[hbirth]

Under DLM the server owns the size, blocks, mtime and ctime of a file another node writes through its writeback cache. To answer a statx exactly, it has to revoke or downgrade that node's write grant, forcing a flush and a later re-acquire. A plain ls -l, find or monitoring agent on one node can so keep breaking the grant of a writer on another.

Add FUSE_STATX_LIGHT to fuse_statx_in.sx_flags, telling the server that the caller does not need cluster-coherent size, blocks and timestamps and that it may answer from its current values without recalling locks. It is only sent for a user stat() of a regular file without AT_STATX_FORCE_SYNC. Internal attribute updates (SEEK_END, O_APPEND, EOF checks, permission checks, setattr) keep sending full requests.

A light reply may lag behind writes, so it does not update the inode: a stale size must not shrink i_size or the page cache, and i_time, inval_mask and FUSE_I_DLM_LOCKED stay as they were. It only fills the kstat, taking the larger size and the later mtime/ctime of the reply and the inode, so this node still sees its own unflushed writes.

[hazhou-ddn]

Thanks, Horst, please hold on the merge, and let me create a page to make the whole story clear (from fuse kernel to redfsd client and server), then we can move forward.

[hbirth]

@hazhou-ddn don't worry! I'm not merging this. I am just testing this, so that we know that we have the code and you can see what I meant.