Re: [PATCH v6 16/21] mingw: try to work around issues with the test cleanup

From: "SZEDER Gábor" <szeder.dev@gmail.com>
To: "Ævar Arnfjörð Bjarmason" <avarab@gmail.com>
Cc: Johannes Schindelin <Johannes.Schindelin@gmx.de>,
	Johannes Schindelin via GitGitGadget <gitgitgadget@gmail.com>,
	git@vger.kernel.org, Junio C Hamano <gitster@pobox.com>,
	Jeff King <peff@peff.net>
Subject: Re: [PATCH v6 16/21] mingw: try to work around issues with the test cleanup
Date: Fri, 26 Mar 2021 20:54:17 +0100	[thread overview]
Message-ID: <20210326195417.GB2271@szeder.dev> (raw)
In-Reply-To: <87lfacp3bp.fsf@evledraar.gmail.com>

On Wed, Mar 24, 2021 at 11:57:46PM +0100, Ævar Arnfjörð Bjarmason wrote:
> 
> On Wed, Mar 24 2021, SZEDER Gábor wrote:
> 
> > On Wed, Mar 24, 2021 at 01:01:49PM +0100, Johannes Schindelin wrote:
> >> Hi Ævar,
> >> 
> >> On Sat, 20 Mar 2021, Ævar Arnfjörð Bjarmason wrote:
> >> 
> >> > On Fri, Mar 19 2021, Johannes Schindelin wrote:
> >> >
> >> > >> On Tue, Jan 29 2019, Johannes Schindelin via GitGitGadget wrote:
> >> > >>
> >> > >> > From: Johannes Schindelin <johannes.schindelin@gmx.de>
> >> > >> >
> >> > >> > It seems that every once in a while in the Git for Windows SDK, there
> >> > >> > are some transient file locking issues preventing the test clean up to
> >> > >> > delete the trash directory. Let's be gentle and try again five seconds
> >> > >> > later, and only error out if it still fails the second time.
> >> > >> >
> >> > >> > This change helps Windows, and does not hurt any other platform
> >> > >> > (normally, it is highly unlikely that said deletion fails, and if it
> >> > >> > does, normally it will fail again even 5 seconds later).
> >> > >> >
> >> > >> > Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
> >> > >> > ---
> >> > >> >  t/test-lib.sh | 6 +++++-
> >> > >> >  1 file changed, 5 insertions(+), 1 deletion(-)
> >> > >> >
> >> > >> > diff --git a/t/test-lib.sh b/t/test-lib.sh
> >> > >> > index f31a1c8f79..9c0ca5effb 100644
> >> > >> > --- a/t/test-lib.sh
> >> > >> > +++ b/t/test-lib.sh
> >> > >> > @@ -1104,7 +1104,11 @@ test_done () {
> >> > >> >  			error "Tests passed but trash directory already removed before test cleanup; aborting"
> >> > >> >
> >> > >> >  			cd "$TRASH_DIRECTORY/.." &&
> >> > >> > -			rm -fr "$TRASH_DIRECTORY" ||
> >> > >> > +			rm -fr "$TRASH_DIRECTORY" || {
> >> > >> > +				# try again in a bit
> >> > >> > +				sleep 5;
> >> > >> > +				rm -fr "$TRASH_DIRECTORY"
> >> > >> > +			} ||
> >> > >> >  			error "Tests passed but test cleanup failed; aborting"
> >> > >> >  		fi
> >> > >> >  		test_at_end_hook_
> >> > >>
> >> > >> I saw this sleep while reading some test-lib.sh code, doesn't this break
> >> > >> df4c0d1a79 (test-lib: abort when can't remove trash directory,
> >> > >> 2017-04-20) for non-Windows platforms?
> >> > >
> >> > > It does not really break it, it just delays the inevitable failure.
> >> 
> >> I still think this is the best answer to this (implicit) question:
> >> 
> >> > In any case, your patch clearly undoes whatever canary for gc issues
> >> > df4c0d1a792 was trying to put into the test-lib, but didn't say so in
> >> > its commit message.
> >> 
> >> I was not _really_ paying attention to that commit when I implemented the
> >> work-around you mentioned above. At the same time I think it does _not_
> >> undo the canary. If the trash directory cannot be removed via `rm -fr`,
> >> and if that is an indicator for something fishy going on, chances are that
> >> the second `rm -fr` a couple seconds later will _also_ fail, and we still
> >> get that error message.
> >> 
> >> The only reason why the second `rm` should succeed, at least that I can
> >> think of, is that something on Windows blocked those files from being
> >> deleted, and it is no longer blocking after a couple seconds, and that
> >> usually means that an anti-malware scanned those files.
> >
> > Both commits referenced in df4c0d1a79's log message fixed races
> > between 'test_done's cleanup and a still running background 'git gc',
> > and df4c0d1a79 was meant to draw our attention to similar issues in
> > the future.  And it did:
> >
> >   https://public-inbox.org/git/20190602091919.GN951@szeder.dev/
> >   
> > So no, the failure is not inevitable, there are other reasons why the
> > second 'rm' might still succeed after the first failed, even just a
> > fraction of a second later.  And yes, that 'sleep' added in the patch
> > above did unquestionably break that canary,
> 
> Having read that thread now I agree, but I also left with a "who cares?"
> and "so let's keep the sleep then?".
> 
> I.e. is this a problem that any of the software we're maintaining is
> going to care about in the wild, it's not like people are expecting gc,
> repack, fast-import etc. to behave well in the face of rm -rfing the
> directory they're operating on.
> 
> So it seems like just an issue that crops up because of how our test
> suite manages and removes per-test trash directories.

Not at all.  The real problem is that some stray background git
process is *still* actively writing to the test repository when the
test script is already supposed to be finished.

> So it seems better
> to:
> 
>  1. Just keep that "sleep a bit" and retry hack
> 
>  2. Maybe on some/most platforms we can use cgroups or whatever passes
>     for a reliable "I started a process tree starting at this PID, kill
>     -9 the whole thing please" before cleanup these days.

What really seems better:

  3. Keep that "sleep a bit" hack only on platforms that can't delete
     opened files.