fscrypt and file truncation on cephfs

tl;dr version: in cephfs, the MDS handles truncating object data when inodes are truncated. This is problematic with fscrypt. Longer version: I've been working on a patchset to add fscrypt support to kcephfs, and have hit a problem with the way that truncation is handled. The main issue is that fscrypt uses block-based ciphers, so we must ensure that we read and write complete crypto blocks on the OSDs. I'm currently using 4k crypto blocks, but we may want to allow this to be tunable eventually (though it will need to be smaller than and align with the OSD object size). For simplicity's sake, I'm planning to disallow custom layouts on encrypted inodes. We could consider adding that later (but it doesn't sound likely to be worthwhile). Normally, when a file is truncated (usually via a SETATTR MDS call), the MDS handles truncating or deleting objects on the OSDs. This is done somewhat lazily in that the MDS replies to the client before this process is complete (AFAICT). Once we add fscrypt support, the MDS handling truncation becomes a problem, in that we need to be able to deal with complete crypto blocks. Letting the MDS truncate away part of a block will leave us with a block that can't be decrypted. There are a number of possible approaches to fixing this, but ultimately the client will have to zero-pad, encrypt and write the blocks at the edges since the MDS doesn't have access to the keys. There are several possible approaches that I've identified: 1/ We could teach the MDS the crypto blocksize, and ensure that it doesn't truncate away partial blocks. The client could tell the MDS what blocksize it's using on the inode and the MDS could ensure that truncates align to the blocks. The client will still need to write partial blocks at the edges of holes or at the EOF, and it probably shouldn't do that until it gets the unstable reply from the MDS. We could handle this by adding a new truncate op or extending the existing one. 2/ We could cede the object truncate/delete to the client altogether. The MDS is aware when an inode is encrypted so it could just not do it for those inodes. We also already handle hole punching completely on the client (though the size doesn't change there). Truncate could be a special case of that. Probably, the client would issue the truncate and then be responsible for deleting/rewriting blocks after that reply comes in. We'd have to consider how to handle delinquent clients that don't clean up correctly. 3/ We could maintain a separate field in the inode for the real inode->i_size that crypto-enabled clients would use. The client would always communicate a size to the MDS that is rounded up to the end of the last crypto block, such that the "true" size of the inode on disk would always be represented in the rstats. Only crypto-enabled clients would care about the "realsize" field. In fact, this value could _itself_ be encrypted too, so that the i_size of the file is masked from clients that don't have keys. Ceph's truncation machinery is pretty complex in general, so I could have missed other approaches or something that makes these ideas impossible. I'm leaning toward #3 here since I think it has the most benefit and keeps the MDS out of the whole business. What should we do here? -- Jeff Layton <jlayton(a)redhat.com>