Weikuan Yu <wkyuwk@xxxxxxxxx> wrote: > So suppose a page is cache by one client of a cluster file system (say > GFS/Lustre), this means if the client VM caches a page in its page cache, it > must hold the correct locks for that page then. Otherwise, it would be a race > condition if another client modifies this page, right? Yes. Assuming you have such locks to play with. AFS monitors state by a combination of two things: (1) maintaining a data version number, and (2) the server holding a temporary lease on behalf of the client - which gains the client notification of change imposed by a third party. NFSv3 and before don't have such things, but NFSv4 does, I believe, as does SMB/CIFS (oplocks). I don't know about GFS and Lustre. > It occurs to me the key is the explanation of page cache being read-through > rather than be looked aside. I was thinking more from file read/write path, > and got lost when I thought a client must confirm with servers about the > state of a file, before reading from the page cache. I guess the locking of > cluster file systems would play a role here to guarantee the consistency of > cached pages then. It would, yes; but you still have to validate the cache against the network when you open a file for caching. David
