Then, I computed the time to create those 4000 empty files in the same directory, whose names were however chosen in order to hash to the same CRC32C value. This operation fails after 5 (!) seconds and creating only 61 files. In other words, this first attack allows an adversary, in a shared directory scenario, to avoid that a victim creates a file with a known-in-advance name. According to the btrfs maintainer, Chris Mason,
Collisions are a known issue with any of the hash based directories. [...] The impact of the DOS is that a malicious user is able to prevent the creation of specific file names. For this to impact other users on the system, it must be done in directories where both the malicious user and the victim user have permission to create files. The obvious example is /tmp, but there are other cases that may be more site-specific. Group writable directories have other security issues, and so we picked the hash knowing this kind of DOS was possible. It is good practice to avoid the shared directories completely if you’re worried about users doing these kinds of attacks.
A bit annoyed by this answer, I tried harder and found the following: I have created several files with random names in a directory (around 500). The time required to remove them is negligible. Then, I have created the same number of files, but giving them only 55 different crc32c values. The time required to remove them is so large that I was not able to figure it out and killed the process after 220 minutes (!). The python script I used is the following, and borrows some code from StalkR:
…continues at Pascal Junod » Like a Hot Knife Through Butter.