suspected race between packing and fetch (single case study)

Hi,

colleagues encouraged me to report a "personal" bug I've stumbled across. Its "personal", because I wasn't able to create a minimal reproducer, or even reproduce it with the same script on other infrastructure. We're suspecting a race between packing and fetch. The script I am using is at the bottom of the email.

The script creates a joint Git/git-annex repository A with a large number of objects. Afterwards, a repository B is created, and A is cloned into it. Cloning fails initially. Errors look like this:

+ git clone --progress ../A /tmp/B/subds
Cloning into '/tmp/B/subds'...
fatal: failed to copy file to
'/tmp/B/subds/.git/objects/44/93d6041a44b5a7280875ec9b6ecd78fbab7b6e':
No such file or directory

Running "ps aux -H | grep git" before and after cloning shows garbage
collection and packing processes in repo A. We're suspecting that there
is a race. Here is script output that shows the processes:
+ cd B
+ ps aux -H
+ grep git
adina     674763  0.0  0.0   6152   836 pts/5    S+   16:38
0:00           grep git
adina     674071  0.0  0.0   9584  2788 ?        Ss   16:38   0:00
/usr/lib/git-core/git gc --auto --no-quiet
adina     674072  0.0  0.0   9584  3884 ?        S    16:38   0:00
/usr/lib/git-core/git repack -d -l --no-write-bitmap-index
adina     674073  149  0.1 583780 20564 ?        R    16:38
0:02         /usr/lib/git-core/git pack-objects --local
--delta-base-offset .git/objects/pack/.tmp-674072-pack
--keep-true-parents --honor-pack-keep --non-empty --all --reflog
--indexed-objects --unpacked  --incremental
+ git clone --progress ../A /tmp/B/subds
Cloning into '/tmp/B/subds'...
fatal: failed to copy file to
'/tmp/B/subds/.git/objects/14/5a4c6775684788ecf51e5d745ac19ad5b204e3':
No such file or directory
+ ps aux -H
+ grep git
adina     674774  0.0  0.0   6152   896 pts/5    S+   16:38
0:00           grep git
adina     674071  0.0  0.0   9584  2788 ?        Ss   16:38   0:00
/usr/lib/git-core/git gc --auto --no-quiet
adina     674072 11.0  0.0  11160  3884 ?        R    16:38   0:00
/usr/lib/git-core/git repack -d -l --no-write-bitmap-index
bash script.sh  65.71s user 29.53s system 94% cpu 1:40.71 total

Both A and B are completely sane repositories, git fsck shows nothing out of the ordinary, I can clone them fine in any situation but the scripted workflow. If I add a short "sleep" between creating A and cloning A into B the error vanishes.

I have been able to trigger this reliably for a month with the script. I am running git version 2.29.2 (but also saw this when downgrading to version 2.24) on Debian testing (bullseye). Other than simply waiting a bit before the clone, setting git config --global gc.autodetach false removes the bug, too.

I wonder if there is a way that Git could guard cases where background gc processes may still be running?

For completeness, here is the script I am using to trigger this on my machine. We didn't manage to reproduce the behavior on another machine, and I didn't find a more minimal example (sorry :( ). The script involves datalad (which uses git-annex):

#!/bin/sh

set -x

# this creates a joint git/git-annex repository
datalad create A && cd A
# this adds adds and extracts a tarball with ~13.000 JPEGs to the
repository. Data is added to git annex.
datalad download-url \
     --archive \
     --message "Download Imagenette dataset" \
     'https://s3.amazonaws.com/fast-ai-imageclas/imagenette2-160.tgz'
# this creates another joint git/git-annex repository
cd ../ && datalad create B
cd B
ps aux -H | grep git
git clone --progress ../A /tmp/B/subds
ps aux -H | grep git

Kind regards, Adina

------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------
Forschungszentrum Juelich GmbH
52425 Juelich
Sitz der Gesellschaft: Juelich
Eingetragen im Handelsregister des Amtsgerichts Dueren Nr. HR B 3498
Vorsitzender des Aufsichtsrats: MinDir Volker Rieke
Geschaeftsfuehrung: Prof. Dr.-Ing. Wolfgang Marquardt (Vorsitzender),
Karsten Beneke (stellv. Vorsitzender), Prof. Dr.-Ing. Harald Bolt
------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------

Hi Adina,

On Fri, Jan 08, 2021 at 05:39:12PM +0100, Adina Wagner wrote:

Indeed, similar races between fetching and repacking are known. For example, this discussion:

  https://lore.kernel.org/git/20200316082348.GA26581@inner.h.apk.li/

is about the .idx going away during a fetch. A similar thing is happening here, but instead of the .idx file going away, your source repository is repacking (and thus getting rid of loose object files).

Here, I think the issue is less complicated. Since you're cloning from a local repository, the 'git clone' command calls 'clone_local()', which in turn calls 'copy_or_link_directory()'. If the directory being copied changes while being iterated over, the receiving end isn't guaranteed to pick up the changes.

Worse, if the source _removes_ a file that hasn't yet been copied, over, then the copy will fail, which is what you're seeing here.

One workaround would be to clone your repositories locally with '--shared', which won't copy any objects from the source repository, but instead mark its object store as an alternate to the newly created one.

Perhaps Git could take some sort of lock when writing to the object store, but an flock wouldn't work since we'd want to allow multiple readers to acquire the lock simultaneously, so long as there is no writer.

Thanks, Taylor

Taylor Blau <me@ttaylorr.com> writes:

And the source that removes a file during a repack would create a new file to keep the contents of the removed file available (if the object still matters after the repack), but because we do not retry our "cp -r" equivalent used in the clone_local(), we may not pick such a new file up.

So, we probalby should document "git clone --local" that the user should expect fallout similar to what may happen when they copy a directory hierarchy with "cp -r src dst" and muck with what is in "src" while the copy is ongoing.

On Sat, Jan 09, 2021 at 02:11:55PM -0800, Junio C Hamano wrote:

Mm, good idea. Below the cut line is a patch to do just that. I thought briefly about documenting it in the pack-protocol page, but it only mentions the local transport in passing, so it seemed inappropriate to add that much more detail there.

--- 8< ---

Subject: [PATCH] Documentation/git-clone.txt: document race with --local

When running 'git clone --local', the operation may fail if another process is modifying the source repository. Document that this race condition is known to hopefully help anyone who may run into it.

Suggested-by: Junio C Hamano <gitster@pobox.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
---
 Documentation/git-clone.txt | 4 ++++
 1 file changed, 4 insertions(+)

On Mon, 11 Jan 2021, Taylor Blau wrote:

Couldn't `gc` be triggered by git in seemingly read-only operations, thus possibly ruining the analogy with `cp` while doing `rm` (explicit intent to modify)?

Moreover, situation is also a bit different since a sane user script would not place `rm` into background to keep operating on original source right before doing `cp` -- and that is what is happening here:

`git` operation is presumably complete (but leaves `gc` running in the background) and script advances to the next step only to run into a race condition with that preceding `git` command which apparently triggered `gc`. Should then any script which operates on local `git` repositories not to forget to add -c gc.autodetach=0 for every git invocation which might be potentially effected?

Cheers,

On Tue, Jan 12, 2021 at 12:46:22PM -0500, yoh@onerussian.com wrote:

If you're suggesting that something is missing from the above patch, I'm not sure I quite understand what you would like added.

All of these (background gc, explicit rm-ing) fall under the category of "concurrent modification": they are changing the source directory in some way while a read operation is taking place.

If your workflow is that you are frequently cloning via the local transport and there is no other synchronization going on between whatever work is happening in the source repository, then yes. (But note of course that you can set gc.autodetach=0 via the source repository's .git/config rather than typing it each time).

Thanks, Taylor

On Tue, 12 Jan 2021, Taylor Blau wrote:

Slept on it. I think your patch (doc disclaimer) is factually correct and probably as good as it can get. Not yet sure if it is worth explicit mentioning `gc` or `repack` as one of such concurrent operations.

yes. My comment was more on how such modifications are triggered: via explicit actions (e.g. `rm`) intended to modify vs as a "house keeping running in the background", which is the case of gc in particular when triggered by seemingly read-only operations.

IMHO it affects efficiency, become cumbersome (for git users), and thus might be error-prone: e.g. gc.autodetach=0 is necessity only to mitigate only for a possible subsequent `clone` invocation operating locally. Higher level constructs siting on top of `git` would not know what is the next command ran in the user script (like in our case of datalad) to set such config variable for their invocations. Adding gc.autodetach=0 to every single `git` invocation would effect our efficiency. User might not be made aware of such necessity for using `git clone` on local repositories, only after having their scripts deployed and at some random points in time start hitting the race condition and go "google" and RTFM mode to figure out what is going on.

That is why I am more in-line with your initial comment in https://lore.kernel.org/git/X%2FipCPFyW3gAWrHo@nand.local/ :

I think it would be nice to have `clone_local()` first check that there is no ongoing modifications happening before proceeding and wait some reasonable amount of time (up to ?0 sec?) if still ongoing, and then fail "informatively" if still cannot clone. Even though it would not prevent race condition in full (`clone_local` might check and initiate, and then some process starts altering while `clone_local` is ongoing), it would mitigate any scripted cases of a local `git clone` following some heavy manipulations of original repository which triggers background gc.