On Thu, Sep 20, 2018 at 4:22 AM, Xuehan Xu <xxhdx1985126@xxxxxxxxx> wrote: > Hi, everyone. > I've been trying to read the source code of cephfs' kernel client, and > I found that the kernel client use ceph_inode_info::i_caps to store > caps issued from different MDSes. Why? Why would a single inode have > different caps returned by different mds? It's really confusing. > Please help me, thanks:-) While MDSes *mostly* are independent with their metadata, they do need to share state on the boundaries. Eg if mds.0 is responsible for directory /foo, and mds.1 is responsible for directory /foo/bar, they both need to know some of the state about /foo/bar (since mds.0 owns its actual dentry and linkage into /foo). So a client may have capabilities from mds.1 that let it work inside of /foo/bar, but need to get another capability from mds.0 in order to move it into a different directory /baz. And other times, metadata migrates between the servers, and the client needs to be able to handle that change in ownership. -Greg
