SMART disk diagnostics on a server: how to read it and what to do

SMART data is read with smartctl from the smartmontools package: smartctl -a /dev/sda for SATA and SAS, smartctl -a /dev/nvme0 for NVMe. You replace a drive based on specific counters, not on the "PASSED" verdict: reallocated sectors, pending sectors, interface CRC errors and the remaining write endurance.

The line SMART overall-health self-assessment test result: PASSED only means no attribute has crossed the vendor threshold. A drive with a hundred reallocated sectors prints it too.
Three things are dangerous: growing Reallocated_Sector_Ct, a non-zero Current_Pending_Sector, and any value in Offline_Uncorrectable.
NVMe has its own fields: percentage_used, media_errors, critical_warning, available_spare.
The smartd daemon runs tests on a schedule and mails you as soon as a counter moves.

Installation and the first read

The package name is identical in both families; add nvme-cli for NVMe.

# Debian/Ubuntu
apt-get install -y smartmontools nvme-cli
# RHEL, AlmaLinux, Rocky
dnf install -y smartmontools nvme-cli

# list the devices smartctl can see
smartctl --scan

# full report for a SATA/SAS drive
smartctl -a /dev/sda

# attributes only, no header
smartctl -A /dev/sda

# NVMe: health log
smartctl -a /dev/nvme0
nvme smart-log /dev/nvme0

If the disks sit behind a RAID controller, add the device type to the command, for example -d megaraid,8. How to tell which array implementation you have is covered in hardware RAID or mdadm.

Which attributes matter

Attribute	Meaning	Healthy	Action
5 Reallocated_Sector_Ct	sectors remapped from the spare area	0 and stable	growth within a week means plan a replacement
197 Current_Pending_Sector	sectors that fail to read but are not remapped yet	0	any non-zero value means run a long test
198 Offline_Uncorrectable	sectors not recovered during the background scan	0	replace the drive
199 UDMA_CRC_Error_Count	transfer errors on the bus	0 and stable	cable or backplane slot, not the disk
187 Reported_Uncorrect	errors the drive reported to the host	0	replace the drive

NVMe reads differently

NVMe has no classic attribute table; it exposes a health log with fixed fields instead. Four of them matter: critical_warning (anything other than zero is an incident), percentage_used (write endurance consumed in percent, where 100 means the rated TBW is exhausted), media_errors (uncorrectable media errors) and available_spare together with available_spare_threshold.

Self-tests

The short test checks the electronics and part of the surface in 1–2 minutes; the long one reads the whole surface and runs for hours. The drive executes both in the background while the server keeps working.

# short test
smartctl -t short /dev/sda
# long test
smartctl -t long /dev/sda
# abort the running test
smartctl -X /dev/sda

# results of the latest tests
smartctl -l selftest /dev/sda

# drive error log
smartctl -l error /dev/sda 2>&1 | head -n 40

Automatic monitoring with smartd

# config: /etc/smartd.conf (RHEL) or /etc/smartmontools/smartd.conf (Debian/Ubuntu)
# one line covering every detected drive:
# DEVICESCAN -a -o on -S on -n standby,q -s (S/../.././02|L/../../6/03) -W 4,45,55 -m root@localhost

systemctl enable --now smartd        # Debian/Ubuntu
systemctl enable --now smartd        # RHEL, AlmaLinux, Rocky
systemctl status smartd

The schedule above means a short test every night at 02:00 and a long one on Saturdays at 03:00. Mail arrives only if the server has a working mail transport; on VDS outgoing SMTP is closed by default and opened on request, so shipping SMART metrics into a monitoring system is more reliable, as described in the guide on server hardware monitoring.

A common mistake is calling a drive healthy because smartctl -H returned PASSED. That verdict fires only on a factory threshold crossing, while a drive with hundreds of reallocated sectors has not reached the threshold yet and keeps returning read errors. Inside a RAID that looks like random I/O stalls with no array degradation message at all. Check it differently: capture smartctl -A once a day and compare the RAW_VALUE of attributes 5, 187, 197 and 198 with the previous snapshot. Any increase is a reason to pull the drive and request a disk replacement in the array before it fails outright.

When to replace a drive

Immediately: non-zero Offline_Uncorrectable or Reported_Uncorrect, a non-zero NVMe critical_warning, read failures in the self-test log.
Within days: a growing Reallocated_Sector_Ct, or a Current_Pending_Sector that does not clear after a long test.
Do not replace: growth in UDMA_CRC_Error_Count alone points at the cable or the cage slot, and a new disk changes nothing.

On ZevsHost dedicated servers the on-duty data centre shift swaps a drive on request: you collect the smartctl -a output and the drive serial number, and the site engineers take it from there.

Key takeaways

Use smartctl -a for SATA and SAS, smartctl -a /dev/nvme0 or nvme smart-log for NVMe.
Decide on the RAW_VALUE of attributes 5, 187, 197, 198 and on their trend, not on the PASSED line.
For NVMe watch percentage_used, media_errors, available_spare and critical_warning.

SMART Disk Diagnostics on a Server: smartctl and NVMe