Bug #129

Updated by ftigeot almost 3 years ago

Hi!

This time the following changes were made:

- Namecache locks are kept by shadowinfo structures which
can be embedded into the namecache structure itself (singleton
groups) or fetched from a pool (for non-singleton groups).

- Shadow group data structure: I ended up at what has been
suggested by Matt earlier: shadow group entries form a circular list.
Double linked in order to support O(1) node deletion, enhanced
with a height counter to retain tree semantics.

- cache_setunresolved(ncp) blows up subtree over ncp to break broken
topologies.

- Deadlock avoiding techniques of the previous patch have been kept.

- The "struct namecache *nc_shadowed" field of namecache structures
still exists, but is nowhere referred by cache code. Now it's sort
of a private field, almost like the "void *" fields in vnodes,
specinfo structures, etc. It could be easily ditched and replaced by
per-mount hash for those fs-es who would use it. I just kept it as
is -- I didn't want to do anything about it without having a
consensus.

- Nullfs adjusted to this API.

Regards,
Csaba

Back