Pull x86 resource control updates from Borislav Petkov:
- Extend the resctrl machinery to support telemetry monitoring on
Intel (Tony Luck)
The practical usage of this is being able to tell how much energy or
how much work can be attributed to a group of tasks tracked under a
single idenitifier. Prepend this work with proper refactoring of
resctrl domains handling code.
* tag 'x86_cache_for_v7.0_rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (30 commits)
x86,fs/resctrl: Update documentation for telemetry events
x86/resctrl: Enable RDT_RESOURCE_PERF_PKG
fs/resctrl: Move RMID initialization to first mount
x86,fs/resctrl: Compute number of RMIDs as minimum across resources
fs/resctrl: Move allocation/free of closid_num_dirty_rmid[]
x86/resctrl: Handle number of RMIDs supported by RDT_RESOURCE_PERF_PKG
x86/resctrl: Add energy/perf choices to rdt boot option
x86,fs/resctrl: Handle domain creation/deletion for RDT_RESOURCE_PERF_PKG
fs/resctrl: Refactor rmdir_mondata_subdir_allrdtgrp()
fs/resctrl: Refactor mkdir_mondata_subdir()
x86/resctrl: Read telemetry events
x86/resctrl: Find and enable usable telemetry events
x86,fs/resctrl: Add architectural event pointer
x86,fs/resctrl: Fill in details of events for performance and energy GUIDs
x86/resctrl: Discover hardware telemetry events
fs/resctrl: Emphasize that L3 monitoring resource is required for summing domains
x86,fs/resctrl: Add and initialize a resource for package scope monitoring
x86,fs/resctrl: Add an architectural hook called for first mount
x86,fs/resctrl: Support binary fixed point event counters
x86,fs/resctrl: Handle events that can be read from any CPU
...
The memory bandwidth calculation relies on reading the hardware counter
and measuring the delta between samples. To ensure accurate measurement,
the software reads the counter frequently enough to prevent it from
rolling over twice between reads.
The default Memory Bandwidth Monitoring (MBM) counter width is 24 bits.
Hygon CPUs provide a 32-bit width counter, but they do not support the
MBM capability CPUID leaf (0xF.[ECX=1]:EAX) to report the width offset
(from 24 bits).
Consequently, the kernel falls back to the 24-bit default counter width,
which causes incorrect overflow handling on Hygon CPUs.
Fix this by explicitly setting the counter width offset to 8 bits (resulting
in a 32-bit total counter width) for Hygon CPUs.
Fixes: d8df126349 ("x86/cpu/hygon: Add missing resctrl_cpu_detect() in bsp_init helper")
Signed-off-by: Xiaochen Shen <shenxiaochen@open-hieco.net>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Cc: stable@vger.kernel.org
Link: https://patch.msgid.link/20251209062650.1536952-3-shenxiaochen@open-hieco.net
Hygon CPUs supporting Platform QoS features currently undergo partial resctrl
initialization through resctrl_cpu_detect() in the Hygon BSP init helper and
AMD/Hygon common initialization code. However, several critical data
structures remain uninitialized for Hygon CPUs in the following paths:
- get_mem_config()-> __rdt_get_mem_config_amd():
rdt_resource::membw,alloc_capable
hw_res::num_closid
- rdt_init_res_defs()->rdt_init_res_defs_amd():
rdt_resource::cache
hw_res::msr_base,msr_update
Add the missing AMD/Hygon common initialization to ensure proper Platform QoS
functionality on Hygon CPUs.
Fixes: d8df126349 ("x86/cpu/hygon: Add missing resctrl_cpu_detect() in bsp_init helper")
Signed-off-by: Xiaochen Shen <shenxiaochen@open-hieco.net>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Cc: stable@vger.kernel.org
Link: https://patch.msgid.link/20251209062650.1536952-2-shenxiaochen@open-hieco.net
Since telemetry events are enumerated on resctrl mount the RDT_RESOURCE_PERF_PKG
resource is not considered "monitoring capable" during early resctrl initialization.
This means that the domain list for RDT_RESOURCE_PERF_PKG is not built when the CPU
hotplug notifiers are registered and run for the first time right after resctrl
initialization.
Mark the RDT_RESOURCE_PERF_PKG as "monitoring capable" upon successful telemetry
event enumeration to ensure future CPU hotplug events include this resource and
initialize its domain list for CPUs that are already online.
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
resctrl assumes that only the L3 resource supports monitor events, so it
simply takes the rdt_resource::num_rmid from RDT_RESOURCE_L3 as the system's
number of RMIDs.
The addition of telemetry events in a different resource breaks that
assumption.
Compute the number of available RMIDs as the minimum value across all
mon_capable resources (analogous to how the number of CLOSIDs is computed
across alloc_capable resources).
Note that mount time enumeration of the telemetry resource means that
this number can be reduced. If this happens, then some memory will
be wasted as the allocations for rdt_l3_mon_domain::mbm_states[] and
rdt_l3_mon_domain::rmid_busy_llc created during resctrl initialization will
be larger than needed.
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
There are now three meanings for "number of RMIDs":
1) The number for legacy features enumerated by CPUID leaf 0xF. This is the
maximum number of distinct values that can be loaded into MSR_IA32_PQR_ASSOC.
Note that systems with Sub-NUMA Cluster mode enabled will force scaling down
the CPUID enumerated value by the number of SNC nodes per L3-cache.
2) The number of registers in MMIO space for each event. This is enumerated in
the XML files and is the value initialized into event_group::num_rmid.
3) The number of "hardware counters" (this isn't a strictly accurate
description of how things work, but serves as a useful analogy that does
describe the limitations) feeding to those MMIO registers. This is enumerated
in telemetry_region::num_rmids returned by intel_pmt_get_regions_by_feature().
Event groups with insufficient "hardware counters" to track all RMIDs are
difficult for users to use, since the system may reassign "hardware counters"
at any time. This means that users cannot reliably collect two consecutive
event counts to compute the rate at which events are occurring.
Disable such event groups by default. The user may override this with
a command line "rdt=" option. In this case limit an under-resourced event
group's number of possible monitor resource groups to the lowest number of
"hardware counters".
Scan all enabled event groups and assign the RDT_RESOURCE_PERF_PKG resource
"num_rmid" value to the smallest of these values as this value will be used
later to compare against the number of RMIDs supported by other resources to
determine how many monitoring resource groups are supported.
N.B. Change type of resctrl_mon::num_rmid to u32 to match its usage and the
type of event_group::num_rmid so that min(r->num_rmid, e->num_rmid) won't
complain about mixing signed and unsigned types.
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
Legacy resctrl features are enumerated by X86_FEATURE_* flags. These may be
overridden by quirks to disable features in the case of errata. Users can use
kernel command line options to either disable a feature, or to force enable
a feature that was disabled by a quirk.
A different approach is needed for hardware features that do not have an
X86_FEATURE_* flag.
Update parsing of the "rdt=" boot parameter to call the telemetry driver
directly to handle new "perf" and "energy" options that controls activation of
telemetry monitoring of the named type. By itself a "perf" or "energy" option
controls the forced enabling or disabling (with ! prefix) of all event groups
of the named type. A ":guid" suffix allows for fine grained control per event
group.
[ bp: s/intel_aet_option/intel_handle_aet_option/g ]
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
The L3 resource has several requirements for domains. There are per-domain
structures that hold the 64-bit values of counters, and elements to keep
track of the overflow and limbo threads.
None of these are needed for the PERF_PKG resource. The hardware counters
are wide enough that they do not wrap around for decades.
Define a new rdt_perf_pkg_mon_domain structure which just consists of the
standard rdt_domain_hdr to keep track of domain id and CPU mask.
Update resctrl_online_mon_domain() for RDT_RESOURCE_PERF_PKG. The only action
needed for this resource is to create and populate domain directories if a
domain is added while resctrl is mounted.
Similarly resctrl_offline_mon_domain() only needs to remove domain directories.
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
Introduce intel_aet_read_event() to read telemetry events for resource
RDT_RESOURCE_PERF_PKG. There may be multiple aggregators tracking each
package, so scan all of them and add up all counters. Aggregators may return
an invalid data indication if they have received no records for a given RMID.
The user will see "Unavailable" if none of the aggregators on a package
provide valid counts.
Resctrl now uses readq() so depends on X86_64. Update Kconfig.
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
Every event group has a private copy of the data of all telemetry event
aggregators (aka "telemetry regions") tracking its feature type. Included
may be regions that have the same feature type but tracking different GUID
from the event group's.
Traverse the event group's telemetry region data and mark all regions that
are not usable by the event group as unusable by clearing those regions'
MMIO addresses. A region is considered unusable if:
1) GUID does not match the GUID of the event group.
2) Package ID is invalid.
3) The enumerated size of the MMIO region does not match the expected
value from the XML description file.
Hereafter any telemetry region with an MMIO address is considered valid for
the event group it is associated with.
Enable all the event group's events as long as there is at least one usable
region from where data for its events can be read. Enabling of an event can
fail if the same event has already been enabled as part of another event
group. It should never happen that the same event is described by different
GUID supported by the same system so just WARN (via resctrl_enable_mon_event())
and skip the event.
Note that it is architecturally possible that some telemetry events are only
supported by a subset of the packages in the system. It is not expected that
systems will ever do this. If they do the user will see event files in resctrl
that always return "Unavailable".
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
The resctrl file system layer passes the domain, RMID, and event id to the
architecture to fetch an event counter.
Fetching a telemetry event counter requires additional information that is
private to the architecture, for example, the offset into MMIO space from
where the counter should be read.
Add mon_evt::arch_priv that architecture can use for any private data related
to the event. The resctrl filesystem initializes mon_evt::arch_priv when the
architecture enables the event and passes it back to architecture when needing
to fetch an event counter.
Suggested-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
The telemetry event aggregators of the Intel Clearwater Forest CPU support two
RMID-based feature types: "energy" with GUID 0x26696143¹, and "perf" with
GUID 0x26557651².
The event counter offsets in an aggregator's MMIO space are arranged in groups
for each RMID.
E.g., the "energy" counters for GUID 0x26696143 are arranged like this:
MMIO offset:0x0000 Counter for RMID 0 PMT_EVENT_ENERGY
MMIO offset:0x0008 Counter for RMID 0 PMT_EVENT_ACTIVITY
MMIO offset:0x0010 Counter for RMID 1 PMT_EVENT_ENERGY
MMIO offset:0x0018 Counter for RMID 1 PMT_EVENT_ACTIVITY
...
MMIO offset:0x23F0 Counter for RMID 575 PMT_EVENT_ENERGY
MMIO offset:0x23F8 Counter for RMID 575 PMT_EVENT_ACTIVITY
After all counters there are three status registers that provide indications
of how many times an aggregator was unable to process event counts, the time
stamp for the most recent loss of data, and the time stamp of the most recent
successful update.
MMIO offset:0x2400 AGG_DATA_LOSS_COUNT
MMIO offset:0x2408 AGG_DATA_LOSS_TIMESTAMP
MMIO offset:0x2410 LAST_UPDATE_TIMESTAMP
Define event_group structures for both of these aggregator types and define
the events tracked by the aggregators in the file system code.
PMT_EVENT_ENERGY and PMT_EVENT_ACTIVITY are produced in fixed point format.
File system code must output as floating point values.
¹https://github.com/intel/Intel-PMT/blob/main/xml/CWF/OOBMSM/RMID-ENERGY/cwf_aggregator.xml
²https://github.com/intel/Intel-PMT/blob/main/xml/CWF/OOBMSM/RMID-PERF/cwf_aggregator.xml
[ bp: Massage commit message. ]
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
Each CPU collects data for telemetry events that it sends to the nearest
telemetry event aggregator either when the value of MSR_IA32_PQR_ASSOC.RMID
changes, or when a two millisecond timer expires.
There is a feature type ("energy" or "perf"), GUID, and MMIO region associated
with each aggregator. This combination links to an XML description of the
set of telemetry events tracked by the aggregator. XML files are published
by Intel in a GitHub repository¹.
The telemetry event aggregators maintain per-RMID per-event counts of the
total seen for all the CPUs. There may be multiple telemetry event aggregators
per package.
There are separate sets of aggregators for each feature type. Aggregators
in a set may have different GUIDs. All aggregators with the same feature
type and GUID are symmetric keeping counts for the same set of events for
the CPUs that provide data to them.
The XML file for each aggregator provides the following information:
0) Feature type of the events ("perf" or "energy")
1) Which telemetry events are tracked by the aggregator.
2) The order in which the event counters appear for each RMID.
3) The value type of each event counter (integer or fixed-point).
4) The number of RMIDs supported.
5) Which additional aggregator status registers are included.
6) The total size of the MMIO region for an aggregator.
Introduce struct event_group that condenses the relevant information from
an XML file. Hereafter an "event group" refers to a group of events of a
particular feature type (event_group::pfname set to "energy" or "perf") with
a particular GUID.
Use event_group::pfname to determine the feature id needed to obtain the
aggregator details. It will later be used in console messages and with the
rdt= boot parameter.
The INTEL_PMT_TELEMETRY driver enumerates support for telemetry events.
This driver provides intel_pmt_get_regions_by_feature() to list all available
telemetry event aggregators of a given feature type. The list includes the
"guid", the base address in MMIO space for the region where the event counters
are exposed, and the package id where the all the CPUs that report to this
aggregator are located.
Call INTEL_PMT_TELEMETRY's intel_pmt_get_regions_by_feature() for each event
group to obtain a private copy of that event group's aggregator data. Duplicate
the aggregator data between event groups that have the same feature type
but different GUID. Further processing on this private copy will be unique
to the event group.
¹https://github.com/intel/Intel-PMT
[ bp: Zap text explaining the code, s/guid/GUID/g ]
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
Add a new PERF_PKG resource and introduce package level scope for monitoring
telemetry events so that CPU hotplug notifiers can build domains at the
package granularity.
Use the physical package ID available via topology_physical_package_id()
to identify the monitoring domains with package level scope. This enables
user space to use:
/sys/devices/system/cpu/cpuX/topology/physical_package_id
to identify the monitoring domain a CPU is associated with.
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
Enumeration of Intel telemetry events is an asynchronous process involving
several mutually dependent drivers added as auxiliary devices during the
device_initcall() phase of Linux boot. The process finishes after the probe
functions of these drivers completes. But this happens after
resctrl_arch_late_init() is executed.
Tracing the enumeration process shows that it does complete a full seven
seconds before the earliest possible mount of the resctrl file system (when
included in /etc/fstab for automatic mount by systemd).
Add a hook for use by telemetry event enumeration and initialization and
run it once at the beginning of resctrl mount without any locks held.
The architecture is responsible for any required locking.
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Link: https://lore.kernel.org/r/20260105191711.GBaVwON5nZn-uO6Sqg@fat_crate.local
resctrl assumes that all monitor events can be displayed as unsigned decimal
integers.
Hardware architecture counters may provide some telemetry events with greater
precision where the event is not a simple count, but is a measurement of some
sort (e.g. Joules for energy consumed).
Add a new argument to resctrl_enable_mon_event() for architecture code to
inform the file system that the value for a counter is a fixed-point value
with a specific number of binary places.
Only allow architecture to use floating point format on events that the file
system has marked with mon_evt::is_floating_point which reflects the contract
with user space on how the event values are displayed.
Display fixed point values with values rounded to ceil(binary_bits * log10(2))
decimal places. Special case for zero binary bits to print "{value}.0".
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
resctrl assumes that monitor events can only be read from a CPU in the
cpumask_t set of each domain. This is true for x86 events accessed with an
MSR interface, but may not be true for other access methods such as MMIO.
Introduce and use flag mon_evt::any_cpu, settable by architecture, that
indicates there are no restrictions on which CPU can read that event. This
flag is not supported by the L3 event reading that requires to be run on a CPU
that belongs to the L3 domain of the event being read.
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
With the arrival of monitor events tied to new domains associated with a
different resource it would be clearer if the L3 resource specific functions
are more accurately named.
Rename three groups of functions:
Functions that allocate/free architecture per-RMID MBM state information:
arch_domain_mbm_alloc() -> l3_mon_domain_mbm_alloc()
mon_domain_free() -> l3_mon_domain_free()
Functions that allocate/free filesystem per-RMID MBM state information:
domain_setup_mon_state() -> domain_setup_l3_mon_state()
domain_destroy_mon_state() -> domain_destroy_l3_mon_state()
Initialization/exit:
rdt_get_mon_l3_config() -> rdt_get_l3_mon_config()
resctrl_mon_resource_init() -> resctrl_l3_mon_resource_init()
resctrl_mon_resource_exit() -> resctrl_l3_mon_resource_exit()
Ensure kernel-doc descriptions of these functions' return values are present
and correctly formatted.
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
The upcoming telemetry event monitoring is not tied to the L3 resource and
will have a new domain structure.
Rename the L3 resource specific domain data structures to include "l3_"
in their names to avoid confusion between the different resource specific
domain structures:
rdt_mon_domain -> rdt_l3_mon_domain
rdt_hw_mon_domain -> rdt_hw_l3_mon_domain
No functional change.
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
Convert the whole call sequence from mon_event_read() to resctrl_arch_rmid_read() to
pass resource independent struct rdt_domain_hdr instead of an L3 specific domain
structure to prepare for monitoring events in other resources.
This additional layer of indirection obscures which aspects of event counting depend
on a valid domain. Event initialization, support for assignable counters, and normal
event counting implicitly depend on a valid domain while summing of domains does not.
Split summing domains from the core event counting handling to make their respective
dependencies obvious.
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
Up until now, all monitoring events were associated with the L3 resource and it
made sense to use the L3 specific "struct rdt_mon_domain *" argument to functions
operating on domains.
Telemetry events will be tied to a new resource with its instances represented
by a new domain structure that, just like struct rdt_mon_domain, starts with
the generic struct rdt_domain_hdr.
Prepare to support domains belonging to different resources by changing the
calling convention of functions operating on domains. Pass the generic header
and use that to find the domain specific structure where needed.
Signed-off-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>
Link: https://lore.kernel.org/20251217172121.12030-1-tony.luck@intel.com
