diff options
author | Vineet Gupta <vgupta@synopsys.com> | 2015-04-15 16:14:07 +0200 |
---|---|---|
committer | Vineet Gupta <vgupta@synopsys.com> | 2015-04-20 14:57:30 +0200 |
commit | bde80c237e49983e2b26dfa9925325a070b71de7 (patch) | |
tree | 8b9f0ff0b9e0553665dfab6e35e14afc7212387b /arch/arc/include | |
parent | ARC: perf: make @arc_pmu static global (diff) | |
download | linux-bde80c237e49983e2b26dfa9925325a070b71de7.tar.xz linux-bde80c237e49983e2b26dfa9925325a070b71de7.zip |
ARC: perf: Add some comments/debug stuff
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Diffstat (limited to 'arch/arc/include')
-rw-r--r-- | arch/arc/include/asm/perf_event.h | 53 |
1 files changed, 24 insertions, 29 deletions
diff --git a/arch/arc/include/asm/perf_event.h b/arch/arc/include/asm/perf_event.h index cbf755e32a03..1c45667c1367 100644 --- a/arch/arc/include/asm/perf_event.h +++ b/arch/arc/include/asm/perf_event.h @@ -57,26 +57,7 @@ struct arc_reg_cc_build { #define PERF_COUNT_ARC_HW_MAX (PERF_COUNT_HW_MAX + 6) /* - * The "generalized" performance events seem to really be a copy - * of the available events on x86 processors; the mapping to ARC - * events is not always possible 1-to-1. Fortunately, there doesn't - * seem to be an exact definition for these events, so we can cheat - * a bit where necessary. - * - * In particular, the following PERF events may behave a bit differently - * compared to other architectures: - * - * PERF_COUNT_HW_CPU_CYCLES - * Cycles not in halted state - * - * PERF_COUNT_HW_REF_CPU_CYCLES - * Reference cycles not in halted state, same as PERF_COUNT_HW_CPU_CYCLES - * for now as we don't do Dynamic Voltage/Frequency Scaling (yet) - * - * PERF_COUNT_HW_BUS_CYCLES - * Unclear what this means, Intel uses 0x013c, which according to - * their datasheet means "unhalted reference cycles". It sounds similar - * to PERF_COUNT_HW_REF_CPU_CYCLES, and we use the same counter for it. + * Some ARC pct quirks: * * PERF_COUNT_HW_STALLED_CYCLES_BACKEND * PERF_COUNT_HW_STALLED_CYCLES_FRONTEND @@ -91,21 +72,35 @@ struct arc_reg_cc_build { * Note that I$ cache misses aren't counted by either of the two! */ +/* + * ARC PCT has hardware conditions with fixed "names" but variable "indexes" + * (based on a specific RTL build) + * Below is the static map between perf generic/arc specific event_id and + * h/w condition names. + * At the time of probe, we loop thru each index and find it's name to + * complete the mapping of perf event_id to h/w index as latter is needed + * to program the counter really + */ static const char * const arc_pmu_ev_hw_map[] = { + /* count cycles */ [PERF_COUNT_HW_CPU_CYCLES] = "crun", [PERF_COUNT_HW_REF_CPU_CYCLES] = "crun", [PERF_COUNT_HW_BUS_CYCLES] = "crun", - [PERF_COUNT_HW_INSTRUCTIONS] = "iall", - [PERF_COUNT_HW_BRANCH_MISSES] = "bpfail", - [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = "ijmp", + [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = "bflush", [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = "bstall", - [PERF_COUNT_ARC_DCLM] = "dclm", - [PERF_COUNT_ARC_DCSM] = "dcsm", - [PERF_COUNT_ARC_ICM] = "icm", - [PERF_COUNT_ARC_BPOK] = "bpok", - [PERF_COUNT_ARC_EDTLB] = "edtlb", - [PERF_COUNT_ARC_EITLB] = "eitlb", + + /* counts condition */ + [PERF_COUNT_HW_INSTRUCTIONS] = "iall", + [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = "ijmp", + [PERF_COUNT_ARC_BPOK] = "bpok", /* NP-NT, PT-T, PNT-NT */ + [PERF_COUNT_HW_BRANCH_MISSES] = "bpfail", /* NP-T, PT-NT, PNT-T */ + + [PERF_COUNT_ARC_DCLM] = "dclm", /* D-cache Load Miss */ + [PERF_COUNT_ARC_DCSM] = "dcsm", /* D-cache Store Miss */ + [PERF_COUNT_ARC_ICM] = "icm", /* I-cache Miss */ + [PERF_COUNT_ARC_EDTLB] = "edtlb", /* D-TLB Miss */ + [PERF_COUNT_ARC_EITLB] = "eitlb", /* I-TLB Miss */ }; #define C(_x) PERF_COUNT_HW_CACHE_##_x |