arbiter: return ENOSYS for 'non readable' FOPs

AFR marks the arbiter as 'non readable'. This has been introduced with
commit 8ab87137 (afr: do not set arbiter as a readable subvol in inode
context). arbiter_readv() should not get called anymore, so it could be
removed. However, it is a good defensive approach to have all the inode
read FOPs that can not be handled by the arbiter to return ENOSYS.

> Reviewed-on: https://review.gluster.org/18103
> Smoke: Gluster Build System <jenkins@build.gluster.org>
> CentOS-regression: Gluster Build System <jenkins@build.gluster.org>
(cherry picked from commit b1352d0974328b367afa7360e9523585efb7178d)

Change-Id: I6ea41680832859bd6790dc8d7440ee98d38205fc
BUG: 1489511
Signed-off-by: Niels de Vos <ndevos@redhat.com>
Signed-off-by: Ravishankar N <ravishankar@redhat.com>
Reviewed-on: https://review.gluster.org/18227
Smoke: Gluster Build System <jenkins@build.gluster.org>
CentOS-regression: Gluster Build System <jenkins@build.gluster.org>
Reviewed-by: Pranith Kumar Karampuri <pkarampu@redhat.com>

1 files changed, 35 insertions, 10 deletions

diff --git a/xlators/features/arbiter/src/arbiter.c b/xlators/features/arbiter/src/arbiter.c
index 0e555c84274..cf0841653b7 100644
--- a/xlators/features/arbiter/src/arbiter.c
+++ b/xlators/features/arbiter/src/arbiter.c
@@ -89,15 +89,6 @@ arbiter_lookup (call_frame_t *frame, xlator_t *this, loc_t *loc, dict_t *xdata)
 }
 
 int32_t
-arbiter_readv (call_frame_t *frame,  xlator_t *this, fd_t *fd, size_t size,
-               off_t offset, uint32_t flags, dict_t *xdata)
-{
-        STACK_UNWIND_STRICT (readv, frame, -1, ENOTCONN, NULL, 0, NULL, NULL,
-                             NULL);
-        return 0;
-}
-
-int32_t
 arbiter_truncate (call_frame_t *frame, xlator_t *this, loc_t *loc, off_t offset,
                   dict_t *xdata)
 {
@@ -270,6 +261,23 @@ unwind:
         return 0;
 }
 
+static int32_t
+arbiter_readv (call_frame_t *frame,  xlator_t *this, fd_t *fd, size_t size,
+               off_t offset, uint32_t flags, dict_t *xdata)
+{
+        STACK_UNWIND_STRICT (readv, frame, -1, ENOSYS, NULL, 0, NULL, NULL,
+                             NULL);
+        return 0;
+}
+
+static int32_t
+arbiter_seek (call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset,
+              gf_seek_what_t what, dict_t *xdata)
+{
+        STACK_UNWIND_STRICT (seek, frame, -1, ENOSYS, 0, xdata);
+        return 0;
+}
+
 int32_t
 mem_acct_init (xlator_t *this)
 {
@@ -328,13 +336,30 @@ fini (xlator_t *this)
 
 struct xlator_fops fops = {
         .lookup = arbiter_lookup,
-        .readv  = arbiter_readv,
+
+        /* Return success for these inode write FOPS without winding it down to
+         * posix; this is needed for AFR write transaction logic to work.*/
         .truncate = arbiter_truncate,
         .writev = arbiter_writev,
         .ftruncate = arbiter_ftruncate,
         .fallocate = arbiter_fallocate,
         .discard = arbiter_discard,
         .zerofill = arbiter_zerofill,
+
+        /* AFR is not expected to wind these inode read FOPS initiated by the
+         * application to the arbiter brick. But in case a bug causes them
+         * to be called, we return ENOSYS. */
+        .readv = arbiter_readv,
+        .seek = arbiter_seek,
+
+        /* The following inode read FOPS initiated by the application are not
+         * wound by AFR either but internal logic like  shd, glfsheal and
+         * client side healing in AFR will send them for selfheal/ inode refresh
+         * operations etc.,so we need to wind them down to posix:
+         *
+         * (f)stat, readdir(p), readlink, (f)getxattr.*/
+
+        /* All other FOPs not listed here are safe to be wound down to posix.*/
 };
 
 struct xlator_cbks cbks = {