2 * Linux performance counter support for ARC700 series
4 * Copyright (C) 2013-2015 Synopsys, Inc. (www.synopsys.com)
6 * This code is inspired by the perf support of various other architectures.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 as
10 * published by the Free Software Foundation.
13 #include <linux/errno.h>
14 #include <linux/interrupt.h>
15 #include <linux/module.h>
17 #include <linux/perf_event.h>
18 #include <linux/platform_device.h>
19 #include <asm/arcregs.h>
20 #include <asm/stacktrace.h>
27 int ev_hw_idx[PERF_COUNT_ARC_HW_MAX];
32 * A 1 bit for an index indicates that the counter is being used for
33 * an event. A 0 means that the counter can be used.
35 unsigned long used_mask[BITS_TO_LONGS(ARC_PERF_MAX_COUNTERS)];
38 * The events that are active on the PMU for the given index.
40 struct perf_event *act_counter[ARC_PERF_MAX_COUNTERS];
43 struct arc_callchain_trace {
48 static int callchain_trace(unsigned int addr, void *data)
50 struct arc_callchain_trace *ctrl = data;
51 struct perf_callchain_entry *entry = ctrl->perf_stuff;
52 perf_callchain_store(entry, addr);
54 if (ctrl->depth++ < 3)
61 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
63 struct arc_callchain_trace ctrl = {
68 arc_unwind_core(NULL, regs, callchain_trace, &ctrl);
72 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
75 * User stack can't be unwound trivially with kernel dwarf unwinder
76 * So for now just record the user PC
78 perf_callchain_store(entry, instruction_pointer(regs));
81 static struct arc_pmu *arc_pmu;
82 static DEFINE_PER_CPU(struct arc_pmu_cpu, arc_pmu_cpu);
84 /* read counter #idx; note that counter# != event# on ARC! */
85 static uint64_t arc_pmu_read_counter(int idx)
91 * ARC supports making 'snapshots' of the counters, so we don't
92 * need to care about counters wrapping to 0 underneath our feet
94 write_aux_reg(ARC_REG_PCT_INDEX, idx);
95 tmp = read_aux_reg(ARC_REG_PCT_CONTROL);
96 write_aux_reg(ARC_REG_PCT_CONTROL, tmp | ARC_REG_PCT_CONTROL_SN);
97 result = (uint64_t) (read_aux_reg(ARC_REG_PCT_SNAPH)) << 32;
98 result |= read_aux_reg(ARC_REG_PCT_SNAPL);
103 static void arc_perf_event_update(struct perf_event *event,
104 struct hw_perf_event *hwc, int idx)
106 uint64_t prev_raw_count = local64_read(&hwc->prev_count);
107 uint64_t new_raw_count = arc_pmu_read_counter(idx);
108 int64_t delta = new_raw_count - prev_raw_count;
111 * We don't afaraid of hwc->prev_count changing beneath our feet
112 * because there's no way for us to re-enter this function anytime.
114 local64_set(&hwc->prev_count, new_raw_count);
115 local64_add(delta, &event->count);
116 local64_sub(delta, &hwc->period_left);
119 static void arc_pmu_read(struct perf_event *event)
121 arc_perf_event_update(event, &event->hw, event->hw.idx);
124 static int arc_pmu_cache_event(u64 config)
126 unsigned int cache_type, cache_op, cache_result;
129 cache_type = (config >> 0) & 0xff;
130 cache_op = (config >> 8) & 0xff;
131 cache_result = (config >> 16) & 0xff;
132 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
134 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
136 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
139 ret = arc_pmu_cache_map[cache_type][cache_op][cache_result];
141 if (ret == CACHE_OP_UNSUPPORTED)
144 pr_debug("init cache event: type/op/result %d/%d/%d with h/w %d \'%s\'\n",
145 cache_type, cache_op, cache_result, ret,
146 arc_pmu_ev_hw_map[ret]);
151 /* initializes hw_perf_event structure if event is supported */
152 static int arc_pmu_event_init(struct perf_event *event)
154 struct hw_perf_event *hwc = &event->hw;
157 if (!is_sampling_event(event)) {
158 hwc->sample_period = arc_pmu->max_period;
159 hwc->last_period = hwc->sample_period;
160 local64_set(&hwc->period_left, hwc->sample_period);
165 if (is_isa_arcv2()) {
166 /* "exclude user" means "count only kernel" */
167 if (event->attr.exclude_user)
168 hwc->config |= ARC_REG_PCT_CONFIG_KERN;
170 /* "exclude kernel" means "count only user" */
171 if (event->attr.exclude_kernel)
172 hwc->config |= ARC_REG_PCT_CONFIG_USER;
175 switch (event->attr.type) {
176 case PERF_TYPE_HARDWARE:
177 if (event->attr.config >= PERF_COUNT_HW_MAX)
179 if (arc_pmu->ev_hw_idx[event->attr.config] < 0)
181 hwc->config |= arc_pmu->ev_hw_idx[event->attr.config];
182 pr_debug("init event %d with h/w %d \'%s\'\n",
183 (int) event->attr.config, (int) hwc->config,
184 arc_pmu_ev_hw_map[event->attr.config]);
187 case PERF_TYPE_HW_CACHE:
188 ret = arc_pmu_cache_event(event->attr.config);
191 hwc->config |= arc_pmu->ev_hw_idx[ret];
198 /* starts all counters */
199 static void arc_pmu_enable(struct pmu *pmu)
202 tmp = read_aux_reg(ARC_REG_PCT_CONTROL);
203 write_aux_reg(ARC_REG_PCT_CONTROL, (tmp & 0xffff0000) | 0x1);
206 /* stops all counters */
207 static void arc_pmu_disable(struct pmu *pmu)
210 tmp = read_aux_reg(ARC_REG_PCT_CONTROL);
211 write_aux_reg(ARC_REG_PCT_CONTROL, (tmp & 0xffff0000) | 0x0);
214 static int arc_pmu_event_set_period(struct perf_event *event)
216 struct hw_perf_event *hwc = &event->hw;
217 s64 left = local64_read(&hwc->period_left);
218 s64 period = hwc->sample_period;
223 if (unlikely(left <= -period)) {
224 /* left underflowed by more than period. */
226 local64_set(&hwc->period_left, left);
227 hwc->last_period = period;
229 } else if (unlikely(left <= 0)) {
230 /* left underflowed by less than period. */
232 local64_set(&hwc->period_left, left);
233 hwc->last_period = period;
237 if (left > arc_pmu->max_period)
238 left = arc_pmu->max_period;
240 value = arc_pmu->max_period - left;
241 local64_set(&hwc->prev_count, value);
244 write_aux_reg(ARC_REG_PCT_INDEX, idx);
247 write_aux_reg(ARC_REG_PCT_COUNTL, (u32)value);
248 write_aux_reg(ARC_REG_PCT_COUNTH, (value >> 32));
250 perf_event_update_userpage(event);
256 * Assigns hardware counter to hardware condition.
257 * Note that there is no separate start/stop mechanism;
258 * stopping is achieved by assigning the 'never' condition
260 static void arc_pmu_start(struct perf_event *event, int flags)
262 struct hw_perf_event *hwc = &event->hw;
265 if (WARN_ON_ONCE(idx == -1))
268 if (flags & PERF_EF_RELOAD)
269 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
273 arc_pmu_event_set_period(event);
275 /* Enable interrupt for this counter */
276 if (is_sampling_event(event))
277 write_aux_reg(ARC_REG_PCT_INT_CTRL,
278 read_aux_reg(ARC_REG_PCT_INT_CTRL) | (1 << idx));
280 /* enable ARC pmu here */
281 write_aux_reg(ARC_REG_PCT_INDEX, idx); /* counter # */
282 write_aux_reg(ARC_REG_PCT_CONFIG, hwc->config); /* condition */
285 static void arc_pmu_stop(struct perf_event *event, int flags)
287 struct hw_perf_event *hwc = &event->hw;
290 /* Disable interrupt for this counter */
291 if (is_sampling_event(event)) {
293 * Reset interrupt flag by writing of 1. This is required
294 * to make sure pending interrupt was not left.
296 write_aux_reg(ARC_REG_PCT_INT_ACT, 1 << idx);
297 write_aux_reg(ARC_REG_PCT_INT_CTRL,
298 read_aux_reg(ARC_REG_PCT_INT_CTRL) & ~(1 << idx));
301 if (!(event->hw.state & PERF_HES_STOPPED)) {
302 /* stop ARC pmu here */
303 write_aux_reg(ARC_REG_PCT_INDEX, idx);
305 /* condition code #0 is always "never" */
306 write_aux_reg(ARC_REG_PCT_CONFIG, 0);
308 event->hw.state |= PERF_HES_STOPPED;
311 if ((flags & PERF_EF_UPDATE) &&
312 !(event->hw.state & PERF_HES_UPTODATE)) {
313 arc_perf_event_update(event, &event->hw, idx);
314 event->hw.state |= PERF_HES_UPTODATE;
318 static void arc_pmu_del(struct perf_event *event, int flags)
320 struct arc_pmu_cpu *pmu_cpu = this_cpu_ptr(&arc_pmu_cpu);
322 arc_pmu_stop(event, PERF_EF_UPDATE);
323 __clear_bit(event->hw.idx, pmu_cpu->used_mask);
325 pmu_cpu->act_counter[event->hw.idx] = 0;
327 perf_event_update_userpage(event);
330 /* allocate hardware counter and optionally start counting */
331 static int arc_pmu_add(struct perf_event *event, int flags)
333 struct arc_pmu_cpu *pmu_cpu = this_cpu_ptr(&arc_pmu_cpu);
334 struct hw_perf_event *hwc = &event->hw;
337 if (__test_and_set_bit(idx, pmu_cpu->used_mask)) {
338 idx = find_first_zero_bit(pmu_cpu->used_mask,
339 arc_pmu->n_counters);
340 if (idx == arc_pmu->n_counters)
343 __set_bit(idx, pmu_cpu->used_mask);
347 write_aux_reg(ARC_REG_PCT_INDEX, idx);
349 pmu_cpu->act_counter[idx] = event;
351 if (is_sampling_event(event)) {
352 /* Mimic full counter overflow as other arches do */
353 write_aux_reg(ARC_REG_PCT_INT_CNTL, (u32)arc_pmu->max_period);
354 write_aux_reg(ARC_REG_PCT_INT_CNTH,
355 (arc_pmu->max_period >> 32));
358 write_aux_reg(ARC_REG_PCT_CONFIG, 0);
359 write_aux_reg(ARC_REG_PCT_COUNTL, 0);
360 write_aux_reg(ARC_REG_PCT_COUNTH, 0);
361 local64_set(&hwc->prev_count, 0);
363 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
364 if (flags & PERF_EF_START)
365 arc_pmu_start(event, PERF_EF_RELOAD);
367 perf_event_update_userpage(event);
372 #ifdef CONFIG_ISA_ARCV2
373 static irqreturn_t arc_pmu_intr(int irq, void *dev)
375 struct perf_sample_data data;
376 struct arc_pmu_cpu *pmu_cpu = this_cpu_ptr(&arc_pmu_cpu);
377 struct pt_regs *regs;
381 arc_pmu_disable(&arc_pmu->pmu);
383 active_ints = read_aux_reg(ARC_REG_PCT_INT_ACT);
385 regs = get_irq_regs();
387 for (idx = 0; idx < arc_pmu->n_counters; idx++) {
388 struct perf_event *event = pmu_cpu->act_counter[idx];
389 struct hw_perf_event *hwc;
391 if (!(active_ints & (1 << idx)))
394 /* Reset interrupt flag by writing of 1 */
395 write_aux_reg(ARC_REG_PCT_INT_ACT, 1 << idx);
398 * On reset of "interrupt active" bit corresponding
399 * "interrupt enable" bit gets automatically reset as well.
400 * Now we need to re-enable interrupt for the counter.
402 write_aux_reg(ARC_REG_PCT_INT_CTRL,
403 read_aux_reg(ARC_REG_PCT_INT_CTRL) | (1 << idx));
407 WARN_ON_ONCE(hwc->idx != idx);
409 arc_perf_event_update(event, &event->hw, event->hw.idx);
410 perf_sample_data_init(&data, 0, hwc->last_period);
411 if (!arc_pmu_event_set_period(event))
414 if (perf_event_overflow(event, &data, regs))
415 arc_pmu_stop(event, 0);
418 arc_pmu_enable(&arc_pmu->pmu);
424 static irqreturn_t arc_pmu_intr(int irq, void *dev)
429 #endif /* CONFIG_ISA_ARCV2 */
431 void arc_cpu_pmu_irq_init(void)
433 struct arc_pmu_cpu *pmu_cpu = this_cpu_ptr(&arc_pmu_cpu);
435 arc_request_percpu_irq(arc_pmu->irq, smp_processor_id(), arc_pmu_intr,
436 "ARC perf counters", pmu_cpu);
438 /* Clear all pending interrupt flags */
439 write_aux_reg(ARC_REG_PCT_INT_ACT, 0xffffffff);
442 static int arc_pmu_device_probe(struct platform_device *pdev)
444 struct arc_reg_pct_build pct_bcr;
445 struct arc_reg_cc_build cc_bcr;
446 int i, j, has_interrupts;
447 int counter_size; /* in bits */
451 uint32_t word0, word1;
458 READ_BCR(ARC_REG_PCT_BUILD, pct_bcr);
460 pr_err("This core does not have performance counters!\n");
463 BUG_ON(pct_bcr.c > ARC_PERF_MAX_COUNTERS);
465 READ_BCR(ARC_REG_CC_BUILD, cc_bcr);
466 BUG_ON(!cc_bcr.v); /* Counters exist but No countable conditions ? */
468 arc_pmu = devm_kzalloc(&pdev->dev, sizeof(struct arc_pmu), GFP_KERNEL);
472 has_interrupts = is_isa_arcv2() ? pct_bcr.i : 0;
474 arc_pmu->n_counters = pct_bcr.c;
475 counter_size = 32 + (pct_bcr.s << 4);
477 arc_pmu->max_period = (1ULL << counter_size) / 2 - 1ULL;
479 pr_info("ARC perf\t: %d counters (%d bits), %d conditions%s\n",
480 arc_pmu->n_counters, counter_size, cc_bcr.c,
481 has_interrupts ? ", [overflow IRQ support]":"");
484 for (i = 0; i < PERF_COUNT_ARC_HW_MAX; i++)
485 arc_pmu->ev_hw_idx[i] = -1;
487 /* loop thru all available h/w condition indexes */
488 for (j = 0; j < cc_bcr.c; j++) {
489 write_aux_reg(ARC_REG_CC_INDEX, j);
490 cc_name.indiv.word0 = read_aux_reg(ARC_REG_CC_NAME0);
491 cc_name.indiv.word1 = read_aux_reg(ARC_REG_CC_NAME1);
493 /* See if it has been mapped to a perf event_id */
494 for (i = 0; i < ARRAY_SIZE(arc_pmu_ev_hw_map); i++) {
495 if (arc_pmu_ev_hw_map[i] &&
496 !strcmp(arc_pmu_ev_hw_map[i], cc_name.str) &&
497 strlen(arc_pmu_ev_hw_map[i])) {
498 pr_debug("mapping perf event %2d to h/w event \'%8s\' (idx %d)\n",
500 arc_pmu->ev_hw_idx[i] = j;
505 arc_pmu->pmu = (struct pmu) {
506 .pmu_enable = arc_pmu_enable,
507 .pmu_disable = arc_pmu_disable,
508 .event_init = arc_pmu_event_init,
511 .start = arc_pmu_start,
512 .stop = arc_pmu_stop,
513 .read = arc_pmu_read,
516 if (has_interrupts) {
517 int irq = platform_get_irq(pdev, 0);
521 pr_err("Cannot get IRQ number for the platform\n");
528 * arc_cpu_pmu_irq_init() needs to be called on all cores for
529 * their respective local PMU.
530 * However we use opencoded on_each_cpu() to ensure it is called
531 * on core0 first, so that arc_request_percpu_irq() sets up
532 * AUTOEN etc. Otherwise enable_percpu_irq() fails to enable
533 * perf IRQ on non master cores.
534 * see arc_request_percpu_irq()
537 local_irq_save(flags);
538 arc_cpu_pmu_irq_init();
539 local_irq_restore(flags);
540 smp_call_function((smp_call_func_t)arc_cpu_pmu_irq_init, 0, 1);
543 /* Clean all pending interrupt flags */
544 write_aux_reg(ARC_REG_PCT_INT_ACT, 0xffffffff);
546 arc_pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
548 return perf_pmu_register(&arc_pmu->pmu, pdev->name, PERF_TYPE_RAW);
552 static const struct of_device_id arc_pmu_match[] = {
553 { .compatible = "snps,arc700-pct" },
554 { .compatible = "snps,archs-pct" },
557 MODULE_DEVICE_TABLE(of, arc_pmu_match);
560 static struct platform_driver arc_pmu_driver = {
563 .of_match_table = of_match_ptr(arc_pmu_match),
565 .probe = arc_pmu_device_probe,
568 module_platform_driver(arc_pmu_driver);
570 MODULE_LICENSE("GPL");
571 MODULE_AUTHOR("Mischa Jonker <mjonker@synopsys.com>");
572 MODULE_DESCRIPTION("ARC PMU driver");