From: René Scharfe <l.s.r@web.de>
Date: Fri, 20 Mar 2026 18:54:21 GMT
Subject: Re: [PATCH v1] path-walk: fix NULL pointer dereference in error message
Message-ID: <98833ee0-4d63-4d72-9a0c-d668a421ece4@web.de>
In-Reply-To: <CALnO6CDnwYaAPhp67kaYWtV48ULjWAR6ks1khVXmSs1oWUbRDQ@mail.gmail.com>

On 3/20/26 4:16 PM, D. Ben Knoble wrote:
> 
> When we compute "child" in either preceding branch using lookup_tree
> or lookup_blob, we only return NULL if !quiet in the object_as_type
> calls (assuming we hit the "else" case there, anyway). But quiet==0 in
> both callers along this path, so !quiet will be truthy and we'll
> error() out there instead, never returning to add_tree_entries.

error() just prints a message, it doesn't end the program.

> Since I didn't quickly come up with a reproduction, I can't quite
> prove this, anyway. It's also possible my analysis is based on code
> that has since changed (I happened to have a537e3e6e9 (Merge branch
> 'sp/send-email-validate-charset' into next, 2026-03-06) checked out at
> the moment).

We could build a tree referencing an object using a mismatched
type to hit that.  It's possible by removing the type check from
builtin/mktree.c:mktree_line(), then using the resulting twisted tool:

   $ commit=$(git rev-parse HEAD)
   $ tree=$(printf "100644 blob $commit\tcommit\n" | git_evil mktree)

> Still, fixing such obviously wrong dereference is good, but I wonder
> if we should go further?
> 
> You mentioned git-backfill with a tree missing from the local odb; do
> you have a short reproduction script or test-case?

I don't know about backfill, but this would work:

  $ echo $tree | git pack-objects --path-walk --all foo

René