commit 79770ef91411cc2d3f97eb5efdd43a7d39b0070e Author: Benjamin Coddington Date: Wed Nov 18 07:24:19 2026 -0500 nfsdctl: bound the rpcbind calls to one 1s attempt libtirpc's rpcb_set() and rpcb_unset() create a client for every call and wait 60s for each reply. A local rpcbind that accepts the connection and never answers stalls nfsdctl for a minute per call: five UNSETs when the threads stop, and five UNSETs plus one SET per listener netid and program when they start. libtirpc's setter for that timeout is private. Make the SET and UNSET calls on a client of nfsdctl's own, built with nfs_get_rpcclient(): AF_LOCAL on _PATH_RPCBINDSOCK first, then TCP to 127.0.0.1 port 111, as libtirpc's local_rpcb() orders them. Each pass uses one client, and each call is one attempt of 1s, matching 17ceb2fecc7c ("SUNRPC: bound the local rpcbind client timeout to 1s"): "The local rpcbind is on loopback or on an AF_LOCAL socket, and it answers in microseconds." A call that gets no answer ends the pass with one warning, as ba8c47f04942 and df1ff8e4e275 do for the kernel's svc_register() and svc_unregister() sweep. A registration pass whose opening UNSET sweep gets no answer makes no SET. A reply of FALSE is logged and the pass continues, as before. nfs_svc_unregister() is unchanged, and a build without TI-RPC keeps using it. With the local rpcbind SIGSTOPped, on a kernel that leaves the nfsd registration to nfsdctl, "nfsdctl threads 0" returned after 3.4s, where it took 302.7s before. lockd's own registration in the kernel accounts for 2s of that. "nfsdctl autostart" now ends its registration pass after 1s. The THREADS_SET that follows still fails with EIO, because lockd cannot register. Reviewed-by: Jeff Layton Signed-off-by: Benjamin Coddington Signed-off-by: Steve Dickson commit c1f521c4f7e198c5b68dc34e1b3b96c7c332a8af Author: Giulio Benetti Date: Sun Nov 15 14:29:06 2026 -0500 gssd: fix build failure with musl libc gssd prints thread IDs with the '%lx' format specifier, which assumes pthread_t is an unsigned long. This holds on glibc, but on musl pthread_t is an opaque pointer type ('struct __pthread *'), so building with -Werror=format fails: gssd.c:546:78: error: format '%lx' expects argument of type 'long unsigned int', but argument 3 has type 'pthread_t' {aka 'struct __pthread *'} [-Werror=format=] Fix this by explicitly casting pthread_t to unsigned long wherever it is printed with '%lx' in gssd.c, gssd_proc.c and krb5_util.c. The cast is safe since the value is only used for logging and unsigned long is pointer-sized on all Linux ABIs. The output is unchanged on glibc. Co-Authored-By: Claude Opus 5.5 Signed-off-by: Giulio Benetti Signed-off-by: Steve Dickson commit b4df4197ed7bc008e3910a36832efbce3514e13d Author: Minsu Kim Date: Sun Nov 15 14:09:47 2026 -0500 rpc.svcgssd: stop using raw pointers as continuation handles rpc.svcgssd currently serializes gss_ctx_id_t itself as the continuation handle when gss_accept_sec_context() returns GSS_S_CONTINUE_NEEDED. A client then sends that value back in a CONTINUE_INIT request, where the daemon converts it directly into a live GSS context. This exposes a daemon address and trusts a client-controlled value as a live GSS context pointer. An unauthenticated NFS client can replace or replay the handle before rpc.svcgssd passes it to the GSS implementation. Keep pending contexts in a daemon-owned search tree. Return a random 64-bit ID from /dev/urandom instead of the context pointer. Take the matching entry before processing a continuation, return the same ID if negotiation needs another step, and reject unknown IDs with GSS_S_NO_CONTEXT. Discard the saved context if writing its reply fails. The behavior was introduced in commit 62c4197d5b41 and first shipped in nfs-utils 1.0.8-rc1. It remains present in nfs-utils 3.1.1. rpc.svcgssd is built only when --enable-svcgss is selected, which is disabled by default. On an unpatched x86_64 build, an INIT request that required another GSS step returned an 8-byte value matching the in-process gss_ctx_id_t address. Modified and stale continuation handles caused invalid memory accesses and daemon termination. With this patch, an unknown ID and a consumed ID both returned GSS_S_NO_CONTEXT, initial negotiations returned opaque 8-byte IDs, and 63 outstanding negotiations produced 63 distinct IDs. I also completed an NFSv4.2 sec=krb5 mount with the patched daemon, which remained active throughout the tests. The reproducer and complete logs are available privately on request. The vulnerability research and validation were conducted by members of the Tobabz team as part of the Best of the Best 15th program. Fixes: 62c4197d5b41 ("Updates from Kevin Coffman at UMich") Assisted-by: LLM Signed-off-by: Minsu Kim Signed-off-by: Steve Dickson commit 53db916e6351b675a4915f2e83a9756479436210 Author: HeonJe Lee Date: Sun Nov 15 13:59:36 2026 -0500 nfsdcld: signal readiness only after the pipefs watch is installed nfsdcld daemonizes with daemon(0, 0) and nfsdcld.service is Type=forking, so systemd considers the service started when the daemonizing parent exits. The child, however, only installs its inotify watch on the rpc_pipefs directory much later, after the sqlite database is prepared and cld_pipe_init() runs. Until the watch is installed, nfsdcld cannot notice that the kernel has (re)created the cld pipe. nfs-server.service orders itself After=nfsdcld.service, but that ordering is meaningless with Type=forking: on a busy boot or a restart storm, systemd starts nfsd while nfsdcld is still doing schema migration or legacy record import, and the kernel's one second wait for nfsdcld to open the pipe (see nfs4recover.c:nfsd4_cld_tracking_init()) can expire unused. An observed instance of this race took an NFS server down for 44 minutes when nothing restarted the failed nfsd. Fix by converting nfsdcld to the daemon_init()/daemon_ready() helpers used by the other nfs-utils daemons: the parent no longer exits until the child signals, and the child signals right after cld_pipe_init() has installed the inotify watch. With that, the existing After=nfsdcld.service ordering means what it claims, and the kernel's one second window is sufficient. The event base is now created after daemonizing: daemon_init() renumbers fds (its status pipe takes fd 3) and would otherwise clobber the epoll descriptor created before daemonizing. The event base allocation failure is now reported through xlog() instead of fprintf(stderr): stderr is /dev/null after daemonizing, so an fprintf() message would be lost, while xlog() routes it to the syslog. Suggested-by: Chuck Lever Link: https://lore.kernel.org/linux-nfs/20261015114755.33331-1-lhjnano@gmail.com/ Signed-off-by: HeonJe Lee Signed-off-by: Steve Dickson commit 3f3625e8784db10d97f6d9edfbe414707cdb2908 Author: Oleksandr Kulkov Date: Sun Nov 15 13:54:34 2026 -0500 mountd: limit caching after managed-group passwd lookup errors With --manage-gids, a getpwuid() error produces an empty supplementary group list cached for the normal TTL. Access that needs those groups can remain denied after the directory service recovers. Clear errno before the lookup and cap the reply lifetime at five seconds when it reports an error. Preserve the usual not-found errno values, successful lookups, and configured TTLs shorter than five seconds. Apply the same policy to procfs and netlink, including netlink batch retries. Keep the existing empty-group reply rather than leave an upcall pending. This does not deny all access: UID, primary-group and other permissions still apply. Failures hidden by NSS and getgrouplist() error handling are outside this change. Assisted-by: LLM Signed-off-by: Oleksandr Kulkov Signed-off-by: Steve Dickson commit 3282da1d50c43f14fb765584fd7d1cf649af300e Author: Jeff Layton Date: Fri Nov 13 16:32:59 2026 -0500 nfsd-iomode-advisor: recommend read_io= and write_io= for an export Admins cannot easily tell which NFSD IO mode suits an export. Add a tool that samples metrics while clients use the export, and recommends buffered, dontcache, or direct for READ and WRITE, with reasons. Metrics: - miss ratio: export_stats io_read_storage / io_read. Without io_read_storage: block device read bytes / io_read of its exports. - reclaim vs. the export's storage traffic, direct reclaim share, refaults, memory PSI. The tool subtracts cgroups at memory.high or memory.max (not system memory pressure). - CPU time of the busiest kswapd or kcompactd thread - Dirty and Writeback vs. the background writeback threshold, sampled each second, from /sys/kernel/debug/bdi or /proc/meminfo - free memory vs. zone high watermarks, to find a page cache that is not full yet - average READ (incl. READ_PLUS) and WRITE size from /proc/net/rpc/nfsd - optional cachestat(2) of the export's largest files Thresholds are ratios, or relative to export traffic, kernel limits, or storage, so they do not change with server size. - large-memory servers with slow storage never reclaim a large fraction per minute - kswapd CPU sums scale with NUMA nodes - stall counts scale with threads - large servers often set vm.dirty_background_bytes Mode limits: - No mode that NFSD silently falls back from. The tool probes with preadv2(RWF_DONTCACHE) and statx(STATX_DIOALIGN). - No write_io=direct when the block device needs more than the 4-byte memory alignment of TCP WRITE data (e.g. device mapper, MD). - No direct IO on rotational storage. - Buffered IO for WRITEs smaller than optimal_io_size (e.g. parity RAID stripe width). Caveats: - Miss ratio is valid only with buffered IO. Otherwise, or if the mode is not readable, the tool reports "unknown". - io_read_storage counts only reads that NFSD threads submit. The tool ignores it on filesystems that read from their own threads (FUSE, ZFS), and flags it when the block device reads much more. - export_stats counters persist across cache updates. The tool skips entries that appear with older counters or leave the cache during the interval, and flags the result incomplete. - export_stats lists an export only after first client use. The tool also accepts an etab export and reads read_io=/write_io= from etab. Signed-off-by: Jeff Layton Signed-off-by: Steve Dickson commit 65c06312b9008bcba8391327666ee5d122bc5e31 Author: Jeff Layton Date: Fri Nov 13 16:26:37 2026 -0500 exportfs: Display and document read_io= and write_io= Show explicitly-set I/O modes in "exportfs -v" output and document the new options in exports(5). Warn when exporting if the running kernel, or no-netlink, means they will be ignored. Signed-off-by: Jeff Layton Signed-off-by: Steve Dickson commit a3da58d3c6ad3649f16331b36d4a27166a0485a1 Author: Jeff Layton Date: Fri Nov 13 16:18:06 2026 -0500 exports: Add read_io= and write_io= export options Parse read_io=/write_io= (buffered/dontcache/direct) into new e_read_io/e_write_io exportent fields, defaulting to unset, and pass explicitly-set values to the kernel over the netlink svc_export interface. - Kernels without the I/O mode attributes (e.g. v7.2) reject the whole entry. Add nfsd_export_iomode_supported(), which checks the svc_export policy's highest attribute via nfsd_export_policy_query(), and have mountd drop the options with a warning on such kernels instead of leaving the export unmountable. - The /proc channel cannot carry them either; warn there too. - Resync nfsd_netlink.h and bump the configure probe to NFSD_A_SVC_EXPORT_WRITE_IO. Signed-off-by: Jeff Layton Signed-off-by: Steve Dickson commit 1a73e4390ffaa37ee08c8cd436d0674df0b7d4fb Author: Jeff Layton Date: Fri Nov 13 16:09:28 2026 -0500 exportfs/mountd/exportd: track the svc_export policy highest attribute Record max_attr, the highest attribute in the kernel svc_export netlink policy, in struct nfsd_export_policy. Follow-on work needs it to avoid sending attributes that an older kernel would reject the whole entry over. Signed-off-by: Jeff Layton Signed-off-by: Steve Dickson commit ba96fb357f748e8f301ca6a44f16ce5198bab284 Author: Scott Mayhew Date: Fri Nov 13 15:32:45 2026 -0500 sm-notify: use a separate socket per address family Previously sm-notify sent all of its RPC traffic over a single AF_INET6 socket with IPV6_V6ONLY disabled, relying on the network stack to reach IPv4 peers via IPv4-mapped IPv6 addresses. That approach is fragile: - It depends on IPv4-mapped IPv6 addresses being usable, which is not true on all configurations, and requires getaddrinfo(3) to be asked for mapped addresses (AI_V4MAPPED) or the IPv4 results are unsendable. - A socket bound to a specific source address (via -v / outgoing-addr) takes on that address's family, so an IPv6 source address cannot send to IPv4 peers (and vice versa), failing with ENETUNREACH. Create one datagram socket per address family instead, each strictly single-family (IPV6_V6ONLY is now enabled on the IPv6 socket), and send each host's notification on the socket matching the destination's family. Peer names are resolved with native address families (no AI_V4MAPPED), and IPv4 peers once again use PMAP_GETPORT, which is also understood by legacy portmappers. When a specific source address is requested, each family's socket is bound to the matching address from that name; a name that resolves to both families can now notify peers of both families from a single invocation. A host that only resolves to a family for which no socket is available is skipped with a warning, leaving its on-disk monitor record intact so a later run can retry. Assisted-by: LLM Signed-off-by: Scott Mayhew Signed-off-by: Steve Dickson commit 62ece1e9f935bf8d96d00d3bd0d35b02a4ed82ce Author: Scott Mayhew Date: Fri Nov 13 15:28:17 2026 -0500 sm-notify: fix multihomed address rotation and first-attempt address choice notify_host() rotates through a host's addresses after 4 retransmits so that all addresses of a multihomed peer are tried. Two defects prevented this from working for dual-stack hosts: 1. The "find the end of the list" loop started from &first->ai_next, which had just been set to NULL, so instead of appending the removed head to the tail of the (new) list it set first->ai_next = first, orphaning that address. For a 2-address host this collapsed the list to a single entry after the first rotation, after which the ai_next != NULL guard was never true again and no further rotation could occur. Walk from &host->ai so the removed head is re-appended. 2. smn_alloc_host() initialized host->retries to 100 to "force address retry" on the first notify_host() call. That made sense when the code transmitted to 'first' (the original head) after rotating, but since commit e132bb5d ("sm-notify: Replace RPC code") the transmit uses host->ai->ai_addr (the new head). As a result the very first packet was sent to the second address returned by getaddrinfo(3), skipping the preferred first address. Combined with defect #1, a host whose name resolved to both an IPv6 and an IPv4 address had its IPv6 address rotated out and orphaned before any packet was sent, pinning sm-notify to an unreachable IPv4 address. Initialize retries to 0; a freshly resolved addrinfo already has port 0, so the first call still performs an rpcbind lookup without forcing a rotation. Fixes: e132bb5d ("sm-notify: Replace RPC code") Assisted-by: LLM Signed-off-by: Scott Mayhew Signed-off-by: Steve Dickson commit bbb50905446f6b6a53baa708c2ac651ba28ba4d7 Author: Jeff Layton Date: Fri Nov 13 11:14:46 2026 -0500 Add MAINTAINERS file and get_maintainer.pl b4 prep --auto-to-cc looks for scripts/get_maintainer.pl at the top of the tree and runs it to collect To/Cc addresses. Add a kernel-style MAINTAINERS file plus a minimal get_maintainer.pl that implements the F/X/N/K match types, ship both in make dist, and cover the b4 flag sets with a new test. Reviewed-by: Benjamin Coddington Signed-off-by: Jeff Layton Signed-off-by: Steve Dickson commit 8ab8422086f607b5031dab88a3e46ead9cdedaef Author: Jeff Layton Date: Fri Nov 13 11:08:24 2026 -0500 mountd/exportd: always give pseudo-root directories a path-based uuid c5528f40f9db ("Fix NFSv4 export of tmpfs filesystems") made v4root_create() generate a uuid from the path only when export_test() says the directory cannot be exported without one, leaving others on the filesystem's own uuid. In practice export_test() fails nearly every time, so the path-based uuid is what pseudo-root directories get: - export_test() needs the "-test-client-" auth_domain, which only lives as long as an ip_map entry that references it. - With netlink, nothing creates that entry: exportfs's can_test() returns before writing it. The kernel answers ENOENT. - With /proc, can_test() creates it, but exportfs's own cache_flush() frees it before mountd rebuilds the pseudo-root. The fs uuid is only used when mountd rebuilds the pseudo-root during the brief window in an exportfs run on a live server, and then that directory's filehandle differs from every other time. Always use the path-based uuid below "/". That keeps what servers have done since c5528f40f9db, closes that window, and drops a write to /proc/net/rpc/nfsd.export/channel that ran even with netlink. Also keep uuid_s in scope until export_create() has used it; eep.e_uuid pointed at it after the block that declared it had ended. Fixes: c5528f40f9db ("Fix NFSv4 export of tmpfs filesystems") Reviewed-by: Benjamin Coddington Signed-off-by: Jeff Layton Signed-off-by: Steve Dickson commit 6a5b21e5462308d0540128a8abf5cecb7963e5d7 Author: Jeff Layton Date: Fri Nov 13 10:59:31 2026 -0500 nfsref: flush the export caches with cache_flush() junction_flush_exports_cache() writes the /proc/net/rpc/*/flush files itself. Call libexport's cache_flush() instead, which flushes over netlink and falls back to the same /proc files. - nfsref now links libexport.a and libnl, and defines no_netlink for cache_flush(). - Its callers ignore the FedFsStatus; it now always returns FEDFS_OK. Reviewed-by: Benjamin Coddington Signed-off-by: Jeff Layton Signed-off-by: Steve Dickson commit a431a45bc3b17da50f8854f85729444840631dbc Author: Jeff Layton Date: Fri Nov 13 10:35:38 2026 -0500 exportfs/mountd/exportd: get export features from the netlink policy get_export_features() reads /proc/fs/nfsd/export_features. The nfsd SET_REQS policy already says which export flags the kernel takes: NFSD_A_SVC_EXPORT_FLAGS is masked to exactly that set. - Add nfsd_export_policy_query() in a new export_policy.c (so cache_flush.o stays free of exports.o): a CTRL_CMD_GETPOLICY dump of SET_REQS, following the op policy -> REQUESTS -> svc_export nest. It fills struct nfsd_export_policy. A good answer is cached. - The policy cannot give the per-flavor flags (the auth-flavor FLAGS mask is the full set), so use NFSEXP_SECINFO_FLAGS; it is the same on every kernel with the netlink export interface. - Add set_export_features(), and nfsd_nl_load_export_features() to seed it from the policy. exportfs calls it at startup, and mountd/exportd once the nfsd netlink socket opens. - If netlink cannot answer, get_export_features() still reads /proc. Reviewed-by: Benjamin Coddington Signed-off-by: Jeff Layton Signed-off-by: Steve Dickson commit e4b74fcfe4bb8a669bd299183a1bdd132a26213b Author: Jeff Layton Date: Fri Nov 13 10:29:54 2026 -0500 exports: describe the kernel's export features with struct nfsd_export_policy Rename struct export_features, filled from /proc/fs/nfsd/export_features, to struct nfsd_export_policy, ahead of filling it from the nfsd netlink policy as well. No functional change. Reviewed-by: Benjamin Coddington Signed-off-by: Jeff Layton Signed-off-by: Steve Dickson commit f25f63c193f0108b9d5d9ca77cbf287d29a6e150 Author: Jeff Layton Date: Fri Nov 13 10:24:07 2026 -0500 mountd: prime the kernel caches over netlink on MOUNT cache_export() writes the ip_map and svc_export entries for a MNT request to the /proc channels even when mountd is using netlink. Any export state that only netlink can carry is then lost: the /proc write replaces the entry the netlink upcall filled. When the netlink command sockets are open, send both entries as unsolicited SET_REQS instead. The kernel's doit handlers look up or create the entry, so they do not need a pending request. - nl_add_ip_map() takes the expiry ttl, so the MOUNT path keeps using the export's e_ttl. - crossmnt submounts go over netlink too. - The /proc path is unchanged for no-netlink and fallback. Reviewed-by: Benjamin Coddington Signed-off-by: Jeff Layton Signed-off-by: Steve Dickson commit 9b47f25b54e7061ece864a7b86ba068235d19215 Author: Jeff Layton Date: Fri Nov 13 10:19:27 2026 -0500 mountd: accept -L for --no-netlink The -L short option was documented and handled but left out of shortopts, so "rpc.mountd -L" prints the usage and exits 1. Fixes: 60c383830537 ("mountd/exportd/exportfs: add --no-netlink option to disable netlink") Reviewed-by: Benjamin Coddington Signed-off-by: Jeff Layton Signed-off-by: Steve Dickson commit c2dc3ff598dec9786217d22e40eaca3f295e8b45 Author: Sven Püschel Date: Sat Nov 7 10:58:47 2026 -0500 configure: use AS_IF when PKG_CHECK_MODULES is used Use AS_IF when PKG_CHECK_MODULES is used to avoid it's required macros being expanded inside the test. Otherwise these required macros are not available to subsequent PKG_CHECK_MODULES invocations, if the first invocation is disabled, which therefore fails to find the package. Without this running `./configure --disable-nfsrahead` fails: checking for libnl-3.0 >= 3.1... no configure: error: in '[redacted]/nfs-utils': configure: error: A pkg-config implementation could not be found or is too old. Make sure it is in your PATH or set the PKG_CONFIG environment variable to the full path to pkg-config. Signed-off-by: Sven Püschel Signed-off-by: Steve Dickson commit f11125bfdf4ba7968729f7b7123b882c3841835c Author: Karim Manaouil Date: Fri Oct 30 09:24:36 2026 -0500 nfsiostat: display READ_PLUS statistics when available NFSv4.2 mounts may report READ_PLUS operations separately from READ in /proc/self/mountstats. nfsiostat parses these counters but does not display them, leaving read activity out of its reports. Display READ_PLUS in the default and page cache reports when the operation is present. Document the conditional output. Signed-off-by: Karim Manaouil Signed-off-by: Steve Dickson commit a246251594b1bfb34d55a81e5f194f68a9646e2c Author: Benjamin Coddington Date: Fri Oct 30 09:18:37 2026 -0500 mountd: hold SIGTERM during rpcbind registration instead of ignoring it mountd ignores HUP, INT and TERM from just before it unregisters and re-registers with rpcbind until it installs its real handling. A SIGTERM in that window is simply lost, and mountd carries on as if nothing happened. With a local rpcbind, about half of the SIGTERMs sent in the first 30ms of startup were dropped this way. Call cache_block_signals() there instead of ignoring them. A signal that arrives during registration now stays pending until my_svc_run(), which unregisters and exits. Assisted-by: LLM Signed-off-by: Benjamin Coddington Signed-off-by: Steve Dickson commit 7c15f9d413ae743e0ab7a9f996a15fe3011a410f Author: Benjamin Coddington Date: Fri Oct 30 09:14:46 2026 -0500 exportd: stop exiting from a signal handler Same change as for mountd: use cache_block_signals() and call killer() from the process loop once cache_stop_signal() is set. Assisted-by: LLM Signed-off-by: Benjamin Coddington Signed-off-by: Steve Dickson commit 4c264c0ac107a64d6236899e08498926ebafbfd4 Author: Benjamin Coddington Date: Fri Oct 30 09:07:03 2026 -0500 mountd: stop exiting from a signal handler Use cache_block_signals() instead of installing killer() and sig_hup(). my_svc_run() returns once a stop signal is seen and main() calls killer() from normal context. The worker handling in killer() is gone; cache_fork_workers() does that now. Assisted-by: LLM Signed-off-by: Benjamin Coddington Signed-off-by: Steve Dickson commit d149a2ef07e61e98d39d5dff5359dbd6bf91cb6e Author: Benjamin Coddington Date: Fri Oct 30 09:01:18 2026 -0500 support/export: take SIGINT/SIGTERM/SIGHUP via signalfd mountd and exportd exit from a signal handler that unregisters with rpcbind, frees memory, logs and calls exit(). None of that is async-signal-safe. We have a core from a mountd that took SIGTERM while pthread_create() was allocating TLS for the nfsd_path workqueue thread; killer() went on to calloc() in clnt_vc_create() and glibc aborted with "malloc(): unsorted double linked list corrupted". Add cache_block_signals() for the daemons to call in place of installing handlers. It blocks the three signals and opens a signalfd that cache_process() reads in its normal loop; cache_stop_signal() tells the caller when to shut down. SIGHUP is logged and ignored there as before. Workers close the signalfd and unblock, so they still die of SIGTERM. The parent of the workers keeps a handler, but all it does is send SIGTERM to the process group once. cache_wait_for_workers() now retries on EINTR, and the caller's existing "no more workers" path cleans up. ha_callout() clears the signal mask in its child before exec, since a blocked mask survives exec. Assisted-by: LLM Signed-off-by: Benjamin Coddington Signed-off-by: Steve Dickson commit 65ac44e831caf230e863a757d5f405c5e3971e2f Author: Benjamin Coddington Date: Fri Oct 30 08:06:18 2026 -0500 nfs-fh-verify: add tool to validate kNFSD filehandle signatures knfsd optionally signs filehandles with an 8-byte SipHash-2-4 MAC for exports marked with the sign_fh option (see fs/nfsd/nfsfh.c, kernel commit 2a83ffc55750). When triaging captures or filehandle issues it is useful to confirm offline that a given filehandle's MAC verifies under the same key the running server was configured with. Add a small standalone utility, nfs-fh-verify, that: - reads filehandles in ASCII hex from stdin (one per line, tolerant of "0x" prefixes, ':' or whitespace separators, '#' comments, and mixed case); - derives the 128-bit SipHash key from a key file using the existing hash_fh_key_file() helper, matching what nfsdctl pushes to the kernel via NFSD_A_SERVER_FH_KEY; - recomputes the trailing 8-byte MAC over fh[0..n-8] and reports OK, BAD, or MALFORMED for each filehandle; - exits 0 if every parsed filehandle verified, 1 on any BAD/MALFORMED, or 2 on setup error. The SipHash-2-4 implementation is embedded as a small reference port of the kernel's lib/siphash.c, validated at startup against two published Aumasson/Bernstein test vectors so that an incorrect build refuses to emit verdicts rather than silently mis-reporting. Wired in behind --disable-nfs-fh-verify (default on), mirroring the existing nfsrahead pattern. Signed-off-by: Benjamin Coddington Co-Authored-By: Claude Opus 4.7 Signed-off-by: Steve Dickson