On Thu, 7 Mar 2024 at 16:10, Sweet Tea Dorminy <sweettea-kernel@xxxxxxxxxx> wrote: > > All calls into generic vfs functions need to make sure that the inode > attributes used by those functions are up to date, by calling > fuse_update_attributes() as appropriate. > > generic_write_checks() accesses inode size in order to get the > appropriate file offset for files opened with O_APPEND. Currently, in > some cases, fuse_update_attributes() is not called before > generic_write_checks(), potentially resulting in corruption/overwrite of > previously appended data if i_size is out of date in the cached inode. While this all sounds good, I don't think it makes sense. Why? Because doing cached O_APPEND writes without any sort of exclusion with remote writes is just not going to work. Either the server ignores the current size and writes at the offset that the kernel supplied (which will be the cached size of the file) and executes the write at that position, or it appends the write to the current EOF. In the former case the cache will be consistent, but append semantics are not observed, while in the latter case the append semantics are observed, but the cache will be inconsistent. Solution: either exclude remote writes or don't use the cache. Updating the file size before the write does not prevent the race, only makes the window smaller. Thanks, Miklos
