[Skiboot] skibot 6.0 released!
Stewart Smith
stewart at linux.ibm.com
Sat May 12 08:03:44 AEST 2018
skiboot-6.0
***********
skiboot v6.0 was released on Friday May 11th 2018. It is the first
release of skiboot 6.0, which is the new stable release of skiboot
following the 5.11 release, first released April 6th 2018.
Skiboot 6.0 is the basis for op-build v2.0 and will is *required* for
POWER9 systems.
skiboot v6.0 contains all bug fixes as of skiboot-5.11,
skiboot-5.10.5, and *skiboot-5.4.9* (the currently maintained stable
releases). We do *not* expect any further stable releases in the
5.10.x series, nor in the 5.11.x series.
For how the skiboot stable releases work, see Skiboot stable tree
rules and releases for details.
Over skiboot-5.11, we have the following changes:
New Features
============
Since 6.0-rc1:
* Update default stop-state-disable mask to cut only stop11
Stability improvements in microcode for stop4/stop5 are available in
upstream hcode images. Stop4 and stop5 can be safely enabled by
default.
Use ~0xE0000000 to cut all but stop0,1,2 in case there are any
issues with stop4/5.
example:
nvram -p ibm,skiboot --update-config opal-stop-state-disable-mask=0x1FFFFFFF
**Note**: that DD2.1 chips that have a frequency <1867Mhz possible
*need* to run a hcode image *different* than the default in op-build
(set *BR2_HCODE_LATEST_VERSION=y* in your config)
* ibm,firmware-versions: add hcode to device tree
op-build commit 736a08b996e292a449c4996edb264011dfe56a40 added hcode
to the VERSION partition, let’s parse it out and let the user know.
* ipmi: Add BMC firmware version to device tree
BMC Get device ID command gives BMC firmware version details. Lets
add this to device tree. User space tools will use this information
to display BMC version details.
Since 5.11:
* Disable stop states from OPAL
On ZZ, stop4,5,11 are enabled for PowerVM, even though doing so may
cause problems with OPAL due to bugs in hcode.
For other platforms, this isn’t so much of an issue as we can just
control stop states by the MRW. However the rebuild-the-world
approach to changing values there is a bit annoying if you just want
to rule out a specific stop state from being problematic.
Provide an nvram option to override what’s disabled in OPAL.
The OPAL mask is currently ~0xE0000000 (i.e. all but stop 0,1,2)
You can set an NVRAM override with:
nvram -p ibm,skiboot --update-config opal-stop-state-disable-mask=0xFFFFFFF
This nvram override will disable *all* stop states.
* interrupts: Create an “interrupts” property in the OPAL node
Deprecate the old “opal-interrupts”, it’s still there, but the new
property follows the standard and allow us to specify whether an
interrupt is level or edge sensitive.
Similarly create “interrupt-names” whose content is identical to
“opal-interrupts-names”.
* SBE: Add timer support on POWER9
SBE on P9 provides one shot programmable timer facility. We can use
this to implement OPAL timers and hence limit the reliance on the
Linux heartbeat (similar to HW timer facility provided by SLW on
P8).
* Add SBE driver support
SBE (Self Boot Engine) on P9 has two different jobs: - Boot the chip
up to the point the core is functional - Provide various services
like timer, scom, stash MPIPL, etc., at runtime
We will use SBE for various purposes like timer, MPIPL, etc.
* opal:hmi: Add missing processor recovery reason string.
With this patch now we see reason string printed for CORE_WOF[43]
bit.
[ 477.352234986,7] HMI: [Loc: U78D3.001.WZS004A-P1-C48]: P:8 C:22 T:3: Processor recovery occurred.
[ 477.352240742,7] HMI: Core WOF = 0x0000000000100000 recovered error:
[ 477.352242181,7] HMI: PC - Thread hang recovery
* Add DIMM actual speed to device tree
Recent HDAT provides DIMM actuall speed. Lets add this to device
tree.
* Fix DIMM size property
Today we parse vpd blob to get DIMM size information. This is
limited to FSP based system. HDAT provides DIMM size value. Lets use
that to populate device tree. So that we can get size information on
BMC based system as well.
* PCI: Set slot power limit when supported
The PCIe slot capability can be implemented in a root or switch
downstream port to set the maximum power a card is allowed to draw
from the system. This patch adds support for setting the power limit
when the platform has defined one.
* hdata/spira: parse vpd to add part-number and serial-number to
xscom@ node
Expected by FWTS and associates our processor with the part/serial
number, which is obviously a good thing for one’s own sanity.
Improved HMI Handling
---------------------
* opal/hmi: Add documentation for opal_handle_hmi2 call
* opal/hmi: Generate hmi event for recovered HDEC parity error.
* opal/hmi: check thread 0 tfmr to validate latched tfmr errors.
Due to P9 errata, HDEC parity and TB residue errors are latched for
non-zero threads 1-3 even if they are cleared. But these are not
latched on thread 0. Hence, use xscom SCOMC/SCOMD to read thread 0
tfmr value and ignore them on non-zero threads if they are not
present on thread 0.
* opal/hmi: Print additional debug information in rendezvous.
* opal/hmi: Fix handling of TFMR parity/corrupt error.
While testing TFMR parity/corrupt error it has been observed that
HMIs are delivered twice for this error
* First time HMI is delivered with HMER[4,5]=1 and TFMR[60]=1.
* Second time HMI is delivered with HMER[4,5]=1 and TFMR[60]=0
with valid TB.
On second HMI we end up throwing “HMI: TB invalid without core error
reported” even though TB is in a valid state.
* opal/hmi: Stop flooding HMI event for TOD errors.
Fix the issue where every thread on the chip sends HMI event to host
for TOD errors. TOD errors are reported to all the core/threads on
the chip. Any one thread can fix the error and send event. Rest of
the threads don’t need to send HMI event unnecessarily.
* opal/hmi: Fix soft lockups during TOD errors
There are some TOD errors which do not affect working of TOD and TB.
They stay in valid state. Hence we don’t need rendez vous for TOD
errors that does not affect TB working.
TOD errors that affects TOD/TB will report a global error on
TFMR[44] alongwith bit 51, and they will go in rendez vous path as
expected.
But the TOD errors that does not affect TB register sets only TFMR
bit 51. The TFMR bit 51 is cleared when any single thread clears the
TOD error. Once cleared, the bit 51 is reflected to all the cores on
that chip. Any thread that reads the TFMR register after the error
is cleared will see TFMR bit 51 reset. Hence the threads that see
TFMR[51]=1, falls through rendez-vous path and threads that see
TFMR[51]=0, returns doing nothing. This ends up in a soft lockups in
host kernel.
This patch fixes this issue by not considering TOD interrupt
(TFMR[51]) as a core-global error and hence avoiding rendez-vous
path completely. Instead threads that see TFMR[51]=1 will now take
different path that just do the TOD error recovery.
* opal/hmi: Do not send HMI event if no errors are found.
For TOD errors, all the cores in the chip get HMIs. Any one thread
from any core can fix the issue and TFMR will have error conditions
cleared. Rest of the threads need take any action if TOD errors are
already cleared. Hence thread 0 of every core should get a fresh
copy of TFMR before going ahead recovery path. Initialize recover =
-1, so that if no errors found that thread need not send a HMI event
to linux. This helps in stop flooding host with hmi event by every
thread even there are no errors found.
* opal/hmi: Initialize the hmi event with old value of HMER.
Do this before we check for TFAC errors. Otherwise the event at host
console shows no error reported in HMER register.
Without this patch the console event show HMER with all zeros
[ 216.753417] Severe Hypervisor Maintenance interrupt [Recovered]
[ 216.753498] Error detail: Timer facility experienced an error
[ 216.753509] HMER: 0000000000000000
[ 216.753518] TFMR: 3c12000870e04000
After this patch it shows old HMER values on host console:
[ 2237.652533] Severe Hypervisor Maintenance interrupt [Recovered]
[ 2237.652651] Error detail: Timer facility experienced an error
[ 2237.652766] HMER: 0840000000000000
[ 2237.652837] TFMR: 3c12000870e04000
* opal/hmi: Rework HMI handling of TFAC errors
This patch reworks the HMI handling for TFAC errors by introducing 4
rendez-vous points improve the thread synchronization while handling
timebase errors that requires all thread to clear dirty data from
TB/HDEC register before clearing the errors.
* opal/hmi: Don’t bother passing HMER to pre-recovery cleanup
The test for TFAC error is now redundant so we remove it and remove
the HMER argument.
* opal/hmi: Move timer related error handling to a separate function
Currently no functional change. This is a first step to completely
rewriting how these things are handled.
* opal/hmi: Add a new opal_handle_hmi2 that returns direct info to
Linux
It returns a 64-bit flags mask currently set to provide info about
which timer facilities were lost, and whether an event was
generated.
* opal/hmi: Remove races in clearing HMER
Writing to HMER acts as an “AND”. The current code writes back the
value we originally read with the bits we handled cleared. This is
racy, if a new bit gets set in HW after the original read, we’ll end
up clearing it without handling it.
Instead, use an all 1’s mask with only the bit handled cleared.
* opal/hmi: Don’t re-read HMER multiple times
We want to make sure all reporting and actions are based upon the
same snapshot of HMER in case bits get added by HW while we are in
OPAL.
libflash and ffspart
--------------------
Many improvements to the *ffspart* utility and *libflash* have come in
this release, making *ffspart* suitable for building bit-identical
PNOR images as the existing tooling used by *op-build*. The plan is to
switch *op-build* to use this infrastructure in the not too distant
future.
* libflash/blocklevel: Make read/write be ECC agnostic for callers
The blocklevel abstraction allows for regions of the backing store
to be marked as ECC protected so that blocklevel can decode/encode
the ECC bytes into the buffer automatically without the caller
having to be ECC aware.
Unfortunately this abstraction is far from perfect, this is only
useful if reads and writes are performed at the start of the ECC
region or in some circumstances at an ECC aligned position - which
requires the caller be aware of the ECC regions.
The problem that has arisen is that the blocklevel abstraction is
initialised somewhere but when it is later called the caller is
unaware if ECC exists in the region it wants to arbitrarily read and
write to. This should not have been a problem since blocklevel
knows. Currently misaligned reads will fail ECC checks and
misaligned writes will overwrite ECC bytes and the backing store
will become corrupted.
This patch add the smarts to blocklevel_read() and
blocklevel_write() to cope with the problem. Note that ECC can
always be bypassed by calling blocklevel_raw_() functions.
All this work means that the gard tool can can safely call
blocklevel_read() and blocklevel_write() and as long as the
blocklevel knows of the presence of ECC then it will deal with all
cases.
This also commit removes code in the gard tool which compensated for
inadequacies no longer present in blocklevel.
* libflash/blocklevel: Return region start from ecc_protected()
Currently all ecc_protected() does is say if a region is ECC
protected or not. Knowing a region is ECC protected is one thing but
there isn’t much that can be done afterwards if this is the only
known fact. A lot more can be done if the caller is told where the
ECC region begins.
Knowing where the ECC region start it allows to caller to align its
read/and writes. This allows for more flexibility calling read and
write without knowing exactly how the backing store is organised.
* libflash/ecc: Add helpers to align a position within an ecc buffer
As part of ongoing work to make ECC invisible to higher levels up
the stack this function converts a ‘position’ which should be ECC
agnostic to the equivalent position within an ECC region starting at
a specified location.
* libflash/ecc: Add functions to deal with unaligned ECC memcpy
* external/ffspart: Improve error output
* libffs: Fix bad checks for partition overlap
Not all TOCs are written at zero
* libflash/libffs: Allow caller to specifiy header partition
An FFS TOC is comprised of two parts. A small header which has a
magic and very minimmal information about the TOC which will be
common to all partitions, things like number of patritions, block
sizes and the like. Following this small header are a series of
entries. Importantly there is always an entry which encompases the
TOC its self, this is usually called the ‘part’ partition.
Currently libffs always assumes that the ‘part’ partition is at
zero. While there is always a TOC and zero there doesn’t actually
have to be. PNORs may have multiple TOCs within them, therefore
libffs needs to be flexible enough to allow callers to specify TOCs
not at zero.
The ‘part’ partition is otherwise a regular partition which may have
flags associated with it. libffs should allow the user to set the
flags for the ‘part’ partition.
This patch achieves both by allowing the caller to specify the
‘part’ partition. The caller can not and libffs will provide a
sensible default.
* libflash/libffs: Refcount ffs entries
Currently consumers can add an new ffs entry to multiple headers,
this is fine but freeing any of the headers will cause the entry to
be freed, this causes double free problems.
Even if only one header is uses, the consumer of the library still
has a reference to the entry, which they may well reuse at some
other point.
libffs will now refcount entries and only free when there are no
more references.
This patch also removes the pointless return value of ffs_hdr_free()
* libflash/libffs: Switch to storing header entries in an array
Since the libffs no longer needs to sort the entries as they get
added it makes little sense to have the complexity of a linked list
when an array will suffice.
* libflash/libffs: Remove backup partition from TOC generation code
It turns out this code was messy and not all that reliable. Doing it
at the library level adds complexity to the library and restrictions
to the caller.
A simpler approach can be achived with the just instantiating
multiple ffs_header structures pointing to different parts of the
same file.
* libflash/libffs: Remove the ‘sides’ from the FFS TOC generation
code
It turns out this code was messy and not all that reliable. Doing it
at the library level adds complexity to the library and restrictions
to the caller.
A simpler approach can be achived with the just instantiating
multiple ffs_header structures pointing to different parts of the
same file.
* libflash/libffs: Always add entries to the end of the TOC
It turns out that sorted order isn’t the best idea. This removes
flexibility from the caller. If the user wants their partitions in
sorted order, they should insert them in sorted order.
* external/ffspart: Remove side, order and backup options
These options are currently flakey in libflash/libffs so there isn’t
much point to being able to use them in ffspart.
Future reworks planned for libflash/libffs will render these options
redundant anyway.
* libflash/libffs: ffs_close() should use ffs_hdr_free()
* libflash/libffs: Add setter for a partitions actual size
* pflash: Use ffs_entry_user_to_string() to standardise flag strings
* libffs: Standardise ffs partition flags
It seems we’ve developed a character respresentation for ffs
partition flags. Currently only pflash really prints them so it
hasn’t been a problem but now ffspart wants to read them in from
user input.
It is important that what libffs reads and what pflash prints remain
consistent, we should move the code into libffs to avoid problems.
* external/ffspart: Allow # comments in input file
p9dsu Platform changes
======================
The p9dsu platform from SuperMicro (also known as ‘Boston’) has
received a number of updates, and the patches once carried by
SuperMicro are now upstream.
Since 6.0-rc1:
* p9dsu: timeout for variant detection, default to 2uess
Since 5.11:
* p9dsu: detect p9dsu variant even when hostboot doesn’t tell us
The SuperMicro BMC can tell us what riser type we have, which
dictates the PCI slot tables. Usually, in an environment that a
customer would experience, Hostboot will do the query with an SMC
specific patch (not upstream as there’s no platform specific code in
hostboot) and skiboot knows what variant it is based on the
compatible string.
However, if you’re using upstream hostboot, you only get the bare
‘p9dsu’ compatible type. We can work around this by asking the BMC
ourselves and setting the slot table appropriately. We do this
syncronously in platform init so that we don’t start probing PCI
before we setup the slot table.
* p9dsu: add slot power limit.
* p9dsu: add pci slot table for Boston LC 1U/2U and Boston LA/ESS.
* p9dsu HACK: fix system-vpd eeprom
* p9dsu: change esel command from AMI to IBM 0x3a.
ZZ Platform Changes
===================
* hdata/i2c: Fix up pci hotplug labels
These labels are used on the devices used to do PCIe slot power
control for implementing PCIe hotplug. I’m not sure how they ended
up as “eeprom-pgood” and “eeprom-controller” since that doesn’t make
any sense.
* hdata/i2c: Ignore multi-port I2C devices
Recent FSP firmware builds add support for multi-port I2C devices
such as the GPIO expanders used for the presence detect of OpenCAPI
devices and the PCIe hotplug controllers used to power cycle PCIe
slots on ZZ.
The OpenCAPI driver inside of skiboot currently uses a platform-
specific method to talk to the relevant I2C device rather than
relying on HDAT since not all platforms correctly report the I2C
devices (hello Zaius). Additionally the nature of multi-port devices
require that we a device specific handler so that we generate the
correct DT bindings. Currently we don’t and there is no immediate
need for this support so just ignore the multi-port devices for now.
* hdata/i2c: Replace *i2c_* prefix with *dev_*
The current naming scheme makes it easy to conflate “i2cm_port” and
“i2c_port.” The latter is used to describe multi-port I2C devices
such as GPIO expanders and multi-channel PCIe hotplug controllers.
Rename i2c_port to dev_port to make the two a bit more distinct.
Also rename i2c_addr to dev_addr for consistency.
* hdata/i2c: Ignore CFAM I2C master
Recent FSP firmware builds put in information about the CFAM I2C
master in addition the to host I2C masters accessible via XSCOM.
Odds are this information should not be there since there’s no
handshaking between the FSP/BMC and the host over who controls that
I2C master, but it is so we need to deal with it.
This patch adds filtering to the HDAT parser so it ignores the CFAM
I2C master. Without this it will create a bogus i2cm@<addr> which
migh cause issues.
* ZZ: hw/imc: Add support to load imc catalog lid file
Add support to load the imc catalog from a lid file packaged as part
of the system firmware. Lid number allocated is 0x80f00103.lid.
Bugs Fixed
==========
Since 6.0-rc2:
* core/opal: Fix recursion check in opal_run_pollers()
An earlier commit introduced a counter variable poller_recursion to
limit to the number number of error messages shown when opal_pollers
are run recursively. However the check for the counter value was
placed in a way that the poller recursion was only detected first 16
times and then allowed afterwards.
This patch fixes this by moving the check for the counter value
inside the conditional branch with some re-factoring so that
opal_poller recursion is not erroneously allowed after
poll_recursion is detected first 16 times.
* phb4: Print WOF registers on fence detect
Without the WOF registers it’s hard to figure out what went wrong
first, so print those when we print the FIRs when a fence is
detected.
* p9dsu: detect variant in init only if probe fails to found.
Currently the slot table init happens twice in both probe and init
functions due to the variant detection logic called with in-correct
condition check.
Since 6.0-rc1:
* core/direct-controls: improve p9_stop_thread error handling
p9_stop_thread should fail the operation if it finds the thread was
already quiescd. This implies something else is doing direct
controls on the thread (e.g., pdbg) or there is some exceptional
condition we don’t know how to deal with. Proceeding here would
cause things to trample on each other, for example the hard lockup
watchdog trying to send a sreset to the core while it is stopped for
debugging with pdbg will end in tears.
If p9_stop_thread times out waiting for the thread to quiesce, do
not hit it with a core_start direct control, because we don’t know
what state things are in and doing more things at this point is
worse than doing nothing. There is no good recipe described in the
workbook to de-assert the core_stop control if it fails to quiesce
the thread. After timing out here, the thread may eventually quiesce
and get stuck, but that’s simpler to debug than undefied behaviour.
* core/direct-controls: fix p9_cont_thread for stopped/inactive
threads
Firstly, p9_cont_thread should check that the thread actually was
quiesced before it tries to resume it. Anything could happen if we
try this from an arbitrary thread state.
Then when resuming a quiesced thread that is inactive or stopped (in
a stop idle state), we must not send a core_start direct control,
clear_maint must be used in these cases.
* hmi: Clear unknown debug trigger
On some systems, seeing hangs like this when Linux starts:
[ 170.027252763,5] OCC: All Chip Rdy after 0 ms
[ 170.062930145,5] INIT: Starting kernel at 0x20011000, fdt at 0x30ae0530 366247 bytes)
[ 171.238270428,5] OPAL: Switch to little-endian OS
If you look at the in memory skiboot console (or do *nvram -p
ibm,skiboot –update-config log-level-driver=7*) we see the console
get spammed with:
[ 5209.109790675,7] HMI: Received HMI interrupt: HMER = 0x0000400000000000
[ 5209.109792716,7] HMI: Received HMI interrupt: HMER = 0x0000400000000000
[ 5209.109794695,7] HMI: Received HMI interrupt: HMER = 0x0000400000000000
[ 5209.109796689,7] HMI: Received HMI interrupt: HMER = 0x0000400000000000
We’re taking the debug trigger (bit 17) early on, before the
hmi_debug_trigger function in the kernel is set up.
This clears the HMI in Skiboot and reports to the kernel instead of
bringing down the machine.
* core/hmi: assign flags=0 in case nothing set by
handle_hmi_exception
Theoretically we could have returned junk to the OS in this
parameter.
* SLW: Fix mambo boot to use stop states
After commit 35c66b8ce5a2 (“SLW: Move MAMBO simulator checks to
slw_init”), mambo boot no longer calls
add_cpu_idle_state_properties() and as such we never enable stop
states.
After adding the call back, we get more testing coverage as well as
faster mambo SMT boots.
* phb4: Hardware init updates
CFG Write Request Timeout was incorrectly set to informational and
not fatal for both non-CAPI and CAPI, so set it to fatal. This was
a mistake in the specification. Correcting this fixes a niche bug
in escalation (which is necessary on pre-DD2.2) that can cause a
checkstop due to a NCU timeout.
In addition, set the values in the timeout control registers to
match. This fixes an extremely rare and unreproducible bug, though
the current timings don’t make sense since they’re higher than the
NCU timeout (16) which will checkstop the machine anyway.
* SLW: quieten ‘Configuring self-restore’ for DARN,NCU_SPEC_BAR and
HRMOR
Since 5.11:
* core: Fix iteration condition to skip garded cpu
* uart: fix uart_opal_flush to take console lock over uart_con_flush
This bug meant that OPAL_CONSOLE_FLUSH didn’t take the appropriate
locks. Luckily, since this call is only currently used in the crash
path.
* xive: fix missing unlock in error path
* OPAL_PCI_SET_POWER_STATE: fix locking in error paths
Otherwise we could exit OPAL holding locks, potentially leading to
all sorts of problems later on.
* hw/slw: Don’t assert on a unknown chip
For some reason skiboot populates nodes in /cpus/ for the cores on
chips that are deconfigured. As a result Linux includes the threads
of those cores in it’s set of possible CPUs in the system and
attempts to set the SPR values that should be used when waking a
thread from a deep sleep state.
However, in the case where we have deconfigured chip we don’t create
a xscom node for that chip and as a result we don’t have a proc_chip
structure for that chip either. In turn, this results in an
assertion failure when calling opal_slw_set_reg() since it expects
the chip structure to exist. Fix this up and print an error instead.
* opal/hmi: Generate one event per core for processor recovery.
Processor recovery is per core error. All threads on that core
receive HMI. All threads don’t need to generate HMI event for same
error.
Let thread 0 only generate the event.
* sensors: Dont add DTS sensors when OCC inband sensors are
available
There are two sets of core temperature sensors today. One is DTS
scom based core temperature sensors and the second group is the
sensors provided by OCC. DTS is the highest temperature among the
different temperature zones in the core while OCC core temperature
sensors are the average temperature of the core. DTS sensors are
read directly by the host by SCOMing the DTS sensors while OCC
sensors are read and updated by OCC to main memory.
Reading DTS sensors by SCOMing is a heavy and slower operation as
compared to reading OCC sensors which is as good as reading memory.
So dont add DTS sensors when OCC sensors are available.
* core/fast-reboot: Increase timeout for dctl sreset to 1sec
Direct control xscom can take more time to complete. We seem to wait
too little on Boston failing fast-reboot for no good reason.
Increase timeout to 1 sec as a reasonable value for sreset to be
delivered and core to start executing instructions.
* occ: sensors-groups: Add DT properties to mark HWMON sensor groups
Fix the sensor type to match HWMON sensor types. Add compatible flag
to indicate the environmental sensor groups so that operations on
these groups can be handled by HWMON linux interface.
* core: Correctly load initramfs in stb container
Skiboot does not calculate the actual size and start location of the
initramfs if it is wrapped by an STB container (for example if
loading an initramfs from the ROOTFS partition).
Check if the initramfs is in an STB container and determine the size
and location correctly in the same manner as the kernel. Since
load_initramfs() is called after load_kernel() move the call to
trustedboot_exit_boot_services() into load_and_boot_kernel() so it
is called after both of these.
* hdat/i2c.c: quieten “v2 found, parsing as v1”
* hw/imc: Check for pause_microcode_at_boot() return status
pause_microcode_at_boot() loops through all the chip’s ucode control
block and pause the ucode if it is in the running state. But it does
not fail if any of the chip’s ucode is not initialised.
Add code to return a failure if ucode is not initialized in any of
the chip. Since pause_microcode_at_boot() is called just before
attaching the IMC device nodes in imc_init(), add code to check for
the function return.
Slot location code fixes:
* npu2: Use ibm, loc-code rather than ibm, slot-label
The ibm,slot-label property is to name the slot that appears under a
PCIe bridge. In the past we (ab)used the slot tables to attach names
to GPU devices and their corresponding NVLinks which resulted in
npu2.c using slot-label as a location code rather than as a way to
name slots.
Fix this up since it’s confusing.
* hdata/slots: Apply slot label to the parent slot
Slot names only really make sense when applied to an actual slot
rather than a device. On witherspoon the GPU devices have a name
associated with the device rather than the slot for the GPUs. Add a
hack that moves the slot label to the parent slot rather than on the
device itself.
* pci-dt-slot: Big ol’ cleanup
The underlying data that we get from HDAT can only really describe a
PCIe system. As such we can simplify the devicetree slot lookup code
by only caring about the important cases, namly, root ports and
switch downstream ports.
This also fixes a bug where root port didn’t get a Slot label
applied which results in devices under that port not having ibm,loc-
code set. This results in the EEH core being unable to report the
location of EEHed devices under that port.
opal-prd
--------
* opal-prd: Insert powernv_flash module
Explictly load powernv_flash module on BMC based system so that we
are sure that flash device is created before starting opal-prd
daemon.
Note that I have replaced pnor_available() check with
is_fsp_system(). As we want to load module on BMC system only. Also
pnor_init has enough logic to detect flash device. Hence
pnor_available() becomes redundant check.
NPU2/NVLINK2
------------
* npu2/hw-procedures: fence bricks on GPU reset
The NPU workbook defines a way of fencing a brick and getting the
brick out of fence state. We do have an implementation of bringing
the brick out of fenced/quiesced state. We do the latter in our
procedures, but to support run time reset we need to do the former.
The fencing ensures that access to memory behind the links will not
lead to HMI’s, but instead SUE’s will be populated in cache (in the
case of speculation). The expectation is then that prior to and
after reset, the operating system components will flush the cache
for the region of memory behind the GPU.
This patch does the following:
1. Implements a npu2_dev_fence_brick() function to set/clear
fence state
2. Clear FIR bits prior to clearing the fence status
3. Clear’s the fence status
4. We take the powerbus out of CQ fence much later now, in
credits_check() which is the last hardware procedure called after
link training.
* hw/npu2.c: Remove static configuration of NPU2 register
The NPU_SM_CONFIG0 register currently needs to be configured in
Skiboot to select NVLink mode, however Hostboot should configure
other bits in this register.
For some reason Skiboot was explicitly clearing bit-6
(CONFIG_DISABLE_VG_NOT_SYS). It is unclear why this bit was getting
cleared as recent Hostboot versions explicitly set it to the correct
value based on the specific system configuration. Therefore Skiboot
should not alter it.
Bit-58 (CONFIG_NVLINK_MODE) selects if NVLink mode should be enabled
or not. Hostboot does not configure this bit so Skiboot should
continue to configure it.
* npu2: Improve log output of GPU-to-link mapping
Debugging issues related to unconnected NVLinks can be a little less
irritating if we use the NPU2DEV{DBG,INF}() macros instead of
prlog().
In short, change this:
NPU2: comparing GPU 'GPU2' and NPU2 'GPU1'
NPU2: comparing GPU 'GPU3' and NPU2 'GPU1'
NPU2: comparing GPU 'GPU4' and NPU2 'GPU1'
NPU2: comparing GPU 'GPU5' and NPU2 'GPU1'
:
npu2_dev_bind_pci_dev: No PCI device for NPU2 device 0006:00:01.0 to bind to. If you expect a GPU to be there, this is a problem.
to this:
NPU6:0:1.0 Comparing GPU 'GPU2' and NPU2 'GPU1'
NPU6:0:1.0 Comparing GPU 'GPU3' and NPU2 'GPU1'
NPU6:0:1.0 Comparing GPU 'GPU4' and NPU2 'GPU1'
NPU6:0:1.0 Comparing GPU 'GPU5' and NPU2 'GPU1'
:
NPU6:0:1.0 No PCI device found for slot 'GPU1'
* npu2: Move NPU2_XTS_BDF_MAP_VALID assignment to context init
A bad GPU or other condition may leave us with a subset of links
that never get initialized. If an ATSD is sent to one of those
bricks, it will never complete, leaving us waiting forever for a
response:
watchdog: BUG: soft lockup - CPU#23 stuck for 23s! [acos:2050]
...
Modules linked in: nvidia_uvm(O) nvidia(O)
CPU: 23 PID: 2050 Comm: acos Tainted: G W O 4.14.0 #2
task: c0000000285cfc00 task.stack: c000001fea860000
NIP: c0000000000abdf0 LR: c0000000000acc48 CTR: c0000000000ace60
REGS: c000001fea863550 TRAP: 0901 Tainted: G W O (4.14.0)
MSR: 9000000000009033 <SF,HV,EE,ME,IR,DR,RI,LE> CR: 28004484 XER: 20040000
CFAR: c0000000000abdf4 SOFTE: 1
GPR00: c0000000000acc48 c000001fea8637d0 c0000000011f7c00 c000001fea863820
GPR04: 0000000002000000 0004100026000000 c0000000012778c8 c00000000127a560
GPR08: 0000000000000001 0000000000000080 c000201cc7cb7750 ffffffffffffffff
GPR12: 0000000000008000 c000000003167e80
NIP [c0000000000abdf0] mmio_invalidate_wait+0x90/0xc0
LR [c0000000000acc48] mmio_invalidate.isra.11+0x158/0x370
ATSDs are only sent to bricks which have a valid entry in the
XTS_BDF table. So to prevent the hang, don’t set
NPU2_XTS_BDF_MAP_VALID unless we make it all the way to creating a
context for the BDF.
Secure and Trusted Boot
-----------------------
* hdata/tpmrel: detect tpm not present by looking up the
stinfo->status
Skiboot detects if tpm is present by checking if a
secureboot_tpm_info entry exists. However, if a tpm is not present,
hostboot also creates a secureboot_tpm_info entry. In this case,
hostboot creates an empty entry, but setting the field tpm_status to
TPM_NOT_PRESENT.
This detects if tpm is not present by looking up the stinfo->status.
This fixes the “TPMREL: TPM node not found for chip_id=0 (HB bug)”
issue, reproduced when skiboot is running on a system that has no
tpm.
PCI
---
* phb4: Restore bus numbers after CRS
Currently we restore PCIe bus numbers right after the link is up.
Unfortunately as this point we haven’t done CRS so config space may
not be accessible.
This moves the bus number restore till after CRS has happened.
* romulus: Add a barebones slot table
* phb4: Quieten and improve “Timeout waiting for electrical link”
This happens normally if a slot doesn’t have a working HW presence
detect and relies instead of inband presence detect.
The message we display is scary and not very useful unless ou are
debugging, so quiten it up and change it to something more
meaningful.
* pcie-slot: Don’t fail powering on an already on switch
If the power state is already the required value, return
OPAL_SUCCESS rather than OPAL_PARAMETER to avoid spurrious errors
during boot.
CAPI/OpenCAPI
-------------
* capi: Keep the current mmio windows in the mbt cache table.
When the phb is used as a CAPI interface, the current mmio windows
list is cleaned before adding the capi and the prefetchable memory
(M64) windows, which implies that the non-prefetchable BAR is no
more configured. This patch allows to set only the mbt bar to pass
capi mmio window and to keep, as defined, the other mmio values (M32
and M64).
* npu2-opencapi: Fix ‘link internal error’ FIR, take 2
When setting up an opencapi link, we set the transport muxes first,
then set the PHY training config register, which includes disabling
nvlink mode for the bricks. That’s the order of the init sequence,
as found in the NPU workbook.
In reality, doing so works, but it raises 2 FIR bits in the PowerBus
OLL FIR Register for the 2 links when we configure the transport
muxes. Presumably because nvlink is not disabled yet and we are
configuring the transport muxes for opencapi.
bit 60:
link0 internal error
bit 61:
link1 internal error
Overall the current setup ends up being correct and everything
works, but we raise 2 FIR bits.
So tweak the order of operations to disable nvlink before
configuring the transport muxes. Incidentally, this is what the
scripts from the opencapi enablement team were doing all along.
* npu2-opencapi: Fix ‘link internal error’ FIR, take 1
When we setup a link, we always enable ODL0 and ODL1 at the same
time in the PHY training config register, even though we are setting
up only one OTL/ODL, so it raises a “link internal error” FIR bit in
the PowerBus OLL FIR Register for the second link. The error is
harmless, as we’ll eventually setup the second link, but there’s no
reason to raise that FIR bit.
The fix is simply to only enable the ODL we are using for the link.
* phb4: Do not set the PBCQ Tunnel BAR register when enabling capi
mode.
The cxl driver will set the capi value, like other drivers already
do.
* phb4: set TVT1 for tunneled operations in capi mode
The ASN indication is used for tunneled operations (as_notify and
atomics). Tunneled operation messages can be sent in PCI mode as
well as CAPI mode.
The address field of as_notify messages is hijacked to encode the
LPID/PID/TID of the target thread, so those messages should not go
through address translation. Therefore bit 59 is part of the ASN
indication.
This patch sets TVT#1 in bypass mode when capi mode is enabled, to
prevent as_notify messages from being dropped.
Debugging/Testing improvements
==============================
Since 6.0-rc1: - mambo: Enable XER CA32 and OV32 bits on P9
POWER9 adds 32 bit carry and overflow bits to the XER, but we need
to set the relevant CTRL1 bit to enable them.
* Makefile: Fix building natively on ppc64le
When on ppc64le and CROSS is not set by the environment, make
assumes ppc64 and sets a default CROSS. Check for ppc64le as well,
so that ‘make’ works out of the box on ppc64le.
* Experimental support for building with Clang
* Improvements to testing and Travis CI
Since 5.11:
* core/stack: backtrace unwind basic OPAL call details
Put OPAL callers’ r1 into the stack back chain, and then use that to
unwind back to the OPAL entry frame (as opposed to boot entry, which
has a 0 back chain).
From there, dump the OPAL call token and the caller’s r1. A
backtrace looks like this:
CPU 0000 Backtrace:
S: 0000000031c03ba0 R: 000000003001a548 ._abort+0x4c
S: 0000000031c03c20 R: 000000003001baac .opal_run_pollers+0x3c
S: 0000000031c03ca0 R: 000000003001bcbc .opal_poll_events+0xc4
S: 0000000031c03d20 R: 00000000300051dc opal_entry+0x12c
--- OPAL call entry token: 0xa caller R1: 0xc0000000006d3b90 ---
This is pretty basic for the moment, but it does give you the bottom
of the Linux stack. It will allow some interesting improvements in
future.
First, with the eframe, all the call’s parameters can be printed out
as well. The ___backtrace / ___print_backtrace API needs to be
reworked in order to support this, but it’s otherwise very simple
(see opal_trace_entry()).
Second, it will allow Linux’s stack to be passed back to Linux via a
debugging opal call. This will allow Linux’s BUG() or xmon to also
print the Linux back trace in case of a NMI or MCE or watchdog
lockup that hits in OPAL.
* asm/head: implement quiescing without stack or clobbering regs
Quiescing currently is implmeented in C in opal_entry before the
opal call handler is called. This works well enough for simple cases
like fast reset when one CPU wants all others out of the way.
Linux would like to use it to prevent an sreset IPI from
interrupting firmware, which could lead to deadlocks when crash
dumping or entering the debugger. Linux interrupts do not recover
well when returning back to general OPAL code, due to r13 not being
restored. OPAL also can’t be re-entered, which may happen e.g., from
the debugger.
So move the quiesce hold/reject to entry code, beore the stack or r1
or r13 registers are switched. OPAL can be interrupted and returned
to or re-entered during this period.
This does not completely solve all such problems. OPAL will be
interrupted with sreset if the quiesce times out, and it can be
interrupted by MCEs as well. These still have the issues above.
* core/opal: Allow poller re-entry if OPAL was re-entered
If an NMI interrupts the middle of running pollers and the OS
invokes pollers again (e.g., for console output), the poller re-
entrancy check will prevent it from running and spam the console.
That check was designed to catch a poller calling opal_run_pollers,
OPAL re-entrancy is something different and is detected elsewhere.
Avoid the poller recursion check if OPAL has been re-entered. This
is a best-effort attempt to cope with errors.
* core/opal: Emergency stack for re-entry
This detects OPAL being re-entered by the OS, and switches to an
emergency stack if it was. This protects the firmware’s main stack
from re-entrancy and allows the OS to use NMI facilities for crash /
debug functionality.
Further nested re-entry will destroy the previous emergency stack
and prevent returning, but those should be rare cases.
This stack is sized at 16kB, which doubles the size of CPU stacks,
so as not to introduce a regression in primary stack size. The 16kB
stack originally had a 4kB machine check stack at the top, which was
removed by 80eee1946 (“opal: Remove machine check interrupt patching
in OPAL.”). So it is possible the size could be tightened again, but
that would require further analysis.
* hdat_to_dt: hash_prop the same on all platforms Fixes this unit
test on ppc64le hosts.
* mambo: Add persistent memory disk support
This adds support to for mapping disks images using persistent
memory. Disks can be added by setting this ENV variable:
PMEM_DISK=”/mydisks/disk1.img,/mydisks/disk2.img”
These will show up in Linux as /dev/pmem0 and /dev/pmem1.
This uses a new feature in mambo “mysim memory mmap ..” which is
only available since mambo commit 0131f0fc08 (from 24/4/2018).
This also needs the of_pmem.c driver in Linux which is only
available since v4.17. It works with powernv_defconfig +
CONFIG_OF_PMEM.
* external/mambo: Add di command to decode instructions
By default you get 16 instructions but you can specify the number
you want. i.e.
systemsim % di 0x100 4
0x0000000000000100: Enc:0xA64BB17D : mtspr HSPRG1,r13
0x0000000000000104: Enc:0xA64AB07D : mfspr r13,HSPRG0
0x0000000000000108: Enc:0xF0092DF9 : std r9,0x9F0(r13)
0x000000000000010C: Enc:0xA6E2207D : mfspr r9,PPR
Using di since it’s what xmon uses.
* mambo/mambo_utils.tcl: Inject an MCE at a specified address
Currently we don’t support injecting an MCE on a specific address.
This is useful for testing functionality like memcpy_mcsafe() (see
https://patchwork.ozlabs.org/cover/893339/)
The core of the functionality is a routine called
inject_mce_ue_on_addr, which takes an addr argument and injects an
MCE (load/store with UE) when the specified address is accessed by
code. This functionality can easily be enhanced to cover instruction
UE’s as well.
A sample use case to create an MCE on stack access would be
set addr [mysim display gpr 1]
inject_mce_ue_on_addr $addr
This would cause an mce on any r1 or r1 based access
* external/mambo: improve helper for machine checks
Improve workarounds for stop injection, because mambo often will
trigger on 0x104/204 when injecting sreset/mces.
This also adds a workaround to skip injecting on reservations to
avoid infinite loops when doing inject_mce_step.
* travis: Enable ppc64le builds
At least on the IBM Travis Enterprise instance, we can now do
ppc64le builds!
We can only build a subset of our matrix due to availability of
ppc64le distros. The Dockerfiles need some tweaking to only attempt
to install (x86_64 only) Mambo binaries, as well as the build
scripts.
* external: Add “lpc” tool
This is a little front-end to the lpc debugfs files to access the
LPC bus from userspace on the host.
* core/test/run-trace: fix on ppc64el
--
Stewart Smith
OPAL Architect, IBM.
More information about the Skiboot
mailing list