On Fri, 2015-10-16 at 09:04 -0700, Junio C Hamano wrote: > David Turner <dturner@xxxxxxxxxxxxxxxx> writes: > > > We also do dozens or hundreds of merges per day and only saw this quite > > rarely. Interestingly, we could only trigger it with an alternate > > hashing function for git's hashmap implementation, and only once a > > certain bug in that implementation was *fixed*. But that was just a > > trigger; it was certainly not the cause. The bug would, in general, > > have caused more hash collisions due to worse mixing, but I believe it > > is possible that some particular collision would have been present in > > the non-bugged version of the code but not in the bugged version. > > > > It may have also had something to do with a case-insensitive filesystem; > > we never saw anyone reproduce it on anything but OS X, and even there it > > was quite difficult to reproduce. > > > > In short, I don't think we know why the bug was not seen widely. > > It has been a long time since I looked at name-hash.c and I am fuzzy > on what the four functions (cache|index)_(file|dir)_exists are meant > for, but I have this feeling that the original premise of the patch, > "we may free a ce because we no longer use it in the index, but it > may still need to keep a reference to it in name-hash" may be wrong > in the first place. The proposed "fix" conceptually feels wrong. > > The whole point of the name-hash is so that we can detect collisions > in two names, one of which wants to have a file in one place while > the other wants to have a directory, at the same path in the index. > The pathnames and cache entries registered in the name-hash have to > be the ones that are relevant to the index in quesition. If a new > ce will be added to the index, the name-hash will have to know about > its path (and that is what CE_HASHED bit is about). On the other > hand, if you are going to remove an existing ce from the index, its > sub-paths should no longer prevent other cache entries to be there. > > E.g. if you have "a/b", it must prevent "A" from entering the index > and the name-hash helps you to do so; when you remove "a/b", then > name-hash must now allow "A" to enter the index. So "a/b" must be > removed from the name-hash by calling remove_name_hash() and normal > codepaths indeed do so. > > I do not doubt the existence of "use-after-free bug" you observed, > but I am not convinced that refcounting is "fixing" the problem; it > smells like papering over a different bug that is the root cause of > the use-after-free. > > For example, if we forget to "unhash" a ce from name-hash when we > remove a ce from the index (or after we "hashed" it, expecting to > add it to the index, but in the end decided not to add to the index, > perhaps), we would see a now-freed ce still in the name-hash. > Checking a path against the name-hash in such a state would have to > use the ce->name from that stale ce, which is a use-after-free bug. > > In such a situation, isn't the real cause of the bug the fact that > the stale ce that is no longer relevant to the true index contents > still in name-hash? The refcounting does not fix the fact that a > ce->name of a stale ce that is no longer relevant being used for D/F > collision checking. > > I am not saying that I found such a codepath that forgets to unhash, > but from the overall design and purpose of the name-hash subsystem, > anything that deliberately _allows_ a stale ce that does not exist > in the index in there smells like a workaround going in a wrong > direction. I've spent some time looking into this, but I don't quite have a repro. I do have some comments which might be interesting. The problem is not the name_hash, but with the dir_hash. The dir_hash is only even populated on case-insensitive filesystems (which is why you and Linus don't see this bug). The dir_hash is not designed to catch d/f conflicts, but rather case conflicts (one side of a merge has foo/bar; the other has FOO/bar). For each directory, it maintains a pointer to a cache_entry. Then add_to_index uses that cache_entry to rewrite incoming entries' parent directories to the expected case. Weirdly, that part of add_to_index is, so far as I can tell, never called during a merge. So where's are we using the freed value? Probably in dir_entry_cmp, while adding another entry to the hashmap; that's where our crash seems to have happened. And our failure depended on the details of the hash function as well; if a certain collision did not happen, we would not see it. There is, I think, another way to handle this: we could *copy* the path into the struct dir_entry instead of pointing to a cache_entry. But this seems like a bunch of extra work; the refcounting is simpler. -- To unsubscribe from this list: send the line "unsubscribe git" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html