Re: False positives in git diff-index

From: Jeff King <peff@peff.net>
Date: Jan 5, 2011, 05:48 UTC
Message-ID: <20110105054825.GC5884@sigill.intra.peff.net>
In-Reply-To: <AANLkTinDSCPz-oukxzn24hj94d9WpzZ8_64TBHeNTmoG@mail.gmail.com>

On Tue, Jan 04, 2011 at 12:45:33PM +0300, Alexander Gladysh wrote:

Show 23 quoted lines

> > Anyway, I'm ready to debug this issue if someone will guide me.
> >
> > Workflow:
> >
> > <...change files in /path/dir1/...>
> > (cd /path && git add </path/dir1/>)
> > (cd /path && git commit -m <message1>)
> >
> > ... repeat change-add-commit several times for various directories
> > (can be the same directory or not) ...
> >
> > <...generate file /path/dirN/foo...>
> > # Accidentally the file is generated the same as it was
> >
> > (cd /path && git add </path/dirN/>)
> > (cd /path && git status) # Refresh index
> > (cd /path && git diff-index --exit-code --quiet HEAD -- /path/dirN) #
> > Incorrectly reports that there are some changes
> > (cd /path && git commit -m <messageN>) # fails, saying that there is
> > nothing to commit
> >
> > If I insert sleep 10 between git status and git diff-index, the
> > problem goes away.

If adding a sleep makes it work, that sounds like a race condition in git. But from the description of your workflow, it should be easy to make a minimal example:

-- >8 -- #!/bin/sh

random() {
  perl -e 'print int(rand(5))+1, "\n"'
}

rm -rf repo mkdir repo && cd repo && git init

for i in 1 2 3 4 5; do
  mkdir dir$i
  echo initial >dir$i/file
done
git add .
git commit -m initial

while true; do
  for i in 1 2 3 4 5; do
    random >dir$i/file
    git add dir$i
    git update-index --refresh
    if ! git diff-index --exit-code --quiet HEAD -- dir$i; then
      if ! git commit -m foo; then
        echo breakage
        exit 1
      fi
    else
      echo not bothering to commit
    fi
  done
done
-- 8< --

Basically, we generate random data which has a 20% chance of being the same as what's there. When it is, we should get "not bothering to commit", but in your error case, we would try to commit (and get "no changes").

But using that script, I can't replicate your problem. Can you try running it on the same box you're having trouble with? That might at least tell us if it's your environment or something more complex going on.

-Peff