2
0
mirror of https://github.com/xcat2/xcat-core.git synced 2026-11-19 17:02:15 +00:00
Commit Graph

29412 Commits

Author SHA1 Message Date
Daniel Hilst bc2be657b6 fix(xcat-core): a ppc64le node is reserved for a DHCPv6 identifier its firmware never sends
SLOF, the open firmware of a PowerKVM guest and a PowerVM LPAR, identifies
itself with an RFC 3315 DUID-LL built from the MAC of its interface, and it
reports no UUID at all. makedhcp derives every DHCPv6 reservation from
vpd.uuid, so the reservation of a ppc64le node can never match: measured on an
IPv6-only node, SLOF sent 00:03:00:01:<mac> against the 00:04:<uuid>
reservation xCAT wrote. The node gets no reserved address and no RFC 5970
option 59, and its firmware has nothing to load.

dhcp6_client_duid now takes the MAC and the architecture. kea_duid_ll_from_mac
builds the DUID-LL from the first address of mac.mac, which is the interface
that netboots, and dhcp6_firmware_duid_is_link_layer says which architectures
send one. A ppc64le node takes the DUID-LL with no attribute set, and the
byte-order warning about vpd.uuid is not reported for it, because that firmware
does not read the value. mac.duid gains ll, which selects the DUID-LL on any
architecture. firmware and canonical are unchanged and still override.

One reservation covers the whole netboot: the grub2 ieee1275 network stack
builds its DUID from the same interface, so the loader sends what the firmware
sent. That is why no second reservation is written, unlike a chainloading UEFI
node.

dhcp_dhcpv6_duid.t gains five cases. On the unfixed tree 15 of its 195
assertions fail, cn11 with the firmware-order DUID-UUID where the DUID-LL is
expected. A mutation that gives every architecture the DUID-LL fails 42,
naming the x86_64 cases that must keep the DUID-UUID.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-11-03 19:32:00 -03:00
Daniel Hilst 6b11e0a419 test(xcat-core): a ppc64le node is reserved for a DUID its firmware never sends
SLOF, the open firmware of a PowerKVM guest and a PowerVM LPAR, identifies
itself with a DUID-LL built from the MAC and reports no UUID at all. makedhcp
derives the DHCPv6 reservation from vpd.uuid, so the reservation of a ppc64le
node can never match it: the node gets no reserved address and no RFC 5970
option 59, and its firmware has nothing to load.

Five cases are added to the identifier table. cn11 fails with the
firmware-order DUID-UUID where the DUID-LL of its MAC is expected; cn13 to cn15
fail with no DUID at all. cn12 sets mac.duid=firmware on a ppc64le node and
passes either way, so a change that gave every node a DUID-LL would not satisfy
the table. The MAC each DUID-LL is built from is asserted, so the expected
values are checkable by hand.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-11-03 19:29:50 -03:00
Daniel Hilst 4aafe12a0a fix(xcat-core): iPXE gets no DHCPv6 boot file, so an IPv6-only node cannot netboot
On an IPv6-only x86_64 node with netboot=ipxe the firmware fetches the shim and
starts iPXE, and iPXE then prints "Nothing to boot: No such file or directory".
The node falls through to "BdsDxe: No bootable option or device was found". The
same node boots over IPv4.

A UEFI node that chainloads is two DHCPv6 clients on one NIC and they send
different DUIDs. EDK2 copies the SMBIOS UUID into its DUID-UUID in the order
SMBIOS stores it, little-endian in the first three fields from SMBIOS 2.6 on,
and iPXE converts the same bytes to RFC 4122 order first. Measured on
xcat147-cn-tmp-333 with vpd.uuid 5d2df358-d529-11f1-9757-4233c0a89316: the
firmware sends 00:04:58:f3:2d:5d:29:d5:... and iPXE sends
00:04:5d:2d:f3:58:d5:29:... kea_node_reservations6 and addnode6 write one
reservation, keyed on the DUID that mac.duid selects, so the other client
matches nothing and gets no RFC 5970 option 59. On IPv4 the reservation is
keyed on the MAC and both clients share it.

dhcp6_client_duid_list now returns every DUID the clients of a node send, and
both backends reserve for each of them when the netboot method chainloads. The
first entry keeps the address and the hostname: Kea refuses a second
reservation for one address, and _reservation_matches treats a shared hostname
as the same reservation. dhcp6_bootfile_url_for_node takes the client, because
iPXE needs the boot script of the node and not the shim again, which is the
split the DHCPv4 client classes make with the iPXE feature options.

A MAC-keyed reservation was measured and rejected. Kea 2.4.1 derived no MAC for
these clients even with mac-sources ipv6-link-local and the EUI-64 link-local
on the wire: the lease kept an empty hwaddr column and the firmware took a pool
address. ISC DHCP has no hardware-address match for DHCPv6 at all. Two
reservations for one address are refused with "There's already reservation for
this address/prefix", and one reservation cannot hold both identifiers:
"the 'hw-address' and 'duid' are mutually exclusive".

xCAT-test/unit/dhcp_dhcpv6_duid.t covers both backends. Seven of its cases fail
without this change.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-11-02 14:58:44 -03:00
Daniel Hilst d202094813 test(xcat-core): an IPv6-only node that chainloads iPXE never gets a boot file
Over DHCPv6 a UEFI node that chainloads iPXE is two DHCP clients on one NIC, and
the two send different DUIDs. EDK2 copies the SMBIOS UUID into its DUID-UUID in
the order SMBIOS stores it, little-endian in the first three fields from SMBIOS
2.6 on. iPXE converts the same bytes to RFC 4122 order first. makedhcp reserves
the address for one DUID, so iPXE matches nothing, receives no RFC 5970 option
59 and prints "Nothing to boot: No such file or directory".

The second client also needs a different boot file. The shim loads ipxe.efi from
the directory it came from, which iPXE cannot do, so a chainloaded iPXE given
the shim again stops at "Failed to open ipxe.efi".

This adds the cases to xCAT-test/unit/dhcp_dhcpv6_duid.t. They assert on the
reservations that kea_build_node_reservations6 returns and on the omshell hosts
that makedhcp creates on ISC DHCP. Seven fail against the unfixed plugin.

One existing case moves from cn03 to cn01. It asserts that an ipxe node whose
shim is absent gets no option 59, and cn03 asks for the canonical DUID order,
which is the order iPXE sends, so the boot file of its first reservation is the
boot script and not the shim.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-11-02 14:58:30 -03:00
Daniel Hilst b195fea1a7 fix(yaboot): a dual-stack node is linked to a boot configuration name no firmware asks for
yaboot links its configuration file to one name for each address of the node in
the mac table. The name is sprintf("%02x%02x%02x%02x", split(/\./, $ip)), and an
IPv6 address has no four octets: sprintf reads the first field as a number and
fills the rest with zeroes, so 2001:db8:156::10 gives 7d1000000. A dual-stack
node got one correct link and one link to a name no firmware requests, with no
warning. getipaddr returns the IPv6 address first on such a node, so this is
reachable without an IPv6-only network.

yaboot_config_name returns a name only for an address with four octets in range,
and the loop passes over anything else. A node that produced no name at all is
reported with what to use instead, rather than reporting success.

The error for a node with no facing address and no service node interpolated
$myname, which is declared in the branch above it and not in that one. The file
has no "use strict", so the message read ": " and the name of the management node
was lost. It now determines the name where it uses it.

This commit no longer routes the facing address by family. yaboot is reachable
only at osarch=ppc64 with RHEL earlier than 7 or SLES earlier than 11.4, and the
support matrix carries no ppc64 column, so there is no supported configuration
whose yaboot node could be IPv6-only. petitboot keeps that work, because
nodediscover selects it for PowerNV and noderes.5 documents it for bare-metal
ppc64le.

yaboot_config_name.t covers the guard. Removing it fails the four name checks and
leaves the two IPv4 checks passing, which tells the defect apart from a test that
dies.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-31 09:07:37 -03:00
Daniel Hilst 9804168ca9 Merge upstream master into feat/ipv6
Integration test of the 51 commits that landed on upstream master after this
branch was built. Four conflicts, all from edits to the same lines:

perl-XCAT/xCAT/BootUtils.pm: both sides add a new sub at the end of the file.
Master adds expand_node_template, this branch adds grub2_kernel_commands and
grub2_image_provides. All three are kept.

xCAT/xCAT.spec and xCATsn/xCATsn.spec: master wraps the chrony and dhcpd
requirements in a suse_version test, and its %endif lands on the line this
branch uses for radvd. The requirement stays outside the test, because an
openSUSE management node writes radvd.conf too.

xCAT-server/xCAT-server.spec: master's openSUSE requirement list is a superset
of this branch's, so master's is taken whole.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-31 08:55:22 -03:00
Daniel Hilst a85baeaf3d fix(template): the Subiquity install fetches from an unbracketed IPv6 server
The Subiquity templates build three URLs as
"http://#XCATVAR:XCATMASTER##COLONHTTPPORT#/install/autoinst/...", and that tag
renders the bare value of the master. For an IPv6-only node that is a literal
address, so wget read the last group as a port:

    wget -T 30 -O /tmp/getinstdisk http://2001:db8:147::1/install/autoinst/getinstdisk
    http://2001:db8:147::1/install/autoinst/getinstdisk: Bad port number.

The early-commands script runs under "set -eux", so that one line ended it.
subiquity reported CalledProcessError from Early/run, wrote a crash report of
kind "unknown", and stopped before curtin ran. Nothing in the installer logs
named the cause: curtin-install.log was empty, cloud-init was healthy, and the
error report that looked like the failure was the Error CONTROLLER loading its
own configuration.

Template.pm already computed $master_uri_host, bracketed for an IPv6-only
node, and used it for the SLES media URL and the repo baseurl. It is now
exposed as #MASTERURIHOST#, and the three URLs take it.

#XCATVAR:XCATMASTER# keeps the bare value where it is right: bash reads
/dev/tcp/<host>/<port> with no brackets, and the updateflag notification uses
that form.

subiquity_uri_host.t reads both templates, which are the artifact: every
http:// URL has to take the URI host, the /dev/tcp line has to keep the bare
one, and Template.pm has to substitute the tag. Restoring the bare tag in one
URL fails two checks; dropping the substitution fails the third.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 15:56:24 -03:00
Daniel Hilst 319a20106e fix(template): the Subiquity autoinstall gives an IPv6-only node DHCPv4 only
Both Subiquity templates wrote "dhcp4: true" for every node, in the network
section the installer uses and in the netplan the late command writes into the
installed system. An IPv6-only node has no DHCPv4 server to answer, so netplan
brought the provisioning NIC up with no address.

Measured on an IPv6-only node: the installer fetched its autoinstall seed over
IPv6, then could not configure the interface and fell back to asking at the
console. On a serial console that waits for a keypress, so the install stopped
at the mode prompt with no sign of why. The installed system would have had no
address either, which is the defect xcatroot had for a diskless node and
cd6719946 fixed there.

#SUBIQUITYDHCP:<indent># and #SUBIQUITYTARGETDHCP# render from the address
family of the node, beside the #SUBIQUITYINSTALLNIC# and #SUBIQUITYINSTALLMAC#
tags that were already there. An IPv6-only node gets dhcp6 with dhcp4 turned
off explicitly, because netplan defaults dhcp4 to false only when nothing
names it, and the overrides follow the family in use. An IPv4 node gets what
it had, byte for byte.

$node_ipv6_only becomes 'our' so a test can set it. subvars already computes
it before these substitutions run, so nothing recomputes the lookup.

subiquity_ipv6_netplan.t covers both families and both sites, and asserts the
templates carry the tags rather than a literal. ubuntu_subiquity_installnic.t
renders the new tag and gains an IPv6 case: it runs the late command for real,
and an unrendered tag is a shell comment that swallows the redirect, so that
test would have written no file at all.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 14:59:20 -03:00
Daniel Hilst 518ddee895 test(service-node): a check matches the path of the checkout it runs in
service_node_artifacts_el.t asserts the EL service template is not a
Subiquity one, and it matched the pattern against the whole absolute path. A
checkout directory named for the thing under test therefore matched itself:
in a worktree called ipv6-subiquity-net the check failed on
share/xcat/install/rh/service.rhels9.tmpl, which is the right template.

It now matches the basename. A real Subiquity template still fails it, and an
EL one in a path carrying the word still passes.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 14:59:19 -03:00
Daniel Hilst f92df9bc9b fix(debian): an IPv6-only Ubuntu node never configures its interface
subiquity_url_kcmdline left ip= out of the command line, on the reasoning that
casper treats an absent IPv4 half as already done. It does, and it also skips
the whole networking stage: casper calls configure_networking only inside

    if [ -n "${STATICIP}" -a "${STATICIP}" != "frommedia" ]

and STATICIP comes from ip=. With no ip= the interface was never brought up,
so wget answered "Cannot assign requested address" and casper panicked with
"Unable to find a live file system on the network". The URL, the brackets and
the served ISO were all correct; nothing had an address to reach them from.

ip=off takes that branch while asking for no IPv4: configure_networking reads
ip= into IP and treats off as already done, then runs dhcpcd -6 on ip6=dhcp.

Measured on an IPv6-only node. With ip=off the console shows dhcpcd-10.0.6,
then "Trying to download and mount http://[2001:db8:147::1]:80/.../install.iso",
and the management node's access log records

    2001:db8:147::1003 "GET /install/ubuntu24.04.4/x86_64/install.iso" 200 3405470119 "Wget"

the whole 3.2G ISO over IPv6, after which subiquity starts. Without ip=off the
same node fails at the wget every time.

debian_subiquity_boot_params.t asserts ip= is present. Removing it from the
command line fails that check alone.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 14:21:32 -03:00
Daniel Hilst 12e1deb45a fix(xcat-core): nothing installs radvd, so the IPv6 networks announce nothing
makedhcp writes /etc/radvd.conf for every IPv6 network in the networks table
and starts the service, and no package declared radvd. On a management node
that did not happen to carry it the file was written and nothing read it. A
client that follows RFC 4861 starts DHCPv6 only when an advertisement sets the
managed flag, and takes its on-link prefix and default route from the
advertisement rather than from the lease, so an IPv6-only node received one
address with a prefix length of 128 and no route to its server.

Measured on an Ubuntu management node: radvd.service answered "not-found"
while /etc/radvd.conf held the right configuration, the compute node held a
/128 DHCPv6 lease, and its installer answered "wget: can't connect to remote
host: Cannot assign requested address" when it tried to fetch its root over
HTTP. Installing radvd made the service active and makedhcp kept it that way.

xCAT and xCATsn now require it, in the rpm specs and in the Debian control
files. A service node answers DHCP for its own subnet and writes the same
file, so it needs the daemon too.

Depends rather than Recommends on Ubuntu, because the apt 3.0 solver drops a
Recommends: that is what took bind9 out of the 26.04 install and left makedns
with no named.

radvd_packaging.t reads the four manifests, which are the artifact. Removing
any one declaration fails its own check, and moving the Debian one to
Recommends fails the check that names the difference.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 13:45:04 -03:00
Daniel Hilst b1aac24b0a fix(dhcp): router advertisements reach an interface DHCP does not serve
configure_router_advertisements took the mgtifname of every IPv6 network in
the networks table. site.dhcpinterfaces names the interfaces this server
answers on, and the advertisement ignored it, so a management node restricted
to eth1 still advertised itself on eth0. Every client on that LAN can then
select it as a router, which is what restricting the interface was meant to
prevent.

The sub now takes the interface set the DHCP configuration was built from and
skips a network whose interface is not in it. All three call sites already had
the set: the ISC path in scope, and both Kea paths as a parameter. An empty or
absent set is no restriction, so a cluster that sets no dhcpinterfaces keeps
the behaviour it had.

There are also two fixes to how radvd is started:

- The service is examined on every call rather than only when the
  configuration file changed. A start that failed left the file already
  correct, so every later makedhcp returned on the file alone and never tried
  again while radvd stayed down. The call also moved out of the dhcpd6 restart
  branch of the ISC path, which a repeat makedhcp does not enter at all.
- restartservice answers -1 when the service is not installed and the exit
  status of the command otherwise. Only the -1 was reported, so a radvd that
  failed to start was silent. Every nonzero result is now reported, and the
  message says which of the two it was.

dhcp_dhcpv6_duid.t drives the sub directly for the service cases, because
process_request reaches it only on some paths. Dropping the interface filter
fails the check on an unserved network; restoring the early return on an
unchanged file fails the restart check and leaves the one for a service that
is already up passing; ignoring a positive exit status fails the two checks
that name it.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 13:18:35 -03:00
Daniel Hilst 2ac19d7202 fix(debian): an IPv6-only node is given [<xcatmaster>] as its gateway
mknetboot read networks.gateway and passed it to dracut_static_ip6 as it
stood. '<xcatmaster>' is the common value of that column, and it reached the
kernel command line as "[<xcatmaster>]", which is not an address. A node whose
image server sits off its own link then had no route to it.

static_ip6_boot_arg resolves the gateway through ipv6_server_for_node, which
already knew the placeholder and which the AutoYaST path already used. A
gateway that does not resolve is dropped rather than written as a bad address:
an address with no route still reaches a server on the same link.

The composition is its own sub, taking its two lookups as coderefs, because
the helpers were covered while nothing connected them to the call site.
debian_static_ip6_gateway.t drives it with the real dracut_static_ip6.
Passing the gateway unresolved puts "[<xcatmaster>]" in the argument, which
two of its checks report.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 13:18:14 -03:00
Daniel Hilst 7f8fc5a937 fix(debian): copycds deletes the install media when it is given the ISO in the tree
link_install_iso unlinked its destination and then created the link. copycds
can be given the ISO that is already in the install tree, or the link that a
previous run left there, and the destination is then the same file as the
source. The unlink deleted it, and the symlink that followed pointed at a path
that no longer existed. An ISO copied into /install by hand was lost, and the
tree was left with a dangling link that answers 404 to the node fetching it.

The source and the destination are now compared by device and inode before
anything is removed, which also follows the link, so a second copycds of the
same media leaves the link alone instead of recreating it.

debian_subiquity_boot_params.t covers an ISO at the destination, a link that
resolves to the media, and a different ISO that replaces the link. Without the
comparison the first case leaves a dangling link and the file is gone, which
three of the new checks report.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 13:18:14 -03:00
Daniel Hilst 3fbf0b2a85 fix(xcat-test): a conf value is cut at its first semicolon
The value class of the conf parser held ':' and '/' and not ';', so
"othersettings=cpumode:host-passthrough;firmware:efi-secureboot" was read as
"cpumode:host-passthrough" and the rest was dropped with no warning. chtab
wrote the row with the first setting alone, and the node was defined without
the firmware it needs to boot over IPv6. The [Object_*] branch had the same
class and the same effect on vmothersetting.

The chtab command was also built by interpolation, with the row selector
unquoted, so a row name or a key carrying shell syntax would have ended the
command. chdef left its type and object name unquoted for the same reason.

parse_conf_assignment is now the one place the class is written, for both
section kinds. table_row_command builds the chtab command with every argument
quoted, the row selector among them, and object_definition_command quotes its
type and name.

xcattest_object_quoting.t asserts the argument list each command delivers, not
only its text, so a quoting change that splits an argument is caught. Removing
';' from the class fails the one check that names a semicolon value and leaves
the other fifteen passing.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 13:17:39 -03:00
Daniel Hilst 102dc80920 fix(postscripts): an installed node ignores the DUID its reservation holds
configdhcpv6duid always derived the DHCPv6 client identifier from the SMBIOS
UUID in firmware byte order. makedhcp does that only when mac.duid is unset or
holds 'firmware': 'canonical' reserves the order product_uuid prints, and a
literal DUID is reserved as it stands. A node with either setting wrote an
identifier that matched no reservation, so it took a pool address and had no
boot file URL.

mypostscript.tmpl now exports the attribute as DHCP6DUID, and the postscript
reads it. 'canonical' keeps the order product_uuid printed. A literal is
written as it stands for NetworkManager, which takes a whole DUID.
systemd-networkd composes one from a named type, and only the UUID type can be
expressed that way, so a literal of another type is written for
NetworkManager alone and the reason is logged. A value that is neither keyword
nor DUID writes nothing, which is what makedhcp refuses as well.

The derived values are byte-identical to kea_duid_from_uuid for both orders.
postscript_dhcpv6_duid.bats goes from 9 checks to 16; ignoring the attribute
fails the five that name a setting and leaves the firmware default passing.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 13:17:39 -03:00
Daniel Hilst 8c704104ab fix(template): the AutoYaST profile gives a node its gateway as a resolver
_autoyast_network6 replaced '<xcatmaster>' in site.nameservers with the
gateway of the node, and fell back to the gateway when the list held no IPv6
address. '<xcatmaster>' is the management node, which runs named. The gateway
is not: a cluster whose gateway is a separate router gave every node a
resolver that answers nothing, and a cluster with no gateway gave none at all,
so the installed node resolved no hostname.

The placeholder now resolves through ipv6_server_for_node, which is what the
gateway itself already used. A value that is not an address is not written,
so an unresolvable placeholder leaves the list empty rather than putting
"<xcatmaster>" in the profile.

autoyast_ipv6_network.t gains a management node that is nobody's gateway, so
the two can be told apart, and the check that asserted the old behaviour now
asserts the new one. Restoring the gateway fails five checks and leaves the
other nineteen passing.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 13:17:23 -03:00
Daniel Hilst 7f46d8dab3 fix(ipv6): an NFS root loses the last group of its IPv6 server
substitute_facing_address brackets the facing address only where a port
follows the placeholder. An NFS root carries the server between two colons,
as in "root=nfs:!myipfn!:/install/netboot", so an IPv6 server came out as
"root=nfs:2001:db8::1:/install/netboot". dracut reads the server up to the
last colon, so it took "2001:db8:" and the statelite node could not mount its
root.

The substitution now also brackets a placeholder that an NFS path follows. A
placeholder with nothing after it still takes the bare address, because
xcatd=!myipfn! names a host with no port and the postscripts read that value
as it stands.

networkutils_ipv6_helpers.t covers both forms in one command line. Removing
the new substitution fails the two IPv6 checks and leaves the IPv4 and
bare-placeholder checks passing.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 13:17:23 -03:00
Daniel Hilst c0f8fd42b0 feat(debian): an IPv6-only node installs Ubuntu from the ISO over HTTP
An IPv6-only node could not install Ubuntu. subiquity_boot_params refused it,
because casper mounts the live filesystem with klibc nfsmount, which has no
IPv6: it answers "need a path" for "[2001:db8::1]:/export" and for
"2001:db8::1:/export", while an IPv4 server gets past parsing to the mount.
casper takes nfsroot= as an opaque path, so no option changes the family.

casper has a second route. do_urlmount downloads an ISO with casper's own
wget, which speaks IPv6, and loop-mounts it, and the network itself was never
the limit: init parses ip6= and configure_networking runs "dhcpcd -1KL -6" on
it. The route needs an ISO to fetch, and xCAT serves the extracted media and
kept no ISO.

copycds now passes the ISO it was given to the distribution plugin, which
links it into the install tree as install.iso. A symbolic link, because the
Ubuntu 24.04 ISO is 3.2G, and the xcat.conf xCAT ships gives /install
"Options FollowSymLinks". A block device has no path to link and passes none.

An IPv6-only node then gets "ip6=dhcp boot=casper netboot=url
url=http://[server]:port/install/<os>/<arch>/install.iso", with no nfsroot
and no toram: do_urlmount already holds the ISO in the initramfs tmpfs, so
copying the live filesystem again only doubles the memory the node needs.
That memory is the cost of the route -- 3.2G for the ISO on top of the
installer -- so a node with 4G cannot take it.

An IPv4 node is untouched and still mounts its root over NFS. A tree copied
before this change carries no ISO, so such a node is refused as before, with
a message naming install.iso and copycds.

debian_subiquity_boot_params.t goes from 17 checks to 36. Routing an IPv6-only
node back to the NFS form fails the six HTTP checks; dropping the bracketing
fails exactly the two bracketing checks; and returning success from
link_install_iso without creating the link fails the three that read the link
back, while the one that reads the return value still passes.

Not yet booted on a cluster.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 10:42:54 -03:00
Daniel Hilst 8cfb0c60a9 fix(debian): the Subiquity refusal names the wrong component
subiquity_boot_params refuses an IPv6-only node and says the live installer
"configures its network and mounts its root with klibc, which has no IPv6
support". The first half is wrong. casper reads ip6= from the kernel command
line and configures the interface with dhcpcd -6, which speaks DHCPv6:
scripts/functions runs "dhcpcd -1KL -t $ROUNDTTT -6" on the ip6= value, and
init parses ip6= at line 156. klibc ipconfig is only the IPv4 branch.

The one component with no IPv6 is klibc nfsmount, which casper uses for the
live filesystem. Measured on noble klibc-utils 2.0.13: it answers "need a
path" for "[2001:db8::1]:/export" and for "2001:db8::1:/export", while an
IPv4 server gets past parsing to the mount, and the binary carries no inet6
string against seven nfs and mount strings.

The refusal stands, because casper takes nfsroot= as an opaque path and has
no way to pass it an address family. Only its reason changes. The comment
also records the route that would close this: netboot=url fetches an ISO with
casper's own wget, which does IPv6, and that is a different provisioning
method than the extracted tree xCAT exports.

The check on the message matched /klibc/, which is the family and not the
component, so it passed for any klibc tool. It now matches /nfsmount/, and
the previous wording fails it.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 10:31:41 -03:00
Daniel Hilst ec5acc96a1 fix(yaboot): a dual-stack node is linked to a boot configuration name no firmware asks for
yaboot links its configuration file to one name for each address of the node
in the mac table. The name is sprintf("%02x%02x%02x%02x", split(/\./, $ip)),
and an IPv6 address has no four octets: sprintf reads the first field as a
number and fills the rest with zeroes, so 2001:db8:156::10 gives 7d1000000.
A dual-stack node therefore got one correct link and one link to a name no
firmware requests, with no warning. An IPv6-only node got only the second.

yaboot_config_name returns a name only for an address with four octets in
range, and the loop passes over anything else. A node that produced no name
at all is reported with what to use instead, rather than reporting success.

The facing address was found with my_ip_facing, which resolves with inet_aton
and reports an IPv6-only node unresolvable. yaboot_facing_address asks
node_address_family first, the way grub2.pm and petitboot.pm do, and the
!myipfn! substitution goes through substitute_facing_address so a placeholder
that a port follows is bracketed.

The error for a node with no facing address and no service node interpolated
$myname, which is declared in the branch above it and not in that one. The
file has no "use strict", so the message read ": " and the name of the
management node was lost. It now determines the name where it uses it.

yaboot_ipv6.t covers both subs. Restoring the unguarded sprintf fails the
four name checks; calling my_ip_facing unconditionally fails the three family
checks. The IPv4 checks pass under both, which is what tells the two defects
apart from a test that dies.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 10:23:38 -03:00
Daniel Hilst 6dcde28dc9 fix(petitboot): an IPv6-only node is given a boot configuration name no firmware asks for
petitboot asks for a configuration file named after the four octets of the
IPv4 address of the node, in hexadecimal. The name is built with
sprintf("%02x%02x%02x%02x", split(/\./, $ip)), and an IPv6 address has no
four octets: sprintf reads the first field as a number and fills the rest
with zeroes, so 2001:db8:156::10 gives 7D1000000. nodeset then links the
configuration to a name no firmware requests, reports success, and the node
boots nothing.

petitboot_config_name returns the name only for an address that has four
octets in range, and nodeset reports the node instead of linking a name that
cannot work. The message says what to use: an IPv4 address on the node, or
netboot=grub2, which carries the server in the boot file URL and needs no
name of this form.

This is a refusal, not support. What name petitboot asks for over DHCPv6 is
not established here, and no cell in the lab runs petitboot, so the value
cannot be measured. An error that names the limit is what an operator can
act on; a wrong name is not.

Five checks in petitboot_ipv6.t cover it. Restoring the unguarded sprintf
fails the three IPv6 and undef checks and leaves the two IPv4 checks passing,
including the name of 192.0.2.50, which stays C0000232.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 10:15:49 -03:00
Daniel Hilst 9cc9503a96 fix(petitboot): an IPv6 node loses the server from its kernel and initrd URL
petitboot builds the URL it fetches a kernel and an initrd from as
"$httpmethod://$installsrv:$httpport$tftpdir/", so an IPv6 install server
gives "http://2001:db8::1:80/tftpboot/...", which cannot be split into a host
and a port. The node then has no server to fetch from.

The facing address is found with my_ip_facing, which resolves with inet_aton.
An IPv6-only node is reported unresolvable, and the node gets no
configuration at all.

petitboot_boot_url builds the URL through uri_authority, which brackets an
IPv6 address. petitboot_facing_address asks the family of the node which
helper answers, the way grub2.pm does. The three !myipfn! substitutions go
through substitute_facing_address, which brackets the placeholder that a port
follows and leaves every other one bare. petitboot.pm also now loads
xCAT::NetworkUtils, which it has always used and never named.

An IPv4 node keeps the URL it had, byte for byte, including the port that the
old form always wrote.

petitboot_ipv6.t covers both. Writing "$server:$port" back in place of
uri_authority fails the two IPv6 URL checks. Calling my_ip_facing
unconditionally fails the three family checks. Neither mutation touches the
IPv4 checks, which is what separates the two defects from a test that dies.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 10:15:36 -03:00
Daniel Hilst 38a60074e0 fix(dhcp): a unit test writes the radvd configuration of the host it runs on
configure_router_advertisements writes /etc/radvd.conf and runs sysctl on the
management interface. dhcp_dhcpv6_duid.t drives makedhcp through
process_request, so the test reaches both. The suite runs as root in the
GitHub CI, where the test therefore replaces the real /etc/radvd.conf of the
runner and turns IPv6 forwarding on for its interfaces.

Where the file is not writable the warning also reaches the callback that
other checks read, so the test failed on an unprivileged host:

  Structures begin differing at:
       $got->[2] = 'Cannot write /etc/radvd.conf, so the IPv6 networks
                    announce no route: Permission denied'
  $expected->[2] = Does not exist

The path is now the package variable $radvdconffile, and the sysctl call is
the sub enable_ipv6_forwarding. The test localizes both, the way it already
localizes $dhcpconffile and xCAT::Utils::restartservice.

Six checks now cover the behaviour the configuration carries: the file is
written, it names the interface of the IPv6 network, the interface sends
advertisements, the managed flag is on, the prefix is announced, and
forwarding is enabled. Removing the call from the ISC path fails exactly
those six and leaves the other 127 passing.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-30 10:15:08 -03:00
Daniel Hilst c0698fa253 Merge pull request #7995 from VersatusHPC/fix/opensuse-mn-package-requires
fix(xcat-core): xCAT cannot be installed on an openSUSE management node
2026-10-30 09:57:51 -03:00
Daniel Hilst 10c30c80ba Merge pull request #7993 from VersatusHPC/fix/getpostscript-client-identity-v2
fix(xcat-core): getpostscript serves no node when the client reverse lookup is empty
2026-10-30 09:57:22 -03:00
Daniel Hilst 2aceb80102 Merge pull request #7991 from VersatusHPC/fix/xcat-test-scapy-optional
fix(xcat-test): scapy is a hard Depends on Debian and only a Recommends on EL
2026-10-30 09:56:54 -03:00
Daniel Hilst 0312d28117 fix(dhcp): makedhcp -n writes no router advertisement
kea_process_request has two paths. The full regeneration writes the whole
configuration, restarts the services, unlocks and returns; the incremental one
falls through to the end. The call added in the previous commit sat only on the
incremental path, so makedhcp -n, which is what configures a new cluster, wrote
no /etc/radvd.conf at all.

Measured on an IPv6-only alma10.1 management node: with radvd.conf removed,
radvd stopped and forwarding set to 0, makedhcp -n left all three unchanged.
The generator itself was correct the whole time; it was never called.

The full regeneration now configures the advertisements before it returns.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-29 18:14:38 -03:00
Daniel Hilst 113463a191 feat(dhcp): an IPv6 network announces no prefix and no route
makedhcp writes the DHCPv6 reservations and nothing announces the network. xCAT
configures no router advertisement anywhere: a fresh IPv6 provision network has
no radvd and forwarding off.

Two failures follow from that one gap. A client that follows RFC 4861 uses
DHCPv6 only when an advertisement sets the managed flag, so iPXE reported
"No configuration methods succeeded" and never sent a solicit: the DHCPv6 log
held the firmware DUID alone. And a client that does solicit gets one address
with a prefix length of 128, because the on-link prefix and the default route
come from the advertisement, not from DHCPv6.

makedhcp now writes /etc/radvd.conf for every IPv6 network in the networks
table and restarts radvd, for both the ISC and the Kea backend. The
advertisement sets the managed flag and marks the prefix on-link but not
autonomous, so the address still comes from DHCPv6, which holds the
reservations. An IPv4 cluster gets no file and no service. A management node
without radvd gets a warning naming what the nodes lose.

Measured on an IPv6-only alma10.1 node with netboot=ipxe. Without the
advertisement iPXE stopped at "No configuration methods succeeded". With it
iPXE reported "Configuring (net0 ...) ok" and
"net0: 2001:db8:149::1002/64 gw fe80::4054:ff:fec9:95d3", loaded the kernel,
and the node booted and answered ssh, xdsh and updatenode over IPv6.

Captured by xCAT-test/unit/radvd_config.t. Removing the managed flag fails the
assertion that names it; making the prefix autonomous fails the assertion that
names that.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-29 17:55:55 -03:00
Daniel Hilst c2c5bc8bb6 docs(dhcp): the iPXE loader cannot finish an IPv6 boot yet
The commit that added ipxe to the DHCPv6 gate justified it with strings found
in snponly.efi. That was not evidence that the feature is built in, and a boot
disproved it.

Measured on an IPv6-only alma10.1 node with ipxe-xcat 2.0.0: the firmware does
fetch the shim over IPv6 from the option this code sends, the shim chains to
snponly.efi, and iPXE then reports "No configuration methods succeeded" on a
network whose only DHCP server is DHCPv6. iPXE never reaches the script.

The option stays, because it is what xCAT owes a client that asks, and the node
fails at the same point without it. The comment now records what the loader
cannot do, so the next reader does not repeat the strings test.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-29 17:44:32 -03:00
Daniel Hilst adeb938962 fix(xnba): the kernel and initrd URLs of an IPv6 node lose their server
The previous commit replaced !myipfn! with the facing address, which fixed the
kernel command line. It left the eleven other places where setstate writes
${next-server} into the script directly, among them the imgfetch lines for the
kernel and the initial ramdisk. An IPv6 node resolved those to
"http:///tftpboot/..." and had no server to fetch its kernel from.

Measured on an IPv6-only alma10.1 node: the script carried
"http://[2001:db8:149::1]:80/install/..." for the command line and
"http://${next-server}/tftpboot/.../vmlinuz" for the kernel.

The server is now computed once in setstate and used at every site, so the
whole script names one server. An IPv4 node still gets ${next-server}, which
iPXE fills from its own offer.

netboot_ipxe_method.t and secureboot_ipxe_node_shim.t load xnba.pm against a
stub of xCAT::NetworkUtils. The stub gains the two methods setstate now calls,
and both tests describe their nodes as IPv4, which keeps ${next-server}.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-29 17:26:26 -03:00
Daniel Hilst 7514b931b3 fix(xnba): an IPv6 node loses the server from every URL in its iPXE script
xnba.pm writes ${next-server} in place of !myipfn! and lets iPXE fill it from
the DHCP offer. next-server is a DHCPv4 field, so an IPv6 node resolves it to
the empty string and every URL in the script loses its host. The fallback does
not help either: my_ip_facing resolves with inet_aton, so it reports an
IPv6-only node unresolvable.

The script now carries the facing address of the management node when the node
has an IPv6 address, bracketed because the script puts it in a URL. The lookup
names the family, the way grub2.pm does. An IPv4 node keeps ${next-server}, so
nothing changes for it.

ipxe.pm runs this code with SCRIPTS and METHOD set to its own, so the change
reaches netboot=ipxe as well as netboot=xnba.

xnba.pm also calls xCAT::NetworkUtils eight times without loading it, and works
only because another module does. The use is now explicit.

Captured by xCAT-test/unit/ipxe_script_server.t. Ignoring the family fails five
of its seven assertions and leaves the two IPv4 ones passing.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-29 16:59:24 -03:00
Daniel Hilst 25720c77a7 feat(dhcp): an ipxe node gets no boot file over DHCPv6
dhcp6_bootfile_url_for_node writes RFC 5970 option 59 for grub2 alone, so a node
with netboot=ipxe takes its reserved IPv6 address and has nothing to load. Its
firmware then releases the address and solicits again. grub2 was the only
loader an IPv6 node could use.

The iPXE build xCAT ships does speak DHCPv6: snponly.efi, the UEFI loader of
ipxe-xcat 2.0.0, carries dhcpv6, ipv6 and four IPv6 strings, against a control
of 26 iPXE and dhcp strings in the same binary. The firmware that solicits over
DHCPv6 is UEFI, so the UEFI loader is the only one such a client can ask for.

An ipxe node now gets tftp://<bracketed server>/<UEFI loader>, and the loader
name comes from BootPolicy->x86_loader, which is the one place it is chosen. No
URL is written when that file is not on disk, because naming a file the TFTP
server does not have costs the client a timeout it cannot diagnose; the Kea
client classes already refuse it for the same reason.

xnba keeps its exclusion. Its loader reports one ipv6 string against a control
of five, where the iPXE loader reports five against 26, so there is no evidence
it reads option 59.

Captured by xCAT-test/unit/dhcp_dhcpv6_duid.t. Removing ipxe from the gate fails
the first assertion; removing the on-disk check fails the second.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-29 16:55:47 -03:00
Daniel Hilst cd67199463 fix(netboot): an IPv6-only diskless node boots with no address
xcatroot writes the netplan of the booted system from the interface it found,
and it always wrote "dhcp4: true". An IPv6-only node has no DHCPv4 server, so it
reached its login prompt with no address: Ubuntu 24.04.5 reported
"[FAILED] Failed to start systemd-networkd-w... Wait for Network to be
Configured." and "[FAILED] Failed to start xcatpostinit.service", and the node
answered no ssh.

The address is already on the kernel command line, in the static form dracut
takes. xcat_static_ip6_fields reads it and the netplan carries the address, the
prefix and a default route through the gateway. A node whose command line holds
an IPv4 form keeps dhcp4, because the fields are read only from a bracketed
literal.

Captured by xCAT-test/bats/ipv6_xcatroot_netplan.bats. Reading the fields by
cutting on the first colon, which is what an IPv6 literal breaks, fails the
three parsing assertions and leaves the four refusals passing.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-29 16:01:34 -03:00
Daniel Hilst 599cd8ac83 Merge pull request #7986 from VersatusHPC/fix/retry-install-transient-icmp
fix(xcat-test): a transient ICMP loss vetoes a node that answers SSH
2026-10-29 15:29:06 -03:00
Daniel Hilst 77eaf4d53c fix(ipv6): a diskless node has an address and no route
mknetboot gave an IPv6-only node ip=either6, which asks dracut for DHCPv6.
DHCPv6 assigns one address with a prefix length of 128 and carries no route, and
an xCAT provision network sends no router advertisement: radvd is not installed
and net.ipv6.conf.<nic>.forwarding is 0. The node came up with the address and
no way to reach the server, and the dracut hook reported
"Connecting to [2001:db8:147::1]:80... failed: Network is unreachable."

debian.pm now builds the static form from the address, gateway and prefix the
node already has in the database, the way sles.pm passes hostip and gateway to
linuxrc. either6 remains the answer when the address cannot be resolved.

Captured by xCAT-test/unit/dracut_static_ip6.t. Removing the prefix length from
the parameter fails four of its nine assertions and leaves the bracketing and
the refusal cases passing.

Measured on an IPv6-only Ubuntu 24.04 node: Kea records prefix_len=128 for its
lease, and dracut logged BOUND6 for eth0 before the fetch failed.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-29 15:27:53 -03:00
Daniel Hilst ddaad73ab4 fix(grub2): a ppc node is given linuxefi when its image carries it
grub2_kernel_commands read the image and not the architecture, so a ppc node
whose grub2 image carries linuxefi was given it. linuxefi loads an EFI binary
and a ppc loader is not one. The previous commit removed the architecture test
along with the defect it carried.

The command now needs both: the architecture whose loader is an EFI binary, and
an image that carries the command.

grub2_netboot_protocol.t covers the three architectures and caught this. Its
loader fixture now records the commands it carries, because the configuration
follows the image.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-29 15:27:53 -03:00
Daniel Hilst dfad7287f0 Merge pull request #7984 from VersatusHPC/fix/genesis-ubuntu-mawk-rebased
fix(xcat-core): the Ubuntu Genesis image runs its awk programs under mawk
2026-10-29 15:24:52 -03:00
Daniel Hilst 89b9213e4d Merge pull request #7977 from VersatusHPC/fix/boot-node-template
fix(boot): stop unsupported node-template keys from looping
2026-10-29 15:00:28 -03:00
Daniel Hilst ac0ec70821 Merge pull request #7979 from VersatusHPC/refactor/debian-profile-lookup
refactor(netboot): share profile asset selection
2026-10-29 14:58:18 -03:00
Daniel Hilst d257780576 Merge pull request #7990 from VersatusHPC/fix/genesis-openssl-config-missing
fix(xcat-core): Genesis node stops at "Getting initial certificate"
2026-10-29 14:50:08 -03:00
Daniel Hilst 075e6801e9 fix(grub2): an x86 node stops at the menu with "can't find command `linuxefi'"
setstate in grub2.pm appends "efi" to the kernel and initrd commands for every
node whose architecture matches x86, so the configuration reads linuxefi and
initrdefi. Both come from a downstream patch. grub2-xcat 2.12-2 for Ubuntu is
built from upstream grub and carries neither, so the node loads the loader and
its own configuration over TFTP, prints "can't find command `linuxefi'", asks
for no kernel, and waits at the menu until it is powered off.

The pair now comes from the image served to the node. BootUtils reads the
image, which records the name of every command its modules register, and
answers linuxefi and initrdefi only when the image carries them. An image that
cannot be read is answered with linux and initrd, which every grub2 image
carries.

This reaches a node only when it netboots with grub2. An IPv6 node has no other
loader, because bootfile-url is written for grub2 alone.

Measured on an IPv6-only Ubuntu 24.04 node: with linuxefi the node made no TFTP
request for the kernel; with linux it loaded the kernel and the initial ramdisk
and printed no error.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-29 14:45:34 -03:00
Daniel Hilst 557c3689f7 test(grub2): a grub2 image without linuxefi is still offered linuxefi
grub2.pm chooses the kernel command from the architecture of the node, so every
x86 node is given linuxefi and initrdefi. Those two come from a downstream patch
and an image built from upstream grub carries neither.

These assertions fail without the fix that follows: the upstream image, the
missing image and the undefined image are all answered with linuxefi.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-29 14:45:33 -03:00
Daniel Hilst b7076359f4 fix(xcat-test): xcat-test can install without python3 on Debian
python3-scapy was the only Depends of xcat-test that pulled in
python3. With scapy moved to Recommends, apt-get install
--no-install-recommends xcat-test can leave python3 absent, and every
blackboxtest command fails before it starts, not only the DHCP steps.

Add python3 to Depends. xcat_test_scapy_optional.t checks that python3
is a hard Depends, and fails without this change.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-29 14:42:47 -03:00
Daniel Hilst e3429791c6 fix(xcat-core): getpostscript serves a node's postscripts to a password or token login
With nodeauthentication off and no peer name, getpostscript used
_xcat_authname as the client identity and expanded it with noderange.
becomeuser and tokens set _xcat_authname to an account name, so a user
named after a node group received the postscripts of the first node in
that group.

xcatd now records the CN of the validated client certificate as
_xcat_certname, and only while no becomeuser or token replaced it. It
deletes any _xcat_certname the client sent. getpostscript uses that
name only when it names exactly one node, with no group or range
expansion. _xcat_authname is no longer a candidate.

getpostscript_client_identity.t drives the plugin with these requests.
On the previous head the group and node login checks fail, and
dropping the exact match fails the group and range checks.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-29 14:41:25 -03:00
Daniel Hilst 001edc1a9f fix(xcat-core): package_relations.t expects perl-DB_File on openSUSE
xCAT-server now asks for the perl(DB_File) capability on openSUSE,
because no openSUSE release ships a package named perl-DB_File. The
native package relation test still counted perl-DB_File in the SLES12
and SLES15 requirements, so both checks failed in xcat_pr_test.

Expect perl(DB_File) on the SLES platforms and keep perl-DB_File for
the others. The test fails against the upstream master spec and
passes with this branch.

Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>
2026-10-29 14:36:35 -03:00
Daniel Hilst 7f4c7dbda0 Merge pull request #7983 from VersatusHPC/fix/vm-machine-type-unsupported-on-host
fix(xcat-test): libvirt refuses the compute node machine type pseries-rhel7.6.0
2026-10-29 14:32:40 -03:00
Daniel Hilst 5260c899a2 Merge pull request #7987 from VersatusHPC/fix/diskless-mount-snapshot-kernel-mounts
fix(xcat-test): the diskless cases fail when the kernel mounts binfmt_misc
2026-10-29 14:32:18 -03:00
Daniel Hilst d150233cfc Merge pull request #7988 from VersatusHPC/fix/genesis-sshd-privsep-upstream
fix(xcat-core): Genesis on Ubuntu answers no ssh, so every case that reaches a node fails
2026-10-29 14:31:56 -03:00
Daniel Hilst 213be1a5b0 Merge pull request #7989 from VersatusHPC/fix/selinux-avc-path-upstream
fix(xcat-test): a path that only starts with an xCAT directory name is counted as xCAT
2026-10-29 14:31:21 -03:00