get_health looked only at each node's State, so an Enabled node whose
Health was Warning or Critical left the chassis reported as ok. It also
skipped a node whose resource could not be read, counting it as
healthy while an unreadable chassis was flagged. Report the node's
Health when it is not OK, and a node that cannot be read as
Unreachable.
The sensor collection has 668 members, and reading each one on its own
takes about 92 seconds and rebuilds on every nodesensors call. Opt in
to $expand=. for that collection, checking once whether the firmware
really inlines the members, so firmware that ignores $expand keeps the
per-member reads.
ForceRestart only restarts the host, so the node BMC kept running and
a reseat did not recover a hung one. The EUREKA firmware provides a
Reseat reset type that removes all power from the slot.
A node whose BMC is not reporting returns Reading 0 from its
temperature sensors, dragging averages down with meaningless
values. The ComputerSystem Oem data flags this via HasBMCMetrics;
skip such nodes, and keep collecting when the flag is absent.
get_health collected an issues list but returned it nowhere, and
flattened every problem to Warning. Return the findings as
badreadings using SensorReading, honor verbose, and map chassis
health through _healthmap so Critical is no longer downgraded.
The sensor URL was built as BMC{N}CpuCPU{X}Temp instead of
BMC{N}CPU{X}Temp, so every request returned 404. The entries also
referenced const.SensorUnits, which does not exist, and used a
dict shape the only consumer, get_average_processor_temperature,
cannot read - it expects thermal-style dicts with ReadingCelsius.
ManagedBy is optional, and when a system does not link a manager
get_bmcurl() answers None. Most callers passed that straight into a
web request, which failed with "Constructor parameter should be str"
from the url library. Route those callers through bmcinfo(), which now
raises UnsupportedFunctionality so confluent reports it plainly. The
event log falls through to its system and chassis fallback, and
list_media lists nothing, since neither needs a manager.
SensorReading computed health and states with .get() and then
overwrote both with a strict lookup, so a sensor whose Status lacks
Health, or carries a value outside the health map, raised KeyError
and aborted the whole sensor listing. Keep the tolerant lookup, and
drop the states of a reading that is OK, as the copy in command.py
already does.
init_sdr assigned self._sdr before initialize() ran, so a failure left the
half built object in the cache. The next call saw a non-None _sdr and handed
back that partial repository rather than trying again.
The visible symptom is a first call raising and the second appearing to
succeed. The real cost is on a bmc where the read fails once: the client
keeps the incomplete sdr for the life of the session and every later sensor
lookup answers from it without complaint.
get_webclient falls off its end when /api/login answers anything but 200, so
it returned None. wc() passes that back, and thirty of the thirty-four call
sites use it unchecked, so a refused login arrived as "'NoneType' object has
no attribute 'grab_json_response'" from wherever it landed.
Raised where the failure is known, and with the status: 404 is firmware with
no web api, or a Redfish-only capture of one, while 401 is credentials it
will not take. Neither was distinguishable before.
The other four call sites are the inventory reads, and they always did check.
They answer partially when the web interface is out of reach, which is why
nodeinventory still says something useful. They ask through wc_if_available
now, so that tolerance is stated rather than resting on a None.
Two places where the code that exists to explain a failure fails instead, and
the caller is shown the second failure rather than the first.
_do_web_request builds its message from the response body when that body is
not the JSON error document the spec asks for. An html 404 page is exactly
that, the body is bytes, and str + bytes raised TypeError. The status and the
body never reached anyone.
LenovoFirmwareConfig raised a bare Exception when python-lxml and
python-eficompressor are absent. Confluent has no handler for one, so a
missing dependency showed as "Unexpected Error" and hid a message that
already said what to install.
Neither changes what fails, only what the caller is told.
Same function, one line above. EthernetInterfaces is optional, and when it is
absent the None went into a request and raised TypeError from the url library
rather than saying what was missing.
Kept beside the ambiguous case because the two are one question asked twice.
Dedicated plus shared is how BMCs are built, so several NICs is ordinary.
_get_bmc_nic_url only reaches the count when the address the session came in
on matches none of them, which is what a tunnel or a NAT does, and it then
raised the bare PyghmiException base class.
Confluent has no handler for the base class, so it fell through to the
generic one and showed "Unexpected Error" plus a traceback, for a machine
doing nothing wrong. UnsupportedFunctionality now, which the redfish plugin
already reports plainly and which stays inside PyghmiException.
The message said "does not have exactly one interface" without saying how
many, which ones, or what to do. Every caller takes a name, so it lists the
candidates, and the empty case reads differently from the ambiguous one.
A Redfish service may publish no Systems collection, and power and cooling
equipment does exactly that. sysurl is then None, and get_power and
get_bootdev handed it straight to a request, raising TypeError from inside
the url library.
sysinfo already guarded the same field. Both now ask through _system_url and
get a refusal naming what is missing. DMTF publish three services of this
shape, which is why they sit commented out in inventory-dmtf.yaml.
PCIeDevices and PCIeFunctions are optional, and get_health indexed both
without checking. _get_adp_urls in the same file already spells it
.get('PCIeDevices', []), so these two were the outliers.
Power and cooling equipment has no PCIe at all. The KeyError escaped the
health read and reached the user as "Unexpected Error" with a traceback
behind it. Reproduces offline against DMTF's public-rackmount1 mockup.
parse_time read the digits after the decimal point as whole milliseconds, so
'.5' became 5ms instead of 500ms. Only a three digit fraction came out right,
and a BMC may write either.
'.5', '.50' and '.500' are all 500ms now, '.125' is 125ms. Every other format
parse_time accepts is untouched.
A read only ever looked at the manager's log services, and fell back to the
system's when the manager published none. A platform that keeps an event log
in both places had the second one invisible: an AMI MegaRAC keeps power unit
and thermal events in a chassis log that nothing read, 103 records that no
command could reach.
Clearing deliberately does not follow. It stays where it was, so a log that
only a read reaches is never destroyed by one, and clearing a platform that
keeps its only event log on the system still works.
The name test now ignores spacing, since a build that calls its post code log
"BIOS POST Code Log" was read as an event log and merged 2719 post codes in.
Reading an event log skipped every log service whose id or name said journal,
dump, post code, host logger or crash, on every implementation. Those names
are bmcweb's: the AMI and Lenovo bmcs call theirs SEL, EventLog, AuditLog and
PlatformLog. bmcweb does need the distinction, keeping its event log under the
system while a clear would destroy its dumps, so it gets a handler that names
the words and generic names none, reading whatever a platform publishes.
Generic added an UpdateParameters part to every multipart firmware push,
because the specification has one carried. Only the AMI firmware was seen to
insist on it, so name it in that handler and let generic send what the caller
passed.
A bmc behind a forward answers Activate Payload with the port it listens on
itself, and the advertised-port check refused that, so a console failed where
command traffic worked. The advertised port is never sent to, so the check now
applies only on the default port. A bmc on another port advertising a third one
is no longer refused outright, which nothing here could have served anyway.
The deadline was set when monitoring began, so a flash that kept
answering for longer than the grace period had already spent it by the
time the reboot it covers arrived, and reported a working update as a
failure.
A device holds its sensor records per lun and scopes the reservation the
same way, so a token taken on lun 0 can be refused for lun 1 with 0xc5.
The retry then took the same wrong token again without end.
Firmware for a supply or a fan matched no fragment and so was called
core, and nothing could ever answer for misc.
Only the collections that mean one thing are matched by url. A Storage
resource is the subsystem, so its firmware is the controller rather than
a drive, and a Processor is a cpu as readily as an accelerator, so that
one is decided by asking the processor what it is.
Jarrod Johnson
Markus Hilger
Jarrod Johnson