On Mon, Apr 29 2019, Derrick Stolee wrote: > On 4/27/2019 9:06 AM, Ævar Arnfjörð Bjarmason wrote: >> >> There's still cases left where we'll exit early, e.g. if you do: >> >> $ git diff -U1 >> diff --git a/commit-graph.c b/commit-graph.c >> index 66865acbd7..63773764ce 100644 >> --- a/commit-graph.c >> +++ b/commit-graph.c >> @@ -1074,3 +1074,3 @@ void write_commit_graph(const char *obj_dir, >> chunk_offsets[1] = chunk_offsets[0] + GRAPH_FANOUT_SIZE; >> - chunk_offsets[2] = chunk_offsets[1] + hashsz * commits.nr; >> + chunk_offsets[2] = chunk_offsets[0] + hashsz * commits.nr; >> chunk_offsets[3] = chunk_offsets[2] + (hashsz + 16) * commits.nr; >> >> Which is obviously bad, but something I encounterd while hacking up [1] >> we'll still hard die as before this patch on: >> >> $ git status >> fatal: invalid parent position 1734910766 >> $ > > I really appreciate you digging in deep into these kinds of issues. You > seem to be hitting corrupted commit-graph files more often than we are > (in VFS for Git world). FWIW I've never encountered any of these in the wild. I just started poking at this in 2ac138d568 ("commit-graph: fix segfault on e.g. "git status"", 2019-03-25) because I was looking at the commit graph, running its tests with -d, and we'd segfault previously on e.g. "git status" in our own graph corruption tests. > However, we should be _very careful_ when turning some of these errors > to warnings. At the very least, we should do some high-level planning > for how to handle this case. Indeed. I should have been explicit, I don't think it's sane to do anything except return NULL up the stack and say "the graph is screwed, we can't use it" when initially parsing it/headers, but reading on... > The biggest issue is that we have some logic that is run after a call to > generation_numbers_enabled(), such as the `git rev-list --topo-order` > logic, that relies on the commit-graph for correctness. If we output a > warning and then stop using the commit-graph, then we will start having > commits with finite generation pointing to commits with infinite generation. > > Perhaps, with some care, we can alert the algorithm to change the "minimum > generation" that limits how far we dequeue the priority-queue. Changing it > to zero will cause the algorithm to behave like the old algorithm. > > But, having an algorithm that usually takes 0.1 seconds suddenly take 10+ > seconds also violates some expectations. > > Q: How should we handle a detectably-invalid commit-graph? ...I don't think we need to do any paranoid algorithm fallback in general. As you point out that's going to be a PITA as we read the actual data in the graph in some cases. > I think most of your patches have done a good job so far of detecting > an invalid header, and responding by ignoring the commit-graph. This case > of a detectable error in the chunk data itself is not something we can > check on the first load without serious performance issues. > > I hope we can decide on a good solution. ...OK, so this is one of those PITA cases. I think it's fine to just leave it. Although maybe we'd still want to be more paranoid with O(n) cases like "--contains" or "--topo-order" and cases that are surely just "look up my immediate commit data", like "status" dying on this particular error. But I think it's fine to just decide to nothing about this. I mainly just wanted to make a note to myself & CC the list in case there was interest...