Re: [PATCH 4/5] safe_create_leading_directories(): fix a mkdir/rmdir race

From: Michael Haggerty <mhagger@alum.mit.edu>
Date: Jan 2, 2014, 00:53 UTC
Message-ID: <52C4B89D.5020902@alum.mit.edu>
In-Reply-To: <20131226230204.GZ20443@google.com>

On 12/27/2013 12:02 AM, Jonathan Nieder wrote:

Show 12 quoted lines

> Michael Haggerty wrote:
> 
>> It could be that some other process is trying to clean up empty
>> directories at the same time that safe_create_leading_directories() is
>> attempting to create them.  In this case, it could happen that
>> directory "a/b" was present at the end of one iteration of the loop
>> (either it was already present or we just created it ourselves), but
>> by the time we try to create directory "a/b/c", directory "a/b" has
>> been deleted.  In fact, directory "a" might also have been deleted.
> 
> When does this happen in practice?  Is this about git racing with
> itself or with some other program?

I think it could be triggered by a reference creation racing with a reference packing. See below.

> I fear that the aggressive directory creator fighting the aggressive
> directory remover might be waging a losing battle.

That may be so, but it would be strange for a directory remover to be aggressive. And even if it were, the worst consequence would be that the director creator would try three times before giving up.

> Is this about a push that creates a ref racing against a push that
> deletes a ref from the same hierarchy?

The race could be triggered in this scenario but I don't think it would result in a spurious error (at least not if there are only two racers...) The reason is that empty loose reference directories are not deleted when a reference is *deleted*, but rather when a new d/f-conflicting reference is *created*. E.g., if

    git branch foo/bar
    git branch -d foo/bar # this leaves directory foo behind
    # this removes directory foo and creates file foo:
    git branch foo &
        git branch foo/baz

The last two commands could fight. However, in this scenario one of the reference creations would ultimately have to fail anyway, so this patch doesn't really help.

However, when packing references, the directories that used to hold the old references are deleted right away. So

    git branch foo/bar
    git pack-refs --all &
        git branch foo/baz

Here, the last two commands could fight.

Show 16 quoted lines

>> So, if a call to mkdir() fails with ENOENT, then try checking/making
>> all directories again from the beginning.  Attempt up to three times
>> before giving up.
> 
> If we are racing against a ref deletion, then we can retry while our
> rival keeps walking up the directory tree and deleting parent
> directories.  As soon as we successfully create a directory, we have
> won the race.
> 
> But what happens if the entire safe_create_leading_directories
> operation succeeds and *then* our racing partner deletes the
> directory?  No one is putting in a file to reserve the directory for
> the directory creator.
>
> If we care enough to retry more than once, I fear this is retrying at
> the wrong level.

I realize that this change doesn't solve the whole problem. But you make a good point, that if the caller is going to retry anyway, then there is no need to retry within this function. It would be sufficient for this function to return a specific error value indicating that "creating the directory failed, but there's a chance of success if you try again".

On the other hand, your argument assumes that all callers really *do* retry, which isn't the case, and I doubt that all callers are going to be fixed. So there might be some value in retrying within this function anyway (it's a game of averages we're playing here anyway).

I'll think some more about it.

> Tests?

I can't think of how to test this short of either instrumenting the code (which I did for my own tests, but didn't include the test code in this submission) or running the test within some kind of malicious virtual filesystem. Ideas?

Michael

-- 
Michael Haggerty
mhagger@alum.mit.edu
http://softwareswirl.blogspot.com/