md_info.sh / md_info_detail.sh: /dev/md/ directory does not exist on newer kernels about node-exporter-textfile-collector-scripts

[dswarbrick]

The /dev/disk/by-{id,partlabel,partuuid,path,uuid} directories are created by udev, and thus probably also somewhat variable from one distro to the next (and also not guaranteed to be present at all).

The only entries that are always going to be there are /dev/md[0-9]* and /sys/block/md[0-9]*, as these are created by the md kernel modules, and not dependent on any userspace daemon.

[candlerb]

In that case, is it better not to attempt to get the array names at all?

--- md_info_detail.sh.orig	2021-03-31 10:16:25.593118296 +0100
+++ md_info_detail.sh.new	2021-03-31 10:41:01.406242215 +0100
@@ -6,16 +6,12 @@

 set -eu

-for MD_DEVICE in /dev/md/*; do
-  if [ -b "$MD_DEVICE" ]; then
+for MD_DEVICE_NUM in /dev/md?*; do
+  if [ -b "$MD_DEVICE_NUM" ]; then
   # Subshell to avoid eval'd variables from leaking between iterations
   (
-    # Resolve symlink to discover device, e.g. /dev/md127
-    MD_DEVICE_NUM=$(readlink -f "${MD_DEVICE}")
-
     # Remove /dev/ prefix
     MD_DEVICE_NUM=${MD_DEVICE_NUM#/dev/}
-    MD_DEVICE=${MD_DEVICE#/dev/md/}

     # Query sysfs for info about md device
     SYSFS_BASE="/sys/devices/virtual/block/${MD_DEVICE_NUM}/md"
@@ -64,13 +60,13 @@
       if echo "$line" | grep -E -q "Devices :|Array Size :| Used Dev Size :|Events :"; then
         MDADM_DETAIL_KEY=$(echo "$line" | cut -d ":" -f 1 | tr -cd '[a-zA-Z0-9]._-')
         MDADM_DETAIL_VALUE=$(echo "$line" | cut -d ":" -f 2 | cut -d " " -f 2 | sed 's:^ ::')
-        echo "node_md_info_${MDADM_DETAIL_KEY}{md_device=\"${MD_DEVICE_NUM}\", md_name=\"${MD_DEVICE}\", raid_level=\"${MD_LEVEL}\", md_num_raid_disks=\"${MD_NUM_RAID_DISKS}\", md_metadata_version=\"${MD_METADATA_VERSION}\"} ${MDADM_DETAIL_VALUE}"
+        echo "node_md_info_${MDADM_DETAIL_KEY}{md_device=\"${MD_DEVICE_NUM}\", raid_level=\"${MD_LEVEL}\", md_num_raid_disks=\"${MD_NUM_RAID_DISKS}\", md_metadata_version=\"${MD_METADATA_VERSION}\"} ${MDADM_DETAIL_VALUE}"
       fi
     done  <<< "$MDADM_DETAIL_OUTPUT"

     # Output RAID detail metrics info from the output of "mdadm --detail"
     # NOTE: Sending this info as labels rather than separate metrics, because some of them can be strings.
-    echo -n "node_md_info{md_device=\"${MD_DEVICE_NUM}\", md_name=\"${MD_DEVICE}\", raid_level=\"${MD_LEVEL}\", md_num_raid_disks=\"${MD_NUM_RAID_DISKS}\", md_metadata_version=\"${MD_METADATA_VERSION}\""
+    echo -n "node_md_info{md_device=\"${MD_DEVICE_NUM}\", raid_level=\"${MD_LEVEL}\", md_num_raid_disks=\"${MD_NUM_RAID_DISKS}\", md_metadata_version=\"${MD_METADATA_VERSION}\""
     while IFS= read -r line ; do
       # Filter for lines with a ":", to use for Key/Value pairs in labels
       if echo "$line" | grep -E -q ":" ; then

Or else the code could work multiple ways (one where /dev/md exists, one where it doesn't)

[dswarbrick]

I think that attempting to obtain the array name is ok, so long as it fails gracefully. But obviously the glob pattern at the start of the for-loop should use a more robust / reliable value than /dev/md/*.

Also I don't think that the lack of a /dev/md directory has got anything to do with kernel versions, as that was always a userspace function of the mdadm tool. The man page describes the various scenarios under which it will create the symlinks in the /dev/md directory. In some ways, this is duplicating the functionality of udev, but IIRC, mdadm predates udev.