blob: 0632b3aad83e7068f663964c77870b9e54cf6826 [file] [log] [blame]
Pawel Molla33b0da2014-07-22 18:32:59 +01001ARM Cache Coherent Network
2==========================
3
4CCN-504 is a ring-bus interconnect consisting of 11 crosspoints
5(XPs), with each crosspoint supporting up to two device ports,
6so nodes (devices) 0 and 1 are connected to crosspoint 0,
7nodes 2 and 3 to crosspoint 1 etc.
8
9PMU (perf) driver
10-----------------
11
12The CCN driver registers a perf PMU driver, which provides
13description of available events and configuration options
14in sysfs, see /sys/bus/event_source/devices/ccn*.
15
16The "format" directory describes format of the config, config1
17and config2 fields of the perf_event_attr structure. The "events"
18directory provides configuration templates for all documented
19events, that can be used with perf tool. For example "xp_valid_flit"
20is an equivalent of "type=0x8,event=0x4". Other parameters must be
21explicitly specified. For events originating from device, "node"
22defines its index. All crosspoint events require "xp" (index),
23"port" (device port number) and "vc" (virtual channel ID) and
24"dir" (direction). Watchpoints (special "event" value 0xfe) also
25require comparator values ("cmp_l" and "cmp_h") and "mask", being
26index of the comparator mask.
27
28Masks are defined separately from the event description
29(due to limited number of the config values) in the "cmp_mask"
30directory, with first 8 configurable by user and additional
314 hardcoded for the most frequent use cases.
32
33Cycle counter is described by a "type" value 0xff and does
34not require any other settings.
35
36Example of perf tool use:
37
38/ # perf list | grep ccn
39 ccn/cycles/ [Kernel PMU event]
40<...>
41 ccn/xp_valid_flit/ [Kernel PMU event]
42<...>
43
44/ # perf stat -C 0 -e ccn/cycles/,ccn/xp_valid_flit,xp=1,port=0,vc=1,dir=1/ \
45 sleep 1
46
47The driver does not support sampling, therefore "perf record" will
48not work. Also notice that only single cpu is being selected
49("-C 0") - this is because perf framework does not support
50"non-CPU related" counters (yet?) so system-wide session ("-a")
51would try (and in most cases fail) to set up the same event
52per each CPU.