprm, &sp) < 0)
 492                 return -EFAULT;
 493 #endif
 494
 495         hwcap = ELF_HWCAP;
 496
 497         /*
 498          * If this architecture has a platform capability string, copy it
 499          * to userspace.  In some cases (Sparc), this info is impossible
 500          * for userspace to get any other way, in others (i386) it is
 501          * merely difficult.
 502          */
 503         k_platform = ELF_PLATFORM;
 504         u_platform = NULL;
 505
 506         if (k_platform) {
 507                 platform_len = strlen(k_platform) + 1;
 508                 sp -= platform_len;
 509                 u_platform = (char __user *) sp;
 510                 if (__copy_to_user(u_platform, k_platform, platform_len) != 0)
 511                         return -EFAULT;
 512         }
 513
 514         /*
 515          * If this architecture has a "base" platform capability
 516          * string, copy it to userspace.
 517          */
 518         k_base_platform = ELF_BASE_PLATFORM;
 519         u_base_platform = NULL;
 520
 521         if (k_base_platform) {
 522                 platform_len = strlen(k_base_platform) + 1;
 523                 sp -= platform_len;
 524                 u_base_platform = (char __user *) sp;
 525                 if (__copy_to_user(u_base_platform, k_base_platform, platform_len) != 0)
 526                         return -EFAULT;
 527         }
 528
 529         sp &= ~7UL;
 530
 531         /* stack the load map(s) */
 532         len = sizeof(struct elf32_fdpic_loadmap);
 533         len += sizeof(struct elf32_fdpic_loadseg) * exec_params->loadmap->nsegs;
 534         sp = (sp - len) & ~7UL;
 535         exec_params->map_addr = sp;
 536
 537         if (copy_to_user((void __user *) sp, exec_params->loadmap, len) != 0)
 538                 return -EFAULT;
 539
 540         current->mm->context.exec_fdpic_loadmap = (unsigned long) sp;
 541
 542         if (interp_params->loadmap) {
 543                 len = sizeof(struct elf32_fdpic_loadmap);
 544                 len += sizeof(struct elf32_fdpic_loadseg) *
 545                         interp_params->loadmap->nsegs;
 546                 sp = (sp - len) & ~7UL;
 547                 interp_params->map_addr = sp;
 548
 549                 if (copy_to_user((void __user *) sp, interp_params->loadmap,
 550                                  len) != 0)
 551                         return -EFAULT;
 552
 553                 current->mm->context.interp_fdpic_loadmap = (unsigned long) sp;
 554         }
 555
 556         /* force 16 byte _final_ alignment here for generality */
 557 #define DLINFO_ITEMS 15
 558
 559         nitems = 1 + DLINFO_ITEMS + (k_platform ? 1 : 0) +
 560                 (k_base_platform ? 1 : 0) + AT_VECTOR_SIZE_ARCH;
 561
 562         if (bprm->interp_flags & BINPRM_FLAGS_EXECFD)
 563                 nitems++;
 564
 565         csp = sp;
 566         sp -= nitems * 2 * sizeof(unsigned long);
 567         sp -= (bprm->envc + 1) * sizeof(char *);        /* envv[] */
 568         sp -= (bprm->argc + 1) * sizeof(char *);        /* argv[] */
 569         sp -= 1 * sizeof(unsigned long);                /* argc */
 570
 571         csp -= sp & 15UL;
 572         sp -= sp & 15UL;
 573
 574         /* put the ELF interpreter info on the stack */
 575 #define NEW_AUX_ENT(id, val)                                            \
 576         do {                                                            \
 577                 struct { unsigned long _id, _val; } __user *ent;        \
 578                                                                         \
 579                 ent = (void __user *) csp;                              \
 580                 __put_user((id), &ent[nr]._id);                         \
 581                 __put_user((val), &ent[nr]._val);                       \
 582                 nr++;                                                   \
 583         } while (0)
 584
 585         nr = 0;
 586         csp -= 2 * sizeof(unsigned long);
 587         NEW_AUX_ENT(AT_NULL, 0);
 588         if (k_platform) {
 589                 nr = 0;
 590                 csp -= 2 * sizeof(unsigned long);
 591                 NEW_AUX_ENT(AT_PLATFORM,
 592                             (elf_addr_t) (unsigned long) u_platform);
 593         }
 594
 595         if (k_base_platform) {
 596                 nr = 0;
 597                 csp -= 2 * sizeof(unsigned long);
 598                 NEW_AUX_ENT(AT_BASE_PLATFORM,
 599                             (elf_addr_t) (unsigned long) u_base_platform);
 600         }
 601
 602         if (bprm->interp_flags & BINPRM_FLAGS_EXECFD) {
 603                 nr = 0;
 604                 csp -= 2 * sizeof(unsigned long);
 605                 NEW_AUX_ENT(AT_EXECFD, bprm->interp_data);
 606         }
 607
 608         nr = 0;
 609         csp -= DLINFO_ITEMS * 2 * sizeof(unsigned long);
 610         NEW_AUX_ENT(AT_HWCAP,   hwcap);
 611         NEW_AUX_ENT(AT_PAGESZ,  PAGE_SIZE);
 612         NEW_AUX_ENT(AT_CLKTCK,  CLOCKS_PER_SEC);
 613         NEW_AUX_ENT(AT_PHDR,    exec_params->ph_addr);
 614         NEW_AUX_ENT(AT_PHENT,   sizeof(struct elf_phdr));
 615         NEW_AUX_ENT(AT_PHNUM,   exec_params->hdr.e_phnum);
 616         NEW_AUX_ENT(AT_BASE,    interp_params->elfhdr_addr);
 617         NEW_AUX_ENT(AT_FLAGS,   0);
 618         NEW_AUX_ENT(AT_ENTRY,   exec_params->entry_addr);
 619         NEW_AUX_ENT(AT_UID,     (elf_addr_t) cred->uid);
 620         NEW_AUX_ENT(AT_EUID,    (elf_addr_t) cred->euid);
 621         NEW_AUX_ENT(AT_GID,     (elf_addr_t) cred->gid);
 622         NEW_AUX_ENT(AT_EGID,    (elf_addr_t) cred->egid);
 623         NEW_AUX_ENT(AT_SECURE,  security_bprm_secureexec(bprm));
 624         NEW_AUX_ENT(AT_EXECFN,  bprm->exec);
 625
 626 #ifdef ARCH_DLINFO
 627         nr = 0;
 628         csp -= AT_VECTOR_SIZE_ARCH * 2 * sizeof(unsigned long);
 629
 630         /* ARCH_DLINFO must come last so platform specific code can enforce
 631          * special alignment requirements on the AUXV if necessary (eg. PPC).
 632          */
 633         ARCH_DLINFO;
 634 #endif
 635 #undef NEW_AUX_ENT
 636
 637         /* allocate room for argv[] and envv[] */
 638         csp -= (bprm->envc + 1) * sizeof(elf_caddr_t);
 639         envp = (elf_caddr_t __user *) csp;
 640         csp -= (bprm->argc + 1) * sizeof(elf_caddr_t);
 641         argv = (elf_caddr_t __user *) csp;
 642
 643         /* stack argc */
 644         csp -= sizeof(unsigned long);
 645         __put_user(bprm->argc, (unsigned long __user *) csp);
 646
 647         BUG_ON(csp != sp);
 648
 649         /* fill in the argv[] array */
 650 #ifdef CONFIG_MMU
 651         current->mm->arg_start = bprm->p;
 652 #else
 653         current->mm->arg_start = current->mm->start_stack -
 654                 (MAX_ARG_PAGES * PAGE_SIZE - bprm->p);
 655 #endif
 656
 657         p = (char __user *) current->mm->arg_start;
 658         for (loop = bprm->argc; loop > 0; loop--) {
 659                 __put_user((elf_caddr_t) p, argv++);
 660                 len = strnlen_user(p, MAX_ARG_STRLEN);
 661                 if (!len || len > MAX_ARG_STRLEN)
 662                         return -EINVAL;
 663                 p += len;
 664         }
 665         __put_user(NULL, argv);
 666         current->mm->arg_end = (unsigned long) p;
 667
 668         /* fill in the envv[] array */
 669         current->mm->env_start = (unsigned long) p;
 670         for (loop = bprm->envc; loop > 0; loop--) {
 671                 __put_user((elf_caddr_t)(unsigned long) p, envp++);
 672                 len = strnlen_user(p, MAX_ARG_STRLEN);
 673                 if (!len || len > MAX_ARG_STRLEN)
 674                         return -EINVAL;
 675                 p += len;
 676         }
 677         __put_user(NULL, envp);
 678         current->mm->env_end = (unsigned long) p;
 679
 680         mm->start_stack = (unsigned long) sp;
 681         return 0;
 682 }
 683
 684 /*****************************************************************************/
 685 /*
 686  * transfer the program arguments and environment from the holding pages onto
 687  * the stack
 688  */
 689 #ifndef CONFIG_MMU
 690 static int elf_fdpic_transfer_args_to_stack(struct linux_binprm *bprm,
 691                                             unsigned long *_sp)
 692 {
 693         unsigned long index, stop, sp;
 694         char *src;
 695         int ret = 0;
 696
 697         stop = bprm->p >> PAGE_SHIFT;
 698         sp = *_sp;
 699
 700         for (index = MAX_ARG_PAGES - 1; index >= stop; index--) {
 701                 src = kmap(bprm->page[index]);
 702                 sp -= PAGE_SIZE;
 703                 if (copy_to_user((void *) sp, src, PAGE_SIZE) != 0)
 704                         ret = -EFAULT;
 705                 kunmap(bprm->page[index]);
 706                 if (ret < 0)
 707                         goto out;
 708         }
 709
 710         *_sp = (*_sp - (MAX_ARG_PAGES * PAGE_SIZE - bprm->p)) & ~15;
 711
 712 out:
 713         return ret;
 714 }
 715 #endif
 716
 717 /*****************************************************************************/
 718 /*
 719  * load the appropriate binary image (executable or interpreter) into memory
 720  * - we assume no MMU is available
 721  * - if no other PIC bits are set in params->hdr->e_flags
 722  *   - we assume that the LOADable segments in the binary are independently relocatable
 723  *   - we assume R/O executable segments are shareable
 724  * - else
 725  *   - we assume the loadable parts of the image to require fixed displacement
 726  *   - the image is not shareable
 727  */
 728 static int elf_fdpic_map_file(struct elf_fdpic_params *params,
 729                               struct file *file,
 730                               struct mm_struct *mm,
 731                               const char *what)
 732 {
 733         struct elf32_fdpic_loadmap *loadmap;
 734 #ifdef CONFIG_MMU
 735         struct elf32_fdpic_loadseg *mseg;
 736 #endif
 737         struct elf32_fdpic_loadseg *seg;
 738         struct elf32_phdr *phdr;
 739         unsigned long load_addr, stop;
 740         unsigned nloads, tmp;
 741         size_t size;
 742         int loop, ret;
 743
 744         /* allocate a load map table */
 745         nloads = 0;
 746         for (loop = 0; loop < params->hdr.e_phnum; loop++)
 747                 if (params->phdrs[loop].p_type == PT_LOAD)
 748                         nloads++;
 749
 750         if (nloads == 0)
 751                 return -ELIBBAD;
 752
 753         size = sizeof(*loadmap) + nloads * sizeof(*seg);
 754         loadmap = kzalloc(size, GFP_KERNEL);
 755         if (!loadmap)
 756                 return -ENOMEM;
 757
 758         params->loadmap = loadmap;
 759
 760         loadmap->version = ELF32_FDPIC_LOADMAP_VERSION;
 761         loadmap->nsegs = nloads;
 762
 763         load_addr = params->load_addr;
 764         seg = loadmap->segs;
 765
 766         /* map the requested LOADs into the memory space */
 767         switch (params->flags & ELF_FDPIC_FLAG_ARRANGEMENT) {
 768         case ELF_FDPIC_FLAG_CONSTDISP:
 769         case ELF_FDPIC_FLAG_CONTIGUOUS:
 770 #ifndef CONFIG_MMU
 771                 ret = elf_fdpic_map_file_constdisp_on_uclinux(params, file, mm);
 772                 if (ret < 0)
 773                         return ret;
 774                 break;
 775 #endif
 776         default:
 777                 ret = elf_fdpic_map_file_by_direct_mmap(params, file, mm);
 778                 if (ret < 0)
 779                         return ret;
 780                 break;
 781         }
 782
 783         /* map the entry point */
 784         if (params->hdr.e_entry) {
 785                 seg = loadmap->segs;
 786                 for (loop = loadmap->nsegs; loop > 0; loop--, seg++) {
 787                         if (params->hdr.e_entry >= seg->p_vaddr &&
 788                             params->hdr.e_entry < seg->p_vaddr + seg->p_memsz) {
 789                                 params->entry_addr =
 790                                         (params->hdr.e_entry - seg->p_vaddr) +
 791                                         seg->addr;
 792                                 break;
 793                         }
 794                 }
 795         }
 796
 797         /* determine where the program header table has wound up if mapped */
 798         stop = params->hdr.e_phoff;
 799         stop += params->hdr.e_phnum * sizeof (struct elf_phdr);
 800         phdr = params->phdrs;
 801
 802         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 803                 if (phdr->p_type != PT_LOAD)
 804                         continue;
 805
 806                 if (phdr->p_offset > params->hdr.e_phoff ||
 807                     phdr->p_offset + phdr->p_filesz < stop)
 808                         continue;
 809
 810                 seg = loadmap->segs;
 811                 for (loop = loadmap->nsegs; loop > 0; loop--, seg++) {
 812                         if (phdr->p_vaddr >= seg->p_vaddr &&
 813                             phdr->p_vaddr + phdr->p_filesz <=
 814                             seg->p_vaddr + seg->p_memsz) {
 815                                 params->ph_addr =
 816                                         (phdr->p_vaddr - seg->p_vaddr) +
 817                                         seg->addr +
 818                                         params->hdr.e_phoff - phdr->p_offset;
 819                                 break;
 820                         }
 821                 }
 822                 break;
 823         }
 824
 825         /* determine where the dynamic section has wound up if there is one */
 826         phdr = params->phdrs;
 827         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 828                 if (phdr->p_type != PT_DYNAMIC)
 829                         continue;
 830
 831                 seg = loadmap->segs;
 832                 for (loop = loadmap->nsegs; loop > 0; loop--, seg++) {
 833                         if (phdr->p_vaddr >= seg->p_vaddr &&
 834                             phdr->p_vaddr + phdr->p_memsz <=
 835                             seg->p_vaddr + seg->p_memsz) {
 836                                 params->dynamic_addr =
 837                                         (phdr->p_vaddr - seg->p_vaddr) +
 838                                         seg->addr;
 839
 840                                 /* check the dynamic section contains at least
 841                                  * one item, and that the last item is a NULL
 842                                  * entry */
 843                                 if (phdr->p_memsz == 0 ||
 844                                     phdr->p_memsz % sizeof(Elf32_Dyn) != 0)
 845                                         goto dynamic_error;
 846
 847                                 tmp = phdr->p_memsz / sizeof(Elf32_Dyn);
 848                                 if (((Elf32_Dyn *)
 849                                      params->dynamic_addr)[tmp - 1].d_tag != 0)
 850                                         goto dynamic_error;
 851                                 break;
 852                         }
 853                 }
 854                 break;
 855         }
 856
 857         /* now elide adjacent segments in the load map on MMU linux
 858          * - on uClinux the holes between may actually be filled with system
 859          *   stuff or stuff from other processes
 860          */
 861 #ifdef CONFIG_MMU
 862         nloads = loadmap->nsegs;
 863         mseg = loadmap->segs;
 864         seg = mseg + 1;
 865         for (loop = 1; loop < nloads; loop++) {
 866                 /* see if we have a candidate for merging */
 867                 if (seg->p_vaddr - mseg->p_vaddr == seg->addr - mseg->addr) {
 868                         load_addr = PAGE_ALIGN(mseg->addr + mseg->p_memsz);
 869                         if (load_addr == (seg->addr & PAGE_MASK)) {
 870                                 mseg->p_memsz +=
 871                                         load_addr -
 872                                         (mseg->addr + mseg->p_memsz);
 873                                 mseg->p_memsz += seg->addr & ~PAGE_MASK;
 874                                 mseg->p_memsz += seg->p_memsz;
 875                                 loadmap->nsegs--;
 876                                 continue;
 877                         }
 878                 }
 879
 880                 mseg++;
 881                 if (mseg != seg)
 882                         *mseg = *seg;
 883         }
 884 #endif
 885
 886         kdebug("Mapped Object [%s]:", what);
 887         kdebug("- elfhdr   : %lx", params->elfhdr_addr);
 888         kdebug("- entry    : %lx", params->entry_addr);
 889         kdebug("- PHDR[]   : %lx", params->ph_addr);
 890         kdebug("- DYNAMIC[]: %lx", params->dynamic_addr);
 891         seg = loadmap->segs;
 892         for (loop = 0; loop < loadmap->nsegs; loop++, seg++)
 893                 kdebug("- LOAD[%d] : %08x-%08x [va=%x ms=%x]",
 894                        loop,
 895                        seg->addr, seg->addr + seg->p_memsz - 1,
 896                        seg->p_vaddr, seg->p_memsz);
 897
 898         return 0;
 899
 900 dynamic_error:
 901         printk("ELF FDPIC %s with invalid DYNAMIC section (inode=%lu)\n",
 902                what, file->f_path.dentry->d_inode->i_ino);
 903         return -ELIBBAD;
 904 }
 905
 906 /*****************************************************************************/
 907 /*
 908  * map a file with constant displacement under uClinux
 909  */
 910 #ifndef CONFIG_MMU
 911 static int elf_fdpic_map_file_constdisp_on_uclinux(
 912         struct elf_fdpic_params *params,
 913         struct file *file,
 914         struct mm_struct *mm)
 915 {
 916         struct elf32_fdpic_loadseg *seg;
 917         struct elf32_phdr *phdr;
 918         unsigned long load_addr, base = ULONG_MAX, top = 0, maddr = 0, mflags;
 919         loff_t fpos;
 920         int loop, ret;
 921
 922         load_addr = params->load_addr;
 923         seg = params->loadmap->segs;
 924
 925         /* determine the bounds of the contiguous overall allocation we must
 926          * make */
 927         phdr = params->phdrs;
 928         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 929                 if (params->phdrs[loop].p_type != PT_LOAD)
 930                         continue;
 931
 932                 if (base > phdr->p_vaddr)
 933                         base = phdr->p_vaddr;
 934                 if (top < phdr->p_vaddr + phdr->p_memsz)
 935                         top = phdr->p_vaddr + phdr->p_memsz;
 936         }
 937
 938         /* allocate one big anon block for everything */
 939         mflags = MAP_PRIVATE;
 940         if (params->flags & ELF_FDPIC_FLAG_EXECUTABLE)
 941                 mflags |= MAP_EXECUTABLE;
 942
 943         down_write(&mm->mmap_sem);
 944         maddr = do_mmap(NULL, load_addr, top - base,
 945                         PROT_READ | PROT_WRITE | PROT_EXEC, mflags, 0);
 946         up_write(&mm->mmap_sem);
 947         if (IS_ERR_VALUE(maddr))
 948                 return (int) maddr;
 949
 950         if (load_addr != 0)
 951                 load_addr += PAGE_ALIGN(top - base);
 952
 953         /* and then load the file segments into it */
 954         phdr = params->phdrs;
 955         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 956                 if (params->phdrs[loop].p_type != PT_LOAD)
 957                         continue;
 958
 959                 fpos = phdr->p_offset;
 960
 961                 seg->addr = maddr + (phdr->p_vaddr - base);
 962                 seg->p_vaddr = phdr->p_vaddr;
 963                 seg->p_memsz = phdr->p_memsz;
 964
 965                 ret = file->f_op->read(file, (void *) seg->addr,
 966                                        phdr->p_filesz, &fpos);
 967                 if (ret < 0)
 968                         return ret;
 969
 970                 /* map the ELF header address if in this segment */
 971                 if (phdr->p_offset == 0)
 972                         params->elfhdr_addr = seg->addr;
 973
 974                 /* clear any space allocated but not loaded */
 975                 if (phdr->p_filesz < phdr->p_memsz) {
 976                         ret = clear_user((void *) (seg->addr + phdr->p_filesz),
 977                                          phdr->p_memsz - phdr->p_filesz);
 978                         if (ret)
 979                                 return ret;
 980                 }
 981
 982                 if (mm) {
 983                         if (phdr->p_flags & PF_X) {
 984                                 if (!mm->start_code) {
 985                                         mm->start_code = seg->addr;
 986                                         mm->end_code = seg->addr +
 987                                                 phdr->p_memsz;
 988                                 }
 989                         } else if (!mm->start_data) {
 990                                 mm->start_data = seg->addr;
 991 #ifndef CONFIG_MMU
 992                                 mm->end_data = seg->addr + phdr->p_memsz;
 993 #endif
 994                         }
 995
 996 #ifdef CONFIG_MMU
 997                         if (seg->addr + phdr->p_memsz > mm->end_data)
 998                                 mm->end_data = seg->addr + phdr->p_memsz;
 999 #endif
1000                 }
1001
1002                 seg++;
1003         }
1004
1005         return 0;
1006 }
1007 #endif
1008
1009 /*****************************************************************************/
1010 /*
1011  * map a binary by direct mmap() of the individual PT_LOAD segments
1012  */
1013 static int elf_fdpic_map_file_by_direct_mmap(struct elf_fdpic_params *params,
1014                                              struct file *file,
1015                                              struct mm_struct *mm)
1016 {
1017         struct elf32_fdpic_loadseg *seg;
1018         struct elf32_phdr *phdr;
1019         unsigned long load_addr, delta_vaddr;
1020         int loop, dvset, ret;
1021
1022         load_addr = params->load_addr;
1023         delta_vaddr = 0;
1024         dvset = 0;
1025
1026         seg = params->loadmap->segs;
1027
1028         /* deal with each load segment separately */
1029         phdr = params->phdrs;
1030         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
1031                 unsigned long maddr, disp, excess, excess1;
1032                 int prot = 0, flags;
1033
1034                 if (phdr->p_type != PT_LOAD)
1035                         continue;
1036
1037                 kdebug("[LOAD] va=%lx of=%lx fs=%lx ms=%lx",
1038                        (unsigned long) phdr->p_vaddr,
1039                        (unsigned long) phdr->p_offset,
1040                        (unsigned long) phdr->p_filesz,
1041                        (unsigned long) phdr->p_memsz);
1042
1043                 /* determine the mapping parameters */
1044                 if (phdr->p_flags & PF_R) prot |= PROT_READ;
1045                 if (phdr->p_flags & PF_W) prot |= PROT_WRITE;
1046                 if (phdr->p_flags & PF_X) prot |= PROT_EXEC;
1047
1048                 flags = MAP_PRIVATE | MAP_DENYWRITE;
1049                 if (params->flags & ELF_FDPIC_FLAG_EXECUTABLE)
1050                         flags |= MAP_EXECUTABLE;
1051
1052                 maddr = 0;
1053
1054                 switch (params->flags & ELF_FDPIC_FLAG_ARRANGEMENT) {
1055                 case ELF_FDPIC_FLAG_INDEPENDENT:
1056                         /* PT_LOADs are independently locatable */
1057                         break;
1058
1059                 case ELF_FDPIC_FLAG_HONOURVADDR:
1060                         /* the specified virtual address must be honoured */
1061                         maddr = phdr->p_vaddr;
1062                         flags |= MAP_FIXED;
1063                         break;
1064
1065                 case ELF_FDPIC_FLAG_CONSTDISP:
1066                         /* constant displacement
1067                          * - can be mapped anywhere, but must be mapped as a
1068                          *   unit
1069                          */
1070                         if (!dvset) {
1071                                 maddr = load_addr;
1072                                 delta_vaddr = phdr->p_vaddr;
1073                                 dvset = 1;
1074                         } else {
1075                                 maddr = load_addr + phdr->p_vaddr - delta_vaddr;
1076                                 flags |= MAP_FIXED;
1077                         }
1078                         break;
1079
1080                 case ELF_FDPIC_FLAG_CONTIGUOUS:
1081                         /* contiguity handled later */
1082                         break;
1083
1084                 default:
1085                         BUG();
1086                 }
1087
1088                 maddr &= PAGE_MASK;
1089
1090                 /* create the mapping */
1091                 disp = phdr->p_vaddr & ~PAGE_MASK;
1092                 down_write(&mm->mmap_sem);
1093                 maddr = do_mmap(file, maddr, phdr->p_memsz + disp, prot, flags,
1094                                 phdr->p_offset - disp);
1095                 up_write(&mm->mmap_sem);
1096
1097                 kdebug("mmap[%d] <file> sz=%lx pr=%x fl=%x of=%lx --> %08lx",
1098                        loop, phdr->p_memsz + disp, prot, flags,
1099                        phdr->p_offset - disp, maddr);
1100
1101                 if (IS_ERR_VALUE(maddr))
1102                         return (int) maddr;
1103
1104                 if ((params->flags & ELF_FDPIC_FLAG_ARRANGEMENT) ==
1105                     ELF_FDPIC_FLAG_CONTIGUOUS)
1106                         load_addr += PAGE_ALIGN(phdr->p_memsz + disp);
1107
1108                 seg->addr = maddr + disp;
1109                 seg->p_vaddr = phdr->p_vaddr;
1110                 seg->p_memsz = phdr->p_memsz;
1111
1112                 /* map the ELF header address if in this segment */
1113                 if (phdr->p_offset == 0)
1114                         params->elfhdr_addr = seg->addr;
1115
1116                 /* clear the bit between beginning of mapping and beginning of
1117                  * PT_LOAD */
1118                 if (prot & PROT_WRITE && disp > 0) {
1119                         kdebug("clear[%d] ad=%lx sz=%lx", loop, maddr, disp);
1120                         ret = clear_user((void __user *) maddr, disp);
1121                         if (ret)
1122                                 return ret;
1123                         maddr += disp;
1124                 }
1125
1126                 /* clear any space allocated but not loaded
1127                  * - on uClinux we can just clear the lot
1128                  * - on MMU linux we'll get a SIGBUS beyond the last page
1129                  *   extant in the file
1130                  */
1131                 excess = phdr->p_memsz - phdr->p_filesz;
1132                 excess1 = PAGE_SIZE - ((maddr + phdr->p_filesz) & ~PAGE_MASK);
1133
1134 #ifdef CONFIG_MMU
1135                 if (excess > excess1) {
1136                         unsigned long xaddr = maddr + phdr->p_filesz + excess1;
1137                         unsigned long xmaddr;
1138
1139                         flags |= MAP_FIXED | MAP_ANONYMOUS;
1140                         down_write(&mm->mmap_sem);
1141                         xmaddr = do_mmap(NULL, xaddr, excess - excess1,
1142                                          prot, flags, 0);
1143                         up_write(&mm->mmap_sem);
1144
1145                         kdebug("mmap[%d] <anon>"
1146                                " ad=%lx sz=%lx pr=%x fl=%x of=0 --> %08lx",
1147                                loop, xaddr, excess - excess1, prot, flags,
1148                                xmaddr);
1149
1150                         if (xmaddr != xaddr)
1151                                 return -ENOMEM;
1152                 }
1153
1154                 if (prot & PROT_WRITE && excess1 > 0) {
1155                         kdebug("clear[%d] ad=%lx sz=%lx",
1156                                loop, maddr + phdr->p_filesz, excess1);
1157                         ret = clear_user((void __user *) maddr + phdr->p_filesz,
1158                                          excess1);
1159                         if (ret)
1160                                 return ret;
1161                 }
1162
1163 #else
1164                 if (excess > 0) {
1165                         kdebug("clear[%d] ad=%lx sz=%lx",
1166                                loop, maddr + phdr->p_filesz, excess);
1167                         ret = clear_user((void *) maddr + phdr->p_filesz, excess);
1168                         if (ret)
1169                                 return ret;
1170                 }
1171 #endif
1172
1173                 if (mm) {
1174                         if (phdr->p_flags & PF_X) {
1175                                 if (!mm->start_code) {
1176                                         mm->start_code = maddr;
1177                                         mm->end_code = maddr + phdr->p_memsz;
1178                                 }
1179                         } else if (!mm->start_data) {
1180                                 mm->start_data = maddr;
1181                                 mm->end_data = maddr + phdr->p_memsz;
1182                         }
1183                 }
1184
1185                 seg++;
1186         }
1187
1188         return 0;
1189 }
1190
1191 /*****************************************************************************/
1192 /*
1193  * ELF-FDPIC core dumper
1194  *
1195  * Modelled on fs/exec.c:aout_core_dump()
1196  * Jeremy Fitzhardinge <jeremy@sw.oz.au>
1197  *
1198  * Modelled on fs/binfmt_elf.c core dumper
1199  */
1200 #if defined(USE_ELF_CORE_DUMP) && defined(CONFIG_ELF_CORE)
1201
1202 /*
1203  * These are the only things you should do on a core-file: use only these
1204  * functions to write out all the necessary info.
1205  */
1206 static int dump_write(struct file *file, const void *addr, int nr)
1207 {
1208         return file->f_op->write(file, addr, nr, &file->f_pos) == nr;
1209 }
1210
1211 static int dump_seek(struct file *file, loff_t off)
1212 {
1213         if (file->f_op->llseek) {
1214                 if (file->f_op->llseek(file, off, SEEK_SET) != off)
1215                         return 0;
1216         } else {
1217                 file->f_pos = off;
1218         }
1219         return 1;
1220 }
1221
1222 /*
1223  * Decide whether a segment is worth dumping; default is yes to be
1224  * sure (missing info is worse than too much; etc).
1225  * Personally I'd include everything, and use the coredump limit...
1226  *
1227  * I think we should skip something. But I am not sure how. H.J.
1228  */
1229 static int maydump(struct vm_area_struct *vma, unsigned long mm_flags)
1230 {
1231         int dump_ok;
1232
1233         /* Do not dump I/O mapped devices or special mappings */
1234         if (vma->vm_flags & (VM_IO | VM_RESERVED)) {
1235                 kdcore("%08lx: %08lx: no (IO)", vma->vm_start, vma->vm_flags);
1236                 return 0;
1237         }
1238
1239         /* If we may not read the contents, don't allow us to dump
1240          * them either. "dump_write()" can't handle it anyway.
1241          */
1242         if (!(vma->vm_flags & VM_READ)) {
1243                 kdcore("%08lx: %08lx: no (!read)", vma->vm_start, vma->vm_flags);
1244                 return 0;
1245         }
1246
1247         /* By default, dump shared memory if mapped from an anonymous file. */
1248         if (vma->vm_flags & VM_SHARED) {
1249                 if (vma->vm_file->f_path.dentry->d_inode->i_nlink == 0) {
1250                         dump_ok = test_bit(MMF_DUMP_ANON_SHARED, &mm_flags);
1251                         kdcore("%08lx: %08lx: %s (share)", vma->vm_start,
1252                                vma->vm_flags, dump_ok ? "yes" : "no");
1253                         return dump_ok;
1254                 }
1255
1256                 dump_ok = test_bit(MMF_DUMP_MAPPED_SHARED, &mm_flags);
1257                 kdcore("%08lx: %08lx: %s (share)", vma->vm_start,
1258                        vma->vm_flags, dump_ok ? "yes" : "no");
1259                 return dump_ok;
1260         }
1261
1262 #ifdef CONFIG_MMU
1263         /* By default, if it hasn't been written to, don't write it out */
1264         if (!vma->anon_vma) {
1265                 dump_ok = test_bit(MMF_DUMP_MAPPED_PRIVATE, &mm_flags);
1266                 kdcore("%08lx: %08lx: %s (!anon)", vma->vm_start,
1267                        vma->vm_flags, dump_ok ? "yes" : "no");
1268                 return dump_ok;
1269         }
1270 #endif
1271
1272         dump_ok = test_bit(MMF_DUMP_ANON_PRIVATE, &mm_flags);
1273         kdcore("%08lx: %08lx: %s", vma->vm_start, vma->vm_flags,
1274                dump_ok ? "yes" : "no");
1275         return dump_ok;
1276 }
1277
1278 /* An ELF note in memory */
1279 struct memelfnote
1280 {
1281         const char *name;
1282         int type;
1283         unsigned int datasz;
1284         void *data;
1285 };
1286
1287 static int notesize(struct memelfnote *en)
1288 {
1289         int sz;
1290
1291         sz = sizeof(struct elf_note);
1292         sz += roundup(strlen(en->name) + 1, 4);
1293         sz += roundup(en->datasz, 4);
1294
1295         return sz;
1296 }
1297
1298 /* #define DEBUG */
1299
1300 #define DUMP_WRITE(addr, nr)    \
1301         do { if (!dump_write(file, (addr), (nr))) return 0; } while(0)
1302 #define DUMP_SEEK(off)  \
1303         do { if (!dump_seek(file, (off))) return 0; } while(0)
1304
1305 static int writenote(struct memelfnote *men, struct file *file)
1306 {
1307         struct elf_note en;
1308
1309         en.n_namesz = strlen(men->name) + 1;
1310         en.n_descsz = men->datasz;
1311         en.n_type = men->type;
1312
1313         DUMP_WRITE(&en, sizeof(en));
1314         DUMP_WRITE(men->name, en.n_namesz);
1315         /* XXX - cast from long long to long to avoid need for libgcc.a */
1316         DUMP_SEEK(roundup((unsigned long)file->f_pos, 4));      /* XXX */
1317         DUMP_WRITE(men->data, men->datasz);
1318         DUMP_SEEK(roundup((unsigned long)file->f_pos, 4));      /* XXX */
1319
1320         return 1;
1321 }
1322 #undef DUMP_WRITE
1323 #undef DUMP_SEEK
1324
1325 #define DUMP_WRITE(addr, nr)    \
1326         if ((size += (nr)) > limit || !dump_write(file, (addr), (nr))) \
1327                 goto end_coredump;
1328 #define DUMP_SEEK(off)  \
1329         if (!dump_seek(file, (off))) \
1330                 goto end_coredump;
1331
1332 static inline void fill_elf_fdpic_header(struct elfhdr *elf, int segs)
1333 {
1334         memcpy(elf->e_ident, ELFMAG, SELFMAG);
1335         elf->e_ident[EI_CLASS] = ELF_CLASS;
1336         elf->e_ident[EI_DATA] = ELF_DATA;
1337         elf->e_ident[EI_VERSION] = EV_CURRENT;
1338         elf->e_ident[EI_OSABI] = ELF_OSABI;
1339         memset(elf->e_ident+EI_PAD, 0, EI_NIDENT-EI_PAD);
1340
1341         elf->e_type = ET_CORE;
1342         elf->e_machine = ELF_ARCH;
1343         elf->e_version = EV_CURRENT;
1344         elf->e_entry = 0;
1345         elf->e_phoff = sizeof(struct elfhdr);
1346         elf->e_shoff = 0;
1347         elf->e_flags = ELF_FDPIC_CORE_EFLAGS;
1348         elf->e_ehsize = sizeof(struct elfhdr);
1349         elf->e_phentsize = sizeof(struct elf_phdr);
1350         elf->e_phnum = segs;
1351         elf->e_shentsize = 0;
1352         elf->e_shnum = 0;
1353         elf->e_shstrndx = 0;
1354         return;
1355 }
1356
1357 static inline void fill_elf_note_phdr(struct elf_phdr *phdr, int sz, loff_t offset)
1358 {
1359         phdr->p_type = PT_NOTE;
1360         phdr->p_offset = offset;
1361         phdr->p_vaddr = 0;
1362         phdr->p_paddr = 0;
1363         phdr->p_filesz = sz;
1364         phdr->p_memsz = 0;
1365         phdr->p_flags = 0;
1366         phdr->p_align = 0;
1367         return;
1368 }
1369
1370 static inline void fill_note(struct memelfnote *note, const char *name, int type,
1371                 unsigned int sz, void *data)
1372 {
1373         note->name = name;
1374         note->type = type;
1375         note->datasz = sz;
1376         note->data = data;
1377         return;
1378 }
1379
1380 /*
1381  * fill up all the fields in prstatus from the given task struct, except
1382  * registers which need to be filled up seperately.
1383  */
1384 static void fill_prstatus(struct elf_prstatus *prstatus,
1385                           struct task_struct *p, long signr)
1386 {
1387         prstatus->pr_info.si_signo = prstatus->pr_cursig = signr;
1388         prstatus->pr_sigpend = p->pending.signal.sig[0];
1389         prstatus->pr_sighold = p->blocked.sig[0];
1390         prstatus->pr_pid = task_pid_vnr(p);
1391         prstatus->pr_ppid = task_pid_vnr(p->parent);
1392         prstatus->pr_pgrp = task_pgrp_vnr(p);
1393         prstatus->pr_sid = task_session_vnr(p);
1394         if (thread_group_leader(p)) {
1395                 struct task_cputime cputime;
1396
1397                 /*
1398                  * This is the record for the group leader.  It shows the
1399                  * group-wide total, not its individual thread total.
1400                  */
1401                 thread_group_cputime(p, &cputime);
1402                 cputime_to_timeval(cputime.utime, &prstatus->pr_utime);
1403                 cputime_to_timeval(cputime.stime, &prstatus->pr_stime);
1404         } else {
1405                 cputime_to_timeval(p->utime, &prstatus->pr_utime);
1406                 cputime_to_timeval(p->stime, &prstatus->pr_stime);
1407         }
1408         cputime_to_timeval(p->signal->cutime, &prstatus->pr_cutime);
1409         cputime_to_timeval(p->signal->cstime, &prstatus->pr_cstime);
1410
1411         prstatus->pr_exec_fdpic_loadmap = p->mm->context.exec_fdpic_loadmap;
1412         prstatus->pr_interp_fdpic_loadmap = p->mm->context.interp_fdpic_loadmap;
1413 }
1414
1415 static int fill_psinfo(struct elf_prpsinfo *psinfo, struct task_struct *p,
1416                        struct mm_struct *mm)
1417 {
1418         const struct cred *cred;
1419         unsigned int i, len;
1420
1421         /* first copy the parameters from user space */
1422         memset(psinfo, 0, sizeof(struct elf_prpsinfo));
1423
1424         len = mm->arg_end - mm->arg_start;
1425         if (len >= ELF_PRARGSZ)
1426                 len = ELF_PRARGSZ - 1;
1427         if (copy_from_user(&psinfo->pr_psargs,
1428                            (const char __user *) mm->arg_start, len))
1429                 return -EFAULT;
1430         for (i = 0; i < len; i++)
1431                 if (psinfo->pr_psargs[i] == 0)
1432                         psinfo->pr_psargs[i] = ' ';
1433         psinfo->pr_psargs[len] = 0;
1434
1435         psinfo->pr_pid = task_pid_vnr(p);
1436         psinfo->pr_ppid = task_pid_vnr(p->parent);
1437         psinfo->pr_pgrp = task_pgrp_vnr(p);
1438         psinfo->pr_sid = task_session_vnr(p);
1439
1440         i = p->state ? ffz(~p->state) + 1 : 0;
1441         psinfo->pr_state = i;
1442         psinfo->pr_sname = (i > 5) ? '.' : "RSDTZW"[i];
1443         psinfo->pr_zomb = psinfo->pr_sname == 'Z';
1444         psinfo->pr_nice = task_nice(p);
1445         psinfo->pr_flag = p->flags;
1446         rcu_read_lock();
1447         cred = __task_cred(p);
1448         SET_UID(psinfo->pr_uid, cred->uid);
1449         SET_GID(psinfo->pr_gid, cred->gid);
1450         rcu_read_unlock();
1451         strncpy(psinfo->pr_fname, p->comm, sizeof(psinfo->pr_fname));
1452
1453         return 0;
1454 }
1455
1456 /* Here is the structure in which status of each thread is captured. */
1457 struct elf_thread_status
1458 {
1459         struct list_head list;
1460         struct elf_prstatus prstatus;   /* NT_PRSTATUS */
1461         elf_fpregset_t fpu;             /* NT_PRFPREG */
1462         struct task_struct *thread;
1463 #ifdef ELF_CORE_COPY_XFPREGS
1464         elf_fpxregset_t xfpu;           /* ELF_CORE_XFPREG_TYPE */
1465 #endif
1466         struct memelfnote notes[3];
1467         int num_notes;
1468 };
1469
1470 /*
1471  * In order to add the specific thread information for the elf file format,
1472  * we need to keep a linked list of every thread's pr_status and then create
1473  * a single section for them in the final core file.
1474  */
1475 static int elf_dump_thread_status(long signr, struct elf_thread_status *t)
1476 {
1477         struct task_struct *p = t->thread;
1478         int sz = 0;
1479
1480         t->num_notes = 0;
1481
1482         fill_prstatus(&t->prstatus, p, signr);
1483         elf_core_copy_task_regs(p, &t->prstatus.pr_reg);
1484
1485         fill_note(&t->notes[0], "CORE", NT_PRSTATUS, sizeof(t->prstatus),
1486                   &t->prstatus);
1487         t->num_notes++;
1488         sz += notesize(&t->notes[0]);
1489
1490         t->prstatus.pr_fpvalid = elf_core_copy_task_fpregs(p, NULL, &t->fpu);
1491         if (t->prstatus.pr_fpvalid) {
1492                 fill_note(&t->notes[1], "CORE", NT_PRFPREG, sizeof(t->fpu),
1493                           &t->fpu);
1494                 t->num_notes++;
1495                 sz += notesize(&t->notes[1]);
1496         }
1497
1498 #ifdef ELF_CORE_COPY_XFPREGS
1499         if (elf_core_copy_task_xfpregs(p, &t->xfpu)) {
1500                 fill_note(&t->notes[2], "LINUX", ELF_CORE_XFPREG_TYPE,
1501                           sizeof(t->xfpu), &t->xfpu);
1502                 t->num_notes++;
1503                 sz += notesize(&t->notes[2]);
1504         }
1505 #endif
1506         return sz;
1507 }
1508
1509 /*
1510  * dump the segments for an MMU process
1511  */
1512 #ifdef CONFIG_MMU
1513 static int elf_fdpic_dump_segments(struct file *file, size_t *size,
1514                            unsigned long *limit, unsigned long mm_flags)
1515 {
1516         struct vm_area_struct *vma;
1517
1518         for (vma = current->mm->mmap; vma; vma = vma->vm_next) {
1519                 unsigned long addr;
1520
1521                 if (!maydump(vma, mm_flags))
1522                         continue;
1523
1524                 for (addr = vma->vm_start;
1525                      addr < vma->vm_end;
1526                      addr += PAGE_SIZE
1527                      ) {
1528                         struct vm_area_struct *vma;
1529                         struct page *page;
1530
1531                         if (get_user_pages(current, current->mm, addr, 1, 0, 1,
1532                                            &page, &vma) <= 0) {
1533                                 DUMP_SEEK(file->f_pos + PAGE_SIZE);
1534                         }
1535                         else if (page == ZERO_PAGE(0)) {
1536                                 page_cache_release(page);
1537                                 DUMP_SEEK(file->f_pos + PAGE_SIZE);
1538                         }
1539                         else {
1540                                 void *kaddr;
1541
1542                                 flush_cache_page(vma, addr, page_to_pfn(page));
1543                                 kaddr = kmap(page);
1544                                 if ((*size += PAGE_SIZE) > *limit ||
1545                                     !dump_write(file, kaddr, PAGE_SIZE)
1546                                     ) {
1547                                         kunmap(page);
1548                                         page_cache_release(page);
1549                                         return -EIO;
1550                                 }
1551                                 kunmap(page);
1552                                 page_cache_release(page);
1553                         }
1554                 }
1555         }
1556
1557         return 0;
1558
1559 end_coredump:
1560         return -EFBIG;
1561 }
1562 #endif
1563
1564 /*
1565  * dump the segments for a NOMMU process
1566  */
1567 #ifndef CONFIG_MMU
1568 static int elf_fdpic_dump_segments(struct file *file, size_t *size,
1569                            unsigned long *limit, unsigned long mm_flags)
1570 {
1571         struct vm_area_struct *vma;
1572
1573         for (vma = current->mm->mmap; vma; vma = vma->vm_next) {
1574                 if (!maydump(vma, mm_flags))
1575                         continue;
1576
1577                 if ((*size += PAGE_SIZE) > *limit)
1578                         return -EFBIG;
1579
1580                 if (!dump_write(file, (void *) vma->vm_start,
1581                                 vma->vm_end - vma->vm_start))
1582                         return -EIO;
1583         }
1584
1585         return 0;
1586 }
1587 #endif
1588
1589 /*
1590  * Actual dumper
1591  *
1592  * This is a two-pass process; first we find the offsets of the bits,
1593  * and then they are actually written out.  If we run out of core limit
1594  * we just truncate.
1595  */
1596 static int elf_fdpic_core_dump(long signr, struct pt_regs *regs,
1597                                struct file *file, unsigned long limit)
1598 {
1599 #define NUM_NOTES       6
1600         int has_dumped = 0;
1601         mm_segment_t fs;
1602         int segs;
1603         size_t size = 0;
1604         int i;
1605         struct vm_area_struct *vma;
1606         struct elfhdr *elf = NULL;
1607         loff_t offset = 0, dataoff;
1608         int numnote;
1609         struct memelfnote *notes = NULL;
1610         struct elf_prstatus *prstatus = NULL;   /* NT_PRSTATUS */
1611         struct elf_prpsinfo *psinfo = NULL;     /* NT_PRPSINFO */
1612         LIST_HEAD(thread_list);
1613         struct list_head *t;
1614         elf_fpregset_t *fpu = NULL;
1615 #ifdef ELF_CORE_COPY_XFPREGS
1616         elf_fpxregset_t *xfpu = NULL;
1617 #endif
1618         int thread_status_size = 0;
1619         elf_addr_t *auxv;
1620         unsigned long mm_flags;
1621
1622         /*
1623          * We no longer stop all VM operations.
1624          *
1625          * This is because those proceses that could possibly change map_count
1626          * or the mmap / vma pages are now blocked in do_exit on current
1627          * finishing this core dump.
1628          *
1629          * Only ptrace can touch these memory addresses, but it doesn't change
1630          * the map_count or the pages allocated. So no possibility of crashing
1631          * exists while dumping the mm->vm_next areas to the core file.
1632          */
1633
1634         /* alloc memory for large data structures: too large to be on stack */
1635         elf = kmalloc(sizeof(*elf), GFP_KERNEL);
1636         if (!elf)
1637                 goto cleanup;
1638         prstatus = kzalloc(sizeof(*prstatus), GFP_KERNEL);
1639         if (!prstatus)
1640                 goto cleanup;
1641         psinfo = kmalloc(sizeof(*psinfo), GFP_KERNEL);
1642         if (!psinfo)
1643                 goto cleanup;
1644         notes = kmalloc(NUM_NOTES * sizeof(struct memelfnote), GFP_KERNEL);
1645         if (!notes)
1646                 goto cleanup;
1647         fpu = kmalloc(sizeof(*fpu), GFP_KERNEL);
1648         if (!fpu)
1649                 goto cleanup;
1650 #ifdef ELF_CORE_COPY_XFPREGS
1651         xfpu = kmalloc(sizeof(*xfpu), GFP_KERNEL);
1652         if (!xfpu)
1653                 goto cleanup;
1654 #endif
1655
1656         if (signr) {
1657                 struct core_thread *ct;
1658                 struct elf_thread_status *tmp;
1659
1660                 for (ct = current->mm->core_state->dumper.next;
1661                                                 ct; ct = ct->next) {
1662                         tmp = kzalloc(sizeof(*tmp), GFP_KERNEL);
1663                         if (!tmp)
1664                                 goto cleanup;
1665
1666                         tmp->thread = ct->task;
1667                         list_add(&tmp->list, &thread_list);
1668                 }
1669
1670                 list_for_each(t, &thread_list) {
1671                         struct elf_thread_status *tmp;
1672                         int sz;
1673
1674                         tmp = list_entry(t, struct elf_thread_status, list);
1675                         sz = elf_dump_thread_status(signr, tmp);
1676                         thread_status_size += sz;
1677                 }
1678         }
1679
1680         /* now collect the dump for the current */
1681         fill_prstatus(prstatus, current, signr);
1682         elf_core_copy_regs(&prstatus->pr_reg, regs);
1683
1684         segs = current->mm->map_count;
1685 #ifdef ELF_CORE_EXTRA_PHDRS
1686         segs += ELF_CORE_EXTRA_PHDRS;
1687 #endif
1688
1689         /* Set up header */
1690         fill_elf_fdpic_header(elf, segs + 1);   /* including notes section */
1691
1692         has_dumped = 1;
1693         current->flags |= PF_DUMPCORE;
1694
1695         /*
1696          * Set up the notes in similar form to SVR4 core dumps made
1697          * with info from their /proc.
1698          */
1699
1700         fill_note(notes + 0, "CORE", NT_PRSTATUS, sizeof(*prstatus), prstatus);
1701         fill_psinfo(psinfo, current->group_leader, current->mm);
1702         fill_note(notes + 1, "CORE", NT_PRPSINFO, sizeof(*psinfo), psinfo);
1703
1704         numnote = 2;
1705
1706         auxv = (elf_addr_t *) current->mm->saved_auxv;
1707
1708         i = 0;
1709         do
1710                 i += 2;
1711         while (auxv[i - 2] != AT_NULL);
1712         fill_note(&notes[numnote++], "CORE", NT_AUXV,
1713                   i * sizeof(elf_addr_t), auxv);
1714
1715         /* Try to dump the FPU. */
1716         if ((prstatus->pr_fpvalid =
1717              elf_core_copy_task_fpregs(current, regs, fpu)))
1718                 fill_note(notes + numnote++,
1719                           "CORE", NT_PRFPREG, sizeof(*fpu), fpu);
1720 #ifdef ELF_CORE_COPY_XFPREGS
1721         if (elf_core_copy_task_xfpregs(current, xfpu))
1722                 fill_note(notes + numnote++,
1723                           "LINUX", ELF_CORE_XFPREG_TYPE, sizeof(*xfpu), xfpu);
1724 #endif
1725
1726         fs = get_fs();
1727         set_fs(KERNEL_DS);
1728
1729         DUMP_WRITE(elf, sizeof(*elf));
1730         offset += sizeof(*elf);                         /* Elf header */
1731         offset += (segs+1) * sizeof(struct elf_phdr);   /* Program headers */
1732
1733         /* Write notes phdr entry */
1734         {
1735                 struct elf_phdr phdr;
1736                 int sz = 0;
1737
1738                 for (i = 0; i < numnote; i++)
1739                         sz += notesize(notes + i);
1740
1741                 sz += thread_status_size;
1742
1743                 fill_elf_note_phdr(&phdr, sz, offset);
1744                 offset += sz;
1745                 DUMP_WRITE(&phdr, sizeof(phdr));
1746         }
1747
1748         /* Page-align dumped data */
1749         dataoff = offset = roundup(offset, ELF_EXEC_PAGESIZE);
1750
1751         /*
1752          * We must use the same mm->flags while dumping core to avoid
1753          * inconsistency between the program headers and bodies, otherwise an
1754          * unusable core file can be generated.
1755          */
1756         mm_flags = current->mm->flags;
1757
1758         /* write program headers for segments dump */
1759         for (vma = current->mm->mmap; vma; vma = vma->vm_next) {
1760                 struct elf_phdr phdr;
1761                 size_t sz;
1762
1763                 sz = vma->vm_end - vma->vm_start;
1764
1765                 phdr.p_type = PT_LOAD;
1766                 phdr.p_offset = offset;
1767                 phdr.p_vaddr = vma->vm_start;
1768                 phdr.p_paddr = 0;
1769                 phdr.p_filesz = maydump(vma, mm_flags) ? sz : 0;
1770                 phdr.p_memsz = sz;
1771                 offset += phdr.p_filesz;
1772                 phdr.p_flags = vma->vm_flags & VM_READ ? PF_R : 0;
1773                 if (vma->vm_flags & VM_WRITE)
1774                         phdr.p_flags |= PF_W;
1775                 if (vma->vm_flags & VM_EXEC)
1776                         phdr.p_flags |= PF_X;
1777                 phdr.p_align = ELF_EXEC_PAGESIZE;
1778
1779                 DUMP_WRITE(&phdr, sizeof(phdr));
1780         }
1781
1782 #ifdef ELF_CORE_WRITE_EXTRA_PHDRS
1783         ELF_CORE_WRITE_EXTRA_PHDRS;
1784 #endif
1785
1786         /* write out the notes section */
1787         for (i = 0; i < numnote; i++)
1788                 if (!writenote(notes + i, file))
1789                         goto end_coredump;
1790
1791         /* write out the thread status notes section */
1792         list_for_each(t, &thread_list) {
1793                 struct elf_thread_status *tmp =
1794                                 list_entry(t, struct elf_thread_status, list);
1795
1796                 for (i = 0; i < tmp->num_notes; i++)
1797                         if (!writenote(&tmp->notes[i], file))
1798                                 goto end_coredump;
1799         }
1800
1801         DUMP_SEEK(dataoff);
1802
1803         if (elf_fdpic_dump_segments(file, &size, &limit, mm_flags) < 0)
1804                 goto end_coredump;
1805
1806 #ifdef ELF_CORE_WRITE_EXTRA_DATA
1807         ELF_CORE_WRITE_EXTRA_DATA;
1808 #endif
1809
1810         if (file->f_pos != offset) {
1811                 /* Sanity check */
1812                 printk(KERN_WARNING
1813                        "elf_core_dump: file->f_pos (%lld) != offset (%lld)\n",
1814                        file->f_pos, offset);
1815         }
1816
1817 end_coredump:
1818         set_fs(fs);
1819
1820 cleanup:
1821         while (!list_empty(&thread_list)) {
1822                 struct list_head *tmp = thread_list.next;
1823                 list_del(tmp);
1824                 kfree(list_entry(tmp, struct elf_thread_status, list));
1825         }
1826
1827         kfree(elf);
1828         kfree(prstatus);
1829         kfree(psinfo);
1830         kfree(notes);
1831         kfree(fpu);
1832 #ifdef ELF_CORE_COPY_XFPREGS
1833         kfree(xfpu);
1834 #endif
1835         return has_dumped;
1836 #undef NUM_NOTES
1837 }
1838
1839 #endif          /* USE_ELF_CORE_DUMP */