libpfm_intel_skx_unc_cha — support for Intel Skylake X Server CHA-Box uncore PMU

Synopsis

#include <perfmon/pfmlib.h>

PMU name: skx_unc_cha[0-27]
PMU desc: Intel Skylake X CHA uncore PMU

Description

The library supports the Intel Skylake X CHA-Box (coherency and home agent engine) uncore PMU. There is one CHA-box PMU per physical core. Therefore there are up to twenty-eight identical CHA-Box PMU instances numbered from 0 up to possibly 27. On dual-socket systems, the number refers to the CHA-Box PMU on the socket where the program runs. For instance, if running on CPU18, then skx_unc_cha0 refers to the CHA-Box for physical core 0 on socket 1. Conversely, if running on CPU0, then the same skx_unc_cha0 refers to the CHA-Box for physical core 0 but on socket 0.

Each CHA-Box PMU implements 4 generic counters and two filter registers used only with certain events and umasks. The filters are either accessed via modifiers (see below) or umasks, such as the opcode or cache state filter.

Modifiers

The following modifiers are supported on Intel Skylake CHA-Box uncore PMU:

e

Enable edge detection, i.e., count only when there is a state transition from no occurrence of the event to at least one occurrence. This modifier must be combined with a threshold modifier (t) with a value greater or equal to one.  This is a boolean modifier.

t

Set the threshold value. When set to a non-zero value, the counter counts the number of C-Box cycles in which the number of occurrences of the event is greater or equal to the threshold.  This is an integer modifier with values in the range [0:255].

i

Invert the meaning of the event. The counter will now count cycles in which the event is not occurring. This is a boolean modifier.

loc

Match on local node target. This filter is only supported on UNC_C_TOR_INSERTS and UNC_C_TOR_OCCUPANCY. This is a boolean filter.

rem

Match on remote node target. This filter is only supported on UNC_C_TOR_INSERTS and UNC_C_TOR_OCCUPANCY. This is a boolean filter.

lmem

Match near memory cacheable. This filter is only supported on UNC_C_TOR_INSERTS and UNC_C_TOR_OCCUPANCY. This is a boolean filter.

rmem

Match not near memory cacheable. This filter is only supported on UNC_C_TOR_INSERTS and UNC_C_TOR_OCCUPANCY. This is a boolean filter.

nc

Match non-coherent requests. This filter is only supported on UNC_C_TOR_INSERTS and UNC_C_TOR_OCCUPANCY. This is a boolean filter.

isoc

Match isochronous requests. This filter is only supported on UNC_C_TOR_INSERTS and UNC_C_TOR_OCCUPANCY. This is a boolean filter.

Opcode filtering

Events UNC_C_TOR_INSERTS and UNC_C_TOR_OCCUPANCY support opcode matching. The processor implements two opcode filters. Both are used at the same time. The OPC0 umasks correspond to the first opcode matcher and OPC1 to the second opcode matcher. If only one opcode must be tracked then the unused filter will be set to 0. The opcode umasks must be used in combination with a specific queue umask otherwise the library will reject the event. The umask description shows which queue umask is required for each opcode. For instance, OPC0_RFO/OPC1_RFO require the IRQ queue and thus the IRQ umask.
 The opcode match umasks can be combined with other modifiers.

Authors

Stephane Eranian <eranian@gmail.com>

Info

January, 2018 Linux Programmer's Manual