Fix false ENOENT on snapshot control dentries

Lookups in the snapshot control directory for an existing snapshot
fail with ENOENT if an earlier lookup failed before the snapshot was
created.  This is because the earlier lookup causes a negative dentry
to be cached which is never invalidated.

The bug can be reproduced as follows (the second ls should succeed):

 $ ls /tank/.zfs/snapshot/s
 ls: cannot access /tank/.zfs/snapshot/s: No such file or directory
 $ zfs snap tank@s
 $ ls /tank/.zfs/snapshot/s
 ls: cannot access /tank/.zfs/snapshot/s: No such file or directory

To remedy this, always invalidate cached dentries in the snapshot
control directory.  Since these entries never exist on disk there is
no significant performance penalty for the extra lookups.

Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes openzfs#1192

config/kernel-dentry-operations.m4

@@ -0,0 +1,64 @@
+dnl #
+dnl # 3.6 API change
+dnl #
+AC_DEFUN([ZFS_AC_KERNEL_D_REVALIDATE_NAMEIDATA], [
+	AC_MSG_CHECKING([whether dops->d_revalidate() takes struct nameidata])
+	ZFS_LINUX_TRY_COMPILE([
+		#include <linux/dcache.h>
+	],[
+		int (*revalidate) (struct dentry *, struct nameidata *) = NULL;
+		struct dentry_operations dops __attribute__ ((unused)) = {
+			.d_revalidate		= revalidate,
+		};
+	],[
+		AC_MSG_RESULT(yes)
+		AC_DEFINE(HAVE_D_REVALIDATE_NAMEIDATA, 1,
+		          [dops->d_revalidate() operation takes nameidata])
+	],[
+		AC_MSG_RESULT(no)
+	])
+])
+
+dnl #
+dnl # 2.6.30 API change
+dnl # The 'struct dentry_operations' was constified in the dentry structure.
+dnl #
+AC_DEFUN([ZFS_AC_KERNEL_CONST_DENTRY_OPERATIONS], [
+	AC_MSG_CHECKING([whether dentry uses const struct dentry_operations])
+	ZFS_LINUX_TRY_COMPILE([
+		#include <linux/dcache.h>
+
+		const struct dentry_operations test_d_op = {
+			.d_revalidate = NULL,
+		};
+	],[
+		struct dentry d __attribute__ ((unused));
+
+		d.d_op = &test_d_op;
+	],[
+		AC_MSG_RESULT(yes)
+		AC_DEFINE(HAVE_CONST_DENTRY_OPERATIONS, 1,
+		          [dentry uses const struct dentry_operations])
+	],[
+		AC_MSG_RESULT(no)
+	])
+])
+
+dnl #
+dnl # 2.6.38 API change
+dnl # Added d_set_d_op() helper function.
+dnl #
+AC_DEFUN([ZFS_AC_KERNEL_D_SET_D_OP],
+	[AC_MSG_CHECKING([whether d_set_d_op() is available])
+	ZFS_LINUX_TRY_COMPILE_SYMBOL([
+		#include <linux/dcache.h>
+	], [
+		d_set_d_op(NULL, NULL);
+	], [d_set_d_op], [fs/dcache.c], [
+		AC_MSG_RESULT(yes)
+		AC_DEFINE(HAVE_D_SET_D_OP, 1,
+		          [d_set_d_op() is available])
+	], [
+		AC_MSG_RESULT(no)
+	])
+])

config/kernel.m4

@@ -63,6 +63,9 @@ AC_DEFUN([ZFS_AC_CONFIG_KERNEL], [
 	ZFS_AC_KERNEL_INSERT_INODE_LOCKED
 	ZFS_AC_KERNEL_D_MAKE_ROOT
 	ZFS_AC_KERNEL_D_OBTAIN_ALIAS
+	ZFS_AC_KERNEL_D_SET_D_OP
+	ZFS_AC_KERNEL_D_REVALIDATE_NAMEIDATA
+	ZFS_AC_KERNEL_CONST_DENTRY_OPERATIONS
 	ZFS_AC_KERNEL_CHECK_DISK_SIZE_CHANGE
 	ZFS_AC_KERNEL_TRUNCATE_SETSIZE
 	ZFS_AC_KERNEL_6ARGS_SECURITY_INODE_INIT_SECURITY

include/linux/dcache_compat.h

@@ -35,4 +35,29 @@
 #define d_make_root(inode)	d_alloc_root(inode)
 #endif /* HAVE_D_MAKE_ROOT */
 
+/*
+ * 2.6.30 API change,
+ * The const keyword was added to the 'struct dentry_operations' in
+ * the dentry structure.  To handle this we define an appropriate
+ * dentry_operations_t typedef which can be used.
+ */
+#ifdef HAVE_CONST_DENTRY_OPERATIONS
+typedef const struct dentry_operations	dentry_operations_t;
+#else
+typedef struct dentry_operations	dentry_operations_t;
+#endif
+
+/*
+ * 2.6.38 API change,
+ * Added d_set_d_op() helper function which sets some flags in
+ * dentry->d_flags based on which operations are defined.
+ */
+#ifndef HAVE_D_SET_D_OP
+static inline void
+d_set_d_op(struct dentry *dentry, dentry_operations_t *op)
+{
+	dentry->d_op = op;
+}
+#endif /* HAVE_D_SET_D_OP */
+
 #endif /* _ZFS_DCACHE_H */

module/zfs/zpl_ctldir.c

@@ -183,6 +183,66 @@ const struct inode_operations zpl_ops_root = {
 	.getattr	= zpl_root_getattr,
 };
 
+#ifdef HAVE_AUTOMOUNT
+static struct vfsmount *
+zpl_snapdir_automount(struct path *path)
+{
+	struct dentry *dentry = path->dentry;
+	int error;
+
+	/*
+	 * We must briefly disable automounts for this dentry because the
+	 * user space mount utility will trigger another lookup on this
+	 * directory.  That will result in zpl_snapdir_automount() being
+	 * called repeatedly.  The DCACHE_NEED_AUTOMOUNT flag can be
+	 * safely reset once the mount completes.
+	 */
+	dentry->d_flags &= ~DCACHE_NEED_AUTOMOUNT;
+	error = -zfsctl_mount_snapshot(path, 0);
+	dentry->d_flags |= DCACHE_NEED_AUTOMOUNT;
+	if (error)
+		return ERR_PTR(error);
+
+	/*
+	 * Rather than returning the new vfsmount for the snapshot we must
+	 * return NULL to indicate a mount collision.  This is done because
+	 * the user space mount calls do_add_mount() which adds the vfsmount
+	 * to the name space.  If we returned the new mount here it would be
+	 * added again to the vfsmount list resulting in list corruption.
+	 */
+	return (NULL);
+}
+#endif /* HAVE_AUTOMOUNT */
+
+/*
+ * Revalidate any dentry in the snapshot directory on lookup, since a snapshot
+ * having the same name have been created or destroyed since it was cached.
+ */
+static int
+#ifdef HAVE_D_REVALIDATE_NAMEIDATA
+zpl_snapdir_revalidate(struct dentry *dentry, struct nameidata *i)
+#else
+zpl_snapdir_revalidate(struct dentry *dentry, unsigned int flags)
+#endif
+{
+	return 0;
+}
+
+dentry_operations_t zpl_dops_snapdirs = {
+/*
+ * Auto mounting of snapshots is only supported for 2.6.37 and
+ * newer kernels.  Prior to this kernel the ops->follow_link()
+ * callback was used as a hack to trigger the mount.  The
+ * resulting vfsmount was then explicitly grafted in to the
+ * name space.  While it might be possible to add compatibility
+ * code to accomplish this it would require considerable care.
+ */
+#ifdef HAVE_AUTOMOUNT
+	.d_automount	= zpl_snapdir_automount,
+#endif /* HAVE_AUTOMOUNT */
+	.d_revalidate	= zpl_snapdir_revalidate,
+};
+
 static struct dentry *
 #ifdef HAVE_LOOKUP_NAMEIDATA
 zpl_snapdir_lookup(struct inode *dip, struct dentry *dentry,
@@ -194,7 +254,7 @@ zpl_snapdir_lookup(struct inode *dip, struct dentry *dentry,
 
 {
 	cred_t *cr = CRED();
-	struct inode *ip;
+	struct inode *ip = NULL;
 	int error;
 
 	crhold(cr);
@@ -203,24 +263,11 @@ zpl_snapdir_lookup(struct inode *dip, struct dentry *dentry,
 	ASSERT3S(error, <=, 0);
 	crfree(cr);
 
-	if (error) {
-		if (error == -ENOENT)
-			return d_splice_alias(NULL, dentry);
-		else
-			return ERR_PTR(error);
-	}
+	if (error && error != -ENOENT)
+		return ERR_PTR(error);
 
-	/*
-	 * Auto mounting of snapshots is only supported for 2.6.37 and
-	 * newer kernels.  Prior to this kernel the ops->follow_link()
-	 * callback was used as a hack to trigger the mount.  The
-	 * resulting vfsmount was then explicitly grafted in to the
-	 * name space.  While it might be possible to add compatibility
-	 * code to accomplish this it would require considerable care.
-	 */
-#ifdef HAVE_AUTOMOUNT
-	dentry->d_op = &zpl_dops_snapdirs;
-#endif /* HAVE_AUTOMOUNT */
+	ASSERT(error == 0 || ip == NULL);
+	d_set_d_op(dentry, &zpl_dops_snapdirs);
 
 	return d_splice_alias(ip, dentry);
 }
@@ -323,9 +370,7 @@ zpl_snapdir_mkdir(struct inode *dip, struct dentry *dentry, zpl_umode_t mode)
 
 	error = -zfsctl_snapdir_mkdir(dip, dname(dentry), vap, &ip, cr, 0);
 	if (error == 0) {
-#ifdef HAVE_AUTOMOUNT
-		dentry->d_op = &zpl_dops_snapdirs;
-#endif /* HAVE_AUTOMOUNT */
+		d_set_d_op(dentry, &zpl_dops_snapdirs);
 		d_instantiate(dentry, ip);
 	}
 
@@ -336,37 +381,6 @@ zpl_snapdir_mkdir(struct inode *dip, struct dentry *dentry, zpl_umode_t mode)
 	return (error);
 }
 
-#ifdef HAVE_AUTOMOUNT
-static struct vfsmount *
-zpl_snapdir_automount(struct path *path)
-{
-	struct dentry *dentry = path->dentry;
-	int error;
-
-	/*
-	 * We must briefly disable automounts for this dentry because the
-	 * user space mount utility will trigger another lookup on this
-	 * directory.  That will result in zpl_snapdir_automount() being
-	 * called repeatedly.  The DCACHE_NEED_AUTOMOUNT flag can be
-	 * safely reset once the mount completes.
-	 */
-	dentry->d_flags &= ~DCACHE_NEED_AUTOMOUNT;
-	error = -zfsctl_mount_snapshot(path, 0);
-	dentry->d_flags |= DCACHE_NEED_AUTOMOUNT;
-	if (error)
-		return ERR_PTR(error);
-
-	/*
-	 * Rather than returning the new vfsmount for the snapshot we must
-	 * return NULL to indicate a mount collision.  This is done because
-	 * the user space mount calls do_add_mount() which adds the vfsmount
-	 * to the name space.  If we returned the new mount here it would be
-	 * added again to the vfsmount list resulting in list corruption.
-	 */
-	return (NULL);
-}
-#endif /* HAVE_AUTOMOUNT */
-
 /*
  * Get snapshot directory attributes.
  */
@@ -413,12 +427,6 @@ const struct inode_operations zpl_ops_snapdir = {
 	.mkdir		= zpl_snapdir_mkdir,
 };
 
-#ifdef HAVE_AUTOMOUNT
-const struct dentry_operations zpl_dops_snapdirs = {
-	.d_automount	= zpl_snapdir_automount,
-};
-#endif /* HAVE_AUTOMOUNT */
-
 static struct dentry *
 #ifdef HAVE_LOOKUP_NAMEIDATA
 zpl_shares_lookup(struct inode *dip, struct dentry *dentry,