On Fri, Sep 15, 2017 at 11:58:41PM -0500, A. Wilcox wrote:
> On 15/09/17 06:30, Jeff King wrote:
> > On Fri, Sep 15, 2017 at 08:37:40AM +0200, Kevin Daudt wrote:
> >
> >> On Thu, Sep 14, 2017 at 09:43:12PM -0500, A. Wilcox wrote:
> >>> -BEGIN PGP SIGNED MESSAGE-
> >>> Hash: SHA256
> >>>
> >>> Hi there,
> >>>
> >>> While bumping Git's version for our Linux distribution to 2.14.1, I've
> >>> run in to a new test failure in t6500-gc.sh. This is the output of
> >>> the failing test with debug=t verbose=t:
> >>
> >> This is a new test introduced by c45af94db
> >> (gc: run pre-detach operations under lock, 2017-07-11) which was
> >> included in v2.14.0.
> >>
> >> So it might be that this was already a problem for a longer time, only
> >> just recently uncovered.
> >
> > The code change there is not all that big. Mostly we're just checking
> > that the lock is actually respected. The lock code doesn't exercise libc
> > all that much. It does use fscanf, which I guess is a little exotic for
> > us. It's also possible that hostname() doesn't behave quite as we
> > expect.
> >
> > If you instrument gc like the patch below, what does it report when you
> > run:
> >
> > GIT_TRACE=1 ./t6500-gc.sh --verbose-only=8
> >
> > I get:
> >
> > [...]
> > trace: built-in: git 'gc' '--auto'
> > Auto packing the repository in background for optimum performance.
> > See "git help gc" for manual housekeeping.
> > debug: gc lock already held by $my_hostname
> > [...]
> >
> > If you get "acquired gc lock", then the problem is in
> > lock_repo_for_gc(), and I'd suspect some problem with fscanf or
> > hostname.
> >
> > -Peff
>
>
> Hey there Peff,
>
> What a corner-y corner case we have here. I believe the actual error is
> in the POSIX standard itself[1], as it is not clear what happens when
> there are not enough characters to 'fill' the width specified with %c in
> fscanf:
ISO C specifies very clearly what happens, in 7.21.6.2 The fscanf
function, paragraph 12:
c
Matches a sequence of characters of exactly the number
specified by the field width...
Note the word "exactly". Thus a read of fewer characters is not a
match.
There is an open glibc bug for this with classic Drepper behavior
until his departure, followed by acknowledgement of the bug, but no
further action I'm aware of:
https://sourceware.org/bugzilla/show_bug.cgi?id=12701
Any applications depending on the buggy glibc behavior should be
fixed.
Rich