Re: PATCH: PR binutils/12283: bfd/doc doesn't support parallel build

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Thu, Dec 2, 2010 at 10:32 AM, Ralf Wildenhues <Ralf.Wildenhues@gmx.de> wrote:
> * Ralf Wildenhues wrote on Thu, Dec 02, 2010 at 07:22:35PM CET:
>> * Ralf Wildenhues wrote on Thu, Dec 02, 2010 at 07:20:50PM CET:
>> > * H.J. Lu wrote on Thu, Dec 02, 2010 at 07:16:07PM CET:
>> > > Failure is very random. There is a race condition. "make -j8" on a 16core
>> > > machine can start making $(MKDOC) at the same time. How can move-if-change
>> > > be 100% atomic?
>> >
>> > Well, 'mv -f' within a mount point is atomic, that's how.
>>
>> Ahh, move-if-change is broken: it doesn't ignore mv -f failure when the
>> target is equal. ?Let's fix move-if-change.
>
> Does this alternative (untested) patch fix the race?
>
> Thanks,
> Ralf
>
> ChangeLog:
> 2010-12-02 ?Ralf Wildenhues ?<Ralf.Wildenhues@gmx.de>
>
> ? ? ? ?PR binutils/12283
> ? ? ? ?* move-if-change: Ignore failure of 'mv -f' if the destination
> ? ? ? ?is equal to the source afterwards.
>
> diff --git a/move-if-change b/move-if-change
> index ff74a55..1d2dd09 100755
> --- a/move-if-change
> +++ b/move-if-change
> @@ -4,6 +4,8 @@
>
> ?usage="$0: usage: $0 SOURCE DEST"
>
> +cmpprog=${CMPPROG-cmp}
> +
> ?case $# in
> ?2) ;;
> ?*) echo "$usage" >&2; exit 1;;
> @@ -18,5 +20,8 @@ done
> ?if test -r "$2" && cmp -s "$1" "$2"; then
> ? rm -f "$1"
> ?else
> - ?mv -f "$1" "$2"
> + ?if mv -f "$1" "$2"; then :; else
> + ? ?# Ignore failure due to a concurrent move-if-change.
> + ? ?$cmpprog "$1" "$2" >/dev/null 2>&1
> + ?fi
> ?fi
>

It is hard to test since it happens very much at random.


-- 
H.J.