Also fix incorrect Trfc[0-3] value on Family 15h.
Change-Id: Iafc233984ae1d44fe6a1cb5b109d36397cbd991a
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: https://review.coreboot.org/12055
Tested-by: build bot (Jenkins)
Reviewed-by: Martin Roth <martinroth@google.com>
Reviewed-by: Patrick Georgi <pgeorgi@google.com>
The silicon in control of CC6 appears to contain minor bugs
and / or deviations from the BKDG; through trial and error
it was found that these issues can be worked around by reserving
the entire possible CC6 save region, regardless of currently
installed node count.
Change-Id: If31140651f25f9c524a824b2da552ce3690eae18
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: https://review.coreboot.org/12054
Tested-by: build bot (Jenkins)
Reviewed-by: Patrick Georgi <pgeorgi@google.com>
The coherent fabric on all Family 10h/15h devices supports
isochronous mode, which is required for IOMMU operation.
Add initial support for isochronous operation.
Change-Id: Idd7c9b94a65f856b0059e1d45f8719d9475771b6
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12042
Tested-by: build bot (Jenkins)
Reviewed-by: Martin Roth <martinroth@google.com>
Certain workloads may evict too many lines of other cores from the
L3 cache if configured as one monolithic shared cache region.
Forcibly partition L3 cache to improve performance.
Change-Id: Ie4e28dd886aaa1c586b0919c5fe87ef1696f47e9
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12036
Reviewed-by: Paul Menzel <paulepanter@users.sourceforge.net>
Tested-by: build bot (Jenkins)
Reviewed-by: Felix Held <felix-coreboot@felixheld.de>
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
The existing code generated an invalid NUMA table
that was rejected by Linux, leading to poor resource
allocation. This was due to system MMIO resources
being inserted into the table when the table should
only contain DRAM resources.
Do not include system MMIO resources (i.e. resources
with an index less than 0x10) in the NUMA table.
Change-Id: I99c200382b52a99687daf266a84873d9ae2df025
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12035
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
Tested-by: build bot (Jenkins)
Stability issues have arisen on multiple Family 15h systems
when configuration restoration is enabled. In all cases these
stability issues resolved by allowing the RAM to go through a
full training cycle.
Change-Id: I017e0dd5120110124d5b5d5276befef6f7740614
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12034
Tested-by: build bot (Jenkins)
Tested-by: Raptor Engineering Automated Test Stand <noreply@raptorengineeringinc.com>
Reviewed-by: Felix Held <felix-coreboot@felixheld.de>
The BKDG is not correct regarding HT Freq write ordering;
indicate this in a comment to avoid confusion.
Change-Id: I37db191c144c81aba5d4a1e6291db5669a35a31a
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12030
Tested-by: build bot (Jenkins)
Reviewed-by: Martin Roth <martinroth@google.com>
All modern Opteron processors support the HT probe filter,
which helps to increase coherent fabric performance by
reducing the number of HT transactions per cache probe.
AMD recommends that the probe filter be enabled on all
systems with more than two nodes, and it does not hurt
to enable it on systems with 2 nodes.
Change-Id: I00a27a828260be8685ae622cfa5a4995add95a8e
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12021
Tested-by: build bot (Jenkins)
Tested-by: Raptor Engineering Automated Test Stand <noreply@raptorengineeringinc.com>
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
This fixes Family 15h multiple package support; the previous code
hung in CAR setup and romstage when more than one CPU package was
installed for a variety of loosely related reasons.
TEST: Booted ASUS KGPE-D16 with two Opteron 6328 processors
and several different RDIMM configurations.
Change-Id: I171197c90f72d3496a385465937b7666cbf7e308
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12020
Tested-by: build bot (Jenkins)
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
The existing MCT support code did not perform any of the requisite
configuration to support registered or x4 DIMMs. Add the needed
configuration per the BKDG for Family 15h.
Change-Id: I9ee0bb7346aa35f564fe535cdd337ec7f6148f2b
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12019
Tested-by: build bot (Jenkins)
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
Sufficient support has been added to allow booting with registered
DIMMs on the KGPE-D16 in certain slots. ECC support needs additional
work; the ECC data lanes appear to cause boot failures in some slots.
Change-Id: Ieaf4cbf351908e5a89760be49a6667dc55dbc575
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12017
Tested-by: build bot (Jenkins)
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
The current code did not define the number of DIMM slots on the
mainboard, which lead to incorrect configuration values and
occassional training failure.
Add preliminary support for DIMM slot count configuration.
Change-Id: I488511d6262ffa8207c442d133314aed0f75acfb
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12016
Tested-by: build bot (Jenkins)
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
CAR space on certain platforms is nearly full. This prevents the
addition of necessary RAM initialization features such as x4 DIMM
support. As the DIMM SPD cache uses a sizeable amount of CAR RAM,
reducing it would free up a significant amount of CAR RAM.
DDR3-based AMD platforms only support up to 3 physical DIMMs on
each channel (6 per node). Reduce the maximum number of DIMMs
on a node from 8 to 6 accordingly.
Change-Id: I38def86da76fc622785318c825670209b2ac9017
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12107
Tested-by: Raptor Engineering Automated Test Stand <noreply@raptorengineeringinc.com>
Tested-by: build bot (Jenkins)
Reviewed-by: Stefan Reinauer <stefan.reinauer@coreboot.org>
In the course of adding full Family 15h MCT support some Family
15h specific settings were inadvertently applied to Family 10h
processors.
Only apply Family15h specific settings to Family 15h processors.
Change-Id: I5dcb333d3a5a49318fe7bddd4c386642205c343e
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12013
Tested-by: build bot (Jenkins)
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
When both DCTs of a node are in use the DRAM clocks should be skewed
with respect to one another in order to reduce cross-channel interference.
Set the clock skew bit according to the BKDG recommendations.
Change-Id: Ibcce54fc53b79beba2f790994bcf87cc0354213a
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12011
Tested-by: build bot (Jenkins)
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
The existing code did not set Rtt timing parameters when registered
DIMMs were used with Family 15h processors. Set the Rtt values
according to the BKDG recommendations.
Change-Id: I80cd7f8aec12951611d802f33e5e167a41dd532e
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12010
Tested-by: build bot (Jenkins)
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
AMD Opteron processors contain a very fragile phy phase detection circuit.
Additionally, the algorithm given in the BKDG does not function as intended;
this was verified both on real hardware via execution trace and on paper
with values read back from multiple CPUs and DIMMs.
As a result, the phy training algorithm given in the BKDG has been
replaced with a phy training algorithm developed at Raptor Engineering.
This particular patch is the first part of that algorithm; the code
is updated in future patches but this should exist in the historical
record in case something breaks down in the later sections of code.
Change-Id: Ic7a19d24954f47c922126e3da7be1f7e85f7396f
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12007
Tested-by: build bot (Jenkins)
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
AMD's automatic phy phase detection hardware is very fragile and often
produces incorrect results. Attempt to recover from obvious phase
locking errors by retrying phy training on the failing link.
Change-Id: Ia2c3022534c9ad44714eef6e118869f054bd9f6b
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12006
Tested-by: build bot (Jenkins)
Reviewed-by: Stefan Reinauer <stefan.reinauer@coreboot.org>
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
The existing MCT code did not properly set up the On Die Termination
(ODT) or timing values for registered DIMMs. Use the BKDG recommended
values when registered DIMMs are installed.
Change-Id: Ia9ee770d9f9c22e18c12e38b5bb4a7bae0a99062
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12005
Tested-by: build bot (Jenkins)
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
There were numerous issues surrounding AMD ECC initialization on
Family 15h processors due to the incomplete derivation from Family
10h MCT code. Bring the Family 15h ECC initialization and supporting
setup code in line with the BKDG recommendations.
Change-Id: I7f009b655f8500aeb22981f7020f1db74cdd6925
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12003
Tested-by: build bot (Jenkins)
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
The existing ECC initialization algorithm contained several bugs on both
Family 10h and Family 15h processors, including activation of ECC scrub
before DRAM setup was completed, in violation of both BKDG and errata
recommendations.
Change-Id: I09a8ea83024186b7ece7d78a4bef1201ab34ff8a
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12002
Tested-by: build bot (Jenkins)
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
The existing prefetcher configuration was incorrect; use the correct
values from the AMD Family 10h and Family 15h BKDGs as appropriate.
Change-Id: I287ffa6345e1f4d232d4b2ea4251650ada3fda92
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/12417
Tested-by: build bot (Jenkins)
Reviewed-by: Stefan Reinauer <stefan.reinauer@coreboot.org>
The existing code enabled ECC before clearing memory. As the
AMD CPUs will generate MCEs on any invalid check bits, this
resulted in random lockups during memory training due to the
uniniailized check bits.
Initialize ECC check bits before enabling ECC hardware.
Change-Id: I992e7040520570893ba6a213138dd57bfa14733b
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/11996
Tested-by: build bot (Jenkins)
Reviewed-by: Martin Roth <martinroth@google.com>
Certain DIMMs, for example DIMMs on which the EEPROM has been modified
by the end user, may not contain a valid SPD checksum. While this is
not a normal condition, it may be useful to allow a checksum override
while memory timing parameters are being altered, e.g. in the course
of overclocking or underclocking, or when recovering from a bad SPD
write.
This is an advanced level feature primarily useful for debugging
and development.
Change-Id: Ia743a13348d0a6e5e4dfffa04ed9582e0f7f3dad
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/11987
Tested-by: build bot (Jenkins)
Reviewed-by: Stefan Reinauer <stefan.reinauer@coreboot.org>
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
Under specific circumstances, for instance in low power or fanless
machines, it may be useful to cap the maximum P-state of the CPU.
Allow the maximum CPU P-state to be set via an NVRAM option.
Change-Id: Ifdbb1ad11a856f855c59702ae0ee99e95b08520e
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/11985
Reviewed-by: Patrick Georgi <pgeorgi@google.com>
Tested-by: build bot (Jenkins)
This patch adds CC6 power save support to the AMD Family 15h
support code. As CC6 is a complex power saving state that
relies heavily on CPU, northbridge, and southbridge cooperation,
this patch alters significant amounts of code throughout the
tree simultaneously.
Allowing the CPU to enter CC6 allows the second level of turbo
boost to be reached, and also provides significant power savings
when the system is idle due to the complete core shutdown.
Change-Id: I44ce157cda97fb85f3e8f3d7262d4712b5410670
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/11979
Tested-by: build bot (Jenkins)
Tested-by: Raptor Engineering Automated Test Stand <noreply@raptorengineeringinc.com>
Reviewed-by: Stefan Reinauer <stefan.reinauer@coreboot.org>
Reviewed-by: Ronald G. Minnich <rminnich@gmail.com>
DRAM training accounts for most of the romstage startup time, yet
if the hardware configuration has not changed from the previous boot
the previously discovered training values are still valid. Use them
if the DIMM configuration has not changed since the last boot.
The SPD values of all installed DIMMs are hashed and stored in the S3
resume data area of the main system Flash device. If a DIMM is changed
the hash will almost certainly change as well, forcing retraining on next
boot.
Change-Id: I37ed277b16476d38e4af76c6ae827a575c6b017d
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/11976
Tested-by: Raptor Engineering Automated Test Stand <noreply@raptorengineeringinc.com>
Tested-by: build bot (Jenkins)
Reviewed-by: Ronald G. Minnich <rminnich@gmail.com>
Upon bootup the hardware reads at minimum 256 * 16 bytes (4Kb, or 32KB) over
I2C on a system with all DIMM slots populated. If even one of those reads
has a single flipped bit in it (due to EMI, cross coupling with another trace,
or just poor margins on some cheap DIMM) the system will hang and require a
hard reset. In practice I've seen failure rates as high as 1 failed boot in
50 due to this issue, granted with cheap DIMMs, but even so retrying the read
resolves the corruption issue.
I2C is not designed for continuous data transmission with high reliability, and
there is no hardware error checking, therefore a single retry when transferring
this amount of data makes sense.
Change-Id: Ifab63eca2233c63a6a42ab8b7e742f8e47fb2a09
Signed-off-by: Timothy Pearson <tpearson@raptorengineeringinc.com>
Reviewed-on: http://review.coreboot.org/11975
Tested-by: build bot (Jenkins)
Reviewed-by: Patrick Georgi <pgeorgi@google.com>
