--- //depot/projects/smpng/sys/kern/vfs_cache.c	2008/09/24 21:56:18
+++ //depot/user/jhb/lock/kern/vfs_cache.c	2008/09/24 22:29:09
@@ -42,9 +42,9 @@
 #include <sys/lock.h>
 #include <sys/malloc.h>
 #include <sys/mount.h>
-#include <sys/mutex.h>
 #include <sys/namei.h>
 #include <sys/proc.h>
+#include <sys/rwlock.h>
 #include <sys/syscallsubr.h>
 #include <sys/sysctl.h>
 #include <sys/sysproto.h>
@@ -109,11 +109,14 @@
 #endif
 struct	nchstats nchstats;		/* cache effectiveness statistics */
 
-static struct mtx cache_lock;
-MTX_SYSINIT(vfscache, &cache_lock, "Name Cache", MTX_DEF);
+static struct rwlock cache_lock;
+RW_SYSINIT(vfscache, &cache_lock, "Name Cache");
 
-#define	CACHE_LOCK()	mtx_lock(&cache_lock)
-#define	CACHE_UNLOCK()	mtx_unlock(&cache_lock)
+#define	CACHE_UPGRADE_LOCK()	rw_try_upgrade(&cache_lock)
+#define	CACHE_RLOCK()		rw_rlock(&cache_lock)
+#define	CACHE_RUNLOCK()		rw_runlock(&cache_lock)
+#define	CACHE_WLOCK()		rw_wlock(&cache_lock)
+#define	CACHE_WUNLOCK()		rw_wunlock(&cache_lock)
 
 /*
  * UMA zones for the VFS cache.
@@ -162,6 +165,7 @@
 static u_long numposhits; STATNODE(CTLFLAG_RD, numposhits, &numposhits);
 static u_long numnegzaps; STATNODE(CTLFLAG_RD, numnegzaps, &numnegzaps);
 static u_long numneghits; STATNODE(CTLFLAG_RD, numneghits, &numneghits);
+static u_long numupgrades; STATNODE(CTLFLAG_RD, numupgrades, &numupgrades);
 
 SYSCTL_OPAQUE(_vfs_cache, OID_AUTO, nchstats, CTLFLAG_RD, &nchstats,
 	sizeof(nchstats), "LU", "VFS cache effectiveness statistics");
@@ -199,12 +203,12 @@
 
 	/* Scan hash tables for applicable entries */
 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
-		CACHE_LOCK();
+		CACHE_RLOCK();
 		count = 0;
 		LIST_FOREACH(ncp, ncpp, nc_hash) {
 			count++;
 		}
-		CACHE_UNLOCK();
+		CACHE_RUNLOCK();
 		error = SYSCTL_OUT(req, &count, sizeof(count));
 		if (error)
 			return (error);
@@ -233,11 +237,11 @@
 	/* Scan hash tables for applicable entries */
 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
 		count = 0;
-		CACHE_LOCK();
+		CACHE_RLOCK();
 		LIST_FOREACH(ncp, ncpp, nc_hash) {
 			count++;
 		}
-		CACHE_UNLOCK();
+		CACHE_RUNLOCK();
 		if (count)
 			used++;
 		if (maxlength < count)
@@ -274,7 +278,7 @@
 {
 	struct vnode *vp;
 
-	mtx_assert(&cache_lock, MA_OWNED);
+	rw_assert(&cache_lock, RA_WLOCKED);
 	CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp, ncp->nc_vp);
 	vp = NULL;
 	LIST_REMOVE(ncp, nc_hash);
@@ -321,16 +325,19 @@
 {
 	struct namecache *ncp;
 	u_int32_t hash;
-	int error, ltype;
+	int error, ltype, wlocked;
 
 	if (!doingcache) {
 		cnp->cn_flags &= ~MAKEENTRY;
 		return (0);
 	}
 retry:
-	CACHE_LOCK();
+	CACHE_RLOCK();
+	wlocked = 0;
 	numcalls++;
+	error = 0;
 
+retry_wlocked:
 	if (cnp->cn_nameptr[0] == '.') {
 		if (cnp->cn_namelen == 1) {
 			*vpp = dvp;
@@ -343,8 +350,7 @@
 			dotdothits++;
 			if (dvp->v_dd == NULL ||
 			    (cnp->cn_flags & MAKEENTRY) == 0) {
-				CACHE_UNLOCK();
-				return (0);
+				goto unlock;
 			}
 			*vpp = dvp->v_dd;
 			CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
@@ -363,23 +369,24 @@
 	}
 
 	/* We failed to find an entry */
-	if (ncp == 0) {
+	if (ncp == NULL) {
 		if ((cnp->cn_flags & MAKEENTRY) == 0) {
 			nummisszap++;
 		} else {
 			nummiss++;
 		}
 		nchstats.ncs_miss++;
-		CACHE_UNLOCK();
-		return (0);
+		goto unlock;
 	}
 
 	/* We don't want to have an entry, so dump it */
 	if ((cnp->cn_flags & MAKEENTRY) == 0) {
 		numposzaps++;
 		nchstats.ncs_badhits++;
+		if (!wlocked && !CACHE_UPGRADE_LOCK())
+			goto wlock;
 		cache_zap(ncp);
-		CACHE_UNLOCK();
+		CACHE_WUNLOCK();
 		return (0);
 	}
 
@@ -397,11 +404,15 @@
 	if (cnp->cn_nameiop == CREATE) {
 		numnegzaps++;
 		nchstats.ncs_badhits++;
+		if (!wlocked && !CACHE_UPGRADE_LOCK())
+			goto wlock;
 		cache_zap(ncp);
-		CACHE_UNLOCK();
+		CACHE_WUNLOCK();
 		return (0);
 	}
 
+	if (!wlocked && !CACHE_UPGRADE_LOCK())
+		goto wlock;
 	numneghits++;
 	/*
 	 * We found a "negative" match, so we shift it to the end of
@@ -414,9 +425,20 @@
 	nchstats.ncs_neghits++;
 	if (ncp->nc_flag & NCF_WHITE)
 		cnp->cn_flags |= ISWHITEOUT;
-	CACHE_UNLOCK();
+	CACHE_WUNLOCK();
 	return (ENOENT);
 
+wlock:
+	/*
+	 * We need to update the cache after our lookup, so upgrade to
+	 * a write lock and retry the operation.
+	 */
+	CACHE_RUNLOCK();
+	CACHE_WLOCK();
+	numupgrades++;
+	wlocked = 1;
+	goto retry_wlocked;
+
 success:
 	/*
 	 * On success we return a locked and ref'd vnode as per the lookup
@@ -424,7 +446,10 @@
 	 */
 	if (dvp == *vpp) {   /* lookup on "." */
 		VREF(*vpp);
-		CACHE_UNLOCK();
+		if (wlocked)
+			CACHE_WUNLOCK();
+		else
+			CACHE_RUNLOCK();
 		/*
 		 * When we lookup "." we still can be asked to lock it
 		 * differently...
@@ -450,7 +475,10 @@
 		VOP_UNLOCK(dvp, 0);
 	}
 	VI_LOCK(*vpp);
-	CACHE_UNLOCK();
+	if (wlocked)
+		CACHE_WUNLOCK();
+	else
+		CACHE_RUNLOCK();
 	error = vget(*vpp, cnp->cn_lkflags | LK_INTERLOCK, cnp->cn_thread);
 	if (cnp->cn_flags & ISDOTDOT)
 		vn_lock(dvp, ltype | LK_RETRY);
@@ -463,6 +491,13 @@
 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
 	}
 	return (-1);
+
+unlock:
+	if (wlocked)
+		CACHE_WUNLOCK();
+	else
+		CACHE_RUNLOCK();
+	return (0);
 }
 
 /*
@@ -500,10 +535,10 @@
 		 * cache_purge() time.
 		 */
 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
-			CACHE_LOCK();
+			CACHE_WLOCK();
 			if (!TAILQ_EMPTY(&dvp->v_cache_dst))
 				dvp->v_dd = vp;
-			CACHE_UNLOCK();
+			CACHE_WUNLOCK();
 			return;
 		}
 	}
@@ -522,7 +557,7 @@
 	hash = fnv_32_buf(cnp->cn_nameptr, len, FNV1_32_INIT);
 	bcopy(cnp->cn_nameptr, ncp->nc_name, len);
 	hash = fnv_32_buf(&dvp, sizeof(dvp), hash);
-	CACHE_LOCK();
+	CACHE_WLOCK();
 
 	/*
 	 * See if this vnode is already in the cache with this name.
@@ -536,7 +571,7 @@
 			if (n2->nc_dvp == dvp &&
 			    n2->nc_nlen == cnp->cn_namelen &&
 			    !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
-				CACHE_UNLOCK();
+				CACHE_WUNLOCK();
 				cache_free(ncp);
 				return;
 			}
@@ -582,7 +617,7 @@
 		vhold(dvp);
 	if (zap)
 		cache_zap(ncp);
-	CACHE_UNLOCK();
+	CACHE_WUNLOCK();
 }
 
 /*
@@ -613,13 +651,13 @@
 {
 
 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
-	CACHE_LOCK();
+	CACHE_WLOCK();
 	while (!LIST_EMPTY(&vp->v_cache_src))
 		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;
-	CACHE_UNLOCK();
+	CACHE_WUNLOCK();
 }
 
 /*
@@ -633,14 +671,14 @@
 	struct namecache *ncp, *nnp;
 
 	/* Scan hash tables for applicable entries */
-	CACHE_LOCK();
+	CACHE_WLOCK();
 	for (ncpp = &nchashtbl[nchash]; ncpp >= nchashtbl; ncpp--) {
 		LIST_FOREACH_SAFE(ncp, ncpp, nc_hash, nnp) {
 			if (ncp->nc_dvp->v_mount == mp)
 				cache_zap(ncp);
 		}
 	}
-	CACHE_UNLOCK();
+	CACHE_WUNLOCK();
 }
 
 /*
@@ -833,20 +871,20 @@
 	error = 0;
 	slash_prefixed = 0;
 
-	CACHE_LOCK();
+	CACHE_RLOCK();
 	numfullpathcalls++;
 	if (vp->v_type != VDIR) {
 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
 		if (!ncp) {
 			numfullpathfail2++;
-			CACHE_UNLOCK();
+			CACHE_RUNLOCK();
 			return (ENOENT);
 		}
 		for (i = ncp->nc_nlen - 1; i >= 0 && bp > buf; i--)
 			*--bp = ncp->nc_name[i];
 		if (bp == buf) {
 			numfullpathfail4++;
-			CACHE_UNLOCK();
+			CACHE_RUNLOCK();
 			return (ENOMEM);
 		}
 		*--bp = '/';
@@ -886,20 +924,20 @@
 		vp = ncp->nc_dvp;
 	}
 	if (error) {
-		CACHE_UNLOCK();
+		CACHE_RUNLOCK();
 		return (error);
 	}
 	if (!slash_prefixed) {
 		if (bp == buf) {
 			numfullpathfail4++;
-			CACHE_UNLOCK();
+			CACHE_RUNLOCK();
 			return (ENOMEM);
 		} else {
 			*--bp = '/';
 		}
 	}
 	numfullpathfound++;
-	CACHE_UNLOCK();
+	CACHE_RUNLOCK();
 
 	*retbuf = bp;
 	return (0);
@@ -911,15 +949,15 @@
 	struct namecache *ncp;
 	int l;
 
-	CACHE_LOCK();
+	CACHE_RLOCK();
 	ncp = TAILQ_FIRST(&vp->v_cache_dst);
 	if (!ncp) {
-		CACHE_UNLOCK();
+		CACHE_RUNLOCK();
 		return (ENOENT);
 	}
 	l = min(ncp->nc_nlen, buflen - 1);
 	memcpy(buf, ncp->nc_name, l);
-	CACHE_UNLOCK();
+	CACHE_RUNLOCK();
 	buf[l] = '\0';
 	return (0);
 }