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