2013-08-28 20:13:51

by Mischa Jonker

[permalink] [raw]
Subject: [PATCH] ARC: Add perf support

This adds basic perf support for ARC. Most PERF_COUNT_HW* events are
supported now. Cache events and callchains are not (yet).

Signed-off-by: Mischa Jonker <[email protected]>
---
arch/arc/boot/dts/angel4.dts | 4 +
arch/arc/include/asm/perf_event.h | 190 +++++++++++++++++++++-
arch/arc/kernel/Makefile | 1 +
arch/arc/kernel/perf_event.c | 323 +++++++++++++++++++++++++++++++++++++
4 files changed, 517 insertions(+), 1 deletion(-)
create mode 100644 arch/arc/kernel/perf_event.c

diff --git a/arch/arc/boot/dts/angel4.dts b/arch/arc/boot/dts/angel4.dts
index 4fb2d6f..fd34b5d 100644
--- a/arch/arc/boot/dts/angel4.dts
+++ b/arch/arc/boot/dts/angel4.dts
@@ -67,5 +67,9 @@
reg = <1>;
};
};
+
+ arcpmu0: pmu {
+ compatible = "snps,arc_pmu";
+ };
};
};
diff --git a/arch/arc/include/asm/perf_event.h b/arch/arc/include/asm/perf_event.h
index 115ad96..e56c5bb 100644
--- a/arch/arc/include/asm/perf_event.h
+++ b/arch/arc/include/asm/perf_event.h
@@ -1,5 +1,7 @@
/*
- * Copyright (C) 2011-2012 Synopsys, Inc. (http://www.synopsys.com)
+ * Linux performance counter support for ARC
+ *
+ * Copyright (C) 2011-2013 Synopsys, Inc. (http://www.synopsys.com)
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License version 2 as
@@ -10,4 +12,190 @@
#ifndef __ASM_PERF_EVENT_H
#define __ASM_PERF_EVENT_H

+/* real maximum varies per CPU, this is the maximum supported by the driver */
+#define ARC_PMU_MAX_HWEVENTS 64
+
+#define ARC_REG_CC_BUILD 0xF6
+#define ARC_REG_CC_INDEX 0x240
+#define ARC_REG_CC_NAME0 0x241
+#define ARC_REG_CC_NAME1 0x242
+
+#define ARC_REG_PCT_BUILD 0xF5
+#define ARC_REG_PCT_COUNTL 0x250
+#define ARC_REG_PCT_COUNTH 0x251
+#define ARC_REG_PCT_SNAPL 0x252
+#define ARC_REG_PCT_SNAPH 0x253
+#define ARC_REG_PCT_CONFIG 0x254
+#define ARC_REG_PCT_CONTROL 0x255
+#define ARC_REG_PCT_INDEX 0x256
+
+#define ARC_REG_PCT_CONTROL_CC (1 << 16) /* clear counts */
+#define ARC_REG_PCT_CONTROL_SN (1 << 17) /* snapshot */
+
+struct arc_reg_pct_build {
+#ifdef CONFIG_CPU_BIG_ENDIAN
+ unsigned int m:8, c:8, r:6, s:2, v:8;
+#else
+ unsigned int v:8, s:2, r:6, c:8, m:8;
+#endif
+};
+
+struct arc_reg_cc_build {
+#ifdef CONFIG_CPU_BIG_ENDIAN
+ unsigned int c:16, r:8, v:8;
+#else
+ unsigned int v:8, r:8, c:16;
+#endif
+};
+
+#define PERF_COUNT_ARC_DCLM (PERF_COUNT_HW_MAX + 0)
+#define PERF_COUNT_ARC_DCSM (PERF_COUNT_HW_MAX + 1)
+#define PERF_COUNT_ARC_ICM (PERF_COUNT_HW_MAX + 2)
+#define PERF_COUNT_ARC_BPOK (PERF_COUNT_HW_MAX + 3)
+#define PERF_COUNT_ARC_EDTLB (PERF_COUNT_HW_MAX + 4)
+#define PERF_COUNT_ARC_EITLB (PERF_COUNT_HW_MAX + 5)
+#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 */
+static const char * const arc_pmu_ev_hw_map[] = {
+ /* cycles not in halted state */
+ [PERF_COUNT_HW_CPU_CYCLES] = "crun",
+ /* reference cycles not in halted state; also "crun" as we don't
+ do Dynamic Voltage/Frequency Scaling (yet) */
+ [PERF_COUNT_HW_REF_CPU_CYCLES] = "crun",
+ /* Unclear what this means, Intel uses 0x013c, which according to
+ their datasheet means "unhalted reference cycles". Don't ask
+ me what the difference is with PERF_COUNT_HW_REF_CPU_CYCLES */
+ [PERF_COUNT_HW_BUS_CYCLES] = "crun",
+ [PERF_COUNT_HW_INSTRUCTIONS] = "iall",
+ [PERF_COUNT_HW_BRANCH_MISSES] = "bpfail",
+ [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = "ijmp",
+ /* ARC can either measure stalls per pipeline stage, or all stalls
+ combined; for now we assign all stalls to STALLED_CYCLES_BACKEND
+ and all pipeline flushes (e.g. caused by mispredicts, etc.) to
+ STALLED_CYCLES_FRONTEND.
+
+ We could start multiple performance counters and combine everything
+ afterwards, but that makes it complicated, and users can always
+ use PERF_TYPE_RAW.
+
+ Note that I$ cache misses aren't counted by either of the two! */
+ [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",
+};
+
+#define C(_x) PERF_COUNT_HW_CACHE_##_x
+#define CACHE_OP_UNSUPPORTED 0xffff
+
+
+static const unsigned arc_pmu_cache_map[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
+ [C(L1D)] = {
+ [C(OP_READ)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = PERF_COUNT_ARC_DCLM,
+ },
+ [C(OP_WRITE)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = PERF_COUNT_ARC_DCSM,
+ },
+ [C(OP_PREFETCH)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ },
+ [C(L1I)] = {
+ [C(OP_READ)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = PERF_COUNT_ARC_ICM,
+ },
+ [C(OP_WRITE)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ [C(OP_PREFETCH)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ },
+ [C(LL)] = {
+ [C(OP_READ)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ [C(OP_WRITE)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ [C(OP_PREFETCH)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ },
+ [C(DTLB)] = {
+ [C(OP_READ)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = PERF_COUNT_ARC_EDTLB,
+ },
+ [C(OP_WRITE)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ [C(OP_PREFETCH)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ },
+ [C(ITLB)] = {
+ [C(OP_READ)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = PERF_COUNT_ARC_EITLB,
+ },
+ [C(OP_WRITE)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ [C(OP_PREFETCH)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ },
+ [C(BPU)] = {
+ [C(OP_READ)] = {
+ [C(RESULT_ACCESS)] = PERF_COUNT_HW_BRANCH_INSTRUCTIONS,
+ [C(RESULT_MISS)] = PERF_COUNT_HW_BRANCH_MISSES,
+ },
+ [C(OP_WRITE)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ [C(OP_PREFETCH)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ },
+ [C(NODE)] = {
+ [C(OP_READ)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ [C(OP_WRITE)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ [C(OP_PREFETCH)] = {
+ [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
+ [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
+ },
+ },
+};
#endif /* __ASM_PERF_EVENT_H */
diff --git a/arch/arc/kernel/Makefile b/arch/arc/kernel/Makefile
index c242ef0..8004b4f 100644
--- a/arch/arc/kernel/Makefile
+++ b/arch/arc/kernel/Makefile
@@ -19,6 +19,7 @@ obj-$(CONFIG_KPROBES) += kprobes.o
obj-$(CONFIG_ARC_MISALIGN_ACCESS) += unaligned.o
obj-$(CONFIG_KGDB) += kgdb.o
obj-$(CONFIG_ARC_METAWARE_HLINK) += arc_hostlink.o
+obj-$(CONFIG_PERF_EVENTS) += perf_event.o

obj-$(CONFIG_ARC_FPU_SAVE_RESTORE) += fpu.o
CFLAGS_fpu.o += -mdpfp
diff --git a/arch/arc/kernel/perf_event.c b/arch/arc/kernel/perf_event.c
new file mode 100644
index 0000000..2733006
--- /dev/null
+++ b/arch/arc/kernel/perf_event.c
@@ -0,0 +1,323 @@
+/*
+ * Linux performance counter support for ARC
+ *
+ * Copyright (C) 2013 Synopsys, Inc. (http://www.synopsys.com)
+ *
+ * This code is inspired by the perf support of various other architectures.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ *
+ * TODO:
+ * - support for RAW events
+ * - SMP
+ */
+#include <linux/errno.h>
+#include <linux/module.h>
+#include <linux/of.h>
+#include <linux/perf_event.h>
+#include <linux/platform_device.h>
+#include <asm/arcregs.h>
+#define DEBUG
+
+struct arc_pmu {
+ struct pmu pmu;
+ int counter_size; /* in bits */
+ int n_counters;
+ unsigned long used_mask[BITS_TO_LONGS(ARC_PMU_MAX_HWEVENTS)];
+ int ev_hw_idx[PERF_COUNT_ARC_HW_MAX];
+};
+
+/* read counter #idx; note that counter# != event# on ARC! */
+static uint64_t arc_pmu_read_counter(int idx)
+{
+ uint32_t tmp;
+ uint64_t result;
+
+ /* ARC supports making 'snapshots' of the counters, so we don't
+ need to care about counters wrapping to 0 underneath our feet */
+ write_aux_reg(ARC_REG_PCT_INDEX, idx);
+ tmp = read_aux_reg(ARC_REG_PCT_CONTROL);
+ write_aux_reg(ARC_REG_PCT_CONTROL, tmp | ARC_REG_PCT_CONTROL_SN);
+ result = (uint64_t) (read_aux_reg(ARC_REG_PCT_SNAPH)) << 32;
+ result |= read_aux_reg(ARC_REG_PCT_SNAPL);
+
+ return result;
+}
+
+static void arc_perf_event_update(struct perf_event *event,
+ struct hw_perf_event *hwc, int idx)
+{
+ struct arc_pmu *arc_pmu = container_of(event->pmu, struct arc_pmu, pmu);
+ uint64_t prev_raw_count, new_raw_count;
+ int64_t delta;
+
+ do {
+ prev_raw_count = local64_read(&hwc->prev_count);
+ new_raw_count = arc_pmu_read_counter(idx);
+ } while (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
+ new_raw_count) != prev_raw_count);
+
+ delta = (new_raw_count - prev_raw_count) &
+ ((1ULL << arc_pmu->counter_size) - 1ULL);
+
+ local64_add(delta, &event->count);
+ local64_sub(delta, &hwc->period_left);
+}
+
+static void arc_pmu_read(struct perf_event *event)
+{
+ arc_perf_event_update(event, &event->hw, event->hw.idx);
+}
+
+static int arc_pmu_cache_event(u64 config)
+{
+ unsigned int cache_type, cache_op, cache_result;
+ int ret;
+
+ cache_type = (config >> 0) & 0xff;
+ cache_op = (config >> 8) & 0xff;
+ cache_result = (config >> 16) & 0xff;
+ if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
+ return -EINVAL;
+ if (cache_type >= PERF_COUNT_HW_CACHE_OP_MAX)
+ return -EINVAL;
+ if (cache_type >= PERF_COUNT_HW_CACHE_RESULT_MAX)
+ return -EINVAL;
+
+ ret = arc_pmu_cache_map[cache_type][cache_op][cache_result];
+
+ if (ret == CACHE_OP_UNSUPPORTED)
+ return -ENOENT;
+
+ return ret;
+}
+
+/* initializes hw_perf_event structure if event is supported */
+static int arc_pmu_event_init(struct perf_event *event)
+{
+ struct arc_pmu *arc_pmu = container_of(event->pmu, struct arc_pmu, pmu);
+ struct hw_perf_event *hwc = &event->hw;
+ int ret;
+
+ switch (event->attr.type) {
+ case PERF_TYPE_HARDWARE:
+ if (event->attr.config >= PERF_COUNT_HW_MAX)
+ return -ENOENT;
+ if (arc_pmu->ev_hw_idx[event->attr.config] < 0)
+ return -ENOENT;
+ hwc->config = arc_pmu->ev_hw_idx[event->attr.config];
+ pr_debug("initializing event %d with cfg %d\n",
+ (int) event->attr.config, (int) hwc->config);
+ return 0;
+ case PERF_TYPE_HW_CACHE:
+ ret = arc_pmu_cache_event(event->attr.config);
+ if (ret < 0)
+ return ret;
+ hwc->config = arc_pmu->ev_hw_idx[ret];
+ return 0;
+ default:
+ return -ENOENT;
+ }
+}
+
+/* starts all counters */
+static void arc_pmu_enable(struct pmu *pmu)
+{
+ uint32_t tmp;
+ tmp = read_aux_reg(ARC_REG_PCT_CONTROL);
+ write_aux_reg(ARC_REG_PCT_CONTROL, (tmp & 0xffff0000) | 0x1);
+}
+
+/* stops all counters */
+static void arc_pmu_disable(struct pmu *pmu)
+{
+ uint32_t tmp;
+ tmp = read_aux_reg(ARC_REG_PCT_CONTROL);
+ write_aux_reg(ARC_REG_PCT_CONTROL, (tmp & 0xffff0000) | 0x0);
+}
+
+/* assigns hardware counter to hardware condition.
+ note that there is no separate start/stop mechanism;
+ stopping is achieved by assigning the 'never' condition */
+static void arc_pmu_start(struct perf_event *event, int flags)
+{
+ struct hw_perf_event *hwc = &event->hw;
+ int idx = hwc->idx;
+
+ if (WARN_ON_ONCE(idx == -1))
+ return;
+
+ if (flags & PERF_EF_RELOAD)
+ WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
+
+ event->hw.state = 0;
+
+ /* enable ARC pmu here */
+ write_aux_reg(ARC_REG_PCT_INDEX, idx);
+ write_aux_reg(ARC_REG_PCT_CONFIG, hwc->config);
+}
+
+static void arc_pmu_stop(struct perf_event *event, int flags)
+{
+ struct hw_perf_event *hwc = &event->hw;
+ int idx = hwc->idx;
+
+ if (!(event->hw.state & PERF_HES_STOPPED)) {
+ /* stop ARC pmu here */
+ write_aux_reg(ARC_REG_PCT_INDEX, idx);
+ /* condition code #0 is always "never" */
+ write_aux_reg(ARC_REG_PCT_CONFIG, 0);
+
+ event->hw.state |= PERF_HES_STOPPED;
+ }
+
+ if ((flags & PERF_EF_UPDATE) &&
+ !(event->hw.state & PERF_HES_UPTODATE)) {
+ arc_perf_event_update(event, &event->hw, idx);
+ event->hw.state |= PERF_HES_UPTODATE;
+ }
+}
+
+static void arc_pmu_del(struct perf_event *event, int flags)
+{
+ struct arc_pmu *arc_pmu = container_of(event->pmu, struct arc_pmu, pmu);
+
+ arc_pmu_stop(event, PERF_EF_UPDATE);
+ __clear_bit(event->hw.idx, arc_pmu->used_mask);
+
+ perf_event_update_userpage(event);
+}
+
+/* allocate hardware counter and optionally start counting */
+static int arc_pmu_add(struct perf_event *event, int flags)
+{
+ struct arc_pmu *arc_pmu = container_of(event->pmu, struct arc_pmu, pmu);
+ struct hw_perf_event *hwc = &event->hw;
+ int idx = hwc->idx;
+
+ if (__test_and_set_bit(idx, arc_pmu->used_mask)) {
+ idx = find_first_zero_bit(arc_pmu->used_mask,
+ arc_pmu->n_counters);
+ if (idx == arc_pmu->n_counters)
+ return -EAGAIN;
+
+ __set_bit(idx, arc_pmu->used_mask);
+ hwc->idx = idx;
+ }
+
+ write_aux_reg(ARC_REG_PCT_INDEX, idx);
+ write_aux_reg(ARC_REG_PCT_CONFIG, 0);
+ write_aux_reg(ARC_REG_PCT_COUNTL, 0);
+ write_aux_reg(ARC_REG_PCT_COUNTH, 0);
+ local64_set(&hwc->prev_count, 0);
+
+ hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
+ if (flags & PERF_EF_START)
+ arc_pmu_start(event, PERF_EF_RELOAD);
+
+ perf_event_update_userpage(event);
+
+ return 0;
+}
+
+static int arc_pmu_device_probe(struct platform_device *pdev)
+{
+ int ret;
+ struct arc_pmu *arc_pmu;
+ struct arc_reg_pct_build pct_bcr;
+ struct arc_reg_cc_build cc_bcr;
+ int i, j;
+
+ union cc_name {
+ struct {
+ uint32_t word0, word1;
+ char sentinel;
+ } indiv;
+ char str[9];
+ } cc_name;
+
+
+ READ_BCR(ARC_REG_PCT_BUILD, pct_bcr);
+ if (!pct_bcr.v) {
+ pr_err("This core does not have performance counters!\n");
+ return -ENODEV;
+ }
+
+ arc_pmu = devm_kzalloc(&pdev->dev, sizeof(struct arc_pmu),
+ GFP_KERNEL);
+ if (!arc_pmu)
+ return -ENOMEM;
+
+ arc_pmu->n_counters = pct_bcr.c;
+ BUG_ON(arc_pmu->n_counters > ARC_PMU_MAX_HWEVENTS);
+
+ arc_pmu->counter_size = 32 + (pct_bcr.s << 4);
+ pr_info("ARC PMU found with %d counters of size %d bits\n",
+ arc_pmu->n_counters, arc_pmu->counter_size);
+
+ READ_BCR(ARC_REG_CC_BUILD, cc_bcr);
+
+ if (!cc_bcr.v)
+ pr_err("Strange! Performance counters exist, but no countable conditions?\n");
+
+ pr_info("ARC PMU has %d countable conditions\n", cc_bcr.c);
+
+ cc_name.str[8] = 0;
+ for (i = 0; i < PERF_COUNT_HW_MAX; i++)
+ arc_pmu->ev_hw_idx[i] = -1;
+
+ for (j = 0; j < cc_bcr.c; j++) {
+ write_aux_reg(ARC_REG_CC_INDEX, j);
+ cc_name.indiv.word0 = read_aux_reg(ARC_REG_CC_NAME0);
+ cc_name.indiv.word1 = read_aux_reg(ARC_REG_CC_NAME1);
+ for (i = 0; i < ARRAY_SIZE(arc_pmu_ev_hw_map); i++) {
+ if (arc_pmu_ev_hw_map[i] &&
+ !strcmp(arc_pmu_ev_hw_map[i], cc_name.str) &&
+ strlen(arc_pmu_ev_hw_map[i])) {
+ pr_debug("mapping %d to idx %d with name %s\n",
+ i, j, cc_name.str);
+ arc_pmu->ev_hw_idx[i] = j;
+ }
+ }
+ }
+
+ arc_pmu->pmu = (struct pmu) {
+ .pmu_enable = arc_pmu_enable,
+ .pmu_disable = arc_pmu_disable,
+ .event_init = arc_pmu_event_init,
+ .add = arc_pmu_add,
+ .del = arc_pmu_del,
+ .start = arc_pmu_start,
+ .stop = arc_pmu_stop,
+ .read = arc_pmu_read,
+ };
+
+ ret = perf_pmu_register(&arc_pmu->pmu, pdev->name, PERF_TYPE_RAW);
+
+ return ret;
+}
+
+#ifdef CONFIG_OF
+static const struct of_device_id arc_pmu_match[] = {
+ { .compatible = "snps,arc_pmu" },
+ {},
+};
+MODULE_DEVICE_TABLE(of, arc_pmu_match);
+#endif
+
+static struct platform_driver arc_pmu_driver = {
+ .driver = {
+ .name = "arc_pmu",
+ .of_match_table = of_match_ptr(arc_pmu_match),
+ },
+ .probe = arc_pmu_device_probe,
+};
+
+module_platform_driver(arc_pmu_driver);
+
+MODULE_LICENSE("GPL");
+MODULE_AUTHOR("Mischa Jonker <[email protected]>");
+MODULE_DESCRIPTION("ARC PMU driver");
--
1.7.9.5


2013-08-29 06:41:30

by Ingo Molnar

[permalink] [raw]
Subject: Re: [PATCH] ARC: Add perf support


* Mischa Jonker <[email protected]> wrote:

> +/* 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 */
> +static const char * const arc_pmu_ev_hw_map[] = {
> + /* cycles not in halted state */
> + [PERF_COUNT_HW_CPU_CYCLES] = "crun",
> + /* reference cycles not in halted state; also "crun" as we don't
> + do Dynamic Voltage/Frequency Scaling (yet) */
> + [PERF_COUNT_HW_REF_CPU_CYCLES] = "crun",
> + /* Unclear what this means, Intel uses 0x013c, which according to
> + their datasheet means "unhalted reference cycles". Don't ask
> + me what the difference is with PERF_COUNT_HW_REF_CPU_CYCLES */
> + [PERF_COUNT_HW_BUS_CYCLES] = "crun",
> + [PERF_COUNT_HW_INSTRUCTIONS] = "iall",
> + [PERF_COUNT_HW_BRANCH_MISSES] = "bpfail",
> + [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = "ijmp",
> + /* ARC can either measure stalls per pipeline stage, or all stalls
> + combined; for now we assign all stalls to STALLED_CYCLES_BACKEND
> + and all pipeline flushes (e.g. caused by mispredicts, etc.) to
> + STALLED_CYCLES_FRONTEND.
> +
> + We could start multiple performance counters and combine everything
> + afterwards, but that makes it complicated, and users can always
> + use PERF_TYPE_RAW.
> +
> + Note that I$ cache misses aren't counted by either of the two! */

Haven't really looked at the patch in detail, just noticed the following
nit, please use the customary (multi-line) comment style:

/*
* Comment .....
* ...... goes here.
*/

specified in Documentation/CodingStyle.

Thanks,

Ingo