[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: patch for mmap and friends

From: Maxime Devos
Subject: Re: patch for mmap and friends
Date: Sat, 14 Jan 2023 23:42:30 +0100
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101 Thunderbird/102.6.0

On 14-01-2023 23:08, Matt Wette wrote:
2) had to copy/modify dynwind_acquire_port and release_port from ports.c

Is it because of the 'static' qualifier? If so, you could use the 'SCM_INTERNAL [...] scm_i_[...]' pattern, e.g. see 'scm_i_is_mutable_bitvector' in libguile/bitvectors.h and libguile/bitvectors.c.

   if (SCM_UNBNDP (file))
    c_fd = -1;
  else {
    /* Use the fd under clobber protection from GC or another thread. */
    if (SCM_PORTP (file))
      c_fd = scm_to_int (scm_fileno (file));
    else {
      c_fd = scm_to_int (file);
      file = SCM_CAR (scm_fdes_to_ports (file));
    scm_dynwind_acquire_port (file);

You need to acquire the port _before_ taking its file descriptor! Otherwise, it is protected too late. I.e., scm_dynwind_acquire_port needs to be moved before the 'scm_fileno'.

On the second branch, in particular 'c_fd = scm_to_int (file);':
IIUC, the idea is to, when a raw fd is passed, look up the corresponding port to lock it, right?

If so, I think it's too late for that -- another thread might change things between 'c_fd = ...' and 'file = SCM_CAR (scm_fdes_to_ports (file))'.

More generally, when a raw fd is passed, I think it's impossible to solve the 'other thread/GC interfering' problem.

As such, to simplify things, I propose to only do the 'scm_dynwind_acquire_port' when a port is passed, instead of failing to solve the interference problems (if the user passed a raw fd, then only they can make sure there are no problems, e.g. by changing their code to use ports or by not using move->fdes stuff).:

Proposed code (untested):

  void *c_mem, *c_addr;
  size_t c_len;
  int c_prot, c_flags, c_fd;
  scm_t_off c_offset;
  SCM pointer, bvec;

  if (SCM_POINTER_P (addr))
    c_addr = SCM_POINTER_VALUE (addr);
  else if (scm_is_integer (addr))
    c_addr = (void*) scm_to_uintptr_t (addr);
    scm_misc_error ("mmap", "bad addr", addr);

  c_len = scm_to_size_t (len);

  if (SCM_UNBNDP (prot))
    c_prot = PROT_READ | PROT_WRITE;
    c_prot = scm_to_int (prot);

  if (SCM_UNBNDP (flags))
    c_flags = scm_to_int (flags);

  scm_dynwind_begin (0);
  if (SCM_UNBNDP (file))
    c_fd = -1;
  else if (scm_is_integer (file))
    c_fd = scm_to_int (file);
      /* Use the fd of the port under clobber protection from
         concurrency. As scm_dynwind_acquire_port assumes that
         FILE is a port, check that first. */
      scm_dynwind_acquire_port (file);
      c_fd = scm_fileno (file);

  if (SCM_UNBNDP (offset))
    c_offset = 0;
    c_offset = scm_to_off_t (offset);

  if ((c_addr == NULL) && (c_flags & MAP_FIXED))
    scm_misc_error ("mmap", "cannot have NULL addr w/ MAP_FIXED", SCM_EOL);

SCM_SYSCALL (c_mem = mmap(c_addr, c_len, c_prot, c_flags, c_fd, c_offset));
  if (c_mem == MAP_FAILED)
    scm_syserror ("mmap");              /* errno set */

  /* The fd is free to go now. */
  scm_dynwind_end ();

  pointer = scm_cell (scm_tc7_pointer, (scm_t_bits) c_mem);
bvec = scm_c_take_typed_bytevector((signed char *) c_mem + c_offset, c_len,
                     SCM_ARRAY_ELEMENT_TYPE_VU8, pointer);
  scm_i_set_finalizer (SCM2PTR (bvec), mmap_finalizer, (void*) c_len);
  return bvec;
#undef FUNC_NAME

Attachment: OpenPGP_0x49E3EE22191725EE.asc
Description: OpenPGP public key

Attachment: OpenPGP_signature
Description: OpenPGP digital signature

reply via email to

[Prev in Thread] Current Thread [Next in Thread]