Ceph » CephFS

Greg Farnum wrote:

Am I reading it correctly that this is just going to be doing the config and wrapper work to not call set_trace() in Server::set_trace_dist()?

I think so, although there may be more to it than that. We need things from the client to look as much like they would when requests are replayed, but without screwing up the MDS's notion of shared state. Hopefully no set_trace_dist() is enough.

Server::set_trace_dist() sets several things on the reply:
*snapbl
*head.is_dentry
*head.is_target
*trace_bl

However, none of these are accessed outside of Client::insert_trace(), and they're only looked at if trace_bl is non-empty. So from the client's perspective, just skipping set_trace_dist is fine.
Happily, it also appears that nothing else is going to do anything weird with the shared state that simply skipping set_trace_dist() will drop on the floor — it calls Locker::issue_client_lease() directly, etc.

That means that we can just have the MDS skip adding the trace on update operations (some configurable percentage of the time), and that should be pretty simple and satisfy this ticket. Hurray!

wip-4036 (commit:4ebba50a15584c89e0c5e4c6e48618055ceb96d8). Testing it now with pjd on a vstart cluster with no traceless replies, and all traceless replies (the latter crashes pretty quickly). Assuming my addition of a dout() won't break it. ;)