1 ===========
2 NFS LOCALIO
3 ===========
4
5 Overview
6 ========
7
8 The LOCALIO auxiliary RPC protocol allows the Linux NFS client and
9 server to reliably handshake to determine if they are on the same
10 host. Select "NFS client and server support for LOCALIO auxiliary
11 protocol" in menuconfig to enable CONFIG_NFS_LOCALIO in the kernel
12 config (both CONFIG_NFS_FS and CONFIG_NFSD must also be enabled).
13
14 Once an NFS client and server handshake as "local", the client will
15 bypass the network RPC protocol for read, write and commit operations.
16 Due to this XDR and RPC bypass, these operations will operate faster.
17
18 The LOCALIO auxiliary protocol's implementation, which uses the same
19 connection as NFS traffic, follows the pattern established by the NFS
20 ACL protocol extension.
21
22 The LOCALIO auxiliary protocol is needed to allow robust discovery of
23 clients local to their servers. In a private implementation that
24 preceded use of this LOCALIO protocol, a fragile sockaddr network
25 address based match against all local network interfaces was attempted.
26 But unlike the LOCALIO protocol, the sockaddr-based matching didn't
27 handle use of iptables or containers.
28
29 The robust handshake between local client and server is just the
30 beginning, the ultimate use case this locality makes possible is the
31 client is able to open files and issue reads, writes and commits
32 directly to the server without having to go over the network. The
33 requirement is to perform these loopback NFS operations as efficiently
34 as possible, this is particularly useful for container use cases
35 (e.g. kubernetes) where it is possible to run an IO job local to the
36 server.
37
38 The performance advantage realized from LOCALIO's ability to bypass
39 using XDR and RPC for reads, writes and commits can be extreme, e.g.:
40
41 fio for 20 secs with directio, qd of 8, 16 libaio threads:
42 - With LOCALIO:
43 4K read: IOPS=979k, BW=3825MiB/s (4011MB/s)(74.7GiB/20002msec)
44 4K write: IOPS=165k, BW=646MiB/s (678MB/s)(12.6GiB/20002msec)
45 128K read: IOPS=402k, BW=49.1GiB/s (52.7GB/s)(982GiB/20002msec)
46 128K write: IOPS=11.5k, BW=1433MiB/s (1503MB/s)(28.0GiB/20004msec)
47
48 - Without LOCALIO:
49 4K read: IOPS=79.2k, BW=309MiB/s (324MB/s)(6188MiB/20003msec)
50 4K write: IOPS=59.8k, BW=234MiB/s (245MB/s)(4671MiB/20002msec)
51 128K read: IOPS=33.9k, BW=4234MiB/s (4440MB/s)(82.7GiB/20004msec)
52 128K write: IOPS=11.5k, BW=1434MiB/s (1504MB/s)(28.0GiB/20011msec)
53
54 fio for 20 secs with directio, qd of 8, 1 libaio thread:
55 - With LOCALIO:
56 4K read: IOPS=230k, BW=898MiB/s (941MB/s)(17.5GiB/20001msec)
57 4K write: IOPS=22.6k, BW=88.3MiB/s (92.6MB/s)(1766MiB/20001msec)
58 128K read: IOPS=38.8k, BW=4855MiB/s (5091MB/s)(94.8GiB/20001msec)
59 128K write: IOPS=11.4k, BW=1428MiB/s (1497MB/s)(27.9GiB/20001msec)
60
61 - Without LOCALIO:
62 4K read: IOPS=77.1k, BW=301MiB/s (316MB/s)(6022MiB/20001msec)
63 4K write: IOPS=32.8k, BW=128MiB/s (135MB/s)(2566MiB/20001msec)
64 128K read: IOPS=24.4k, BW=3050MiB/s (3198MB/s)(59.6GiB/20001msec)
65 128K write: IOPS=11.4k, BW=1430MiB/s (1500MB/s)(27.9GiB/20001msec)
66
67 FAQ
68 ===
69
70 1. What are the use cases for LOCALIO?
71
72 a. Workloads where the NFS client and server are on the same host
73 realize improved IO performance. In particular, it is common when
74 running containerised workloads for jobs to find themselves
75 running on the same host as the knfsd server being used for
76 storage.
77
78 2. What are the requirements for LOCALIO?
79
80 a. Bypass use of the network RPC protocol as much as possible. This
81 includes bypassing XDR and RPC for open, read, write and commit
82 operations.
83 b. Allow client and server to autonomously discover if they are
84 running local to each other without making any assumptions about
85 the local network topology.
86 c. Support the use of containers by being compatible with relevant
87 namespaces (e.g. network, user, mount).
88 d. Support all versions of NFS. NFSv3 is of particular importance
89 because it has wide enterprise usage and pNFS flexfiles makes use
90 of it for the data path.
91
92 3. Why doesn’t LOCALIO just compare IP addresses or hostnames when
93 deciding if the NFS client and server are co-located on the same
94 host?
95
96 Since one of the main use cases is containerised workloads, we cannot
97 assume that IP addresses will be shared between the client and
98 server. This sets up a requirement for a handshake protocol that
99 needs to go over the same connection as the NFS traffic in order to
100 identify that the client and the server really are running on the
101 same host. The handshake uses a secret that is sent over the wire,
102 and can be verified by both parties by comparing with a value stored
103 in shared kernel memory if they are truly co-located.
104
105 4. Does LOCALIO improve pNFS flexfiles?
106
107 Yes, LOCALIO complements pNFS flexfiles by allowing it to take
108 advantage of NFS client and server locality. Policy that initiates
109 client IO as closely to the server where the data is stored naturally
110 benefits from the data path optimization LOCALIO provides.
111
112 5. Why not develop a new pNFS layout to enable LOCALIO?
113
114 A new pNFS layout could be developed, but doing so would put the
115 onus on the server to somehow discover that the client is co-located
116 when deciding to hand out the layout.
117 There is value in a simpler approach (as provided by LOCALIO) that
118 allows the NFS client to negotiate and leverage locality without
119 requiring more elaborate modeling and discovery of such locality in a
120 more centralized manner.
121
122 6. Why is having the client perform a server-side file OPEN, without
123 using RPC, beneficial? Is the benefit pNFS specific?
124
125 Avoiding the use of XDR and RPC for file opens is beneficial to
126 performance regardless of whether pNFS is used. Especially when
127 dealing with small files its best to avoid going over the wire
128 whenever possible, otherwise it could reduce or even negate the
129 benefits of avoiding the wire for doing the small file I/O itself.
130 Given LOCALIO's requirements the current approach of having the
131 client perform a server-side file open, without using RPC, is ideal.
132 If in the future requirements change then we can adapt accordingly.
133
134 7. Why is LOCALIO only supported with UNIX Authentication (AUTH_UNIX)?
135
136 Strong authentication is usually tied to the connection itself. It
137 works by establishing a context that is cached by the server, and
138 that acts as the key for discovering the authorisation token, which
139 can then be passed to rpc.mountd to complete the authentication
140 process. On the other hand, in the case of AUTH_UNIX, the credential
141 that was passed over the wire is used directly as the key in the
142 upcall to rpc.mountd. This simplifies the authentication process, and
143 so makes AUTH_UNIX easier to support.
144
145 8. How do export options that translate RPC user IDs behave for LOCALIO
146 operations (eg. root_squash, all_squash)?
147
148 Export options that translate user IDs are managed by nfsd_setuser()
149 which is called by nfsd_setuser_and_check_port() which is called by
150 __fh_verify(). So they get handled exactly the same way for LOCALIO
151 as they do for non-LOCALIO.
152
153 9. How does LOCALIO make certain that object lifetimes are managed
154 properly given NFSD and NFS operate in different contexts?
155
156 See the detailed "NFS Client and Server Interlock" section below.
157
158 RPC
159 ===
160
161 The LOCALIO auxiliary RPC protocol consists of a single "UUID_IS_LOCAL"
162 RPC method that allows the Linux NFS client to verify the local Linux
163 NFS server can see the nonce (single-use UUID) the client generated and
164 made available in nfs_common. This protocol isn't part of an IETF
165 standard, nor does it need to be considering it is Linux-to-Linux
166 auxiliary RPC protocol that amounts to an implementation detail.
167
168 The UUID_IS_LOCAL method encodes the client generated uuid_t in terms of
169 the fixed UUID_SIZE (16 bytes). The fixed size opaque encode and decode
170 XDR methods are used instead of the less efficient variable sized
171 methods.
172
173 The RPC program number for the NFS_LOCALIO_PROGRAM is 400122 (as assigned
174 by IANA, see https://www.iana.org/assignments/rpc-program-numbers/ ):
175 Linux Kernel Organization 400122 nfslocalio
176
177 The LOCALIO protocol spec in rpcgen syntax is::
178
179 /* raw RFC 9562 UUID */
180 #define UUID_SIZE 16
181 typedef u8 uuid_t<UUID_SIZE>;
182
183 program NFS_LOCALIO_PROGRAM {
184 version LOCALIO_V1 {
185 void
186 NULL(void) = 0;
187
188 void
189 UUID_IS_LOCAL(uuid_t) = 1;
190 } = 1;
191 } = 400122;
192
193 LOCALIO uses the same transport connection as NFS traffic. As such,
194 LOCALIO is not registered with rpcbind.
195
196 NFS Common and Client/Server Handshake
197 ======================================
198
199 fs/nfs_common/nfslocalio.c provides interfaces that enable an NFS client
200 to generate a nonce (single-use UUID) and associated short-lived
201 nfs_uuid_t struct, register it with nfs_common for subsequent lookup and
202 verification by the NFS server and if matched the NFS server populates
203 members in the nfs_uuid_t struct. The NFS client then uses nfs_common to
204 transfer the nfs_uuid_t from its nfs_uuids to the nn->nfsd_serv
205 clients_list from the nfs_common's uuids_list. See:
206 fs/nfs/localio.c:nfs_local_probe()
207
208 nfs_common's nfs_uuids list is the basis for LOCALIO enablement, as such
209 it has members that point to nfsd memory for direct use by the client
210 (e.g. 'net' is the server's network namespace, through it the client can
211 access nn->nfsd_serv with proper rcu read access). It is this client
212 and server synchronization that enables advanced usage and lifetime of
213 objects to span from the host kernel's nfsd to per-container knfsd
214 instances that are connected to nfs client's running on the same local
215 host.
216
217 NFS Client and Server Interlock
218 ===============================
219
220 LOCALIO provides the nfs_uuid_t object and associated interfaces to
221 allow proper network namespace (net-ns) and NFSD object refcounting:
222
223 We don't want to keep a long-term counted reference on each NFSD's
224 net-ns in the client because that prevents a server container from
225 completely shutting down.
226
227 So we avoid taking a reference at all and rely on the per-cpu
228 reference to the server (detailed below) being sufficient to keep
229 the net-ns active. This involves allowing the NFSD's net-ns exit
230 code to iterate all active clients and clear their ->net pointers
231 (which are needed to find the per-cpu-refcount for the nfsd_serv).
232
233 Details:
234
235 - Embed nfs_uuid_t in nfs_client. nfs_uuid_t provides a list_head
236 that can be used to find the client. It does add the 16-byte
237 uuid_t to nfs_client so it is bigger than needed (given that
238 uuid_t is only used during the initial NFS client and server
239 LOCALIO handshake to determine if they are local to each other).
240 If that is really a problem we can find a fix.
241
242 - When the nfs server confirms that the uuid_t is local, it moves
243 the nfs_uuid_t onto a per-net-ns list in NFSD's nfsd_net.
244
245 - When each server's net-ns is shutting down - in a "pre_exit"
246 handler, all these nfs_uuid_t have their ->net cleared. There is
247 an rcu_synchronize() call between pre_exit() handlers and exit()
248 handlers so any caller that sees nfs_uuid_t ->net as not NULL can
249 safely manage the per-cpu-refcount for nfsd_serv.
250
251 - The client's nfs_uuid_t is passed to nfsd_open_local_fh() so it
252 can safely dereference ->net in a private rcu_read_lock() section
253 to allow safe access to the associated nfsd_net and nfsd_serv.
254
255 So LOCALIO required the introduction and use of NFSD's percpu_ref to
256 interlock nfsd_destroy_serv() and nfsd_open_local_fh(), to ensure each
257 nn->nfsd_serv is not destroyed while in use by nfsd_open_local_fh(), and
258 warrants a more detailed explanation:
259
260 nfsd_open_local_fh() uses nfsd_serv_try_get() before opening its
261 nfsd_file handle and then the caller (NFS client) must drop the
262 reference for the nfsd_file and associated nn->nfsd_serv using
263 nfs_file_put_local() once it has completed its IO.
264
265 This interlock working relies heavily on nfsd_open_local_fh() being
266 afforded the ability to safely deal with the possibility that the
267 NFSD's net-ns (and nfsd_net by association) may have been destroyed
268 by nfsd_destroy_serv() via nfsd_shutdown_net() -- which is only
269 possible given the nfs_uuid_t ->net pointer managemenet detailed
270 above.
271
272 All told, this elaborate interlock of the NFS client and server has been
273 verified to fix an easy to hit crash that would occur if an NFSD
274 instance running in a container, with a LOCALIO client mounted, is
275 shutdown. Upon restart of the container and associated NFSD the client
276 would go on to crash due to NULL pointer dereference that occurred due
277 to the LOCALIO client's attempting to nfsd_open_local_fh(), using
278 nn->nfsd_serv, without having a proper reference on nn->nfsd_serv.
279
280 NFS Client issues IO instead of Server
281 ======================================
282
283 Because LOCALIO is focused on protocol bypass to achieve improved IO
284 performance, alternatives to the traditional NFS wire protocol (SUNRPC
285 with XDR) must be provided to access the backing filesystem.
286
287 See fs/nfs/localio.c:nfs_local_open_fh() and
288 fs/nfsd/localio.c:nfsd_open_local_fh() for the interface that makes
289 focused use of select nfs server objects to allow a client local to a
290 server to open a file pointer without needing to go over the network.
291
292 The client's fs/nfs/localio.c:nfs_local_open_fh() will call into the
293 server's fs/nfsd/localio.c:nfsd_open_local_fh() and carefully access
294 both the associated nfsd network namespace and nn->nfsd_serv in terms of
295 RCU. If nfsd_open_local_fh() finds that the client no longer sees valid
296 nfsd objects (be it struct net or nn->nfsd_serv) it returns -ENXIO
297 to nfs_local_open_fh() and the client will try to reestablish the
298 LOCALIO resources needed by calling nfs_local_probe() again. This
299 recovery is needed if/when an nfsd instance running in a container were
300 to reboot while a LOCALIO client is connected to it.
301
302 Once the client has an open nfsd_file pointer it will issue reads,
303 writes and commits directly to the underlying local filesystem (normally
304 done by the nfs server). As such, for these operations, the NFS client
305 is issuing IO to the underlying local filesystem that it is sharing with
306 the NFS server. See: fs/nfs/localio.c:nfs_local_doio() and
307 fs/nfs/localio.c:nfs_local_commit().
308
309 Security
310 ========
311
312 Localio is only supported when UNIX-style authentication (AUTH_UNIX, aka
313 AUTH_SYS) is used.
314
315 Care is taken to ensure the same NFS security mechanisms are used
316 (authentication, etc) regardless of whether LOCALIO or regular NFS
317 access is used. The auth_domain established as part of the traditional
318 NFS client access to the NFS server is also used for LOCALIO.
319
320 Relative to containers, LOCALIO gives the client access to the network
321 namespace the server has. This is required to allow the client to access
322 the server's per-namespace nfsd_net struct. With traditional NFS, the
323 client is afforded this same level of access (albeit in terms of the NFS
324 protocol via SUNRPC). No other namespaces (user, mount, etc) have been
325 altered or purposely extended from the server to the client.
326
327 Testing
328 =======
329
330 The LOCALIO auxiliary protocol and associated NFS LOCALIO read, write
331 and commit access have proven stable against various test scenarios:
332
333 - Client and server both on the same host.
334
335 - All permutations of client and server support enablement for both
336 local and remote client and server.
337
338 - Testing against NFS storage products that don't support the LOCALIO
339 protocol was also performed.
340
341 - Client on host, server within a container (for both v3 and v4.2).
342 The container testing was in terms of podman managed containers and
343 includes successful container stop/restart scenario.
344
345 - Formalizing these test scenarios in terms of existing test
346 infrastructure is on-going. Initial regular coverage is provided in
347 terms of ktest running xfstests against a LOCALIO-enabled NFS loopback
348 mount configuration, and includes lockdep and KASAN coverage, see:
349 https://evilpiepirate.org/~testdashboard/ci?user=snitzer&branch=snitm-nfs-next
350 https://github.com/koverstreet/ktest
351
352 - Various kdevops testing (in terms of "Chuck's BuildBot") has been
353 performed to regularly verify the LOCALIO changes haven't caused any
354 regressions to non-LOCALIO NFS use cases.
355
356 - All of Hammerspace's various sanity tests pass with LOCALIO enabled
357 (this includes numerous pNFS and flexfiles tests).
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.