Put back the pedro diff for VOP_FSYNC of softdep when its

limit is reached.

Added twist, clamp the max_softdeps to 64 as it is still
possible to exhaust bufkva.

3 files changed, 51 insertions, 8 deletions

diff --git a/sys/ufs/ffs/ffs_inode.c b/sys/ufs/ffs/ffs_inode.c
index 3e47f0a98f3..597bd4ecca4 100644
--- a/sys/ufs/ffs/ffs_inode.c
+++ b/sys/ufs/ffs/ffs_inode.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: ffs_inode.c,v 1.59 2011/08/03 20:21:19 beck Exp $	*/
+/*	$OpenBSD: ffs_inode.c,v 1.60 2011/08/16 14:36:39 thib Exp $	*/
 /*	$NetBSD: ffs_inode.c,v 1.10 1996/05/11 18:27:19 mycroft Exp $	*/
 
 /*
@@ -164,8 +164,10 @@ ffs_truncate(struct inode *oip, off_t length, int flags, struct ucred *cred)
 	struct buf *bp;
 	int offset, size, level;
 	long count, nblocks, vflags, blocksreleased = 0;
-	int i, aflags, error, allerror;
+	int i, aflags, error, allerror, indirect = 0;
 	off_t osize;
+	extern int num_indirdep;
+	extern int max_indirdep;
 
 	if (length < 0)
 		return (EINVAL);
@@ -241,6 +243,8 @@ ffs_truncate(struct inode *oip, off_t length, int flags, struct ucred *cred)
 				   cred, aflags, &bp);
 		if (error)
 			return (error);
+		if (bp->b_lblkno >= NDADDR)
+			indirect = 1;
 		DIP_ASSIGN(oip, size, length);
 		uvm_vnp_setsize(ovp, length);
 		(void) uvm_vnp_uncache(ovp);
@@ -249,7 +253,20 @@ ffs_truncate(struct inode *oip, off_t length, int flags, struct ucred *cred)
 		else
 			bawrite(bp);
 		oip->i_flag |= IN_CHANGE | IN_UPDATE;
-		return (UFS_UPDATE(oip, MNT_WAIT));
+		error = UFS_UPDATE(oip, MNT_WAIT);
+		if (DOINGSOFTDEP(ovp) && num_indirdep > max_indirdep)
+			if (indirect) {
+				/*
+				 * If the number of pending indirect block
+				 * dependencies is sufficiently close to the
+				 * maximum number of simultaneously mappable
+				 * buffers force a sync on the vnode to prevent
+				 * buffer cache exhaustion.
+				 */
+				VOP_FSYNC(ovp, curproc->p_ucred, MNT_WAIT,
+				    curproc);
+			}
+		return (error);
 	}
 	uvm_vnp_setsize(ovp, length);
 
diff --git a/sys/ufs/ffs/ffs_softdep.c b/sys/ufs/ffs/ffs_softdep.c
index c42c64499bf..80d9da3d25f 100644
--- a/sys/ufs/ffs/ffs_softdep.c
+++ b/sys/ufs/ffs/ffs_softdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: ffs_softdep.c,v 1.108 2011/08/03 20:21:19 beck Exp $	*/
+/*	$OpenBSD: ffs_softdep.c,v 1.109 2011/08/16 14:36:39 thib Exp $	*/
 
 /*
  * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
@@ -526,6 +526,8 @@ workitem_free(struct worklist *item)
 STATIC struct workhead softdep_workitem_pending;
 STATIC struct worklist *worklist_tail;
 STATIC int num_on_worklist;	/* number of worklist items to be processed */
+STATIC int num_indirdep;	/* number of indirdep items to be processed */
+STATIC int max_indirdep;	/* maximum number of indirdep items allowed */
 STATIC int softdep_worklist_busy; /* 1 => trying to do unmount */
 STATIC int softdep_worklist_req; /* serialized waiters */
 STATIC int max_softdeps;	/* maximum number of structs before slowdown */
@@ -1140,6 +1142,8 @@ top:
 void 
 softdep_initialize(void)
 {
+	extern vsize_t bufkvm;
+	max_indirdep = (int)bufkvm / MAXPHYS * 80 / 100;
 
 	bioops.io_start = softdep_disk_io_initiation;
 	bioops.io_complete = softdep_disk_write_complete;
@@ -1155,6 +1159,7 @@ softdep_initialize(void)
 #else
 	max_softdeps = desiredvnodes * 4;
 #endif
+	max_softdeps = 64;
 	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, M_WAITOK,
 	    &pagedep_hash);
 	sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0);
@@ -1815,12 +1820,14 @@ setup_allocindir_phase2(struct buf *bp, struct inode *ip,
 			if (indirdep->ir_savebp != NULL)
 				brelse(newindirdep->ir_savebp);
 			WORKITEM_FREE(newindirdep, D_INDIRDEP);
+			num_indirdep--;
 		}
 		if (indirdep)
 			break;
 		newindirdep = pool_get(&indirdep_pool, PR_WAITOK);
 		newindirdep->ir_list.wk_type = D_INDIRDEP;
 		newindirdep->ir_state = ATTACHED;
+		num_indirdep++;
 		if (ip->i_ump->um_fstype == UM_UFS1)
 			newindirdep->ir_state |= UFS1FMT;
 		LIST_INIT(&newindirdep->ir_deplisthd);
@@ -2440,6 +2447,7 @@ indir_trunc(struct inode *ip, daddr64_t dbn, int level, daddr64_t lbn,
 		}
 		WORKLIST_REMOVE(wk);
 		WORKITEM_FREE(indirdep, D_INDIRDEP);
+		num_indirdep--;
 		if (LIST_FIRST(&bp->b_dep) != NULL) {
 			FREE_LOCK(&lk);
 			panic("indir_trunc: dangling dep");
@@ -3281,6 +3289,7 @@ softdep_disk_io_initiation(struct buf *bp)
 				wk->wk_state &= ~ONWORKLIST;
 				LIST_REMOVE(wk, wk_list);
 				WORKITEM_FREE(indirdep, D_INDIRDEP);
+				num_indirdep--;
 				FREE_LOCK(&lk);
 				brelse(sbp);
 				ACQUIRE_LOCK(&lk);
diff --git a/sys/ufs/ffs/ffs_vnops.c b/sys/ufs/ffs/ffs_vnops.c
index 94bd8788910..66a7cc0e43a 100644
--- a/sys/ufs/ffs/ffs_vnops.c
+++ b/sys/ufs/ffs/ffs_vnops.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: ffs_vnops.c,v 1.66 2011/08/03 20:21:19 beck Exp $	*/
+/*	$OpenBSD: ffs_vnops.c,v 1.67 2011/08/16 14:36:39 thib Exp $	*/
 /*	$NetBSD: ffs_vnops.c,v 1.7 1996/05/11 18:27:24 mycroft Exp $	*/
 
 /*
@@ -299,6 +299,7 @@ ffs_write(void *v)
 	daddr64_t lbn;
 	off_t osize;
 	int blkoffset, error, extended, flags, ioflag, resid, size, xfersize;
+	extern int num_indirdep, max_indirdep;
 
 	extended = 0;
 	ioflag = ap->a_ioflag;
@@ -372,7 +373,11 @@ ffs_write(void *v)
 		if (uio->uio_offset + xfersize > DIP(ip, size)) {
 			DIP_ASSIGN(ip, size, uio->uio_offset + xfersize);
 			uvm_vnp_setsize(vp, DIP(ip, size));
-			extended = 1;
+			/* Are we extending into an indirect block? */
+			if (bp->b_lblkno < NDADDR)
+				extended = 1;
+			else
+				extended = 2;
 		}
 		(void)uvm_vnp_uncache(vp);
 
@@ -416,8 +421,20 @@ ffs_write(void *v)
 			uio->uio_offset -= resid - uio->uio_resid;
 			uio->uio_resid = resid;
 		}
-	} else if (resid > uio->uio_resid && (ioflag & IO_SYNC)) {
-		error = UFS_UPDATE(ip, MNT_WAIT);
+	} else if (resid > uio->uio_resid) {
+		if (ioflag & IO_SYNC)
+			error = UFS_UPDATE(ip, MNT_WAIT);
+		if (DOINGSOFTDEP(vp) && num_indirdep > max_indirdep)
+			if (extended > 1) {
+				/*
+				 * If the number of pending indirect block
+				 * dependencies is sufficiently close to the
+				 * maximum number of simultaneously mappable
+				 * buffers force a sync on the vnode to prevent
+				 * buffer cache exhaustion.
+				 */
+				VOP_FSYNC(vp, NULL, MNT_WAIT, p);
+			}
 	}
 	return (error);
 }