Diagnosing 'No space left on device' when df shows free space

Goal

Find which of the four causes of a write failure with ENOSPC applies, then free the right resource instead of deleting files at random.

Prerequisites

Root or sudo on the affected host, the error message with the path that failed, and df, du, lsof or /proc access.

Steps

1. Locate the filesystem: df -h /path/that/failed prints the mount point; a full /var/log on its own partition looks the same to the application as a full root.
2. Check bytes and inodes separately: df -h and df -i (--inodes) on that mount. An ext4 filesystem is created with a fixed number of inodes derived from bytes-per-inode (mke2fs(8)); a directory of millions of tiny files (sessions, cache shards, mail queues) can exhaust inodes while most of the bytes are still free. If IUse% is 100, find the directory with find /mount -xdev -type d -exec sh -c 'echo $(ls -1A "$1" | wc -l) "$1"' _ {} \; | sort -n | tail.
3. Check for deleted-but-open files: unlink(2) states that a file whose last name is removed but that is still open remains in existence until the last descriptor is closed. lsof +L1 or find /proc/*/fd -lname '*(deleted)' lists them; the usual holder is a logging process or a database with a rotated file. Restart or signal that process (kill -HUP for daemons that reopen logs) rather than deleting more.
4. Check reservation: ext filesystems keep a reserved-blocks-percentage (default 5% per mke2fs(8)) usable only by root, so unprivileged processes fail while df still shows a few percent free. tune2fs -m 1 lowers it on data volumes; keep it on the root filesystem.
5. If space is genuinely used, find it with du -xh --max-depth=1 /mount | sort -h (-x stays on one filesystem) and prefer the owners' own clean-up (journalctl --vacuum-size, container image pruning, log rotation) over rm.
6. Record cause and fix; if a growth trend is visible, add an alert on inode and byte usage per mount, not just on the root filesystem.

Expected result

The write succeeds again, the cause is named (inodes, held file, reservation or real usage) and the monitoring covers the dimension that failed.

Limits and test basis

Filesystem behaviour cited is for the ext family; XFS and btrfs allocate inodes differently and have their own accounting tools. Container overlay filesystems and tmpfs mounts add further limits (tmpfs size, container quotas) not covered here. No measurement or field observation is claimed.