Kernel arguments could only be set per profile, baked into the static
boot files of the profile. Nodes or groups that need extra arguments,
such as a different console or driver options, required a copy of the
whole profile.
When deployment.kernelargs is set, the iPXE boot offer now points at the
existing /confluent-api/boot/by-node endpoint, which serves the profile
boot.ipxe or boot.ipxe.shim with the node arguments appended to the
kernel line and the paths made absolute. UEFI HTTP boot takes the same
path, as it chainloads the iPXE shim when that is installed. A UEFI
HTTP boot that loads the static boot.img instead logs that the node
arguments are not applied.
Nodes without the attribute keep receiving the static profile files.
A client sends its Request to the server whose Advertise it picked,
named by the server identifier. confluent answered every Request, so
with several DHCPv6 servers on the segment, such as the members of a
collective, all of them replied to it.
IPv6 UEFI HTTP boot always served boot.img. Offer the iPXE shim when
the profile has boot.ipxe.shim and the shim of the booting
architecture is installed, as the IPv4 path does.
iPXE then sends its own DHCPv6 request with user class iPXE and a
PXEClient vendor class, which was refused as IPv6 PXE. Answer it with
the boot.ipxe.shim URL, or boot.ipxe when the profile has no shim
script. With only the other configuration flag in the router
advertisement, iPXE sends an Information-request, which was not
handled at all and carries no IA_NA. Handle it and only assign an
address when the request asks for one.
Before it starts the shim, the firmware releases the address and
retransmits the Release while it is unanswered, which delayed the
handoff by over a minute. Answer a Release sent to this server.
UEFI HTTP boot chainloads the iPXE shim only when it is installed, but
the check always looked for the x86_64 shim. An aarch64 node on a
server with only the x86_64 shim was offered the missing aarch64 one
and failed to boot, and on a server with only the aarch64 shim it got
the static boot.img instead.
On Debian and Ubuntu 18.04 the installedargs line of profile.yaml was
added to GRUB_CMDLINE_LINUX with its "installedargs:" key still in
front, and update-grub was not run for it, so the arguments only took
effect with a later grub update, as a broken first argument.
All Debian and Ubuntu profiles also put the arguments into a sed
replacement unescaped, so an argument containing a slash, such as
rd.lvm.lv=vg/root, made sed fail and left the grub defaults unchanged.
SUSE 16 put the arguments into a JSON string unescaped, so an argument
containing a double quote or backslash produced an invalid
autoinstall.json, and one containing "!" or "&" broke the sed command
that fills it in.
SUSE 15 put them into the AutoYaST XML unescaped, so an argument
containing "&" or "<" made the whole profile invalid and stopped the
install.
When the image has no /etc/default/grub, as images built by imgutil do
not, installing it to disk on EL9 and EL10 wrote one with
rd.lvm.lv=vg/root rd.lvm.lv=vg/swap, volumes that do not exist. A
missing /etc/kernel/cmdline likewise got root=/dev/mapper/localstorage-root,
which is wrong for an install without LVM or with a volume group named
after the node. The boot loader entries were rewritten correctly, so
the first boot worked, but grub2-mkconfig --update-bls-cmdline or a new
kernel would take the wrong names from these files.
Both defaults now name the root device and the logical volumes for root
and swap that the install created.
A Giga Computing BMC reports the system UUID as null, which made
nodeinventory fail with 'NoneType' object has no attribute 'lower'.
Treat a null UUID like a missing one.
A SOL keepalive reply arriving after the console was closed failed with
an AttributeError on the released session. A privilege level request on
a session that was logged out or given up on during login raised
IpmiException out of the packet handler into the event loop shared by
all sessions. Ignore the late keepalive reply, and report the failed
login to its waiters instead.
A request sent from within a console session, such as the power state
check confetty makes, went through a handler that logged every failure
as an unexpected 500 with a stack trace. A node with an unreachable BMC
filled the trace log with TargetEndpointUnreachable. Share the error
reporting of the main request loop, so the client gets the real error
code and only unexpected failures are traced.
A client that reset the connection or closed it before finishing the
exchange left a stack trace in the trace log (ConnectionResetError, or
an IndexError on an empty read). Read each field in full, and treat a
client going away as the end of that exchange rather than a server fault.
Every other OS labels its stock profile "(Default Profile)", including
SUSE 15, whose server profile also has the directory name server. SUSE
16 said "(Server Profile)" instead.
It is verified that iPXE leverages the UEFI CA and does not let untrusted connections.
So let it do https as well as http, relegating boot.img for virtual media instead.
This like the plain http boot works nicely with a broader set of UEFI.
Some implementations claim to have a system, but have useless data.
We can't tell programattically that the data is useless, so provide chassis for the user to decide.
Since we have a Seccomp cBPF going anyway, add insmod/modprobe operations to block.
Also reframe things to be a bit more readable using enum and less magical numbers.
image2disk only took the boot UUID from a /boot volume. Images captured
from a system without a separate /boot left it unset, and rewriting the
EFI stub grub.cfg failed with "TypeError: sequence item 5: expected str
instance, NoneType found". Fall back to the UUID of the / filesystem,
which holds /boot in that case.
Profile scripts had to call apiclient and parse /self/myattribs to read
a node attribute. confluent_attrib prints one attribute, comma joining
lists, and returns 1 when it is unset. Non-shell scripts can run it as
'bash /etc/confluent/functions confluent_attrib <name>'.
groups is stored as a plain list rather than a value dict, so the
value check skipped it and a node could not see its own groups. Return
it as a list, the same way ansible plays get it as host vars.
nodedefine with a noderange passed one dict for all nodes, so they
shared one stored groups list. Later group membership changes then
updated inheritance for only the first node, leaving the others with
values inherited from groups they no longer belong to.
post.sh runs post.custom, which the profiles lacked, so every install
to disk logged a failed download. firstboot.sh copied
/etc/confluent/tls/*.pem, which diskless nodes never have: installimage
already adds the confluent CA to the ca-trust anchors.