afr: don't treat all cases all bricks being blamed as split-brain

Problem:
We currently don't have a roll-back/undoing of post-ops if quorum is not
met. Though the FOP is still unwound with failure, the xattrs remain on
the disk.  Due to these partial post-ops and partial heals (healing only when
2 bricks are up), we can end up in split-brain purely from the afr
xattrs point of view i.e each brick is blamed by atleast one of the
others. These scenarios are hit when there is frequent
connect/disconnect of the client/shd to the bricks while I/O or heal
are in progress.

Fix:
Instead of undoing the post-op, pick a source based on the xattr values.
If 2 bricks blame one, the blamed one must be treated as sink.
If there is no majority, all are sources. Once we pick a source,
self-heal will then do the heal instead of erroring out due to
split-brain.

Change-Id: I3d0224b883eb0945785ade0e9697a1c828aec0ae
BUG: 1539358
Signed-off-by: Ravishankar N <ravishankar@redhat.com>

5 files changed, 165 insertions, 9 deletions

diff --git a/tests/basic/afr/arbiter-add-brick.t b/tests/basic/afr/arbiter-add-brick.t
index fe919de0ab4..77b93d9a210 100644
--- a/tests/basic/afr/arbiter-add-brick.t
+++ b/tests/basic/afr/arbiter-add-brick.t
@@ -12,6 +12,8 @@ TEST $CLI volume set $V0 performance.stat-prefetch off
 TEST $CLI volume start $V0
 TEST $CLI volume set $V0 self-heal-daemon off
 TEST $GFS --volfile-id=$V0 --volfile-server=$H0 $M0;
+EXPECT_WITHIN $PROCESS_UP_TIMEOUT "1" afr_child_up_status $V0 0
+EXPECT_WITHIN $PROCESS_UP_TIMEOUT "1" afr_child_up_status $V0 1
 TEST mkdir  $M0/dir1
 TEST dd if=/dev/urandom of=$M0/file1 bs=1024 count=1
 
@@ -24,6 +26,7 @@ TEST dd if=/dev/urandom of=$M0/file1 bs=1024 count=1024
 #convert replica 2 to arbiter volume
 TEST $CLI volume start $V0 force
 EXPECT_WITHIN $PROCESS_UP_TIMEOUT "1" brick_up_status $V0 $H0 $B0/${V0}1
+EXPECT_WITHIN $PROCESS_UP_TIMEOUT "1" afr_child_up_status $V0 1
 
 #syntax check for add-brick.
 TEST ! $CLI volume add-brick $V0 replica 2 arbiter 1 $H0:$B0/${V0}2
@@ -31,6 +34,19 @@ TEST ! $CLI volume add-brick $V0 replica 3 arbiter 2 $H0:$B0/${V0}2
 
 TEST $CLI volume add-brick $V0 replica 3 arbiter 1 $H0:$B0/${V0}2
 EXPECT_WITHIN $PROCESS_UP_TIMEOUT "1" brick_up_status $V0 $H0 $B0/${V0}2
+EXPECT_WITHIN $PROCESS_UP_TIMEOUT "1" afr_child_up_status $V0 2
+
+#Trigger name heals from client. If we just rely on index heal, the first index
+#crawl on B0 fails for /, dir2 and /file either due to lock collision or files
+#not being present on the other 2 bricks yet. It is getting healed only in the
+#next crawl after priv->shd.timeout (600 seconds) or by manually launching
+#index heal again.
+TEST $CLI volume set $V0 data-self-heal off
+TEST $CLI volume set $V0 metadata-self-heal off
+TEST $CLI volume set $V0 entry-self-heal off
+TEST stat $M0/dir1
+TEST stat $M0/dir2
+TEST stat $M0/file1
 
 #Heal files
 TEST $CLI volume set $V0 self-heal-daemon on
diff --git a/tests/bugs/replicate/bug-1539358-split-brain-detection.t b/tests/bugs/replicate/bug-1539358-split-brain-detection.t
new file mode 100755
index 00000000000..7b71a7a9e7d
--- /dev/null
+++ b/tests/bugs/replicate/bug-1539358-split-brain-detection.t
@@ -0,0 +1,89 @@
+#!/bin/bash
+
+. $(dirname $0)/../../include.rc
+. $(dirname $0)/../../volume.rc
+. $(dirname $0)/../../afr.rc
+
+cleanup;
+
+## Start and create a volume
+TEST glusterd;
+TEST pidof glusterd;
+TEST $CLI volume info;
+
+TEST $CLI volume create $V0 replica 3 $H0:$B0/${V0}{0,1,2};
+TEST $CLI volume start $V0;
+EXPECT 'Started' volinfo_field $V0 'Status';
+TEST $CLI volume set $V0 self-heal-daemon off
+TEST glusterfs --volfile-id=/$V0 --volfile-server=$H0 $M0 --attribute-timeout=0 --entry-timeout=0
+
+###############################################################################yy
+# Case of 2 bricks blaming the third and the third blaming the other two.
+
+TEST `echo "hello" >> $M0/file`
+
+# B0 and B2 must blame B1
+TEST kill_brick $V0 $H0 $B0/$V0"1"
+TEST `echo "append" >> $M0/file`
+EXPECT "00000001" afr_get_specific_changelog_xattr $B0/${V0}0/file trusted.afr.$V0-client-1 data
+EXPECT "00000001" afr_get_specific_changelog_xattr $B0/${V0}2/file trusted.afr.$V0-client-1 data
+CLIENT_MD5=$(md5sum $M0/file | cut -d\  -f1)
+
+# B1 must blame B0 and B2
+setfattr -n trusted.afr.$V0-client-0 -v 0x000000010000000000000000 $B0/$V0"1"/file
+setfattr -n trusted.afr.$V0-client-2 -v 0x000000010000000000000000 $B0/$V0"1"/file
+
+# Launch heal
+TEST $CLI volume start $V0 force
+EXPECT_WITHIN $PROCESS_UP_TIMEOUT "1" brick_up_status $V0 $H0 $B0/${V0}1
+TEST $CLI volume set $V0 self-heal-daemon on
+EXPECT_WITHIN $PROCESS_UP_TIMEOUT "Y" glustershd_up_status
+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
+EXPECT_WITHIN $CHILD_UP_TIMEOUT "1" afr_child_up_status_in_shd $V0 2
+TEST $CLI volume heal $V0
+EXPECT_WITHIN $HEAL_TIMEOUT "^0$" get_pending_heal_count $V0
+B0_MD5=$(md5sum $B0/${V0}0/file | cut -d\  -f1)
+B1_MD5=$(md5sum $B0/${V0}1/file | cut -d\  -f1)
+B2_MD5=$(md5sum $B0/${V0}2/file | cut -d\  -f1)
+TEST [ "$CLIENT_MD5" == "$B0_MD5" ]
+TEST [ "$CLIENT_MD5" == "$B1_MD5" ]
+TEST [ "$CLIENT_MD5" == "$B2_MD5" ]
+
+TEST rm $M0/file
+
+###############################################################################yy
+# Case of each brick blaming the next one in a cyclic manner
+
+TEST `echo "hello" >> $M0/file`
+# Mark cyclic xattrs and modify file content directly on the bricks.
+TEST $CLI volume set $V0 self-heal-daemon off
+setfattr -n trusted.afr.$V0-client-1 -v 0x000000010000000000000000 $B0/$V0"0"/file
+setfattr -n trusted.afr.dirty -v 0x000000010000000000000000 $B0/$V0"0"/file
+setfattr -n trusted.afr.$V0-client-2 -v 0x000000010000000000000000 $B0/$V0"1"/file
+setfattr -n trusted.afr.dirty -v 0x000000010000000000000000 $B0/$V0"1"/file
+setfattr -n trusted.afr.$V0-client-0 -v 0x000000010000000000000000 $B0/$V0"2"/file
+setfattr -n trusted.afr.dirty -v 0x000000010000000000000000 $B0/$V0"2"/file
+
+TEST `echo "ab" >> $B0/$V0"0"/file`
+TEST `echo "cdef" >> $B0/$V0"1"/file`
+TEST `echo "ghi" >> $B0/$V0"2"/file`
+
+# Add entry to xattrop dir to trigger index heal.
+xattrop_dir0=$(afr_get_index_path $B0/$V0"0")
+base_entry_b0=`ls $xattrop_dir0`
+gfid_str=$(gf_gfid_xattr_to_str $(gf_get_gfid_xattr $B0/$V0"0"/file))
+ln $xattrop_dir0/$base_entry_b0 $xattrop_dir0/$gfid_str
+EXPECT_WITHIN $HEAL_TIMEOUT "^1$" get_pending_heal_count $V0
+
+# Launch heal
+TEST $CLI volume set $V0 self-heal-daemon on
+TEST $CLI volume heal $V0
+EXPECT_WITHIN $HEAL_TIMEOUT "^0$" get_pending_heal_count $V0
+B0_MD5=$(md5sum $B0/${V0}0/file | cut -d\  -f1)
+B1_MD5=$(md5sum $B0/${V0}1/file | cut -d\  -f1)
+B2_MD5=$(md5sum $B0/${V0}2/file | cut -d\  -f1)
+TEST [ "$B0_MD5" == "$B1_MD5" ]
+TEST [ "$B0_MD5" == "$B2_MD5" ]
+###############################################################################yy
+cleanup
diff --git a/tests/bugs/replicate/bug-802417.t b/tests/bugs/replicate/bug-802417.t
index c5ba98b65fd..f213439401e 100755
--- a/tests/bugs/replicate/bug-802417.t
+++ b/tests/bugs/replicate/bug-802417.t
@@ -10,6 +10,18 @@ function write_file()
 }
 
 cleanup;
+
+#####################################################
+# We are currently not triggering data heal unless all bricks of the replica are
+# up. We will need to modify this .t once the fix for preventing stale reads
+# being served to clients for files in spurious split-brains is done. Spurious
+# split-brains here means afr xattrs indicates sbrain but it is actually not.
+# Self-heal will heal such files automatically but before the heal completes,
+# reads can be served which needs fixing.
+#G_TESTDEF_TEST_STATUS_NETBSD7=BAD_TEST,BUG=000000
+#G_TESTDEF_TEST_STATUS_CENTOS6=BAD_TEST,BUG=000000
+######################################################
+
 TEST glusterd
 TEST pidof glusterd
 TEST $CLI volume info;
diff --git a/xlators/cluster/afr/src/afr-self-heal-common.c b/xlators/cluster/afr/src/afr-self-heal-common.c
index 311da68e8ee..7e2a6027126 100644
--- a/xlators/cluster/afr/src/afr-self-heal-common.c
+++ b/xlators/cluster/afr/src/afr-self-heal-common.c
@@ -1455,6 +1455,36 @@ afr_does_witness_exist (xlator_t *this, uint64_t *witness)
         return _gf_false;
 }
 
+unsigned int
+afr_get_quorum_count (afr_private_t *priv)
+{
+        if (priv->quorum_count == AFR_QUORUM_AUTO) {
+                return priv->child_count/2 + 1;
+        } else {
+                return priv->quorum_count;
+        }
+}
+
+void
+afr_selfheal_post_op_failure_accounting (afr_private_t *priv, char *accused,
+                                         unsigned char *sources,
+                                         unsigned char *locked_on)
+{
+        int i = 0;
+        unsigned int  quorum_count = 0;
+
+        if (AFR_COUNT (sources, priv->child_count) != 0)
+                return;
+
+        quorum_count = afr_get_quorum_count (priv);
+        for (i = 0; i < priv->child_count; i++) {
+                if ((accused[i] < quorum_count) && locked_on[i]) {
+                        sources[i] = 1;
+                }
+        }
+        return;
+}
+
 /*
  * This function determines if a self-heal is required for a given inode,
  * and if needed, in what direction.
@@ -1490,6 +1520,7 @@ afr_selfheal_find_direction (call_frame_t *frame, xlator_t *this,
         char *accused = NULL;/* Accused others without any self-accusal */
         char *pending = NULL;/* Have pending operations on others */
         char *self_accused = NULL; /* Accused itself */
+        int min_participants = -1;
 
 	priv = this->private;
 
@@ -1513,8 +1544,13 @@ afr_selfheal_find_direction (call_frame_t *frame, xlator_t *this,
                 }
         }
 
+        if (type == AFR_DATA_TRANSACTION) {
+                min_participants = priv->child_count;
+        } else {
+                min_participants = AFR_SH_MIN_PARTICIPANTS;
+        }
         if (afr_success_count (replies,
-                               priv->child_count) < AFR_SH_MIN_PARTICIPANTS) {
+                               priv->child_count) < min_participants) {
                 /* Treat this just like locks not being acquired */
                 return -ENOTCONN;
         }
@@ -1530,11 +1566,10 @@ afr_selfheal_find_direction (call_frame_t *frame, xlator_t *this,
 	for (i = 0; i < priv->child_count; i++) {
 		for (j = 0; j < priv->child_count; j++) {
                         if (matrix[i][j]) {
-                                 if (!self_accused[i])
-                                         accused[j] = 1;
-
-                                 if (i != j)
-                                         pending[i] = 1;
+                                if (!self_accused[i])
+                                        accused[j] += 1;
+                                if (i != j)
+                                        pending[i] += 1;
                          }
 		}
 	}
@@ -1575,6 +1610,10 @@ afr_selfheal_find_direction (call_frame_t *frame, xlator_t *this,
                 }
         }
 
+        if (type == AFR_DATA_TRANSACTION)
+                afr_selfheal_post_op_failure_accounting (priv, accused,
+                                                         sources, locked_on);
+
          /* If no sources, all locked nodes are sinks - split brain */
          if (AFR_COUNT (sources, priv->child_count) == 0) {
                 for (i = 0; i < priv->child_count; i++) {
diff --git a/xlators/cluster/afr/src/afr-self-heal-data.c b/xlators/cluster/afr/src/afr-self-heal-data.c
index 7880a611592..c033a8afc07 100644
--- a/xlators/cluster/afr/src/afr-self-heal-data.c
+++ b/xlators/cluster/afr/src/afr-self-heal-data.c
@@ -684,7 +684,7 @@ __afr_selfheal_data (call_frame_t *frame, xlator_t *this, fd_t *fd,
 	ret = afr_selfheal_inodelk (frame, this, fd->inode, this->name, 0, 0,
 				    data_lock);
 	{
-		if (ret < AFR_SH_MIN_PARTICIPANTS) {
+		if (ret < priv->child_count) {
                         gf_msg_debug (this->name, 0, "%s: Skipping "
                                       "self-heal as only %d number "
                                       "of subvolumes "
@@ -749,7 +749,7 @@ restore_time:
         if (!is_arbiter_the_only_sink) {
                 ret = afr_selfheal_inodelk (frame, this, fd->inode, this->name,
                                             0, 0, data_lock);
-                if (ret < AFR_SH_MIN_PARTICIPANTS) {
+                if (ret < priv->child_count) {
                         ret = -ENOTCONN;
                         did_sh = _gf_false;
                         goto skip_undo_pending;
@@ -878,7 +878,7 @@ afr_selfheal_data (call_frame_t *frame, xlator_t *this, inode_t *inode)
 	                                        priv->sh_domain, 0, 0,
 				                locked_on);
 	{
-		if (ret < AFR_SH_MIN_PARTICIPANTS) {
+		if (ret < priv->child_count) {
                         gf_msg_debug (this->name, 0, "%s: Skipping "
                                       "self-heal as only %d number of "
                                       "subvolumes could be locked",