fs/binfmt_elf_fdpic.c

   1 /* binfmt_elf_fdpic.c: FDPIC ELF binary format
   2  *
   3  * Copyright (C) 2003, 2004, 2006 Red Hat, Inc. All Rights Reserved.
   4  * Written by David Howells (dhowells@redhat.com)
   5  * Derived from binfmt_elf.c
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public License
   9  * as published by the Free Software Foundation; either version
  10  * 2 of the License, or (at your option) any later version.
  11  */
  12
  13 #include <linux/module.h>
  14
  15 #include <linux/fs.h>
  16 #include <linux/stat.h>
  17 #include <linux/sched.h>
  18 #include <linux/mm.h>
  19 #include <linux/mman.h>
  20 #include <linux/errno.h>
  21 #include <linux/signal.h>
  22 #include <linux/binfmts.h>
  23 #include <linux/string.h>
  24 #include <linux/file.h>
  25 #include <linux/fcntl.h>
  26 #include <linux/slab.h>
  27 #include <linux/pagemap.h>
  28 #include <linux/security.h>
  29 #include <linux/highmem.h>
  30 #include <linux/highuid.h>
  31 #include <linux/personality.h>
  32 #include <linux/ptrace.h>
  33 #include <linux/init.h>
  34 #include <linux/elf.h>
  35 #include <linux/elf-fdpic.h>
  36 #include <linux/elfcore.h>
  37 #include <linux/coredump.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
  45 #if 0
  46 #define kdebug(fmt, ...) printk("FDPIC "fmt"\n" ,##__VA_ARGS__ )
  47 #else
  48 #define kdebug(fmt, ...) do {} while(0)
  49 #endif
  50
  51 #if 0
  52 #define kdcore(fmt, ...) printk("FDPIC "fmt"\n" ,##__VA_ARGS__ )
  53 #else
  54 #define kdcore(fmt, ...) do {} while(0)
  55 #endif
  56
  57 MODULE_LICENSE("GPL");
  58
  59 static int load_elf_fdpic_binary(struct linux_binprm *);
  60 static int elf_fdpic_fetch_phdrs(struct elf_fdpic_params *, struct file *);
  61 static int elf_fdpic_map_file(struct elf_fdpic_params *, struct file *,
  62                               struct mm_struct *, const char *);
  63
  64 static int create_elf_fdpic_tables(struct linux_binprm *, struct mm_struct *,
  65                                    struct elf_fdpic_params *,
  66                                    struct elf_fdpic_params *);
  67
  68 #ifndef CONFIG_MMU
  69 static int elf_fdpic_transfer_args_to_stack(struct linux_binprm *,
  70                                             unsigned long *);
  71 static int elf_fdpic_map_file_constdisp_on_uclinux(struct elf_fdpic_params *,
  72                                                    struct file *,
  73                                                    struct mm_struct *);
  74 #endif
  75
  76 static int elf_fdpic_map_file_by_direct_mmap(struct elf_fdpic_params *,
  77                                              struct file *, struct mm_struct *);
  78
  79 #ifdef CONFIG_ELF_CORE
  80 static int elf_fdpic_core_dump(struct coredump_params *cprm);
  81 #endif
  82
  83 static struct linux_binfmt elf_fdpic_format = {
  84         .module         = THIS_MODULE,
  85         .load_binary    = load_elf_fdpic_binary,
  86 #ifdef CONFIG_ELF_CORE
  87         .core_dump      = elf_fdpic_core_dump,
  88 #endif
  89         .min_coredump   = ELF_EXEC_PAGESIZE,
  90 };
  91
  92 static int __init init_elf_fdpic_binfmt(void)
  93 {
  94         register_binfmt(&elf_fdpic_format);
  95         return 0;
  96 }
  97
  98 static void __exit exit_elf_fdpic_binfmt(void)
  99 {
 100         unregister_binfmt(&elf_fdpic_format);
 101 }
 102
 103 core_initcall(init_elf_fdpic_binfmt);
 104 module_exit(exit_elf_fdpic_binfmt);
 105
 106 static int is_elf(struct elfhdr *hdr, struct file *file)
 107 {
 108         if (memcmp(hdr->e_ident, ELFMAG, SELFMAG) != 0)
 109                 return 0;
 110         if (hdr->e_type != ET_EXEC && hdr->e_type != ET_DYN)
 111                 return 0;
 112         if (!elf_check_arch(hdr))
 113                 return 0;
 114         if (!file->f_op->mmap)
 115                 return 0;
 116         return 1;
 117 }
 118
 119 #ifndef elf_check_fdpic
 120 #define elf_check_fdpic(x) 0
 121 #endif
 122
 123 #ifndef elf_check_const_displacement
 124 #define elf_check_const_displacement(x) 0
 125 #endif
 126
 127 static int is_constdisp(struct elfhdr *hdr)
 128 {
 129         if (!elf_check_fdpic(hdr))
 130                 return 1;
 131         if (elf_check_const_displacement(hdr))
 132                 return 1;
 133         return 0;
 134 }
 135
 136 /*****************************************************************************/
 137 /*
 138  * read the program headers table into memory
 139  */
 140 static int elf_fdpic_fetch_phdrs(struct elf_fdpic_params *params,
 141                                  struct file *file)
 142 {
 143         struct elf32_phdr *phdr;
 144         unsigned long size;
 145         int retval, loop;
 146
 147         if (params->hdr.e_phentsize != sizeof(struct elf_phdr))
 148                 return -ENOMEM;
 149         if (params->hdr.e_phnum > 65536U / sizeof(struct elf_phdr))
 150                 return -ENOMEM;
 151
 152         size = params->hdr.e_phnum * sizeof(struct elf_phdr);
 153         params->phdrs = kmalloc(size, GFP_KERNEL);
 154         if (!params->phdrs)
 155                 return -ENOMEM;
 156
 157         retval = kernel_read(file, params->hdr.e_phoff,
 158                              (char *) params->phdrs, size);
 159         if (unlikely(retval != size))
 160                 return retval < 0 ? retval : -ENOEXEC;
 161
 162         /* determine stack size for this binary */
 163         phdr = params->phdrs;
 164         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 165                 if (phdr->p_type != PT_GNU_STACK)
 166                         continue;
 167
 168                 if (phdr->p_flags & PF_X)
 169                         params->flags |= ELF_FDPIC_FLAG_EXEC_STACK;
 170                 else
 171                         params->flags |= ELF_FDPIC_FLAG_NOEXEC_STACK;
 172
 173                 params->stack_size = phdr->p_memsz;
 174                 break;
 175         }
 176
 177         return 0;
 178 }
 179
 180 /*****************************************************************************/
 181 /*
 182  * load an fdpic binary into various bits of memory
 183  */
 184 static int load_elf_fdpic_binary(struct linux_binprm *bprm)
 185 {
 186         struct elf_fdpic_params exec_params, interp_params;
 187         struct pt_regs *regs = current_pt_regs();
 188         struct elf_phdr *phdr;
 189         unsigned long stack_size, entryaddr;
 190 #ifdef ELF_FDPIC_PLAT_INIT
 191         unsigned long dynaddr;
 192 #endif
 193 #ifndef CONFIG_MMU
 194         unsigned long stack_prot;
 195 #endif
 196         struct file *interpreter = NULL; /* to shut gcc up */
 197         char *interpreter_name = NULL;
 198         int executable_stack;
 199         int retval, i;
 200
 201         kdebug("____ LOAD %d ____", current->pid);
 202
 203         memset(&exec_params, 0, sizeof(exec_params));
 204         memset(&interp_params, 0, sizeof(interp_params));
 205
 206         exec_params.hdr = *(struct elfhdr *) bprm->buf;
 207         exec_params.flags = ELF_FDPIC_FLAG_PRESENT | ELF_FDPIC_FLAG_EXECUTABLE;
 208
 209         /* check that this is a binary we know how to deal with */
 210         retval = -ENOEXEC;
 211         if (!is_elf(&exec_params.hdr, bprm->file))
 212                 goto error;
 213         if (!elf_check_fdpic(&exec_params.hdr)) {
 214 #ifdef CONFIG_MMU
 215                 /* binfmt_elf handles non-fdpic elf except on nommu */
 216                 goto error;
 217 #else
 218                 /* nommu can only load ET_DYN (PIE) ELF */
 219                 if (exec_params.hdr.e_type != ET_DYN)
 220                         goto error;
 221 #endif
 222         }
 223
 224         /* read the program header table */
 225         retval = elf_fdpic_fetch_phdrs(&exec_params, bprm->file);
 226         if (retval < 0)
 227                 goto error;
 228
 229         /* scan for a program header that specifies an interpreter */
 230         phdr = exec_params.phdrs;
 231
 232         for (i = 0; i < exec_params.hdr.e_phnum; i++, phdr++) {
 233                 switch (phdr->p_type) {
 234                 case PT_INTERP:
 235                         retval = -ENOMEM;
 236                         if (phdr->p_filesz > PATH_MAX)
 237                                 goto error;
 238                         retval = -ENOENT;
 239                         if (phdr->p_filesz < 2)
 240                                 goto error;
 241
 242                         /* read the name of the interpreter into memory */
 243                         interpreter_name = kmalloc(phdr->p_filesz, GFP_KERNEL);
 244                         if (!interpreter_name)
 245                                 goto error;
 246
 247                         retval = kernel_read(bprm->file,
 248                                              phdr->p_offset,
 249                                              interpreter_name,
 250                                              phdr->p_filesz);
 251                         if (unlikely(retval != phdr->p_filesz)) {
 252                                 if (retval >= 0)
 253                                         retval = -ENOEXEC;
 254                                 goto error;
 255                         }
 256
 257                         retval = -ENOENT;
 258                         if (interpreter_name[phdr->p_filesz - 1] != '\0')
 259                                 goto error;
 260
 261                         kdebug("Using ELF interpreter %s", interpreter_name);
 262
 263                         /* replace the program with the interpreter */
 264                         interpreter = open_exec(interpreter_name);
 265                         retval = PTR_ERR(interpreter);
 266                         if (IS_ERR(interpreter)) {
 267                                 interpreter = NULL;
 268                                 goto error;
 269                         }
 270
 271                         /*
 272                          * If the binary is not readable then enforce
 273                          * mm->dumpable = 0 regardless of the interpreter's
 274                          * permissions.
 275                          */
 276                         would_dump(bprm, interpreter);
 277
 278                         retval = kernel_read(interpreter, 0, bprm->buf,
 279                                              BINPRM_BUF_SIZE);
 280                         if (unlikely(retval != BINPRM_BUF_SIZE)) {
 281                                 if (retval >= 0)
 282                                         retval = -ENOEXEC;
 283                                 goto error;
 284                         }
 285
 286                         interp_params.hdr = *((struct elfhdr *) bprm->buf);
 287                         break;
 288
 289                 case PT_LOAD:
 290 #ifdef CONFIG_MMU
 291                         if (exec_params.load_addr == 0)
 292                                 exec_params.load_addr = phdr->p_vaddr;
 293 #endif
 294                         break;
 295                 }
 296
 297         }
 298
 299         if (is_constdisp(&exec_params.hdr))
 300                 exec_params.flags |= ELF_FDPIC_FLAG_CONSTDISP;
 301
 302         /* perform insanity checks on the interpreter */
 303         if (interpreter_name) {
 304                 retval = -ELIBBAD;
 305                 if (!is_elf(&interp_params.hdr, interpreter))
 306                         goto error;
 307
 308                 interp_params.flags = ELF_FDPIC_FLAG_PRESENT;
 309
 310                 /* read the interpreter's program header table */
 311                 retval = elf_fdpic_fetch_phdrs(&interp_params, interpreter);
 312                 if (retval < 0)
 313                         goto error;
 314         }
 315
 316         stack_size = exec_params.stack_size;
 317         if (exec_params.flags & ELF_FDPIC_FLAG_EXEC_STACK)
 318                 executable_stack = EXSTACK_ENABLE_X;
 319         else if (exec_params.flags & ELF_FDPIC_FLAG_NOEXEC_STACK)
 320                 executable_stack = EXSTACK_DISABLE_X;
 321         else
 322                 executable_stack = EXSTACK_DEFAULT;
 323
 324         if (stack_size == 0) {
 325                 stack_size = interp_params.stack_size;
 326                 if (interp_params.flags & ELF_FDPIC_FLAG_EXEC_STACK)
 327                         executable_stack = EXSTACK_ENABLE_X;
 328                 else if (interp_params.flags & ELF_FDPIC_FLAG_NOEXEC_STACK)
 329                         executable_stack = EXSTACK_DISABLE_X;
 330                 else
 331                         executable_stack = EXSTACK_DEFAULT;
 332         }
 333
 334         retval = -ENOEXEC;
 335         if (stack_size == 0)
 336                 stack_size = 131072UL; /* same as exec.c's default commit */
 337
 338         if (is_constdisp(&interp_params.hdr))
 339                 interp_params.flags |= ELF_FDPIC_FLAG_CONSTDISP;
 340
 341         /* flush all traces of the currently running executable */
 342         retval = flush_old_exec(bprm);
 343         if (retval)
 344                 goto error;
 345
 346         /* there's now no turning back... the old userspace image is dead,
 347          * defunct, deceased, etc.
 348          */
 349         if (elf_check_fdpic(&exec_params.hdr))
 350                 set_personality(PER_LINUX_FDPIC);
 351         else
 352                 set_personality(PER_LINUX);
 353         if (elf_read_implies_exec(&exec_params.hdr, executable_stack))
 354                 current->personality |= READ_IMPLIES_EXEC;
 355
 356         setup_new_exec(bprm);
 357
 358         set_binfmt(&elf_fdpic_format);
 359
 360         current->mm->start_code = 0;
 361         current->mm->end_code = 0;
 362         current->mm->start_stack = 0;
 363         current->mm->start_data = 0;
 364         current->mm->end_data = 0;
 365         current->mm->context.exec_fdpic_loadmap = 0;
 366         current->mm->context.interp_fdpic_loadmap = 0;
 367
 368 #ifdef CONFIG_MMU
 369         elf_fdpic_arch_lay_out_mm(&exec_params,
 370                                   &interp_params,
 371                                   &current->mm->start_stack,
 372                                   &current->mm->start_brk);
 373
 374         retval = setup_arg_pages(bprm, current->mm->start_stack,
 375                                  executable_stack);
 376         if (retval < 0)
 377                 goto error;
 378 #endif
 379
 380         /* load the executable and interpreter into memory */
 381         retval = elf_fdpic_map_file(&exec_params, bprm->file, current->mm,
 382                                     "executable");
 383         if (retval < 0)
 384                 goto error;
 385
 386         if (interpreter_name) {
 387                 retval = elf_fdpic_map_file(&interp_params, interpreter,
 388                                             current->mm, "interpreter");
 389                 if (retval < 0) {
 390                         printk(KERN_ERR "Unable to load interpreter\n");
 391                         goto error;
 392                 }
 393
 394                 allow_write_access(interpreter);
 395                 fput(interpreter);
 396                 interpreter = NULL;
 397         }
 398
 399 #ifdef CONFIG_MMU
 400         if (!current->mm->start_brk)
 401                 current->mm->start_brk = current->mm->end_data;
 402
 403         current->mm->brk = current->mm->start_brk =
 404                 PAGE_ALIGN(current->mm->start_brk);
 405
 406 #else
 407         /* create a stack area and zero-size brk area */
 408         stack_size = (stack_size + PAGE_SIZE - 1) & PAGE_MASK;
 409         if (stack_size < PAGE_SIZE * 2)
 410                 stack_size = PAGE_SIZE * 2;
 411
 412         stack_prot = PROT_READ | PROT_WRITE;
 413         if (executable_stack == EXSTACK_ENABLE_X ||
 414             (executable_stack == EXSTACK_DEFAULT && VM_STACK_FLAGS & VM_EXEC))
 415                 stack_prot |= PROT_EXEC;
 416
 417         current->mm->start_brk = vm_mmap(NULL, 0, stack_size, stack_prot,
 418                                          MAP_PRIVATE | MAP_ANONYMOUS |
 419                                          MAP_UNINITIALIZED | MAP_GROWSDOWN,
 420                                          0);
 421
 422         if (IS_ERR_VALUE(current->mm->start_brk)) {
 423                 retval = current->mm->start_brk;
 424                 current->mm->start_brk = 0;
 425                 goto error;
 426         }
 427
 428         current->mm->brk = current->mm->start_brk;
 429         current->mm->context.end_brk = current->mm->start_brk;
 430         current->mm->start_stack = current->mm->start_brk + stack_size;
 431 #endif
 432
 433         install_exec_creds(bprm);
 434         if (create_elf_fdpic_tables(bprm, current->mm,
 435                                     &exec_params, &interp_params) < 0)
 436                 goto error;
 437
 438         kdebug("- start_code  %lx", current->mm->start_code);
 439         kdebug("- end_code    %lx", current->mm->end_code);
 440         kdebug("- start_data  %lx", current->mm->start_data);
 441         kdebug("- end_data    %lx", current->mm->end_data);
 442         kdebug("- start_brk   %lx", current->mm->start_brk);
 443         kdebug("- brk         %lx", current->mm->brk);
 444         kdebug("- start_stack %lx", current->mm->start_stack);
 445
 446 #ifdef ELF_FDPIC_PLAT_INIT
 447         /*
 448          * The ABI may specify that certain registers be set up in special
 449          * ways (on i386 %edx is the address of a DT_FINI function, for
 450          * example.  This macro performs whatever initialization to
 451          * the regs structure is required.
 452          */
 453         dynaddr = interp_params.dynamic_addr ?: exec_params.dynamic_addr;
 454         ELF_FDPIC_PLAT_INIT(regs, exec_params.map_addr, interp_params.map_addr,
 455                             dynaddr);
 456 #endif
 457
 458         /* everything is now ready... get the userspace context ready to roll */
 459         entryaddr = interp_params.entry_addr ?: exec_params.entry_addr;
 460         start_thread(regs, entryaddr, current->mm->start_stack);
 461
 462         retval = 0;
 463
 464 error:
 465         if (interpreter) {
 466                 allow_write_access(interpreter);
 467                 fput(interpreter);
 468         }
 469         kfree(interpreter_name);
 470         kfree(exec_params.phdrs);
 471         kfree(exec_params.loadmap);
 472         kfree(interp_params.phdrs);
 473         kfree(interp_params.loadmap);
 474         return retval;
 475 }
 476
 477 /*****************************************************************************/
 478
 479 #ifndef ELF_BASE_PLATFORM
 480 /*
 481  * AT_BASE_PLATFORM indicates the "real" hardware/microarchitecture.
 482  * If the arch defines ELF_BASE_PLATFORM (in asm/elf.h), the value
 483  * will be copied to the user stack in the same manner as AT_PLATFORM.
 484  */
 485 #define ELF_BASE_PLATFORM NULL
 486 #endif
 487
 488 /*
 489  * present useful information to the program by shovelling it onto the new
 490  * process's stack
 491  */
 492 static int create_elf_fdpic_tables(struct linux_binprm *bprm,
 493                                    struct mm_struct *mm,
 494                                    struct elf_fdpic_params *exec_params,
 495                                    struct elf_fdpic_params *interp_params)
 496 {
 497         const struct cred *cred = current_cred();
 498         unsigned long sp, csp, nitems;
 499         elf_caddr_t __user *argv, *envp;
 500         size_t platform_len = 0, len;
 501         char *k_platform, *k_base_platform;
 502         char __user *u_platform, *u_base_platform, *p;
 503         int loop;
 504         int nr; /* reset for each csp adjustment */
 505
 506 #ifdef CONFIG_MMU
 507         /* In some cases (e.g. Hyper-Threading), we want to avoid L1 evictions
 508          * by the processes running on the same package. One thing we can do is
 509          * to shuffle the initial stack for them, so we give the architecture
 510          * an opportunity to do so here.
 511          */
 512         sp = arch_align_stack(bprm->p);
 513 #else
 514         sp = mm->start_stack;
 515
 516         /* stack the program arguments and environment */
 517         if (elf_fdpic_transfer_args_to_stack(bprm, &sp) < 0)
 518                 return -EFAULT;
 519 #endif
 520
 521         /*
 522          * If this architecture has a platform capability string, copy it
 523          * to userspace.  In some cases (Sparc), this info is impossible
 524          * for userspace to get any other way, in others (i386) it is
 525          * merely difficult.
 526          */
 527         k_platform = ELF_PLATFORM;
 528         u_platform = NULL;
 529
 530         if (k_platform) {
 531                 platform_len = strlen(k_platform) + 1;
 532                 sp -= platform_len;
 533                 u_platform = (char __user *) sp;
 534                 if (__copy_to_user(u_platform, k_platform, platform_len) != 0)
 535                         return -EFAULT;
 536         }
 537
 538         /*
 539          * If this architecture has a "base" platform capability
 540          * string, copy it to userspace.
 541          */
 542         k_base_platform = ELF_BASE_PLATFORM;
 543         u_base_platform = NULL;
 544
 545         if (k_base_platform) {
 546                 platform_len = strlen(k_base_platform) + 1;
 547                 sp -= platform_len;
 548                 u_base_platform = (char __user *) sp;
 549                 if (__copy_to_user(u_base_platform, k_base_platform, platform_len) != 0)
 550                         return -EFAULT;
 551         }
 552
 553         sp &= ~7UL;
 554
 555         /* stack the load map(s) */
 556         len = sizeof(struct elf32_fdpic_loadmap);
 557         len += sizeof(struct elf32_fdpic_loadseg) * exec_params->loadmap->nsegs;
 558         sp = (sp - len) & ~7UL;
 559         exec_params->map_addr = sp;
 560
 561         if (copy_to_user((void __user *) sp, exec_params->loadmap, len) != 0)
 562                 return -EFAULT;
 563
 564         current->mm->context.exec_fdpic_loadmap = (unsigned long) sp;
 565
 566         if (interp_params->loadmap) {
 567                 len = sizeof(struct elf32_fdpic_loadmap);
 568                 len += sizeof(struct elf32_fdpic_loadseg) *
 569                         interp_params->loadmap->nsegs;
 570                 sp = (sp - len) & ~7UL;
 571                 interp_params->map_addr = sp;
 572
 573                 if (copy_to_user((void __user *) sp, interp_params->loadmap,
 574                                  len) != 0)
 575                         return -EFAULT;
 576
 577                 current->mm->context.interp_fdpic_loadmap = (unsigned long) sp;
 578         }
 579
 580         /* force 16 byte _final_ alignment here for generality */
 581 #define DLINFO_ITEMS 15
 582
 583         nitems = 1 + DLINFO_ITEMS + (k_platform ? 1 : 0) +
 584                 (k_base_platform ? 1 : 0) + AT_VECTOR_SIZE_ARCH;
 585
 586         if (bprm->interp_flags & BINPRM_FLAGS_EXECFD)
 587                 nitems++;
 588
 589         csp = sp;
 590         sp -= nitems * 2 * sizeof(unsigned long);
 591         sp -= (bprm->envc + 1) * sizeof(char *);        /* envv[] */
 592         sp -= (bprm->argc + 1) * sizeof(char *);        /* argv[] */
 593         sp -= 1 * sizeof(unsigned long);                /* argc */
 594
 595         csp -= sp & 15UL;
 596         sp -= sp & 15UL;
 597
 598         /* put the ELF interpreter info on the stack */
 599 #define NEW_AUX_ENT(id, val)                                            \
 600         do {                                                            \
 601                 struct { unsigned long _id, _val; } __user *ent;        \
 602                                                                         \
 603                 ent = (void __user *) csp;                              \
 604                 __put_user((id), &ent[nr]._id);                         \
 605                 __put_user((val), &ent[nr]._val);                       \
 606                 nr++;                                                   \
 607         } while (0)
 608
 609         nr = 0;
 610         csp -= 2 * sizeof(unsigned long);
 611         NEW_AUX_ENT(AT_NULL, 0);
 612         if (k_platform) {
 613                 nr = 0;
 614                 csp -= 2 * sizeof(unsigned long);
 615                 NEW_AUX_ENT(AT_PLATFORM,
 616                             (elf_addr_t) (unsigned long) u_platform);
 617         }
 618
 619         if (k_base_platform) {
 620                 nr = 0;
 621                 csp -= 2 * sizeof(unsigned long);
 622                 NEW_AUX_ENT(AT_BASE_PLATFORM,
 623                             (elf_addr_t) (unsigned long) u_base_platform);
 624         }
 625
 626         if (bprm->interp_flags & BINPRM_FLAGS_EXECFD) {
 627                 nr = 0;
 628                 csp -= 2 * sizeof(unsigned long);
 629                 NEW_AUX_ENT(AT_EXECFD, bprm->interp_data);
 630         }
 631
 632         nr = 0;
 633         csp -= DLINFO_ITEMS * 2 * sizeof(unsigned long);
 634         NEW_AUX_ENT(AT_HWCAP,   ELF_HWCAP);
 635 #ifdef ELF_HWCAP2
 636         NEW_AUX_ENT(AT_HWCAP2,  ELF_HWCAP2);
 637 #endif
 638         NEW_AUX_ENT(AT_PAGESZ,  PAGE_SIZE);
 639         NEW_AUX_ENT(AT_CLKTCK,  CLOCKS_PER_SEC);
 640         NEW_AUX_ENT(AT_PHDR,    exec_params->ph_addr);
 641         NEW_AUX_ENT(AT_PHENT,   sizeof(struct elf_phdr));
 642         NEW_AUX_ENT(AT_PHNUM,   exec_params->hdr.e_phnum);
 643         NEW_AUX_ENT(AT_BASE,    interp_params->elfhdr_addr);
 644         NEW_AUX_ENT(AT_FLAGS,   0);
 645         NEW_AUX_ENT(AT_ENTRY,   exec_params->entry_addr);
 646         NEW_AUX_ENT(AT_UID,     (elf_addr_t) from_kuid_munged(cred->user_ns, cred->uid));
 647         NEW_AUX_ENT(AT_EUID,    (elf_addr_t) from_kuid_munged(cred->user_ns, cred->euid));
 648         NEW_AUX_ENT(AT_GID,     (elf_addr_t) from_kgid_munged(cred->user_ns, cred->gid));
 649         NEW_AUX_ENT(AT_EGID,    (elf_addr_t) from_kgid_munged(cred->user_ns, cred->egid));
 650         NEW_AUX_ENT(AT_SECURE,  security_bprm_secureexec(bprm));
 651         NEW_AUX_ENT(AT_EXECFN,  bprm->exec);
 652
 653 #ifdef ARCH_DLINFO
 654         nr = 0;
 655         csp -= AT_VECTOR_SIZE_ARCH * 2 * sizeof(unsigned long);
 656
 657         /* ARCH_DLINFO must come last so platform specific code can enforce
 658          * special alignment requirements on the AUXV if necessary (eg. PPC).
 659          */
 660         ARCH_DLINFO;
 661 #endif
 662 #undef NEW_AUX_ENT
 663
 664         /* allocate room for argv[] and envv[] */
 665         csp -= (bprm->envc + 1) * sizeof(elf_caddr_t);
 666         envp = (elf_caddr_t __user *) csp;
 667         csp -= (bprm->argc + 1) * sizeof(elf_caddr_t);
 668         argv = (elf_caddr_t __user *) csp;
 669
 670         /* stack argc */
 671         csp -= sizeof(unsigned long);
 672         __put_user(bprm->argc, (unsigned long __user *) csp);
 673
 674         BUG_ON(csp != sp);
 675
 676         /* fill in the argv[] array */
 677 #ifdef CONFIG_MMU
 678         current->mm->arg_start = bprm->p;
 679 #else
 680         current->mm->arg_start = current->mm->start_stack -
 681                 (MAX_ARG_PAGES * PAGE_SIZE - bprm->p);
 682 #endif
 683
 684         p = (char __user *) current->mm->arg_start;
 685         for (loop = bprm->argc; loop > 0; loop--) {
 686                 __put_user((elf_caddr_t) p, argv++);
 687                 len = strnlen_user(p, MAX_ARG_STRLEN);
 688                 if (!len || len > MAX_ARG_STRLEN)
 689                         return -EINVAL;
 690                 p += len;
 691         }
 692         __put_user(NULL, argv);
 693         current->mm->arg_end = (unsigned long) p;
 694
 695         /* fill in the envv[] array */
 696         current->mm->env_start = (unsigned long) p;
 697         for (loop = bprm->envc; loop > 0; loop--) {
 698                 __put_user((elf_caddr_t)(unsigned long) p, envp++);
 699                 len = strnlen_user(p, MAX_ARG_STRLEN);
 700                 if (!len || len > MAX_ARG_STRLEN)
 701                         return -EINVAL;
 702                 p += len;
 703         }
 704         __put_user(NULL, envp);
 705         current->mm->env_end = (unsigned long) p;
 706
 707         mm->start_stack = (unsigned long) sp;
 708         return 0;
 709 }
 710
 711 /*****************************************************************************/
 712 /*
 713  * transfer the program arguments and environment from the holding pages onto
 714  * the stack
 715  */
 716 #ifndef CONFIG_MMU
 717 static int elf_fdpic_transfer_args_to_stack(struct linux_binprm *bprm,
 718                                             unsigned long *_sp)
 719 {
 720         unsigned long index, stop, sp;
 721         char *src;
 722         int ret = 0;
 723
 724         stop = bprm->p >> PAGE_SHIFT;
 725         sp = *_sp;
 726
 727         for (index = MAX_ARG_PAGES - 1; index >= stop; index--) {
 728                 src = kmap(bprm->page[index]);
 729                 sp -= PAGE_SIZE;
 730                 if (copy_to_user((void *) sp, src, PAGE_SIZE) != 0)
 731                         ret = -EFAULT;
 732                 kunmap(bprm->page[index]);
 733                 if (ret < 0)
 734                         goto out;
 735         }
 736
 737         *_sp = (*_sp - (MAX_ARG_PAGES * PAGE_SIZE - bprm->p)) & ~15;
 738
 739 out:
 740         return ret;
 741 }
 742 #endif
 743
 744 /*****************************************************************************/
 745 /*
 746  * load the appropriate binary image (executable or interpreter) into memory
 747  * - we assume no MMU is available
 748  * - if no other PIC bits are set in params->hdr->e_flags
 749  *   - we assume that the LOADable segments in the binary are independently relocatable
 750  *   - we assume R/O executable segments are shareable
 751  * - else
 752  *   - we assume the loadable parts of the image to require fixed displacement
 753  *   - the image is not shareable
 754  */
 755 static int elf_fdpic_map_file(struct elf_fdpic_params *params,
 756                               struct file *file,
 757                               struct mm_struct *mm,
 758                               const char *what)
 759 {
 760         struct elf32_fdpic_loadmap *loadmap;
 761 #ifdef CONFIG_MMU
 762         struct elf32_fdpic_loadseg *mseg;
 763 #endif
 764         struct elf32_fdpic_loadseg *seg;
 765         struct elf32_phdr *phdr;
 766         unsigned long load_addr, stop;
 767         unsigned nloads, tmp;
 768         size_t size;
 769         int loop, ret;
 770
 771         /* allocate a load map table */
 772         nloads = 0;
 773         for (loop = 0; loop < params->hdr.e_phnum; loop++)
 774                 if (params->phdrs[loop].p_type == PT_LOAD)
 775                         nloads++;
 776
 777         if (nloads == 0)
 778                 return -ELIBBAD;
 779
 780         size = sizeof(*loadmap) + nloads * sizeof(*seg);
 781         loadmap = kzalloc(size, GFP_KERNEL);
 782         if (!loadmap)
 783                 return -ENOMEM;
 784
 785         params->loadmap = loadmap;
 786
 787         loadmap->version = ELF32_FDPIC_LOADMAP_VERSION;
 788         loadmap->nsegs = nloads;
 789
 790         load_addr = params->load_addr;
 791         seg = loadmap->segs;
 792
 793         /* map the requested LOADs into the memory space */
 794         switch (params->flags & ELF_FDPIC_FLAG_ARRANGEMENT) {
 795         case ELF_FDPIC_FLAG_CONSTDISP:
 796         case ELF_FDPIC_FLAG_CONTIGUOUS:
 797 #ifndef CONFIG_MMU
 798                 ret = elf_fdpic_map_file_constdisp_on_uclinux(params, file, mm);
 799                 if (ret < 0)
 800                         return ret;
 801                 break;
 802 #endif
 803         default:
 804                 ret = elf_fdpic_map_file_by_direct_mmap(params, file, mm);
 805                 if (ret < 0)
 806                         return ret;
 807                 break;
 808         }
 809
 810         /* map the entry point */
 811         if (params->hdr.e_entry) {
 812                 seg = loadmap->segs;
 813                 for (loop = loadmap->nsegs; loop > 0; loop--, seg++) {
 814                         if (params->hdr.e_entry >= seg->p_vaddr &&
 815                             params->hdr.e_entry < seg->p_vaddr + seg->p_memsz) {
 816                                 params->entry_addr =
 817                                         (params->hdr.e_entry - seg->p_vaddr) +
 818                                         seg->addr;
 819                                 break;
 820                         }
 821                 }
 822         }
 823
 824         /* determine where the program header table has wound up if mapped */
 825         stop = params->hdr.e_phoff;
 826         stop += params->hdr.e_phnum * sizeof (struct elf_phdr);
 827         phdr = params->phdrs;
 828
 829         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 830                 if (phdr->p_type != PT_LOAD)
 831                         continue;
 832
 833                 if (phdr->p_offset > params->hdr.e_phoff ||
 834                     phdr->p_offset + phdr->p_filesz < stop)
 835                         continue;
 836
 837                 seg = loadmap->segs;
 838                 for (loop = loadmap->nsegs; loop > 0; loop--, seg++) {
 839                         if (phdr->p_vaddr >= seg->p_vaddr &&
 840                             phdr->p_vaddr + phdr->p_filesz <=
 841                             seg->p_vaddr + seg->p_memsz) {
 842                                 params->ph_addr =
 843                                         (phdr->p_vaddr - seg->p_vaddr) +
 844                                         seg->addr +
 845                                         params->hdr.e_phoff - phdr->p_offset;
 846                                 break;
 847                         }
 848                 }
 849                 break;
 850         }
 851
 852         /* determine where the dynamic section has wound up if there is one */
 853         phdr = params->phdrs;
 854         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 855                 if (phdr->p_type != PT_DYNAMIC)
 856                         continue;
 857
 858                 seg = loadmap->segs;
 859                 for (loop = loadmap->nsegs; loop > 0; loop--, seg++) {
 860                         if (phdr->p_vaddr >= seg->p_vaddr &&
 861                             phdr->p_vaddr + phdr->p_memsz <=
 862                             seg->p_vaddr + seg->p_memsz) {
 863                                 params->dynamic_addr =
 864                                         (phdr->p_vaddr - seg->p_vaddr) +
 865                                         seg->addr;
 866
 867                                 /* check the dynamic section contains at least
 868                                  * one item, and that the last item is a NULL
 869                                  * entry */
 870                                 if (phdr->p_memsz == 0 ||
 871                                     phdr->p_memsz % sizeof(Elf32_Dyn) != 0)
 872                                         goto dynamic_error;
 873
 874                                 tmp = phdr->p_memsz / sizeof(Elf32_Dyn);
 875                                 if (((Elf32_Dyn *)
 876                                      params->dynamic_addr)[tmp - 1].d_tag != 0)
 877                                         goto dynamic_error;
 878                                 break;
 879                         }
 880                 }
 881                 break;
 882         }
 883
 884         /* now elide adjacent segments in the load map on MMU linux
 885          * - on uClinux the holes between may actually be filled with system
 886          *   stuff or stuff from other processes
 887          */
 888 #ifdef CONFIG_MMU
 889         nloads = loadmap->nsegs;
 890         mseg = loadmap->segs;
 891         seg = mseg + 1;
 892         for (loop = 1; loop < nloads; loop++) {
 893                 /* see if we have a candidate for merging */
 894                 if (seg->p_vaddr - mseg->p_vaddr == seg->addr - mseg->addr) {
 895                         load_addr = PAGE_ALIGN(mseg->addr + mseg->p_memsz);
 896                         if (load_addr == (seg->addr & PAGE_MASK)) {
 897                                 mseg->p_memsz +=
 898                                         load_addr -
 899                                         (mseg->addr + mseg->p_memsz);
 900                                 mseg->p_memsz += seg->addr & ~PAGE_MASK;
 901                                 mseg->p_memsz += seg->p_memsz;
 902                                 loadmap->nsegs--;
 903                                 continue;
 904                         }
 905                 }
 906
 907                 mseg++;
 908                 if (mseg != seg)
 909                         *mseg = *seg;
 910         }
 911 #endif
 912
 913         kdebug("Mapped Object [%s]:", what);
 914         kdebug("- elfhdr   : %lx", params->elfhdr_addr);
 915         kdebug("- entry    : %lx", params->entry_addr);
 916         kdebug("- PHDR[]   : %lx", params->ph_addr);
 917         kdebug("- DYNAMIC[]: %lx", params->dynamic_addr);
 918         seg = loadmap->segs;
 919         for (loop = 0; loop < loadmap->nsegs; loop++, seg++)
 920                 kdebug("- LOAD[%d] : %08x-%08x [va=%x ms=%x]",
 921                        loop,
 922                        seg->addr, seg->addr + seg->p_memsz - 1,
 923                        seg->p_vaddr, seg->p_memsz);
 924
 925         return 0;
 926
 927 dynamic_error:
 928         printk("ELF FDPIC %s with invalid DYNAMIC section (inode=%lu)\n",
 929                what, file_inode(file)->i_ino);
 930         return -ELIBBAD;
 931 }
 932
 933 /*****************************************************************************/
 934 /*
 935  * map a file with constant displacement under uClinux
 936  */
 937 #ifndef CONFIG_MMU
 938 static int elf_fdpic_map_file_constdisp_on_uclinux(
 939         struct elf_fdpic_params *params,
 940         struct file *file,
 941         struct mm_struct *mm)
 942 {
 943         struct elf32_fdpic_loadseg *seg;
 944         struct elf32_phdr *phdr;
 945         unsigned long load_addr, base = ULONG_MAX, top = 0, maddr = 0, mflags;
 946         int loop, ret;
 947
 948         load_addr = params->load_addr;
 949         seg = params->loadmap->segs;
 950
 951         /* determine the bounds of the contiguous overall allocation we must
 952          * make */
 953         phdr = params->phdrs;
 954         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 955                 if (params->phdrs[loop].p_type != PT_LOAD)
 956                         continue;
 957
 958                 if (base > phdr->p_vaddr)
 959                         base = phdr->p_vaddr;
 960                 if (top < phdr->p_vaddr + phdr->p_memsz)
 961                         top = phdr->p_vaddr + phdr->p_memsz;
 962         }
 963
 964         /* allocate one big anon block for everything */
 965         mflags = MAP_PRIVATE;
 966         if (params->flags & ELF_FDPIC_FLAG_EXECUTABLE)
 967                 mflags |= MAP_EXECUTABLE;
 968
 969         maddr = vm_mmap(NULL, load_addr, top - base,
 970                         PROT_READ | PROT_WRITE | PROT_EXEC, mflags, 0);
 971         if (IS_ERR_VALUE(maddr))
 972                 return (int) maddr;
 973
 974         if (load_addr != 0)
 975                 load_addr += PAGE_ALIGN(top - base);
 976
 977         /* and then load the file segments into it */
 978         phdr = params->phdrs;
 979         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
 980                 if (params->phdrs[loop].p_type != PT_LOAD)
 981                         continue;
 982
 983                 seg->addr = maddr + (phdr->p_vaddr - base);
 984                 seg->p_vaddr = phdr->p_vaddr;
 985                 seg->p_memsz = phdr->p_memsz;
 986
 987                 ret = read_code(file, seg->addr, phdr->p_offset,
 988                                        phdr->p_filesz);
 989                 if (ret < 0)
 990                         return ret;
 991
 992                 /* map the ELF header address if in this segment */
 993                 if (phdr->p_offset == 0)
 994                         params->elfhdr_addr = seg->addr;
 995
 996                 /* clear any space allocated but not loaded */
 997                 if (phdr->p_filesz < phdr->p_memsz) {
 998                         if (clear_user((void *) (seg->addr + phdr->p_filesz),
 999                                        phdr->p_memsz - phdr->p_filesz))
1000                                 return -EFAULT;
1001                 }
1002
1003                 if (mm) {
1004                         if (phdr->p_flags & PF_X) {
1005                                 if (!mm->start_code) {
1006                                         mm->start_code = seg->addr;
1007                                         mm->end_code = seg->addr +
1008                                                 phdr->p_memsz;
1009                                 }
1010                         } else if (!mm->start_data) {
1011                                 mm->start_data = seg->addr;
1012                                 mm->end_data = seg->addr + phdr->p_memsz;
1013                         }
1014                 }
1015
1016                 seg++;
1017         }
1018
1019         return 0;
1020 }
1021 #endif
1022
1023 /*****************************************************************************/
1024 /*
1025  * map a binary by direct mmap() of the individual PT_LOAD segments
1026  */
1027 static int elf_fdpic_map_file_by_direct_mmap(struct elf_fdpic_params *params,
1028                                              struct file *file,
1029                                              struct mm_struct *mm)
1030 {
1031         struct elf32_fdpic_loadseg *seg;
1032         struct elf32_phdr *phdr;
1033         unsigned long load_addr, delta_vaddr;
1034         int loop, dvset;
1035
1036         load_addr = params->load_addr;
1037         delta_vaddr = 0;
1038         dvset = 0;
1039
1040         seg = params->loadmap->segs;
1041
1042         /* deal with each load segment separately */
1043         phdr = params->phdrs;
1044         for (loop = 0; loop < params->hdr.e_phnum; loop++, phdr++) {
1045                 unsigned long maddr, disp, excess, excess1;
1046                 int prot = 0, flags;
1047
1048                 if (phdr->p_type != PT_LOAD)
1049                         continue;
1050
1051                 kdebug("[LOAD] va=%lx of=%lx fs=%lx ms=%lx",
1052                        (unsigned long) phdr->p_vaddr,
1053                        (unsigned long) phdr->p_offset,
1054                        (unsigned long) phdr->p_filesz,
1055                        (unsigned long) phdr->p_memsz);
1056
1057                 /* determine the mapping parameters */
1058                 if (phdr->p_flags & PF_R) prot |= PROT_READ;
1059                 if (phdr->p_flags & PF_W) prot |= PROT_WRITE;
1060                 if (phdr->p_flags & PF_X) prot |= PROT_EXEC;
1061
1062                 flags = MAP_PRIVATE | MAP_DENYWRITE;
1063                 if (params->flags & ELF_FDPIC_FLAG_EXECUTABLE)
1064                         flags |= MAP_EXECUTABLE;
1065
1066                 maddr = 0;
1067
1068                 switch (params->flags & ELF_FDPIC_FLAG_ARRANGEMENT) {
1069                 case ELF_FDPIC_FLAG_INDEPENDENT:
1070                         /* PT_LOADs are independently locatable */
1071                         break;
1072
1073                 case ELF_FDPIC_FLAG_HONOURVADDR:
1074                         /* the specified virtual address must be honoured */
1075                         maddr = phdr->p_vaddr;
1076                         flags |= MAP_FIXED;
1077                         break;
1078
1079                 case ELF_FDPIC_FLAG_CONSTDISP:
1080                         /* constant displacement
1081                          * - can be mapped anywhere, but must be mapped as a
1082                          *   unit
1083                          */
1084                         if (!dvset) {
1085                                 maddr = load_addr;
1086                                 delta_vaddr = phdr->p_vaddr;
1087                                 dvset = 1;
1088                         } else {
1089                                 maddr = load_addr + phdr->p_vaddr - delta_vaddr;
1090                                 flags |= MAP_FIXED;
1091                         }
1092                         break;
1093
1094                 case ELF_FDPIC_FLAG_CONTIGUOUS:
1095                         /* contiguity handled later */
1096                         break;
1097
1098                 default:
1099                         BUG();
1100                 }
1101
1102                 maddr &= PAGE_MASK;
1103
1104                 /* create the mapping */
1105                 disp = phdr->p_vaddr & ~PAGE_MASK;
1106                 maddr = vm_mmap(file, maddr, phdr->p_memsz + disp, prot, flags,
1107                                 phdr->p_offset - disp);
1108
1109                 kdebug("mmap[%d] <file> sz=%lx pr=%x fl=%x of=%lx --> %08lx",
1110                        loop, phdr->p_memsz + disp, prot, flags,
1111                        phdr->p_offset - disp, maddr);
1112
1113                 if (IS_ERR_VALUE(maddr))
1114                         return (int) maddr;
1115
1116                 if ((params->flags & ELF_FDPIC_FLAG_ARRANGEMENT) ==
1117                     ELF_FDPIC_FLAG_CONTIGUOUS)
1118                         load_addr += PAGE_ALIGN(phdr->p_memsz + disp);
1119
1120                 seg->addr = maddr + disp;
1121                 seg->p_vaddr = phdr->p_vaddr;
1122                 seg->p_memsz = phdr->p_memsz;
1123
1124                 /* map the ELF header address if in this segment */
1125                 if (phdr->p_offset == 0)
1126                         params->elfhdr_addr = seg->addr;
1127
1128                 /* clear the bit between beginning of mapping and beginning of
1129                  * PT_LOAD */
1130                 if (prot & PROT_WRITE && disp > 0) {
1131                         kdebug("clear[%d] ad=%lx sz=%lx", loop, maddr, disp);
1132                         if (clear_user((void __user *) maddr, disp))
1133                                 return -EFAULT;
1134                         maddr += disp;
1135                 }
1136
1137                 /* clear any space allocated but not loaded
1138                  * - on uClinux we can just clear the lot
1139                  * - on MMU linux we'll get a SIGBUS beyond the last page
1140                  *   extant in the file
1141                  */
1142                 excess = phdr->p_memsz - phdr->p_filesz;
1143                 excess1 = PAGE_SIZE - ((maddr + phdr->p_filesz) & ~PAGE_MASK);
1144
1145 #ifdef CONFIG_MMU
1146                 if (excess > excess1) {
1147                         unsigned long xaddr = maddr + phdr->p_filesz + excess1;
1148                         unsigned long xmaddr;
1149
1150                         flags |= MAP_FIXED | MAP_ANONYMOUS;
1151                         xmaddr = vm_mmap(NULL, xaddr, excess - excess1,
1152                                          prot, flags, 0);
1153
1154                         kdebug("mmap[%d] <anon>"
1155                                " ad=%lx sz=%lx pr=%x fl=%x of=0 --> %08lx",
1156                                loop, xaddr, excess - excess1, prot, flags,
1157                                xmaddr);
1158
1159                         if (xmaddr != xaddr)
1160                                 return -ENOMEM;
1161                 }
1162
1163                 if (prot & PROT_WRITE && excess1 > 0) {
1164                         kdebug("clear[%d] ad=%lx sz=%lx",
1165                                loop, maddr + phdr->p_filesz, excess1);
1166                         if (clear_user((void __user *) maddr + phdr->p_filesz,
1167                                        excess1))
1168                                 return -EFAULT;
1169                 }
1170
1171 #else
1172                 if (excess > 0) {
1173                         kdebug("clear[%d] ad=%lx sz=%lx",
1174                                loop, maddr + phdr->p_filesz, excess);
1175                         if (clear_user((void *) maddr + phdr->p_filesz, excess))
1176                                 return -EFAULT;
1177                 }
1178 #endif
1179
1180                 if (mm) {
1181                         if (phdr->p_flags & PF_X) {
1182                                 if (!mm->start_code) {
1183                                         mm->start_code = maddr;
1184                                         mm->end_code = maddr + phdr->p_memsz;
1185                                 }
1186                         } else if (!mm->start_data) {
1187                                 mm->start_data = maddr;
1188                                 mm->end_data = maddr + phdr->p_memsz;
1189                         }
1190                 }
1191
1192                 seg++;
1193         }
1194
1195         return 0;
1196 }
1197
1198 /*****************************************************************************/
1199 /*
1200  * ELF-FDPIC core dumper
1201  *
1202  * Modelled on fs/exec.c:aout_core_dump()
1203  * Jeremy Fitzhardinge <jeremy@sw.oz.au>
1204  *
1205  * Modelled on fs/binfmt_elf.c core dumper
1206  */
1207 #ifdef CONFIG_ELF_CORE
1208
1209 /*
1210  * Decide whether a segment is worth dumping; default is yes to be
1211  * sure (missing info is worse than too much; etc).
1212  * Personally I'd include everything, and use the coredump limit...
1213  *
1214  * I think we should skip something. But I am not sure how. H.J.
1215  */
1216 static int maydump(struct vm_area_struct *vma, unsigned long mm_flags)
1217 {
1218         int dump_ok;
1219
1220         /* Do not dump I/O mapped devices or special mappings */
1221         if (vma->vm_flags & VM_IO) {
1222                 kdcore("%08lx: %08lx: no (IO)", vma->vm_start, vma->vm_flags);
1223                 return 0;
1224         }
1225
1226         /* If we may not read the contents, don't allow us to dump
1227          * them either. "dump_write()" can't handle it anyway.
1228          */
1229         if (!(vma->vm_flags & VM_READ)) {
1230                 kdcore("%08lx: %08lx: no (!read)", vma->vm_start, vma->vm_flags);
1231                 return 0;
1232         }
1233
1234         /* By default, dump shared memory if mapped from an anonymous file. */
1235         if (vma->vm_flags & VM_SHARED) {
1236                 if (file_inode(vma->vm_file)->i_nlink == 0) {
1237                         dump_ok = test_bit(MMF_DUMP_ANON_SHARED, &mm_flags);
1238                         kdcore("%08lx: %08lx: %s (share)", vma->vm_start,
1239                                vma->vm_flags, dump_ok ? "yes" : "no");
1240                         return dump_ok;
1241                 }
1242
1243                 dump_ok = test_bit(MMF_DUMP_MAPPED_SHARED, &mm_flags);
1244                 kdcore("%08lx: %08lx: %s (share)", vma->vm_start,
1245                        vma->vm_flags, dump_ok ? "yes" : "no");
1246                 return dump_ok;
1247         }
1248
1249 #ifdef CONFIG_MMU
1250         /* By default, if it hasn't been written to, don't write it out */
1251         if (!vma->anon_vma) {
1252                 dump_ok = test_bit(MMF_DUMP_MAPPED_PRIVATE, &mm_flags);
1253                 kdcore("%08lx: %08lx: %s (!anon)", vma->vm_start,
1254                        vma->vm_flags, dump_ok ? "yes" : "no");
1255                 return dump_ok;
1256         }
1257 #endif
1258
1259         dump_ok = test_bit(MMF_DUMP_ANON_PRIVATE, &mm_flags);
1260         kdcore("%08lx: %08lx: %s", vma->vm_start, vma->vm_flags,
1261                dump_ok ? "yes" : "no");
1262         return dump_ok;
1263 }
1264
1265 /* An ELF note in memory */
1266 struct memelfnote
1267 {
1268         const char *name;
1269         int type;
1270         unsigned int datasz;
1271         void *data;
1272 };
1273
1274 static int notesize(struct memelfnote *en)
1275 {
1276         int sz;
1277
1278         sz = sizeof(struct elf_note);
1279         sz += roundup(strlen(en->name) + 1, 4);
1280         sz += roundup(en->datasz, 4);
1281
1282         return sz;
1283 }
1284
1285 /* #define DEBUG */
1286
1287 static int writenote(struct memelfnote *men, struct coredump_params *cprm)
1288 {
1289         struct elf_note en;
1290         en.n_namesz = strlen(men->name) + 1;
1291         en.n_descsz = men->datasz;
1292         en.n_type = men->type;
1293
1294         return dump_emit(cprm, &en, sizeof(en)) &&
1295                 dump_emit(cprm, men->name, en.n_namesz) && dump_align(cprm, 4) &&
1296                 dump_emit(cprm, men->data, men->datasz) && dump_align(cprm, 4);
1297 }
1298
1299 static inline void fill_elf_fdpic_header(struct elfhdr *elf, int segs)
1300 {
1301         memcpy(elf->e_ident, ELFMAG, SELFMAG);
1302         elf->e_ident[EI_CLASS] = ELF_CLASS;
1303         elf->e_ident[EI_DATA] = ELF_DATA;
1304         elf->e_ident[EI_VERSION] = EV_CURRENT;
1305         elf->e_ident[EI_OSABI] = ELF_OSABI;
1306         memset(elf->e_ident+EI_PAD, 0, EI_NIDENT-EI_PAD);
1307
1308         elf->e_type = ET_CORE;
1309         elf->e_machine = ELF_ARCH;
1310         elf->e_version = EV_CURRENT;
1311         elf->e_entry = 0;
1312         elf->e_phoff = sizeof(struct elfhdr);
1313         elf->e_shoff = 0;
1314         elf->e_flags = ELF_FDPIC_CORE_EFLAGS;
1315         elf->e_ehsize = sizeof(struct elfhdr);
1316         elf->e_phentsize = sizeof(struct elf_phdr);
1317         elf->e_phnum = segs;
1318         elf->e_shentsize = 0;
1319         elf->e_shnum = 0;
1320         elf->e_shstrndx = 0;
1321         return;
1322 }
1323
1324 static inline void fill_elf_note_phdr(struct elf_phdr *phdr, int sz, loff_t offset)
1325 {
1326         phdr->p_type = PT_NOTE;
1327         phdr->p_offset = offset;
1328         phdr->p_vaddr = 0;
1329         phdr->p_paddr = 0;
1330         phdr->p_filesz = sz;
1331         phdr->p_memsz = 0;
1332         phdr->p_flags = 0;
1333         phdr->p_align = 0;
1334         return;
1335 }
1336
1337 static inline void fill_note(struct memelfnote *note, const char *name, int type,
1338                 unsigned int sz, void *data)
1339 {
1340         note->name = name;
1341         note->type = type;
1342         note->datasz = sz;
1343         note->data = data;
1344         return;
1345 }
1346
1347 /*
1348  * fill up all the fields in prstatus from the given task struct, except
1349  * registers which need to be filled up separately.
1350  */
1351 static void fill_prstatus(struct elf_prstatus *prstatus,
1352                           struct task_struct *p, long signr)
1353 {
1354         prstatus->pr_info.si_signo = prstatus->pr_cursig = signr;
1355         prstatus->pr_sigpend = p->pending.signal.sig[0];
1356         prstatus->pr_sighold = p->blocked.sig[0];
1357         rcu_read_lock();
1358         prstatus->pr_ppid = task_pid_vnr(rcu_dereference(p->real_parent));
1359         rcu_read_unlock();
1360         prstatus->pr_pid = task_pid_vnr(p);
1361         prstatus->pr_pgrp = task_pgrp_vnr(p);
1362         prstatus->pr_sid = task_session_vnr(p);
1363         if (thread_group_leader(p)) {
1364                 struct task_cputime cputime;
1365
1366                 /*
1367                  * This is the record for the group leader.  It shows the
1368                  * group-wide total, not its individual thread total.
1369                  */
1370                 thread_group_cputime(p, &cputime);
1371                 cputime_to_timeval(cputime.utime, &prstatus->pr_utime);
1372                 cputime_to_timeval(cputime.stime, &prstatus->pr_stime);
1373         } else {
1374                 cputime_t utime, stime;
1375
1376                 task_cputime(p, &utime, &stime);
1377                 cputime_to_timeval(utime, &prstatus->pr_utime);
1378                 cputime_to_timeval(stime, &prstatus->pr_stime);
1379         }
1380         cputime_to_timeval(p->signal->cutime, &prstatus->pr_cutime);
1381         cputime_to_timeval(p->signal->cstime, &prstatus->pr_cstime);
1382
1383         prstatus->pr_exec_fdpic_loadmap = p->mm->context.exec_fdpic_loadmap;
1384         prstatus->pr_interp_fdpic_loadmap = p->mm->context.interp_fdpic_loadmap;
1385 }
1386
1387 static int fill_psinfo(struct elf_prpsinfo *psinfo, struct task_struct *p,
1388                        struct mm_struct *mm)
1389 {
1390         const struct cred *cred;
1391         unsigned int i, len;
1392
1393         /* first copy the parameters from user space */
1394         memset(psinfo, 0, sizeof(struct elf_prpsinfo));
1395
1396         len = mm->arg_end - mm->arg_start;
1397         if (len >= ELF_PRARGSZ)
1398                 len = ELF_PRARGSZ - 1;
1399         if (copy_from_user(&psinfo->pr_psargs,
1400                            (const char __user *) mm->arg_start, len))
1401                 return -EFAULT;
1402         for (i = 0; i < len; i++)
1403                 if (psinfo->pr_psargs[i] == 0)
1404                         psinfo->pr_psargs[i] = ' ';
1405         psinfo->pr_psargs[len] = 0;
1406
1407         rcu_read_lock();
1408         psinfo->pr_ppid = task_pid_vnr(rcu_dereference(p->real_parent));
1409         rcu_read_unlock();
1410         psinfo->pr_pid = task_pid_vnr(p);
1411         psinfo->pr_pgrp = task_pgrp_vnr(p);
1412         psinfo->pr_sid = task_session_vnr(p);
1413
1414         i = p->state ? ffz(~p->state) + 1 : 0;
1415         psinfo->pr_state = i;
1416         psinfo->pr_sname = (i > 5) ? '.' : "RSDTZW"[i];
1417         psinfo->pr_zomb = psinfo->pr_sname == 'Z';
1418         psinfo->pr_nice = task_nice(p);
1419         psinfo->pr_flag = p->flags;
1420         rcu_read_lock();
1421         cred = __task_cred(p);
1422         SET_UID(psinfo->pr_uid, from_kuid_munged(cred->user_ns, cred->uid));
1423         SET_GID(psinfo->pr_gid, from_kgid_munged(cred->user_ns, cred->gid));
1424         rcu_read_unlock();
1425         strncpy(psinfo->pr_fname, p->comm, sizeof(psinfo->pr_fname));
1426
1427         return 0;
1428 }
1429
1430 /* Here is the structure in which status of each thread is captured. */
1431 struct elf_thread_status
1432 {
1433         struct list_head list;
1434         struct elf_prstatus prstatus;   /* NT_PRSTATUS */
1435         elf_fpregset_t fpu;             /* NT_PRFPREG */
1436         struct task_struct *thread;
1437 #ifdef ELF_CORE_COPY_XFPREGS
1438         elf_fpxregset_t xfpu;           /* ELF_CORE_XFPREG_TYPE */
1439 #endif
1440         struct memelfnote notes[3];
1441         int num_notes;
1442 };
1443
1444 /*
1445  * In order to add the specific thread information for the elf file format,
1446  * we need to keep a linked list of every thread's pr_status and then create
1447  * a single section for them in the final core file.
1448  */
1449 static int elf_dump_thread_status(long signr, struct elf_thread_status *t)
1450 {
1451         struct task_struct *p = t->thread;
1452         int sz = 0;
1453
1454         t->num_notes = 0;
1455
1456         fill_prstatus(&t->prstatus, p, signr);
1457         elf_core_copy_task_regs(p, &t->prstatus.pr_reg);
1458
1459         fill_note(&t->notes[0], "CORE", NT_PRSTATUS, sizeof(t->prstatus),
1460                   &t->prstatus);
1461         t->num_notes++;
1462         sz += notesize(&t->notes[0]);
1463
1464         t->prstatus.pr_fpvalid = elf_core_copy_task_fpregs(p, NULL, &t->fpu);
1465         if (t->prstatus.pr_fpvalid) {
1466                 fill_note(&t->notes[1], "CORE", NT_PRFPREG, sizeof(t->fpu),
1467                           &t->fpu);
1468                 t->num_notes++;
1469                 sz += notesize(&t->notes[1]);
1470         }
1471
1472 #ifdef ELF_CORE_COPY_XFPREGS
1473         if (elf_core_copy_task_xfpregs(p, &t->xfpu)) {
1474                 fill_note(&t->notes[2], "LINUX", ELF_CORE_XFPREG_TYPE,
1475                           sizeof(t->xfpu), &t->xfpu);
1476                 t->num_notes++;
1477                 sz += notesize(&t->notes[2]);
1478         }
1479 #endif
1480         return sz;
1481 }
1482
1483 static void fill_extnum_info(struct elfhdr *elf, struct elf_shdr *shdr4extnum,
1484                              elf_addr_t e_shoff, int segs)
1485 {
1486         elf->e_shoff = e_shoff;
1487         elf->e_shentsize = sizeof(*shdr4extnum);
1488         elf->e_shnum = 1;
1489         elf->e_shstrndx = SHN_UNDEF;
1490
1491         memset(shdr4extnum, 0, sizeof(*shdr4extnum));
1492
1493         shdr4extnum->sh_type = SHT_NULL;
1494         shdr4extnum->sh_size = elf->e_shnum;
1495         shdr4extnum->sh_link = elf->e_shstrndx;
1496         shdr4extnum->sh_info = segs;
1497 }
1498
1499 /*
1500  * dump the segments for an MMU process
1501  */
1502 static bool elf_fdpic_dump_segments(struct coredump_params *cprm)
1503 {
1504         struct vm_area_struct *vma;
1505
1506         for (vma = current->mm->mmap; vma; vma = vma->vm_next) {
1507                 unsigned long addr;
1508
1509                 if (!maydump(vma, cprm->mm_flags))
1510                         continue;
1511
1512 #ifdef CONFIG_MMU
1513                 for (addr = vma->vm_start; addr < vma->vm_end;
1514                                                         addr += PAGE_SIZE) {
1515                         bool res;
1516                         struct page *page = get_dump_page(addr);
1517                         if (page) {
1518                                 void *kaddr = kmap(page);
1519                                 res = dump_emit(cprm, kaddr, PAGE_SIZE);
1520                                 kunmap(page);
1521                                 page_cache_release(page);
1522                         } else {
1523                                 res = dump_skip(cprm, PAGE_SIZE);
1524                         }
1525                         if (!res)
1526                                 return false;
1527                 }
1528 #else
1529                 if (!dump_emit(cprm, (void *) vma->vm_start,
1530                                 vma->vm_end - vma->vm_start))
1531                         return false;
1532 #endif
1533         }
1534         return true;
1535 }
1536
1537 static size_t elf_core_vma_data_size(unsigned long mm_flags)
1538 {
1539         struct vm_area_struct *vma;
1540         size_t size = 0;
1541
1542         for (vma = current->mm->mmap; vma; vma = vma->vm_next)
1543                 if (maydump(vma, mm_flags))
1544                         size += vma->vm_end - vma->vm_start;
1545         return size;
1546 }
1547
1548 /*
1549  * Actual dumper
1550  *
1551  * This is a two-pass process; first we find the offsets of the bits,
1552  * and then they are actually written out.  If we run out of core limit
1553  * we just truncate.
1554  */
1555 static int elf_fdpic_core_dump(struct coredump_params *cprm)
1556 {
1557 #define NUM_NOTES       6
1558         int has_dumped = 0;
1559         mm_segment_t fs;
1560         int segs;
1561         int i;
1562         struct vm_area_struct *vma;
1563         struct elfhdr *elf = NULL;
1564         loff_t offset = 0, dataoff;
1565         int numnote;
1566         struct memelfnote *notes = NULL;
1567         struct elf_prstatus *prstatus = NULL;   /* NT_PRSTATUS */
1568         struct elf_prpsinfo *psinfo = NULL;     /* NT_PRPSINFO */
1569         LIST_HEAD(thread_list);
1570         struct list_head *t;
1571         elf_fpregset_t *fpu = NULL;
1572 #ifdef ELF_CORE_COPY_XFPREGS
1573         elf_fpxregset_t *xfpu = NULL;
1574 #endif
1575         int thread_status_size = 0;
1576         elf_addr_t *auxv;
1577         struct elf_phdr *phdr4note = NULL;
1578         struct elf_shdr *shdr4extnum = NULL;
1579         Elf_Half e_phnum;
1580         elf_addr_t e_shoff;
1581         struct core_thread *ct;
1582         struct elf_thread_status *tmp;
1583
1584         /*
1585          * We no longer stop all VM operations.
1586          *
1587          * This is because those proceses that could possibly change map_count
1588          * or the mmap / vma pages are now blocked in do_exit on current
1589          * finishing this core dump.
1590          *
1591          * Only ptrace can touch these memory addresses, but it doesn't change
1592          * the map_count or the pages allocated. So no possibility of crashing
1593          * exists while dumping the mm->vm_next areas to the core file.
1594          */
1595
1596         /* alloc memory for large data structures: too large to be on stack */
1597         elf = kmalloc(sizeof(*elf), GFP_KERNEL);
1598         if (!elf)
1599                 goto cleanup;
1600         prstatus = kzalloc(sizeof(*prstatus), GFP_KERNEL);
1601         if (!prstatus)
1602                 goto cleanup;
1603         psinfo = kmalloc(sizeof(*psinfo), GFP_KERNEL);
1604         if (!psinfo)
1605                 goto cleanup;
1606         notes = kmalloc(NUM_NOTES * sizeof(struct memelfnote), GFP_KERNEL);
1607         if (!notes)
1608                 goto cleanup;
1609         fpu = kmalloc(sizeof(*fpu), GFP_KERNEL);
1610         if (!fpu)
1611                 goto cleanup;
1612 #ifdef ELF_CORE_COPY_XFPREGS
1613         xfpu = kmalloc(sizeof(*xfpu), GFP_KERNEL);
1614         if (!xfpu)
1615                 goto cleanup;
1616 #endif
1617
1618         for (ct = current->mm->core_state->dumper.next;
1619                                         ct; ct = ct->next) {
1620                 tmp = kzalloc(sizeof(*tmp), GFP_KERNEL);
1621                 if (!tmp)
1622                         goto cleanup;
1623
1624                 tmp->thread = ct->task;
1625                 list_add(&tmp->list, &thread_list);
1626         }
1627
1628         list_for_each(t, &thread_list) {
1629                 struct elf_thread_status *tmp;
1630                 int sz;
1631
1632                 tmp = list_entry(t, struct elf_thread_status, list);
1633                 sz = elf_dump_thread_status(cprm->siginfo->si_signo, tmp);
1634                 thread_status_size += sz;
1635         }
1636
1637         /* now collect the dump for the current */
1638         fill_prstatus(prstatus, current, cprm->siginfo->si_signo);
1639         elf_core_copy_regs(&prstatus->pr_reg, cprm->regs);
1640
1641         segs = current->mm->map_count;
1642         segs += elf_core_extra_phdrs();
1643
1644         /* for notes section */
1645         segs++;
1646
1647         /* If segs > PN_XNUM(0xffff), then e_phnum overflows. To avoid
1648          * this, kernel supports extended numbering. Have a look at
1649          * include/linux/elf.h for further information. */
1650         e_phnum = segs > PN_XNUM ? PN_XNUM : segs;
1651
1652         /* Set up header */
1653         fill_elf_fdpic_header(elf, e_phnum);
1654
1655         has_dumped = 1;
1656         /*
1657          * Set up the notes in similar form to SVR4 core dumps made
1658          * with info from their /proc.
1659          */
1660
1661         fill_note(notes + 0, "CORE", NT_PRSTATUS, sizeof(*prstatus), prstatus);
1662         fill_psinfo(psinfo, current->group_leader, current->mm);
1663         fill_note(notes + 1, "CORE", NT_PRPSINFO, sizeof(*psinfo), psinfo);
1664
1665         numnote = 2;
1666
1667         auxv = (elf_addr_t *) current->mm->saved_auxv;
1668
1669         i = 0;
1670         do
1671                 i += 2;
1672         while (auxv[i - 2] != AT_NULL);
1673         fill_note(&notes[numnote++], "CORE", NT_AUXV,
1674                   i * sizeof(elf_addr_t), auxv);
1675
1676         /* Try to dump the FPU. */
1677         if ((prstatus->pr_fpvalid =
1678              elf_core_copy_task_fpregs(current, cprm->regs, fpu)))
1679                 fill_note(notes + numnote++,
1680                           "CORE", NT_PRFPREG, sizeof(*fpu), fpu);
1681 #ifdef ELF_CORE_COPY_XFPREGS
1682         if (elf_core_copy_task_xfpregs(current, xfpu))
1683                 fill_note(notes + numnote++,
1684                           "LINUX", ELF_CORE_XFPREG_TYPE, sizeof(*xfpu), xfpu);
1685 #endif
1686
1687         fs = get_fs();
1688         set_fs(KERNEL_DS);
1689
1690         offset += sizeof(*elf);                         /* Elf header */
1691         offset += segs * sizeof(struct elf_phdr);       /* Program headers */
1692
1693         /* Write notes phdr entry */
1694         {
1695                 int sz = 0;
1696
1697                 for (i = 0; i < numnote; i++)
1698                         sz += notesize(notes + i);
1699
1700                 sz += thread_status_size;
1701
1702                 phdr4note = kmalloc(sizeof(*phdr4note), GFP_KERNEL);
1703                 if (!phdr4note)
1704                         goto end_coredump;
1705
1706                 fill_elf_note_phdr(phdr4note, sz, offset);
1707                 offset += sz;
1708         }
1709
1710         /* Page-align dumped data */
1711         dataoff = offset = roundup(offset, ELF_EXEC_PAGESIZE);
1712
1713         offset += elf_core_vma_data_size(cprm->mm_flags);
1714         offset += elf_core_extra_data_size();
1715         e_shoff = offset;
1716
1717         if (e_phnum == PN_XNUM) {
1718                 shdr4extnum = kmalloc(sizeof(*shdr4extnum), GFP_KERNEL);
1719                 if (!shdr4extnum)
1720                         goto end_coredump;
1721                 fill_extnum_info(elf, shdr4extnum, e_shoff, segs);
1722         }
1723
1724         offset = dataoff;
1725
1726         if (!dump_emit(cprm, elf, sizeof(*elf)))
1727                 goto end_coredump;
1728
1729         if (!dump_emit(cprm, phdr4note, sizeof(*phdr4note)))
1730                 goto end_coredump;
1731
1732         /* write program headers for segments dump */
1733         for (vma = current->mm->mmap; vma; vma = vma->vm_next) {
1734                 struct elf_phdr phdr;
1735                 size_t sz;
1736
1737                 sz = vma->vm_end - vma->vm_start;
1738
1739                 phdr.p_type = PT_LOAD;
1740                 phdr.p_offset = offset;
1741                 phdr.p_vaddr = vma->vm_start;
1742                 phdr.p_paddr = 0;
1743                 phdr.p_filesz = maydump(vma, cprm->mm_flags) ? sz : 0;
1744                 phdr.p_memsz = sz;
1745                 offset += phdr.p_filesz;
1746                 phdr.p_flags = vma->vm_flags & VM_READ ? PF_R : 0;
1747                 if (vma->vm_flags & VM_WRITE)
1748                         phdr.p_flags |= PF_W;
1749                 if (vma->vm_flags & VM_EXEC)
1750                         phdr.p_flags |= PF_X;
1751                 phdr.p_align = ELF_EXEC_PAGESIZE;
1752
1753                 if (!dump_emit(cprm, &phdr, sizeof(phdr)))
1754                         goto end_coredump;
1755         }
1756
1757         if (!elf_core_write_extra_phdrs(cprm, offset))
1758                 goto end_coredump;
1759
1760         /* write out the notes section */
1761         for (i = 0; i < numnote; i++)
1762                 if (!writenote(notes + i, cprm))
1763                         goto end_coredump;
1764
1765         /* write out the thread status notes section */
1766         list_for_each(t, &thread_list) {
1767                 struct elf_thread_status *tmp =
1768                                 list_entry(t, struct elf_thread_status, list);
1769
1770                 for (i = 0; i < tmp->num_notes; i++)
1771                         if (!writenote(&tmp->notes[i], cprm))
1772                                 goto end_coredump;
1773         }
1774
1775         if (!dump_skip(cprm, dataoff - cprm->written))
1776                 goto end_coredump;
1777
1778         if (!elf_fdpic_dump_segments(cprm))
1779                 goto end_coredump;
1780
1781         if (!elf_core_write_extra_data(cprm))
1782                 goto end_coredump;
1783
1784         if (e_phnum == PN_XNUM) {
1785                 if (!dump_emit(cprm, shdr4extnum, sizeof(*shdr4extnum)))
1786                         goto end_coredump;
1787         }
1788
1789         if (cprm->file->f_pos != offset) {
1790                 /* Sanity check */
1791                 printk(KERN_WARNING
1792                        "elf_core_dump: file->f_pos (%lld) != offset (%lld)\n",
1793                        cprm->file->f_pos, offset);
1794         }
1795
1796 end_coredump:
1797         set_fs(fs);
1798
1799 cleanup:
1800         while (!list_empty(&thread_list)) {
1801                 struct list_head *tmp = thread_list.next;
1802                 list_del(tmp);
1803                 kfree(list_entry(tmp, struct elf_thread_status, list));
1804         }
1805         kfree(phdr4note);
1806         kfree(elf);
1807         kfree(prstatus);
1808         kfree(psinfo);
1809         kfree(notes);
1810         kfree(fpu);
1811         kfree(shdr4extnum);
1812 #ifdef ELF_CORE_COPY_XFPREGS
1813         kfree(xfpu);
1814 #endif
1815         return has_dumped;
1816 #undef NUM_NOTES
1817 }
1818
1819 #endif          /* CONFIG_ELF_CORE */