libio fflush() bug

[Ulrich Drepper]

Thorsten Kukuk <kukuk@weber-eb.uni-paderborn.de> writes:

> I have found a big problem. If you do fseek/fread/fflush, the file pointer
> is set to the end of the file. If you don't call fseek, the pointer is set
> correct to the next byte after the last read one.

I've looked at this closer and the first thing I came across is a
kernel bug is the 2.1.xxx kernels.  Linus already changed his tree.
The problem is in fs/read_write.c:109:

	retval = (int)offset & INT_MAX;

This should read

	retval = (int)offset;

In case of an error the & removes the MSB and therefore the libc does
not recognize this as an error.

But even with this the result is a bit suspecious (though Thorsten's
test case works now):

_llseek(3, 0, {0}, SEEK_SET)            = 0
read(3, "#include <stdio.h>\n\nint\nmain "..., 4096) = 405
_llseek(3, -401, 0xbffff690, SEEK_CUR)  = -1 EINVAL (Invalid argument)

Why should the second _llseek fail?

(The above result were produced with a new version of strace which
understands _llseek and which I"ll upload to
ftp.cygnus.com/pub/home/drepper in a few minutes).

-- 
---------------.      drepper at gnu.org  ,-.   1325 Chesapeake Terrace
Ulrich Drepper  \    ,-------------------'   \  Sunnyvale, CA 94089 USA
Cygnus Solutions `--' drepper at cygnus.com   `------------------------