The function `commit_quick_parse` provides a way to quickly parse
parts of a commit without storing or verifying most of its
metadata. The first thing it does is calculating the number of
parents by skipping "parent " lines until it finds the first
non-parent line. Afterwards, this parent count is passed to
`alloc_parents`, which will allocate an array to store all the
parent.
To calculate the amount of storage required for the parents
array, `alloc_parents` simply multiplicates the number of parents
with the respective elements's size. This already screams "buffer
overflow", and in fact this problem is getting worse by the
result being cast to an `uint32_t`.
In fact, triggering this is possible: git-hash-object(1) will
happily write a commit with multiple millions of parents for you.
I've stopped at 67,108,864 parents as git-hash-object(1)
unfortunately soaks up the complete object without streaming
anything to disk and thus will cause an OOM situation at a later
point. The point here is: this commit was about 4.1GB of size but
compressed down to 24MB and thus easy to distribute.
The above doesn't yet trigger the buffer overflow, thus. As the
array's elements are all pointers which are 8 bytes on 64 bit, we
need a total of 536,870,912 parents to trigger the overflow to
`0`. The effect is that we're now underallocating the array
and do an out-of-bound writes. As the buffer is kindly provided
by the adversary, this may easily result in code execution.
Extrapolating from the test file with 67m commits to the one with
536m commits results in a factor of 8. Thus the uncompressed
contents would be about 32GB in size and the compressed ones
192MB. While still easily distributable via the network, only
servers will have that amount of RAM and not cause an
out-of-memory condition previous to triggering the overflow. This
at least makes this attack not an easy vector for client-side use
of libgit2.