- Complete part of the unfinished bufobj work by consistently using

   BO_LOCK/UNLOCK/MTX when manipulating the bufobj.
 - Create a new lock in the bufobj to lock bufobj fields independently.
   This leaves the vnode interlock as an 'identity' lock while the bufobj
   is an io lock.  The bufobj lock is ordered before the vnode interlock
   and also before the mnt ilock.
 - Exploit this new lock order to simplify softdep_check_suspend().
 - A few sync related functions are marked with a new XXX to note that
   we may not properly interlock against a non-zero bv_cnt when
   attempting to sync all vnodes on a mountlist.  I do not believe this
   race is important.  If I'm wrong this will make these locations easier
   to find.

Reviewed by:	kib (earlier diff)
Tested by:	kris, pho (earlier diff)

2 files changed, 24 insertions, 29 deletions

diff --git a/sys/nfs4client/nfs4_vfsops.c b/sys/nfs4client/nfs4_vfsops.c
index e587d53..6de47a9 100644
--- a/sys/nfs4client/nfs4_vfsops.c
+++ b/sys/nfs4client/nfs4_vfsops.c
@@ -749,6 +749,7 @@ loop:
 	MNT_VNODE_FOREACH(vp, mp, mvp) {
 		VI_LOCK(vp);
 		MNT_IUNLOCK(mp);
+		/* XXX racy bv_cnt check. */
 		if (VOP_ISLOCKED(vp) || vp->v_bufobj.bo_dirty.bv_cnt == 0 ||
 		    waitfor == MNT_LAZY) {
 			VI_UNLOCK(vp);
diff --git a/sys/nfs4client/nfs4_vnops.c b/sys/nfs4client/nfs4_vnops.c
index 7a6db66..0d6ad9a 100644
--- a/sys/nfs4client/nfs4_vnops.c
+++ b/sys/nfs4client/nfs4_vnops.c
@@ -2486,11 +2486,12 @@ nfs4_flush(struct vnode *vp, int waitfor, struct thread *td,
     int commit)
 {
 	struct nfsnode *np = VTONFS(vp);
+	struct bufobj *bo;
 	struct buf *bp;
 	int i;
 	struct buf *nbp;
 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
-	int s, error = 0, slptimeo = 0, slpflag = 0, retv, bvecpos;
+	int error = 0, slptimeo = 0, slpflag = 0, retv, bvecpos;
 	int passone = 1;
 	u_quad_t off, endoff, toff;
 	struct ucred* wcred = NULL;
@@ -2500,6 +2501,7 @@ nfs4_flush(struct vnode *vp, int waitfor, struct thread *td,
 #endif
 	struct buf *bvec_on_stack[NFS_COMMITBVECSIZ];
 	int bvecsize = 0, bveccount;
+	bo = &vp->v_bufobj;
 
 	if (nmp->nm_flag & NFSMNT_INT)
 		slpflag = PCATCH;
@@ -2517,15 +2519,14 @@ again:
 	endoff = 0;
 	bvecpos = 0;
 	if (NFS_ISV3(vp) && commit) {
-		s = splbio();
 		if (bvec != NULL && bvec != bvec_on_stack)
 			free(bvec, M_TEMP);
 		/*
 		 * Count up how many buffers waiting for a commit.
 		 */
 		bveccount = 0;
-		VI_LOCK(vp);
-		TAILQ_FOREACH_SAFE(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs, nbp) {
+		BO_LOCK(bo);
+		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
 			if (!BUF_ISLOCKED(bp) &&
 			    (bp->b_flags & (B_DELWRI | B_NEEDCOMMIT))
 				== (B_DELWRI | B_NEEDCOMMIT))
@@ -2542,11 +2543,11 @@ again:
 			 * Release the vnode interlock to avoid a lock
 			 * order reversal.
 			 */
-			VI_UNLOCK(vp);
+			BO_UNLOCK(bo);
 			bvec = (struct buf **)
 				malloc(bveccount * sizeof(struct buf *),
 				       M_TEMP, M_NOWAIT);
-			VI_LOCK(vp);
+			BO_LOCK(bo);
 			if (bvec == NULL) {
 				bvec = bvec_on_stack;
 				bvecsize = NFS_COMMITBVECSIZ;
@@ -2556,7 +2557,7 @@ again:
 			bvec = bvec_on_stack;
 			bvecsize = NFS_COMMITBVECSIZ;
 		}
-		TAILQ_FOREACH_SAFE(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs, nbp) {
+		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
 			if (bvecpos >= bvecsize)
 				break;
 			if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
@@ -2569,7 +2570,7 @@ again:
 				nbp = TAILQ_NEXT(bp, b_bobufs);
 				continue;
 			}
-			VI_UNLOCK(vp);
+			BO_UNLOCK(bo);
 			bremfree(bp);
 			/*
 			 * Work out if all buffers are using the same cred
@@ -2588,7 +2589,7 @@ again:
 				wcred = NOCRED;
 			vfs_busy_pages(bp, 1);
 
-			VI_LOCK(vp);
+			BO_LOCK(bo);
 			/*
 			 * bp is protected by being locked, but nbp is not
 			 * and vfs_busy_pages() may sleep.  We have to
@@ -2612,8 +2613,7 @@ again:
 			if (toff > endoff)
 				endoff = toff;
 		}
-		splx(s);
-		VI_UNLOCK(vp);
+		BO_UNLOCK(bo);
 	}
 	if (bvecpos > 0) {
 		/*
@@ -2665,14 +2665,12 @@ again:
 				 * specific.  We should probably move that
 				 * into bundirty(). XXX
 				 */
-				s = splbio();
-				bufobj_wref(&vp->v_bufobj);
+				bufobj_wref(bo);
 				bp->b_flags |= B_ASYNC;
 				bundirty(bp);
 				bp->b_flags &= ~B_DONE;
 				bp->b_ioflags &= ~BIO_ERROR;
 				bp->b_dirtyoff = bp->b_dirtyend = 0;
-				splx(s);
 				bufdone(bp);
 			}
 		}
@@ -2682,17 +2680,15 @@ again:
 	 * Start/do any write(s) that are required.
 	 */
 loop:
-	s = splbio();
-	VI_LOCK(vp);
-	TAILQ_FOREACH_SAFE(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs, nbp) {
+	BO_LOCK(bo);
+	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
 			if (waitfor != MNT_WAIT || passone)
 				continue;
 
 			error = BUF_TIMELOCK(bp,
 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
-			    VI_MTX(vp), "nfsfsync", slpflag, slptimeo);
-			splx(s);
+			    BO_MTX(bo), "nfsfsync", slpflag, slptimeo);
 			if (error == 0)
 				panic("nfs4_fsync: inconsistent lock");
 			if (error == ENOLCK)
@@ -2713,27 +2709,25 @@ loop:
 			BUF_UNLOCK(bp);
 			continue;
 		}
-		VI_UNLOCK(vp);
+		BO_LOCK(bo);
 		bremfree(bp);
 		if (passone || !commit)
 		    bp->b_flags |= B_ASYNC;
 		else
 		    bp->b_flags |= B_ASYNC;
-		splx(s);
 		bwrite(bp);
 		goto loop;
 	}
-	splx(s);
 	if (passone) {
 		passone = 0;
-		VI_UNLOCK(vp);
+		BO_UNLOCK(bo);
 		goto again;
 	}
 	if (waitfor == MNT_WAIT) {
-		while (vp->v_bufobj.bo_numoutput) {
-			error = bufobj_wwait(&vp->v_bufobj, slpflag, slptimeo);
+		while (bo->bo_numoutput) {
+			error = bufobj_wwait(bo, slpflag, slptimeo);
 			if (error) {
-			    VI_UNLOCK(vp);
+			    BO_UNLOCK(bo);
 			    if (nfs4_sigintr(nmp, NULL, td)) {
 				error = EINTR;
 				goto done;
@@ -2742,15 +2736,15 @@ loop:
 				slpflag = 0;
 				slptimeo = 2 * hz;
 			    }
-			    VI_LOCK(vp);
+			    BO_LOCK(bo);
 			}
 		}
 		if (vp->v_bufobj.bo_dirty.bv_cnt > 0 && commit) {
-			VI_UNLOCK(vp);
+			BO_UNLOCK(bo);
 			goto loop;
 		}
 	}
-	VI_UNLOCK(vp);
+	BO_UNLOCK(bo);
 	if (np->n_flag & NWRITEERR) {
 		error = np->n_error;
 		np->n_flag &= ~NWRITEERR;