TRIM & discard configuration on Linux

TRIM informs SSDs and NVMe drives which blocks are no longer in use, allowing the garbage collector to proactively erase them. Without TRIM, the drive doesn't know which blocks contain deleted data, leading to write amplification and performance degradation over time.

How TRIM works

Flash memory is organized in pages (typically 4-16 KiB) that can be written individually but must be erased in blocks (typically 128-512 pages). The erase-before-write cycle causes write amplification: writing 4 KiB may require reading, erasing, and rewriting an entire 4 MiB block.

TRIM solves this by telling the drive which pages are free, so garbage collection can erase them ahead of time:

• ATA TRIM: SATA SSDs use the TRIM command (ATA standard)
• NVMe Dataset Management: NVMe uses the DSM command for the same purpose
• SCSI UNMAP: SCSI/SAS drives use UNMAP

fstrim (periodic batch TRIM, recommended)

# Trim a specific mounted filesystem
fstrim -v /

# Trim all mounted filesystems that support it
fstrim --all -v

# Trim with offset and length
fstrim -o 10G -l 50G /mnt/data

# Minimum extent size to trim
fstrim -m 64M /mnt/data

fstrim.timer (systemd weekly timer)

# Enable weekly TRIM
sudo systemctl enable --now fstrim.timer

# Check status
systemctl status fstrim.timer

# View timer configuration
systemctl cat fstrim.timer

The default timer runs weekly with a randomized delay. This is the recommended approach, Red Hat, Debian, Arch Wiki, and Ubuntu (enabled by default since 18.04) all recommend periodic fstrim over continuous discard.

discard mount option (continuous TRIM)

# Mount with continuous discard
mount -o discard /dev/sdb1 /mnt/data

# In /etc/fstab
/dev/sdb1 /mnt/data ext4 defaults,discard 0 1

Filesystem support

Btrfs async discard

Since kernel 5.6, Btrfs supports discard=async, which moves TRIM out of the transaction commit path and rate-limits the discard I/O. This eliminates the latency penalty of synchronous discard while still reclaiming space promptly. It is the default since kernel 6.2 when the device supports it.

# Explicit async discard
mount -o discard=async /dev/sdb1 /mnt

# Synchronous (legacy behavior)
mount -o discard=sync /dev/sdb1 /mnt

blkdiscard (whole-device TRIM)

# Discard entire device (DESTROYS ALL DATA)
blkdiscard /dev/sdX

# Discard a range
blkdiscard -o 10G -l 50G /dev/sdX

# Secure discard (if supported)
blkdiscard -s /dev/sdX

# Zero-fill discard (write zeroes to discarded blocks, if supported)
blkdiscard -z /dev/sdX

Use cases: wiping an SSD before repartitioning, returning space to thin-provisioned storage. Note that TRIM doesn't guarantee immediate erasure. The SSD controller decides when to erase blocks.

NVMe format (secure erase)

# User data erase
nvme format /dev/nvme0n1 --ses=1

# Cryptographic erase (delete encryption key)
nvme format /dev/nvme0n1 --ses=2

# Format specific namespace
nvme format /dev/nvme0 --namespace-id=1 --ses=1

# Check if crypto-erase is supported
nvme id-ctrl /dev/nvme0 | grep fna

RAID and TRIM

mdadm

Since kernel 3.7, md supports TRIM for linear, RAID 0, RAID 1, RAID 5, and RAID 10. RAID 1 is safe, both mirrors receive the same TRIM. RAID 5/6 requires a kernel parameter because drives must return zeroes on discarded regions:

# Kernel command line for RAID 5/6 TRIM
raid456.devices_handle_discards_safely=y

# Test before enabling in production
cat /sys/block/md0/md/mismatch_cnt
mdadm --action=check /dev/md0 && mdadm --wait /dev/md0
fstrim -v /
mdadm --action=check /dev/md0 && mdadm --wait /dev/md0
cat /sys/block/md0/md/mismatch_cnt   # should still be 0

dm-crypt and TRIM

# Enable discard on LUKS mapping
cryptsetup --allow-discards open /dev/sdb1 crypt

# Refresh an existing mapping
cryptsetup refresh --allow-discards crypt

LVM thin pools and TRIM

LVM thin pools can pass discard commands through to underlying devices. Three modes control this:

# Set discard mode when creating a thin pool
lvconvert --type thin-pool --discards passdown \
  --poolmetadata VG/ThinMetaLV VG/ThinDataLV

# Change mode of existing thin pool
lvchange --discards passdown VG/ThinPoolLV

# Display current mode
lvs -o+discards VG/ThinPoolLV

Note: thin snapshots keep chunks from being freed, so fstrim on a thin LV with snapshots may not reclaim as much space as expected.

Recommendation summary

1. Use periodic fstrim (weekly timer). This is the universal recommendation.
2. Avoid the continuous discard mount option except on Btrfs with discard=async.
3. Enable allow_discards on dm-crypt if the security trade-off is acceptable for your threat model.
4. Use --discards passdown on LVM thin pools to reclaim space on underlying devices.
5. Test RAID 5/6 TRIM with mismatch_cnt before enabling it in production.
6. Use blkdiscard or nvme format when repurposing or disposing of drives.