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_cvirt.h>
  38
  39 #include <asm/uaccess.h>
  40 #include <asm/param.h>
  41 #include <asm/pgalloc.h>
  42
  43 typedef char *elf_caddr_t;
  44 #ifndef elf_addr_t
  45 #define elf_addr_t unsigned long
  46 #endif
  47
  48 #if 0
  49 #define kdebug(fmt, ...) printk("FDPIC "fmt"\n" ,##__VA_ARGS__ )
  50 #else
  51 #define kdebug(fmt, ...) do {} while(0)
  52 #endif
  53
  54 #if 0
  55 #define kdcore(fmt, ...) printk("FDPIC "fmt"\n" ,##__VA_ARGS__ )
  56 #else
  57 #define kdcore(fmt, ...) do {} while(0)
  58 #endif
  59
  60 MODULE_LICENSE("GPL");
  61
  62 static int load_elf_fdpic_binary(struct linux_binprm *, struct pt_regs *);
  63 static int elf_fdpic_fetch_phdrs(struct elf_fdpic_params *, struct file *);
  64 static int elf_fdpic_map_file(struct elf_fdpic_params *, struct file *,
  65                               struct mm_struct *, const char *);
  66
  67 static int create_elf_fdpic_tables(struct linux_binprm *, struct mm_struct *,
  68                                    struct elf_fdpic_params *,
  69                                    struct elf_fdpic_params *);
  70
  71 #ifndef CONFIG_MMU
  72 static int elf_fdpic_transfer_args_to_stack(struct linux_binprm *,
  73                                             unsigned long *);
  74 static int elf_fdpic_map_file_constdisp_on_uclinux(struct elf_fdpic_params *,
  75                                                    struct file *,
  76                                                    struct mm_struct *);
  77 #endif
  78
  79 static int elf_fdpic_map_file_by_direct_mmap(struct elf_fdpic_params *,
  80                                              struct file *, struct mm_struct *);
  81
  82 #if defined(USE_ELF_CORE_DUMP) && defined(CONFIG_ELF_CORE)
  83 static int elf_fdpic_core_dump(long, struct pt_regs *, struct file *);
  84 #endif
  85
  86 static struct linux_binfmt elf_fdpic_format = {
  87         .module         = THIS_MODULE,
  88         .load_binary    = load_elf_fdpic_binary,
  89 #if defined(USE_ELF_CORE_DUMP) && defined(CONFIG_ELF_CORE)
  90         .core_dump      = elf_fdpic_core_dump,
  91 #endif
  92         .min_coredump   = ELF_EXEC_PAGESIZE,
  93 };
  94
  95 static int __init init_elf_fdpic_binfmt(void)
  96 {
  97         return register_binfmt(&elf_fdpic_format);
  98 }
  99
 100 static void __exit exit_elf_fdpic_binfmt(void)
 101 {
 102         unregister_binfmt(&elf_fdpic_format);
 103 }
 104
 105 core_initcall(init_elf_fdpic_binfmt);
 106 module_exit(exit_elf_fdpic_binfmt);
 107
 108 static int is_elf_fdpic(struct elfhdr *hdr, struct file *file)
 109 {
 110         if (memcmp(hdr->e_ident, ELFMAG, SELFMAG) != 0)
 111                 return 0;
 112         if (hdr->e_type != ET_EXEC && hdr->e_type != ET_DYN)
 113                 return 0;
 114         if (!elf_check_arch(hdr) || !elf_check_fdpic(hdr))
 115                 return 0;
 116         if (!file->f_op || !file->f_op->mmap)
 117                 return 0;
 118         return 1;
 119 }
 120
 121 /*****************************************************************************/
 122 /*
 123  * read the program headers table into memory
 124  */
 125 static int elf_fdpic_fetch_phdrs(struct elf_fdpic_params *params,
 126                                  struct file *file)
 127 {
 128         struct elf32_phdr *phdr;
 129         unsigned long size;
 130         int retval, loop;
 131
 132         if (params->hdr.e_phentsize != sizeof(struct elf_phdr))
 133                 return -ENOMEM;
 134         if (params->hdr.e_phnum > 65536U / sizeof(struct elf_phdr))
 135                 return -ENOMEM;
 136
 137         size = params->hdr.e_phnum * sizeof(struct elf_phdr);
 138         params->phdrs = kmalloc(size, GFP_KERNEL);
 139         if (!params->phdrs)
 140                 return -ENOMEM;
 141
 142         retval = kernel_read(file, params->hdr.e_phoff,
 143                              (char *) params->phdrs, size);
 144         if (retval < 0)
 145                 return retval;
 146
 147         /* determine stack size for this binary */
 148         phdr = params->phdrs;
 149         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 150                 if (phdr->p_type != PT_GNU_STACK)
 151                         continue;
 152
 153                 if (phdr->p_flags & PF_X)
 154                         params->flags |= ELF_FDPIC_FLAG_EXEC_STACK;
 155                 else
 156                         params->flags |= ELF_FDPIC_FLAG_NOEXEC_STACK;
 157
 158                 params->stack_size = phdr->p_memsz;
 159                 break;
 160         }
 161
 162         return 0;
 163 }
 164
 165 /*****************************************************************************/
 166 /*
 167  * load an fdpic binary into various bits of memory
 168  */
 169 static int load_elf_fdpic_binary(struct linux_binprm *bprm,
 170                                  struct pt_regs *regs)
 171 {
 172         struct elf_fdpic_params exec_params, interp_params;
 173         struct elf_phdr *phdr;
 174         unsigned long stack_size, entryaddr;
 175 #ifndef CONFIG_MMU
 176         unsigned long fullsize;
 177 #endif
 178 #ifdef ELF_FDPIC_PLAT_INIT
 179         unsigned long dynaddr;
 180 #endif
 181         struct file *interpreter = NULL; /* to shut gcc up */
 182         char *interpreter_name = NULL;
 183         int executable_stack;
 184         int retval, i;
 185
 186         memset(&exec_params, 0, sizeof(exec_params));
 187         memset(&interp_params, 0, sizeof(interp_params));
 188
 189         exec_params.hdr = *(struct elfhdr *) bprm->buf;
 190         exec_params.flags = ELF_FDPIC_FLAG_PRESENT | ELF_FDPIC_FLAG_EXECUTABLE;
 191
 192         /* check that this is a binary we know how to deal with */
 193         retval = -ENOEXEC;
 194         if (!is_elf_fdpic(&exec_params.hdr, bprm->file))
 195                 goto error;
 196
 197         /* read the program header table */
 198         retval = elf_fdpic_fetch_phdrs(&exec_params, bprm->file);
 199         if (retval < 0)
 200                 goto error;
 201
 202         /* scan for a program header that specifies an interpreter */
 203         phdr = exec_params.phdrs;
 204
 205         for (i = 0; i < exec_params.hdr.e_phnum; i++, phdr++) {
 206                 switch (phdr->p_type) {
 207                 case PT_INTERP:
 208                         retval = -ENOMEM;
 209                         if (phdr->p_filesz > PATH_MAX)
 210                                 goto error;
 211                         retval = -ENOENT;
 212                         if (phdr->p_filesz < 2)
 213                                 goto error;
 214
 215                         /* read the name of the interpreter into memory */
 216                         interpreter_name = kmalloc(phdr->p_filesz, GFP_KERNEL);
 217                         if (!interpreter_name)
 218                                 goto error;
 219
 220                         retval = kernel_read(bprm->file,
 221                                              phdr->p_offset,
 222                                              interpreter_name,
 223                                              phdr->p_filesz);
 224                         if (retval < 0)
 225                                 goto error;
 226
 227                         retval = -ENOENT;
 228                         if (interpreter_name[phdr->p_filesz - 1] != '\0')
 229                                 goto error;
 230
 231                         kdebug("Using ELF interpreter %s", interpreter_name);
 232
 233                         /* replace the program with the interpreter */
 234                         interpreter = open_exec(interpreter_name);
 235                         retval = PTR_ERR(interpreter);
 236                         if (IS_ERR(interpreter)) {
 237                                 interpreter = NULL;
 238                                 goto error;
 239                         }
 240
 241                         retval = kernel_read(interpreter, 0, bprm->buf,
 242                                              BINPRM_BUF_SIZE);
 243                         if (retval < 0)
 244                                 goto error;
 245
 246                         interp_params.hdr = *((struct elfhdr *) bprm->buf);
 247                         break;
 248
 249                 case PT_LOAD:
 250 #ifdef CONFIG_MMU
 251                         if (exec_params.load_addr == 0)
 252                                 exec_params.load_addr = phdr->p_vaddr;
 253 #endif
 254                         break;
 255                 }
 256
 257         }
 258
 259         if (elf_check_const_displacement(&exec_params.hdr))
 260                 exec_params.flags |= ELF_FDPIC_FLAG_CONSTDISP;
 261
 262         /* perform insanity checks on the interpreter */
 263         if (interpreter_name) {
 264                 retval = -ELIBBAD;
 265                 if (!is_elf_fdpic(&interp_params.hdr, interpreter))
 266                         goto error;
 267
 268                 interp_params.flags = ELF_FDPIC_FLAG_PRESENT;
 269
 270                 /* read the interpreter's program header table */
 271                 retval = elf_fdpic_fetch_phdrs(&interp_params, interpreter);
 272                 if (retval < 0)
 273                         goto error;
 274         }
 275
 276         stack_size = exec_params.stack_size;
 277         if (stack_size < interp_params.stack_size)
 278                 stack_size = interp_params.stack_size;
 279
 280         if (exec_params.flags & ELF_FDPIC_FLAG_EXEC_STACK)
 281                 executable_stack = EXSTACK_ENABLE_X;
 282         else if (exec_params.flags & ELF_FDPIC_FLAG_NOEXEC_STACK)
 283                 executable_stack = EXSTACK_DISABLE_X;
 284         else if (interp_params.flags & ELF_FDPIC_FLAG_EXEC_STACK)
 285                 executable_stack = EXSTACK_ENABLE_X;
 286         else if (interp_params.flags & ELF_FDPIC_FLAG_NOEXEC_STACK)
 287                 executable_stack = EXSTACK_DISABLE_X;
 288         else
 289                 executable_stack = EXSTACK_DEFAULT;
 290
 291         retval = -ENOEXEC;
 292         if (stack_size == 0)
 293                 goto error;
 294
 295         if (elf_check_const_displacement(&interp_params.hdr))
 296                 interp_params.flags |= ELF_FDPIC_FLAG_CONSTDISP;
 297
 298         /* flush all traces of the currently running executable */
 299         retval = flush_old_exec(bprm);
 300         if (retval)
 301                 goto error;
 302
 303         /* there's now no turning back... the old userspace image is dead,
 304          * defunct, deceased, etc. after this point we have to exit via
 305          * error_kill */
 306         set_personality(PER_LINUX_FDPIC);
 307         set_binfmt(&elf_fdpic_format);
 308
 309         current->mm->start_code = 0;
 310         current->mm->end_code = 0;
 311         current->mm->start_stack = 0;
 312         current->mm->start_data = 0;
 313         current->mm->end_data = 0;
 314         current->mm->context.exec_fdpic_loadmap = 0;
 315         current->mm->context.interp_fdpic_loadmap = 0;
 316
 317         current->flags &= ~PF_FORKNOEXEC;
 318
 319 #ifdef CONFIG_MMU
 320         elf_fdpic_arch_lay_out_mm(&exec_params,
 321                                   &interp_params,
 322                                   &current->mm->start_stack,
 323                                   &current->mm->start_brk);
 324
 325         retval = setup_arg_pages(bprm, current->mm->start_stack,
 326                                  executable_stack);
 327         if (retval < 0) {
 328                 send_sig(SIGKILL, current, 0);
 329                 goto error_kill;
 330         }
 331 #endif
 332
 333         /* load the executable and interpreter into memory */
 334         retval = elf_fdpic_map_file(&exec_params, bprm->file, current->mm,
 335                                     "executable");
 336         if (retval < 0)
 337                 goto error_kill;
 338
 339         if (interpreter_name) {
 340                 retval = elf_fdpic_map_file(&interp_params, interpreter,
 341                                             current->mm, "interpreter");
 342                 if (retval < 0) {
 343                         printk(KERN_ERR "Unable to load interpreter\n");
 344                         goto error_kill;
 345                 }
 346
 347                 allow_write_access(interpreter);
 348                 fput(interpreter);
 349                 interpreter = NULL;
 350         }
 351
 352 #ifdef CONFIG_MMU
 353         if (!current->mm->start_brk)
 354                 current->mm->start_brk = current->mm->end_data;
 355
 356         current->mm->brk = current->mm->start_brk =
 357                 PAGE_ALIGN(current->mm->start_brk);
 358
 359 #else
 360         /* create a stack and brk area big enough for everyone
 361          * - the brk heap starts at the bottom and works up
 362          * - the stack starts at the top and works down
 363          */
 364         stack_size = (stack_size + PAGE_SIZE - 1) & PAGE_MASK;
 365         if (stack_size < PAGE_SIZE * 2)
 366                 stack_size = PAGE_SIZE * 2;
 367
 368         down_write(&current->mm->mmap_sem);
 369         current->mm->start_brk = do_mmap(NULL, 0, stack_size,
 370                                          PROT_READ | PROT_WRITE | PROT_EXEC,
 371                                          MAP_PRIVATE | MAP_ANON | MAP_GROWSDOWN,
 372                                          0);
 373
 374         if (IS_ERR_VALUE(current->mm->start_brk)) {
 375                 up_write(&current->mm->mmap_sem);
 376                 retval = current->mm->start_brk;
 377                 current->mm->start_brk = 0;
 378                 goto error_kill;
 379         }
 380
 381         /* expand the stack mapping to use up the entire allocation granule */
 382         fullsize = ksize((char *) current->mm->start_brk);
 383         if (!IS_ERR_VALUE(do_mremap(current->mm->start_brk, stack_size,
 384                                     fullsize, 0, 0)))
 385                 stack_size = fullsize;
 386         up_write(&current->mm->mmap_sem);
 387
 388         current->mm->brk = current->mm->start_brk;
 389         current->mm->context.end_brk = current->mm->start_brk;
 390         current->mm->context.end_brk +=
 391                 (stack_size > PAGE_SIZE) ? (stack_size - PAGE_SIZE) : 0;
 392         current->mm->start_stack = current->mm->start_brk + stack_size;
 393 #endif
 394
 395         compute_creds(bprm);
 396         current->flags &= ~PF_FORKNOEXEC;
 397         if (create_elf_fdpic_tables(bprm, current->mm,
 398                                     &exec_params, &interp_params) < 0)
 399                 goto error_kill;
 400
 401         kdebug("- start_code  %lx", current->mm->start_code);
 402         kdebug("- end_code    %lx", current->mm->end_code);
 403         kdebug("- start_data  %lx", current->mm->start_data);
 404         kdebug("- end_data    %lx", current->mm->end_data);
 405         kdebug("- start_brk   %lx", current->mm->start_brk);
 406         kdebug("- brk         %lx", current->mm->brk);
 407         kdebug("- start_stack %lx", current->mm->start_stack);
 408
 409 #ifdef ELF_FDPIC_PLAT_INIT
 410         /*
 411          * The ABI may specify that certain registers be set up in special
 412          * ways (on i386 %edx is the address of a DT_FINI function, for
 413          * example.  This macro performs whatever initialization to
 414          * the regs structure is required.
 415          */
 416         dynaddr = interp_params.dynamic_addr ?: exec_params.dynamic_addr;
 417         ELF_FDPIC_PLAT_INIT(regs, exec_params.map_addr, interp_params.map_addr,
 418                             dynaddr);
 419 #endif
 420
 421         /* everything is now ready... get the userspace context ready to roll */
 422         entryaddr = interp_params.entry_addr ?: exec_params.entry_addr;
 423         start_thread(regs, entryaddr, current->mm->start_stack);
 424
 425         if (unlikely(current->ptrace & PT_PTRACED)) {
 426                 if (current->ptrace & PT_TRACE_EXEC)
 427                         ptrace_notify((PTRACE_EVENT_EXEC << 8) | SIGTRAP);
 428                 else
 429                         send_sig(SIGTRAP, current, 0);
 430         }
 431
 432         retval = 0;
 433
 434 error:
 435         if (interpreter) {
 436                 allow_write_access(interpreter);
 437                 fput(interpreter);
 438         }
 439         kfree(interpreter_name);
 440         kfree(exec_params.phdrs);
 441         kfree(exec_params.loadmap);
 442         kfree(interp_params.phdrs);
 443         kfree(interp_params.loadmap);
 444         return retval;
 445
 446         /* unrecoverable error - kill the process */
 447 error_kill:
 448         send_sig(SIGSEGV, current, 0);
 449         goto error;
 450
 451 }
 452
 453 /*****************************************************************************/
 454 /*
 455  * present useful information to the program
 456  */
 457 static int create_elf_fdpic_tables(struct linux_binprm *bprm,
 458                                    struct mm_struct *mm,
 459                                    struct elf_fdpic_params *exec_params,
 460                                    struct elf_fdpic_params *interp_params)
 461 {
 462         unsigned long sp, csp, nitems;
 463         elf_caddr_t __user *argv, *envp;
 464         size_t platform_len = 0, len;
 465         char *k_platform;
 466         char __user *u_platform, *p;
 467         long hwcap;
 468         int loop;
 469
 470         /* we're going to shovel a whole load of stuff onto the stack */
 471 #ifdef CONFIG_MMU
 472         sp = bprm->p;
 473 #else
 474         sp = mm->start_stack;
 475
 476         /* stack the program arguments and environment */
 477         if (elf_fdpic_transfer_args_to_stack(bprm, &sp) < 0)
 478                 return -EFAULT;
 479 #endif
 480
 481         /* get hold of platform and hardware capabilities masks for the machine
 482          * we are running on.  In some cases (Sparc), this info is impossible
 483          * to get, in others (i386) it is merely difficult.
 484          */
 485         hwcap = ELF_HWCAP;
 486         k_platform = ELF_PLATFORM;
 487         u_platform = NULL;
 488
 489         if (k_platform) {
 490                 platform_len = strlen(k_platform) + 1;
 491                 sp -= platform_len;
 492                 u_platform = (char __user *) sp;
 493                 if (__copy_to_user(u_platform, k_platform, platform_len) != 0)
 494                         return -EFAULT;
 495         }
 496
 497 #if defined(__i386__) && defined(CONFIG_SMP)
 498         /* in some cases (e.g. Hyper-Threading), we want to avoid L1 evictions
 499          * by the processes running on the same package. One thing we can do is
 500          * to shuffle the initial stack for them.
 501          *
 502          * the conditionals here are unneeded, but kept in to make the code
 503          * behaviour the same as pre change unless we have hyperthreaded
 504          * processors. This keeps Mr Marcelo Person happier but should be
 505          * removed for 2.5
 506          */
 507         if (smp_num_siblings > 1)
 508                 sp = sp - ((current->pid % 64) << 7);
 509 #endif
 510
 511         sp &= ~7UL;
 512
 513         /* stack the load map(s) */
 514         len = sizeof(struct elf32_fdpic_loadmap);
 515         len += sizeof(struct elf32_fdpic_loadseg) * exec_params->loadmap->nsegs;
 516         sp = (sp - len) & ~7UL;
 517         exec_params->map_addr = sp;
 518
 519         if (copy_to_user((void __user *) sp, exec_params->loadmap, len) != 0)
 520                 return -EFAULT;
 521
 522         current->mm->context.exec_fdpic_loadmap = (unsigned long) sp;
 523
 524         if (interp_params->loadmap) {
 525                 len = sizeof(struct elf32_fdpic_loadmap);
 526                 len += sizeof(struct elf32_fdpic_loadseg) *
 527                         interp_params->loadmap->nsegs;
 528                 sp = (sp - len) & ~7UL;
 529                 interp_params->map_addr = sp;
 530
 531                 if (copy_to_user((void __user *) sp, interp_params->loadmap,
 532                                  len) != 0)
 533                         return -EFAULT;
 534
 535                 current->mm->context.interp_fdpic_loadmap = (unsigned long) sp;
 536         }
 537
 538         /* force 16 byte _final_ alignment here for generality */
 539 #define DLINFO_ITEMS 13
 540
 541         nitems = 1 + DLINFO_ITEMS + (k_platform ? 1 : 0);
 542 #ifdef DLINFO_ARCH_ITEMS
 543         nitems += DLINFO_ARCH_ITEMS;
 544 #endif
 545
 546         csp = sp;
 547         sp -= nitems * 2 * sizeof(unsigned long);
 548         sp -= (bprm->envc + 1) * sizeof(char *);        /* envv[] */
 549         sp -= (bprm->argc + 1) * sizeof(char *);        /* argv[] */
 550         sp -= 1 * sizeof(unsigned long);                /* argc */
 551
 552         csp -= sp & 15UL;
 553         sp -= sp & 15UL;
 554
 555         /* put the ELF interpreter info on the stack */
 556 #define NEW_AUX_ENT(nr, id, val)                                        \
 557         do {                                                            \
 558                 struct { unsigned long _id, _val; } __user *ent;        \
 559                                                                         \
 560                 ent = (void __user *) csp;                              \
 561                 __put_user((id), &ent[nr]._id);                         \
 562                 __put_user((val), &ent[nr]._val);                       \
 563         } while (0)
 564
 565         csp -= 2 * sizeof(unsigned long);
 566         NEW_AUX_ENT(0, AT_NULL, 0);
 567         if (k_platform) {
 568                 csp -= 2 * sizeof(unsigned long);
 569                 NEW_AUX_ENT(0, AT_PLATFORM,
 570                             (elf_addr_t) (unsigned long) u_platform);
 571         }
 572
 573         csp -= DLINFO_ITEMS * 2 * sizeof(unsigned long);
 574         NEW_AUX_ENT( 0, AT_HWCAP,       hwcap);
 575         NEW_AUX_ENT( 1, AT_PAGESZ,      PAGE_SIZE);
 576         NEW_AUX_ENT( 2, AT_CLKTCK,      CLOCKS_PER_SEC);
 577         NEW_AUX_ENT( 3, AT_PHDR,        exec_params->ph_addr);
 578         NEW_AUX_ENT( 4, AT_PHENT,       sizeof(struct elf_phdr));
 579         NEW_AUX_ENT( 5, AT_PHNUM,       exec_params->hdr.e_phnum);
 580         NEW_AUX_ENT( 6, AT_BASE,        interp_params->elfhdr_addr);
 581         NEW_AUX_ENT( 7, AT_FLAGS,       0);
 582         NEW_AUX_ENT( 8, AT_ENTRY,       exec_params->entry_addr);
 583         NEW_AUX_ENT( 9, AT_UID,         (elf_addr_t) current->uid);
 584         NEW_AUX_ENT(10, AT_EUID,        (elf_addr_t) current->euid);
 585         NEW_AUX_ENT(11, AT_GID,         (elf_addr_t) current->gid);
 586         NEW_AUX_ENT(12, AT_EGID,        (elf_addr_t) current->egid);
 587
 588 #ifdef ARCH_DLINFO
 589         /* ARCH_DLINFO must come last so platform specific code can enforce
 590          * special alignment requirements on the AUXV if necessary (eg. PPC).
 591          */
 592         ARCH_DLINFO;
 593 #endif
 594 #undef NEW_AUX_ENT
 595
 596         /* allocate room for argv[] and envv[] */
 597         csp -= (bprm->envc + 1) * sizeof(elf_caddr_t);
 598         envp = (elf_caddr_t __user *) csp;
 599         csp -= (bprm->argc + 1) * sizeof(elf_caddr_t);
 600         argv = (elf_caddr_t __user *) csp;
 601
 602         /* stack argc */
 603         csp -= sizeof(unsigned long);
 604         __put_user(bprm->argc, (unsigned long __user *) csp);
 605
 606         BUG_ON(csp != sp);
 607
 608         /* fill in the argv[] array */
 609 #ifdef CONFIG_MMU
 610         current->mm->arg_start = bprm->p;
 611 #else
 612         current->mm->arg_start = current->mm->start_stack -
 613                 (MAX_ARG_PAGES * PAGE_SIZE - bprm->p);
 614 #endif
 615
 616         p = (char __user *) current->mm->arg_start;
 617         for (loop = bprm->argc; loop > 0; loop--) {
 618                 __put_user((elf_caddr_t) p, argv++);
 619                 len = strnlen_user(p, PAGE_SIZE * MAX_ARG_PAGES);
 620                 if (!len || len > PAGE_SIZE * MAX_ARG_PAGES)
 621                         return -EINVAL;
 622                 p += len;
 623         }
 624         __put_user(NULL, argv);
 625         current->mm->arg_end = (unsigned long) p;
 626
 627         /* fill in the envv[] array */
 628         current->mm->env_start = (unsigned long) p;
 629         for (loop = bprm->envc; loop > 0; loop--) {
 630                 __put_user((elf_caddr_t)(unsigned long) p, envp++);
 631                 len = strnlen_user(p, PAGE_SIZE * MAX_ARG_PAGES);
 632                 if (!len || len > PAGE_SIZE * MAX_ARG_PAGES)
 633                         return -EINVAL;
 634                 p += len;
 635         }
 636         __put_user(NULL, envp);
 637         current->mm->env_end = (unsigned long) p;
 638
 639         mm->start_stack = (unsigned long) sp;
 640         return 0;
 641 }
 642
 643 /*****************************************************************************/
 644 /*
 645  * transfer the program arguments and environment from the holding pages onto
 646  * the stack
 647  */
 648 #ifndef CONFIG_MMU
 649 static int elf_fdpic_transfer_args_to_stack(struct linux_binprm *bprm,
 650                                             unsigned long *_sp)
 651 {
 652         unsigned long index, stop, sp;
 653         char *src;
 654         int ret = 0;
 655
 656         stop = bprm->p >> PAGE_SHIFT;
 657         sp = *_sp;
 658
 659         for (index = MAX_ARG_PAGES - 1; index >= stop; index--) {
 660                 src = kmap(bprm->page[index]);
 661                 sp -= PAGE_SIZE;
 662                 if (copy_to_user((void *) sp, src, PAGE_SIZE) != 0)
 663                         ret = -EFAULT;
 664                 kunmap(bprm->page[index]);
 665                 if (ret < 0)
 666                         goto out;
 667         }
 668
 669         *_sp = (*_sp - (MAX_ARG_PAGES * PAGE_SIZE - bprm->p)) & ~15;
 670
 671 out:
 672         return ret;
 673 }
 674 #endif
 675
 676 /*****************************************************************************/
 677 /*
 678  * load the appropriate binary image (executable or interpreter) into memory
 679  * - we assume no MMU is available
 680  * - if no other PIC bits are set in params->hdr->e_flags
 681  *   - we assume that the LOADable segments in the binary are independently relocatable
 682  *   - we assume R/O executable segments are shareable
 683  * - else
 684  *   - we assume the loadable parts of the image to require fixed displacement
 685  *   - the image is not shareable
 686  */
 687 static int elf_fdpic_map_file(struct elf_fdpic_params *params,
 688                               struct file *file,
 689                               struct mm_struct *mm,
 690                               const char *what)
 691 {
 692         struct elf32_fdpic_loadmap *loadmap;
 693 #ifdef CONFIG_MMU
 694         struct elf32_fdpic_loadseg *mseg;
 695 #endif
 696         struct elf32_fdpic_loadseg *seg;
 697         struct elf32_phdr *phdr;
 698         unsigned long load_addr, stop;
 699         unsigned nloads, tmp;
 700         size_t size;
 701         int loop, ret;
 702
 703         /* allocate a load map table */
 704         nloads = 0;
 705         for (loop = 0; loop < params->hdr.e_phnum; loop++)
 706                 if (params->phdrs[loop].p_type == PT_LOAD)
 707                         nloads++;
 708
 709         if (nloads == 0)
 710                 return -ELIBBAD;
 711
 712         size = sizeof(*loadmap) + nloads * sizeof(*seg);
 713         loadmap = kmalloc(size, GFP_KERNEL);
 714         if (!loadmap)
 715                 return -ENOMEM;
 716
 717         params->loadmap = loadmap;
 718         memset(loadmap, 0, size);
 719
 720         loadmap->version = ELF32_FDPIC_LOADMAP_VERSION;
 721         loadmap->nsegs = nloads;
 722
 723         load_addr = params->load_addr;
 724         seg = loadmap->segs;
 725
 726         /* map the requested LOADs into the memory space */
 727         switch (params->flags & ELF_FDPIC_FLAG_ARRANGEMENT) {
 728         case ELF_FDPIC_FLAG_CONSTDISP:
 729         case ELF_FDPIC_FLAG_CONTIGUOUS:
 730 #ifndef CONFIG_MMU
 731                 ret = elf_fdpic_map_file_constdisp_on_uclinux(params, file, mm);
 732                 if (ret < 0)
 733                         return ret;
 734                 break;
 735 #endif
 736         default:
 737                 ret = elf_fdpic_map_file_by_direct_mmap(params, file, mm);
 738                 if (ret < 0)
 739                         return ret;
 740                 break;
 741         }
 742
 743         /* map the entry point */
 744         if (params->hdr.e_entry) {
 745                 seg = loadmap->segs;
 746                 for (loop = loadmap->nsegs; loop > 0; loop--, seg++) {
 747                         if (params->hdr.e_entry >= seg->p_vaddr &&
 748                             params->hdr.e_entry < seg->p_vaddr + seg->p_memsz) {
 749                                 params->entry_addr =
 750                                         (params->hdr.e_entry - seg->p_vaddr) +
 751                                         seg->addr;
 752                                 break;
 753                         }
 754                 }
 755         }
 756
 757         /* determine where the program header table has wound up if mapped */
 758         stop = params->hdr.e_phoff;
 759         stop += params->hdr.e_phnum * sizeof (struct elf_phdr);
 760         phdr = params->phdrs;
 761
 762         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 763                 if (phdr->p_type != PT_LOAD)
 764                         continue;
 765
 766                 if (phdr->p_offset > params->hdr.e_phoff ||
 767                     phdr->p_offset + phdr->p_filesz < stop)
 768                         continue;
 769
 770                 seg = loadmap->segs;
 771                 for (loop = loadmap->nsegs; loop > 0; loop--, seg++) {
 772                         if (phdr->p_vaddr >= seg->p_vaddr &&
 773                             phdr->p_vaddr + phdr->p_filesz <=
 774                             seg->p_vaddr + seg->p_memsz) {
 775                                 params->ph_addr =
 776                                         (phdr->p_vaddr - seg->p_vaddr) +
 777                                         seg->addr +
 778                                         params->hdr.e_phoff - phdr->p_offset;
 779                                 break;
 780                         }
 781                 }
 782                 break;
 783         }
 784
 785         /* determine where the dynamic section has wound up if there is one */
 786         phdr = params->phdrs;
 787         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 788                 if (phdr->p_type != PT_DYNAMIC)
 789                         continue;
 790
 791                 seg = loadmap->segs;
 792                 for (loop = loadmap->nsegs; loop > 0; loop--, seg++) {
 793                         if (phdr->p_vaddr >= seg->p_vaddr &&
 794                             phdr->p_vaddr + phdr->p_memsz <=
 795                             seg->p_vaddr + seg->p_memsz) {
 796                                 params->dynamic_addr =
 797                                         (phdr->p_vaddr - seg->p_vaddr) +
 798                                         seg->addr;
 799
 800                                 /* check the dynamic section contains at least
 801                                  * one item, and that the last item is a NULL
 802                                  * entry */
 803                                 if (phdr->p_memsz == 0 ||
 804                                     phdr->p_memsz % sizeof(Elf32_Dyn) != 0)
 805                                         goto dynamic_error;
 806
 807                                 tmp = phdr->p_memsz / sizeof(Elf32_Dyn);
 808                                 if (((Elf32_Dyn *)
 809                                      params->dynamic_addr)[tmp - 1].d_tag != 0)
 810                                         goto dynamic_error;
 811                                 break;
 812                         }
 813                 }
 814                 break;
 815         }
 816
 817         /* now elide adjacent segments in the load map on MMU linux
 818          * - on uClinux the holes between may actually be filled with system
 819          *   stuff or stuff from other processes
 820          */
 821 #ifdef CONFIG_MMU
 822         nloads = loadmap->nsegs;
 823         mseg = loadmap->segs;
 824         seg = mseg + 1;
 825         for (loop = 1; loop < nloads; loop++) {
 826                 /* see if we have a candidate for merging */
 827                 if (seg->p_vaddr - mseg->p_vaddr == seg->addr - mseg->addr) {
 828                         load_addr = PAGE_ALIGN(mseg->addr + mseg->p_memsz);
 829                         if (load_addr == (seg->addr & PAGE_MASK)) {
 830                                 mseg->p_memsz +=
 831                                         load_addr -
 832                                         (mseg->addr + mseg->p_memsz);
 833                                 mseg->p_memsz += seg->addr & ~PAGE_MASK;
 834                                 mseg->p_memsz += seg->p_memsz;
 835                                 loadmap->nsegs--;
 836                                 continue;
 837                         }
 838                 }
 839
 840                 mseg++;
 841                 if (mseg != seg)
 842                         *mseg = *seg;
 843         }
 844 #endif
 845
 846         kdebug("Mapped Object [%s]:", what);
 847         kdebug("- elfhdr   : %lx", params->elfhdr_addr);
 848         kdebug("- entry    : %lx", params->entry_addr);
 849         kdebug("- PHDR[]   : %lx", params->ph_addr);
 850         kdebug("- DYNAMIC[]: %lx", params->dynamic_addr);
 851         seg = loadmap->segs;
 852         for (loop = 0; loop < loadmap->nsegs; loop++, seg++)
 853                 kdebug("- LOAD[%d] : %08x-%08x [va=%x ms=%x]",
 854                        loop,
 855                        seg->addr, seg->addr + seg->p_memsz - 1,
 856                        seg->p_vaddr, seg->p_memsz);
 857
 858         return 0;
 859
 860 dynamic_error:
 861         printk("ELF FDPIC %s with invalid DYNAMIC section (inode=%lu)\n",
 862                what, file->f_dentry->d_inode->i_ino);
 863         return -ELIBBAD;
 864 }
 865
 866 /*****************************************************************************/
 867 /*
 868  * map a file with constant displacement under uClinux
 869  */
 870 #ifndef CONFIG_MMU
 871 static int elf_fdpic_map_file_constdisp_on_uclinux(
 872         struct elf_fdpic_params *params,
 873         struct file *file,
 874         struct mm_struct *mm)
 875 {
 876         struct elf32_fdpic_loadseg *seg;
 877         struct elf32_phdr *phdr;
 878         unsigned long load_addr, base = ULONG_MAX, top = 0, maddr = 0, mflags;
 879         loff_t fpos;
 880         int loop, ret;
 881
 882         load_addr = params->load_addr;
 883         seg = params->loadmap->segs;
 884
 885         /* determine the bounds of the contiguous overall allocation we must
 886          * make */
 887         phdr = params->phdrs;
 888         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 889                 if (params->phdrs[loop].p_type != PT_LOAD)
 890                         continue;
 891
 892                 if (base > phdr->p_vaddr)
 893                         base = phdr->p_vaddr;
 894                 if (top < phdr->p_vaddr + phdr->p_memsz)
 895                         top = phdr->p_vaddr + phdr->p_memsz;
 896         }
 897
 898         /* allocate one big anon block for everything */
 899         mflags = MAP_PRIVATE;
 900         if (params->flags & ELF_FDPIC_FLAG_EXECUTABLE)
 901                 mflags |= MAP_EXECUTABLE;
 902
 903         down_write(&mm->mmap_sem);
 904         maddr = do_mmap(NULL, load_addr, top - base,
 905                         PROT_READ | PROT_WRITE | PROT_EXEC, mflags, 0);
 906         up_write(&mm->mmap_sem);
 907         if (IS_ERR_VALUE(maddr))
 908                 return (int) maddr;
 909
 910         if (load_addr != 0)
 911                 load_addr += PAGE_ALIGN(top - base);
 912
 913         /* and then load the file segments into it */
 914         phdr = params->phdrs;
 915         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 916                 if (params->phdrs[loop].p_type != PT_LOAD)
 917                         continue;
 918
 919                 fpos = phdr->p_offset;
 920
 921                 seg->addr = maddr + (phdr->p_vaddr - base);
 922                 seg->p_vaddr = phdr->p_vaddr;
 923                 seg->p_memsz = phdr->p_memsz;
 924
 925                 ret = file->f_op->read(file, (void *) seg->addr,
 926                                        phdr->p_filesz, &fpos);
 927                 if (ret < 0)
 928                         return ret;
 929
 930                 /* map the ELF header address if in this segment */
 931                 if (phdr->p_offset == 0)
 932                         params->elfhdr_addr = seg->addr;
 933
 934                 /* clear any space allocated but not loaded */
 935                 if (phdr->p_filesz < phdr->p_memsz)
 936                         clear_user((void *) (seg->addr + phdr->p_filesz),
 937                                    phdr->p_memsz - phdr->p_filesz);
 938
 939                 if (mm) {
 940                         if (phdr->p_flags & PF_X) {
 941                                 mm->start_code = seg->addr;
 942                                 mm->end_code = seg->addr + phdr->p_memsz;
 943                         } else if (!mm->start_data) {
 944                                 mm->start_data = seg->addr;
 945 #ifndef CONFIG_MMU
 946                                 mm->end_data = seg->addr + phdr->p_memsz;
 947 #endif
 948                         }
 949
 950 #ifdef CONFIG_MMU
 951                         if (seg->addr + phdr->p_memsz > mm->end_data)
 952                                 mm->end_data = seg->addr + phdr->p_memsz;
 953 #endif
 954                 }
 955
 956                 seg++;
 957         }
 958
 959         return 0;
 960 }
 961 #endif
 962
 963 /*****************************************************************************/
 964 /*
 965  * map a binary by direct mmap() of the individual PT_LOAD segments
 966  */
 967 static int elf_fdpic_map_file_by_direct_mmap(struct elf_fdpic_params *params,
 968                                              struct file *file,
 969                                              struct mm_struct *mm)
 970 {
 971         struct elf32_fdpic_loadseg *seg;
 972         struct elf32_phdr *phdr;
 973         unsigned long load_addr, delta_vaddr;
 974         int loop, dvset;
 975
 976         load_addr = params->load_addr;
 977         delta_vaddr = 0;
 978         dvset = 0;
 979
 980         seg = params->loadmap->segs;
 981
 982         /* deal with each load segment separately */
 983         phdr = params->phdrs;
 984         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 985                 unsigned long maddr, disp, excess, excess1;
 986                 int prot = 0, flags;
 987
 988                 if (phdr->p_type != PT_LOAD)
 989                         continue;
 990
 991                 kdebug("[LOAD] va=%lx of=%lx fs=%lx ms=%lx",
 992                        (unsigned long) phdr->p_vaddr,
 993                        (unsigned long) phdr->p_offset,
 994                        (unsigned long) phdr->p_filesz,
 995                        (unsigned long) phdr->p_memsz);
 996
 997                 /* determine the mapping parameters */
 998                 if (phdr->p_flags & PF_R) prot |= PROT_READ;
 999                 if (phdr->p_flags & PF_W) prot |= PROT_WRITE;
1000                 if (phdr->p_flags & PF_X) prot |= PROT_EXEC;
1001
1002                 flags = MAP_PRIVATE | MAP_DENYWRITE;
1003                 if (params->flags & ELF_FDPIC_FLAG_EXECUTABLE)
1004                         flags |= MAP_EXECUTABLE;
1005
1006                 maddr = 0;
1007
1008                 switch (params->flags & ELF_FDPIC_FLAG_ARRANGEMENT) {
1009                 case ELF_FDPIC_FLAG_INDEPENDENT:
1010                         /* PT_LOADs are independently locatable */
1011                         break;
1012
1013                 case ELF_FDPIC_FLAG_HONOURVADDR:
1014                         /* the specified virtual address must be honoured */
1015                         maddr = phdr->p_vaddr;
1016                         flags |= MAP_FIXED;
1017                         break;
1018
1019                 case ELF_FDPIC_FLAG_CONSTDISP:
1020                         /* constant displacement
1021                          * - can be mapped anywhere, but must be mapped as a
1022                          *   unit
1023                          */
1024                         if (!dvset) {
1025                                 maddr = load_addr;
1026                                 delta_vaddr = phdr->p_vaddr;
1027                                 dvset = 1;
1028                         } else {
1029                                 maddr = load_addr + phdr->p_vaddr - delta_vaddr;
1030                                 flags |= MAP_FIXED;
1031                         }
1032                         break;
1033
1034                 case ELF_FDPIC_FLAG_CONTIGUOUS:
1035                         /* contiguity handled later */
1036                         break;
1037
1038                 default:
1039                         BUG();
1040                 }
1041
1042                 maddr &= PAGE_MASK;
1043
1044                 /* create the mapping */
1045                 disp = phdr->p_vaddr & ~PAGE_MASK;
1046                 down_write(&mm->mmap_sem);
1047                 maddr = do_mmap(file, maddr, phdr->p_memsz + disp, prot, flags,
1048                                 phdr->p_offset - disp);
1049                 up_write(&mm->mmap_sem);
1050
1051                 kdebug("mmap[%d] <file> sz=%lx pr=%x fl=%x of=%lx --> %08lx",
1052                        loop, phdr->p_memsz + disp, prot, flags,
1053                        phdr->p_offset - disp, maddr);
1054
1055                 if (IS_ERR_VALUE(maddr))
1056                         return (int) maddr;
1057
1058                 if ((params->flags & ELF_FDPIC_FLAG_ARRANGEMENT) ==
1059                     ELF_FDPIC_FLAG_CONTIGUOUS)
1060                         load_addr += PAGE_ALIGN(phdr->p_memsz + disp);
1061
1062                 seg->addr = maddr + disp;
1063                 seg->p_vaddr = phdr->p_vaddr;
1064                 seg->p_memsz = phdr->p_memsz;
1065
1066                 /* map the ELF header address if in this segment */
1067                 if (phdr->p_offset == 0)
1068                         params->elfhdr_addr = seg->addr;
1069
1070                 /* clear the bit between beginning of mapping and beginning of
1071                  * PT_LOAD */
1072                 if (prot & PROT_WRITE && disp > 0) {
1073                         kdebug("clear[%d] ad=%lx sz=%lx", loop, maddr, disp);
1074                         clear_user((void __user *) maddr, disp);
1075                         maddr += disp;
1076                 }
1077
1078                 /* clear any space allocated but not loaded
1079                  * - on uClinux we can just clear the lot
1080                  * - on MMU linux we'll get a SIGBUS beyond the last page
1081                  *   extant in the file
1082                  */
1083                 excess = phdr->p_memsz - phdr->p_filesz;
1084                 excess1 = PAGE_SIZE - ((maddr + phdr->p_filesz) & ~PAGE_MASK);
1085
1086 #ifdef CONFIG_MMU
1087                 if (excess > excess1) {
1088                         unsigned long xaddr = maddr + phdr->p_filesz + excess1;
1089                         unsigned long xmaddr;
1090
1091                         flags |= MAP_FIXED | MAP_ANONYMOUS;
1092                         down_write(&mm->mmap_sem);
1093                         xmaddr = do_mmap(NULL, xaddr, excess - excess1,
1094                                          prot, flags, 0);
1095                         up_write(&mm->mmap_sem);
1096
1097                         kdebug("mmap[%d] <anon>"
1098                                " ad=%lx sz=%lx pr=%x fl=%x of=0 --> %08lx",
1099                                loop, xaddr, excess - excess1, prot, flags,
1100                                xmaddr);
1101
1102                         if (xmaddr != xaddr)
1103                                 return -ENOMEM;
1104                 }
1105
1106                 if (prot & PROT_WRITE && excess1 > 0) {
1107                         kdebug("clear[%d] ad=%lx sz=%lx",
1108                                loop, maddr + phdr->p_filesz, excess1);
1109                         clear_user((void __user *) maddr + phdr->p_filesz,
1110                                    excess1);
1111                 }
1112
1113 #else
1114                 if (excess > 0) {
1115                         kdebug("clear[%d] ad=%lx sz=%lx",
1116                                loop, maddr + phdr->p_filesz, excess);
1117                         clear_user((void *) maddr + phdr->p_filesz, excess);
1118                 }
1119 #endif
1120
1121                 if (mm) {
1122                         if (phdr->p_flags & PF_X) {
1123                                 mm->start_code = maddr;
1124                                 mm->end_code = maddr + phdr->p_memsz;
1125                         } else if (!mm->start_data) {
1126                                 mm->start_data = maddr;
1127                                 mm->end_data = maddr + phdr->p_memsz;
1128                         }
1129                 }
1130
1131                 seg++;
1132         }
1133
1134         return 0;
1135 }
1136
1137 /*****************************************************************************/
1138 /*
1139  * ELF-FDPIC core dumper
1140  *
1141  * Modelled on fs/exec.c:aout_core_dump()
1142  * Jeremy Fitzhardinge <jeremy@sw.oz.au>
1143  *
1144  * Modelled on fs/binfmt_elf.c core dumper
1145  */
1146 #if defined(USE_ELF_CORE_DUMP) && defined(CONFIG_ELF_CORE)
1147
1148 /*
1149  * These are the only things you should do on a core-file: use only these
1150  * functions to write out all the necessary info.
1151  */
1152 static int dump_write(struct file *file, const void *addr, int nr)
1153 {
1154         return file->f_op->write(file, addr, nr, &file->f_pos) == nr;
1155 }
1156
1157 static int dump_seek(struct file *file, loff_t off)
1158 {
1159         if (file->f_op->llseek) {
1160                 if (file->f_op->llseek(file, off, SEEK_SET) != off)
1161                         return 0;
1162         } else {
1163                 file->f_pos = off;
1164         }
1165         return 1;
1166 }
1167
1168 /*
1169  * Decide whether a segment is worth dumping; default is yes to be
1170  * sure (missing info is worse than too much; etc).
1171  * Personally I'd include everything, and use the coredump limit...
1172  *
1173  * I think we should skip something. But I am not sure how. H.J.
1174  */
1175 static int maydump(struct vm_area_struct *vma)
1176 {
1177         /* Do not dump I/O mapped devices or special mappings */
1178         if (vma->vm_flags & (VM_IO | VM_RESERVED)) {
1179                 kdcore("%08lx: %08lx: no (IO)", vma->vm_start, vma->vm_flags);
1180                 return 0;
1181         }
1182
1183         /* If we may not read the contents, don't allow us to dump
1184          * them either. "dump_write()" can't handle it anyway.
1185          */
1186         if (!(vma->vm_flags & VM_READ)) {
1187                 kdcore("%08lx: %08lx: no (!read)", vma->vm_start, vma->vm_flags);
1188                 return 0;
1189         }
1190
1191         /* Dump shared memory only if mapped from an anonymous file. */
1192         if (vma->vm_flags & VM_SHARED) {
1193                 if (vma->vm_file->f_dentry->d_inode->i_nlink == 0) {
1194                         kdcore("%08lx: %08lx: no (share)", vma->vm_start, vma->vm_flags);
1195                         return 1;
1196                 }
1197
1198                 kdcore("%08lx: %08lx: no (share)", vma->vm_start, vma->vm_flags);
1199                 return 0;
1200         }
1201
1202 #ifdef CONFIG_MMU
1203         /* If it hasn't been written to, don't write it out */
1204         if (!vma->anon_vma) {
1205                 kdcore("%08lx: %08lx: no (!anon)", vma->vm_start, vma->vm_flags);
1206                 return 0;
1207         }
1208 #endif
1209
1210         kdcore("%08lx: %08lx: yes", vma->vm_start, vma->vm_flags);
1211         return 1;
1212 }
1213
1214 /* An ELF note in memory */
1215 struct memelfnote
1216 {
1217         const char *name;
1218         int type;
1219         unsigned int datasz;
1220         void *data;
1221 };
1222
1223 static int notesize(struct memelfnote *en)
1224 {
1225         int sz;
1226
1227         sz = sizeof(struct elf_note);
1228         sz += roundup(strlen(en->name) + 1, 4);
1229         sz += roundup(en->datasz, 4);
1230
1231         return sz;
1232 }
1233
1234 /* #define DEBUG */
1235
1236 #define DUMP_WRITE(addr, nr)    \
1237         do { if (!dump_write(file, (addr), (nr))) return 0; } while(0)
1238 #define DUMP_SEEK(off)  \
1239         do { if (!dump_seek(file, (off))) return 0; } while(0)
1240
1241 static int writenote(struct memelfnote *men, struct file *file)
1242 {
1243         struct elf_note en;
1244
1245         en.n_namesz = strlen(men->name) + 1;
1246         en.n_descsz = men->datasz;
1247         en.n_type = men->type;
1248
1249         DUMP_WRITE(&en, sizeof(en));
1250         DUMP_WRITE(men->name, en.n_namesz);
1251         /* XXX - cast from long long to long to avoid need for libgcc.a */
1252         DUMP_SEEK(roundup((unsigned long)file->f_pos, 4));      /* XXX */
1253         DUMP_WRITE(men->data, men->datasz);
1254         DUMP_SEEK(roundup((unsigned long)file->f_pos, 4));      /* XXX */
1255
1256         return 1;
1257 }
1258 #undef DUMP_WRITE
1259 #undef DUMP_SEEK
1260
1261 #define DUMP_WRITE(addr, nr)    \
1262         if ((size += (nr)) > limit || !dump_write(file, (addr), (nr))) \
1263                 goto end_coredump;
1264 #define DUMP_SEEK(off)  \
1265         if (!dump_seek(file, (off))) \
1266                 goto end_coredump;
1267
1268 static inline void fill_elf_fdpic_header(struct elfhdr *elf, int segs)
1269 {
1270         memcpy(elf->e_ident, ELFMAG, SELFMAG);
1271         elf->e_ident[EI_CLASS] = ELF_CLASS;
1272         elf->e_ident[EI_DATA] = ELF_DATA;
1273         elf->e_ident[EI_VERSION] = EV_CURRENT;
1274         elf->e_ident[EI_OSABI] = ELF_OSABI;
1275         memset(elf->e_ident+EI_PAD, 0, EI_NIDENT-EI_PAD);
1276
1277         elf->e_type = ET_CORE;
1278         elf->e_machine = ELF_ARCH;
1279         elf->e_version = EV_CURRENT;
1280         elf->e_entry = 0;
1281         elf->e_phoff = sizeof(struct elfhdr);
1282         elf->e_shoff = 0;
1283         elf->e_flags = ELF_FDPIC_CORE_EFLAGS;
1284         elf->e_ehsize = sizeof(struct elfhdr);
1285         elf->e_phentsize = sizeof(struct elf_phdr);
1286         elf->e_phnum = segs;
1287         elf->e_shentsize = 0;
1288         elf->e_shnum = 0;
1289         elf->e_shstrndx = 0;
1290         return;
1291 }
1292
1293 static inline void fill_elf_note_phdr(struct elf_phdr *phdr, int sz, loff_t offset)
1294 {
1295         phdr->p_type = PT_NOTE;
1296         phdr->p_offset = offset;
1297         phdr->p_vaddr = 0;
1298         phdr->p_paddr = 0;
1299         phdr->p_filesz = sz;
1300         phdr->p_memsz = 0;
1301         phdr->p_flags = 0;
1302         phdr->p_align = 0;
1303         return;
1304 }
1305
1306 static inline void fill_note(struct memelfnote *note, const char *name, int type,
1307                 unsigned int sz, void *data)
1308 {
1309         note->name = name;
1310         note->type = type;
1311         note->datasz = sz;
1312         note->data = data;
1313         return;
1314 }
1315
1316 /*
1317  * fill up all the fields in prstatus from the given task struct, except
1318  * registers which need to be filled up seperately.
1319  */
1320 static void fill_prstatus(struct elf_prstatus *prstatus,
1321                           struct task_struct *p, long signr)
1322 {
1323         prstatus->pr_info.si_signo = prstatus->pr_cursig = signr;
1324         prstatus->pr_sigpend = p->pending.signal.sig[0];
1325         prstatus->pr_sighold = p->blocked.sig[0];
1326         prstatus->pr_pid = p->pid;
1327         prstatus->pr_ppid = p->parent->pid;
1328         prstatus->pr_pgrp = process_group(p);
1329         prstatus->pr_sid = p->signal->session;
1330         if (thread_group_leader(p)) {
1331                 /*
1332                  * This is the record for the group leader.  Add in the
1333                  * cumulative times of previous dead threads.  This total
1334                  * won't include the time of each live thread whose state
1335                  * is included in the core dump.  The final total reported
1336                  * to our parent process when it calls wait4 will include
1337                  * those sums as well as the little bit more time it takes
1338                  * this and each other thread to finish dying after the
1339                  * core dump synchronization phase.
1340                  */
1341                 cputime_to_timeval(cputime_add(p->utime, p->signal->utime),
1342                                    &prstatus->pr_utime);
1343                 cputime_to_timeval(cputime_add(p->stime, p->signal->stime),
1344                                    &prstatus->pr_stime);
1345         } else {
1346                 cputime_to_timeval(p->utime, &prstatus->pr_utime);
1347                 cputime_to_timeval(p->stime, &prstatus->pr_stime);
1348         }
1349         cputime_to_timeval(p->signal->cutime, &prstatus->pr_cutime);
1350         cputime_to_timeval(p->signal->cstime, &prstatus->pr_cstime);
1351
1352         prstatus->pr_exec_fdpic_loadmap = p->mm->context.exec_fdpic_loadmap;
1353         prstatus->pr_interp_fdpic_loadmap = p->mm->context.interp_fdpic_loadmap;
1354 }
1355
1356 static int fill_psinfo(struct elf_prpsinfo *psinfo, struct task_struct *p,
1357                        struct mm_struct *mm)
1358 {
1359         unsigned int i, len;
1360
1361         /* first copy the parameters from user space */
1362         memset(psinfo, 0, sizeof(struct elf_prpsinfo));
1363
1364         len = mm->arg_end - mm->arg_start;
1365         if (len >= ELF_PRARGSZ)
1366                 len = ELF_PRARGSZ - 1;
1367         if (copy_from_user(&psinfo->pr_psargs,
1368                            (const char __user *) mm->arg_start, len))
1369                 return -EFAULT;
1370         for (i = 0; i < len; i++)
1371                 if (psinfo->pr_psargs[i] == 0)
1372                         psinfo->pr_psargs[i] = ' ';
1373         psinfo->pr_psargs[len] = 0;
1374
1375         psinfo->pr_pid = p->pid;
1376         psinfo->pr_ppid = p->parent->pid;
1377         psinfo->pr_pgrp = process_group(p);
1378         psinfo->pr_sid = p->signal->session;
1379
1380         i = p->state ? ffz(~p->state) + 1 : 0;
1381         psinfo->pr_state = i;
1382         psinfo->pr_sname = (i > 5) ? '.' : "RSDTZW"[i];
1383         psinfo->pr_zomb = psinfo->pr_sname == 'Z';
1384         psinfo->pr_nice = task_nice(p);
1385         psinfo->pr_flag = p->flags;
1386         SET_UID(psinfo->pr_uid, p->uid);
1387         SET_GID(psinfo->pr_gid, p->gid);
1388         strncpy(psinfo->pr_fname, p->comm, sizeof(psinfo->pr_fname));
1389
1390         return 0;
1391 }
1392
1393 /* Here is the structure in which status of each thread is captured. */
1394 struct elf_thread_status
1395 {
1396         struct list_head list;
1397         struct elf_prstatus prstatus;   /* NT_PRSTATUS */
1398         elf_fpregset_t fpu;             /* NT_PRFPREG */
1399         struct task_struct *thread;
1400 #ifdef ELF_CORE_COPY_XFPREGS
1401         elf_fpxregset_t xfpu;           /* NT_PRXFPREG */
1402 #endif
1403         struct memelfnote notes[3];
1404         int num_notes;
1405 };
1406
1407 /*
1408  * In order to add the specific thread information for the elf file format,
1409  * we need to keep a linked list of every thread's pr_status and then create
1410  * a single section for them in the final core file.
1411  */
1412 static int elf_dump_thread_status(long signr, struct elf_thread_status *t)
1413 {
1414         struct task_struct *p = t->thread;
1415         int sz = 0;
1416
1417         t->num_notes = 0;
1418
1419         fill_prstatus(&t->prstatus, p, signr);
1420         elf_core_copy_task_regs(p, &t->prstatus.pr_reg);
1421
1422         fill_note(&t->notes[0], "CORE", NT_PRSTATUS, sizeof(t->prstatus),
1423                   &t->prstatus);
1424         t->num_notes++;
1425         sz += notesize(&t->notes[0]);
1426
1427         t->prstatus.pr_fpvalid = elf_core_copy_task_fpregs(p, NULL, &t->fpu);
1428         if (t->prstatus.pr_fpvalid) {
1429                 fill_note(&t->notes[1], "CORE", NT_PRFPREG, sizeof(t->fpu),
1430                           &t->fpu);
1431                 t->num_notes++;
1432                 sz += notesize(&t->notes[1]);
1433         }
1434
1435 #ifdef ELF_CORE_COPY_XFPREGS
1436         if (elf_core_copy_task_xfpregs(p, &t->xfpu)) {
1437                 fill_note(&t->notes[2], "LINUX", NT_PRXFPREG, sizeof(t->xfpu),
1438                           &t->xfpu);
1439                 t->num_notes++;
1440                 sz += notesize(&t->notes[2]);
1441         }
1442 #endif
1443         return sz;
1444 }
1445
1446 /*
1447  * dump the segments for an MMU process
1448  */
1449 #ifdef CONFIG_MMU
1450 static int elf_fdpic_dump_segments(struct file *file, struct mm_struct *mm,
1451                                    size_t *size, unsigned long *limit)
1452 {
1453         struct vm_area_struct *vma;
1454
1455         for (vma = current->mm->mmap; vma; vma = vma->vm_next) {
1456                 unsigned long addr;
1457
1458                 if (!maydump(vma))
1459                         continue;
1460
1461                 for (addr = vma->vm_start;
1462                      addr < vma->vm_end;
1463                      addr += PAGE_SIZE
1464                      ) {
1465                         struct vm_area_struct *vma;
1466                         struct page *page;
1467
1468                         if (get_user_pages(current, current->mm, addr, 1, 0, 1,
1469                                            &page, &vma) <= 0) {
1470                                 DUMP_SEEK(file->f_pos + PAGE_SIZE);
1471                         }
1472                         else if (page == ZERO_PAGE(addr)) {
1473                                 DUMP_SEEK(file->f_pos + PAGE_SIZE);
1474                                 page_cache_release(page);
1475                         }
1476                         else {
1477                                 void *kaddr;
1478
1479                                 flush_cache_page(vma, addr, page_to_pfn(page));
1480                                 kaddr = kmap(page);
1481                                 if ((*size += PAGE_SIZE) > *limit ||
1482                                     !dump_write(file, kaddr, PAGE_SIZE)
1483                                     ) {
1484                                         kunmap(page);
1485                                         page_cache_release(page);
1486                                         return -EIO;
1487                                 }
1488                                 kunmap(page);
1489                                 page_cache_release(page);
1490                         }
1491                 }
1492         }
1493
1494         return 0;
1495
1496 end_coredump:
1497         return -EFBIG;
1498 }
1499 #endif
1500
1501 /*
1502  * dump the segments for a NOMMU process
1503  */
1504 #ifndef CONFIG_MMU
1505 static int elf_fdpic_dump_segments(struct file *file, struct mm_struct *mm,
1506                                    size_t *size, unsigned long *limit)
1507 {
1508         struct vm_list_struct *vml;
1509
1510         for (vml = current->mm->context.vmlist; vml; vml = vml->next) {
1511         struct vm_area_struct *vma = vml->vma;
1512
1513                 if (!maydump(vma))
1514                         continue;
1515
1516                 if ((*size += PAGE_SIZE) > *limit)
1517                         return -EFBIG;
1518
1519                 if (!dump_write(file, (void *) vma->vm_start,
1520                                 vma->vm_end - vma->vm_start))
1521                         return -EIO;
1522         }
1523
1524         return 0;
1525 }
1526 #endif
1527
1528 /*
1529  * Actual dumper
1530  *
1531  * This is a two-pass process; first we find the offsets of the bits,
1532  * and then they are actually written out.  If we run out of core limit
1533  * we just truncate.
1534  */
1535 static int elf_fdpic_core_dump(long signr, struct pt_regs *regs,
1536                                struct file *file)
1537 {
1538 #define NUM_NOTES       6
1539         int has_dumped = 0;
1540         mm_segment_t fs;
1541         int segs;
1542         size_t size = 0;
1543         int i;
1544         struct vm_area_struct *vma;
1545         struct elfhdr *elf = NULL;
1546         loff_t offset = 0, dataoff;
1547         unsigned long limit = current->signal->rlim[RLIMIT_CORE].rlim_cur;
1548         int numnote;
1549         struct memelfnote *notes = NULL;
1550         struct elf_prstatus *prstatus = NULL;   /* NT_PRSTATUS */
1551         struct elf_prpsinfo *psinfo = NULL;     /* NT_PRPSINFO */
1552         struct task_struct *g, *p;
1553         LIST_HEAD(thread_list);
1554         struct list_head *t;
1555         elf_fpregset_t *fpu = NULL;
1556 #ifdef ELF_CORE_COPY_XFPREGS
1557         elf_fpxregset_t *xfpu = NULL;
1558 #endif
1559         int thread_status_size = 0;
1560 #ifndef CONFIG_MMU
1561         struct vm_list_struct *vml;
1562 #endif
1563         elf_addr_t *auxv;
1564
1565         /*
1566          * We no longer stop all VM operations.
1567          *
1568          * This is because those proceses that could possibly change map_count
1569          * or the mmap / vma pages are now blocked in do_exit on current
1570          * finishing this core dump.
1571          *
1572          * Only ptrace can touch these memory addresses, but it doesn't change
1573          * the map_count or the pages allocated. So no possibility of crashing
1574          * exists while dumping the mm->vm_next areas to the core file.
1575          */
1576
1577         /* alloc memory for large data structures: too large to be on stack */
1578         elf = kmalloc(sizeof(*elf), GFP_KERNEL);
1579         if (!elf)
1580                 goto cleanup;
1581         prstatus = kzalloc(sizeof(*prstatus), GFP_KERNEL);
1582         if (!prstatus)
1583                 goto cleanup;
1584         psinfo = kmalloc(sizeof(*psinfo), GFP_KERNEL);
1585         if (!psinfo)
1586                 goto cleanup;
1587         notes = kmalloc(NUM_NOTES * sizeof(struct memelfnote), GFP_KERNEL);
1588         if (!notes)
1589                 goto cleanup;
1590         fpu = kmalloc(sizeof(*fpu), GFP_KERNEL);
1591         if (!fpu)
1592                 goto cleanup;
1593 #ifdef ELF_CORE_COPY_XFPREGS
1594         xfpu = kmalloc(sizeof(*xfpu), GFP_KERNEL);
1595         if (!xfpu)
1596                 goto cleanup;
1597 #endif
1598
1599         if (signr) {
1600                 struct elf_thread_status *tmp;
1601                 read_lock(&tasklist_lock);
1602                 do_each_thread(g,p)
1603                         if (current->mm == p->mm && current != p) {
1604                                 tmp = kzalloc(sizeof(*tmp), GFP_ATOMIC);
1605                                 if (!tmp) {
1606                                         read_unlock(&tasklist_lock);
1607                                         goto cleanup;
1608                                 }
1609                                 INIT_LIST_HEAD(&tmp->list);
1610                                 tmp->thread = p;
1611                                 list_add(&tmp->list, &thread_list);
1612                         }
1613                 while_each_thread(g,p);
1614                 read_unlock(&tasklist_lock);
1615                 list_for_each(t, &thread_list) {
1616                         struct elf_thread_status *tmp;
1617                         int sz;
1618
1619                         tmp = list_entry(t, struct elf_thread_status, list);
1620                         sz = elf_dump_thread_status(signr, tmp);
1621                         thread_status_size += sz;
1622                 }
1623         }
1624
1625         /* now collect the dump for the current */
1626         fill_prstatus(prstatus, current, signr);
1627         elf_core_copy_regs(&prstatus->pr_reg, regs);
1628
1629 #ifdef CONFIG_MMU
1630         segs = current->mm->map_count;
1631 #else
1632         segs = 0;
1633         for (vml = current->mm->context.vmlist; vml; vml = vml->next)
1634             segs++;
1635 #endif
1636 #ifdef ELF_CORE_EXTRA_PHDRS
1637         segs += ELF_CORE_EXTRA_PHDRS;
1638 #endif
1639
1640         /* Set up header */
1641         fill_elf_fdpic_header(elf, segs + 1);   /* including notes section */
1642
1643         has_dumped = 1;
1644         current->flags |= PF_DUMPCORE;
1645
1646         /*
1647          * Set up the notes in similar form to SVR4 core dumps made
1648          * with info from their /proc.
1649          */
1650
1651         fill_note(notes + 0, "CORE", NT_PRSTATUS, sizeof(*prstatus), prstatus);
1652         fill_psinfo(psinfo, current->group_leader, current->mm);
1653         fill_note(notes + 1, "CORE", NT_PRPSINFO, sizeof(*psinfo), psinfo);
1654
1655         numnote = 2;
1656
1657         auxv = (elf_addr_t *) current->mm->saved_auxv;
1658
1659         i = 0;
1660         do
1661                 i += 2;
1662         while (auxv[i - 2] != AT_NULL);
1663         fill_note(&notes[numnote++], "CORE", NT_AUXV,
1664                   i * sizeof(elf_addr_t), auxv);
1665
1666         /* Try to dump the FPU. */
1667         if ((prstatus->pr_fpvalid =
1668              elf_core_copy_task_fpregs(current, regs, fpu)))
1669                 fill_note(notes + numnote++,
1670                           "CORE", NT_PRFPREG, sizeof(*fpu), fpu);
1671 #ifdef ELF_CORE_COPY_XFPREGS
1672         if (elf_core_copy_task_xfpregs(current, xfpu))
1673                 fill_note(notes + numnote++,
1674                           "LINUX", NT_PRXFPREG, sizeof(*xfpu), xfpu);
1675 #endif
1676
1677         fs = get_fs();
1678         set_fs(KERNEL_DS);
1679
1680         DUMP_WRITE(elf, sizeof(*elf));
1681         offset += sizeof(*elf);                         /* Elf header */
1682         offset += (segs+1) * sizeof(struct elf_phdr);   /* Program headers */
1683
1684         /* Write notes phdr entry */
1685         {
1686                 struct elf_phdr phdr;
1687                 int sz = 0;
1688
1689                 for (i = 0; i < numnote; i++)
1690                         sz += notesize(notes + i);
1691
1692                 sz += thread_status_size;
1693
1694                 fill_elf_note_phdr(&phdr, sz, offset);
1695                 offset += sz;
1696                 DUMP_WRITE(&phdr, sizeof(phdr));
1697         }
1698
1699         /* Page-align dumped data */
1700         dataoff = offset = roundup(offset, ELF_EXEC_PAGESIZE);
1701
1702         /* write program headers for segments dump */
1703         for (
1704 #ifdef CONFIG_MMU
1705                 vma = current->mm->mmap; vma; vma = vma->vm_next
1706 #else
1707                         vml = current->mm->context.vmlist; vml; vml = vml->next
1708 #endif
1709              ) {
1710                 struct elf_phdr phdr;
1711                 size_t sz;
1712
1713 #ifndef CONFIG_MMU
1714                 vma = vml->vma;
1715 #endif
1716
1717                 sz = vma->vm_end - vma->vm_start;
1718
1719                 phdr.p_type = PT_LOAD;
1720                 phdr.p_offset = offset;
1721                 phdr.p_vaddr = vma->vm_start;
1722                 phdr.p_paddr = 0;
1723                 phdr.p_filesz = maydump(vma) ? sz : 0;
1724                 phdr.p_memsz = sz;
1725                 offset += phdr.p_filesz;
1726                 phdr.p_flags = vma->vm_flags & VM_READ ? PF_R : 0;
1727                 if (vma->vm_flags & VM_WRITE)
1728                         phdr.p_flags |= PF_W;
1729                 if (vma->vm_flags & VM_EXEC)
1730                         phdr.p_flags |= PF_X;
1731                 phdr.p_align = ELF_EXEC_PAGESIZE;
1732
1733                 DUMP_WRITE(&phdr, sizeof(phdr));
1734         }
1735
1736 #ifdef ELF_CORE_WRITE_EXTRA_PHDRS
1737         ELF_CORE_WRITE_EXTRA_PHDRS;
1738 #endif
1739
1740         /* write out the notes section */
1741         for (i = 0; i < numnote; i++)
1742                 if (!writenote(notes + i, file))
1743                         goto end_coredump;
1744
1745         /* write out the thread status notes section */
1746         list_for_each(t, &thread_list) {
1747                 struct elf_thread_status *tmp =
1748                                 list_entry(t, struct elf_thread_status, list);
1749
1750                 for (i = 0; i < tmp->num_notes; i++)
1751                         if (!writenote(&tmp->notes[i], file))
1752                                 goto end_coredump;
1753         }
1754
1755         DUMP_SEEK(dataoff);
1756
1757         if (elf_fdpic_dump_segments(file, current->mm, &size, &limit) < 0)
1758                 goto end_coredump;
1759
1760 #ifdef ELF_CORE_WRITE_EXTRA_DATA
1761         ELF_CORE_WRITE_EXTRA_DATA;
1762 #endif
1763
1764         if (file->f_pos != offset) {
1765                 /* Sanity check */
1766                 printk(KERN_WARNING
1767                        "elf_core_dump: file->f_pos (%lld) != offset (%lld)\n",
1768                        file->f_pos, offset);
1769         }
1770
1771 end_coredump:
1772         set_fs(fs);
1773
1774 cleanup:
1775         while (!list_empty(&thread_list)) {
1776                 struct list_head *tmp = thread_list.next;
1777                 list_del(tmp);
1778                 kfree(list_entry(tmp, struct elf_thread_status, list));
1779         }
1780
1781         kfree(elf);
1782         kfree(prstatus);
1783         kfree(psinfo);
1784         kfree(notes);
1785         kfree(fpu);
1786 #ifdef ELF_CORE_COPY_XFPREGS
1787         kfree(xfpu);
1788 #endif
1789         return has_dumped;
1790 #undef NUM_NOTES
1791 }
1792
1793 #endif          /* USE_ELF_CORE_DUMP */