more cap notes

diff --git a/src/TODO b/src/TODO
index 0879bacd39d7290a9a37c7b3f012c5122f813c74..5f50c7e52ff93ab810fc730ed3d4c90eeb6b1d3a 100644 (file)
--- a/src/TODO
+++ b/src/TODO
@@ -43,16 +43,20 @@ v0.8
 
 kclient caps
 - explicit release
-- pin inode only if dirty caps
-  - release on destroy_inode
+  - this makes open on client work better, since we care less about the mds 'wanted' value.
+    (we only need to update it if/when we release the caps we still want, or async on open
+     if we don't have them yet)
 - two session lists: dirty_caps, clean_caps
 
-
 alternative #1:
-- keep CAP_PIN on _all_ inodes
+- keep CAP_PIN on _all_ inodes, and don't pin inodes with caps.
 - release on destroy_inode
+  - this will require some message preallocation trickery to avoid allocating on iput where possible.
+  - maybe maintain a pool of N release messages, each able to release M caps.. and add to the pool whenever we reach
+    N * M inodes with caps?  and periodically send the release messages early IFF we can allocate a new message in its place...
 - we only need to worry about flushing dirty caps on a timely basis
 - getattr guaranteed safe
+- higher memory utilization on the mds
 - on client reconnect we have TONS of caps to reassert... :/ what then?
   - the mds can't keep them all in the journal, so it'll need to reload some, and that's not reliable.  so we'll still end up with inodes without valid caps.
     - but we'll still get all _open file_ caps.  and everything else (except CWD) will need to be rediscovered via lookup
@@ -62,8 +66,9 @@ alternative #2:
 - make clean_caps an lru
 - touch caps on revalidate
   - that should avoid releasing caps just prior to getattr
-- maybe still pin cap inodes
-- periodic trimmer to efficiently release caps in big batches
+- maybe still pin cap inodes for now.
+  - eventually avoid it, using the release on destroy_inode strategy above
+- periodic trimmer to efficiently release the oldest clean_caps in big batches