Replace v_dd vnode pointer with v_cache_dd pointer to struct namecache

in directory vnodes. Allow namecache dotdot entry to be created pointing
from child vnode to parent vnode if no existing links in opposite
direction exist. Use direct link from parent to child for dotdot lookups
otherwise.

This restores more efficient dotdot caching in NFS filesystems which
was lost when vnodes stoppped being type stable.

Reviewed by:	kib

2 files changed, 91 insertions, 33 deletions

diff --git a/sys/kern/vfs_cache.c b/sys/kern/vfs_cache.c
index 77cf2c4..6427f65 100644
--- a/sys/kern/vfs_cache.c
+++ b/sys/kern/vfs_cache.c
@@ -187,7 +187,8 @@ static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
 /*
  * Flags in namecache.nc_flag
  */
-#define NCF_WHITE	1
+#define NCF_WHITE	0x01
+#define NCF_ISDOTDOT	0x02
 
 #ifdef DIAGNOSTIC
 /*
@@ -292,14 +293,20 @@ cache_zap(ncp)
 	CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp, ncp->nc_vp);
 	vp = NULL;
 	LIST_REMOVE(ncp, nc_hash);
-	LIST_REMOVE(ncp, nc_src);
-	if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
-		vp = ncp->nc_dvp;
-		numcachehv--;
+	if (ncp->nc_flag & NCF_ISDOTDOT) {
+		if (ncp == ncp->nc_dvp->v_cache_dd)
+			ncp->nc_dvp->v_cache_dd = NULL;
+	} else {
+		LIST_REMOVE(ncp, nc_src);
+		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
+			vp = ncp->nc_dvp;
+			numcachehv--;
+		}
 	}
 	if (ncp->nc_vp) {
 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
-		ncp->nc_vp->v_dd = NULL;
+		if (ncp == ncp->nc_vp->v_cache_dd)
+			ncp->nc_vp->v_cache_dd = NULL;
 	} else {
 		TAILQ_REMOVE(&ncneg, ncp, nc_dst);
 		numneg--;
@@ -358,11 +365,18 @@ retry_wlocked:
 		}
 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
 			dotdothits++;
-			if (dvp->v_dd == NULL ||
-			    (cnp->cn_flags & MAKEENTRY) == 0) {
+			if (dvp->v_cache_dd == NULL)
+				goto unlock;
+			if ((cnp->cn_flags & MAKEENTRY) == 0) {
+				if (dvp->v_cache_dd->nc_flag & NCF_ISDOTDOT)
+					cache_zap(dvp->v_cache_dd);
+				dvp->v_cache_dd = NULL;
 				goto unlock;
 			}
-			*vpp = dvp->v_dd;
+			if (dvp->v_cache_dd->nc_flag & NCF_ISDOTDOT)
+				*vpp = dvp->v_cache_dd->nc_vp;
+			else
+				*vpp = dvp->v_cache_dd->nc_dvp;
 			CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
 			    dvp, cnp->cn_nameptr, *vpp);
 			goto success;
@@ -522,6 +536,7 @@ cache_enter(dvp, vp, cnp)
 	struct namecache *ncp, *n2;
 	struct nchashhead *ncpp;
 	u_int32_t hash;
+	int flag;
 	int hold;
 	int zap;
 	int len;
@@ -539,23 +554,33 @@ cache_enter(dvp, vp, cnp)
 	if (numcache >= desiredvnodes * 2)
 		return;
 
+	flag = 0;
 	if (cnp->cn_nameptr[0] == '.') {
-		if (cnp->cn_namelen == 1) {
+		if (cnp->cn_namelen == 1)
 			return;
-		}
-		/*
-		 * For dotdot lookups only cache the v_dd pointer if the
-		 * directory has a link back to its parent via v_cache_dst.
-		 * Without this an unlinked directory would keep a soft
-		 * reference to its parent which could not be NULLd at
-		 * cache_purge() time.
-		 */
 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
 			CACHE_WLOCK();
-			if (!TAILQ_EMPTY(&dvp->v_cache_dst))
-				dvp->v_dd = vp;
+			/*
+			 * If dotdot entry already exists, just retarget it
+			 * to new parent vnode, otherwise continue with new
+			 * namecache entry allocation.
+			 */
+			if ((ncp = dvp->v_cache_dd) != NULL) {
+				if (ncp->nc_flag & NCF_ISDOTDOT) {
+					KASSERT(ncp->nc_dvp == dvp,
+					    ("wrong isdotdot parent"));
+					TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst,
+					    ncp, nc_dst);
+					TAILQ_INSERT_HEAD(&vp->v_cache_dst,
+					    ncp, nc_dst);
+					ncp->nc_vp = vp;
+					CACHE_WUNLOCK();
+					return;
+				}
+			}
+			dvp->v_cache_dd = NULL;
 			CACHE_WUNLOCK();
-			return;
+			flag = NCF_ISDOTDOT;
 		}
 	}
 
@@ -569,6 +594,7 @@ cache_enter(dvp, vp, cnp)
 	ncp = cache_alloc(cnp->cn_namelen);
 	ncp->nc_vp = vp;
 	ncp->nc_dvp = dvp;
+	ncp->nc_flag = flag;
 	len = ncp->nc_nlen = cnp->cn_namelen;
 	hash = fnv_32_buf(cnp->cn_nameptr, len, FNV1_32_INIT);
 	bcopy(cnp->cn_nameptr, ncp->nc_name, len);
@@ -591,14 +617,34 @@ cache_enter(dvp, vp, cnp)
 		}
 	}
 
+	/*
+	 * See if we are trying to add .. entry, but some other lookup
+	 * has populated v_cache_dd pointer already.
+	 */
+	if (flag == NCF_ISDOTDOT && dvp->v_cache_dd != NULL) {
+		CACHE_WUNLOCK();
+		cache_free(ncp);
+		return;
+	}
+
 	numcache++;
 	if (!vp) {
 		numneg++;
-		ncp->nc_flag = cnp->cn_flags & ISWHITEOUT ? NCF_WHITE : 0;
+		if (cnp->cn_flags & ISWHITEOUT)
+			ncp->nc_flag |= NCF_WHITE;
 	} else if (vp->v_type == VDIR) {
-		vp->v_dd = dvp;
+		if (flag == NCF_ISDOTDOT) {
+			KASSERT(dvp->v_cache_dd == NULL,
+			    ("dangling v_cache_dd"));
+			dvp->v_cache_dd = ncp;
+		} else {
+			if ((n2 = vp->v_cache_dd) != NULL &&
+			    (n2->nc_flag & NCF_ISDOTDOT) != 0)
+				cache_zap(n2);
+			vp->v_cache_dd = ncp;
+		}
 	} else {
-		vp->v_dd = NULL;
+		vp->v_cache_dd = NULL;
 	}
 
 	/*
@@ -606,11 +652,14 @@ cache_enter(dvp, vp, cnp)
 	 * within the cache entries table.
 	 */
 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
-	if (LIST_EMPTY(&dvp->v_cache_src)) {
-		hold = 1;
-		numcachehv++;
+	if (flag != NCF_ISDOTDOT) {
+		if (LIST_EMPTY(&dvp->v_cache_src)) {
+			hold = 1;
+			numcachehv++;
+		}
+		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
 	}
-	LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
+
 	/*
 	 * If the entry is "negative", we place it into the
 	 * "negative" cache queue, otherwise, we place it into the
@@ -665,7 +714,12 @@ cache_purge(vp)
 		cache_zap(LIST_FIRST(&vp->v_cache_src));
 	while (!TAILQ_EMPTY(&vp->v_cache_dst))
 		cache_zap(TAILQ_FIRST(&vp->v_cache_dst));
-	vp->v_dd = NULL;
+	if (vp->v_cache_dd != NULL) {
+		KASSERT(vp->v_cache_dd->nc_flag & NCF_ISDOTDOT,
+		   ("lost dotdot link"));
+		cache_zap(vp->v_cache_dd);
+	}
+	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
 	CACHE_WUNLOCK();
 }
 
@@ -995,9 +1049,10 @@ vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
 			error = ENOTDIR;
 			break;
 		}
-		ncp = TAILQ_FIRST(&vp->v_cache_dst);
+ 		TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
+ 			if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
+ 				break;
 		if (ncp != NULL) {
-			MPASS(vp->v_dd == NULL || ncp->nc_dvp == vp->v_dd);
 			buflen -= ncp->nc_nlen;
 			for (i = ncp->nc_nlen - 1; i >= 0 && bp != buf; i--)
 				*--bp = ncp->nc_name[i];
@@ -1047,8 +1102,10 @@ vn_commname(struct vnode *vp, char *buf, u_int buflen)
 	int l;
 
 	CACHE_RLOCK();
-	ncp = TAILQ_FIRST(&vp->v_cache_dst);
-	if (!ncp) {
+	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
+		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
+			break;
+	if (ncp == NULL) {
 		CACHE_RUNLOCK();
 		return (ENOENT);
 	}
diff --git a/sys/kern/vfs_subr.c b/sys/kern/vfs_subr.c
index 9411e81..fcaf514 100644
--- a/sys/kern/vfs_subr.c
+++ b/sys/kern/vfs_subr.c
@@ -860,6 +860,7 @@ vdestroy(struct vnode *vp)
 	VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL"));
 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
+	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
 	VI_UNLOCK(vp);
 #ifdef MAC
 	mac_vnode_destroy(vp);