cluster/dht: Fix min-free-disk calculations when quota-deem-statfs is on

PROBLEM:

As part of file creation, DHT sends a statfs call to all of its
sub-volumes and expects in return the local space consumption and
availability on each one of them. This information is used by DHT to
ensure that atleast min-free-disk amount of space is left on each
sub-volume in the event that there ARE other sub-volumes with more
space available.
But when quota-deem-statfs is enabled, quota xlator on every brick
unwinds the statfs call with volume-wide consumption of disk space.
This leads to miscalculation in min-free-disk algo, thereby misleading
DHT at some point, into thinking all sub-volumes have equal available
space, in which case DHT keeps sending new file creates to subvol-0,
causing it to become 100% full at some point although there ARE other
subvols with ample space available.

FIX:

The fix is to make quota_statfs() behave as if quota xlator weren't
enabled, thereby making every brick return only its local consumption
and disk space availability.

Change-Id: I211371a1eddb220037bd36a128973938ea8124c2
BUG: 1099890
Signed-off-by: Krutika Dhananjay <kdhananj@redhat.com>
Reviewed-on: http://review.gluster.org/7845
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Raghavendra G <rgowdapp@redhat.com>
Reviewed-by: Vijay Bellur <vbellur@redhat.com>

2 files changed, 139 insertions, 0 deletions

diff --git a/tests/bugs/bug-1099890.t b/tests/bugs/bug-1099890.t
new file mode 100644
index 00000000000..1c07b62c02f
--- /dev/null
+++ b/tests/bugs/bug-1099890.t
@@ -0,0 +1,120 @@
+#!/bin/bash
+
+. $(dirname $0)/../include.rc
+. $(dirname $0)/../volume.rc
+. $(dirname $0)/../dht.rc
+
+## TO-DO: Fix the following once the dht du refresh interval issue is fixed:
+## 1. Do away with sleep(1).
+## 2. Do away with creation of empty files.
+
+cleanup;
+
+TEST   glusterd;
+TEST   pidof glusterd;
+
+# Create 2 loop devices, one per brick.
+TEST   truncate -s 100M $B0/brick1
+TEST   truncate -s 100M $B0/brick2
+
+TEST   L1=`losetup --find --show $B0/brick1`
+TEST   mkfs.xfs $L1
+
+TEST   L2=`losetup --find --show $B0/brick2`
+TEST   mkfs.xfs $L2
+
+TEST   mkdir -p $B0/${V0}{1,2}
+
+TEST   mount -t xfs $L1 $B0/${V0}1
+TEST   mount -t xfs $L2 $B0/${V0}2
+
+# Create a plain distribute volume with 2 subvols.
+TEST   $CLI volume create $V0 $H0:$B0/${V0}{1,2};
+
+TEST   $CLI volume start $V0;
+EXPECT "Started" volinfo_field $V0 'Status';
+
+TEST   $CLI volume quota $V0 enable;
+
+TEST   $CLI volume set $V0 features.quota-deem-statfs on
+
+TEST   $CLI volume quota $V0 limit-usage / 150MB;
+
+TEST   $CLI volume set $V0 cluster.min-free-disk 50%
+
+TEST   glusterfs -s $H0 --volfile-id=$V0 $M0
+
+# Make sure quota-deem-statfs is working as expected
+EXPECT "150M" echo `df -h $M0 -P | tail -1 | awk {'print $2'}`
+
+# Create a new file 'foo' under the root of the volume, which hashes to subvol-0
+# of DHT, that consumes 40M
+TEST   dd if=/dev/zero of=$M0/foo bs=5M count=8
+
+TEST   stat $B0/${V0}1/foo
+TEST ! stat $B0/${V0}2/foo
+
+# Create a new file 'bar' under the root of the volume, which hashes to subvol-1
+# of DHT, that consumes 40M
+TEST   dd if=/dev/zero of=$M0/bar bs=5M count=8
+
+TEST ! stat $B0/${V0}1/bar
+TEST   stat $B0/${V0}2/bar
+
+# Touch a zero-byte file on the root of the volume to make sure the statfs data
+# on DHT is refreshed
+sleep 1;
+TEST   touch $M0/empty1;
+
+# At this point, the available space on each subvol {60M,60M} is greater than
+# their min-free-disk {50M,50M}, but if this bug still exists, then
+# the total available space on the volume as perceived by DHT should be less
+# than min-free-disk, i.e.,
+#
+# consumed space returned per subvol by quota = (40M + 40M) = 80M
+#
+# Therefore, consumed space per subvol computed by DHT WITHOUT the fix would be:
+# (80M/150M)*100 = 53%
+#
+# Available space per subvol as perceived by DHT with the bug = 47%
+# which is less than min-free-disk
+
+# Now I create a file that hashes to subvol-1 (counting from 0) of DHT.
+# If this bug still exists,then DHT should be routing this creation to subvol-0.
+# If this bug is fixed, then DHT should be routing the creation to subvol-1 only
+# as it has more than min-free-disk space available.
+
+TEST   dd if=/dev/zero of=$M0/file bs=1K count=1
+sleep 1;
+TEST ! stat $B0/${V0}1/file
+TEST   stat $B0/${V0}2/file
+
+# Touch another zero-byte file on the root of the volume to refresh statfs
+# values stored by DHT.
+
+TEST touch $M0/empty2;
+
+# Now I create a new file that hashes to subvol-0, at the end of which, there
+# will be less than min-free-disk space available on it.
+TEST   dd if=/dev/zero of=$M0/fil bs=5M count=4
+sleep 1;
+TEST   stat $B0/${V0}1/fil
+TEST ! stat $B0/${V0}2/fil
+
+# Touch to refresh statfs info cached by DHT
+
+TEST   touch $M0/empty3;
+
+# Now I create a file that hashes to subvol-0 but since it has less than
+# min-free-disk space available, its data will be cached on subvol-1.
+
+TEST   dd if=/dev/zero of=$M0/zz bs=5M count=1
+
+TEST   stat $B0/${V0}1/zz
+TEST   stat $B0/${V0}2/zz
+
+EXPECT "$V0-client-1" dht_get_linkto_target "$B0/${V0}1/zz"
+
+EXPECT "1" is_dht_linkfile "$B0/${V0}1/zz"
+
+cleanup;
diff --git a/tests/dht.rc b/tests/dht.rc
index a11bbfd8a97..50c4532617e 100644
--- a/tests/dht.rc
+++ b/tests/dht.rc
@@ -91,3 +91,22 @@ function remove_brick_completed()
        echo $val
        return $val
 }
+
+function dht_get_linkto_target()
+{
+        local path=$1;
+        echo `getfattr -d -m . -e text --only-values --absolute-names --name=trusted.glusterfs.dht.linkto $path`
+}
+
+function is_dht_linkfile()
+{
+        local path=$1
+        retval=0
+        local output=`stat --format=%a $path`
+        if [ $output -eq 1000 ]; then
+                 retval=1
+        fi
+
+        echo $retval
+        return $retval
+}