Date: Sat, 16 Sep 2017 12:13:22 -0400 From: Rich Felker <dalias@...c.org> To: "A. Wilcox" <awilfox@...lielinux.org> Cc: Jeff King <peff@...f.net>, Kevin Daudt <me@...e.info>, git@...r.kernel.org, musl@...ts.openwall.com Subject: Re: Re: Git 2.14.1: t6500: error during test on musl libc 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, 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 126.96.36.199 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
Powered by blists - more mailing lists
Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.