thanks for checking into this
the next release ksh93/tests/io.sh will take the conventions into account
---
set \
/proc/self/fd open \
/proc/$$/fd open
fdfs=/dev/fd
semantics=dup
while (( $# >= 2 ))
do if [[ -e $1 ]]
then fdfs=$1
semantics=$2
break
fi
shift 2
done
exec 3<> file1
if command exec 4< $fdfs/3
then read -u3 line
read -u4 line
exp=foo
case $line in
foo) got=open ;;
bar) got=dup ;;
*) got=error ;;
esac
if [[ $semantics != $got ]]
then err_exit "'4< $fdfs/3' uses $got(2) instead of $semantics(2) --
expected '$exp', got '$line'"
fi
fi
# 2004-11-25 ancient /dev/fd/N redirection bug fix
x=$(
{
print -n 1
print -n 2 > $fdfs/2
print -n 3
print -n 4 > $fdfs/2
} 2>&1
)
case $x in
1234) got=dup ;;
4) got=open ;;
*) got=error ;;
esac
case $semantics in
dup) exp=1234 ;;
*) exp=4 ;;
esac
[[ $x == $exp ]] || err_exit "$fdfs/N uses $got(2) instead of $semantics(2) --
expected '$exp', got '$x'"
---
On Fri, 12 Dec 2008 17:50:39 -0600 Nicolas Williams wrote:
> On Fri, Dec 12, 2008 at 05:36:36PM -0600, Nicolas Williams wrote:
> > So, yes, open(2) of /dev/fd/N is as a dup(2) of N.
> Also, every reference I can find says that open(2) of /dev/fd/N is
> supposed to be a dup(2) of file descriptor N, on Solaris, on Linux, on
> BSD, on Plan 9.
> And, apparently also the convention is that open(2) of /proc/self/fd/N
> is NOT equivalent to dup(2) of fd N (it opens the same file, device,
> whatever, but does not keep the offset from fildes N).
> % (read line <#((200)) ; print $line ; read line
> </proc/self/fd/9 ; print $line) <&9
> 1041
> 1001
> %
> And that's supposed to be so for Linux and Solaris (haven't checked
> BSD). An open(2) of /proc/self/fd/N is supposed to check file
> permissions too:
> % pfexec chown root /tmp/foo
> % pfexec chmod 600 /tmp/foo
> % (read line <#((200)) ; print $line ; read line
> </proc/self/fd/9 ; print $line) <&9
> 1041
> ksh93: /proc/self/fd/9: cannot open [Permission denied]
> 1041
> %
