Bug 653054

i just had a look at the tcsh src code and found that the authors r aware of this bug (they suggest to do certain things (open, create, dump) atomically)...

a fix seems to be complicated (e g dosource() doesnt allow a "do it atomically" parameter)... :-)

i wasnt aware of that RHEL-5 bug report... maybe i should have waited just some years longer (that bug is very ooooold)... :-)) sorry for the trouble...

-arne

As you said - to do that atomically is not so easy - and authors are aware of the issue. Vojta is new tcsh maintainer and he is now trying to make a patch and propose it to upstream (independently on this report) ... so hopefully it's not a matter of years...

Hopefully not :)

i made a workaround: (a little blunt *blush* but it works)

> grep histfile .cshrc
set histfile=`/usr/local/bin/csh-hist-workaround.sh $$`
> cat devel/chw/csh-hist-workaround.sh 
#!/bin/sh
# $Id$

cd /tmp
lockfile history.lock
touch history.fst
ls -tr history.[1-9]*.?????????? 2> /dev/null | \
while read f ; do
	pid=`echo $f | sed -e 's:^history\.::' -e 's:\..*::'`
	if pgrep tcsh | fgrep -q $pid ; then
		continue
	fi
	if [ -f history.fst ] ; then
		rm history.fst
		mv $f history
	else
		cat $f >> history
		rm $f
	fi
done
fn=`mktemp /tmp/history.$1.XXXXXXXXXX`
if [ -f history ] ; then
	cp history $fn
fi
rm -f history.fst
rm -f history.lock
echo -n $fn
> 

-arne

it should be
grep -q ^$pid'$'
instead of
fgrep -q $pid

-arne

here is an update of the workaround (its disk complexity is slightly better):
lockfile /tmp/history.lock
set histfile=`mktemp /tmp/history.$$.XXXXXXXXXX`
set history=2000
set histdup=erase
set savehist=(2000 merge)
test -f /tmp/history && history -L /tmp/history
foreach f ( `ls -tr /tmp/history.[1-9]*.??????????` ) >& /dev/null
   history -M $f
   history -S /tmp/history
   test -e /proc/`echo $f | sed -e 's:^/tmp/history\.::' -e 's:\..*::'` || rm -f $f
end
rm -f /tmp/history.lock

-arne

i found a further optimization:
lockfile -1 /tmp/history.lock
set history=2000
set histdup=erase
set savehist=(2000 merge)
test -f /tmp/history && history -L /tmp/history
set hsn=""
foreach f ( `ls -tr /tmp/history.[1-9]*.??????????` ) >& /dev/null
   test -e /proc/`echo $f|sed -e 's:^/tmp/history\.::' -e 's:\..*::'` && continue
   history -M $f
   rm -f $f
   set hsn=1
end
set histfile=`mktemp /tmp/history.$$.XXXXXXXXXX`
test -z "$hsn" || history -S /tmp/history
rm -f /tmp/history.lock

-arne

Arne, thank you for your help. Unfortunately, we can't use such workarounds as a final solution.

Created attachment 530274 [details]
solution that needs an additional lock file

i found a new workaround...
it uses an additional lockfile, because: dosource (or so) seems to drop the lock to the history file...

Created attachment 562503 [details]
fcntl .history file locking - shared readers, exclusive writer

congratulations!
looks like a voluminous fix... :-)
but it is more elegant with a lock on the history file...
(i mean: an extra file (like in my workaround) is quite clumsy...)

-arne

Fixed in Rawhide. Builds for stable Fedora releases coming probably with few other fixes..

(In reply to comment #13)
> congratulations!
> looks like a voluminous fix... :-)
> but it is more elegant with a lock on the history file...
> (i mean: an extra file (like in my workaround) is quite clumsy...)

Thanks Arne! And thank you again for your help, I really appreciated that.

oops
it still happens that the .history file is empty...
i think it happens when the X session is killed (e. g. with ctrl+alt+backspace)...

is it possible, that tcsh doesnt ignore signals during saving history?
or does it get a SIGKILL (which cant be ignored IIRC) (but why does it delete the history file then)?

currently i use tcsh.x86_64 (6.17-18.fc17)...

Arne, I'm going to release new tcsh Fedora builds with the latest .history file locking changes that came from the RHEL-6 QA/testing phase. Stay tunned for the new builds - I'd be happy to hear your feedback then again.

oh - ok - i thought the fix would be part of fc17 already... :-) -arne

is it fixed in tcsh.x86_64 6.17-18.fc17?
i lost my history file again...

did the patch reach fc18?
or what does "status: MODIFIED" mean?

i had to switch to bash, because: tcsh continued to destroy my .history file...

Vojtech no longer maintains the package, I moved it to current maintainer.

Hi, the patch made by Vojtech is not in f17, only in f18.