SSD health: read SMART on Windows and Linux

Your SSD will not send a postcard before it dies. It does keep a few counters, though, and reading them takes five minutes. Just do not mistake a smoke detector for a fortune teller.

The short answer

SSD health: how do you read SMART on Windows and Linux? Back up first, then read the drive's own counters. On Windows, Settings shows them for NVMe drives only [8]; PowerShell shows a few more [9][10]; smartctl from smartmontools reads the full report on Windows and Linux [1]; on Linux, nvme-cli reads the NVMe log [6]. On an NVMe drive, look at four fields first: Critical Warning, Available Spare, Percentage Used and Media and Data Integrity Errors [3][4]. A clean report is not a guarantee, and a high number is not a death date. The NVMe standard says a Percentage Used of 100 may not indicate a failure [3][4], and Google's field studies found drives that failed without any SMART warning [15] and flash drives with frequent uncorrectable errors [16]. Our assessment: treat SMART as a smoke detector, log it monthly, and let the backup do the real protecting. We make no lifetime predictions from a single reading.

How it works

1. Back up first, then test. Reading SMART is a light query. A long self-test or a recovery attempt is not. The author of CrystalDiskInfo says it plainly: if the tool shows "Caution" or "Bad", back up your important data to another storage device before you investigate the cause [13]. The same page says that a "Caution" does not mean the drive stops today, but also gives no guarantee for months or years [13]. Our assessment: the order is backup, read, log, and only then self-test. How to run a backup you can restore is on our 3-2-1 page.
2. What SMART is. Self-Monitoring, Analysis and Reporting Technology is the set of counters a drive keeps about itself. SATA drives report numbered attributes with a normalized VALUE, a WORST value, a THRESHOLD and a RAW value; NVMe drives report one standardized page, the SMART / Health Information log (log identifier 02h) [3][4]. The meaning of SATA attributes is up to the vendor. smartctl says so itself: an attribute of type "Pre-fail" does not mean the disk is about to fail, and it only matters when the current normalized value is at or below the threshold, which smartctl then marks FAILING_NOW (or In_the_past if only the worst value ever got there) [1]. CrystalDiskInfo shows how vendor-specific it is: it reads Micron's attribute 0xCA, Intel's 0xE8, SandForce's 0xE7 or Samsung's 0xB4 as the "life" reference, depending on the controller [12].

The NVMe fields that matter

They come from the log page that the NVMe standard defines [3]; the descriptions below follow the libnvme documentation, which quotes the standard [4][5].

SATA SSDs

Windows Settings monitors NVMe SSDs only, not SATA SSDs or hard disks [8]. For a SATA SSD use smartctl or a tool such as CrystalDiskInfo. CrystalDiskInfo marks an SSD "Caution" at a remaining life of 10 percent or less for the controller families it knows [12], and its author notes that it does not support every drive and may misjudge some [13]. smartctl reports its own exit status as a bit mask: bit 3 means the SMART status check returned "DISK FAILING", bit 4 means pre-fail attributes at or below the threshold, bit 6 and bit 7 mean that the error log or the self-test log contain records of errors [1]. The man page defines these bits for ATA disks; for other drive types, check the page of your version. Read bits 6 and 7 as "look at the logs", not as "the drive is dead".

Windows: three ways

First, the Settings app. Open Start > Settings > System > Storage > Advanced storage settings > Disks & volumes, then Properties of the drive [8]. For NVMe drives Windows shows critical warnings, "Estimated remaining life", "Available spare" and "Temperature" [8]. Second, PowerShell. The first line lists your drives with bus type and health, and the second reads the reliability counters [9][10][11]:

Get-PhysicalDisk | Select-Object FriendlyName, MediaType, BusType, HealthStatus, OperationalStatus, Size
Get-PhysicalDisk | Get-StorageReliabilityCounter | Format-List DeviceId, Temperature, Wear, PowerOnHours, ReadErrorsTotal, ReadErrorsUncorrected, WriteErrorsTotal, WriteErrorsUncorrected

Not every drive, driver or USB bridge fills every field. Microsoft's own example output shows an empty Wear value [9], so read a blank as "not reported", never as zero. HealthStatus is Windows' summary and no replacement for the full report (our assessment). Third, smartctl. Install smartmontools, open a terminal with administrator rights (our assessment: Windows usually requires it) and let smartctl list the devices; use the name it prints for your drive, and do not copy a name from another computer [1]:

smartctl --scan
smartctl -a <device name as printed by --scan>

For NVMe, -a is equivalent to -H -i -c -A -l error -l selftest [1].

Linux: three ways

Find the drive first. Then read it with smartctl, with nvme-cli, or as JSON for a log [1][6][17][18]:

lsblk -d -o NAME,MODEL,SERIAL,SIZE,TRAN
sudo smartctl -a /dev/nvme0
sudo nvme list
sudo nvme log smart /dev/nvme0
sudo nvme log error /dev/nvme0

sudo nvme smart-log /dev/nvme0 and sudo nvme error-log /dev/nvme0 are the names before nvme-cli 3.0; the old commands still work in 3.0, print a deprecation warning and may disappear in a future release [6][7]. Check your version with nvme version. For a SATA SSD the device is usually /dev/sda (check the lsblk output), and sudo smartctl -a /dev/sda is the same call. smartctl can also answer in JSON: smartctl -j -A /dev/nvme0 returns the NVMe values under nvme_smart_health_information_log, for example percentage_used, data_units_written, media_errors, available_spare and critical_warning [18].

USB enclosures and adapters

A USB bridge sits between you and the drive and may block the SMART commands. For SATA SSDs, the SCSI/ATA Translation standard defines a vendor-independent pass-through that many USB-SATA bridges support, which smartctl uses with -d sat [1][2]. For NVMe, a smartmontools developer wrote in 2018 that no vendor-independent equivalent exists [2]; smartctl therefore has bridge-specific types such as sntasmedia, sntjmicron and sntrealtek, and since smartctl 7.5 an experimental /sat suffix that falls back to SATA for combined M.2 adapters [1]. Use a type only when the scan result or the adapter documentation supports it, and be careful with usbjmicron,x: the man page warns of I/O errors and a disconnecting drive when the bridge does not support it [1]. If a query fails, the cause may be the adapter, and the failure alone does not show a defective SSD (our assessment). When in doubt, read the drive directly in a computer.

Self-tests

smartctl -t short and smartctl -t long start the drive's own tests, and smartctl -l selftest prints the results [1]. On NVMe drives, the man page marks -t short and -t long as an experimental feature since smartctl 7.4, and the drive has to support them [1]. A passed self-test is no guarantee either. Run it after the backup, not instead of it.

From counters to terabytes

Data Units Written counts thousands of 512-byte units, so one unit stands for 512,000 bytes [14]. Terabytes written (decimal) = Data Units Written × 512,000 ÷ 10¹². The datasheet's TBW is the manufacturer's rating for host writes before the drive wears out; the Write Amplification Factor is NAND writes divided by host writes, and DWPD is TBW divided by 365, the warranty years and the capacity [14]. Because of write amplification, the drive's own Percentage Used and the share of the datasheet TBW you have written need not match (our assessment). Treat TBW as a warranty figure and Percentage Used as the drive's estimate. Neither is a countdown.

Log it: one line per month

A single reading says little; a series shows a direction. This line appends one row to a CSV file (run it with the right permissions for smartctl):

smartctl -j -A /dev/nvme0 | jq -r '.nvme_smart_health_information_log | [(now|todate), .percentage_used, .data_units_written, .media_errors, .available_spare, .critical_warning, .power_on_hours] | @csv' >> ssd-log.csv

We tested the jq filter on a hand-made sample, not on a real drive, and the key names follow the smartctl_exporter source [18]. Google's flash study found that previous errors of various types predict later uncorrectable errors [16], which is why a rising error count matters more than a single number (our assessment).

What to do with the numbers

A scenario, not our data

Imagine an NVMe drive that has run 3,120 power-on hours and reports these values (invented for this example): Critical Warning 0, Available Spare 100 percent with a threshold of 10 percent, Percentage Used 2, Media and Data Integrity Errors 0, Error Information Log Entries 0, Unsafe Shutdowns 14, Data Units Written 7,935,401 and Data Units Read 10,584,293. The arithmetic: 7,935,401 × 512,000 bytes ≈ 4.06 TB written, and 10,584,293 × 512,000 bytes ≈ 5.42 TB read (decimal terabytes). If the datasheet said 600 TBW (an invented number), 4.06 TB would be well under one percent of that rating, while the drive itself reports 2 percent used: a gap you would expect from the drive's own estimate and write amplification, and no cause for alarm. The 14 unsafe shutdowns are worth a look at the power and the habit of switching off at the wall. The plan: back up, note the date and the values in the log, read again in a month. No forecast of the drive's remaining life comes out of this single reading.

Checklist

• Back up your data before you run a self-test or open a case with a vendor.
• Identify the drive and its bus: NVMe or SATA, internal or behind USB.
• Read the report with smartctl, plus Settings or nvme-cli as a second view.
• Check Critical Warning, Available Spare, Percentage Used and Media and Data Integrity Errors first.
• Convert Data Units Written to terabytes and write the number and the date down.
• Treat blank values as "not reported", never as zero.
• Run -d sat or the bridge type only when the scan or the adapter documentation supports it.
• Run a self-test after the backup and read it with -l selftest.
• Log the key values once a month and compare with the last line.
• Replace the drive when a warning, a media error or a rising error count appears, not when you have guessed a date.

Related pages: Understanding NVMe SSDs; 3-2-1 backup and restore testing; Reading Backblaze Drive Stats before you buy; Offsite backup for a homelab: the third copy; S.M.A.R.T.: drive health report; TRIM for SSDs; TBW: SSD endurance rating. Our assessment: a drive with clean SMART values still fails sometimes, so the question that decides is whether a restore from your backup works. We cite no prices here.