[PATCH v13 0/5] In-Memory Collection Counter instrumentation Support

Madhavan Srinivasan maddy at linux.vnet.ibm.com
Wed Jul 19 07:36:31 AEST 2017

Power9 has In-Memory-Collection (IMC) infrastructure which contains
various Performance Monitoring Units (PMUs) at Nest level (these are
on-chip but off-core), Core level and Thread level.

The Nest PMU counters are handled by a Nest IMC microcode which runs
in the OCC (On-Chip Controller) complex. The microcode collects the
counter data and moves the nest IMC counter data to memory.

The Core and Thread IMC PMU counters are handled in the core. Core
level PMU counters give us the IMC counters' data per core and thread
level PMU counters give us the IMC counters' data per CPU thread.

This patchset enables the nest IMC, core IMC and thread IMC
PMUs and is based on the initial work done by Madhavan Srinivasan.
"Nest Instrumentation Support" :

v1 for this patchset can be found here :

Nest events:
Per-chip nest instrumentation provides various per-chip metrics
such as memory, powerbus, Xlink and Alink bandwidth.

Core events:
Per-core IMC instrumentation provides various per-core metrics
such as non-idle cycles, non-idle instructions, various cache and
memory related metrics etc.

Thread events:
All the events for thread level are same as core level with the
difference being in the domain. These are per-cpu metrics.

PMU Events' Information:
OPAL obtains the IMC PMU and event information from the IMC Catalog
and passes on to the kernel via the device tree. The events' information
contains :
 - Event name
 - Event Offset
 - Event description
and, maybe :
 - Event scale
 - Event unit

Some PMUs may have a common scale and unit values for all their
supported events. For those cases, the scale and unit properties for
those events must be inherited from the PMU.

The event offset in the memory is where the counter data gets

The kernel discovers the IMC counters information in the device tree
at the "imc-counters" device node which has a compatible field

Parsing of the Events' information:
To parse the IMC PMUs and events information, the kernel has to
discover the "imc-counters" node and walk through the pmu and event

Here is an excerpt of the dt showing the imc-counters with
mcs (nest), core and thread node:


/ {
        name = "";
        compatible = "ibm,opal-in-memory-counters";
        #address-cells = <0x1>;
        #size-cells = <0x1>;
        version-id = "";

        NEST_MCS: nest-mcs-events {
                #address-cells = <0x1>;
                #size-cells = <0x1>;

                event at 0 {
                        event-name = "RRTO_QFULL_NO_DISP" ;
                        reg = <0x0 0x8>;
                        desc = "RRTO not dispatched in MCS0 due to capacity - pulses once for each time a valid RRTO op is not dispatched due to a command list full condition" ;
                event at 8 {
                        event-name = "WRTO_QFULL_NO_DISP" ;
                        reg = <0x8 0x8>;
                        desc = "WRTO not dispatched in MCS0 due to capacity - pulses once for each time a valid WRTO op is not dispatched due to a command list full condition" ;
        mcs01 {
                compatible = "ibm,imc-counters";
                events-prefix = "PM_MCS01_";
                unit = "";
                scale = "";
                reg = <0x118 0x8>;
                events = < &NEST_MCS >;
                type = <0x10>;
        mcs23 {
                compatible = "ibm,imc-counters";
                events-prefix = "PM_MCS23_";
                unit = "";
                scale = "";
                reg = <0x198 0x8>;
                events = < &NEST_MCS >;
                type = <0x10>;

        CORE_EVENTS: core-events {
                #address-cells = <0x1>;
                #size-cells = <0x1>;

                event at e0 {
                        event-name = "0THRD_NON_IDLE_PCYC" ;
                        reg = <0xe0 0x8>;
                        desc = "The number of processor cycles when all threads are idle" ;
                event at 120 {
                        event-name = "1THRD_NON_IDLE_PCYC" ;
                        reg = <0x120 0x8>;
                        desc = "The number of processor cycles when exactly one SMT thread is executing non-idle code" ;
        core {
                compatible = "ibm,imc-counters";
                events-prefix = "CPM_";
                unit = "";
                scale = "";
                reg = <0x0 0x8>;
                events = < &CORE_EVENTS >;
                type = <0x4>;

        thread {
                compatible = "ibm,imc-counters";
                events-prefix = "CPM_";
                unit = "";
                scale = "";
                reg = <0x0 0x8>;
                events = < &CORE_EVENTS >;
                type = <0x1>;

>From the device tree, the kernel parses the PMUs and their events'

After parsing the IMC PMUs and their events, the PMUs and their
attributes are registered in the kernel.

This patchset (patches 9 and 10) configure the thread level IMC PMUs
to count for tasks, which give us the thread level metric values per
Example Usage :
 # perf list

  nest_mcs01_imc/PM_MCS01_64B_RD_DISP_PORT01/            [Kernel PMU event]
  nest_mcs01_imc/PM_MCS01_64B_RD_DISP_PORT23/            [Kernel PMU event]

  core_imc/CPM_0THRD_NON_IDLE_PCYC/                  [Kernel PMU event]
  core_imc/CPM_1THRD_NON_IDLE_INST/                  [Kernel PMU event]
  thread_imc/CPM_0THRD_NON_IDLE_PCYC/                [Kernel PMU event]
  thread_imc/CPM_1THRD_NON_IDLE_INST/                [Kernel PMU event]

To see per chip data for nest_mcs0/PM_MCS_DOWN_128B_DATA_XFER_MC0/ :
 # ./perf stat -e "nest_mcs01_imc/PM_MCS01_64B_WR_DISP_PORT01/" -a --per-socket

To see non-idle instructions for core 0 :
 # ./perf stat -e "core_imc/CPM_NON_IDLE_INST/" -C 0 -I 1000

To see non-idle instructions for a "make" :
 # ./perf stat -e "thread_imc/CPM_NON_IDLE_PCYC/" make

v12 -> v13
 - merged all the patches and have split them into 5.
 - reworked the event parser functions
 - made changes to cleanup code
 - fixed error handling path
 - Made changes to commit messages
 - Added more inline comments and documents
v11 -> v12
 - cleanup_all_core_imc_memory() function updated.
 - is_core_imc_mem_inited function is made static.
 - code rearrangement is done
 - event_init functions for nest, core and thread are updated
   with a new logic to obtain the lock.
 - Updated the comments.

v10 -> v11

 - cpuhotplug call unregistration for nest counters is handled.
 - nest counters are also disable in case of kdump.
 - alloc_pages_node is used for memory allocation for core and thread,
   instead of alloc_pages_exact_nid.
 - base_addr calculations for nest, core, thread events are modified, as the
   'config' now has more fields .
 - event config fields are updated for nest,core and thread.
 - cpuhotplug function for nest,core and thread are modified.
 - opal-call api for start and stop is changed.

v9 -> v10
 - reworked the cpu hot plug functions for nest and core
 - Updated imc_get_mem_addr_nest
 - Changed u64 vbase[IMC_MAX_PAGES]; to u64 *vbase[IMC_MAX_PAGES]; in struct imc_mem_info

v8 -> v9
 - Updated nest, core, thread cpuhotplug functions.
 - PMU node parsing logic is changed as there is change in
   the ima-catalog file. PMU nodes are identified based on the
   "type" property.
 - Since imc-counters subtree accomodates the memory base
   address and offset for nest counter data, logic to get
   memory address for nest counters data is updated.
 - Memory allocation functions for core and thread are updated.
 - Data structures for imc instrumentation are updated.
 - pmu reserve/release functions for nest,core,thread are
   moved to *_imc_event_init.
 - Updated the comments.
 - Included necessary checks in core_imc_change_cpu_context()

v7 -> v8:
 - opal-call API for nest and core is changed.
 - thread_ima doesn't have CPUMASK_ATTR, hence added a
   fix in patch 09/10, which will swap the IMC_EVENT_ATTR
   slot with IMC_CPUMASK_ATTR.

v6 -> v7:
 - Updated the commit message and code comments.
 - Changed the counter init code to disable the
   nest/core counters by default and enable only
   when it is used.
 - Updated the pmu-setup code to register the
   PMUs which doesn't have events.
 - replaced imc_event_info_val() to imc_event_prop_update()
 - Updated the imc_pmu_setup() code, by checking for the "value"
   of compatible property instead of merely checking for compatible.
 - removed imc_get_domain().
 - init_imc_pmu() and imc_pmu_setup() are made  __init.
 - update_max_val() is invoked immediately after updating the offset value.

v5 -> v6:
 - merged few patches for the readability and code flow
 - Updated the commit message and code comments.
 - updated cpuhotplug code and added checks for perf migration context
 - Added READ_ONCE() when reading the counter data.
 - replaced of_property_read_u32() with of_get_address() for "reg" property read

 v4 -> v5:
 - Updated opal call numbers
 - Added a patch to disable Core-IMC device using shutdown callback
 - Added patch to support cpuhotplug for thread-imc
 - Added patch to disable and enable core imc engine in cpuhot plug path

 v3 -> v4 :
 - Changed the events parser code to discover the PMU and events because
   of the changed format of the IMC DTS file (Patch 3).
 - Implemented the two TODOs to include core and thread IMC support with
   this patchset (Patches 7 through 10).
 - Changed the CPU hotplug code of Nest IMC PMUs to include a new state

 v2 -> v3 :
 - Changed all references for IMA (In-Memory Accumulation) to IMC (In-Memory

v1 -> v2 :
 - Account for the cases where a PMU can have a common scale and unit
   values for all its supported events (Patch 3/6).
 - Fixed a Build error (for maple_defconfig) by enabling imc_pmu.o
   only for CONFIG_PPC_POWERNV=y (Patch 4/6)
 - Read from the "event-name" property instead of "name" for an event
   node (Patch 3/6).

Madhavan Srinivasan (5):
  powerpc/platform/powernv: Update IMC OPAL APIs
  powerpc/platform/powernv: Detect and create IMC device
  powerpc/perf: Add nest imc pmu support
  powerpc/perf: Add core imc pmu support
  powerpc/perf: Add thread imc pmu support

 arch/powerpc/include/asm/imc-pmu.h             |  128 +++
 arch/powerpc/include/asm/opal-api.h            |   11 +-
 arch/powerpc/include/asm/opal.h                |    5 +
 arch/powerpc/perf/Makefile                     |    1 +
 arch/powerpc/perf/imc-pmu.c                    | 1302 ++++++++++++++++++++++++
 arch/powerpc/platforms/powernv/Makefile        |    2 +-
 arch/powerpc/platforms/powernv/opal-imc.c      |  226 ++++
 arch/powerpc/platforms/powernv/opal-wrappers.S |    3 +
 arch/powerpc/platforms/powernv/opal.c          |   14 +
 include/linux/cpuhotplug.h                     |    3 +
 10 files changed, 1693 insertions(+), 2 deletions(-)
 create mode 100644 arch/powerpc/include/asm/imc-pmu.h
 create mode 100644 arch/powerpc/perf/imc-pmu.c
 create mode 100644 arch/powerpc/platforms/powernv/opal-imc.c


More information about the Linuxppc-dev mailing list