xlab: persist CAPWAP DF-clear so Cisco APs can join the WLC over the wg tunnel ...

The lab 3802 (b08b.cfa8.4fa8) joins the C9800 (10.0.10.10) via opt43/DNS, but its
DTLS cert is a ~1469B DON'T-FRAGMENT packet that blackholes on the 1420-MTU
wg-to-wgnet tunnel (AP ignores PMTUD + DHCP interface-mtu). Add a systemd service
that clears DF on UDP->10.0.10.10 at bond.lan254 ingress (tc pedit) so the cert
fragments through and the WLC reassembles it. With this live the AP reached
Operation Status: Registered. See memory cisco-wlc-discovery.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

xlab/kea: pin AP wired MTU to 1420 (interface-mtu) so CAPWAP joins over WG ...

The 3802 sends 1441-byte DF DTLS packets to the WLC and ignores PMTUD
frag-needed (mtu 1420), so the join blackholes on the 1420 wg-to-wgnet tunnel
and cycles every 30s. Add DHCP option-26 interface-mtu=1420 to the cisco-ap
class so APs cap their wired MTU and never exceed the path MTU.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

xlab/kea: announce Cisco WLC 10.0.10.10 via DHCP option 43 (cisco-ap class) ...

Mirrors door1's kea-dhcp4 cisco-ap client-class so remote APs on the lab
(10.253.254.0/24) discover the Catalyst 9800-CL WMI: option-60 "Cisco AP*"
clients get vendor-encapsulated-options F1:04:0A:00:0A:0A (opt43 sub-opt 241).
Both scopes already hand out domain dev.skyw.top + DNS 10.0.0.1; the
CISCO-CAPWAP-CONTROLLER.dev.skyw.top A record (10.0.10.10) is the DNS-discovery
backup. Lab->WLC reachability over wg-to-wgnet confirmed (1.24ms).

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

skydick/samba: veto macOS .DS_Store / AppleDouble junk at the SMB layer ...

Add `veto files` + `delete veto files` to the Samba global so Finder turds
(.DS_Store, ._*, .Trashes, .TemporaryItems, .fseventsd, .apdisk) stop
accumulating on the rw shares. fruit:metadata=stream keeps genuine Apple
metadata in xattr streams, so vetoing ._* is safe.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

skydick: read-only [torrents] SMB share + stale-doc fixes ...

Expose /srv/media/data (raw torrent payload) read-only & browseable to
@storage as \\SKYDICK\torrents, so every storage user can see the per-tracker
download tree and re-seed the same files without re-downloading. Seeding needs
only read access, so the share grants no write: door-pek's qBittorrent (uid
900), writing /srv/media over the rw,all_squash NFS export, stays the sole
writer. On-disk perms already permit @storage reads (payload 0640, dirs
0750/2775, group-readable by storage gid 997).

Docs:
- README: deploy-rs is the canonical push workflow (was "not consistently
  used"); skydick is bonded 2x40G ConnectX-3 / MTU 9200 (was "bonded 100G");
  IPv6 is enabled with stable token ::d1c0 (was the old enableIPv6=false gotcha).
- DATAPOOL.md: network header, new Torrents share row + "### Torrents" section
  with a re-seed runbook, connect examples, the remote door-pek writer note,
  and MTU 9200.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

Added ajax to skydick

skydick: pin local CERNET NTP (ntp.tuna.tsinghua.edu.cn) instead of abroad pool ...

Default 0.nixos.pool.ntp.org resolved to a Vultr server abroad (137ms delay,
42ms jitter) -> ~60ms offset swings that flapped the clock-drift alert. tuna is
on-campus (0.58ms RTT); ntsc stratum-1 (~19ms) backup. Fixes the offset at the
source instead of widening the alert.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

skydick: lock ldx NFS export to single client (10.0.75.15 + v6 /128) ...

Was exported to the whole 10.0.0.0/16 LAN; restrict to ldx's host only,
per request. v6 uses the /128 (fd99:23eb:1682::75:15), NOT a /64 — the
LAN ULA /64 is shared by every host, so a /64 would defeat the lock
(unlike zhuyz24's /64, which is their own private remote network).

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

Merge PR #2 (xuelin): skydisk allow zzy via xuelin ip

skydisk: allow zzy via xuelin ip

skydick/networking: disable IPv6 temporary addresses on bond40g ...

The IPv6 SLAAC config aimed to pin the stable ::d1c0 token address as the
source, but NixOS emits net.ipv6.conf.bond40g.use_tempaddr=2 by default,
so the kernel generated and preferred a rotating temporary privacy
address. networkd's IPv6PrivacyExtensions=false does not override that
sysctl. Set the interface tempAddress = "disabled" so NixOS emits
use_tempaddr=0 — a storage server needs a fixed IPv6 source for ACLs,
logging, and reverse DNS.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

skydick/networking: mkForce IPv6PrivacyExtensions to fix eval conflict ...

The merged IPv6 PR set IPv6PrivacyExtensions = false on the 40-bond40g
networkd unit, but NixOS unconditionally sets it to "kernel" at normal
priority for every interface in networking.interfaces. Equal-priority
double definition aborted evaluation. mkForce resolves it while keeping
the intent: disable privacy extensions so the stable ::d1c0 token
address is the source for outbound IPv6.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

skydisk: IPv6 不加 via 路由

skydisk: add IPv6 address

nfs-rdma: bind listener lifecycle to nfs-server.service ...

The nfsd-rdma-listener oneshot writes "rdma 20049" to
/proc/fs/nfsd/portlist exactly once (at boot or first activation).
It then stays in `active (exited)`. But the nfsd kernel module's
portlist is wiped when nfs-server.service STOPS — and every
`nixos-rebuild switch` that touches services.nfs.server or its
config cycles nfs-server. After the cycle, port 20049 silently
goes dead while the listener unit shows "active" (it never re-ran).

Caught today (2026-05-26):
  01:05:31  ldx ran nixos-rebuild switch on skydick
  01:05:49  nfs-server stopped → portlist wiped
  01:05:52  nfs-server started, port 2049 back up
            nfsd-rdma-listener.service NOT restarted (still in
            "active exited"), so 20049 stays unbound
  ~14h later: door-pek's NFS-RDMA mount finally hangs hard, df
            blocks, telegraf disk input stops emitting, my
            disk-data-full alert goes NoData and pages with
            "[no value]" labels. qBittorrent unhealthy from
            its /mnt/media-touching healthcheck. *arr health
            issues climb.

Fix: add `partOf = [ "nfs-server.service" ]`. systemd then
propagates restart from the parent unit to the listener oneshot,
which re-runs its ExecStart (re-writes "rdma 20049" into
portlist) on every nfs-server cycle. Also added the listener
to nfs-server.service's wantedBy so a stop-then-start sequence
brings the listener back too.

Manual mitigation already applied — port 20049 is back up after
`sudo systemctl restart nfsd-rdma-listener.service`. This commit
is the structural fix so the next nfs-server cycle doesn't
silently break NFS-RDMA again.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

skydick: onboard zhuyz24 to datapool ...

LDAP user (uidNumber 2200000020). Local override pins UID; GID stays at
100 (users) per existing convention so on-disk ownership matches what
NFS exports anongid to.

Revert "skydick/samba: advertise RSS + speed for SMB Multichannel" ...

The `interfaces = "lo;capability=DYNAMIC,speed=1 10.0.1.1;capability=RSS,speed=..."`
directive from 4f21721 is malformed for Samba's parser. Samba's
interfaces list uses spaces between entries and semicolons to attach
multichannel options to a single interface, but the parser in 4.22
splits on semicolons FIRST, producing 6 invalid tokens instead of 2
tagged interfaces:
  lo capability=DYNAMIC speed=1 10.0.1.1 capability=RSS speed=80000000000

Symptom chain caused by this:
  - getaddrinfo failed for "capability=DYNAMIC" (logged in smbd debug)
  - interfaces table corrupted → rpcd_classic crash-loops with
    NT_STATUS_CONNECTION_DISCONNECTED on svcctl endpoint init
  - SMB auth from real clients rejected as "Authentication error"
    even with valid credentials (LDAP backend was fine; the proximate
    cause was the broken RPC fabric below the auth layer)
  - Both ldx@MacBook and ldx@Mac-mini couldn't connect via Finder
    SMB to \\10.0.1.1\ldx (verified with smbutil view -N //[email protected]
    → "server rejected the authentication")
  - LDAP entries were intact (sambaNTPassword still 981fb5a6...,
    POSIX userPassword still SSHA-hashed, account flags [U  ])

Reverting drops `interfaces`, `bind interfaces only`, and keeps only
`server multi channel support = yes` so clients can still negotiate
multichannel (just without the server-side RSS advertise hint).

Re-enabling capability advertising can be tried later with verified
syntax. Candidate per the Samba wiki examples:
  interfaces = bond40g;capability=RSS,speed=80000000000
(without loopback's DYNAMIC tag, which may be what tripped the parser).

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

skydick/samba: advertise RSS + speed for SMB Multichannel ...

Closest thing to "use the whole bond from SMB" without the
ksmbd/userspace fork dilemma. `server multi channel support = yes`
was already on but the client side (macOS, Windows) needs an
explicit capability hint to actually open more than one TCP
stream — without `interfaces ... ;capability=RSS,speed=...`,
Sequoia clients negotiate a single channel and that's it.

Now advertised:
  lo         capability=DYNAMIC speed=1                (loopback)
  10.0.1.1   capability=RSS     speed=80_000_000_000   (bond40g aggregate)

Effect: Sequoia 15.4+ opens up to 32 channels per SMB3 session;
LACP layer3+4 xmit hash distributes those streams across both
40 GbE bond slaves. Expected SMB throughput improvement is 2-4×
on bulk Finder copies vs single-channel TCP, with zero feature
loss (still userspace smbd: LDAP, Spotlight, fruit, recycle, TM
all intact).

Paired with `bind interfaces only = yes` so smbd doesn't listen
on any incidental interface (the existing `hosts allow` IP filter
stays as a second layer).

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

skydick/networking: tidy bond40g comments + reference cisco Po5

skydick/networking: drop bond0, rename references to bond40g ...

bond0 (ConnectX-4 LX 25G, active-backup) was carrying 10.0.1.1/16
until the 2026-05-15 cutover onto bond40g (2× 40G ConnectX-3 LACP
layer3+4, MTU 9200). With cutover done and verified — Aggregator ID
1 on both slaves, jumbo end-to-end to gateway, traffic flowing —
the old bond is dead weight.

* Remove the `bonds.bond0` and `interfaces.bond0` blocks.
* Rename the remaining active `bond0` references to `bond40g`:
  - `systemd.network.networks."40-bond0"` → `."40-bond40g"`
  - `"net.ipv6.conf.bond0.accept_ra"` sysctl
  - `skyworks.monitoring.netInterfaces = [ "bond0" ]`
  - wait-online and RA-leak comments
* The freed enp4s0f0np0/enp4s0f1np1 are now standalone DOWN,
  available for future use.

The live kernel `bond0` device persisted past nixos-rebuild
because networkd doesn't destroy unmanaged ifaces; cleaned up
manually with `ip link set <slave> nomaster; ip link del bond0`.

Revert "skydick/samba: enable SMB-Direct" ...

The previous commit (407a0b3) was based on a wrong premise. Userspace
Samba's smbd does NOT implement an SMB-Direct (RDMA) transport even
with --with-smb-direct passed to waf — the flag is silently accepted
but the resulting binary contains no ibverbs code (verified post-
deploy: ldd /bin/smbd shows no libibverbs linkage, smbd doesn't
listen on port 5445, and testparm rejects "smb direct" as an unknown
parameter).

SMB Direct in Linux is implemented in the kernel server `ksmbd`
(net/smb/server/ in the kernel tree), which is a separate
implementation from Samba. ksmbd would lose us:
  - passdb backend = ldapsam (LDAP-backed posix users)
  - Spotlight + tinysparql tracker integration
  - vfs_fruit (metadata stream / macOS attrs / Time Machine sparse-
    bundle support — central to ldx-timemachine share)

Not a worthwhile trade for the SMB workload, which is interactive
Finder browsing not bulk throughput. NFS-over-RDMA on the same
RoCE fabric (mlx4_ib via bond40g) covers the bulk-throughput case
already.

Replaced the misleading "SMB Direct" comment block with an explicit
"why this is NOT enabled" note so this doesn't get re-tried.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

skydick/samba: enable SMB-Direct (SMB3 over RDMA, port 5445) ...

Two coordinated changes:

1. sambaFull overlay extended to build with SMB-Direct support:
   - rdma-core added to buildInputs (provides libibverbs + librdmacm)
   - --with-smb-direct passed via configureFlags so waf wires up the
     transport layer at compile time

2. settings.global gains `smb direct = yes` + 8 MiB read/write knobs
   matching the NFS rsize/wsize on the same fabric. smbd now advertises
   capability 0x40 on protocol negotiate; clients that speak SMB-Direct
   (Win Server / Win Pro for Workstations / macOS Sequoia 15.4+) can
   upgrade SMB3 sessions onto the bond40g RoCE fabric. Clients without
   SMB-Direct silently fall back to plain TCP on 445.

The 2×40 GbE bond40g (ConnectX-3, post-cutover 2026-05-15) is the same
RDMA fabric NFS uses; SMB-Direct shares it without contention since
the queue-pair fanout is per-session. The "10 GbE NIC" comment in the
settings block is stale — replaced with the current 80 Gbps reality.

Build cost: sambaFull overlay forces a local rebuild on deploy
(~10-15 min, one CPU bound on smbd compilation).

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

skydick/nfs: enable NFS-over-RDMA listener on port 20049 ...

Additive to the existing TCP listener — clients choose one transport
per mount, so adding RDMA doesn't disrupt anything. The hardware path
exists: mlx5_bond_0 (the LACP bond's RDMA representation) is ACTIVE
with link_layer=Ethernet (RoCEv2). Bonded RoCE on ConnectX-5 surfaces
both 25 GbE slaves as a single RDMA device, so RDMA traffic uses the
full 50 Gbps aggregate via the firmware's own LAG handling.

Clients (door-pek) use proto=rdma,port=20049 in nfs.nix to opt in.
RDMA transports have intrinsic parallelism (queue pairs), so nconnect
becomes a no-op — drop it from the mount options when switching.

Idempotent listener registration: nfsd's portlist accepts duplicate
adds with EINVAL, so the oneshot pre-checks before writing.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

skydick/networking: skyw VLAN MTU 9000 → 9200 jumbo frames ...

Match the skyw storage VLAN end-to-end:
  cisco Po4 (switch port-channel):  9216
  skydick (bond0 + skyw VLAN):      9200  ← this commit
  door-pek (bond0 + skyw VLAN):     9200

The 9000 → 9200 bump leaves 16 bytes of headroom under cisco Po4 9216
for VLAN tag + L2 overhead.

Pairs with nix-infra commit 0xxxxxxx (door-pek/networking: skyw VLAN
MTU 9200 jumbo frames).

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

skydick/nfs: crossmnt on per-user exports so child datasets are reachable ...

Per-user namespace is structured as:
  dick/users/ldx                — parent (quota boundary, no content of its own)
  dick/users/ldx/files          — SMB-exposed personal files (\\SKYDICK\ldx)
  dick/users/ldx/bt-state       — *arr / qBT runtime state
  dick/users/ldx/timemachine    — macOS sparsebundle target (\\SKYDICK\ldx-timemachine)
  dick/users/ldx/vm             — VM disk roots

Without crossmnt on the parent export, NFS clients mounting
/srv/users/ldx only see the parent dataset and hit empty placeholders
where the children mount. 2026-05-14 incident: door-pek's baidunetdisk
container bound /mnt/users/ldx/baidu (top-level placeholder location)
because /mnt/users/ldx/files showed empty over NFSv3 — downloads landed
outside the SMB-visible namespace until the dataset boundary was
diagnosed.

Adding crossmnt makes the children visible from the existing parent
export with no separate export entries; equivalent to `nohide` on each
child. Options (all_squash, anonuid=1000) inherit naturally — exactly
the behaviour the parent already provides.

Also applied to /srv/users/ye-lw21 for parity (same dataset shape).

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

monitoring: add nodeExporter option, enable on skydick ...

Replaces telegraf-as-only-monitoring with a declarative node-exporter that
the skyw-gw Prometheus scrapes directly. Telegraf->InfluxDB(door1) keeps
running until door1 retirement so the legacy skydick.json grafana
dashboard does not go dark mid-migration.

xlab-gw: fix MSS clamp — match SYN-ACK too, use rt mtu ...

Old rule `tcp flags & (syn|ack) == syn` only matched plain SYN.
SYN-ACK from the server has SYN+ACK both set, so masking with
syn|ack and comparing == syn FAILED for SYN-ACK. Result: server
responses came back unclamped, full-MTU TCP segments overflowed
the WG path's effective MTU (1420 inner), large pages silently
stalled — YouTube didn't load, Microsoft pages partial-loaded,
Google was slow. Browsers retried indefinitely, looked like
"the network is broken" from a user perspective.

Replaced with `& (syn|rst) == syn` which matches both plain SYN
and SYN-ACK (only excludes RST, which carries no data). Combined
with `set rt mtu` instead of the hard 1280 — lets the kernel
pick the right MSS per egress interface (wg-to-wgnet → 1380 v4 /
1360 v6) instead of pessimistically clamping everything.

User's commented-out line had the right idea (rt mtu) but wrong
flag mask; fixed both at once.

skydick: also disable RA in systemd-networkd userspace ...

Sysctl accept_ra=0 only stops the kernel — systemd-networkd does
its own RA processing in userspace and was caching the link-DNS
even after the kernel sysctl was applied. Override the auto-
generated 40-bond0.network with networkConfig.IPv6AcceptRA=false.

skydick: suppress IPv6 RA processing on bond0 ...

`networking.enableIPv6 = false` only disables IPv6 forwarding/use;
the kernel still accepts router advertisements unless told otherwise.
The gateway's radvd was seeding fd99:23eb:1682::1 as a per-link DNS
on bond0, which then took precedence in systemd-resolved for AAAA
queries — making blocked names error as 'Connection refused' instead
of returning a clean NXDOMAIN through 10.0.0.1's mosdns.

Set accept_ra=0 globally + on bond0 explicitly. Existing 'enableIPv6
= false' continues to handle the higher-level disable.

skydick: route DNS via 10.0.0.1 only, AliDNS as fallback ...

Was: nameservers = [ "10.0.0.1" "223.5.5.5" ] — both treated as
primary by systemd-resolved, which then load-balanced to AliDNS
and bypassed mosdns's analytics blocking (resolvectl confirmed
hm.baidu.com / google-analytics.com leaking through).

Now: 10.0.0.1 only as primary, AliDNS demoted to fallbackDns so
it activates only when 10.0.0.1 is unreachable.