Ceph » CephFS

I don't think we can alloc CACHE caps in the excl->mix state because caching isn't allowed in mix. Generally speaking, and A->B state shouldn't allow anything that is not allowed in B or else we can wait indefinitely; the point of the inbetween state is to wait for a revocation or lock state or something like that.

I think the problem here is that it went to MIX even though nobody wanted to write. I pushed a patch to wip-3559b that only calls scatter_mix() if somebody wants WR; if they only want RD (as is the case here) we should move to the SYNC state because that lets everyone read (and cache).

Yeah your right - the fix I proposed doesn't work because it ends up in mix state eventually. Testing your patch.

Sweet.

I'm a bit leary about exposing the caps via libcephfs.. that's an implementation detail that users shouldn't rely on. Maybe we can stick debug in the function name or something? ceph_debug_get_fd_caps()? Something along those lines?

Noah has created a libcephtools library for the admin socket stuff. Should we add it there, or a separate libcephadmin library? Alternatively, we could add a ceph_ioctl() call and bury it in there (along with the other ceph specific calls like ceph_get_local_osd()), so that we have some code in place if we want to support ioctls in fuse. Or I can just call it debug. :-)

libcephtools is wrapping the monitor interaction stuff; i think this belongs in libcephfs, we just need to advertise that it is not something users should use or rely on. putting debug in the name still seems like the simplest route to me, and is cleaner and easier than an awkward ioctl interface.

I changed those calls to ceph_debug_get_* and update the wip-3559 branch.