While investigating another grub issue, I was confused by the path
taken in the logs reported, and noticed that on a ramdisk, we might
not actually have a valid response to os.path.ismount, I'm guessing
depending on what in memory filesystem is in use while also coupled
with attempting to check a filesystem.
Adds a test to validate that exceptions raised on these commands
where this issue can be encountered, are properly bypassed, and also
adds additional logging to make it easier to figure out what is
going on in the entire bootloader setup sequence.
Change-Id: Ibd3060bef2e56468ada6b1a5c1cc1632a42803c3
Check if the ESP is already mounted before attempting to mount it
for the bootloader installation.
Change-Id: Ifd738b2c5663f1a211d7e13b5ba386be631d8db1
Adds support to identify and utilize a CSV file to signal which
bootloader to utilize, and set it when the OS is running as opposed
to when EFI is running. This works around EFI loader potentially
crashing some vendors hardware types when entry stored in the
image does not match the EFI loader record which was utilzied to
boot.
Grub2+shim specifically specifically needs the CSV file name
and entry label to match what the system was booted with in order
to prevent the machine from potentially crashing.
See https://storyboard.openstack.org/#!/story/2008962
and https://bugzilla.redhat.com/show_bug.cgi?id=1966129#c37
for more information.
Change-Id: Ibf1ef4fe0764c0a6f1a39cb7eebc23ecc0ee177d
Story: 2008962
Task: 42598
Co-Authored-By: Bob Fournier <bfournie@redhat.com>
Recent releases of redhat grub2 will always fail when installing to
EFI paths, to encourage a transition to the signed shim bootloader.
Partition image deploys avoid calling grub2-install with the
preserve-efi-assets functions. Deploying whole disk images doesn't
require grub2-install. This leaves whole disk images installed onto
softraid devices, which still attempts to call grub2-install.
This change will still attempt to run grub2-install in this
one remaining case, but will ignore any failure.
A future enhancement can avoid calling grub2-install entirely so that
non-redhat secure-boot capable images can keep their signed
bootloaders.
Story: 2008923
Task: 42521
Change-Id: If432ef795d64d76442d739eb4f7d155ff847041e
We're currently requiring it twice: in image_info and in a separate
configdrive argument. I think we should eventually settle on separate
arguments for separate entities, so this change makes the value in
image_info optional with a goal to stop accepting it.
We could probably just remove the handling in image_info, but a
deprecation is safer.
The (unused in ironic) cache_image call is updated with an optional
configdrive arguments.
Story: #2008904
Task: #42480
Change-Id: I1e2efa28efa3ea7e389774cb7633d916757bc6ed
qemu-img attempts to launch multiple threads by default *and*
attempts to have multiple memory allocation arenas to operate
from. While multithreading can be good for performance, this
pattern and the memory footprint for process launch and
dependencies can turn the memory footprint for a cirros image
conversion (16MB) into 1.2GB of memory being asked for by the
qemu-img tool.
In order to limit this impact, as the default number of arenas
is governed by the number of CPUs times the number 8, it seems
reasonable to lower this to a more reasonable number which
also helps keep our possible memory footprint from being exceeded.
Change-Id: I71a28ec59ec31c691205eb34d9fcab63a2ccb682
Story: 2008928
Task: 42528
To make this function useful for purposes other than efibootmgr
entries, this change moves the path manipulation to _run_efibootmgr.
This change also adds boot*.efi entries to BOOTLOADERS_EFI so that it
includes every entry in the UEFI Spec 2.9[1] Table 3-2 UEFI Image
Types.
[1] https://uefi.org/sites/default/files/resources/UEFI_Spec_2_9_2021_03_18.pdf
Story: 2008923
Task: 42521
Change-Id: Ibe02786609aa0de65115897d8f4a9b4f36c8aed2
The command params can be huge when configdrive is used. There is no
point in sending them back, Ironic does not use them anyhow.
Story: #2008904
Task: #42479
Change-Id: I6e3db5db2042ca3fb5dafacfacf036fd7fc2fc4c
The _manage_uefi code has a check where it attempts to just
identify the precise partition number of the device, in order
for configuration to be parsed and passed. However, the same code
did not handle the existence of a `p1` partition instead of just a
partition #1. This is because the device naming format is different
with NVMe and Software RAID.
Likely, this wasn't an issue with software raid due to how complex the
code interaction is, but the docs also indicate to use only whole disk
images in that case.
This patch was pulled down my one RH's professional services folks
who has confirmed it does indeed fix the issue at hand. This is noted
as a public comment on the Red Hat bugzilla.
https://bugzilla.redhat.com/show_bug.cgi?id=1954096
Story: 2008881
Task: 42426
Related: rhbz#1954096
Change-Id: Ie3bd49add9a57fabbcdcbae4b73309066b620d02
The two functions work_on_disk and create_config_drive_partition contain
a substantial part of the deployment logic. Previously we placed them in
ironic-lib for re-using on the conductor side in the iSCSI deploy
interface. Since the iSCSI deploy is going away, we can move this code
to ironic-python-agent to simplify maintenance.
Imports code from ironic_lib commit 9fb5be348202f4854a455cd08f400ae12b99e1f2.
Change-Id: I6cbcd81533f135208b57746cb0e33ffdfaf94eee
For software RAID in UEFI mode, we create ESPs on all holder disks
and copy the bootloader there. Since there is no mechanism to keep
the ESPs in sync, e.g. on kernel upgrades or when kernel parameters
are updated, the ESPs will get out of sync eventually. This may lead
to a situation where a node boots with outdated parameters or does
not have any of the installed kernels in the boot menu anymore.
This change proposes to RAID the ESPs. While the UEFI firmware will
find an ESP partition (one leg of the mirror), the node will see
an md device and all subsequent updates will go to all member disks.
Also, remove the source ESP after copying in order to avoid mount
confusion (same UUID!).
Story: #2008745
Task: #42103
Change-Id: I9078ef37f1e94382c645ae98ce724ac9ed87c287
The line we're looking for is not there when IPA is in a container, at least
for CentOS based containers. Just fall back to sysrq on errors.
Change-Id: Ie4ee605ad9c6cda58808512a563247175859c71e
The root UUID changes after a streamed partition image is written to
the block device, causing later deployment failure when assuming the
old UUID.
This change updates the root UUID after streaming the partition image
is complete.
This issue may have been missed in local testing because deploying the
same image repeatedly will result in stable root UUID across runs.
Change-Id: Ice4630c16fc216980488d1427f3b02e1b8a417fa
IPA is not properly checking if the root partition is already
mounted. Device is being passed to os.path.ismount() instead
of the mount point.
Story: 2008631
Task: 41839
Change-Id: I37a6e7e6bbe0bbbb0317c6e55bb822dafe7cce20
It's somewhat confusing at the moment, since we're trying to find
a UEFI partition by UUID "None". Don't search for partition if
we don't know its UUID, and provide a better error message.
Change-Id: Ief874084132797a445ddae8009264712a05facfd
Previously, partition images were hard coded to be bios based
as opposed to consulting all of the values AND the node itself
before making the most appropriate determination. Now the agent
utilises the internal helper to properly determine the boot
mode when calling ironic-lib.
Story: 2008070
Task: 41265
Change-Id: Id5eeda69d5b9de2b393af414472d57b0d4380c43
The partition image support has been telling ironic-lib
that the machine will be local booted. While this is likely
harmless, and doesn't seem to break anythign, we should have
it match moving forward just to be on the safe side so we don't
accidently break things down the road.
Change-Id: I33e5d583964ef8c21aa04d7427bcd3957b89d449
Adds support for the EFI partition to be appended to fstab so the
filesystem can be automounted and EFI loader updated should the
deployed operating system need to do so.
This should enable bootloaders to be upgraded by linux based
operating systems after the instance has been deployed when
a partition image was utilized for the initial deployment.
Change-Id: Iec28a8841cc01ec8b01a3f5cca070c934c7a2531
Story: 2008070
Task: 40754
Partition images can sometimes contain a /boot folder structure
event he assets for EFI booting on that filesystem. Which is a
good thing. The conundrum is that Ironic does not handle this
properly and potentially replaces the bootloader in this sequence
such that grub2-install is used instead of signed bootloader assets.
As such, we should be preserving the assets and using them from
a partition image much like we do when we have a wholedisk
image and can identify the assets.
Now we will preserve the EFI boot assets, copy them to the new EFI
boot partition, and call the EFI setup methods to manage the EFI
nvram.
Note, this change also splits the logic path out that performs the
end call of the EFI boot manager into a reusable method but does
not retool all of the testing as it is intertwined in the
install_grub2 testing.
Also adds some additional debug logging, as much of the bootloader
installation code has multiple fallback/cleanup points which makes
it difficult to debug from logs.
Story: 2008070
Task: 40753
Change-Id: If17d4b4c06df5504987e61a1fde6662e9acd6989
Partition images through the agent have the unfortunate
side effect of being executed without full node context
by default. Luckilly we've had a similar problem and
cache the node.
This patch changes the lookup from a default of msdos
partitions to use the cached node object.
Change-Id: I002816c9372fdf1cc32f3c67f420073551479fd9
Some hardware is very well intentioned. However this intention
can result in the UEFI NVRAM table being full which prevents us
from adding new records to the table. We can't be sure what to
delete, so in this case some operators just need the ability to
tell ironic "it is okay if this fails, it will still work."
The added ``ignore_bootloader_failure`` option adds
this capability which can be set per-node either in the agent
configuation via the ramdisk image, or in the pxe_append_params
configuration parameter for the node itself with a
``ipa-ignore-bootloader-failure`` option in order to prevent
the failure from being raised.
Change-Id: If3c83fb2ea2025fce092d495a64f32077c70d2d6
Story: 2008386
Task: 41309
Add possibility to use disk LABEL to identify rootfs uuid for
Software RAID deployment
Change-Id: I77f36e70ddc539af0190db1c1abe0fb2c66f34b4
Story: 2008303
Task: 41188
By default, grub2-mkconfig scans everything to look for other
environments and then load those into the grub configuration.
It makes sense, but on newer versions of grub2 in distribution
images, os-prober is taking an exceptionally long time in some
cases where more than one storage device exists with other
filesystems.
As a result, of the os-prober execution by grub2-mkconfig, the
bootloader installation can completely time out and fail the
deployment. This is presently experienced with metalsmith on
centos8.
There are numerous sporatic reports of issues like this issue
where grub2-mkconfig hangs for some period of time, and this is
observable on Centos8.2 in our CI. While one report[0] mentions
this issue, Another bug [1] has the dialog that actually helps us
frame the context as to what we likely should do.
Also, fixes the unit testing so we actually test if we're running
with grub2. :\
[0]: https://bugzilla.redhat.com/show_bug.cgi?id=1744693
[1]: https://bugzilla.redhat.com/show_bug.cgi?id=1709682
Depends-On: https://review.opendev.org/#/c/748315
Change-Id: I14bf299afef3a1ddb2006fe5f182d7f0d249e734
Calling join() does not raise, we need to explicitly check the result.
Change-Id: I81d3d727af220c2b50358edab8139f07874611f0
Story: #2008240
Task: #41083
We don't have a really working way to detect root UUID for whole
disk images at the moment, which results in an ignored traceback
every time install_bootloader is called with whole disk images in
UEFI mode. Avoid it by skipping GRUB2 if root UUID is unknown.
Change-Id: I84245538f59c664b72d1cafbca8d61be0978f489
Or at least try to.
Some deployments just don't use root device hints, and this is okay.
However, other deployments need root device hints, and with fast
track mode in ramdisks, we created a situation where the node cache
could be updated by a human or software between the time the agent
was started, and the deployment was requested.
As a result, the agent has been updated to check if we have a hint
and if we don't, update the cache from the node lookup endpoint.
This is not needed when the inband deploy steps are executed, as
the process of updating the steps does force the node cache to be
updated.
Change-Id: I27201319f31cdc01605a3c5ae9ef4b4218e4a3f6
Story: 2008039
Task: 40701
When we added software raid support, we started calling bootloader
installation. As time went on, we ehnanced that code path for non
RAID cases in order to ensure that UEFI nvram was setup
for the instance to boot properly.
Somewhere in this process, we missed a possible failure case where
the iscsi client tgtadm may return failures. Obviously, the correct
path is to not call iscsi teardown if we don't need to.
Since it was always semi-opportunistic teardown, we can't blindly
catch any error, and if we started iSCSI and failed to tear the
connection down, we might want to still fail, so this change
moves the logic over to use a flag on the agent object which
one extension to set the flag and the other to read it and take
action based upon that.
Change-Id: Id3b1ae5e59282f4109f6246d5614d44c93aefa7c
Story: 2007937
Task: 40395
Shuffle some functions around and reduce size of _is_bootloader_loaded
moving logic out to a new function.
Change-Id: I9c10bf05186dcebb37f175d61bf4ac9ff86b6510
