cluster/afr: Fix data loss due to race between sh and ongoing write.

        Backport of: http://review.gluster.org/#/c/13001/

Problem:

When IO is happening on a file and a brick goes down comes back up
during this time, protocol/client translator attempts reopening of the
fd on the gfid handle of the file. But if another client renames this
file while a brick was down && writes were in progress on it, once this
brick is back up, there can be a race between reopening of the fd and
entry self-heal replaying the effect of the rename() on the sink brick.
If the reopening of the fd happens first, the application's writes
continue to go into the data blocks associated with the gfid.
Now entry-self-heal deletes 'src' and creates 'dst' file on the sink,
marking dst as a 'newentry'.  Data self-heal is also completed on 'dst'
as a result and self-heal terminates. If at this point the application
is still writing to this fd, all writes on the file after self-heal
would go into the data blocks associated with this fd, which would be
lost once the fd is closed. The result - the 'dst' file on the source
and sink are not the same and there is no pending heal on the file,
leading to silent corruption on the sink.

Fix:

Leverage http://review.gluster.org/#/c/12816/ to ensure the gfid handle
path gets saved in .glusterfs/unlink until the fd is closed on the file.
During this time, when self-heal sends mknod() with gfid of the file,
do the following:
link() the gfid handle under .glusterfs/unlink to the new path to be
created in mknod() and
rename() the gfid handle to go back under .glusterfs/ab/cd/.


Change-Id: I5dc49c127ef0a1bf3cf4ce1b24610b1527f84d6f
BUG: 1293265
Signed-off-by: Krutika Dhananjay <kdhananj@redhat.com>
Reviewed-on: http://review.gluster.org/13036
Reviewed-by: Pranith Kumar Karampuri <pkarampu@redhat.com>
Tested-by: Pranith Kumar Karampuri <pkarampu@redhat.com>
Smoke: Gluster Build System <jenkins@build.gluster.com>
NetBSD-regression: NetBSD Build System <jenkins@build.gluster.org>
CentOS-regression: Gluster Build System <jenkins@build.gluster.com>

1 files changed, 57 insertions, 0 deletions

diff --git a/tests/bugs/replicate/bug-1292379.t b/tests/bugs/replicate/bug-1292379.t
new file mode 100644
index 00000000000..f0865020d54
--- /dev/null
+++ b/tests/bugs/replicate/bug-1292379.t
@@ -0,0 +1,57 @@
+#!/bin/bash
+
+. $(dirname $0)/../../include.rc
+. $(dirname $0)/../../volume.rc
+. $(dirname $0)/../../afr.rc
+. $(dirname $0)/../../fileio.rc
+
+cleanup
+
+TEST glusterd
+TEST pidof glusterd
+TEST $CLI volume info
+
+TEST $CLI volume create $V0 replica 2 $H0:$B0/${V0}{0,1}
+TEST $CLI volume start $V0
+EXPECT 'Started' volinfo_field $V0 'Status'
+TEST $CLI volume set $V0 cluster.self-heal-daemon off
+TEST $CLI volume set $V0 cluster.entry-self-heal off
+TEST $CLI volume set $V0 cluster.data-self-heal off
+TEST $CLI volume set $V0 cluster.metadata-self-heal off
+TEST $CLI volume set $V0 cluster.eager-lock off
+
+TEST glusterfs --volfile-id=$V0 --volfile-server=$H0 $M0
+
+TEST wfd=`fd_available`
+TEST fd_open $wfd "w" $M0/a
+
+TEST fd_write $wfd "abcd"
+
+# Kill brick-0
+TEST kill_brick $V0 $H0 $B0/${V0}0
+
+# While brick-0 is down, rename 'a' to 'b'
+TEST mv $M0/a $M0/b
+
+TEST fd_write $wfd "lmn"
+
+TEST $CLI volume start $V0 force
+EXPECT_WITHIN $CHILD_UP_TIMEOUT "1" afr_child_up_status $V0 0
+TEST fd_write $wfd "pqrs"
+TEST $CLI volume set $V0 self-heal-daemon on
+EXPECT_WITHIN $CHILD_UP_TIMEOUT "1" afr_child_up_status_in_shd $V0 0
+EXPECT_WITHIN $CHILD_UP_TIMEOUT "1" afr_child_up_status_in_shd $V0 1
+
+TEST $CLI volume heal $V0
+
+EXPECT_WITHIN $HEAL_TIMEOUT "^0$" get_pending_heal_count $V0
+
+TEST fd_write $wfd "xyz"
+TEST fd_close $wfd
+
+md5sum_b0=$(md5sum $B0/${V0}0/b | awk '{print $1}')
+
+EXPECT "$md5sum_b0" echo `md5sum $B0/${V0}1/b | awk '{print $1}'`
+
+TEST umount $M0
+cleanup