An Introduction to NFS
Avishay Traeger
IBM Haifa Research Lab Internal Storage Course ― November 2010 v1.2 Outline
The Basics NFSv2 NFSv3 NFSv4 NFSv4.1
2 Typical Use
ws-avishay ws-bob ws-carl mount -t nfs (NFS Client) (NFS Client) (NFS Client) nfsserv:/home /home
10.0.2.56 10.0.2.103 10.0.2.81
nfsserv (NFS Server) /etc/exports: / /home 10.0.2.*(rw)
home …
avishay bob carl Some benefits of NFS: 1. All clients have the same view 2. Centralized storage management RAID Storage
3 NFS Evolution
NFS is a standardized protocol
Version Year RFC # Pages Status
NFSv2 1989 1094 27 Obsolete
NFSv3 1995 1813 126 Most popular
Available on several OSs, slowly NFSv4 2003 3530 275 but surely replacing NFSv3
NFSv4.1 2010 5661 617 Early adopters only
4 Design Goals
OS independence & interoperability Simple crash recovery for clients and servers Transparent access (client programs do not know files are remote) Maintain local file system semantics Reasonable performance
5 Remote Procedure Call (RPC)
NFS is defined as a set of RPCs – their arguments, results, and effects RPCs are synchronous The use of RPCs makes the protocol easier to understand
6 NFS Client/Server
7 Outline
The Basics NFSv2 NFSv3 NFSv4 NFSv4.1
8 Stateless Protocol
The server does not keep state for RPCs Each RPC contains the necessary information to complete the call This makes crash recovery easy Server crash: server does no crash recovery, clients resubmit requests Client crash: no crash recovery for client or server
This is nice in theory, but Adds complexity Not really stateless... File locking adds state, provided by separate protocol & daemon Server keeps an RPC reply cache to handle duplicate non-
idempotent RPC 9 File Handles
The most common NFS procedure parameter is a structure called a file handle (fh, fhandle) Provided by the server and used by the client to reference a file The fhandle is opaque to the client New fhandles returned by LOOKUP, CREATE, MKDIR, ... The fhandle for the root of the file system is obtained by the client when it mounts the file system
10 Operations
NULL() returns () Do nothing procedure to used for pinging the server LOOKUP(dirfh, name) returns (fh, attr) Returns a new fh and attributes for the named file in the directory specified by dirfh CREATE(dirfh, name, attr) returns (newfh, attr) Creates a new file name in the directory dirfh and returns the new fh and attributes. REMOVE(dirfh, name) returns (status) Removes the file name from directory dirfh.
11 Operations
GETATTR(fh) returns (attr) Returns file attributes (similar to stat syscall) SETATTR(fh, attr) returns (attr) Sets the mode, uid, gid, size, access time, and modify time of a file. Setting the size to zero truncates the file.
12 Operations
READ(fh, offset, count) returns (attr, data) Returns up to count bytes of data from a file starting offset bytes into the file. Returns the attributes of the file. WRITE(fh, offset, count, data) returns (attr) Writes count bytes of data to a file beginning at offset bytes from the beginning of the file. Returns the new attributes of the file after the write.
13 Operations
RENAME(dirfh, name, tofh, toname) returns (status) Renames name in directory dirfh, to toname in directory tofh. LINK(dirfh, name, tofh, toname) returns (status) Creates a hard link toname in directory tofh, that points to name in directory dirfh.
14 Operations
SYMLINK(dirfh, name, string) returns (status) Creates a symlink name in the directory dirfh with value string. The server does not interpret the string argument in any way, just saves it and makes an association to the new symlink file. READLINK(fh) returns (string) Returns the string which is associated with the symlink file.
15 Operations
MKDIR(dirfh, name, attr) returns (fh, newattr) Creates a new directory name in the directory dirfh and returns the new fh and attributes. RMDIR(dirfh, name) returns (status) Removes the empty directory name from the parent directory dirfh.
STATFS(fh) returns (fsstats) Returns file system information such as block size, number of free blocks, etc.
16 Operations
READDIR (dirfh, cookie, count) returns (entries) Returns up to count bytes of directory entries from the directory dirfh. Each entry contains a file name, file id, and an opaque pointer to the next directory entry called a cookie. The cookie is used in subsequent readdir calls to start reading at a specific entry in the directory. A readdir call with the cookie of zero returns entries starting with the first entry in the directory.
17 The MOUNT Protocol
The MOUNT protocol takes a directory pathname and returns an fhandle if the client has permissions to mount the file system Separate protocol Easier to plug in new permission check methods Separates the OS-dependent aspects of the protocol Other OS implementations can change the MOUNT protocol without having to change the NFS protocol
18 The Linux File Handle
Remember that information contained in the fhandle is only meaningful on the server If the local FS on the server reuses an inode number, an NFS client could mistakenly use an old file handle and access the new file. File systems include generation numbers in the inode to avoid this. The value is usually taken from a counter used across the file system. Important file handle fields: Major/minor number of the exported device Inode number Generation number 19 Security
NFSv2 uses UNIX-style permission checks The client passes uid/gid info in RPCs, and the server performs permission checks as if the user was performing the operation locally Problem – the mapping from uid/gid to user must be the same on the client and server Can be solved via Network Information Service (NIS) Another problem – should root on the client have root access to files on the server? Server specifies policy
20 Cache Consistency Problems
Clients use caching and write buffering to improve performance, but this causes issues Problem: Update visibility; If client C1 buffers writes in its cache, client C2 will see the old version NFSv2 solution: Close-to-open consistency – Clients flush on close(), so other clients will see the latest version on open() Problem: Stale cache; If C1 has a file cached, it will see old data even if the file is updated by C2 NFSv2 solution: Send a GETATTR and check the file's modification time to see if it has been updated. Cache attributes for a few seconds to reduce the number of GETATTR calls. 21 Strong Semantics for Write
Because the NFS server is stateless, when servicing an NFS request it must commit any modified data to stable storage before returning results The implication for UNIX based servers is that requests which modify the file system must flush all modified data & metadata to disk before returning from the call This can be a big performance bottleneck unless something is done to improve write performance (e.g., NetApp's WAFL file system)
22 Outline
The Basics NFSv2 NFSv3 NFSv4 NFSv4.1
23 Major Changes from NFSv2 to v3
Sizes and offsets are widened from 32 bits to 64 bits A new COMMIT RPC allows for reliable asynchronous writes A new ACCESS RPC improves support for ACLs and super-user All operations now return attributes to reduce the number of subsequent GETATTR procedure calls The 8KB data size limitation on the READ and WRITE procedures is relaxed
24 Major Changes from NFSv2 to v3
A new READDIRPLUS RPC returns both file handle and attributes to eliminate LOOKUP calls when scanning a directory
25 Asynchronous Writes
In NFSv3, the server can reply to WRITE RPCs immediately, without syncing to disk When the client wants to ensure that data is on stable storage, it sends a COMMIT RPC Asynchronous writes are optional, and negotiated at mount time
26 Asynchronous Writes: Crash Recovery
The client must keep all uncommitted data in case of a server crash Replies for WRITE and COMMIT RPCs include a write verifier for server crash detection Write verifier: 8-byte value that the server must change if it crashes (many use boot time) The client must save verifiers returned by async WRITE RPCs, and compare them to the verifier returned by a leter COMMIT RPC If the verifiers don't match, the client must rewrite all uncommitted data
27 Outline
The Basics NFSv2 NFSv3 NFSv4 NFSv4.1
28 Additional Goals for NFSv4
Improved access and good performance on the Internet Only TCP Easy to transit firewalls: uses one port (mount & lock protocols merged into NFS) COMPOUNDs, delegations, uid/gid issue resolved Strong security with negotiation built in Better cross-platform interoperability Better extensibility New security types, new attributes, etc.
Big design change – NFSv4 is stateful 29 Security
For previous versions, only UNIX permissions were widely adopted NFSv4 mandates the use of strong RPC security flavors that depend on cryptography Security type negotiation is done securely and in-band User and groups are identified with strings, not numbers Access control policies compatible with both UNIX and Windows The problematic MOUNT protocol is removed
30 RPCSEC_GSS
A framework adopted by NFSv4 to provide authentication, integrity, and privacy at the RPC level The following mechanisms must be implemented: Kerberos v5, LIPKEY, SPKM3 Security options are negotiated at mount time The SECINFO operation allows a client to determine the security policy (usually on mount, but can be on a per-filehandle basis) RPCSEC_GSS can be used with previous versions of NFS, but in NFSv4 support is
mandatory 31 Identifying Users
In v2 and v3, users and groups were represented as integers This required all clients and the server to agree on user and group assignments - not practical (especially over the Internet) NFSv4 uses strings ‘user@domain’ and ‘group@domain’, where domain represents a registered DNS domain or a sub-domain On Linux, idmapd translates NFSv4 IDs
32 COMPOUND Procedure
NFSv4 has 2 procedures: NULL and COMPOUND The COMPOUND procedure can contain several operations (similar to previous NFS procedures) Possible example: {LOOKUP, OPEN, READ} Operations are evaluated in order, and each can have a return value If an operation fails, the server stops evaluating the COMPOUND and returns
33 Filehandles
Current filehandle: used by most operations Saved filehandle: used as an additional operand
Example from Linux #1: WRITE request PUTFH(fh): set CURFH to the target file WRITE: write the data to the current file GETATTR: get attributes for the current file
34 Filehandles
Example from Linux #2: CREATE request PUTFH(dirfh): set CURFH to the directory SAVEFH: save CURFH (SAVEDFH=CURFH) CREATE: create the file (CURFH=NEWFH) GETFH: return CURFH to the client GETATTR: get the attributes of the new file RESTOREFH: (CURFH=SAVEDFH) GETATTR: get the attributes of the directory
35 Some Differences in Operations
CREATE now creates file, directories, and special files LOOKUPP was introduced to look up the parent directory – no special meaning for ‘.’ and ‘..’ as in previous NFS versions (better cross-platform interoperability) READDIRPLUS removed - READDIR now returns requested attributes
36 Filehandle Types
In previous NFS versions, the fhandle was valid for the lifetime of the file system object Now these fhandles are called “persistent filehandles” “Volatile filehandles” may become invalid, but the client is prepared to deal with these semantics
37 File System Migration/Replication
Migration The file system locations attribute provides a method for the client to probe the server about the location of a file system In the event of a file system migration, the client will receive an error when operating on the file system and it can then query as to the new location Replication The client is able to query the server for the multiple available locations of a particular file system From this information, the client can use its own policies to access the appropriate file system location 38 Attribute Types
Mandatory: minimal set of file or file system attributes that must be provided by the server type, filehandle expiration type, change indicator, size, fsid, lease duration, etc. Recommended: represent different file system types and operating environments case insensitive, hidden, max file size, max read size, max write size, UNIX mode bits, owner string, group string, modify/create/access time, etc. Named: Similar to extended attributes, implemented as hidden directories ACLs: implemented as recommended attribute 39 Pseudo Filesystems
In NFSv4, the server presents a single seamless view of all the exported file systems to a client The client can use the fsid to notice changes mount -t nfs4 servername:/ /mnt/dir
40 Client Caching
File, attribute, and directory caching is similar to previous versions: clients determine what to cache and for how long, and when to see if an update occurred Close-to-open consistency Client checks if cached data is valid on OPEN Client writes modified data on CLOSE Sufficient for most applications and users
41 Leases
A lease is a time-bounded grant of control of the state of a file, from the server to the client (lock or delegation) During a lease interval a server may not grant conflicting control to another client The client may assume that a lock granted by the server will remain valid for a fixed (server-specified) interval and is subject to renewal by the client The client is responsible for refreshing the lease If the lease interval expires without a refresh from the client, the server assumes the client has failed and may allow other clients to acquire the same lock If the server fails, on reboot the server waits a duration equal to a lease interval for clients to reclaim the locks that they may still hold, before allowing any new lock requests 42 File Locking
Support for byte-range file locking part of protocol Lease-based model: lease state is stored on the server Clients must either explicitly renew leases (RENEW), or implicitly renew them (usually READ) A refresh of any lock by the client validates all locks held by the client to a particular server (reduces the number of refreshes)
43 Delegations
The server may grant a read or write delegation for a file to a client Read delegation: client is assured that no other client will write to the file for the duration of the delegation Write delegation: like read delegation, but other clients may not read or write Delegations may be recalled using a callback A callback is a server → client RPC A client must support callbacks in order to get a delegation - tested with CB_NULL request Delegations allow clients to service operations like OPEN, CLOSE, LOCK, READ, WRITE
without immediate interaction with the server 44 Outline
The Basics NFSv2 NFSv3 NFSv4 NFSv4.1
45 Parallel NFS (pNFS)
Clients may now access storage devices directly and in parallel Eliminates the classic NFS bottleneck of having only one server The management protocol is NFSv4.1 The data protocol can be NFSv4.1, OSD, or FC
46 Other NFSv4.1 Highlights
Sessions Session layer on top of the transport layer Solves many issues with dropped connections, as well as client and server crashes
Delegation support for directories
47 References
http://pages.cs.wisc.edu/~cs537-1/notes/34_file-nfs.pdf
RFC1094 - NFS version 2 RFC1813 - NFS version 3 RFC1831 - RPC: Remote Procedure Call Protocol Specification Version 2 RFC1832 - XDR: External Data Representation Standard RFC1964 - The Kerberos Version 5 GSS-API Mechanism RFC2025 - The Simple Public-Key GSS-API Mechanism (SPKM) RFC2054 - WebNFS Client Specification RFC2055 - WebNFS Server Specification RFC2203 - RPCSEC_GSS Protocol Specification RFC2224 - NFS URL Scheme RFC2581 - TCP Congestion Control RFC2623 - NFS Version 2 and Version 3 Security Issues and the NFS Protocol's Use of RPCSEC_GSS and Kerberos V5 RFC2624 - NFS Version 4 Design Considerations RFC2224 - Security Negotiation for WebNFS RFC2743 - Generic Security Service Application Program Interface, Version 2, Update 1 RFC2847 - LIPKEY - A Low Infrastructure Public Key Mechanism Using SPKM RFC3010 - NFS version 4 Protocol (Obsoleted by RFC3530) RFC3530 - NFS version 4 Protocol RFC5661 - NFS version 4 Minor Version 1 Protocol
48