kern/164793: 'write' system call violates POSIX standard

Bruce Evans brde at optusnet.com.au
Sun Feb 5 18:54:54 UTC 2012


On Sun, 5 Feb 2012, Nicolas Bourdaud wrote:

>> Description:
> When a write() cannot transfer as many bytes as requested (because of a file
> limit), it fails instead of transferring as many bytes as there is room to
> write.
>
> This is a violation of the POSIX standard:
> http://pubs.opengroup.org/onlinepubs/007904975/functions/write.html

FreeBSD's handling of the maxfilesize limits is similar, so it has the
same bug.

This affects many fileystems which copied the buggy code from ffs.
(Both truncate() and write() fail if extending to or writing the full
number of bytes would exceed the limit.  This is correct for truncate(),
but write() is required to creep up on the limit.)

I think this is actually a bug in POSIX (XSI).  Most programs aren't
prepared to deal with short writes, and returning an error like
truncate() is specified to is adequate.  For regular files, most file
systems in FreeBSD back out of writes after an i/o error, using
ftruncate() (some truncation is necessary for security, since the place
at which the error occurred is usually not known precisely), so the
following bug in the upper layer rarely matters.  From an old version
of sys_generic.c, for writing (reading has a similar bug):

% 	if ((error = fo_write(fp, &auio, td->td_ucred, flags, td))) {
% 		/* XXX short write botch. */
% 		if (auio.uio_resid != cnt && (error == ERESTART ||
% 		    error == EINTR || error == EWOULDBLOCK))
% 			error = 0;

The XXX comment is only in my version.  Here (auio.uio_resid != cnt)
means that some i/o was done.  In that case, write() is required to
return the amount done, with no error, which is implemented by setting
`error' to 0.  But this is only done if `error' is one of ERESTART,
EINTR or EWOULDBLOCK.  At least the case of the most common error that
is not one of these, namely EIO, is broken.  The handling of the special
3 here is delicate:
- ERESTART: hopefully can't happen, since if it happens then we should
   restart.  This error is a non-error that in most cases means that the
   we handled a signal but are not returning with EINTR because SA_RESTART
   says to restart instead of returning.
- EINTR: since we have this and not ERESTART, it is clearly correct to
   return, but if we did some i/o then we must return its amount and there
   is no way to return EINTR.
- EWOULDBLOCK: similar to EINTR for a SIGALRM, but more precise.  I guess
   this is here since it is the only other common error, and it is not
   really an error so failing for it would be obviously wrong (except
   when no i/o was done, EWOULDBLOCK = EAGAIN is the standard way to
   indicate this).

The flag that controls backing out of writes is IO_UNIT.  This is always
set for write(2), and probably should be set unconditionally (so it
shouldn't exist), since not setting it mainly asks for security holes
and most cases are write(2) anyway.  IO_UNIT means that the i/o is done
as an "atomic unit".  The semantics of "unit" probably includes doing
all of it or none of it, so it would have to be broken to match the
POSIX spec.

> Patch attached with submission follows:
> ...
> int main(void)
> {
> 	struct rlimit lim;
> 	int fd;
> 	ssize_t retc;
> 	size_t count = 0;
> 	const char pattern[PATTSIZE] = "Hello world!";
>
> 	signal(SIGXFSZ, SIG_IGN);
> 	lim.rlim_cur = LIMSIZE;
> 	setrlimit(RLIMIT_FSIZE, &lim);

This is missing initialization of at least lim.rlim_max in lim.  This
gave the bizarre behaviour that when the program was statically linked,
it failed for the first write, because the stack garbage for
lim.rlim_max happened to be 0.

Bruce


More information about the freebsd-bugs mailing list