Replace the global buffer hash table with per-vnode splay trees using a

methodology similar to the vm_map_entry splay and the VM splay that Alan
Cox is working on.  Extensive testing has appeared to have shown no
increase in overhead.

Disadvantages
    Dirties more cache lines during lookups.

    Not as fast as a hash table lookup (but still N log N and optimal
    when there is locality of reference).

Advantages
    vnode->v_dirtyblkhd is now perfectly sorted, making fsync/sync/filesystem
    syncer operate more efficiently.

    I get to rip out all the old hacks (some of which were mine) that tried
    to keep the v_dirtyblkhd tailq sorted.

    The per-vnode splay tree should be easier to lock / SMPng pushdown on
    vnodes will be easier.

    This commit along with another that Alan is working on for the VM page
    global hash table will allow me to implement ranged fsync(), optimize
    server-side nfs commit rpcs, and implement partial syncs by the
    filesystem syncer (aka filesystem syncer would detect that someone is
    trying to get the vnode lock, remembers its place, and skip to the
    next vnode).

Note that the buffer cache splay is somewhat more complex then other splays
due to special handling of background bitmap writes (multiple buffers with
the same lblkno in the same vnode), and B_INVAL discontinuities between the
old hash table and the existence of the buffer on the v_cleanblkhd list.

Suggested by: alc

1 files changed, 7 insertions, 1 deletions

diff --git a/sys/nfsserver/nfs_serv.c b/sys/nfsserver/nfs_serv.c
index 73bee42..131c0b4 100644
--- a/sys/nfsserver/nfs_serv.c
+++ b/sys/nfsserver/nfs_serv.c
@@ -3695,8 +3695,14 @@ nfsrv_commit(struct nfsrv_descript *nfsd, struct nfssvc_sock *slp,
 			 * If we have a buffer and it is marked B_DELWRI we
 			 * have to lock and write it.  Otherwise the prior
 			 * write is assumed to have already been committed.
+			 *
+			 * gbincore() can return invalid buffers now so we
+			 * have to check that bit as well (though B_DELWRI
+			 * should not be set if B_INVAL is set there could be
+			 * a race here since we haven't locked the buffer).
 			 */
-			if ((bp = gbincore(vp, lblkno)) != NULL && (bp->b_flags & B_DELWRI)) {
+			if ((bp = gbincore(vp, lblkno)) != NULL &&
+			    (bp->b_flags & (B_DELWRI|B_INVAL)) == B_DELWRI) {
 				if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
 					BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL);
 					continue; /* retry */