From: Patrick Steinhardt Date: Tue, 16 Jun 2026 05:45:43 GMT Subject: Re: [PATCH] builtin/history: unuse the commit buffer after use Message-ID: In-Reply-To: <20260615172946.GD91269@coredump.intra.peff.net> On Mon, Jun 15, 2026 at 01:29:46PM -0400, Jeff King wrote: > On Mon, Jun 15, 2026 at 11:48:10AM +0200, Patrick Steinhardt wrote: > > > On Sun, Jun 14, 2026 at 02:15:40PM +0000, Kaartic Sivaraam wrote: > > > While running `git history reword` using a Git built with `SANITIZE` flag set > > > to `address,leak`, we could observe the following leak being reported: > > > > Huh, curious. That seems to hint that we're missing test coverage for > > this specific scenario, as our test suite doesn't detect this leak. > > I think it will only leak when the commit object has an "encoding" > header. See below. > > > > As part of rewording a commit in `git history`, we get the commit message > > > buffer in the `commit_tree_ext` function. This in turn obtains the buffer > > > from `repo_logmsg_reencode`. Given how `commit_tree_ext` is invoking the > > > function with the last two parameters as NULL, we are clearly not expecting > > > a reencode to happen. In this case, the buffer that we receive from > > > `repo_logmsg_reencode` ends up always being obtained from a call to > > > `repo_get_commit_buffer`. > > > > > > This buffer is expected to be released with an accompanying call to > > > `repo_unuse_commit_buffer` which takes care of freeing it. This call > > > is missing in the `commit_tree_ext` flow thus resulting in the leak. > > > > So this doesn't really read specific at all, and I would have expected > > us to hit this leak. Puzzling. > > The first paragraph is accurate here. We'd generally just get a pointer > to the buffer cached in the slab, because no re-encoding occurs. And in > that case you _don't_ need to call unuse_commit_buffer(), because you > have a read-only copy, and the slab cache will hold it forever[1]. > Calling the unuse function will be a noop. > > But when we _do_ re-encode, then you get a new buffer which must be > freed. And that is when you have to call the unuse function. And the > reason it is "unuse" and not just "free" is that you don't necessarily > know which you have, but that function figures it out (and frees it only > if necessary). > > So what the patch is doing is correct, but the explanation is a little > confused. We see the leak only when re-encoding, so we'd probably want a > test case that triggers that. Which I assume implies rewriting a commit > that was previously generated with an encoding header. > > Now back to that [1] note. Even if we didn't re-encode, we'll still hold > onto that buffer forever. It's not a "leak" in the traditional sense > because it's still referenced in the commit slab cache. But if you are > going to walk over a million commits (like git-log does), you probably > don't want to hold a million commit messages in memory at once. > > For that you'd want to call free_commit_buffer() when you know you're > totally done with it (again, like git-log does after it finishes showing > the commit). That might be the case here in commit_tree_ext(), or it > might happen later (I'm not familiar with the git-history code). > > But note that you need to do _both_ the unuse and free calls. If we did > re-encode, the former is needed to free the newly allocated buffer. The > latter only drops the original buffer in the cache. Thanks for the explanation! Patrick