path: root/exec/exec.c

/* GNU Hurd standard exec server.
   Copyright (C) 1992,93,94,95,96,98,99,2000,01,02,04
   	Free Software Foundation, Inc.
   Written by Roland McGrath.

   Can exec ELF format directly.
   #ifdef GZIP
   Can gunzip executables into core on the fly.
   #endif
   #ifdef BFD
   Can exec any executable format the BFD library understands
   to be for this flavor of machine.
   #endif
   #ifdef BZIP2
   Can bunzip2 executables into core on the fly.
   #endif

This file is part of the GNU Hurd.

The GNU Hurd is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2, or (at your option)
any later version.

The GNU Hurd is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU General Public License for more details.

You should have received a copy of the GNU General Public License
along with the GNU Hurd; see the file COPYING.  If not, write to
the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.  */



#include "priv.h"
#include <hurd.h>
#include <hurd/exec.h>
#include <sys/stat.h>
#include <sys/param.h>
#include <unistd.h>
#include <stdbool.h>

mach_port_t procserver;	/* Our proc port.  */

/* Standard exec data for secure execs.  */
mach_port_t *std_ports;
int *std_ints;
size_t std_nports, std_nints;
pthread_rwlock_t std_lock = PTHREAD_RWLOCK_INITIALIZER;


#ifdef	BFD
/* Return a Hurd error code corresponding to the most recent BFD error.  */
static error_t
b2he (error_t deflt)
{
  switch (bfd_get_error ())
    {
    case bfd_error_system_call:
      return errno;

    case bfd_error_no_memory:
      return ENOMEM;

    default:
      return deflt;
    }
}
#else
#define	b2he()	a2he (errno)
#endif

#ifdef GZIP
static void check_gzip (struct execdata *);
#endif

#ifdef BZIP2
static void check_bzip2 (struct execdata *);
#endif

#ifdef	BFD

/* Check a section, updating the `locations' vector [BFD].  */
static void
check_section (bfd *bfd, asection *sec, void *userdata)
{
  struct execdata *u = userdata;
  vm_address_t addr;
  static const union
    {
      char string[8];
      unsigned int quadword __attribute__ ((mode (DI)));
    } interp = { string: ".interp" };

  if (u->error)
    return;

  /* Fast strcmp for this 8-byte constant string.  */
  if (*(const __typeof (interp.quadword) *) sec->name == interp.quadword)
    u->interp.section = sec;

  if (!(sec->flags & (SEC_ALLOC|SEC_LOAD)) ||
      (sec->flags & SEC_NEVER_LOAD))
    /* Nothing to do for this section.  */
    return;

  addr = (vm_address_t) sec->vma;

  if (sec->flags & SEC_LOAD)
    {
      u->info.bfd_locations[sec->index] = sec->filepos;
      if ((off_t) sec->filepos < 0 || (off_t) sec->filepos > u->file_size)
	u->error = ENOEXEC;
    }
}
#endif


/* Zero the specified region but don't crash the server if it faults.  */

#include <hurd/sigpreempt.h>

static error_t
safe_bzero (void *ptr, size_t size)
{
  return hurd_safe_memset (ptr, 0, size);
}


/* Load or allocate a section.  */
static void
load_section (void *section, struct execdata *u)
{
  vm_address_t addr = 0;
  vm_offset_t filepos = 0;
  vm_size_t filesz = 0, memsz = 0;
  vm_prot_t vm_prot;
  int anywhere;
  vm_address_t mask = 0;
#ifdef BFD
  asection *const sec = section;
#endif
  const ElfW(Phdr) *const ph = section;

  if (u->error)
    return;

#ifdef BFD
  if (u->bfd && sec->flags & SEC_NEVER_LOAD)
    /* Nothing to do for this section.  */
    return;
#endif

  vm_prot = VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE;

#ifdef	BFD
  if (u->bfd)
    {
      addr = (vm_address_t) sec->vma;
      filepos = u->info.bfd_locations[sec->index];
      memsz = sec->_raw_size;
      filesz = (sec->flags & SEC_LOAD) ? memsz : 0;
      if (sec->flags & (SEC_READONLY|SEC_ROM))
	vm_prot &= ~VM_PROT_WRITE;
      anywhere = 0;
    }
  else
#endif
    {
      addr = ph->p_vaddr & ~(ph->p_align - 1);
      memsz = ph->p_vaddr + ph->p_memsz - addr;
      filepos = ph->p_offset & ~(ph->p_align - 1);
      filesz = ph->p_offset + ph->p_filesz - filepos;
      if ((ph->p_flags & PF_R) == 0)
	vm_prot &= ~VM_PROT_READ;
      if ((ph->p_flags & PF_W) == 0)
	vm_prot &= ~VM_PROT_WRITE;
      if ((ph->p_flags & PF_X) == 0)
	vm_prot &= ~VM_PROT_EXECUTE;
      anywhere = u->info.elf.anywhere;
      if (! anywhere)
	addr += u->info.elf.loadbase;
      else
#if 0
	switch (elf_machine)
	  {
	  case EM_386:
	  case EM_486:
	    /* On the i386, programs normally load at 0x08000000, and
	       expect their data segment to be able to grow dynamically
	       upward from its start near that address.  We need to make
	       sure that the dynamic linker is not mapped in a conflicting
	       address.  */
	    /* mask = 0xf8000000UL; */ /* XXX */
	    break;
	  default:
	    break;
	  }
#endif
      if (anywhere && addr < vm_page_size)
	addr = vm_page_size;
    }

  if (memsz == 0)
    /* This section is empty; ignore it.  */
    return;

  if (filesz != 0)
    {
      vm_address_t mapstart = round_page (addr);

      /* Allocate space in the task and write CONTENTS into it.  */
      void write_to_task (vm_address_t mapstart, vm_size_t size,
			  vm_prot_t vm_prot, vm_address_t contents)
	{
	  vm_size_t off = size % vm_page_size;
	  /* Allocate with vm_map to set max protections.  */
	  u->error = vm_map (u->task,
			     &mapstart, size, mask, anywhere,
			     MACH_PORT_NULL, 0, 1,
			     vm_prot|VM_PROT_WRITE,
			     VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE,
			     VM_INHERIT_COPY);
	  if (! u->error && size >= vm_page_size)
	    u->error = vm_write (u->task, mapstart, contents, size - off);
	  if (! u->error && off != 0)
	    {
	      vm_address_t page = 0;
	      page = (vm_address_t) mmap (0, vm_page_size,
					  PROT_READ|PROT_WRITE, MAP_ANON,
					  0, 0);
	      u->error = (page == -1) ? errno : 0;
	      if (! u->error)
		{
		  u->error = hurd_safe_copyin ((void *) page, /* XXX/fault */
			  (void *) (contents + (size - off)),
			  off);
		  if (! u->error)
		    u->error = vm_write (u->task, mapstart + (size - off),
				         page, vm_page_size);
		  munmap ((caddr_t) page, vm_page_size);
		}
	    }
	  /* Reset the current protections to the desired state.  */
	  if (! u->error && (vm_prot & VM_PROT_WRITE) == 0)
	    u->error = vm_protect (u->task, mapstart, size, 0, vm_prot);
	}

      if (mapstart - addr < filesz)
	{
	  /* MAPSTART is the first page that starts inside the section.
	     Map all the pages that start inside the section.  */

#define SECTION_IN_MEMORY_P	(u->file_data != NULL)
#define SECTION_CONTENTS	(u->file_data + filepos)
	  if (SECTION_IN_MEMORY_P)
	    /* Data is already in memory; write it into the task.  */
	    write_to_task (mapstart, filesz - (mapstart - addr), vm_prot,
			   (vm_address_t) SECTION_CONTENTS
			   + (mapstart - addr));
	  else if (u->filemap != MACH_PORT_NULL)
	    /* Map the data into the task directly from the file.  */
	    u->error = vm_map (u->task,
			       &mapstart, filesz - (mapstart - addr),
			       mask, anywhere,
			       u->filemap, filepos + (mapstart - addr), 1,
			       vm_prot,
			       VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE,
			       VM_INHERIT_COPY);
	  else
	    {
	      /* Cannot map the data.  Read it into a buffer and vm_write
		 it into the task.  */
	      const vm_size_t size = filesz - (mapstart - addr);
	      void *buf = map (u, filepos + (mapstart - addr), size);
	      if (buf)
		write_to_task (mapstart, size, vm_prot, (vm_address_t) buf);
	    }
	  if (u->error)
	    return;

	  if (anywhere)
	    {
	      /* We let the kernel choose the location of the mapping.
		 Now record where it ended up.  Later sections cannot
		 be mapped anywhere, they must come after this one.  */
	      u->info.elf.loadbase = mapstart;
	      addr = mapstart + (addr % vm_page_size);
	      anywhere = u->info.elf.anywhere = 0;
	      mask = 0;
	    }
	}

      if (mapstart > addr)
	{
	  /* We must read and copy in the space in the section before the
             first page boundary.  */
	  vm_address_t overlap_page = trunc_page (addr);
	  vm_address_t ourpage = 0;
	  vm_size_t size = 0;
	  void *readaddr;
	  size_t readsize;

	  u->error = vm_read (u->task, overlap_page, vm_page_size,
			      &ourpage, &size);
	  if (u->error)
	    {
	      if (u->error == KERN_INVALID_ADDRESS)
		{
		  /* The space is unallocated.  */
		  u->error = vm_allocate (u->task,
					  &overlap_page, vm_page_size, 0);
		  size = vm_page_size;
		  if (!u->error)
		    {
		      ourpage = (vm_address_t) mmap (0, vm_page_size,
						     PROT_READ|PROT_WRITE,
						     MAP_ANON, 0, 0);
		      u->error = (ourpage == -1) ? errno : 0;
		    }
		}
	      if (u->error)
		{
		maplose:
		  vm_deallocate (u->task, mapstart, filesz);
		  return;
		}
	    }

	  readaddr = (void *) (ourpage + (addr - overlap_page));
	  readsize = size - (addr - overlap_page);
	  if (readsize > filesz)
	    readsize = filesz;

	  if (SECTION_IN_MEMORY_P)
	    memcpy (readaddr, SECTION_CONTENTS, readsize);
	  else
	    {
	      const void *contents = map (u, filepos, readsize);
	      if (!contents)
		goto maplose;
	      u->error = hurd_safe_copyin (readaddr, contents,
	                                   readsize); /* XXX/fault */
	      if (u->error)
	        goto maplose;
	    }
	  u->error = vm_write (u->task, overlap_page, ourpage, size);
	  if (u->error == KERN_PROTECTION_FAILURE)
	    {
	      /* The overlap page is not writable; the section
		 that appears in preceding memory is read-only.
		 Change the page's protection so we can write it.  */
	      u->error = vm_protect (u->task, overlap_page, size,
				     0, vm_prot | VM_PROT_WRITE);
	      if (!u->error)
		u->error = vm_write (u->task, overlap_page, ourpage, size);
	      /* If this section is not supposed to be writable either,
		 restore the page's protection to read-only.  */
	      if (!u->error && !(vm_prot & VM_PROT_WRITE))
		u->error = vm_protect (u->task, overlap_page, size,
				       0, vm_prot);
	    }
	  munmap ((caddr_t) ourpage, size);
	  if (u->error)
	    goto maplose;
	}

      if (u->cntl)
	u->cntl->accessed = 1;

      /* Tell the code below to zero-fill the remaining area.  */
      addr += filesz;
      memsz -= filesz;
    }

  if (memsz != 0)
    {
      /* SEC_ALLOC: Allocate zero-filled memory for the section.  */

      vm_address_t mapstart = round_page (addr);

      if (mapstart - addr < memsz)
	{
	  /* MAPSTART is the first page that starts inside the section.
	     Allocate all the pages that start inside the section.  */
	  u->error = vm_map (u->task, &mapstart, memsz - (mapstart - addr),
			     mask, anywhere, MACH_PORT_NULL, 0, 1,
			     vm_prot, VM_PROT_ALL, VM_INHERIT_COPY);
	  if (u->error)
	    return;
	}

      if (anywhere)
	{
	  /* We let the kernel choose the location of the zero space.
	     Now record where it ended up.  Later sections cannot
	     be mapped anywhere, they must come after this one.  */
	  u->info.elf.loadbase = mapstart;
	  addr = mapstart + (addr % vm_page_size);
	  anywhere = u->info.elf.anywhere = 0;
	  mask = 0;
	}

      if (mapstart > addr)
	{
	  /* Zero space in the section before the first page boundary.  */
	  vm_address_t overlap_page = trunc_page (addr);
	  vm_address_t ourpage = 0;
	  vm_size_t size = 0;
	  u->error = vm_read (u->task, overlap_page, vm_page_size,
			      &ourpage, &size);
	  if (u->error)
	    {
	      vm_deallocate (u->task, mapstart, memsz);
	      return;
	    }
	  u->error = safe_bzero ((void *) (ourpage + (addr - overlap_page)),
				 size - (addr - overlap_page));
	  if (! u->error && !(vm_prot & VM_PROT_WRITE))
	    u->error = vm_protect (u->task, overlap_page, size,
				   0, VM_PROT_WRITE);
	  if (! u->error)
	    u->error = vm_write (u->task, overlap_page, ourpage, size);
	  if (! u->error && !(vm_prot & VM_PROT_WRITE))
	    u->error = vm_protect (u->task, overlap_page, size, 0, vm_prot);
	  munmap ((caddr_t) ourpage, size);
	}
    }
}

/* XXX all accesses of the mapped data need to use fault handling
   to abort the RPC when mapped file data generates bad page faults.
   I've marked some accesses with XXX/fault comments.
   --roland  */

void *
map (struct execdata *e, off_t posn, size_t len)
{
  const size_t size = e->file_size;
  size_t offset;

  if ((map_filepos (e) & ~(map_vsize (e) - 1)) == (posn & ~(map_vsize (e) - 1))
      && posn + len - map_filepos (e) <= map_fsize (e))
    /* The current mapping window covers it.  */
    offset = posn & (map_vsize (e) - 1);
  else if (posn + len > size)
    /* The requested data wouldn't fit in the file.  */
    return NULL;
  else if (e->file_data != NULL) {
    return e->file_data + posn;
  } else if (e->filemap == MACH_PORT_NULL)
    {
      /* No mapping for the file.  Read the data by RPC.  */
      char *buffer = map_buffer (e);
      mach_msg_type_number_t nread = map_vsize (e);

      assert (e->file_data == NULL); /* Must be first or second case.  */

      /* Read as much as we can get into the buffer right now.  */
      e->error = io_read (e->file, &buffer, &nread, posn, round_page (len));
      if (e->error)
	return NULL;
      if (buffer != map_buffer (e))
	{
	  /* The data was returned out of line.  Discard the old buffer.  */
	  if (map_vsize (e) != 0)
	    munmap (map_buffer (e), map_vsize (e));
	  map_buffer (e) = buffer;
	  map_vsize (e) = round_page (nread);
	}

      map_filepos (e) = posn;
      map_set_fsize (e, nread);
      offset = 0;
    }
  else
    {
      /* Deallocate the old mapping area.  */
      if (map_buffer (e) != NULL)
	munmap (map_buffer (e), map_vsize (e));
      map_buffer (e) = NULL;

      /* Make sure our mapping is page-aligned in the file.  */
      offset = posn & (vm_page_size - 1);
      map_filepos (e) = trunc_page (posn);
      map_vsize (e) = round_page (posn + len) - map_filepos (e);

      /* Map the data from the file.  */
      if (vm_map (mach_task_self (),
		  (vm_address_t *) &map_buffer (e), map_vsize (e), 0, 1,
		  e->filemap, map_filepos (e), 1, VM_PROT_READ, VM_PROT_READ,
		  VM_INHERIT_NONE))
	{
	  e->error = EIO;
	  return NULL;
	}

      if (e->cntl)
	e->cntl->accessed = 1;

      map_set_fsize (e, MIN (map_vsize (e), size - map_filepos (e)));
    }

  return map_buffer (e) + offset;
}


/* Initialize E's stdio stream.  */
static void prepare_stream (struct execdata *e);

/* Point the stream at the buffer of file data in E->file_data.  */
static void prepare_in_memory (struct execdata *e);


#ifndef EXECDATA_STREAM

/* We don't have a stdio stream, but we have a mapping window
   we need to initialize.  */
static void
prepare_stream (struct execdata *e)
{
  e->map_buffer = NULL;
  e->map_vsize = e->map_fsize = 0;
  e->map_filepos = 0;
}

static void prepare_in_memory (struct execdata *e)
{
  prepare_stream(e);
}

#else

#ifdef _STDIO_USES_IOSTREAM

# error implement me for libio!

#else  /* old GNU stdio */

#if 0
void *
map (struct execdata *e, off_t posn, size_t len)
{
  FILE *f = &e->stream;
  const size_t size = e->file_size;
  size_t offset;

  if ((f->__offset & ~(f->__bufsize - 1)) == (posn & ~(f->__bufsize - 1)) &&
      f->__buffer + (posn + len - f->__offset) < f->__get_limit)
    /* The current mapping window covers it.  */
    offset = posn & (f->__bufsize - 1);
  else if (e->file_data != NULL)
    {
      /* The current "mapping window" is in fact the whole file contents.
	 So if it's not in there, it's not in there.  */
      f->__eof = 1;
      return NULL;
    }
  else if (e->filemap == MACH_PORT_NULL)
    {
      /* No mapping for the file.  Read the data by RPC.  */
      char *buffer = f->__buffer;
      mach_msg_type_number_t nread = f->__bufsize;
      /* Read as much as we can get into the buffer right now.  */
      e->error = io_read (e->file, &buffer, &nread, posn, round_page (len));
      if (e->error)
	{
	  errno = e->error;
	  f->__error = 1;
	  return NULL;
	}
      if (buffer != f->__buffer)
	{
	  /* The data was returned out of line.  Discard the old buffer.  */
	  if (f->__bufsize != 0)
	    munmap (f->__buffer, f->__bufsize);
	  f->__buffer = buffer;
	  f->__bufsize = round_page (nread);
	}

      f->__offset = posn;
      f->__get_limit = f->__buffer + nread;
      offset = 0;
    }
  else
    {
      /* Deallocate the old mapping area.  */
      if (f->__buffer != NULL)
	munmap (f->__buffer, f->__bufsize);
      f->__buffer = NULL;

      /* Make sure our mapping is page-aligned in the file.  */
      offset = posn & (vm_page_size - 1);
      f->__offset = trunc_page (posn);
      f->__bufsize = round_page (posn + len) - f->__offset;

      /* Map the data from the file.  */
      if (vm_map (mach_task_self (),
		  (vm_address_t *) &f->__buffer, f->__bufsize, 0, 1,
		  e->filemap, f->__offset, 1, VM_PROT_READ, VM_PROT_READ,
		  VM_INHERIT_NONE))
	{
	  errno = e->error = EIO;
	  f->__error = 1;
	  return NULL;
	}

      if (e->cntl)
	e->cntl->accessed = 1;

      if (f->__offset + f->__bufsize > size)
	f->__get_limit = f->__buffer + (size - f->__offset);
      else
	f->__get_limit = f->__buffer + f->__bufsize;
    }

  f->__target = f->__offset;
  f->__bufp = f->__buffer + offset;

  if (f->__bufp + len > f->__get_limit)
    {
      f->__eof = 1;
      return NULL;
    }

  return f->__bufp;
}
#endif

/* stdio input-room function.
   XXX/fault in the stdio case (or libio replacement), i.e. for bfd
   (if ever revived), need to check all the mapping fault issues  */
static int
input_room (FILE *f)
{
  struct execdata *e = f->__cookie;
  char *p = map (e, f->__target, 1);
  if (p == NULL)
    {
      (e->error ? f->__error : f->__eof) = 1;
      return EOF;
    }

  f->__target = f->__offset;
  f->__bufp = p;

  return (unsigned char) *f->__bufp++;
}

static int
close_exec_stream (void *cookie)
{
  struct execdata *e = cookie;

  if (e->stream.__buffer != NULL)
    munmap (e->stream.__buffer, e->stream.__bufsize);

  return 0;
}

/* stdio seek function. */
static int
fake_seek (void *cookie, fpos_t *pos, int whence)
{
  struct execdata *e = cookie;

  /* Set __target to match the specifed seek location */
  switch (whence)
    {
    case SEEK_END:
      e->stream.__target = e->file_size + *pos;
      break;

    case SEEK_CUR:
      e->stream.__target += *pos;
      break;

    case SEEK_SET:
      e->stream.__target = *pos;
      break;
    }
  *pos = e->stream.__target;
  return 0;
}

/* Initialize E's stdio stream.  */
static void
prepare_stream (struct execdata *e)
{
  memset (&e->stream, 0, sizeof (e->stream));
  e->stream.__magic = _IOMAGIC;
  e->stream.__mode.__read = 1;
  e->stream.__userbuf = 1;
  e->stream.__room_funcs.__input = input_room;
  e->stream.__io_funcs.seek = fake_seek;
  e->stream.__io_funcs.close = close_exec_stream;
  e->stream.__cookie = e;
  e->stream.__seen = 1;
}

/* Point the stream at the buffer of file data.  */
static void
prepare_in_memory (struct execdata *e)
{
  memset (&e->stream, 0, sizeof (e->stream));
  e->stream.__magic = _IOMAGIC;
  e->stream.__mode.__read = 1;
  e->stream.__buffer = e->file_data;
  e->stream.__bufsize = e->file_size;
  e->stream.__get_limit = e->stream.__buffer + e->stream.__bufsize;
  e->stream.__bufp = e->stream.__buffer;
  e->stream.__seen = 1;
}
#endif

#endif


/* Prepare to check and load FILE.  */
static void
prepare (file_t file, struct execdata *e)
{
  memory_object_t rd, wr;

  e->file = file;

#ifdef	BFD
  e->bfd = NULL;
#endif
  e->file_data = NULL;
  e->cntl = NULL;
  e->filemap = MACH_PORT_NULL;
  e->cntlmap = MACH_PORT_NULL;

  e->interp.section = NULL;

  /* Initialize E's stdio stream.  */
  prepare_stream (e);

  /* Try to mmap FILE.  */
  e->error = io_map (file, &rd, &wr);
  if (! e->error)
    /* Mapping is O.K.  */
    {
      if (wr != MACH_PORT_NULL)
	mach_port_deallocate (mach_task_self (), wr);
      if (rd == MACH_PORT_NULL)
	{
	  e->error = EBADF;	/* ? XXX */
	  return;
	}
      e->filemap = rd;

      e->error = /* io_map_cntl (file, &e->cntlmap) */ EOPNOTSUPP; /* XXX */
      if (!e->error)
	e->error = vm_map (mach_task_self (), (vm_address_t *) &e->cntl,
			   vm_page_size, 0, 1, e->cntlmap, 0, 0,
			   VM_PROT_READ|VM_PROT_WRITE,
			   VM_PROT_READ|VM_PROT_WRITE, VM_INHERIT_NONE);

      if (e->cntl)
	while (1)
	  {
	    pthread_spin_lock (&e->cntl->lock);
	    switch (e->cntl->conch_status)
	      {
	      case USER_COULD_HAVE_CONCH:
		e->cntl->conch_status = USER_HAS_CONCH;
	      case USER_HAS_CONCH:
		pthread_spin_unlock (&e->cntl->lock);
		/* Break out of the loop.  */
		break;
	      case USER_RELEASE_CONCH:
	      case USER_HAS_NOT_CONCH:
	      default:		/* Oops.  */
		pthread_spin_unlock (&e->cntl->lock);
		e->error = io_get_conch (e->file);
		if (e->error)
		  return;
		/* Continue the loop.  */
		continue;
	      }

	    /* Get here if we are now IT.  */
	    e->file_size = 0;
	    if (e->cntl->use_file_size)
	      e->file_size = e->cntl->file_size;
	    if (e->cntl->use_read_size && e->cntl->read_size > e->file_size)
	      e->file_size = e->cntl->read_size;
	    break;
	  }
    }

  if (!e->cntl && (!e->error || e->error == EOPNOTSUPP))
    {
      /* No shared page.  Do a stat to find the file size.  */
      struct stat st;
      e->error = io_stat (file, &st);
      if (e->error)
	return;
      e->file_size = st.st_size;
      e->optimal_block = st.st_blksize;
    }
}

/* Check the magic number, etc. of the file.
   On successful return, the caller must allocate the
   E->locations vector, and map check_section over the BFD.  */

#ifdef BFD
static void
check_bfd (struct execdata *e)
{
  bfd_set_error (bfd_error_no_error);

  e->bfd = bfd_openstreamr (NULL, NULL, &e->stream);
  if (e->bfd == NULL)
    {
      e->error = b2he (ENOEXEC);
      return;
    }

  if (!bfd_check_format (e->bfd, bfd_object))
    {
      e->error = b2he (ENOEXEC);
      return;
    }
  else if (/* !(e->bfd->flags & EXEC_P) || XXX */
	   (host_bfd.arch_info->compatible = e->bfd->arch_info->compatible,
	    bfd_arch_get_compatible (&host_bfd, e->bfd)) != host_bfd.arch_info)
    {
      /* This file is of a recognized binary file format, but it is not
	 executable on this machine.  */
      e->error = b2he (ENOEXEC);
      return;
    }

  e->entry = e->bfd->start_address;
}
#endif

#include <endian.h>
#if BYTE_ORDER == BIG_ENDIAN
#define host_ELFDATA ELFDATA2MSB
#endif
#if BYTE_ORDER == LITTLE_ENDIAN
#define host_ELFDATA ELFDATA2LSB
#endif

static void
check_elf (struct execdata *e)
{
  ElfW(Ehdr) *ehdr = map (e, 0, sizeof (ElfW(Ehdr)));
  ElfW(Phdr) *phdr;

  if (! ehdr)
    {
      if (!e->error)
	e->error = ENOEXEC;
      return;
    }

  if (*(ElfW(Word) *) ehdr != ((union { ElfW(Word) word;
				        unsigned char string[SELFMAG]; })
			       { string: ELFMAG }).word)
    {
      e->error = ENOEXEC;
      return;
    }

  if (ehdr->e_ident[EI_CLASS] != ELFCLASS32 ||
      ehdr->e_ident[EI_DATA] != host_ELFDATA ||
      ehdr->e_ident[EI_VERSION] != EV_CURRENT ||
      ehdr->e_version != EV_CURRENT ||
      ehdr->e_ehsize < sizeof *ehdr ||
      ehdr->e_phentsize != sizeof (ElfW(Phdr)))
    {
      e->error = ENOEXEC;
      return;
    }
  e->error = elf_machine_matches_host (ehdr->e_machine);
  if (e->error)
    return;

  /* Extract all this information now, while EHDR is mapped.
     The `map' call below for the phdrs may reuse the mapping window.  */
  e->entry = ehdr->e_entry;
  e->info.elf.anywhere = (ehdr->e_type == ET_DYN ||
			  ehdr->e_type == ET_REL);
  e->info.elf.loadbase = 0;
  e->info.elf.phnum = ehdr->e_phnum;

  phdr = map (e, ehdr->e_phoff, ehdr->e_phnum * sizeof (ElfW(Phdr)));
  if (! phdr)
    {
      if (!e->error)
	e->error = ENOEXEC;
      return;
    }
  e->info.elf.phdr = phdr;
  e->info.elf.phdr_addr = ehdr->e_phoff;
}

/* Copy MAPPED_PHDR into E->info.elf.phdr, filling in E->interp.phdr
   in the process.  */
static void
check_elf_phdr (struct execdata *e, const ElfW(Phdr) *mapped_phdr)
{
  const ElfW(Phdr) *phdr;
  bool seen_phdr = false;

  memcpy (e->info.elf.phdr, mapped_phdr,
	  e->info.elf.phnum * sizeof (ElfW(Phdr)));

  /* Default state if we do not see PT_GNU_STACK telling us what to do.
     Executable stack is the compatible default.
     (XXX should be machine-dependent??)
  */
  e->info.elf.execstack = 1;

  for (phdr = e->info.elf.phdr;
       phdr < &e->info.elf.phdr[e->info.elf.phnum];
       ++phdr)
    switch (phdr->p_type)
      {
      case PT_INTERP:
	e->interp.phdr = phdr;
	break;
      case PT_LOAD:
	if (e->file_size <= (off_t) (phdr->p_offset +
				     phdr->p_filesz))
	  {
	    e->error = ENOEXEC;
	    return;
	  }
	/* Check if this is the segment that contains the phdr image.  */
	if (!seen_phdr
	    && (phdr->p_offset & -phdr->p_align) == 0 /* Sanity check.  */
	    && phdr->p_offset <= e->info.elf.phdr_addr
	    && e->info.elf.phdr_addr - phdr->p_offset < phdr->p_filesz)
	  {
	    e->info.elf.phdr_addr += phdr->p_vaddr - phdr->p_offset;
	    seen_phdr = true;
	  }
	break;
      case PT_GNU_STACK:
	e->info.elf.execstack = phdr->p_flags & PF_X;
	break;
      }

  if (!seen_phdr)
    e->info.elf.phdr_addr = 0;
}


static void
check (struct execdata *e)
{
  check_elf (e);		/* XXX/fault */
#ifdef BFD
  if (e->error == ENOEXEC)
    {
      e->error = 0;
      check_bfd (e);
    }
#endif
}


/* Release the conch and clean up mapping the file and control page.  */
static void
finish_mapping (struct execdata *e)
{
  if (e->cntl != NULL)
    {
      pthread_spin_lock (&e->cntl->lock);
      if (e->cntl->conch_status == USER_RELEASE_CONCH)
	{
	  pthread_spin_unlock (&e->cntl->lock);
	  io_release_conch (e->file);
	}
      else
	{
	  e->cntl->conch_status = USER_HAS_NOT_CONCH;
	  pthread_spin_unlock (&e->cntl->lock);
	}
      munmap (e->cntl, vm_page_size);
      e->cntl = NULL;
    }
  if (e->filemap != MACH_PORT_NULL)
    {
      mach_port_deallocate (mach_task_self (), e->filemap);
      e->filemap = MACH_PORT_NULL;
    }
  if (e->cntlmap != MACH_PORT_NULL)
    {
      mach_port_deallocate (mach_task_self (), e->cntlmap);
      e->cntlmap = MACH_PORT_NULL;
    }
}

/* Clean up after reading the file (need not be completed).
   Note: this may be called several times for the E, so it must take care
   of checking what was already freed.  */
void
finish (struct execdata *e, int dealloc_file)
{
  finish_mapping (e);
#ifdef	BFD
  if (e->bfd != NULL)
    {
      bfd_close (e->bfd);
      e->bfd = NULL;
    }
  else
#endif
    {
#ifdef EXECDATA_STREAM
      fclose (&e->stream);
#else
      if (e->file_data != NULL) {
	free (e->file_data);
	e->file_data = NULL;
      } else if (map_buffer (e) != NULL) {
	munmap (map_buffer (e), map_vsize (e));
	map_buffer (e) = NULL;
      }
#endif
    }
  if (dealloc_file && e->file != MACH_PORT_NULL)
    {
      mach_port_deallocate (mach_task_self (), e->file);
      e->file = MACH_PORT_NULL;
    }
}


/* Load the file.  */
static void
load (task_t usertask, struct execdata *e)
{
  e->task = usertask;

  if (! e->error)
    {
#ifdef	BFD
      if (e->bfd)
	{
	  void load_bfd_section (bfd *bfd, asection *sec, void *userdata)
	    {
	      load_section (sec, userdata);
	    }
	  bfd_map_over_sections (e->bfd, &load_bfd_section, e);
	}
      else
#endif
	{
	  ElfW(Word) i;
	  for (i = 0; i < e->info.elf.phnum; ++i)
	    if (e->info.elf.phdr[i].p_type == PT_LOAD)
	      load_section (&e->info.elf.phdr[i], e);

	  /* The entry point address is relative to wherever we loaded the
	     program text.  */
	  e->entry += e->info.elf.loadbase;
	}
    }

  /* Release the conch for the file.  */
  finish_mapping (e);

  if (! e->error)
    {
      /* Do post-loading processing on the task.  */

#ifdef	BFD
      if (e->bfd)
	{
	  /* Do post-loading processing for a section.  This consists of
	     peeking the pages of non-demand-paged executables.  */

	  void postload_section (bfd *bfd, asection *sec, void *userdata)
	    {
	      struct execdata *u = userdata;
	      vm_address_t addr = 0;
	      vm_size_t secsize = 0;

	      addr = (vm_address_t) sec->vma;
	      secsize = sec->_raw_size;

	      if ((sec->flags & SEC_LOAD) && !(bfd->flags & D_PAGED))
		{
		  /* Pre-load the section by peeking every mapped page.  */
		  vm_address_t myaddr, a;
		  vm_size_t mysize;
		  myaddr = 0;

		  /* We have already mapped the file into the task in
		     load_section.  Now read from the task's memory into our
		     own address space so we can peek each page and cause it to
		     be paged in.  */
		  u->error = vm_read (u->task, trunc_page (addr),
				      round_page (secsize), &myaddr, &mysize);
		  if (u->error)
		    return;

		  /* Peek at the first word of each page.  */
		  for (a = ((myaddr + mysize) & ~(vm_page_size - 1));
		       a >= myaddr; a -= vm_page_size)
		    /* Force it to be paged in.  */
		    (void) *(volatile int *) a;

		  munmap ((caddr_t) myaddr, mysize);
		}
	    }

	  bfd_map_over_sections (e->bfd, postload_section, e);
	}
#endif
    }
}

#ifdef GZIP
/* Check the file for being a gzip'd image.  Return with ENOEXEC means not
   a valid gzip file; return with another error means lossage in decoding;
   return with zero means the file was uncompressed into memory which E now
   points to, and `check' can be run again.  */

static void
check_gzip (struct execdata *earg)
{
  struct execdata *e = earg;
  /* Entry points to unzip engine.  */
  int get_method (int);
  void unzip (int, int);
  extern long int bytes_out;
  /* Callbacks from unzip for I/O and error interface.  */
  extern int (*unzip_read) (char *buf, size_t maxread);
  extern void (*unzip_write) (const char *buf, size_t nwrite);
  extern void (*unzip_read_error) (void);
  extern void (*unzip_error) (const char *msg);

  char *zipdata = NULL;
  size_t zipdatasz = 0;
  FILE *zipout = NULL;
  jmp_buf ziperr;
  off_t zipread_pos = 0;
  int zipread (char *buf, size_t maxread)
    {
      char *contents = map (e, zipread_pos, 1);
      size_t n;
      if (contents == NULL)
	{
	  errno = e->error;
	  return -1;
	}
      n = MIN (maxread, map_buffer (e) + map_fsize (e) - contents);
      errno = hurd_safe_copyin (buf, contents, n); /* XXX/fault */
      if (errno)
        longjmp (ziperr, 2);
        
      zipread_pos += n;
      return n;
    }
  void zipwrite (const char *buf, size_t nwrite)
    {
      if (fwrite (buf, nwrite, 1, zipout) != 1)
	longjmp (ziperr, 1);
    }
  void ziprderr (void)
    {
      errno = ENOEXEC;
      longjmp (ziperr, 2);
    }
  void ziperror (const char *msg)
    {
      errno = ENOEXEC;
      longjmp (ziperr, 2);
    }

  unzip_read = zipread;
  unzip_write = zipwrite;
  unzip_read_error = ziprderr;
  unzip_error = ziperror;

  if (setjmp (ziperr))
    {
      /* Error in unzipping jumped out.  */
      if (zipout)
	{
	  fclose (zipout);
	  free (zipdata);
	}
      e->error = errno;
      return;
    }

  if (get_method (0) != 0)
    {
      /* Not a happy gzip file.  */
      e->error = ENOEXEC;
      return;
    }

  /* Matched gzip magic number.  Ready to unzip.
     Set up the output stream and let 'er rip.  */

  zipout = open_memstream (&zipdata, &zipdatasz);
  if (! zipout)
    {
      e->error = errno;
      return;
    }

  /* Call the gunzip engine.  */
  bytes_out = 0;
  unzip (17, 23);		/* Arguments ignored.  */

  /* The output is complete.  Clean up the stream and store its resultant
     buffer and size in the execdata as the file contents.  */
  fclose (zipout);

  /* Clean up the old exec file stream's state.
     Now that we have the contents all in memory (in E->file_data),
     nothing will in fact ever try to use E->stream again.  */
  finish (e, 0);

  /* Prepare the stream state to use the file contents already in memory.  */
  e->file_data = zipdata;
  e->file_size = zipdatasz;
  prepare_in_memory (e);
}
#endif

#ifdef BZIP2
/* Check the file for being a bzip2'd image.  Return with ENOEXEC means not
   a valid bzip2 file; return with another error means lossage in decoding;
   return with zero means the file was uncompressed into memory which E now
   points to, and `check' can be run again.  */

static void
check_bzip2 (struct execdata *earg)
{
  struct execdata *e = earg;
  /* Entry points to bunzip2 engine.  */
  void do_bunzip2 (void);
  /* Callbacks from unzip for I/O and error interface.  */
  extern int (*unzip_read) (char *buf, size_t maxread);
  extern void (*unzip_write) (const char *buf, size_t nwrite);
  extern void (*unzip_read_error) (void);
  extern void (*unzip_error) (const char *msg);

  char *zipdata = NULL;
  size_t zipdatasz = 0;
  FILE *zipout = NULL;
  jmp_buf ziperr;
  off_t zipread_pos = 0;
  int zipread (char *buf, size_t maxread)
    {
      char *contents = map (e, zipread_pos, 1);
      size_t n;
      if (contents == NULL)
	{
	  errno = e->error;
	  return -1;
	}
      n = MIN (maxread, map_buffer (e) + map_fsize (e) - contents);
      errno = hurd_safe_copyin (buf, contents, n); /* XXX/fault */
      if (errno)
        longjmp (ziperr, 2);

      zipread_pos += n;
      return n;
    }
  void zipwrite (const char *buf, size_t nwrite)
    {
      if (fwrite (buf, nwrite, 1, zipout) != 1)
	longjmp (ziperr, 1);
    }
  void ziprderr (void)
    {
      errno = ENOEXEC;
      longjmp (ziperr, 2);
    }
  void ziperror (const char *msg)
    {
      errno = ENOEXEC;
      longjmp (ziperr, 2);
    }

  unzip_read = zipread;
  unzip_write = zipwrite;
  unzip_read_error = ziprderr;
  unzip_error = ziperror;

  if (setjmp (ziperr))
    {
      /* Error in unzipping jumped out.  */
      if (zipout)
	{
	  fclose (zipout);
	  free (zipdata);
	}
      e->error = errno;
      return;
    }

  zipout = open_memstream (&zipdata, &zipdatasz);
  if (! zipout)
    {
      e->error = errno;
      return;
    }

  /* Call the bunzip2 engine.  */
  do_bunzip2 ();

  /* The output is complete.  Clean up the stream and store its resultant
     buffer and size in the execdata as the file contents.  */
  fclose (zipout);

  /* Clean up the old exec file stream's state.
     Now that we have the contents all in memory (in E->file_data),
     nothing will in fact ever try to use E->stream again.  */
  finish (e, 0);

  /* Prepare the stream state to use the file contents already in memory.  */
  e->file_data = zipdata;
  e->file_size = zipdatasz;
  prepare_in_memory (e);
}
#endif


static inline void *
servercopy (void *arg, mach_msg_type_number_t argsize, boolean_t argcopy,
	    error_t *errorp)
{
  if (! argcopy)
    return arg;

  /* ARG came in-line, so we must copy it.  */
  void *copy;
  copy = mmap (0, argsize, PROT_READ|PROT_WRITE, MAP_ANON, 0, 0);
  if (copy == MAP_FAILED)
    {
      *errorp = errno;
      return NULL;
    }
  memcpy (copy, arg, argsize);
  return copy;
}


static error_t
do_exec (file_t file,
	 task_t oldtask,
	 int flags,
	 char *argv, mach_msg_type_number_t argvlen, boolean_t argv_copy,
	 char *envp, mach_msg_type_number_t envplen, boolean_t envp_copy,
	 mach_port_t *dtable, mach_msg_type_number_t dtablesize,
	 boolean_t dtable_copy,
	 mach_port_t *portarray, mach_msg_type_number_t nports,
	 boolean_t portarray_copy,
	 int *intarray, mach_msg_type_number_t nints, boolean_t intarray_copy,
	 mach_port_t *deallocnames, mach_msg_type_number_t ndeallocnames,
	 mach_port_t *destroynames, mach_msg_type_number_t ndestroynames)
{
  struct execdata e, interp;
  task_t newtask = MACH_PORT_NULL;
  thread_t thread = MACH_PORT_NULL;
  struct bootinfo *boot = 0;
  int *ports_replaced;
  int secure, defaults;
  mach_msg_type_number_t i;
  int intarray_dealloc = 0;	/* Dealloc INTARRAY before returning?  */
  int oldtask_trashed = 0;	/* Have we trashed the old task?  */

  /* Prime E for executing FILE and check its validity.  This must be an
     inline function because it stores pointers into alloca'd storage in E
     for later use in `load'.  */
  void prepare_and_check (file_t file, struct execdata *e)
    {
      /* Prepare E to read the file.  */
      prepare (file, e);
      if (e->error)
	return;

      /* Check the file for validity first.  */
      check (e);

#ifdef GZIP
      if (e->error == ENOEXEC)
	{
	  /* See if it is a compressed image.  */
	  static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
	  /* The gzip code is really cheesy, not even close to thread-safe.
	     So we serialize all uses of it.  */
	  pthread_mutex_lock (&lock);
	  e->error = 0;
	  check_gzip (e);
	  pthread_mutex_unlock (&lock);
	  if (e->error == 0)
	    /* The file was uncompressed into memory, and now E describes the
	       uncompressed image rather than the actual file.  Check it again
	       for a valid magic number.  */
	    check (e);
	}
#endif
#ifdef BZIP2
      if (e->error == ENOEXEC)
	{
	  /* See if it is a compressed image.  */
	  static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
	  /* The bzip2 code is really cheesy, not even close to thread-safe.
	     So we serialize all uses of it.  */
	  pthread_mutex_lock (&lock);
	  e->error = 0;
	  check_bzip2 (e);
	  pthread_mutex_unlock (&lock);
	  if (e->error == 0)
	    /* The file was uncompressed into memory, and now E describes the
	       uncompressed image rather than the actual file.  Check it again
	       for a valid magic number.  */
	    check (e);
	}
#endif
    }


  /* Here is the main body of the function.  */

  interp.file = MACH_PORT_NULL;

  /* Catch this error now, rather than later.  */
  /* XXX For EXEC_DEFAULTS, this is only an error if one of the user's
     ports is null; if they are all provided, then EXEC_DEFAULTS would
     have no effect, and the lack of installed standard ports should
     not cause an error.  -mib */
  if ((!std_ports || !std_ints) && (flags & (EXEC_SECURE|EXEC_DEFAULTS)))
    return EIEIO;

  /* Suspend the existing task before frobnicating it.  */
  if (oldtask != MACH_PORT_NULL && (e.error = task_suspend (oldtask)))
    return e.error;

  /* Prime E for executing FILE and check its validity.  */
  prepare_and_check (file, &e);

  if (e.error == ENOEXEC)
    {
      /* Check for a #! executable file.  */
      check_hashbang (&e,
		      file, oldtask, flags,
		      argv, argvlen, argv_copy,
		      envp, envplen, envp_copy,
		      dtable, dtablesize, dtable_copy,
		      portarray, nports, portarray_copy,
		      intarray, nints, intarray_copy,
		      deallocnames, ndeallocnames,
		      destroynames, ndestroynames);
      if (! e.error)
	/* The #! exec succeeded; nothing more to do.  */
	return 0;
    }

  if (e.error)
    /* The file is not a valid executable.  */
    goto out;

#ifdef	BFD
  if (e.bfd)
    {
      e.info.bfd_locations = alloca (e.bfd->section_count *
				     sizeof (vm_offset_t));
      bfd_map_over_sections (e.bfd, check_section, &e);
    }
  else
#endif
    {
      const ElfW(Phdr) *phdr = e.info.elf.phdr;
      e.info.elf.phdr = alloca (e.info.elf.phnum * sizeof (ElfW(Phdr)));
      check_elf_phdr (&e, phdr);
    }

  if (oldtask == MACH_PORT_NULL)
    flags |= EXEC_NEWTASK;

  if (flags & (EXEC_NEWTASK|EXEC_SECURE))
    {
      /* Create the new task.  If we are not being secure, then use OLDTASK
	 for the task_create RPC, in case it is something magical.  */
      e.error = task_create (((flags & EXEC_SECURE) ||
			      oldtask == MACH_PORT_NULL) ?
			     mach_task_self () : oldtask,
#ifdef KERN_INVALID_LEDGER
			       NULL, 0,	/* OSF Mach */
#endif
			     0, &newtask);
      if (e.error)
	goto out;
    }
  else
    newtask = oldtask;


  pthread_rwlock_rdlock (&std_lock);
  {
    /* Store the data that we will give in response
       to the RPC on the new task's bootstrap port.  */

    /* Set boot->portarray[IDX] to NEW.  If REAUTH is nonzero,
       io_reauthenticate NEW and set it to the authenticated port.
       If CONSUME is nonzero, a reference on NEW is consumed;
       it is invalid to give nonzero values to both REAUTH and CONSUME.  */
#define use(idx, new, reauth, consume) \
  do { use1 (idx, new, reauth, consume); \
       if (e.error) goto stdout; } while (0)
    void use1 (unsigned int idx, mach_port_t new,
	       int reauth, int consume)
      {
	if (new != MACH_PORT_NULL && reauth)
	  {
	    mach_port_t ref = mach_reply_port (), authed;
	    /* MAKE_SEND is safe here because we destroy REF ourselves. */
	    e.error = io_reauthenticate (new, ref, MACH_MSG_TYPE_MAKE_SEND);
	    if (! e.error)
	      e.error = auth_user_authenticate
		(boot->portarray[INIT_PORT_AUTH],
		 ref, MACH_MSG_TYPE_MAKE_SEND, &authed);
	    mach_port_destroy (mach_task_self (), ref);
	    if (e.error)
	      return;
	    new = authed;
	  }
	else
	  {
	    if (!consume && new != MACH_PORT_NULL)
	      mach_port_mod_refs (mach_task_self (),
				  new, MACH_PORT_RIGHT_SEND, 1);
	  }

	boot->portarray[idx] = new;
	ports_replaced[idx] = 1;
      }

    e.error = ports_create_port (execboot_portclass, port_bucket,
				 sizeof *boot, &boot);
    if (boot == NULL)
      {
      stdout:
	pthread_rwlock_unlock (&std_lock);
	goto out;
      }
    bzero (&boot->pi + 1, (char *) &boot[1] - (char *) (&boot->pi + 1));

    /* These flags say the information we pass through to the new program
       may need to be modified.  */
    secure = (flags & EXEC_SECURE);
    defaults = (flags & EXEC_DEFAULTS);

    /* Now record the big blocks of data we shuffle around unchanged.
       Whatever arrived inline, we must allocate space for so it can
       survive after this RPC returns.  */

    boot->flags = flags;

    argv = servercopy (argv, argvlen, argv_copy, &e.error);
    if (e.error)
      goto stdout;
    boot->argv = argv;
    boot->argvlen = argvlen;
    envp = servercopy (envp, envplen, envp_copy, &e.error);
    if (e.error)
      goto stdout;
    boot->envp = envp;
    boot->envplen = envplen;
    dtable = servercopy (dtable, dtablesize * sizeof (mach_port_t),
			 dtable_copy, &e.error);
    if (e.error)
      goto stdout;
    boot->dtable = dtable;
    boot->dtablesize = dtablesize;

    if ((secure || defaults) && nints < INIT_INT_MAX)
      {
	/* Make sure the intarray is at least big enough.  */
	if (intarray_copy || (round_page (nints * sizeof (int)) <
			      round_page (INIT_INT_MAX * sizeof (int))))
	  {
	    /* Allocate a new vector that is big enough.  */
	    boot->intarray = mmap (0, INIT_INT_MAX * sizeof (int),
				   PROT_READ|PROT_WRITE, MAP_ANON, 0, 0);
	    memcpy (boot->intarray, intarray, nints * sizeof (int));
	    intarray_dealloc = !intarray_copy;
	  }
	else
	  boot->intarray = intarray;
	boot->nints = INIT_INT_MAX;
      }
    else
      {
	intarray = servercopy (intarray, nints * sizeof (int), intarray_copy,
			       &e.error);
	if (e.error)
	  goto stdout;
	boot->intarray = intarray;
	boot->nints = nints;
      }

    if (secure)
      boot->intarray[INIT_UMASK] = std_ints ? std_ints[INIT_UMASK] : CMASK;

    /* Now choose the ports to give the new program.  */

    boot->nports = nports < INIT_PORT_MAX ? INIT_PORT_MAX : nports;
    boot->portarray = mmap (0, boot->nports * sizeof (mach_port_t),
			    PROT_READ|PROT_WRITE, MAP_ANON, 0, 0);
    /* Start by copying the array as passed.  */
    for (i = 0; i < nports; ++i)
      boot->portarray[i] = portarray[i];
    if (MACH_PORT_NULL != 0)
      for (; i < boot->nports; ++i)
	boot->portarray[i] = MACH_PORT_NULL;
    /* Keep track of which ports in BOOT->portarray come from the original
       PORTARRAY, and which we replace.  */
    ports_replaced = alloca (boot->nports * sizeof *ports_replaced);
    bzero (ports_replaced, boot->nports * sizeof *ports_replaced);

    if (portarray[INIT_PORT_BOOTSTRAP] == MACH_PORT_NULL &&
	oldtask != MACH_PORT_NULL)
      {
	if (! task_get_bootstrap_port (oldtask,
				       &boot->portarray[INIT_PORT_BOOTSTRAP]))
	  ports_replaced[INIT_PORT_BOOTSTRAP] = 1;
      }

    /* Note that the parentheses on this first test are different from the
       others below it. */
    if ((secure || defaults)
	&& boot->portarray[INIT_PORT_AUTH] == MACH_PORT_NULL)
      /* Q: Doesn't this let anyone run a program and make it
	 get a root auth port?
	 A: No; the standard port for INIT_PORT_AUTH has no UID's at all.
	 See init.trim/init.c (init_stdarrays).  */
      use (INIT_PORT_AUTH, std_ports[INIT_PORT_AUTH], 0, 0);
    if (secure || (defaults
		   && boot->portarray[INIT_PORT_PROC] == MACH_PORT_NULL))
      {
	/* Ask the proc server for the proc port for this task.  */
	mach_port_t new;

	e.error = proc_task2proc (procserver, newtask, &new);
	if (e.error)
	  goto stdout;
	use (INIT_PORT_PROC, new, 0, 1);
      }
    else if (oldtask != newtask && oldtask != MACH_PORT_NULL
	     && boot->portarray[INIT_PORT_PROC] != MACH_PORT_NULL)
      {
	mach_port_t new;
	/* This task port refers to the old task; use it to fetch a new
	   one for the new task.  */
	e.error = proc_task2proc (boot->portarray[INIT_PORT_PROC],
				  newtask, &new);
	if (e.error)
	  goto stdout;
	use (INIT_PORT_PROC, new, 0, 1);
      }
    if (secure || (defaults
		   && boot->portarray[INIT_PORT_CRDIR] == MACH_PORT_NULL))
      use (INIT_PORT_CRDIR, std_ports[INIT_PORT_CRDIR], 1, 0);
    if ((secure || defaults)
	&& boot->portarray[INIT_PORT_CWDIR] == MACH_PORT_NULL)
      use (INIT_PORT_CWDIR, std_ports[INIT_PORT_CWDIR], 1, 0);
  }
  pthread_rwlock_unlock (&std_lock);


  /* We have now concocted in BOOT the complete Hurd context (ports and
     ints) that the new program image will run under.  We will use these
     ports for looking up the interpreter file if there is one.  */

  if (! e.error && e.interp.section)
    {
      /* There is an interpreter section specifying another file to load
	 along with this executable.  Find the name of the file and open
	 it.  */

#ifdef BFD
      char namebuf[e.bfd ? e.interp.section->_raw_size : 0];
#endif
      char *name;

#ifdef BFD
      if (e.bfd)
	{
	  if (! bfd_get_section_contents (e.bfd, e.interp.section,
					  namebuf, 0,
					  e.interp.section->_raw_size))
	    {
	      e.error = b2he (errno);
	      name = NULL;
	    }
	  else
	    name = namebuf;
	}
      else
#endif
	{
	  name = map (&e, (e.interp.phdr->p_offset
			   & ~(e.interp.phdr->p_align - 1)),
		      e.interp.phdr->p_filesz);
	  if (! name && ! e.error)
	    e.error = ENOEXEC;
	}

      if (! name)
	e.interp.section = NULL;
      else
	{
	  /* Open the named file using the appropriate directory ports for
	     the user.  */
	  error_t user_port (int which, error_t (*operate) (mach_port_t))
	    {
	      return (*operate) (boot->nports > which ?
				 boot->portarray[which] :
				 MACH_PORT_NULL);
	    }
	  file_t user_fd (int fd)
	    {
	      if (fd < 0 || fd >= boot->dtablesize ||
		  boot->dtable[fd] == MACH_PORT_NULL)
		{
		  errno = EBADF;
		  return MACH_PORT_NULL;
		}
	      mach_port_mod_refs (mach_task_self (), boot->dtable[fd],
				  MACH_PORT_RIGHT_SEND, +1);
	      return boot->dtable[fd];
	    }
				/* XXX/fault */
	  e.error = hurd_file_name_lookup (&user_port, &user_fd, 0,
					   name, O_READ, 0, &interp.file);
	}
    }

  if (interp.file != MACH_PORT_NULL)
    {
      /* We opened an interpreter file.  Prepare it for loading too.  */
      prepare_and_check (interp.file, &interp);
      if (! interp.error)
	{
#ifdef	BFD
	  if (interp.bfd)
	    {
	      interp.info.bfd_locations = alloca (interp.bfd->section_count *
						  sizeof (vm_offset_t));
	      bfd_map_over_sections (interp.bfd, check_section, &e);
	    }
	  else
#endif
	    {
	      const ElfW(Phdr) *phdr = interp.info.elf.phdr;
	      interp.info.elf.phdr = alloca (interp.info.elf.phnum *
					     sizeof (ElfW(Phdr)));
	      check_elf_phdr (&interp, phdr);
	    }
	}
      e.error = interp.error;
    }

  if (e.error)
    goto out;


  /* We are now committed to the exec.  It "should not fail".
     If it does fail now, the task will be hopelessly munged.  */

  if (newtask == oldtask)
    {
      thread_t *threads;
      mach_msg_type_number_t nthreads, i;

      /* Terminate all the threads of the old task.  */

      e.error = task_threads (oldtask, &threads, &nthreads);
      if (e.error)
	goto out;
      for (i = 0; i < nthreads; ++i)
	{
	  thread_terminate (threads[i]);
	  mach_port_deallocate (mach_task_self (), threads[i]);
	}
      munmap ((caddr_t) threads, nthreads * sizeof (thread_t));

      /* Deallocate the entire virtual address space of the task.  */

      vm_deallocate (oldtask,
		     VM_MIN_ADDRESS, VM_MAX_ADDRESS - VM_MIN_ADDRESS);

      /* Nothing is supposed to go wrong any more.  If anything does, the
	 old task is now in a hopeless state and must be killed.  */
      oldtask_trashed = 1;

      /* Deallocate and destroy the ports requested by the caller.
	 These are ports the task wants not to lose if the exec call
	 fails, but wants removed from the new program task.  */

      for (i = 0; i < ndeallocnames; ++i)
	mach_port_deallocate (oldtask, deallocnames[i]);

      for (i = 0; i < ndestroynames; ++i)
	mach_port_destroy (oldtask, destroynames[i]);
    }

/* XXX this should be below
   it is here to work around a vm_map kernel bug. */
  if (interp.file != MACH_PORT_NULL)
    {
      /* Load the interpreter file.  */
      load (newtask, &interp);
      if (interp.error)
	{
	  e.error = interp.error;
	  goto out;
	}
      finish (&interp, 1);
    }


  /* Load the file into the task.  */
  load (newtask, &e);
  if (e.error)
    goto out;

  /* XXX loading of interp belongs here */

  /* Clean up.  */
  finish (&e, 0);

  /* Now record some essential addresses from the image itself that the
     program's startup code will need to know.  We do this after loading
     the image so that a load-anywhere image gets the adjusted addresses.  */
    if (e.info.elf.phdr_addr != 0)
      {
#ifdef BFD
	if (!e.bfd)
#endif
	  e.info.elf.phdr_addr += e.info.elf.loadbase;
	boot->phdr_addr = e.info.elf.phdr_addr;
	boot->phdr_size = e.info.elf.phnum * sizeof (ElfW(Phdr));
      }
  boot->user_entry = e.entry;	/* already adjusted in `load' */

  /* Create the initial thread.  */
  e.error = thread_create (newtask, &thread);
  if (e.error)
    goto out;

  /* Start up the initial thread at the entry point.  */
  boot->stack_base = 0, boot->stack_size = 0; /* Don't care about values.  */
  e.error = mach_setup_thread (newtask, thread,
			       (void *) (e.interp.section ? interp.entry :
					 e.entry),
			       &boot->stack_base, &boot->stack_size);
  if (e.error)
    goto out;
#ifdef BFD
  if (!e.bfd)
#endif
    {
      /* It would probably be better to change mach_setup_thread so
	 it does a vm_map with the right permissions to start with.  */
      if (!e.info.elf.execstack)
	e.error = vm_protect (newtask, boot->stack_base, boot->stack_size,
			      0, VM_PROT_READ | VM_PROT_WRITE);
    }


  if (oldtask != newtask && oldtask != MACH_PORT_NULL)
    {
      /* The program is on its way.  The old task can be nuked.  */
      process_t proc;
      process_t psrv;

      /* Use the canonical proc server if secure, or there is none other.
	 When not secure, it is nice to let processes associate with
	 whatever proc server turns them on, regardless of which exec
	 itself is using.  */
      if (secure
	  || boot->nports <= INIT_PORT_PROC
	  || boot->portarray[INIT_PORT_PROC] == MACH_PORT_NULL)
	psrv = procserver;
      else
	psrv = boot->portarray[INIT_PORT_PROC];

      /* XXX there is a race here for SIGKILLing the process. -roland
         I don't think it matters.  -mib */
      if (! proc_task2proc (psrv, oldtask, &proc))
	{
	  proc_reassign (proc, newtask);
	  mach_port_deallocate (mach_task_self (), proc);
	}
    }

  /* Make sure the proc server has the right idea of our identity. */
  if (secure)
    {
      uid_t euidbuf[10], egidbuf[10], auidbuf[10], agidbuf[10];
      uid_t *euids, *egids, *auids, *agids;
      size_t neuids, negids, nauids, nagids;
      error_t err;

      /* Find out what our UID is from the auth server. */
      neuids = negids = nauids = nagids = 10;
      euids = euidbuf, egids = egidbuf;
      auids = auidbuf, agids = agidbuf;
      err = auth_getids (boot->portarray[INIT_PORT_AUTH],
			     &euids, &neuids, &auids, &nauids,
			     &egids, &negids, &agids, &nagids);

      if (!err)
	{
	  /* Set the owner with the proc server */
	  /* Not much we can do about errors here; caller is responsible
	     for making sure that the provided proc port is correctly
	     authenticated anyhow. */
	  proc_setowner (boot->portarray[INIT_PORT_PROC],
			 neuids ? euids[0] : 0, !neuids);

	  /* Clean up */
	  if (euids != euidbuf)
	    munmap (euids, neuids * sizeof (uid_t));
	  if (egids != egidbuf)
	    munmap (egids, negids * sizeof (uid_t));
	  if (auids != auidbuf)
	    munmap (auids, nauids * sizeof (uid_t));
	  if (agids != agidbuf)
	    munmap (agids, nagids * sizeof (uid_t));
	}
    }

  {
    mach_port_t btport = ports_get_send_right (boot);
    e.error = task_set_bootstrap_port (newtask, btport);
    mach_port_deallocate (mach_task_self (), btport);
  }

 out:
  if (interp.file != MACH_PORT_NULL)
    finish (&interp, 1);
  finish (&e, !e.error);

  if (!e.error && (flags & EXEC_SIGTRAP)) /* XXX && !secure ? */
    {
      /* This is a "traced" exec, i.e. the new task is to be debugged.  The
	 caller has requested that the new process stop with SIGTRAP before
	 it starts.  Since the process has no signal thread yet to do its
	 own POSIX signal mechanics, we simulate it by notifying the proc
	 server of the signal and leaving the initial thread with a suspend
	 count of one, as it would be if the process were stopped by a
	 POSIX signal.  */
      mach_port_t proc;
      if (boot->nports > INIT_PORT_PROC)
	proc = boot->portarray[INIT_PORT_PROC];
      else
	/* Ask the proc server for the proc port for this task.  */
	e.error = proc_task2proc (procserver, newtask, &proc);
      if (!e.error)
	/* Tell the proc server that the process has stopped with the
	   SIGTRAP signal.  Don't bother to check for errors from the RPC
	   here; for non-secure execs PROC may be the user's own proc
	   server its confusion shouldn't make the exec fail.  */
	proc_mark_stop (proc, SIGTRAP, 0);
    }

  if (boot)
    {
      /* Release the original reference.  Now there is only one
	 reference, which will be released on no-senders notification.
	 If we are bailing out due to error before setting the task's
	 bootstrap port, this will be the last reference and BOOT
	 will get cleaned up here.  */

      if (e.error)
	/* Kill the pointers to the argument information so the cleanup
	   of BOOT doesn't deallocate it.  It will be deallocated my MiG
	   when we return the error.  */
	bzero (&boot->pi + 1, (char *) &boot[1] - (char *) (&boot->pi + 1));
      else
	/* Do this before we release the last reference.  */
	if (boot->nports > INIT_PORT_PROC)
	  proc_mark_exec (boot->portarray[INIT_PORT_PROC]);

      ports_port_deref (boot);
    }

  if (thread != MACH_PORT_NULL)
    {
      if (!e.error && !(flags & EXEC_SIGTRAP))
	thread_resume (thread);
      mach_port_deallocate (mach_task_self (), thread);
    }

  if (e.error)
    {
      if (oldtask != newtask)
	{
	  /* We created a new task but failed to set it up.  Kill it.  */
	  task_terminate (newtask);
	  mach_port_deallocate (mach_task_self (), newtask);
	}
      if (oldtask_trashed)
	/* The old task is hopelessly trashed; there is no way it
	   can resume execution.  Coup de grace.  */
	task_terminate (oldtask);
      else
	/* Resume the old task, which we suspended earlier.  */
	task_resume (oldtask);
    }
  else
    {
      if (oldtask != newtask)
	{
	  /* We successfully set the new task up.
	     Terminate the old task and deallocate our right to it.  */
	  task_terminate (oldtask);
	  mach_port_deallocate (mach_task_self (), oldtask);
	}
      else
	/* Resume the task, it is ready to run the new program.  */
	task_resume (oldtask);
      /* Deallocate the right to the new task we created.  */
      mach_port_deallocate (mach_task_self (), newtask);

      for (i = 0; i < nports; ++i)
	if (ports_replaced[i] && portarray[i] != MACH_PORT_NULL)
	  /* This port was replaced, so the reference that arrived in the
	     original portarray is not being saved in BOOT for transfer to
	     the user task.  Deallocate it; we don't want it, and MiG will
	     leave it for us on successful return.  */
	  mach_port_deallocate (mach_task_self (), portarray[i]);

      /* If there is vm_allocate'd space for the original intarray and/or
	 portarray, and we are not saving those pointers in BOOT for later
	 transfer, deallocate the original space now.  */
      if (intarray_dealloc)
	munmap (intarray, nints * sizeof intarray[0]);
      if (!portarray_copy)
	munmap (portarray, nports * sizeof portarray[0]);
    }

  return e.error;
}

kern_return_t
S_exec_exec (struct trivfs_protid *protid,
	     file_t file,
	     task_t oldtask,
	     int flags,
	     char *argv, mach_msg_type_number_t argvlen, boolean_t argv_copy,
	     char *envp, mach_msg_type_number_t envplen, boolean_t envp_copy,
	     mach_port_t *dtable, mach_msg_type_number_t dtablesize,
	     boolean_t dtable_copy,
	     mach_port_t *portarray, mach_msg_type_number_t nports,
	     boolean_t portarray_copy,
	     int *intarray, mach_msg_type_number_t nints,
	     boolean_t intarray_copy,
	     mach_port_t *deallocnames, mach_msg_type_number_t ndeallocnames,
	     mach_port_t *destroynames, mach_msg_type_number_t ndestroynames)
{
  if (! protid)
    return EOPNOTSUPP;

#if 0
  if (!(flags & EXEC_SECURE))
    {
      char *list = envz_get (envp, envplen, "EXECSERVERS");

      if (list)
	{
	  int tried = 0;
	  list = strdupa (list);
	  while ((p = strsep (&list, ":")))
	    {
	      /* Open the named file using the appropriate directory ports for
		 the user.  */
	      error_t user_port (int which, error_t (*operate) (mach_port_t))
		{
		  return (*operate) (nports > which
				     ? portarray[which] : MACH_PORT_NULL);
		}
	      file_t user_fd (int fd)
		{
		  if (fd < 0 || fd >= dtablesize ||
		      dtable[fd] == MACH_PORT_NULL)
		    {
		      errno = EBADF;
		      return MACH_PORT_NULL;
		    }
		  return dtable[fd];
		}
	      file_t server;
	      if (!hurd_file_name_lookup (user_port, user_fd, 0, p, 0,0, &server))
		{
		  error_t err;
		  struct trivfs_protid *protid
		    = ports_lookup_port (port_bucket, server,
					 trivfs_protid_portclasses[0]);
		  if (protid)
		    {
		      err = do_exec (file, oldtask, 0,
				     argv, argvlen, argv_copy,
				     envp, envplen, envp_copy,
				     dtable, dtablesize, dtable_copy,
				     portarray, nports, portarray_copy,
				     intarray, nints, intarray_copy,
				     deallocnames, ndeallocnames,
				     destroynames, ndestroynames);
		      ports_port_deref (protid);
		    }
		  else
		    {
		      int n;
		      err = exec_exec (server,
				       file, MACH_MSG_TYPE_COPY_SEND,
				       oldtask, 0,
				       argv, argvlen,
				       envp, envplen,
				       dtable, MACH_MSG_TYPE_COPY_SEND,
				       dtablesize,
				       portarray, MACH_MSG_TYPE_COPY_SEND,
				       nports,
				       intarray, nints,
				       deallocnames, ndeallocnames,
				       destroynames, ndestroynames);
		      mach_port_deallocate (mach_task_self (), file);
		      for (n = 0; n < dtablesize; n++)
			mach_port_deallocate (mach_task_self (), dtable[n]);
		      for (n = 0; n < nports; n++)
			mach_port_deallocate (mach_task_self (), portarray[n]);
		    }
		  mach_port_deallocate (mach_task_self (), server);
		  if (err != ENOEXEC)
		    return err;
		  tried = 1;
		}
	    }
	  if (tried)
	    /* At least one exec server got a crack at it and gave up.  */
	    return ENOEXEC;
	}
    }
#endif

  /* There were no user-specified exec servers,
     or none of them could be found.  */

  return do_exec (file, oldtask, flags,
		  argv, argvlen, argv_copy,
		  envp, envplen, envp_copy,
		  dtable, dtablesize, dtable_copy,
		  portarray, nports, portarray_copy,
		  intarray, nints, intarray_copy,
		  deallocnames, ndeallocnames,
		  destroynames, ndestroynames);
}

kern_return_t
S_exec_setexecdata (struct trivfs_protid *protid,
		    mach_port_t *ports, mach_msg_type_number_t nports, int ports_copy,
		    int *ints, mach_msg_type_number_t nints, int ints_copy)
{
  error_t err;

  if (! protid || (protid->realnode != MACH_PORT_NULL && ! protid->isroot))
    return EPERM;

  if (nports < INIT_PORT_MAX || nints < INIT_INT_MAX)
    return EINVAL;		/*  */

  err = 0;
  ports = servercopy (ports, nports * sizeof (mach_port_t), ports_copy, &err);
  if (err)
    return err;
  ints = servercopy (ints, nints * sizeof (int), ints_copy, &err);
  if (err)
    {
      munmap (ports, nports * sizeof (mach_port_t));
      return err;
    }

  pthread_rwlock_wrlock (&std_lock);

  if (std_ports)
    {
      mach_msg_type_number_t i;
      for (i = 0; i < std_nports; ++i)
	mach_port_deallocate (mach_task_self (), std_ports[i]);
      munmap (std_ports, std_nports * sizeof (mach_port_t));
    }

  std_ports = ports;
  std_nports = nports;

  if (std_ints)
    munmap (std_ints, std_nints * sizeof (int));

  std_ints = ints;
  std_nints = nints;

  pthread_rwlock_unlock (&std_lock);

  return 0;
}


#include "exec_startup_S.h"

/* RPC sent on the bootstrap port.  */

kern_return_t
S_exec_startup_get_info (mach_port_t port,
			 vm_address_t *user_entry,
			 vm_address_t *phdr_data, vm_size_t *phdr_size,
			 vm_address_t *stack_base, vm_size_t *stack_size,
			 int *flags,
			 char **argvp, mach_msg_type_number_t *argvlen,
			 char **envpp, mach_msg_type_number_t *envplen,
			 mach_port_t **dtable,
			 mach_msg_type_name_t *dtablepoly,
			 mach_msg_type_number_t *dtablesize,
			 mach_port_t **portarray,
			 mach_msg_type_name_t *portpoly,
			 mach_msg_type_number_t *nports,
			 int **intarray, mach_msg_type_number_t *nints)
{
  struct bootinfo *boot = ports_lookup_port (port_bucket, port,
					     execboot_portclass);
  if (! boot)
    return EOPNOTSUPP;
  ports_port_deref (boot);

  /* Pass back all the information we are storing.  */

  *user_entry = boot->user_entry;
  *phdr_data = boot->phdr_addr;
  *phdr_size = boot->phdr_size;
  *stack_base = boot->stack_base;
  *stack_size = boot->stack_size;

  *argvp = boot->argv;
  *argvlen = boot->argvlen;
  boot->argvlen = 0;

  *envpp = boot->envp;
  *envplen = boot->envplen;
  boot->envplen = 0;

  *dtable = boot->dtable;
  *dtablesize = boot->dtablesize;
  *dtablepoly = MACH_MSG_TYPE_MOVE_SEND;
  boot->dtablesize = 0;

  *intarray = boot->intarray;
  *nints = boot->nints;
  boot->nints = 0;

  *portarray = boot->portarray;
  *nports = boot->nports;
  *portpoly = MACH_MSG_TYPE_MOVE_SEND;
  boot->nports = 0;

  *flags = boot->flags;

  return 0;
}