starlingx/integ
Code Issues Proposed changes

Add Ceph mds client hung detection

Browse Source 
When there is a buggy cephfs client, the ceph health detail output
will show a message like the one below:

HEALTH_WARN 1 clients failing to respond to capability release; 1 \
              clients failing to advance oldest client/flush tid

MDS_CLIENT_LATE_RELEASE 1 clients failing to respond to capability \
              release
    mds.controller-0(mds.0): Client controller-0 failing to respond\
              to capability release client_id: 774246

MDS_CLIENT_OLDEST_TID 1 clients failing to advance oldest \
              client/flush tid
    mds.controller-0(mds.0): Client controller-0 failing to advance \
              its oldest client/flush tid.  client_id: 774246

When this happens, the cephfs client cannot read or write to
the volume. To restore the communication, it is necessary to force a
client reconnection.

To force this reconnection, the client must be evicted by Ceph. The
client will be disconnected and added to the Ceph blacklist. After
clearing the blacklist, the client will reconnect to the Ceph cluster.

The client hung detection and the eviction procedure are implemented
in the /etc/init.d/ceph script when checking the status of the MDS
process. The script will look for the error output like this one:

 mds.controller-0(mds.0): Client controller-0: failing to respond to \
     capability release client_id: 774246

Test-Plan:
  PASS: Start a pod reading from and writing to a cephfs pvc in a loop
  PASS: Inject the error line to the Ceph health detail output, verify
        the detection appears in the ceph-process-states.log log file
        and check if the client has been evicted and then reconnected.

Closes-bug: 2085648

Signed-off-by: Felipe Sanches Zanoni <Felipe.SanchesZanoni@windriver.com>
Change-Id: I2fad851652cf269b4ebb758b2dfdbe994f2a7b0c
This commit is contained in:
Felipe Sanches Zanoni 2024-10-25 16:54:56 -03:00
parent 2b0c73a1c8
commit df9ae04e54
1 changed files with 118 additions and 39 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

157
ceph/ceph/debian/deb_folder/ceph-base.ceph.init
View File

@ -611,6 +611,59 @@ osd_has_stuck_peering() {
fi
}
mds_has_blocked_clients() {
local name=$1
# Abort if we had previous errors with Ceph
if [ "$CEPH_FAILURE" = "true" ]; then
log $name "WARN" "Ceph cluster is marked as failed, aborting blocked MDS clients check"
return 1
fi
# Cache Ceph Health for later use as calling Ceph takes time This is
# initially cached from the hang check but check and call again here if
# needed
get_ceph_health_detail
if [ $? -ne 0 ]; then
log $name "WARN" "Aborting blocked MDS clients check"
return 1
fi
# Ignore health check if OSDs are administratively down
# Note this can be done with: 'ceph osd set noup; ceph osd down <osd.id>'
$(echo "$CEPH_HEALTH_DETAIL" | grep -q "noup.*set")
if [ $? -eq 0 ]; then
log $name "WARN" "Ceph 'noup' flag is set, aborting blocked ops check"
return 1
fi
# Look for and parse:' mds.controller-0(mds.0): Client controller-0: failing to respond to capability release client_id: 1737491'
local client_id_list=($(echo "$CEPH_HEALTH_DETAIL" | grep "failing to respond to capability release" | sed -rn 's/.*client_id: ([[:digit:]]*).*/\1/p'))
log $name "INFO" "${client_id_list[@]}"
if [[ "$client_id_list" != "" ]]; then
log $name "WARN" "Detected blocked MDS clients: ${client_id_list[@]}"
# Extract the active mds
local active_mds_list=($(echo "$CEPH_HEALTH_DETAIL" | grep "failing to respond to capability release" | sed -rn 's/[[:space:]]+(mds\..*)\(mds.*client_id:.*/\1/p'))
MDS_EVICTION_CMD_LIST=()
local list_end=$(( ${#client_id_list[@]} - 1 ))
# only evict from the active mds
for i in $(seq 0 ${list_end}); do
if [[ ${active_mds_list[$i]} =~ $(hostname) ]]; then
# Form eviction string and add it to the list
MDS_EVICTION_CMD_LIST+=("${active_mds_list[$i]} session evict ${client_id_list[$i]}")
fi
done
if [ ${#MDS_EVICTION_CMD_LIST[@]} -gt 0 ]; then
log $name "INFO" "${MDS_EVICTION_CMD_LIST[@]}"
return 0
fi
fi
return 1
}
######################
#### StarlingX END ###
######################
@ -1196,7 +1249,7 @@ EOF
fi
;;
status)
status)
if daemon_is_running $name ceph-$type $id $pid_file; then
# log ceph osd state
@ -1217,33 +1270,33 @@ EOF
fi
fi
# check if daemon is hung
is_process_hung $name $type
if [ $? -eq 0 ]; then
echo "$name: hung."
# based on http://refspecs.linuxbase.org/LSB_3.1.0/LSB-Core-generic/LSB-Core-generic/iniscrptact.html
# exit codes from 150 to 199 are application specific, therefore we define one here
EXIT_STATUS=150
else
# Wait a period of time prior to OSD start before restarting based on slow/blocked requests
if [ "$type" = "osd" ] && [ $BLOCKED_OPS_DETECTION_ENABLED = "true" ]; then
up_time=$(get_proc_run_time $name)
if [ $up_time -gt $BLOCKED_OPS_START_DETECTION ]; then
osd_has_blocked_ops $name
if [ $? -eq 0 ]; then
echo "$name: blocked ops."
# based on http://refspecs.linuxbase.org/LSB_3.1.0/LSB-Core-generic/LSB-Core-generic/iniscrptact.html
# exit codes from 150 to 199 are application specific, therefore we define one here
EXIT_STATUS=151
else
echo "$name: running."
fi
else
echo "$name: running."
fi
else
echo "$name: running."
fi
# check if daemon is hung
is_process_hung $name $type
if [ $? -eq 0 ]; then
echo "$name: hung."
# based on http://refspecs.linuxbase.org/LSB_3.1.0/LSB-Core-generic/LSB-Core-generic/iniscrptact.html
# exit codes from 150 to 199 are application specific, therefore we define one here
EXIT_STATUS=150
else
# Wait a period of time prior to OSD start before restarting based on slow/blocked requests
if [ "$type" = "osd" ] && [ $BLOCKED_OPS_DETECTION_ENABLED = "true" ]; then
up_time=$(get_proc_run_time $name)
if [ $up_time -gt $BLOCKED_OPS_START_DETECTION ]; then
osd_has_blocked_ops $name
if [ $? -eq 0 ]; then
echo "$name: blocked ops."
# based on http://refspecs.linuxbase.org/LSB_3.1.0/LSB-Core-generic/LSB-Core-generic/iniscrptact.html
# exit codes from 150 to 199 are application specific, therefore we define one here
EXIT_STATUS=151
else
echo "$name: running."
fi
else
echo "$name: running."
fi
else
echo "$name: running."
fi
# Wait a period of time prior to OSD start before restarting based on stuck peering
if [ "$type" = "osd" ] && [ $STUCK_PEERING_DETECTION_ENABLED = "true" ]; then
@ -1264,18 +1317,44 @@ EOF
else
echo "$name: running."
fi
fi
fi
elif [ -e "$pid_file" ]; then
# daemon is dead, but pid file still exists
echo "$name: dead."
EXIT_STATUS=1
else
# daemon is dead, and pid file is gone
echo "$name: not running."
EXIT_STATUS=3
fi
;;
# Check mds daemon
if [ "$type" = "mds" ]; then
log $name "DEBUG" "checking $name for blocked clients"
mds_has_blocked_clients $name
if [ $? -eq 0 ]; then
list_end=$(( ${#MDS_EVICTION_CMD_LIST[@]} - 1 ))
for i in $(seq 0 $list_end); do
log $name "INFO" "Evicting client $(echo ${MDS_EVICTION_CMD_LIST[$i]} | awk '{ print $NF }')"
CEPH_EVICT_CLIENT=""
execute_ceph_cmd CEPH_EVICT_CLIENT $name "ceph daemon ${MDS_EVICTION_CMD_LIST[$i]} && echo success"
rc=$?
if [ ${rc} -ne 0 ]; then
log $name "ERROR" "MDS Client eviction failed: ceph daemon ${MDS_EVICTION_CMD_LIST[$i]}: ${rc} - '${CEPH_EVICT_CLIENT}'"
fi
done
# Clear the Ceph blacklist
log $name "INFO" "Clear ceph blacklist"
CEPH_BLKLIST_CLEAR=""
execute_ceph_cmd CEPH_BLKLIST_CLEAR $name "ceph osd blacklist clear"
rc=$?
if [ ${rc} -ne 0 ]; then
log $name "ERROR" "OSD blacklist clear failed: ${rc} - '${CEPH_BLKLIST_CLEAR}'"
fi
fi
fi
elif [ -e "$pid_file" ]; then
# daemon is dead, but pid file still exists
echo "$name: dead."
EXIT_STATUS=1
else
# daemon is dead, and pid file is gone
echo "$name: not running."
EXIT_STATUS=3
fi
;;
ssh)
$ssh