NFS/トラブルシューティング

NFS

よくある問題と解決法の記事。

サーバー側の問題

exportfs: /etc/exports:2: syntax error: bad option list

/etc/exports 内のオプションリストから空白を全て削除してください。

グループ/GID パーミッションの問題

NFS 共有のマウントが問題なく動作し、所有者からアクセスすることができるのに、グループメンバーからアクセスできない場合、ユーザーが属しているグループの数を確認してください。NFS にはユーザーが所属できるグループは最大で16までという制限があります。16以上のグループにユーザーが所属している場合、NFS サーバーで rpc.mountd の起動フラグとして --manage-gids を有効にする必要があります。

/etc/conf.d/nfs-server.conf

# Options for rpc.mountd.
# If you have a port-based firewall, you might want to set up
# a fixed port here using the --port option.
# See rpc.mountd(8) for more details.

MOUNTD_OPTS="--manage-gids"

ファイルを書き込もうとしたときに "Permission denied"

root で共有をマウントして、クライアントから完全に読み書きできるようにしたい場合、/etc/exports のエクスポートに no_root_squash オプションを追加してください:

/var/cache/pacman/pkg 192.168.1.0/24(rw,no_subtree_check,no_root_squash)

showmount -e コマンドを実行したときに "RPC: Program not registered"

サーバー側で nfs-server.service と rpcbind.service が実行されていることを確認してください。systemd を参照。実行されてない場合、サービスを起動して有効化してください。

NFSv3 が有効になっていることも確認してください。showmount は NFSv4 だけのサーバーでは機能しません。

クライアント側の問題

mount.nfs4: No such device

nfs モジュールがロードされているか確認してください:

lsmod | grep nfs

上記コマンドで何も返ってこなかったり nfsd-stuff しか表示されない場合、次を実行:

# modprobe nfs

mount.nfs4: Invalid argument

nfs-client.target を起動・有効化してサーバー上で適切なデーモン (nfs-idmapd や rpc-gssd など) が動作していることを確認してください。

macOS クライアントから接続できない

macOS クライアントから接続したときに、サーバーのログでは問題ないようにみえるのに、macOS が NFS 共有のマウントを拒否することがあります。以下の方法で解決できます:

NFS サーバーで、/etc/exports を編集して insecure オプションを追加して exportfs -r を再実行する。

macOS クライアントで、mount コマンドに resvport オプションを追加する。/etc/nfs.conf でデフォルトのクライアントマウントオプションとして resvport を設定することもできます:

/etc/nfs.conf

nfs.client.mount.options = resvport

デフォルトのクライアントマウントオプションを変更すると Finder から "Connect to Server..." で共有をマウントする場合にも影響します。

macOS クライアントからの接続が不安定

macOS の NFS クライアントは macOS サーバーに最適化されているため Linux サーバーでは複数の問題が存在します。パフォーマンスが出なかったり、頻繁に切断したり、日本語の文字に問題が起こる場合、デフォルトのマウントオプションを編集して Mac クライアントの /etc/nfs.conf に nfs.client.mount.options = intr,locallocks,nfc という行を追加してください。マウントオプションに関する詳細情報はこちらにあります。

巨大なファイルをコピーするときにクライアントが一時的にフリーズする

クライアントマシンから NFS サーバーに巨大なファイルをコピーする際、転送速度に非常に高速だと、数秒後に速度が落ちて、転送が完了するまで暫くの間、クライアントマシンが一時的にロックアップすることがあります。

クライアントの (/etc/fstab の) マウントオプションに sync を追加することで問題は解決します。

mount.nfs: Operation not permitted

nfs-utils 1.2.1-2 以上にアップデートすると、NFS 共有のマウントは機能を停止します。nfs-utils がデフォルトで NFSv3 ではなく NFSv4 を使うようになるからです。この問題はコマンドラインでマウントオプションとして 'vers=3' や 'nfsvers=3' を使うことで解決できます:

# mount.nfs remote target directory -o ...,vers=3,...
# mount.nfs remote target directory -o ...,nfsvers=3,...

もしくは /etc/fstab でマウントする場合:

remote target directory nfs ...,vers=3,... 0 0
remote target directory nfs ...,nfsvers=3,... 0 0

mount.nfs: Protocol not supported

NFS ソースのパスにエクスポートするルートディレクトリを記述した場合に発生します。

例:

# mount SERVER:/srv/nfs4/media /mnt
mount.nfs4: Protocol not supported

代わりに相対パスを使うようにしてください:

# mount SERVER:/media /mnt

Vagrant と synced_folders の問題

Vagrant のスクリプトで NFS を通してフォルダをマウントできない場合、net-tools パッケージをインストールすることで問題は解決します。

パフォーマンスの問題

この NFS Howto ページにパフォーマンスに関する有益な情報が載っています。以下はその他のヒントです:

問題の診断

Htop should be your first port of call. The most obvious symptom will be a maxed-out CPU.
Press F2, and under "Display options", enable "Detailed CPU time". Press F1 for an explanation of the colours used in the CPU bars. In particular, is the CPU spending most of its time responding to IRQs, or in Wait-IO (wio)?

サーバースレッド

Symptoms: Nothing seems to be very heavily loaded, but some operations on the client take a long time to complete for no apparent reason.

If your workload involves lots of small reads and writes (or if there are a lot of clients), there may not be enough threads running on the server to handle the quantity of queries. To check if this is the case, run the following command on one or more of the clients:

# nfsstat -rc

Client rpc stats:
calls      retrans    authrefrsh
113482     0          113484

If the retrans column contains a number larger than 0, the server is failing to respond to some NFS requests, and the number of threads should be increased.

To increase the number of threads on the server, edit the file /etc/conf.d/nfs-server.conf and set the value in the NFSD_OPTS variable. For example, to set the number of threads to 32:

/etc/conf.d/nfs-server.conf

NFSD_OPTS="32"

The default number of threads is 8. Try doubling this number until retrans remains consistently at zero. Don't be afraid of increasing the number quite substantially. 256 threads may be quite reasonable, depending on the workload. You will need to restart the NFS server daemon each time you modify the configuration file. Bear in mind that the client statistics will only be reset to zero when the client is rebooted.

Use htop (disable the hiding of kernel threads) to keep an eye on how much work each nfsd thread is doing. If you reach a point where the retrans values are non-zero, but you can see nfsd threads on the server doing no work, something different is now causing your bottleneck, and you'll need to re-diagnose this new problem.

Close-to-open/flush-on-close

Symptoms: Your clients are writing many small files. The server CPU is not maxed out, but there is very high wait-IO, and the server disk seems to be churning more than you might expect.

In order to ensure data consistency across clients, the NFS protocol requires that the client's cache is flushed (all data is pushed to the server) whenever a file is closed after writing. Because the server is not allowed to buffer disk writes (if it crashes, the client won't realise the data wasn't written properly), the data is written to disk immediately before the client's request is completed. When you're writing lots of small files from the client, this means that the server spends most of its time waiting for small files to be written to its disk, which can cause a significant reduction in throughput.

See this excellent article or the nfs manpage for more details on the close-to-open policy. There are several approaches to solving this problem:

nocto マウントオプション

ノート: The linux kernel does not seem to honour this option properly. Files are still flushed when they're closed.

Does your situation match these conditions?

The export you have mounted on the client is only going to be used by the one client.
It doesn't matter too much if a file written on one client doesn't immediately appear on other clients.
It doesn't matter if after a client has written a file, and the client thinks the file has been saved, and then the client crashes, the file may be lost.

If you're happy with the above conditions, you can use the nocto mount option, which will disable the close-to-open behaviour. See the nfs manpage for details.

async エクスポートオプション

Does your situation match these conditions?

It's important that when a file is closed after writing on one client, it is:
- Immediately visible on all the other clients.
- Safely stored on the server, even if the client crashes immediately after closing the file.
It's not important to you that if the server crashes:
- You may loose the files that were most recently written by clients.
- When the server is restarted, the clients will believe their recent files exist, even though they were actually lost.

In this situation, you can use async instead of sync in the server's /etc/exports file for those specific exports. See the exports manual page for details. In this case, it does not make sense to use the nocto mount option on the client.

バッファキャッシュサイズと MTU

Symptoms: High kernel or IRQ CPU usage, a very high packet count through the network card.

This is a trickier optimisation. Make sure this is definitely the problem before spending too much time on this. The default values are usually fine for most situations.

See this excellent article for information about I/O buffering in NFS. Essentially, data is accumulated into buffers before being sent. The size of the buffer will affect the way data is transmitted over the network. The Maximum Transmission Unit (MTU) of the network equipment will also affect throughput, as the buffers need to be split into MTU-sized chunks before they're sent over the network. If your buffer size is too big, the kernel or hardware may spend too much time splitting it into MTU-sized chunks. If the buffer size is too small, there will be overhead involved in sending a very large number of small packets. You can use the rsize and wsize mount options on the client to alter the buffer cache size. To achieve the best throughput, you need to experiment and discover the best values for your setup.

It is possible to change the MTU of many network cards. If your clients are on a separate subnet (e.g. for a Beowulf cluster), it may be safe to configure all of the network cards to use a high MTU. This should be done in very-high-bandwidth environments.

rsize と wsize についての詳細は nfs の man ページを参照してください。

デバッグ

rpcdebug を使う

Using rpcdebug is the easiest way to manipulate the kernel interfaces in place of echoing bitmasks to /proc.

オプション	説明
-c	Clear the given debug flags
-s	Set the given debug flags
-m module	Specify which module's flags to set or clear.
-v	Increase the verbosity of rpcdebug's output
-h	Print a help message and exit. When combined with the -v option, also prints the available debug flags.

-m オプションでは、以下のモジュールが利用できます:

モジュール	説明
nfsd	NFS サーバー
nfs	NFS クライアント
nlm	NFS クライアントあるいはサーバーの Network Lock Manager
rpc	NFS クライアントあるいはサーバーの Remote Procedure Call モジュール

例:

rpcdebug -m rpc -s all    # sets all debug flags for RPC
rpcdebug -m rpc -c all    # clears all debug flags for RPC

rpcdebug -m nfsd -s all   # sets all debug flags for NFS Server
rpcdebug -m nfsd -c all   # clears all debug flags for NFS Server

Once the flags are set you can tail the journal for the debug output, usually journalctl -fl or similar.

mountstats を使う

nfs-utils パッケージに含まれている mountstats ツールを使うことで、平均時間やパケットサイズなどの統計情報を得ることができます。

$ mountstats 
Stats for example:/tank mounted on /tank:
  NFS mount options: rw,sync,vers=4.2,rsize=524288,wsize=524288,namlen=255,acregmin=3,acregmax=60,acdirmin=30,acdirmax=60,soft,proto=tcp,port=0,timeo=15,retrans=2,sec=sys,clientaddr=xx.yy.zz.tt,local_lock=none
  NFS server capabilities: caps=0xfbffdf,wtmult=512,dtsize=32768,bsize=0,namlen=255
  NFSv4 capability flags: bm0=0xfdffbfff,bm1=0x40f9be3e,bm2=0x803,acl=0x3,sessions,pnfs=notconfigured
  NFS security flavor: 1  pseudoflavor: 0

NFS byte counts:
  applications read 248542089 bytes via read(2)
  applications wrote 0 bytes via write(2)
  applications read 0 bytes via O_DIRECT read(2)
  applications wrote 0 bytes via O_DIRECT write(2)
  client read 171375125 bytes via NFS READ
  client wrote 0 bytes via NFS WRITE

RPC statistics:
  699 RPC requests sent, 699 RPC replies received (0 XIDs not found)
  average backlog queue length: 0

READ:
	338 ops (48%) 
	avg bytes sent per op: 216	avg bytes received per op: 507131
	backlog wait: 0.005917 	RTT: 548.736686 	total execute time: 548.775148 (milliseconds)
GETATTR:
	115 ops (16%) 
	avg bytes sent per op: 199	avg bytes received per op: 240
	backlog wait: 0.008696 	RTT: 15.756522 	total execute time: 15.843478 (milliseconds)
ACCESS:
	93 ops (13%) 
	avg bytes sent per op: 203	avg bytes received per op: 168
	backlog wait: 0.010753 	RTT: 2.967742 	total execute time: 3.032258 (milliseconds)
LOOKUP:
	32 ops (4%) 
	avg bytes sent per op: 220	avg bytes received per op: 274
	backlog wait: 0.000000 	RTT: 3.906250 	total execute time: 3.968750 (milliseconds)
OPEN_NOATTR:
	25 ops (3%) 
	avg bytes sent per op: 268	avg bytes received per op: 350
	backlog wait: 0.000000 	RTT: 2.320000 	total execute time: 2.360000 (milliseconds)
CLOSE:
	24 ops (3%) 
	avg bytes sent per op: 224	avg bytes received per op: 176
	backlog wait: 0.000000 	RTT: 30.250000 	total execute time: 30.291667 (milliseconds)
DELEGRETURN:
	23 ops (3%) 
	avg bytes sent per op: 220	avg bytes received per op: 160
	backlog wait: 0.000000 	RTT: 6.782609 	total execute time: 6.826087 (milliseconds)
READDIR:
	4 ops (0%) 
	avg bytes sent per op: 224	avg bytes received per op: 14372
	backlog wait: 0.000000 	RTT: 198.000000 	total execute time: 198.250000 (milliseconds)
SERVER_CAPS:
	2 ops (0%) 
	avg bytes sent per op: 172	avg bytes received per op: 164
	backlog wait: 0.000000 	RTT: 1.500000 	total execute time: 1.500000 (milliseconds)
FSINFO:
	1 ops (0%) 
	avg bytes sent per op: 172	avg bytes received per op: 164
	backlog wait: 0.000000 	RTT: 2.000000 	total execute time: 2.000000 (milliseconds)
PATHCONF:
	1 ops (0%) 
	avg bytes sent per op: 164	avg bytes received per op: 116
	backlog wait: 0.000000 	RTT: 1.000000 	total execute time: 1.000000 (milliseconds)

カーネルインターフェイス

A bitmask of the debug flags can be echoed into the interface to enable output to syslog; 0 is the default:

/proc/sys/sunrpc/nfsd_debug
/proc/sys/sunrpc/nfs_debug
/proc/sys/sunrpc/nlm_debug
/proc/sys/sunrpc/rpc_debug

Sysctl controls are registered for these interfaces, so they can be used instead of echo:

sysctl -w sunrpc.rpc_debug=1023
sysctl -w sunrpc.rpc_debug=0

sysctl -w sunrpc.nfsd_debug=1023
sysctl -w sunrpc.nfsd_debug=0

At runtime the server holds information that can be examined:

grep . /proc/net/rpc/*/content
cat /proc/fs/nfs/exports
cat /proc/net/rpc/nfsd
ls -l /proc/fs/nfsd

A rundown of /proc/net/rpc/nfsd (the userspace tool nfsstat pretty-prints this info):

* rc (reply cache): <hits> <misses> <nocache>
- hits: client it's retransmitting
- misses: a operation that requires caching
- nocache: a operation that no requires caching

* fh (filehandle): <stale> <total-lookups> <anonlookups> <dir-not-in-cache> <nodir-not-in-cache>
- stale: file handle errors
- total-lookups, anonlookups, dir-not-in-cache, nodir-not-in-cache
  . always seem to be zeros

* io (input/output): <bytes-read> <bytes-written>
- bytes-read: bytes read directly from disk
- bytes-written: bytes written to disk

* th (threads): <threads> <fullcnt> <10%-20%> <20%-30%> ... <90%-100%> <100%>
- threads: number of nfsd threads
- fullcnt: number of times that the last 10% of threads are busy
- 10%-20%, 20%-30% ... 90%-100%: 10 numbers representing 10-20%, 20-30% to 100%
  . Counts the number of times a given interval are busy

* ra (read-ahead): <cache-size> <10%> <20%> ... <100%> <not-found>
- cache-size: always the double of number threads
- 10%, 20% ... 100%: how deep it found what was looking for
- not-found: not found in the read-ahead cache

* net: <netcnt> <netudpcnt> <nettcpcnt> <nettcpconn>
- netcnt: counts every read
- netudpcnt: counts every UDP packet it receives
- nettcpcnt: counts every time it receives data from a TCP connection
- nettcpconn: count every TCP connection it receives

* rpc: <rpccnt> <rpcbadfmt+rpcbadauth+rpcbadclnt> <rpcbadfmt> <rpcbadauth> <rpcbadclnt>
- rpccnt: counts all rpc operations
- rpcbadfmt: counts if while processing a RPC it encounters the following errors:
  . err_bad_dir, err_bad_rpc, err_bad_prog, err_bad_vers, err_bad_proc, err_bad
- rpcbadauth: bad authentication
  . does not count if you try to mount from a machine that it's not in your exports file
- rpcbadclnt: unused

* procN (N = vers): <vs_nproc> <null> <getattr> <setattr> <lookup> <access> <readlink> <read> <write> <create> <mkdir> <symlink> <mknod> <remove> <rmdir> <rename> <link> <readdir> <readdirplus> <fsstat> <fsinfo> <pathconf> <commit>
- vs_nproc: number of procedures for NFS version
  . v2: nfsproc.c, 18
  . v3: nfs3proc.c, 22
  - v4, nfs4proc.c, 2
- statistics: generated from NFS operations at runtime

* proc4ops: <ops> <x..y>
- ops: the definition of LAST_NFS4_OP, OP_RELEASE_LOCKOWNER = 39, plus 1 (so 40); defined in nfs4.h
- x..y: the array of nfs_opcount up to LAST_NFS4_OP (nfsdstats.nfs4_opcount[i])

NFSD デバッグフラグ

/usr/include/linux/nfsd/debug.h

/*
 * knfsd debug flags
 */
#define NFSDDBG_SOCK            0x0001
#define NFSDDBG_FH              0x0002
#define NFSDDBG_EXPORT          0x0004
#define NFSDDBG_SVC             0x0008
#define NFSDDBG_PROC            0x0010
#define NFSDDBG_FILEOP          0x0020
#define NFSDDBG_AUTH            0x0040
#define NFSDDBG_REPCACHE        0x0080
#define NFSDDBG_XDR             0x0100
#define NFSDDBG_LOCKD           0x0200
#define NFSDDBG_ALL             0x7FFF
#define NFSDDBG_NOCHANGE        0xFFFF

NFS デバッグフラグ

/usr/include/linux/nfs_fs.h

/*
 * NFS debug flags
 */
#define NFSDBG_VFS              0x0001
#define NFSDBG_DIRCACHE         0x0002
#define NFSDBG_LOOKUPCACHE      0x0004
#define NFSDBG_PAGECACHE        0x0008
#define NFSDBG_PROC             0x0010
#define NFSDBG_XDR              0x0020
#define NFSDBG_FILE             0x0040
#define NFSDBG_ROOT             0x0080
#define NFSDBG_CALLBACK         0x0100
#define NFSDBG_CLIENT           0x0200
#define NFSDBG_MOUNT            0x0400
#define NFSDBG_FSCACHE          0x0800
#define NFSDBG_PNFS             0x1000
#define NFSDBG_PNFS_LD          0x2000
#define NFSDBG_STATE            0x4000
#define NFSDBG_ALL              0xFFFF

NLM デバッグフラグ

/usr/include/linux/lockd/debug.h

/*
 * Debug flags
 */
#define NLMDBG_SVC		0x0001
#define NLMDBG_CLIENT		0x0002
#define NLMDBG_CLNTLOCK		0x0004
#define NLMDBG_SVCLOCK		0x0008
#define NLMDBG_MONITOR		0x0010
#define NLMDBG_CLNTSUBS		0x0020
#define NLMDBG_SVCSUBS		0x0040
#define NLMDBG_HOSTCACHE	0x0080
#define NLMDBG_XDR		0x0100
#define NLMDBG_ALL		0x7fff

RPC デバッグフラグ

/usr/include/linux/sunrpc/debug.h

/*
 * RPC debug facilities
 */
#define RPCDBG_XPRT             0x0001
#define RPCDBG_CALL             0x0002
#define RPCDBG_DEBUG            0x0004
#define RPCDBG_NFS              0x0008
#define RPCDBG_AUTH             0x0010
#define RPCDBG_BIND             0x0020
#define RPCDBG_SCHED            0x0040
#define RPCDBG_TRANS            0x0080
#define RPCDBG_SVCXPRT          0x0100
#define RPCDBG_SVCDSP           0x0200
#define RPCDBG_MISC             0x0400
#define RPCDBG_CACHE            0x0800
#define RPCDBG_ALL              0x7fff

一般注記事項

While the number of threads can be increased at runtime via an echo to /proc/fs/nfsd/threads, the cache size (double the threads, see the ra line of /proc/net/rpc/nfsd) is not dynamic. The NFS daemon must be restarted with the new thread size during initialization in order for the thread cache to properly adjust.

参照

他の問題

パーミッションの問題

ファイルのパーミッションが正しく設定できない場合、chown を実行しているユーザー・グループがクライアントとサーバーの両方に存在することを確認してください。

NFSv4 を使用していて全てのファイルの所有者が nobody になる場合:

systemd の場合、nfs-idmapd サービスが起動していることを確認してください。
initscripts の場合、/etc/conf.d/nfs-common.conf の NEED_IDMAPD が YES に設定されていることを確認してください。

On some systems detecting the domain from FQDN minus hostname does not seem to work reliably. If files are still showing as nobody after the above changes, edit /etc/idmapd.conf, ensure that Domain is set to FQDN minus hostname. For example:

/etc/idmapd.conf

[General]

Verbosity = 7
Pipefs-Directory = /var/lib/nfs/rpc_pipefs
Domain = yourdomain.local

[Mapping]

Nobody-User = nobody
Nobody-Group = nobody

[Translation]

Method = nsswitch

If nfs-idmapd.service refuses to start because it cannot open the Pipefs-directory (defined in /etc/idmapd.conf and appended with '/nfs'), issue a mkdir-command and restart the daemon.