Backout the asyncio/aiod change, as it causes buf's to get hung.

problem noticed by deraadt@

ok beck@

10 files changed, 186 insertions, 123 deletions

diff --git a/sys/conf/files b/sys/conf/files
index 7b5b5741fd6..fe983e3a3a5 100644
--- a/sys/conf/files
+++ b/sys/conf/files
@@ -1,4 +1,4 @@
-#	$OpenBSD: files,v 1.473 2009/08/29 21:12:55 kettenis Exp $
+#	$OpenBSD: files,v 1.474 2009/09/02 18:20:54 thib Exp $
 #	$NetBSD: files,v 1.87 1996/05/19 17:17:50 jonathan Exp $
 
 #	@(#)files.newconf	7.5 (Berkeley) 5/10/93
@@ -907,7 +907,7 @@ file netnatm/natm_pcb.c			natm
 file netnatm/natm_proto.c		natm
 file netnatm/natm.c			natm
 file nfs/krpc_subr.c			nfsclient
-file nfs/nfs_aiod.c			nfsclient
+#file nfs/nfs_aiod.c			nfsclient
 file nfs/nfs_bio.c			nfsclient
 file nfs/nfs_boot.c			nfsclient
 file nfs/nfs_debug.c			nfsclient & ddb
diff --git a/sys/nfs/nfs.h b/sys/nfs/nfs.h
index c3bb1ca00a2..1dcaa29cbcf 100644
--- a/sys/nfs/nfs.h
+++ b/sys/nfs/nfs.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: nfs.h,v 1.50 2009/08/25 13:41:29 thib Exp $	*/
+/*	$OpenBSD: nfs.h,v 1.51 2009/09/02 18:20:54 thib Exp $	*/
 /*	$NetBSD: nfs.h,v 1.10.4.1 1996/05/27 11:23:56 fvdl Exp $	*/
 
 /*
@@ -55,8 +55,7 @@
 #define NFS_READDIRSIZE	8192		/* Def. readdir size */
 #define	NFS_DEFRAHEAD	1		/* Def. read ahead # blocks */
 #define	NFS_MAXRAHEAD	4		/* Max. read ahead # blocks */
-#define NFS_MAXASYNCDAEMON	64	/* Max. # of aiods runnable. */
-#define NFS_DEFASYNCDAEMON	4	/* Def. # of aiods runnable. */
+#define	NFS_MAXASYNCDAEMON 	20	/* Max. number async_daemons runable */
 
 /*
  * Ideally, NFS_DIRBLKSIZ should be bigger, but I've seen servers with
@@ -190,6 +189,8 @@ struct nfsstats {
  * by them and break.
  */
 #ifdef _KERNEL
+extern int nfs_niothreads;
+
 struct uio; struct buf; struct vattr; struct nameidata;	/* XXX */
 
 #define	NFSINT_SIGMASK	(sigmask(SIGINT)|sigmask(SIGTERM)|sigmask(SIGKILL)| \
diff --git a/sys/nfs/nfs_bio.c b/sys/nfs/nfs_bio.c
index 45e827946f1..f670a997e02 100644
--- a/sys/nfs/nfs_bio.c
+++ b/sys/nfs/nfs_bio.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: nfs_bio.c,v 1.66 2009/08/27 23:39:46 thib Exp $	*/
+/*	$OpenBSD: nfs_bio.c,v 1.67 2009/09/02 18:20:54 thib Exp $	*/
 /*	$NetBSD: nfs_bio.c,v 1.25.4.2 1996/07/08 20:47:04 jtc Exp $	*/
 
 /*
@@ -57,7 +57,10 @@
 #include <nfs/nfsnode.h>
 #include <nfs/nfs_var.h>
 
+extern int nfs_numasync;
 extern struct nfsstats nfsstats;
+struct nfs_bufqhead nfs_bufq;
+uint32_t nfs_bufqmax, nfs_bufqlen;
 
 /*
  * Vnode op for read using bio
@@ -147,7 +150,7 @@ nfs_bioread(vp, uio, ioflag, cred)
 		/*
 		 * Start the read ahead(s), as required.
 		 */
-		if (nfs_numaiods > 0 && nmp->nm_readahead > 0) {
+		if (nfs_numasync > 0 && nmp->nm_readahead > 0) {
 		    for (nra = 0; nra < nmp->nm_readahead &&
 			(lbn + 1 + nra) * biosize < np->n_size; nra++) {
 			rabn = (lbn + 1 + nra) * (biosize / DEV_BSIZE);
@@ -503,75 +506,25 @@ nfs_vinvalbuf(struct vnode *vp, int flags, struct ucred *cred, struct proc *p)
  * are all hung on a dead server.
  */
 int
-nfs_asyncio(struct buf *bp)
+nfs_asyncio(bp)
+	struct buf *bp;
 {
-	struct nfs_aiod	*aiod;
-	struct nfsmount	*nmp;
-	int		 gotone, error;
-
-	aiod = NULL;
-	nmp = VFSTONFS(bp->b_vp->v_mount);
-	gotone = error = 0;
-
-	mtx_enter(&nfs_aiodl_mtx);
-	aiod = LIST_FIRST(&nfs_aiods_idle);
-	if (aiod) {
-		/*
-		 * Found an available aiod, wake it up and send
-		 * it to work on this mount.
-		 */
-		LIST_REMOVE(aiod, nad_idle);
-		mtx_leave(&nfs_aiodl_mtx);
-
-		aiod->nad_flags |= NFSAIOD_WAKEUP;
-		gotone = 1;
-		KASSERT(aiod->nad_mnt == NULL);
-		aiod->nad_mnt = nmp;
-		nmp->nm_naiods++;
-		wakeup_one(aiod);
-	} else {
-		mtx_leave(&nfs_aiodl_mtx);
-	}
-
-	/*
-	 * If no aiod's are available, check if theres already an
-	 * aiod assoicated with this mount, if so it will process
-	 * this buf.
-	 */
-	if (!gotone && nmp->nm_naiods > 0)
-		gotone = 1;
-
-	/*
-	 * If we still don't have an aiod to process this buffer,
-	 * force it sync.
-	 */
-	if (!gotone)
+	if (nfs_numasync == 0)
 		goto out;
 
+	if (nfs_bufqlen > nfs_bufqmax)
+		goto out; /* too many bufs in use, force sync */
 
-	/*
-	 * Make sure we don't queue up too much.
-	 * TODO: Look into implementing migration for aiods.
-	 */
-	if (nmp->nm_bufqlen >= nfs_aiodbufqmax) {
-		if (aiod != NULL) {
-			aiod->nad_flags &= ~NFSAIOD_WAKEUP;
-			aiod->nad_mnt = NULL;
-			mtx_enter(&nfs_aiodl_mtx);
-			LIST_INSERT_HEAD(&nfs_aiods_idle, aiod, nad_idle);
-			mtx_leave(&nfs_aiodl_mtx);
-		}
-		goto out;
+	if ((bp->b_flags & B_READ) == 0) {
+		bp->b_flags |= B_WRITEINPROG;
 	}
 
-	/* Finally, queue the buffer and return. */
+	TAILQ_INSERT_TAIL(&nfs_bufq, bp, b_freelist);
+	nfs_bufqlen++;
 
-	if ((bp->b_flags & B_READ) == 0)
-		bp->b_flags |= B_WRITEINPROG;
-
-	TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
-	nmp->nm_bufqlen++;
+	wakeup_one(&nfs_bufq);
 	return (0);
+
 out:
 	nfsstats.forcedsync++;
 	return (EIO);
diff --git a/sys/nfs/nfs_subs.c b/sys/nfs/nfs_subs.c
index a55ec8b6158..43e2b0c4e4a 100644
--- a/sys/nfs/nfs_subs.c
+++ b/sys/nfs/nfs_subs.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: nfs_subs.c,v 1.105 2009/08/25 13:41:29 thib Exp $	*/
+/*	$OpenBSD: nfs_subs.c,v 1.106 2009/09/02 18:20:54 thib Exp $	*/
 /*	$NetBSD: nfs_subs.c,v 1.27.4.3 1996/07/08 20:34:24 jtc Exp $	*/
 
 /*
@@ -925,9 +925,7 @@ nfs_vfs_init(struct vfsconf *vfsp)
 {
 	extern struct pool nfs_node_pool;
 
-	LIST_INIT(&nfs_aiods_all);
-	LIST_INIT(&nfs_aiods_idle);
-	mtx_init(&nfs_aiodl_mtx, IPL_BIO);
+	TAILQ_INIT(&nfs_bufq);
 
 	pool_init(&nfs_node_pool, sizeof(struct nfsnode), 0, 0, 0,
 	    "nfsnodepl", NULL);
diff --git a/sys/nfs/nfs_syscalls.c b/sys/nfs/nfs_syscalls.c
index 2820038399c..9e5aa218ef9 100644
--- a/sys/nfs/nfs_syscalls.c
+++ b/sys/nfs/nfs_syscalls.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: nfs_syscalls.c,v 1.86 2009/08/20 15:04:24 thib Exp $	*/
+/*	$OpenBSD: nfs_syscalls.c,v 1.87 2009/09/02 18:20:54 thib Exp $	*/
 /*	$NetBSD: nfs_syscalls.c,v 1.19 1996/02/18 11:53:52 fvdl Exp $	*/
 
 /*
@@ -73,6 +73,7 @@
 #include <nfs/nfs_var.h>
 
 /* Global defs. */
+extern int nfs_numasync;
 extern struct nfsstats nfsstats;
 struct nfssvc_sock *nfs_udpsock;
 int nfsd_waiting = 0;
@@ -117,6 +118,11 @@ struct nfsdhead nfsd_head;
 int nfssvc_sockhead_flag;
 int nfsd_head_flag;
 
+#ifdef NFSCLIENT
+struct proc *nfs_asyncdaemon[NFS_MAXASYNCDAEMON];
+int nfs_niothreads = -1;
+#endif
+
 /*
  * NFS server pseudo system call for the nfsd's
  * Based on the flag value it either:
@@ -550,7 +556,135 @@ nfsrv_init(terminating)
 	pool_init(&nfsrv_descript_pl, sizeof(struct nfsrv_descript), 0, 0, 0,
 	    "ndscpl", &pool_allocator_nointr);
 }
+#endif /* NFSSERVER */
+
+#ifdef NFSCLIENT
+/*
+ * Asynchronous I/O threads for client nfs.
+ * They do read-ahead and write-behind operations on the block I/O cache.
+ * Never returns unless it fails or gets killed.
+ */
+void
+nfssvc_iod(void *arg)
+{
+	struct proc *p = (struct proc *)arg;
+	struct buf *bp, *nbp;
+	int i, myiod;
+	struct vnode *vp;
+	int error = 0, s, bufcount;
+
+	bufcount = 256;	/* XXX: Big enough? sysctl, constant ? */
+
+	/* Assign my position or return error if too many already running. */
+	myiod = -1;
+	for (i = 0; i < NFS_MAXASYNCDAEMON; i++) {
+		if (nfs_asyncdaemon[i] == NULL) {
+			myiod = i;
+			break;
+		}
+	}
+	if (myiod == -1)
+		kthread_exit(EBUSY);
+
+	nfs_asyncdaemon[myiod] = p;
+	nfs_numasync++;
+
+	/* Upper limit on how many bufs we'll queue up for this iod. */
+	if (nfs_bufqmax > bcstats.numbufs / 4) {
+		nfs_bufqmax = bcstats.numbufs / 4; /* limit to 1/4 of bufs */
+		bufcount = 0;
+	}
+
+	nfs_bufqmax += bufcount;
+
+	/* Just loop around doin our stuff until SIGKILL. */
+	for (;;) {
+	    while (TAILQ_FIRST(&nfs_bufq) == NULL && error == 0) {
+		    error = tsleep(&nfs_bufq,
+			PWAIT | PCATCH, "nfsidl", 0);
+	    }
+	    while ((bp = TAILQ_FIRST(&nfs_bufq)) != NULL) {
+		/* Take one off the front of the list */
+		TAILQ_REMOVE(&nfs_bufq, bp, b_freelist);
+		nfs_bufqlen--;
+		if (bp->b_flags & B_READ)
+		    (void) nfs_doio(bp, NULL);
+		else do {
+		    /*
+		     * Look for a delayed write for the same vnode, so I can do 
+		     * it now. We must grab it before calling nfs_doio() to
+		     * avoid any risk of the vnode getting vclean()'d while
+		     * we are doing the write rpc.
+		     */
+		    vp = bp->b_vp;
+		    s = splbio();
+		    LIST_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) {
+			if ((nbp->b_flags &
+			    (B_BUSY|B_DELWRI|B_NEEDCOMMIT|B_NOCACHE))!=B_DELWRI)
+			    continue;
+			nbp->b_flags |= B_ASYNC;
+			bremfree(nbp);
+			buf_acquire(nbp);
+			break;
+		    }
+		    /*
+		     * For the delayed write, do the first part of nfs_bwrite()
+		     * up to, but not including nfs_strategy().
+		     */
+		    if (nbp) {
+			nbp->b_flags &= ~(B_READ|B_DONE|B_ERROR);
+			buf_undirty(nbp);
+			nbp->b_vp->v_numoutput++;
+		    }
+		    splx(s);
+
+		    (void) nfs_doio(bp, NULL);
+		} while ((bp = nbp) != NULL);
+	    }
+	    if (error) {
+		nfs_asyncdaemon[myiod] = NULL;
+		nfs_numasync--;
+		nfs_bufqmax -= bufcount;
+		kthread_exit(error);
+	    }
+	}
+}
+
+void
+nfs_getset_niothreads(set)
+	int set;
+{
+	struct proc *p;
+	int i, have, start;
+	
+	for (have = 0, i = 0; i < NFS_MAXASYNCDAEMON; i++)
+		if (nfs_asyncdaemon[i] != NULL)
+			have++;
+
+	if (set) {
+		/* clamp to sane range */
+		nfs_niothreads = max(0, min(nfs_niothreads, NFS_MAXASYNCDAEMON));
+
+		start = nfs_niothreads - have;
+
+		while (start > 0) {
+			kthread_create(nfssvc_iod, p, &p, "nfsio");
+			start--;
+		}
+
+		for (i = 0; (start < 0) && (i < NFS_MAXASYNCDAEMON); i++)
+			if (nfs_asyncdaemon[i] != NULL) {
+				psignal(nfs_asyncdaemon[i], SIGKILL);
+				start++;
+			}
+	} else {
+		if (nfs_niothreads >= 0)
+			nfs_niothreads = have;
+	}
+}
+#endif /* NFSCLIENT */
 
+#ifdef NFSSERVER
 /*
  * Find an nfssrv_sock for nfsd, sleeping if needed.
  */
diff --git a/sys/nfs/nfs_var.h b/sys/nfs/nfs_var.h
index adaf5e2772f..56fb97a615d 100644
--- a/sys/nfs/nfs_var.h
+++ b/sys/nfs/nfs_var.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: nfs_var.h,v 1.57 2009/08/20 15:04:24 thib Exp $	*/
+/*	$OpenBSD: nfs_var.h,v 1.58 2009/09/02 18:20:54 thib Exp $	*/
 /*	$NetBSD: nfs_var.h,v 1.3 1996/02/18 11:53:54 fvdl Exp $	*/
 
 /*
@@ -44,11 +44,6 @@ struct componentname;
 struct nfs_diskless;
 struct nfsm_info;
 
-/* nfs_aiod.c */
-void	nfs_aiod(void *);
-int	nfs_set_naiod(int);
-void	nfs_init_aiod(void);
-
 /* nfs_bio.c */
 int nfs_bioread(struct vnode *, struct uio *, int, struct ucred *);
 int nfs_write(void *);
@@ -261,6 +256,9 @@ int nfssvc_nfsd(struct nfsd *);
 void nfsrv_zapsock(struct nfssvc_sock *);
 void nfsrv_slpderef(struct nfssvc_sock *);
 void nfsrv_init(int);
+void nfssvc_iod(void *);
+void start_nfsio(void *);
+void nfs_getset_niothreads(int);
 
 /* nfs_kq.c */
 int  nfs_kqfilter(void *);
diff --git a/sys/nfs/nfs_vfsops.c b/sys/nfs/nfs_vfsops.c
index b872cfd2a56..5c570f1efb8 100644
--- a/sys/nfs/nfs_vfsops.c
+++ b/sys/nfs/nfs_vfsops.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: nfs_vfsops.c,v 1.90 2009/08/25 13:41:29 thib Exp $	*/
+/*	$OpenBSD: nfs_vfsops.c,v 1.91 2009/09/02 18:20:54 thib Exp $	*/
 /*	$NetBSD: nfs_vfsops.c,v 1.46.4.1 1996/05/25 22:40:35 fvdl Exp $	*/
 
 /*
@@ -611,10 +611,9 @@ nfs_mount(mp, path, data, ndp, p)
 	if ((args.flags & (NFSMNT_NFSV3|NFSMNT_RDIRPLUS)) == NFSMNT_RDIRPLUS)
 		return (EINVAL);
 
-	if (nfs_numaiods == -1) {
-		error = nfs_set_naiod(NFS_DEFASYNCDAEMON);
-		if (error)
-			return (error);
+	if (nfs_niothreads < 0) {
+		nfs_niothreads = 4;
+		nfs_getset_niothreads(1);
 	}
 
 	if (mp->mnt_flag & MNT_UPDATE) {
@@ -672,9 +671,9 @@ mountnfs(argp, mp, nam, pth, hst)
 		m_freem(nam);
 		return (0);
 	} else {
-		nmp = malloc(sizeof(*nmp), M_NFSMNT, M_WAITOK|M_ZERO);
-		mp->mnt_data = nmp;
-		TAILQ_INIT(&nmp->nm_bufq);
+		nmp = malloc(sizeof(struct nfsmount), M_NFSMNT,
+		    M_WAITOK|M_ZERO);
+		mp->mnt_data = (qaddr_t)nmp;
 	}
 
 	vfs_getnewfsid(mp);
@@ -840,7 +839,7 @@ int
 nfs_sysctl(int *name, u_int namelen, void *oldp, size_t *oldlenp, void *newp,
 	   size_t newlen, struct proc *p)
 {
-	int rv, naiods;
+	int rv;
 
 	/*
 	 * All names at this level are terminal.
@@ -872,13 +871,11 @@ nfs_sysctl(int *name, u_int namelen, void *oldp, size_t *oldlenp, void *newp,
 		return 0;
 
 	case NFS_NIOTHREADS:
-		naiods = nfs_numaiods;
-		rv = sysctl_int(oldp, oldlenp, newp, newlen, &naiods);
-		if (newp && rv == 0) {
-			if (naiods < 0)
-				return (EINVAL);
-			rv = nfs_set_naiod(naiods);
-		}
+		nfs_getset_niothreads(0);
+
+		rv = sysctl_int(oldp, oldlenp, newp, newlen, &nfs_niothreads);
+		if (newp)
+			nfs_getset_niothreads(1);
 
 		return rv;
 
diff --git a/sys/nfs/nfs_vnops.c b/sys/nfs/nfs_vnops.c
index 47eeda6a711..4e32e84259b 100644
--- a/sys/nfs/nfs_vnops.c
+++ b/sys/nfs/nfs_vnops.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: nfs_vnops.c,v 1.123 2009/08/20 15:04:24 thib Exp $	*/
+/*	$OpenBSD: nfs_vnops.c,v 1.124 2009/09/02 18:20:54 thib Exp $	*/
 /*	$NetBSD: nfs_vnops.c,v 1.62.4.1 1996/07/08 20:26:52 jtc Exp $	*/
 
 /*
@@ -184,6 +184,8 @@ extern u_int32_t nfs_true, nfs_false;
 extern u_int32_t nfs_xdrneg1;
 extern struct nfsstats nfsstats;
 extern nfstype nfsv3_type[9];
+int nfs_numasync = 0;
+
 
 void
 nfs_cache_enter(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
diff --git a/sys/nfs/nfsmount.h b/sys/nfs/nfsmount.h
index 2fc13680897..fc6fcf9bc05 100644
--- a/sys/nfs/nfsmount.h
+++ b/sys/nfs/nfsmount.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: nfsmount.h,v 1.23 2009/08/25 13:41:29 thib Exp $	*/
+/*	$OpenBSD: nfsmount.h,v 1.24 2009/09/02 18:20:54 thib Exp $	*/
 /*	$NetBSD: nfsmount.h,v 1.10 1996/02/18 11:54:03 fvdl Exp $	*/
 
 /*
@@ -71,9 +71,6 @@ struct	nfsmount {
 	int	nm_wsize;		/* Max size of write rpc */
 	int	nm_readdirsize;		/* Size of a readdir rpc */
 	int	nm_readahead;		/* Num. of blocks to readahead */
-	TAILQ_HEAD(, buf) nm_bufq;	/* async io buffer queue. */
-	int	nm_bufqlen;		/* number of buffers in nm_bufq */
-	int	nm_naiods;		/* # of aiods processing this mount */
 	u_char	nm_verf[NFSX_V3WRITEVERF]; /* V3 write verifier */
 	u_short	nm_acregmin;		/* Attr cache file recently modified */
 	u_short	nm_acregmax;		/* ac file not recently modified */
diff --git a/sys/nfs/nfsnode.h b/sys/nfs/nfsnode.h
index b0f00b77b73..85d411c9ba5 100644
--- a/sys/nfs/nfsnode.h
+++ b/sys/nfs/nfsnode.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: nfsnode.h,v 1.37 2009/08/27 23:26:56 thib Exp $	*/
+/*	$OpenBSD: nfsnode.h,v 1.38 2009/09/02 18:20:54 thib Exp $	*/
 /*	$NetBSD: nfsnode.h,v 1.16 1996/02/18 11:54:04 fvdl Exp $	*/
 
 /*
@@ -137,27 +137,10 @@ struct nfsnode {
 #define VTONFS(vp)	((struct nfsnode *)(vp)->v_data)
 #define NFSTOV(np)	((np)->n_vnode)
 
-/* nfs aiod datas. */
-struct nfs_aiod {
-	LIST_ENTRY(nfs_aiod)	 nad_all;
-	LIST_ENTRY(nfs_aiod)	 nad_idle;
-	struct nfsmount		*nad_mnt;
-	int			 nad_flags;
-};
-
-/* Flags for nad_flags. */
-#define	NFSAIOD_EXIT	0x0001	/* aiod being asked to exit. */
-#define	NFSAIOD_WAKEUP	0x0002	/* aiod being asked to wakeup. */
-/* used by nfs_set_naiod(), for convience. */
-#define	NFSAIOD_QUIT	(NFSAIOD_EXIT|NFSAIOD_WAKEUP)	/* aiod must quit. */
-
-
-LIST_HEAD(nfs_aiodhead, nfs_aiod);
-
-extern struct mutex				nfs_aiodl_mtx;
-extern struct nfs_aiodhead			nfs_aiods_all;
-extern struct nfs_aiodhead			nfs_aiods_idle;
-extern int					nfs_numaiods;
-extern int					nfs_aiodbufqmax;
+/*
+ * Queue head for nfsiod's
+ */
+extern TAILQ_HEAD(nfs_bufqhead, buf) nfs_bufq;
+extern uint32_t nfs_bufqlen, nfs_bufqmax;
 
 #endif		/* _NFS_NFSNODE_H_ */