nfs-doctor is a small command line tool written in C to help debug NFS servers from the client side.

The idea is simple: you give one IP or hostname, and the tool checks the things that usually break in NFS: network, rpcbind, NFS versions, mountd, exports, permissions, root squash, locking, stale handles and some basic performance.

It is not magic, and it will not replace a good server side analysis. But it helps a lot to understand if the problem is network, NFS config, permissions, UID/GID mapping, or something more strange.

What this tool does

Today nfs-doctor can do these checks:

• test if rpcbind TCP port 111 is reachable
• test if NFS TCP port 2049 is reachable
• query the RPC service map from rpcbind
• detect registered NFS, mountd, lockd/NLM and statd/NSM services
• test NFS v2, v3 and v4 with RPC NULLPROC
• test mountd v1, v2 and v3
• optionally test RPC over UDP with --udp
• enumerate exports using mountd
• mount exports automatically under /tmp/nfsdoctor-*
• try NFS v4 first, then fallback to NFS v3
• run filesystem checks after mount
• test read/traverse permission
• test directory listing
• test ACL xattr visibility
• test create/write/read/fsync when allowed
• test advisory locks with fcntl
• detect practical root_squash behavior
• simulate UID/GID access
• simulate supplemental groups with --groups
• run metadata latency test with create/rename/unlink
• run stale file handle loop looking for ESTALE
• cleanup temporary files, mounts and folders
• generate JSON report for automation
• run Docker fixture tests for regression checks

By default the output is compact. If you want all details, use --verbose.

Important note

NFS problems are very environment dependent. The result can change because of:

• firewall rules
• server export options
• NFS version
• kernel client state
• UID/GID mapping
• root squash
• ACLs
• SELinux/AppArmor on server
• server load
• stale file handles that only happen during real use

So, if the tool says no ESTALE happened, it means only that the tool did not reproduce it during the test window. It does not mean the problem can never happen.

Build requirements

On Debian or Ubuntu:

sudo apt-get update
sudo apt-get install -y build-essential pkg-config libtirpc-dev

On Fedora/RHEL style distros:

sudo dnf install -y gcc make pkgconf-pkg-config libtirpc-devel

For live mount tests you also need NFS client tools.

Debian or Ubuntu:

sudo apt-get install -y nfs-common

Fedora/RHEL style distros:

sudo dnf install -y nfs-utils

Build

Normal build:

make

Clean and rebuild:

make rebuild

Small self-check:

make check

Install:

sudo make install

Install in another prefix:

make PREFIX=/opt/nfs-doctor install

Uninstall:

sudo make uninstall

Manual compile, if you want:

gcc -O2 -Wall -Wextra -I/usr/include/tirpc nfsdiag.c -ltirpc -o nfsdiag

Basic usage

Full diagnostic:

sudo ./nfsdiag 192.168.1.10

Verbose mode:

sudo ./nfsdiag --verbose 192.168.1.10

Only network and RPC checks, without mounting anything:

./nfsdiag --no-mount 192.168.1.10

Test only one export:

sudo ./nfsdiag --export /data 192.168.1.10

Pass mount options:

sudo ./nfsdiag --mount-options soft,timeo=30,retrans=2 192.168.1.10

Do not create/write test files:

sudo ./nfsdiag --read-only 192.168.1.10

Keep the temp folder for manual inspection:

sudo ./nfsdiag --keep-temp 192.168.1.10

Output style

Default output is clean and short. For example, in a healthy server it can be something like:

nfsdiag: 192.168.0.21
[OK] 1 export(s) discovered
summary: ok=13 warn=0 fail=0

If you want to see all probe steps, use:

./nfsdiag --verbose 192.168.0.21

Warnings and failures always appear in normal mode. Informational and low-level OK messages only appear in verbose mode.

JSON output

For automation, use JSON.

JSON to stdout, without human text mixed together:

./nfsdiag --json 192.168.1.10

JSON to file, keeping stdout empty:

./nfsdiag --json=report.json 192.168.1.10

The JSON includes:

• tool name
• host
• timestamp
• summary
• options used
• events
• recommendations

UID/GID and permission tests

Simulate one UID/GID:

sudo ./nfsdiag --uid 1000 --gid 1000 192.168.1.10

Simulate more than one identity:

sudo ./nfsdiag --uid 1000 --gid 1000 --uid 65534 --gid 65534 192.168.1.10

Simulate supplemental groups:

sudo ./nfsdiag --uid 1000 --gid 1000 --groups 10,20,30 192.168.1.10

This is useful because many NFS problems are not really NFS protocol problems. Many times it is UID, GID, groups, ACL, or root squash.

Performance and stale handle tests

Change write/read test size:

sudo ./nfsdiag --bench-bytes 167772160 192.168.1.10

Change metadata latency iterations:

sudo ./nfsdiag --bench-iterations 500 192.168.1.10

Change stale handle loop:

sudo ./nfsdiag --stale-iterations 1000 192.168.1.10

The performance test is only a smoke test. It is not a replacement for fio, nfsiostat, Prometheus, Grafana, or a real benchmark.

Safety options

Timeout for external commands like mount and umount:

sudo ./nfsdiag --command-timeout 15 192.168.1.10

Try to isolate live mounts in a private mount namespace:

sudo ./nfsdiag --mount-namespace 192.168.1.10

This needs root or CAP_SYS_ADMIN.

Network/protocol options

Probe UDP RPC too:

./nfsdiag --no-mount --udp 192.168.1.10

Force IPv4 direct TCP checks:

./nfsdiag --ipv4-only --no-mount 192.168.1.10

Force IPv6 direct TCP checks:

./nfsdiag --ipv6-only --no-mount nfs-server.example.com

Disable NFSv4 pseudo-root fallback:

sudo ./nfsdiag --no-nfs4-discovery 192.168.1.10

The NFSv4 fallback is useful when the server is NFSv4-only and mountd is not available.

Command line reference

Usage: ./nfsdiag [OPTIONS] <server-ip-or-hostname>

Options:
  -e, --export PATH          Test only this export
  -o, --mount-options OPTS   Extra mount options
      --no-mount             Run network/RPC checks only
      --keep-temp            Do not remove /tmp/nfsdoctor-* after tests
      --read-only            Do not create/write test files
      --uid UID              Simulate access as UID
      --gid GID              GID paired with last --uid
      --groups G1,G2         Supplemental groups for simulation
      --timeout SEC          Network/RPC timeout
      --command-timeout SEC  Timeout for mount/umount commands
      --mount-namespace      Use private mount namespace when possible
      --json[=PATH]          Write JSON report
      --udp                  Probe RPC over UDP too
      --ipv4-only            Force IPv4 direct TCP checks
      --ipv6-only            Force IPv6 direct TCP checks
      --no-nfs4-discovery    Disable NFSv4 pseudo-root fallback
      --bench-iterations N   Metadata latency iterations
      --stale-iterations N   ESTALE loop iterations
      --bench-bytes BYTES    Bytes used in read/write smoke test
  -v, --verbose              Show detailed output
  -h, --help                 Show help

Exit codes

• 0: no warnings or failures
• 1: warning or failure found
• 2: usage error or local runtime error

Warnings return 1 because in automation they usually need attention.

Docker fixtures

The project has Docker fixtures to reproduce bad NFS situations.

List fixtures:

make docker-list

Build all fixtures:

make docker-build-all

Build one fixture:

make docker-build-read-only-export

Run automated fixture tests:

make test-fixtures

Run only one test:

make test-fixture-rpcbind-unreachable

Some tests need root because they do real NFS mounts from the host. If the host cannot run kernel NFS inside Docker, the test runner skips those cases instead of failing everything.

The current fixture set includes:

• rpcbind-unreachable
• nfs-port-unreachable
• rpc-map-missing-nfs
• mountd-unavailable
• empty-exports
• mount-denied
• permission-denied
• acl-unsupported
• identity-denied
• read-only-export
• root-squash
• locking-missing
• stale-handle
• slow-performance

What was added in the last big update

This version has these improvements:

1. timeout for external commands
2. cleanup with atexit and signal handling
3. JSON-only output
4. optional mount namespace
5. NFSv4 pseudo-root fallback
6. UDP and IPv4/IPv6 controls
7. supplemental group simulation
8. recommendations based on symptoms
9. metadata latency benchmark
10. automated Docker fixture tests

Security notes

Be careful when running against production exports.

By default the tool may create hidden .nfsdoctor-* files to test write/read behavior. If you do not want this, use:

--read-only

Also, UID/GID simulation requires root because the tool uses setgid, setgroups, and setuid in child processes.

Limitations

Some things are impossible to guarantee from the client side:

• ESTALE only appears if the handle becomes stale during the test
• SELinux/AppArmor problems can look only like generic permission denied
• ACL info depends on what the NFS client exposes
• performance numbers are only smoke-test values
• Docker NFS fixtures depend on host kernel and Docker privileges

So use this tool as a fast diagnostic helper, not as the only source of truth.