NAME

pmc.cmn-600 — Library for accessing the Arm CoreLink CMN-600 Coherent Mesh Network Controller performance counter events

LIBRARY

Performance Counters Library (libpmc, -lpmc)

SYNOPSIS

#include <pmc.h>

DESCRIPTION

CMN-600 PMU counters may be configured to count any one of a defined set of hardware events. Unlike other performance counters, counters for the CMN-600 require the node ID to set up.

Node ID information currently can be obtained one of two ways. Using bootverbose, for example set sysctl debug.bootverbose=1 and then load the hwpmc(4) KLD module. The cmn600 module will be loaded automatically as a dependency:

$ sysctl
  debug.bootverbose=1

$ kldload hwpmc

Another way is to use sysctl to trigger dump of nodes tree to system console:

$ sysctl
  dev.cmn600.0.dump_nodes=1

Some BIOS versions of dual-socket machines have no NUMA domain information in ACPI. In such cases, to get more accurate events statistics, set the kernel environment variable hint.cmn600.{unit}.domain={value}. Where {unit} is a cmn600 device unit number and {value} is the NUMA domain of the CPU package containing that CMN-600 controller. Example:

$ kenv
  hint.cmn600.0.domain=0

$ kenv
  hint.cmn600.1.domain=1

$ kldunload hwpmc cmn600

$ kldload hwpmc

Arm CoreLink CMN-600 Coherent Mesh Network Controller performance counters are documented in Revision: r3p2, Arm CoreLink CMN-600 Coherent Mesh Network Technical Reference Manual, ARM Limited, 2020.

PMC Capabilities

CMN-600 PMU counters support the following capabilities:

Capability	Support
PMC_CAP_CASCADE	No
PMC_CAP_EDGE	No
PMC_CAP_INTERRUPT	Yes
PMC_CAP_INVERT	No
PMC_CAP_READ	Yes
PMC_CAP_PRECISE	No
PMC_CAP_SYSTEM	Yes
PMC_CAP_TAGGING	No
PMC_CAP_THRESHOLD	Yes
PMC_CAP_USER	No
PMC_CAP_WRITE	Yes

Event Qualifiers

Event specifiers for these PMCs support the following common qualifiers:

nodeid=nodeid: Request counting for specific event at node nodeid.
occupancy=value: Filtering by occupancy type.
xpport=port: Count only events matched by port. (East, West, North, South, devport0, devport1 or numeric 0 to 5)
xpchannel=channel: Filter events by XP node channel. (REQ, RSP, SNP, DAT or 0, 1, 2, 3)

Class Name Prefix

These PMCs are named using a class name prefix of “CMN600_PMU_”.

Event Specifiers

The following list of PMC events are available:

DVM node events

dn_rxreq_dvmop: Number of DVMOP requests received. This includes all the sub-types include TLB invalidate, Branch predictor invalidate, instruction cache (physical and virtual) invalidate.
dn_rxreq_dvmsync: Number of DVM Sync requests received.
dn_rxreq_dvmop_vmid_filtered: Number of incoming DVMOP requests that are subject to VMID based filtering. This is a measure of the effectiveness of VMID based filtering and potential reduction in DVM snoops.
dn_rxreq_retried: Number of incoming requests that are retried. This is a measure of the retry rate.
dn_rxreq_trk_occupancy: Counts the tracker occupancy in DN. "occupancy": All, dvmop, dvmsync
dn_rxreq_tlbi_dvmop: Number of DVMOP TLB invalidate requests received.
dn_rxreq_bpi_dvmop: Number of DVMOP Branch predictor invalidate requests received.
dn_rxreq_pici_dvmop: Number of DVMOP physical instruction cache invalidate requests received.
dn_rxreq_vivi_dvmop: Number of DVMOP virtual instruction cache invalidate requests received.
dn_rxreq_dvmop_other_filtered: Number of DVM op requests to RNDs, BPI or PICI/VICI, that were filtered
dn_rxreq_snp_sent: Number of SNPs sent to RNs
dn_rxreq_snp_stalled: Number of SNPs stalled to RNs due to lack of Crds
dn_rxreq_trk_full: DVM tracker full counter

HN-F node events

hnf_cache_miss: Counts total cache misses in first lookup result (high priority)
hnf_slc_sf_cache_access: Counts number of cache accesses in first access (high priority)
hnf_cache_fill: Counts total allocations in HN SLC (all cache line allocations to SLC)
hnf_pocq_retry: Counts number of retried requests
hnf_pocq_reqs_recvd: Counts number of requests received by HN
hnf_sf_hit: Counts number of SF hits
hnf_sf_evictions: Counts number of SF eviction cache invalidations initiated
hnf_dir_snoops_sent: Counts number of directed snoops sent (not including SF back invalidation)
hnf_brd_snoops_sent: Counts number of multicast snoops sent (not including SF back invalidation)
hnf_slc_eviction: Counts number of SLC evictions (dirty only)
hnf_slc_fill_invalid_way: Counts number of SLC fills to an invalid way
hnf_mc_retries: Counts number of retried transactions by the MC
hnf_mc_reqs: Counts number of requests sent to MC
hnf_qos_hh_retry: Counts number of times a HighHigh priority request is protocol-retried at the HN-F
hnf_qos_pocq: Counts the POCQ occupancy in HN-F. Support argument "occupancy". Accept: All, Read, Write, Atomic, Stash. Default: All.
hnf_pocq_addrhaz: Counts number of POCQ address hazards upon allocation
hnf_pocq_atomic_addrhaz: Counts number of POCQ address hazards upon allocation for atomic operations
hnf_ld_st_swp_adq_full: Counts number of times ADQ is full for Ld/St/SWP type atomic operations while POCQ has pending operations
hnf_cmp_adq_full: Counts number of times ADQ is full for CMP type atomic operations while POCQ has pending operations
hnf_txdat_stall: Counts number of times HN-F has a pending TXDAT flit but no credits to upload
hnf_txrsp_stall: Counts number of times HN-F has a pending TXRSP flit but no credits to upload
hnf_seq_full: Counts number of times requests are replayed in SLC pipe due to SEQ being full
hnf_seq_hit: Counts number of times a request in SLC hit a pending SF eviction in SEQ
hnf_snp_sent: Counts number of snoops sent including directed, multicast, and SF back invalidation
hnf_sfbi_dir_snp_sent: Counts number of times directed snoops were sent due to SF back invalidation
hnf_sfbi_brd_snp_sent: Counts number of times multicast snoops were sent due to SF back invalidation
hnf_snp_sent_untrk: Counts number of times snooped were sent due to untracked RNF's
hnf_intv_dirty: Counts number of times SF back invalidation resulted in dirty line intervention from the RN
hnf_stash_snp_sent: Counts number of times stash snoops were sent
hnf_stash_data_pull: Counts number of times stash snoops resulted in data pull from the RN
hnf_snp_fwded: Counts number of times data forward snoops were sent

HN-I node events

hni_rrt_rd_occ_cnt_ovfl: RRT read occupancy count overflow
hni_rrt_wr_occ_cnt_ovfl: RRT write occupancy count overflow
hni_rdt_rd_occ_cnt_ovfl: RDT read occupancy count overflow
hni_rdt_wr_occ_cnt_ovfl: RDT write occupancy count overflow
hni_wdb_occ_cnt_ovfl: WDB occupancy count overflow
hni_rrt_rd_alloc: RRT read allocation
hni_rrt_wr_alloc: RRT write allocation
hni_rdt_rd_alloc: RDT read allocation
hni_rdt_wr_alloc: RDT write allocation
hni_wdb_alloc: WDB allocation
hni_txrsp_retryack: RETRYACK TXRSP flit sent
hni_arvalid_no_arready: ARVALID set without ARREADY event
hni_arready_no_arvalid: ARREADY set without ARVALID event
hni_awvalid_no_awready: AWVALID set without AWREADY event
hni_awready_no_awvalid: AWREADY set without AWVALID event
hni_wvalid_no_wready: WVALID set without WREADY event
hni_txdat_stall: TXDAT stall (TXDAT valid but no link credit available)
hni_nonpcie_serialization: Non-PCIe serialization event
hni_pcie_serialization: PCIe serialization event

XP node events

xp_txflit_valid: Number of flits transmitted on a specified port and CHI channel. This is a measure of the flit transfer bandwidth from an XP. Note: On device ports, this event also includes link flit transfers.
xp_txflit_stall: Number of cycles when a flit is stalled at an XP waiting for link credits at a specified port and CHI channel. This is a measure of the flit traffic congestion on the mesh and at the flit download ports.
xp_partial_dat_flit: Number of times when a partial DAT flit is uploaded onto the mesh from a RN-F_CHIA port. Partial DAT flit transmission occurs when XP is not able to combine two 128b DAT flits and send them over the 256b DAT channel. This can happen under 2 circumstances: 1. Only one 128b DAT flit is received within a transmission time window. 2. Two 128b DAT flits are received but they are not two halves of a
single 256b word.

SBSX node events

sbsx_rd_req: Read request
sbsx_wr_req: Write request
sbsx_cmo_req: CMO request
sbsx_txrsp_retryack: RETRYACK TXRSP flit sent
sbsx_txdat_flitv: TXDAT flit seen
sbsx_txrsp_flitv: TXRSP flit seen
sbsx_rd_req_trkr_occ_cnt_ovfl: Read request tracker occupancy count overflow
sbsx_wr_req_trkr_occ_cnt_ovfl: Write request tracker occupancy count overflow
sbsx_cmo_req_trkr_occ_cnt_ovfl: CMO request tracker occupancy count overflow
sbsx_wdb_occ_cnt_ovfl: WDB occupancy count overflow
sbsx_rd_axi_trkr_occ_cnt_ovfl: Read AXI pending tracker occupancy count overflow
sbsx_cmo_axi_trkr_occ_cnt_ovfl: CMO AXI pending tracker occupancy count overflow
sbsx_arvalid_no_arready: ARVALID set without ARREADY
sbsx_awvalid_no_awready: AWVALID set without AWREADY
sbsx_wvalid_no_wready: WVALID set without WREADY
sbsx_txdat_stall: TXDAT stall (TXDAT valid but no link credit available)
sbsx_txrsp_stall: TXRSP stall (TXRSP valid but no link credit available)

RN-D node events

rnd_s0_rdata_beats: Number of RData beats, RVALID and RREADY, dispatched on port 0. This is a measure of the read bandwidth, including CMO responses.
rnd_s1_rdata_beats: Number of RData beats, RVALID and RREADY, dispatched on port 1. This is a measure of the read bandwidth, including CMO responses.
rnd_s2_rdata_beats: Number of RData beats, RVALID and RREADY, dispatched on port 2. This is a measure of the read bandwidth, including CMO responses.
rnd_rxdat_flits: Number of RXDAT flits received. This is a measure of the true read data bandwidth, excluding CMOs.
rnd_txdat_flits: Number of TXDAT flits dispatched. This is a measure of the write bandwidth.
rnd_txreq_flits_total: Number of TXREQ flits dispatched. This is a measure of the total request bandwidth.
rnd_txreq_flits_retried: Number of retried TXREQ flits dispatched. This is a measure of the retry rate.
rnd_rrt_occ_ovfl: All entries in the read request tracker are occupied. This is a measure of oversubscription in the read request tracker.
rnd_wrt_occ_ovfl: All entries in the write request tracker are occupied. This is a measure of oversubscription in the write request tracker.
rnd_txreq_flits_replayed: Number of replayed TXREQ flits. This is the measure of replay rate.
rnd_wrcancel_sent: Number of write data cancels sent. This is the measure of write cancel rate.
rnd_s0_wdata_beats: Number of WData beats, WVALID and WREADY, dispatched on port 0. This is a measure of write bandwidth on AXI port 0.
rnd_s1_wdata_beats: Number of WData beats, WVALID and WREADY, dispatched on port 1. This is a measure of write bandwidth on AXI port 1.
rnd_s2_wdata_beats: Number of WData beats, WVALID and WREADY, dispatched on port 2. This is a measure of write bandwidth on AXI port 2.
rnd_rrt_alloc: Number of allocations in the read request tracker. This is a measure of read transaction count.
rnd_wrt_alloc: Number of allocations in the write request tracker. This is a measure of write transaction count.
rnd_rdb_unord: Number of cycles for which Read Data Buffer state machine is in Unordered Mode.
rnd_rdb_replay: Number of cycles for which Read Data Buffer state machine is in Replay mode
rnd_rdb_hybrid: Number of cycles for which Read Data Buffer state machine is in hybrid mode. Hybrid mode is where there is mix of ordered/unordered traffic.
rnd_rdb_ord: Number of cycles for which Read Data Buffer state machine is in ordered Mode.

RN-I node events

rni_s0_rdata_beats: Number of RData beats, RVALID and RREADY, dispatched on port 0. This is a measure of the read bandwidth, including CMO responses.
rni_s1_rdata_beats: Number of RData beats, RVALID and RREADY, dispatched on port 1. This is a measure of the read bandwidth, including CMO responses.
rni_s2_rdata_beats: Number of RData beats, RVALID and RREADY, dispatched on port 2. This is a measure of the read bandwidth, including CMO responses.
rni_rxdat_flits: Number of RXDAT flits received. This is a measure of the true read data bandwidth, excluding CMOs.
rni_txdat_flits: Number of TXDAT flits dispatched. This is a measure of the write bandwidth.
rni_txreq_flits_total: Number of TXREQ flits dispatched. This is a measure of the total request bandwidth.
rni_txreq_flits_retried: Number of retried TXREQ flits dispatched. This is a measure of the retry rate.
rni_rrt_occ_ovfl: All entries in the read request tracker are occupied. This is a measure of oversubscription in the read request tracker.
rni_wrt_occ_ovfl: All entries in the write request tracker are occupied. This is a measure of oversubscription in the write request tracker.
rni_txreq_flits_replayed: Number of replayed TXREQ flits. This is the measure of replay rate.
rni_wrcancel_sent: Number of write data cancels sent. This is the measure of write cancel rate
rni_s0_wdata_beats: Number of WData beats, WVALID and WREADY, dispatched on port 0. This is a measure of write bandwidth on AXI port 0.
rni_s1_wdata_beats: Number of WData beats, WVALID and WREADY, dispatched on port 1. This is a measure of write bandwidth on AXI port 1.
rni_s2_wdata_beats: Number of WData beats, WVALID and WREADY, dispatched on port 2. This is a measure of write bandwidth on AXI port 2.
rni_rrt_alloc: Number of allocations in the read request tracker. This is a measure of read transaction count.
rni_wrt_alloc: Number of allocations in the write request tracker. This is a measure of write transaction count
rni_rdb_unord: Number of cycles for which Read Data Buffer state machine is in Unordered Mode.
rni_rdb_replay: Number of cycles for which Read Data Buffer state machine is in Replay mode
rni_rdb_hybrid: Number of cycles for which Read Data Buffer state machine is in hybrid mode. Hybrid mode is where there is mix of ordered/unordered traffic.
rni_rdb_ord: Number of cycles for which Read Data Buffer state machine is in ordered Mode.

CXHA node events

(Note: CXHA events descriptions are guessed)

cxha_rddatbyp: Number of Read DAT Bypass
cxha_chirsp_up_stall: Number of CHI RSP up Stall
cxha_chidat_up_stall: Number of CHI DAT up Stall
cxha_snppcrd_lnk0_stall: Number of Snoop Pcrd Stall on Link 0
cxha_snppcrd_lnk1_stall: Number of Snoop Pcrd Stall on Link 1
cxha_snppcrd_lnk2_stall: Number of Snoop Pcrd Stall on Link 2
cxha_reqtrk_occ: Request Tracker Occupancy
cxha_rdb_occ: Read Data Buffer Occupancy
cxha_rdbbyp_occ: Read Data Buffer Bypass Occupancy
cxha_wdb_occ: Write Data Buffer Occupancy
cxha_snptrk_occ: Snoop Tracker Occupancy
cxha_sdb_occ: SDB Occupancy
cxha_snphaz_occ: Snoop Hazard Occupancy

CXRA node events

cxra_req_trk_occ: Request tracker occupancy
cxra_snp_trk_occ: Snoop tracker occupancy
cxra_rd_dat_buf_occ: Read data buffer occupancy
cxra_wr_dat_buf_occ: Write data buffer occupancy
cxra_snp_sink_buf_occ: Snoop sink buffer occupancy
cxra_snp_bcasts: Snoop broadcasts
cxra_req_chains: Number of request chains formed larger than one
cxra_req_chain_avg_len: Average size of request chains, only for chain sizes larger than one
cxra_chi_rsp_upload_stalls: Local RA upload stalls to CHI because of contention with HA
cxra_chi_dat_upload_stalls: Local RA upload stalls to CHI because of contention with HA
cxra_dat_pcrd_stalls_lnk0: Memory Data Request available, but no DAT Pcrd to send over CCIX per LinkEnd 0
cxra_dat_pcrd_stalls_lnk1: Memory Data Request available, but no DAT Pcrd to send over CCIX per LinkEnd 1
cxra_dat_pcrd_stalls_lnk2: Memory Data Request available, but no DAT Pcrd to send over CCIX per LinkEnd 2
cxra_req_pcrd_stalls_lnk0: Memory Data Request available but no Req Pcrd to send over CCIX per LinkEnd 0
cxra_req_pcrd_stalls_lnk1: Memory Data Request available but no Req Pcrd to send over CCIX per LinkEnd 1
cxra_req_pcrd_stalls_lnk2: Memory Data Request available but no Req Pcrd to send over CCIX per LinkEnd 2
cxra_ext_rsp_stall: CHI external RSP stall
cxra_ext_dat_stall: CHI external DAT stall

CXLA node events

cxla_rx_tlp_link0: RX TLP for Link 0
cxla_rx_tlp_link1: RX TLP for Link 1
cxla_rx_tlp_link2: RX TLP for Link 2
cxla_tx_tlp_link0: TX TLP for Link 0
cxla_tx_tlp_link1: TX TLP for Link 1
cxla_tx_tlp_link2: TX TLP for Link 2
cxla_rx_cxs_link0: RX CXS for Link 0
cxla_rx_cxs_link1: RX CXS for Link 1
cxla_rx_cxs_link2: RX CXS for Link 2
cxla_tx_cxs_link0: TX CXS for Link 0
cxla_tx_cxs_link1: TX CXS for Link 1
cxla_tx_cxs_link2: TX CXS for Link 2
cxla_avg_rx_tlp_sz_dws: Average RX TLP size in DWs
cxla_avg_tx_tlp_sz_dws: Average TX TLP size in DWs
cxla_avg_rx_tlp_sz_ccix_msg: Average RX TLP size in CCIX messages
cxla_avg_tx_tlp_sz_ccix_msg: Average TX TLP size in CCIX messages
cxla_avg_sz_rx_cxs_dw_beat: Average size of RX CXS in DWs within a beat
cxla_avg_sz_tx_cxs_dw_beat: Average size of TX CXS in DWs within a beat
cxla_tx_cxs_link_credit_backpressure: TX CXS link credit backpressure
cxla_rx_tlp_buffer_full: RX TLP buffer full and backpressured
cxla_tx_tlp_buffer_full: TX TLP buffer full and backpressured
cxla_avg_latency_process_rx_tlp: Average latency to process an RX TLP
cxla_avg_latency_form_tx_tlp: Average latency to form a TX TLP

HISTORY

The pmc library first appeared in FreeBSD 6.0.
The pmc.cmn-600 driver was added in FreeBSD 14.0.

AUTHORS

The Performance Counters Library (libpmc, -lpmc) library was written by Joseph Koshy <jkoshy@FreeBSD.org>, Oleksandr Rybalko <ray@FreeBSD.org>.
The CMN-600 PMU driver was sponsored by ARM Ltd. This manual page was written by Oleksandr Rybalko <ray@FreeBSD.org>.