Even though this looks like it, this is not a complex code change.

The interface into the "VMIO" system has changed to be more consistant
and robust.  Essentially, it is now no longer necessary to call vn_open
to get merged VM/Buffer cache operation, and exceptional conditions
such as merged operation of VBLK devices is simpler and more correct.

This code corrects a potentially large set of problems including the
problems with ktrace output and loaded systems, file create/deletes,
etc.

Most of the changes to NFS are cosmetic and name changes, eliminating
a layer of subroutine calls.  The direct calls to vput/vrele have
been re-instituted for better cross platform compatibility.

Reviewed by: davidg

4 files changed, 228 insertions, 78 deletions

diff --git a/sys/kern/vfs_bio.c b/sys/kern/vfs_bio.c
index 675ecab..b278f6c 100644
--- a/sys/kern/vfs_bio.c
+++ b/sys/kern/vfs_bio.c
@@ -18,7 +18,7 @@
  * 5. Modifications may be freely made to this file if the above conditions
  *    are met.
  *
- * $Id: vfs_bio.c,v 1.94 1996/06/30 05:17:08 davidg Exp $
+ * $Id: vfs_bio.c,v 1.95 1996/08/04 20:13:08 phk Exp $
  */
 
 /*
@@ -113,6 +113,9 @@ static struct bqueues bufqueues[BUFFER_QUEUES];
 extern int vm_swap_size;
 
 #define BUF_MAXUSE 8
+/*
+#define NO_B_MALLOC
+*/
 
 /*
  * Initialize buffer headers and related structures.
@@ -844,7 +847,7 @@ fillbuf:
 	bp->b_data = buffers_kva + (bp - buf) * MAXBSIZE;
 	bp->b_dirtyoff = bp->b_dirtyend = 0;
 	bp->b_validoff = bp->b_validend = 0;
-	bp->b_usecount = 2;
+	bp->b_usecount = 4;
 	if (bufspace >= maxbufspace + nbyteswritten) {
 		bp->b_flags |= B_INVAL;
 		brelse(bp);
@@ -1120,12 +1123,15 @@ allocbuf(struct buf * bp, int size)
 		 * Just get anonymous memory from the kernel
 		 */
 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
+#if !defined(NO_B_MALLOC)
 		if (bp->b_flags & B_MALLOC)
 			newbsize = mbsize;
 		else
+#endif
 			newbsize = round_page(size);
 
 		if (newbsize < bp->b_bufsize) {
+#if !defined(NO_B_MALLOC)
 			/*
 			 * malloced buffers are not shrunk
 			 */
@@ -1143,11 +1149,13 @@ allocbuf(struct buf * bp, int size)
 				}
 				return 1;
 			}		
+#endif
 			vm_hold_free_pages(
 			    bp,
 			    (vm_offset_t) bp->b_data + newbsize,
 			    (vm_offset_t) bp->b_data + bp->b_bufsize);
 		} else if (newbsize > bp->b_bufsize) {
+#if !defined(NO_B_MALLOC)
 			/*
 			 * We only use malloced memory on the first allocation.
 			 * and revert to page-allocated memory when the buffer grows.
@@ -1164,8 +1172,10 @@ allocbuf(struct buf * bp, int size)
 				bufmallocspace += mbsize;
 				return 1;
 			}
+#endif
 			origbuf = NULL;
 			origbufsize = 0;
+#if !defined(NO_B_MALLOC)
 			/*
 			 * If the buffer is growing on it's other-than-first allocation,
 			 * then we revert to the page-allocation scheme.
@@ -1180,14 +1190,17 @@ allocbuf(struct buf * bp, int size)
 				bp->b_flags &= ~B_MALLOC;
 				newbsize = round_page(newbsize);
 			}
+#endif
 			vm_hold_load_pages(
 			    bp,
 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
 			    (vm_offset_t) bp->b_data + newbsize);
+#if !defined(NO_B_MALLOC)
 			if (origbuf) {
 				bcopy(origbuf, bp->b_data, origbufsize);
 				free(origbuf, M_BIOBUF);
 			}
+#endif
 		}
 	} else {
 		vm_page_t m;
@@ -1196,8 +1209,10 @@ allocbuf(struct buf * bp, int size)
 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
 		desiredpages = (round_page(newbsize) >> PAGE_SHIFT);
 
+#if !defined(NO_B_MALLOC)
 		if (bp->b_flags & B_MALLOC)
 			panic("allocbuf: VMIO buffer can't be malloced");
+#endif
 
 		if (newbsize < bp->b_bufsize) {
 			if (desiredpages < bp->b_npages) {
diff --git a/sys/kern/vfs_export.c b/sys/kern/vfs_export.c
index af966dd..fbe236d 100644
--- a/sys/kern/vfs_export.c
+++ b/sys/kern/vfs_export.c
@@ -36,7 +36,7 @@
  * SUCH DAMAGE.
  *
  *	@(#)vfs_subr.c	8.13 (Berkeley) 4/18/94
- * $Id: vfs_subr.c,v 1.57 1996/07/30 18:00:25 bde Exp $
+ * $Id: vfs_subr.c,v 1.58 1996/08/15 06:45:01 dyson Exp $
  */
 
 /*
@@ -65,6 +65,8 @@
 #include <vm/vm_param.h>
 #include <vm/vm_object.h>
 #include <vm/vm_extern.h>
+#include <vm/vm_pager.h>
+#include <vm/vnode_pager.h>
 #include <sys/sysctl.h>
 
 #include <miscfs/specfs/specdev.h>
@@ -477,6 +479,8 @@ vinvalbuf(vp, flags, cred, p, slpflag, slptimeo)
 		if (vp->v_dirtyblkhd.lh_first != NULL)
 			panic("vinvalbuf: dirty bufs");
 	}
+
+	s = splbio();
 	for (;;) {
 		if ((blist = vp->v_cleanblkhd.lh_first) && (flags & V_SAVEMETA))
 			while (blist && blist->b_lblkno < 0)
@@ -492,7 +496,6 @@ vinvalbuf(vp, flags, cred, p, slpflag, slptimeo)
 			nbp = bp->b_vnbufs.le_next;
 			if ((flags & V_SAVEMETA) && bp->b_lblkno < 0)
 				continue;
-			s = splbio();
 			if (bp->b_flags & B_BUSY) {
 				bp->b_flags |= B_WANTED;
 				error = tsleep((caddr_t) bp,
@@ -505,7 +508,6 @@ vinvalbuf(vp, flags, cred, p, slpflag, slptimeo)
 			}
 			bremfree(bp);
 			bp->b_flags |= B_BUSY;
-			splx(s);
 			/*
 			 * XXX Since there are no node locks for NFS, I
 			 * believe there is a slight chance that a delayed
@@ -520,6 +522,7 @@ vinvalbuf(vp, flags, cred, p, slpflag, slptimeo)
 			brelse(bp);
 		}
 	}
+	splx(s);
 
 	s = splbio();
 	while (vp->v_numoutput > 0) {
@@ -638,7 +641,6 @@ reassignbuf(bp, newvp)
 	register struct buf *bp;
 	register struct vnode *newvp;
 {
-	register struct buflists *listheadp;
 	int s;
 
 	if (newvp == NULL) {
@@ -670,8 +672,7 @@ reassignbuf(bp, newvp)
 			LIST_INSERT_AFTER(tbp, bp, b_vnbufs);
 		}
 	} else {
-		listheadp = &newvp->v_cleanblkhd;
-		bufinsvn(bp, listheadp);
+		bufinsvn(bp, &newvp->v_cleanblkhd);
 	}
 	splx(s);
 }
@@ -745,6 +746,7 @@ loop:
 			goto loop;
 		break;
 	}
+
 	if (vp == NULL || vp->v_tag != VT_NON) {
 		MALLOC(nvp->v_specinfo, struct specinfo *,
 		    sizeof(struct specinfo), M_VNODE, M_WAITOK);
@@ -804,8 +806,18 @@ vget(vp, lockflag)
 		freevnodes--;
 	}
 	vp->v_usecount++;
+
+	/*
+	 * Create the VM object, if needed
+	 */
+	if ((vp->v_type == VREG) &&
+		((vp->v_object == NULL) ||
+			(vp->v_object->flags & OBJ_VFS_REF) == 0)) {
+		vfs_object_create(vp, curproc, curproc->p_ucred, 0);
+	}
 	if (lockflag)
 		VOP_LOCK(vp);
+
 	return (0);
 }
 
@@ -816,9 +828,21 @@ void
 vref(vp)
 	struct vnode *vp;
 {
-
 	if (vp->v_usecount <= 0)
 		panic("vref used where vget required");
+
+	if ((vp->v_type == VREG) &&
+		((vp->v_object == NULL) ||
+			((vp->v_object->flags & OBJ_VFS_REF) == 0)) ) {
+		/*
+		 * We need to lock to VP during the time that
+		 * the object is created.  This is necessary to
+		 * keep the system from re-entrantly doing it
+		 * multiple times.
+		 */
+		vfs_object_create(vp, curproc, curproc->p_ucred, 0);
+	}
+
 	vp->v_usecount++;
 }
 
@@ -829,7 +853,6 @@ void
 vput(vp)
 	register struct vnode *vp;
 {
-
 	VOP_UNLOCK(vp);
 	vrele(vp);
 }
@@ -847,10 +870,21 @@ vrele(vp)
 	if (vp == NULL)
 		panic("vrele: null vp");
 #endif
+
 	vp->v_usecount--;
+
+	if ((vp->v_usecount == 1) &&
+		vp->v_object &&
+		(vp->v_object->flags & OBJ_VFS_REF)) {
+		vp->v_object->flags &= ~OBJ_VFS_REF;
+		vm_object_deallocate(vp->v_object);
+		return;
+	}
+
 	if (vp->v_usecount > 0)
 		return;
-	if (vp->v_usecount < 0 /* || vp->v_writecount < 0 */ ) {
+
+	if (vp->v_usecount < 0) {
 #ifdef DIAGNOSTIC
 		vprint("vrele: negative ref count", vp);
 #endif
@@ -944,6 +978,11 @@ loop:
 		if ((flags & WRITECLOSE) &&
 		    (vp->v_writecount == 0 || vp->v_type != VREG))
 			continue;
+
+		if ((vp->v_usecount == 1) && vp->v_object) {
+			pager_cache(vp->v_object, FALSE);
+		}
+
 		/*
 		 * With v_usecount == 0, all we need to do is clear out the
 		 * vnode data structures and we are done.
@@ -1546,8 +1585,62 @@ loop:
 		if (VOP_ISLOCKED(vp) && (flags != MNT_WAIT))
 			continue;
 		if (vp->v_object &&
-		   (((vm_object_t) vp->v_object)->flags & OBJ_MIGHTBEDIRTY)) {
+		   (vp->v_object->flags & OBJ_MIGHTBEDIRTY)) {
 			vm_object_page_clean(vp->v_object, 0, 0, TRUE, TRUE);
 		}
 	}
 }
+
+/*
+ * Create the VM object needed for VMIO and mmap support.  This
+ * is done for all VREG files in the system.  Some filesystems might
+ * afford the additional metadata buffering capability of the
+ * VMIO code by making the device node be VMIO mode also.
+ */
+int
+vfs_object_create(vp, p, cred, waslocked)
+	struct vnode *vp;
+	struct proc *p;
+	struct ucred *cred;
+	int waslocked;
+{
+	struct vattr vat;
+	vm_object_t object;
+	int error = 0;
+
+retry:
+	if ((object = vp->v_object) == NULL) {
+		if (vp->v_type == VREG) {
+			if ((error = VOP_GETATTR(vp, &vat, cred, p)) != 0)
+				goto retn;
+			(void) vnode_pager_alloc(vp,
+				OFF_TO_IDX(round_page(vat.va_size)), 0, 0);
+		} else {
+			/*
+			 * This simply allocates the biggest object possible
+			 * for a VBLK vnode.  This should be fixed, but doesn't
+			 * cause any problems (yet).
+			 */
+			(void) vnode_pager_alloc(vp, INT_MAX, 0, 0);
+		}
+		vp->v_object->flags |= OBJ_VFS_REF;
+	} else {
+		if (object->flags & OBJ_DEAD) {
+			if (waslocked)
+				VOP_UNLOCK(vp);
+			tsleep(object, PVM, "vodead", 0);
+			if (waslocked)
+				VOP_LOCK(vp);
+			goto retry;
+		}
+		if ((object->flags & OBJ_VFS_REF) == 0) {
+			object->flags |= OBJ_VFS_REF;
+			vm_object_reference(object);
+		}
+	}
+	if (vp->v_object)
+		vp->v_flag |= VVMIO;
+
+retn:
+	return error;
+}
diff --git a/sys/kern/vfs_subr.c b/sys/kern/vfs_subr.c
index af966dd..fbe236d 100644
--- a/sys/kern/vfs_subr.c
+++ b/sys/kern/vfs_subr.c
@@ -36,7 +36,7 @@
  * SUCH DAMAGE.
  *
  *	@(#)vfs_subr.c	8.13 (Berkeley) 4/18/94
- * $Id: vfs_subr.c,v 1.57 1996/07/30 18:00:25 bde Exp $
+ * $Id: vfs_subr.c,v 1.58 1996/08/15 06:45:01 dyson Exp $
  */
 
 /*
@@ -65,6 +65,8 @@
 #include <vm/vm_param.h>
 #include <vm/vm_object.h>
 #include <vm/vm_extern.h>
+#include <vm/vm_pager.h>
+#include <vm/vnode_pager.h>
 #include <sys/sysctl.h>
 
 #include <miscfs/specfs/specdev.h>
@@ -477,6 +479,8 @@ vinvalbuf(vp, flags, cred, p, slpflag, slptimeo)
 		if (vp->v_dirtyblkhd.lh_first != NULL)
 			panic("vinvalbuf: dirty bufs");
 	}
+
+	s = splbio();
 	for (;;) {
 		if ((blist = vp->v_cleanblkhd.lh_first) && (flags & V_SAVEMETA))
 			while (blist && blist->b_lblkno < 0)
@@ -492,7 +496,6 @@ vinvalbuf(vp, flags, cred, p, slpflag, slptimeo)
 			nbp = bp->b_vnbufs.le_next;
 			if ((flags & V_SAVEMETA) && bp->b_lblkno < 0)
 				continue;
-			s = splbio();
 			if (bp->b_flags & B_BUSY) {
 				bp->b_flags |= B_WANTED;
 				error = tsleep((caddr_t) bp,
@@ -505,7 +508,6 @@ vinvalbuf(vp, flags, cred, p, slpflag, slptimeo)
 			}
 			bremfree(bp);
 			bp->b_flags |= B_BUSY;
-			splx(s);
 			/*
 			 * XXX Since there are no node locks for NFS, I
 			 * believe there is a slight chance that a delayed
@@ -520,6 +522,7 @@ vinvalbuf(vp, flags, cred, p, slpflag, slptimeo)
 			brelse(bp);
 		}
 	}
+	splx(s);
 
 	s = splbio();
 	while (vp->v_numoutput > 0) {
@@ -638,7 +641,6 @@ reassignbuf(bp, newvp)
 	register struct buf *bp;
 	register struct vnode *newvp;
 {
-	register struct buflists *listheadp;
 	int s;
 
 	if (newvp == NULL) {
@@ -670,8 +672,7 @@ reassignbuf(bp, newvp)
 			LIST_INSERT_AFTER(tbp, bp, b_vnbufs);
 		}
 	} else {
-		listheadp = &newvp->v_cleanblkhd;
-		bufinsvn(bp, listheadp);
+		bufinsvn(bp, &newvp->v_cleanblkhd);
 	}
 	splx(s);
 }
@@ -745,6 +746,7 @@ loop:
 			goto loop;
 		break;
 	}
+
 	if (vp == NULL || vp->v_tag != VT_NON) {
 		MALLOC(nvp->v_specinfo, struct specinfo *,
 		    sizeof(struct specinfo), M_VNODE, M_WAITOK);
@@ -804,8 +806,18 @@ vget(vp, lockflag)
 		freevnodes--;
 	}
 	vp->v_usecount++;
+
+	/*
+	 * Create the VM object, if needed
+	 */
+	if ((vp->v_type == VREG) &&
+		((vp->v_object == NULL) ||
+			(vp->v_object->flags & OBJ_VFS_REF) == 0)) {
+		vfs_object_create(vp, curproc, curproc->p_ucred, 0);
+	}
 	if (lockflag)
 		VOP_LOCK(vp);
+
 	return (0);
 }
 
@@ -816,9 +828,21 @@ void
 vref(vp)
 	struct vnode *vp;
 {
-
 	if (vp->v_usecount <= 0)
 		panic("vref used where vget required");
+
+	if ((vp->v_type == VREG) &&
+		((vp->v_object == NULL) ||
+			((vp->v_object->flags & OBJ_VFS_REF) == 0)) ) {
+		/*
+		 * We need to lock to VP during the time that
+		 * the object is created.  This is necessary to
+		 * keep the system from re-entrantly doing it
+		 * multiple times.
+		 */
+		vfs_object_create(vp, curproc, curproc->p_ucred, 0);
+	}
+
 	vp->v_usecount++;
 }
 
@@ -829,7 +853,6 @@ void
 vput(vp)
 	register struct vnode *vp;
 {
-
 	VOP_UNLOCK(vp);
 	vrele(vp);
 }
@@ -847,10 +870,21 @@ vrele(vp)
 	if (vp == NULL)
 		panic("vrele: null vp");
 #endif
+
 	vp->v_usecount--;
+
+	if ((vp->v_usecount == 1) &&
+		vp->v_object &&
+		(vp->v_object->flags & OBJ_VFS_REF)) {
+		vp->v_object->flags &= ~OBJ_VFS_REF;
+		vm_object_deallocate(vp->v_object);
+		return;
+	}
+
 	if (vp->v_usecount > 0)
 		return;
-	if (vp->v_usecount < 0 /* || vp->v_writecount < 0 */ ) {
+
+	if (vp->v_usecount < 0) {
 #ifdef DIAGNOSTIC
 		vprint("vrele: negative ref count", vp);
 #endif
@@ -944,6 +978,11 @@ loop:
 		if ((flags & WRITECLOSE) &&
 		    (vp->v_writecount == 0 || vp->v_type != VREG))
 			continue;
+
+		if ((vp->v_usecount == 1) && vp->v_object) {
+			pager_cache(vp->v_object, FALSE);
+		}
+
 		/*
 		 * With v_usecount == 0, all we need to do is clear out the
 		 * vnode data structures and we are done.
@@ -1546,8 +1585,62 @@ loop:
 		if (VOP_ISLOCKED(vp) && (flags != MNT_WAIT))
 			continue;
 		if (vp->v_object &&
-		   (((vm_object_t) vp->v_object)->flags & OBJ_MIGHTBEDIRTY)) {
+		   (vp->v_object->flags & OBJ_MIGHTBEDIRTY)) {
 			vm_object_page_clean(vp->v_object, 0, 0, TRUE, TRUE);
 		}
 	}
 }
+
+/*
+ * Create the VM object needed for VMIO and mmap support.  This
+ * is done for all VREG files in the system.  Some filesystems might
+ * afford the additional metadata buffering capability of the
+ * VMIO code by making the device node be VMIO mode also.
+ */
+int
+vfs_object_create(vp, p, cred, waslocked)
+	struct vnode *vp;
+	struct proc *p;
+	struct ucred *cred;
+	int waslocked;
+{
+	struct vattr vat;
+	vm_object_t object;
+	int error = 0;
+
+retry:
+	if ((object = vp->v_object) == NULL) {
+		if (vp->v_type == VREG) {
+			if ((error = VOP_GETATTR(vp, &vat, cred, p)) != 0)
+				goto retn;
+			(void) vnode_pager_alloc(vp,
+				OFF_TO_IDX(round_page(vat.va_size)), 0, 0);
+		} else {
+			/*
+			 * This simply allocates the biggest object possible
+			 * for a VBLK vnode.  This should be fixed, but doesn't
+			 * cause any problems (yet).
+			 */
+			(void) vnode_pager_alloc(vp, INT_MAX, 0, 0);
+		}
+		vp->v_object->flags |= OBJ_VFS_REF;
+	} else {
+		if (object->flags & OBJ_DEAD) {
+			if (waslocked)
+				VOP_UNLOCK(vp);
+			tsleep(object, PVM, "vodead", 0);
+			if (waslocked)
+				VOP_LOCK(vp);
+			goto retry;
+		}
+		if ((object->flags & OBJ_VFS_REF) == 0) {
+			object->flags |= OBJ_VFS_REF;
+			vm_object_reference(object);
+		}
+	}
+	if (vp->v_object)
+		vp->v_flag |= VVMIO;
+
+retn:
+	return error;
+}
diff --git a/sys/kern/vfs_vnops.c b/sys/kern/vfs_vnops.c
index 2817da1..abc85c4 100644
--- a/sys/kern/vfs_vnops.c
+++ b/sys/kern/vfs_vnops.c
@@ -36,7 +36,7 @@
  * SUCH DAMAGE.
  *
  *	@(#)vfs_vnops.c	8.2 (Berkeley) 1/21/94
- * $Id: vfs_vnops.c,v 1.24 1996/03/02 03:45:05 dyson Exp $
+ * $Id: vfs_vnops.c,v 1.25 1996/03/09 06:42:15 dyson Exp $
  */
 
 #include <sys/param.h>
@@ -162,12 +162,13 @@ vn_open(ndp, fmode, cmode)
 	if (error)
 		goto bad;
 	/*
-	 * this is here for VMIO support
+	 * Make sure that a VM object is created for VMIO support.
 	 */
 	if (vp->v_type == VREG) {
-		if ((error = vn_vmio_open(vp, p, cred)) != 0)
+		if ((error = vfs_object_create(vp, p, cred, 1)) != 0)
 			goto bad;
 	}
+
 	if (fmode & FWRITE)
 		vp->v_writecount++;
 	return (0);
@@ -211,7 +212,7 @@ vn_close(vp, flags, cred, p)
 	if (flags & FWRITE)
 		vp->v_writecount--;
 	error = VOP_CLOSE(vp, flags, cred, p);
-	vn_vmio_close(vp);
+	vrele(vp);
 	return (error);
 }
 
@@ -462,55 +463,3 @@ vn_closefile(fp, p)
 	return (vn_close(((struct vnode *)fp->f_data), fp->f_flag,
 		fp->f_cred, p));
 }
-
-int
-vn_vmio_open(vp, p, cred)
-	struct vnode *vp;
-	struct proc *p;
-	struct ucred *cred;
-{
-	struct vattr vat;
-	int error;
-	/*
-	 * this is here for VMIO support
-	 */
-	if (vp->v_type == VREG || vp->v_type == VBLK) {
-retry:
-		if ((vp->v_flag & VVMIO) == 0) {
-			if ((error = VOP_GETATTR(vp, &vat, cred, p)) != 0)
-				return error;
-			(void) vnode_pager_alloc(vp, OFF_TO_IDX(round_page(vat.va_size)), 0, 0);
-			vp->v_flag |= VVMIO;
-		} else {
-			vm_object_t object;
-			if ((object = vp->v_object) &&
-				(object->flags & OBJ_DEAD)) {
-				VOP_UNLOCK(vp);
-				tsleep(object, PVM, "vodead", 0);
-				VOP_LOCK(vp);
-				goto retry;
-			}
-			if (!object)
-				panic("vn_open: VMIO object missing");
-			vm_object_reference(object);
-		}
-	}
-	return 0;
-}
-
-void
-vn_vmio_close(vp)
-	struct vnode *vp;
-{
-	/*
-	 * this code is here for VMIO support, will eventually
-	 * be in vfs code.
-	 */
-	if (vp->v_flag & VVMIO) {
-		vrele(vp);
-		if (vp->v_object == NULL)
-			panic("vn_close: VMIO object missing");
-		vm_object_deallocate(vp->v_object);
-	} else
-		vrele(vp);
-}