cluster/afr: Fix resolution issues with afr

Problem with afr:
Lets say there is a directory hierarchy a/b/c/d on the mount and the
user is cd'ed into the directory. Bring down one of the bricks of replica and
remove all directories/files to simulate disk replacement on that brick. Now
this brick is brought back up. Creates on the cd'ed directory fail with ESTALE.
Basically before sending a create of 'f' inside 'd', fuse sends a lookup to
make sure the file is not present.  On one of the bricks  'd' is present and
'f' is not so it sends ENOENT as response. On the new brick 'd' itself is not
present. So it sends ESTALE. In afr ESTALE is considered to be special errno on
witnessing which lookup has to fail. And ESTALE is given more priority than
ENOENT. Due to these reasons lookup fails with ESTALE rather than ENOENT. Since
lookup didn't fail with ENOENT, 'create' can't be issued so the command is
failed with ESTALE.

Solution:
Afr needs to consider ESTALE errno normally and ENOENT needs to
be given more priority so that operations like create can proceed even when
only one of the brick is up and running. Whenever client xlator identifies
that gfid-changed, it sets that information in lookup xdata. Afr uses this
information to fail the lookup with ESTALE so that top xlator can send
fresh lookup.

Change-Id: Ie8e0e327542fd644409eb5dadf451679afa1c0e5
BUG: 1112348
Signed-off-by: Pranith Kumar K <pkarampu@redhat.com>
Reviewed-on: http://review.gluster.org/8154
Tested-by: Justin Clift <justin@gluster.org>
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Ravishankar N <ravishankar@redhat.com>
Reviewed-by: Niels de Vos <ndevos@redhat.com>

2 files changed, 50 insertions, 1 deletions

diff --git a/tests/basic/afr/resolve.t b/tests/basic/afr/resolve.t
new file mode 100644
index 00000000000..7dd432996f6
--- /dev/null
+++ b/tests/basic/afr/resolve.t
@@ -0,0 +1,49 @@
+#!/bin/bash
+
+. $(dirname $0)/../../include.rc
+. $(dirname $0)/../../volume.rc
+
+#Check that operations succeed after changing the disk of the brick while
+#a brick is down
+cleanup;
+
+TEST glusterd
+TEST pidof glusterd
+TEST $CLI volume create $V0 replica 2 $H0:$B0/${V0}{0,1}
+TEST $CLI volume set $V0 cluster.self-heal-daemon off
+TEST $CLI volume start $V0
+TEST glusterfs --volfile-id=/$V0 --volfile-server=$H0 $M0
+TEST glusterfs --volfile-id=/$V0 --volfile-server=$H0 $M1
+TEST cd $M0
+TEST mkdir -p a/b/c/d/e
+TEST cd a/b/c/d/e
+echo abc > g
+
+#Simulate disk replacement
+TEST kill_brick $V0 $H0 $B0/${V0}0
+rm -rf $B0/${V0}0/.glusterfs $B0/${V0}0/a
+
+TEST $CLI volume start $V0 force
+EXPECT_WITHIN $CHILD_UP_TIMEOUT "1" afr_child_up_status $V0 0
+#Test that the lookup returns ENOENT instead of ESTALE
+#If lookup returns ESTALE this command will fail with ESTALE
+TEST touch f
+
+#Test that ESTALE is ignored when there is a good copy
+EXPECT abc cat g
+
+#Simulate file changing only one mount
+#create the file on first mount
+echo ghi > $M0/b
+
+#re-create the file on other mount while one of the bricks is down.
+TEST kill_brick $V0 $H0 $B0/${V0}0
+TEST rm -f $M1/b
+echo jkl > $M1/b
+#Clear the extended attributes on the directory to create a scenario where
+#gfid-mismatch happened. This should result in EIO
+TEST setfattr -x trusted.afr.$V0-client-0 $B0/${V0}1
+TEST $CLI volume start $V0 force
+EXPECT_WITHIN $CHILD_UP_TIMEOUT "1" afr_child_up_status $V0 0
+TEST ! cat $M0/b
+cleanup
diff --git a/tests/include.rc b/tests/include.rc
index 40dd8ac96b9..3017de9e653 100644
--- a/tests/include.rc
+++ b/tests/include.rc
@@ -7,7 +7,7 @@ B0=${B0:=/d/backends};        # top level of brick directories
 H0=${H0:=`hostname --fqdn`};  # hostname
 DEBUG=${DEBUG:=0}             # turn on debugging?
 statedumpdir=`gluster --print-statedumpdir`; # Default directory for statedump
-
+CHILD_UP_TIMEOUT=60
 CLI="gluster --mode=script";
 
 mkdir -p $B0;