On Fri, Aug 16, 2024 at 10:54:37AM +0200, Michal Hocko wrote: > Yes, I think we should kill it before it spreads even more but I would > not like to make the existing user just broken. I have zero visibility > and understanding of the bcachefs code but from a quick look at __bch2_new_inode > it shouldn't be really terribly hard to push GFP_NOWAIT flag there > directly. It would require inode_init_always_gfp variant as well (to not > touch all existing callers that do not have any locking requirements but > I do not see any other nested allocations. I'll probably have to go down into security_inode_alloc as well. That being said there is no explanation for the behavior here in the commit logs or the code itself, so who knows.