X-Git-Url: https://git.camperquake.de/gitweb.cgi?a=blobdiff_plain;f=module%2Fzfs%2Fzfs_znode.c;h=936b950f34be946c3c59a6d7aae0b3e9bc8e0c46;hb=712f8bd87b2d3799107e102652875996fa59647b;hp=0bb9c09e57ad9fa445dcb995550326783f57df64;hpb=5484965ab650a56871a62a3373859a7a36ddafcb;p=zfs.git

diff --git a/module/zfs/zfs_znode.c b/module/zfs/zfs_znode.c
index 0bb9c09..936b950 100644
--- a/module/zfs/zfs_znode.c
+++ b/module/zfs/zfs_znode.c
@@ -174,7 +174,7 @@ zfs_create_share_dir(zfs_sb_t *zsb, dmu_tx_t *tx)
 	vattr.va_uid = crgetuid(kcred);
 	vattr.va_gid = crgetgid(kcred);
 
-	sharezp = kmem_cache_alloc(znode_cache, KM_SLEEP);
+	sharezp = kmem_cache_alloc(znode_cache, KM_PUSHPAGE);
 	sharezp->z_moved = 0;
 	sharezp->z_unlinked = 0;
 	sharezp->z_atime_dirty = 0;
@@ -248,7 +248,7 @@ zfs_inode_alloc(struct super_block *sb, struct inode **ip)
 {
 	znode_t *zp;
 
-	zp = kmem_cache_alloc(znode_cache, KM_SLEEP);
+	zp = kmem_cache_alloc(znode_cache, KM_PUSHPAGE);
 	*ip = ZTOI(zp);
 
 	return (0);
@@ -322,7 +322,8 @@ zfs_inode_set_ops(zfs_sb_t *zsb, struct inode *ip)
  */
 static znode_t *
 zfs_znode_alloc(zfs_sb_t *zsb, dmu_buf_t *db, int blksz,
-    dmu_object_type_t obj_type, uint64_t obj, sa_handle_t *hdl)
+    dmu_object_type_t obj_type, uint64_t obj, sa_handle_t *hdl,
+    struct dentry *dentry)
 {
 	znode_t	*zp;
 	struct inode *ip;
@@ -347,6 +348,7 @@ zfs_znode_alloc(zfs_sb_t *zsb, dmu_buf_t *db, int blksz,
 	zp->z_blksz = blksz;
 	zp->z_seq = 0x7A4653;
 	zp->z_sync_cnt = 0;
+	zp->z_is_zvol = 0;
 
 	zfs_znode_sa_init(zsb, zp, db, obj_type, hdl);
 
@@ -371,13 +373,15 @@ zfs_znode_alloc(zfs_sb_t *zsb, dmu_buf_t *db, int blksz,
 	}
 
 	ip->i_ino = obj;
-	ip->i_mode = zp->z_mode;
-	ip->i_mtime = ip->i_atime = ip->i_ctime = CURRENT_TIME_SEC;
+	zfs_inode_update(zp);
 	zfs_inode_set_ops(zsb, ip);
 
 	if (insert_inode_locked(ip))
 		goto error;
 
+	if (dentry)
+		d_instantiate(dentry, ip);
+
 	mutex_enter(&zsb->z_znodes_lock);
 	list_insert_tail(&zsb->z_all_znodes, zp);
 	membar_producer();
@@ -674,7 +678,8 @@ zfs_mknode(znode_t *dzp, vattr_t *vap, dmu_tx_t *tx, cred_t *cr,
 	VERIFY(sa_replace_all_by_template(sa_hdl, sa_attrs, cnt, tx) == 0);
 
 	if (!(flag & IS_ROOT_NODE)) {
-		*zpp = zfs_znode_alloc(zsb, db, 0, obj_type, obj, sa_hdl);
+		*zpp = zfs_znode_alloc(zsb, db, 0, obj_type, obj, sa_hdl,
+		    vap->va_dentry);
 		ASSERT(*zpp != NULL);
 		ASSERT(dzp != NULL);
 		err = zpl_xattr_security_init(ZTOI(*zpp), ZTOI(dzp));
@@ -805,14 +810,19 @@ zfs_zget(zfs_sb_t *zsb, uint64_t obj_num, znode_t **zpp)
 	znode_t		*zp;
 	int err;
 	sa_handle_t	*hdl;
+	struct inode	*ip;
 
 	*zpp = NULL;
 
+again:
+	ip = ilookup(zsb->z_sb, obj_num);
+
 	ZFS_OBJ_HOLD_ENTER(zsb, obj_num);
 
 	err = sa_buf_hold(zsb->z_os, obj_num, NULL, &db);
 	if (err) {
 		ZFS_OBJ_HOLD_EXIT(zsb, obj_num);
+		iput(ip);
 		return (err);
 	}
 
@@ -823,13 +833,27 @@ zfs_zget(zfs_sb_t *zsb, uint64_t obj_num, znode_t **zpp)
 	    doi.doi_bonus_size < sizeof (znode_phys_t)))) {
 		sa_buf_rele(db, NULL);
 		ZFS_OBJ_HOLD_EXIT(zsb, obj_num);
+		iput(ip);
 		return (EINVAL);
 	}
 
 	hdl = dmu_buf_get_user(db);
 	if (hdl != NULL) {
-		zp  = sa_get_userdata(hdl);
+		if (ip == NULL) {
+			/*
+			 * ilookup returned NULL, which means
+			 * the znode is dying - but the SA handle isn't
+			 * quite dead yet, we need to drop any locks
+			 * we're holding, re-schedule the task and try again.
+			 */
+			sa_buf_rele(db, NULL);
+			ZFS_OBJ_HOLD_EXIT(zsb, obj_num);
+
+			schedule();
+			goto again;
+		}
 
+		zp = sa_get_userdata(hdl);
 
 		/*
 		 * Since "SA" does immediate eviction we
@@ -851,9 +875,12 @@ zfs_zget(zfs_sb_t *zsb, uint64_t obj_num, znode_t **zpp)
 		sa_buf_rele(db, NULL);
 		mutex_exit(&zp->z_lock);
 		ZFS_OBJ_HOLD_EXIT(zsb, obj_num);
+		iput(ip);
 		return (err);
 	}
 
+	ASSERT3P(ip, ==, NULL);
+
 	/*
 	 * Not found create new znode/vnode but only if file exists.
 	 *
@@ -865,7 +892,7 @@ zfs_zget(zfs_sb_t *zsb, uint64_t obj_num, znode_t **zpp)
 	 * bonus buffer.
 	 */
 	zp = zfs_znode_alloc(zsb, db, doi.doi_data_block_size,
-	    doi.doi_bonus_type, obj_num, NULL);
+	    doi.doi_bonus_type, obj_num, NULL, NULL);
 	if (zp == NULL) {
 		err = ENOENT;
 	} else {
@@ -979,13 +1006,24 @@ zfs_zinactive(znode_t *zp)
 {
 	zfs_sb_t *zsb = ZTOZSB(zp);
 	uint64_t z_id = zp->z_id;
+	boolean_t drop_mutex = 0;
 
 	ASSERT(zp->z_sa_hdl);
 
 	/*
-	 * Don't allow a zfs_zget() while were trying to release this znode
+	 * Don't allow a zfs_zget() while were trying to release this znode.
+	 *
+	 * Linux allows direct memory reclaim which means that any KM_SLEEP
+	 * allocation may trigger inode eviction.  This can lead to a deadlock
+	 * through the ->shrink_icache_memory()->evict()->zfs_inactive()->
+	 * zfs_zinactive() call path.  To avoid this deadlock the process
+	 * must not reacquire the mutex when it is already holding it.
 	 */
-	ZFS_OBJ_HOLD_ENTER(zsb, z_id);
+	if (!ZFS_OBJ_HOLD_OWNED(zsb, z_id)) {
+		ZFS_OBJ_HOLD_ENTER(zsb, z_id);
+		drop_mutex = 1;
+	}
+
 	mutex_enter(&zp->z_lock);
 
 	/*
@@ -994,14 +1032,19 @@ zfs_zinactive(znode_t *zp)
 	 */
 	if (zp->z_unlinked) {
 		mutex_exit(&zp->z_lock);
-		ZFS_OBJ_HOLD_EXIT(zsb, z_id);
+
+		if (drop_mutex)
+			ZFS_OBJ_HOLD_EXIT(zsb, z_id);
+
 		zfs_rmnode(zp);
 		return;
 	}
 
 	mutex_exit(&zp->z_lock);
 	zfs_znode_dmu_fini(zp);
-	ZFS_OBJ_HOLD_EXIT(zsb, z_id);
+
+	if (drop_mutex)
+		ZFS_OBJ_HOLD_EXIT(zsb, z_id);
 }
 
 void