tree 9de02f0dea7a699bb7e6e16301ec71a84e507f4b
parent fbab552a53a2558ec84a06034e6657fb95227266
author Jeff King <peff@peff.net> 1568299485 -0400
committer Junio C Hamano <gitster@pobox.com> 1568316608 -0700

upload-pack: disable commit graph more gently for shallow traversal

When the client has asked for certain shallow options like
"deepen-since", we do a custom rev-list walk that pretends to be
shallow. Before doing so, we have to disable the commit-graph, since it
is not compatible with the shallow view of the repository. That's
handled by 829a321569 (commit-graph: close_commit_graph before shallow
walk, 2018-08-20). That commit literally closes and frees our
repo->objects->commit_graph struct.

That creates an interesting problem for commits that have _already_ been
parsed using the commit graph. Their commit->object.parsed flag is set,
their commit->graph_pos is set, but their commit->maybe_tree may still
be NULL. When somebody later calls repo_get_commit_tree(), we see that
we haven't loaded the tree oid yet and try to get it from the commit
graph. But since it has been freed, we segfault!

So the root of the issue is a data dependency between the commit's
lazy-load of the tree oid and the fact that the commit graph can go
away mid-process. How can we resolve it?

There are a couple of general approaches:

  1. The obvious answer is to avoid loading the tree from the graph when
     we see that it's NULL. But then what do we return for the tree oid?
     If we return NULL, our caller in do_traverse() will rightly
     complain that we have no tree. We'd have to fallback to loading the
     actual commit object and re-parsing it. That requires teaching
     parse_commit_buffer() to understand re-parsing (i.e., not starting
     from a clean slate and not leaking any allocated bits like parent
     list pointers).

  2. When we close the commit graph, walk through the set of in-memory
     objects and clear any graph_pos pointers. But this means we also
     have to "unparse" any such commits so that we know they still need
     to open the commit object to fill in their trees. So it's no less
     complicated than (1), and is more expensive (since we clear objects
     we might not later need).

  3. Stop freeing the commit-graph struct. Continue to let it be used
     for lazy-loads of tree oids, but let upload-pack specify that it
     shouldn't be used for further commit parsing.

  4. Push the whole shallow rev-list out to its own sub-process, with
     the commit-graph disabled from the start, giving it a clean memory
     space to work from.

I've chosen (3) here. Options (1) and (2) would work, but are
non-trivial to implement. Option (4) is more expensive, and I'm not sure
how complicated it is (shelling out for the actual rev-list part is
easy, but we do then parse the resulting commits internally, and I'm not
clear which parts need to be handling shallow-ness).

The new test in t5500 triggers this segfault, but see the comments there
for how horribly intimate it has to be with how both upload-pack and
commit graphs work.

Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>