1 // SPDX-License-Identifier: GPL-2.0+
3 // Linux performance counter support for ARC CPUs.
4 // This code is inspired by the perf support of various other architectures.
6 // Copyright (C) 2013-2018 Synopsys, Inc. (www.synopsys.com)
8 #include <linux/errno.h>
9 #include <linux/interrupt.h>
10 #include <linux/module.h>
12 #include <linux/perf_event.h>
13 #include <linux/platform_device.h>
14 #include <asm/arcregs.h>
15 #include <asm/stacktrace.h>
17 /* HW holds 8 symbols + one for null terminator */
18 #define ARCPMU_EVENT_NAME_LEN 9
20 enum arc_pmu_attr_groups
{
21 ARCPMU_ATTR_GR_EVENTS
,
22 ARCPMU_ATTR_GR_FORMATS
,
26 struct arc_pmu_raw_event_entry
{
27 char name
[ARCPMU_EVENT_NAME_LEN
];
36 int ev_hw_idx
[PERF_COUNT_ARC_HW_MAX
];
38 struct arc_pmu_raw_event_entry
*raw_entry
;
39 struct attribute
**attrs
;
40 struct perf_pmu_events_attr
*attr
;
41 const struct attribute_group
*attr_groups
[ARCPMU_NR_ATTR_GR
+ 1];
46 * A 1 bit for an index indicates that the counter is being used for
47 * an event. A 0 means that the counter can be used.
49 unsigned long used_mask
[BITS_TO_LONGS(ARC_PERF_MAX_COUNTERS
)];
52 * The events that are active on the PMU for the given index.
54 struct perf_event
*act_counter
[ARC_PERF_MAX_COUNTERS
];
57 struct arc_callchain_trace
{
62 static int callchain_trace(unsigned int addr
, void *data
)
64 struct arc_callchain_trace
*ctrl
= data
;
65 struct perf_callchain_entry_ctx
*entry
= ctrl
->perf_stuff
;
67 perf_callchain_store(entry
, addr
);
69 if (ctrl
->depth
++ < 3)
75 void perf_callchain_kernel(struct perf_callchain_entry_ctx
*entry
,
78 struct arc_callchain_trace ctrl
= {
83 arc_unwind_core(NULL
, regs
, callchain_trace
, &ctrl
);
86 void perf_callchain_user(struct perf_callchain_entry_ctx
*entry
,
90 * User stack can't be unwound trivially with kernel dwarf unwinder
91 * So for now just record the user PC
93 perf_callchain_store(entry
, instruction_pointer(regs
));
96 static struct arc_pmu
*arc_pmu
;
97 static DEFINE_PER_CPU(struct arc_pmu_cpu
, arc_pmu_cpu
);
99 /* read counter #idx; note that counter# != event# on ARC! */
100 static u64
arc_pmu_read_counter(int idx
)
106 * ARC supports making 'snapshots' of the counters, so we don't
107 * need to care about counters wrapping to 0 underneath our feet
109 write_aux_reg(ARC_REG_PCT_INDEX
, idx
);
110 tmp
= read_aux_reg(ARC_REG_PCT_CONTROL
);
111 write_aux_reg(ARC_REG_PCT_CONTROL
, tmp
| ARC_REG_PCT_CONTROL_SN
);
112 result
= (u64
) (read_aux_reg(ARC_REG_PCT_SNAPH
)) << 32;
113 result
|= read_aux_reg(ARC_REG_PCT_SNAPL
);
118 static void arc_perf_event_update(struct perf_event
*event
,
119 struct hw_perf_event
*hwc
, int idx
)
121 u64 prev_raw_count
= local64_read(&hwc
->prev_count
);
122 u64 new_raw_count
= arc_pmu_read_counter(idx
);
123 s64 delta
= new_raw_count
- prev_raw_count
;
126 * We aren't afraid of hwc->prev_count changing beneath our feet
127 * because there's no way for us to re-enter this function anytime.
129 local64_set(&hwc
->prev_count
, new_raw_count
);
130 local64_add(delta
, &event
->count
);
131 local64_sub(delta
, &hwc
->period_left
);
134 static void arc_pmu_read(struct perf_event
*event
)
136 arc_perf_event_update(event
, &event
->hw
, event
->hw
.idx
);
139 static int arc_pmu_cache_event(u64 config
)
141 unsigned int cache_type
, cache_op
, cache_result
;
144 cache_type
= (config
>> 0) & 0xff;
145 cache_op
= (config
>> 8) & 0xff;
146 cache_result
= (config
>> 16) & 0xff;
147 if (cache_type
>= PERF_COUNT_HW_CACHE_MAX
)
149 if (cache_op
>= PERF_COUNT_HW_CACHE_OP_MAX
)
151 if (cache_result
>= PERF_COUNT_HW_CACHE_RESULT_MAX
)
154 ret
= arc_pmu_cache_map
[cache_type
][cache_op
][cache_result
];
156 if (ret
== CACHE_OP_UNSUPPORTED
)
159 pr_debug("init cache event: type/op/result %d/%d/%d with h/w %d \'%s\'\n",
160 cache_type
, cache_op
, cache_result
, ret
,
161 arc_pmu_ev_hw_map
[ret
]);
166 /* initializes hw_perf_event structure if event is supported */
167 static int arc_pmu_event_init(struct perf_event
*event
)
169 struct hw_perf_event
*hwc
= &event
->hw
;
172 if (!is_sampling_event(event
)) {
173 hwc
->sample_period
= arc_pmu
->max_period
;
174 hwc
->last_period
= hwc
->sample_period
;
175 local64_set(&hwc
->period_left
, hwc
->sample_period
);
180 if (is_isa_arcv2()) {
181 /* "exclude user" means "count only kernel" */
182 if (event
->attr
.exclude_user
)
183 hwc
->config
|= ARC_REG_PCT_CONFIG_KERN
;
185 /* "exclude kernel" means "count only user" */
186 if (event
->attr
.exclude_kernel
)
187 hwc
->config
|= ARC_REG_PCT_CONFIG_USER
;
190 switch (event
->attr
.type
) {
191 case PERF_TYPE_HARDWARE
:
192 if (event
->attr
.config
>= PERF_COUNT_HW_MAX
)
194 if (arc_pmu
->ev_hw_idx
[event
->attr
.config
] < 0)
196 hwc
->config
|= arc_pmu
->ev_hw_idx
[event
->attr
.config
];
197 pr_debug("init event %d with h/w %08x \'%s\'\n",
198 (int)event
->attr
.config
, (int)hwc
->config
,
199 arc_pmu_ev_hw_map
[event
->attr
.config
]);
202 case PERF_TYPE_HW_CACHE
:
203 ret
= arc_pmu_cache_event(event
->attr
.config
);
206 hwc
->config
|= arc_pmu
->ev_hw_idx
[ret
];
207 pr_debug("init cache event with h/w %08x \'%s\'\n",
208 (int)hwc
->config
, arc_pmu_ev_hw_map
[ret
]);
212 if (event
->attr
.config
>= arc_pmu
->n_events
)
215 hwc
->config
|= event
->attr
.config
;
216 pr_debug("init raw event with idx %lld \'%s\'\n",
218 arc_pmu
->raw_entry
[event
->attr
.config
].name
);
227 /* starts all counters */
228 static void arc_pmu_enable(struct pmu
*pmu
)
231 tmp
= read_aux_reg(ARC_REG_PCT_CONTROL
);
232 write_aux_reg(ARC_REG_PCT_CONTROL
, (tmp
& 0xffff0000) | 0x1);
235 /* stops all counters */
236 static void arc_pmu_disable(struct pmu
*pmu
)
239 tmp
= read_aux_reg(ARC_REG_PCT_CONTROL
);
240 write_aux_reg(ARC_REG_PCT_CONTROL
, (tmp
& 0xffff0000) | 0x0);
243 static int arc_pmu_event_set_period(struct perf_event
*event
)
245 struct hw_perf_event
*hwc
= &event
->hw
;
246 s64 left
= local64_read(&hwc
->period_left
);
247 s64 period
= hwc
->sample_period
;
252 if (unlikely(left
<= -period
)) {
253 /* left underflowed by more than period. */
255 local64_set(&hwc
->period_left
, left
);
256 hwc
->last_period
= period
;
258 } else if (unlikely(left
<= 0)) {
259 /* left underflowed by less than period. */
261 local64_set(&hwc
->period_left
, left
);
262 hwc
->last_period
= period
;
266 if (left
> arc_pmu
->max_period
)
267 left
= arc_pmu
->max_period
;
269 value
= arc_pmu
->max_period
- left
;
270 local64_set(&hwc
->prev_count
, value
);
273 write_aux_reg(ARC_REG_PCT_INDEX
, idx
);
276 write_aux_reg(ARC_REG_PCT_COUNTL
, lower_32_bits(value
));
277 write_aux_reg(ARC_REG_PCT_COUNTH
, upper_32_bits(value
));
279 perf_event_update_userpage(event
);
285 * Assigns hardware counter to hardware condition.
286 * Note that there is no separate start/stop mechanism;
287 * stopping is achieved by assigning the 'never' condition
289 static void arc_pmu_start(struct perf_event
*event
, int flags
)
291 struct hw_perf_event
*hwc
= &event
->hw
;
294 if (WARN_ON_ONCE(idx
== -1))
297 if (flags
& PERF_EF_RELOAD
)
298 WARN_ON_ONCE(!(hwc
->state
& PERF_HES_UPTODATE
));
302 arc_pmu_event_set_period(event
);
304 /* Enable interrupt for this counter */
305 if (is_sampling_event(event
))
306 write_aux_reg(ARC_REG_PCT_INT_CTRL
,
307 read_aux_reg(ARC_REG_PCT_INT_CTRL
) | BIT(idx
));
309 /* enable ARC pmu here */
310 write_aux_reg(ARC_REG_PCT_INDEX
, idx
); /* counter # */
311 write_aux_reg(ARC_REG_PCT_CONFIG
, hwc
->config
); /* condition */
314 static void arc_pmu_stop(struct perf_event
*event
, int flags
)
316 struct hw_perf_event
*hwc
= &event
->hw
;
319 /* Disable interrupt for this counter */
320 if (is_sampling_event(event
)) {
322 * Reset interrupt flag by writing of 1. This is required
323 * to make sure pending interrupt was not left.
325 write_aux_reg(ARC_REG_PCT_INT_ACT
, BIT(idx
));
326 write_aux_reg(ARC_REG_PCT_INT_CTRL
,
327 read_aux_reg(ARC_REG_PCT_INT_CTRL
) & ~BIT(idx
));
330 if (!(event
->hw
.state
& PERF_HES_STOPPED
)) {
331 /* stop ARC pmu here */
332 write_aux_reg(ARC_REG_PCT_INDEX
, idx
);
334 /* condition code #0 is always "never" */
335 write_aux_reg(ARC_REG_PCT_CONFIG
, 0);
337 event
->hw
.state
|= PERF_HES_STOPPED
;
340 if ((flags
& PERF_EF_UPDATE
) &&
341 !(event
->hw
.state
& PERF_HES_UPTODATE
)) {
342 arc_perf_event_update(event
, &event
->hw
, idx
);
343 event
->hw
.state
|= PERF_HES_UPTODATE
;
347 static void arc_pmu_del(struct perf_event
*event
, int flags
)
349 struct arc_pmu_cpu
*pmu_cpu
= this_cpu_ptr(&arc_pmu_cpu
);
351 arc_pmu_stop(event
, PERF_EF_UPDATE
);
352 __clear_bit(event
->hw
.idx
, pmu_cpu
->used_mask
);
354 pmu_cpu
->act_counter
[event
->hw
.idx
] = 0;
356 perf_event_update_userpage(event
);
359 /* allocate hardware counter and optionally start counting */
360 static int arc_pmu_add(struct perf_event
*event
, int flags
)
362 struct arc_pmu_cpu
*pmu_cpu
= this_cpu_ptr(&arc_pmu_cpu
);
363 struct hw_perf_event
*hwc
= &event
->hw
;
366 idx
= ffz(pmu_cpu
->used_mask
[0]);
367 if (idx
== arc_pmu
->n_counters
)
370 __set_bit(idx
, pmu_cpu
->used_mask
);
373 write_aux_reg(ARC_REG_PCT_INDEX
, idx
);
375 pmu_cpu
->act_counter
[idx
] = event
;
377 if (is_sampling_event(event
)) {
378 /* Mimic full counter overflow as other arches do */
379 write_aux_reg(ARC_REG_PCT_INT_CNTL
,
380 lower_32_bits(arc_pmu
->max_period
));
381 write_aux_reg(ARC_REG_PCT_INT_CNTH
,
382 upper_32_bits(arc_pmu
->max_period
));
385 write_aux_reg(ARC_REG_PCT_CONFIG
, 0);
386 write_aux_reg(ARC_REG_PCT_COUNTL
, 0);
387 write_aux_reg(ARC_REG_PCT_COUNTH
, 0);
388 local64_set(&hwc
->prev_count
, 0);
390 hwc
->state
= PERF_HES_UPTODATE
| PERF_HES_STOPPED
;
391 if (flags
& PERF_EF_START
)
392 arc_pmu_start(event
, PERF_EF_RELOAD
);
394 perf_event_update_userpage(event
);
399 #ifdef CONFIG_ISA_ARCV2
400 static irqreturn_t
arc_pmu_intr(int irq
, void *dev
)
402 struct perf_sample_data data
;
403 struct arc_pmu_cpu
*pmu_cpu
= this_cpu_ptr(&arc_pmu_cpu
);
404 struct pt_regs
*regs
;
405 unsigned int active_ints
;
408 arc_pmu_disable(&arc_pmu
->pmu
);
410 active_ints
= read_aux_reg(ARC_REG_PCT_INT_ACT
);
414 regs
= get_irq_regs();
417 struct perf_event
*event
;
418 struct hw_perf_event
*hwc
;
420 idx
= __ffs(active_ints
);
422 /* Reset interrupt flag by writing of 1 */
423 write_aux_reg(ARC_REG_PCT_INT_ACT
, BIT(idx
));
426 * On reset of "interrupt active" bit corresponding
427 * "interrupt enable" bit gets automatically reset as well.
428 * Now we need to re-enable interrupt for the counter.
430 write_aux_reg(ARC_REG_PCT_INT_CTRL
,
431 read_aux_reg(ARC_REG_PCT_INT_CTRL
) | BIT(idx
));
433 event
= pmu_cpu
->act_counter
[idx
];
436 WARN_ON_ONCE(hwc
->idx
!= idx
);
438 arc_perf_event_update(event
, &event
->hw
, event
->hw
.idx
);
439 perf_sample_data_init(&data
, 0, hwc
->last_period
);
440 if (arc_pmu_event_set_period(event
)) {
441 if (perf_event_overflow(event
, &data
, regs
))
442 arc_pmu_stop(event
, 0);
445 active_ints
&= ~BIT(idx
);
446 } while (active_ints
);
449 arc_pmu_enable(&arc_pmu
->pmu
);
455 static irqreturn_t
arc_pmu_intr(int irq
, void *dev
)
460 #endif /* CONFIG_ISA_ARCV2 */
462 static void arc_cpu_pmu_irq_init(void *data
)
464 int irq
= *(int *)data
;
466 enable_percpu_irq(irq
, IRQ_TYPE_NONE
);
468 /* Clear all pending interrupt flags */
469 write_aux_reg(ARC_REG_PCT_INT_ACT
, 0xffffffff);
472 /* Event field occupies the bottom 15 bits of our config field */
473 PMU_FORMAT_ATTR(event
, "config:0-14");
474 static struct attribute
*arc_pmu_format_attrs
[] = {
475 &format_attr_event
.attr
,
479 static struct attribute_group arc_pmu_format_attr_gr
= {
481 .attrs
= arc_pmu_format_attrs
,
484 static ssize_t
arc_pmu_events_sysfs_show(struct device
*dev
,
485 struct device_attribute
*attr
,
488 struct perf_pmu_events_attr
*pmu_attr
;
490 pmu_attr
= container_of(attr
, struct perf_pmu_events_attr
, attr
);
491 return sprintf(page
, "event=0x%04llx\n", pmu_attr
->id
);
495 * We don't add attrs here as we don't have pre-defined list of perf events.
496 * We will generate and add attrs dynamically in probe() after we read HW
499 static struct attribute_group arc_pmu_events_attr_gr
= {
503 static void arc_pmu_add_raw_event_attr(int j
, char *str
)
505 memmove(arc_pmu
->raw_entry
[j
].name
, str
, ARCPMU_EVENT_NAME_LEN
- 1);
506 arc_pmu
->attr
[j
].attr
.attr
.name
= arc_pmu
->raw_entry
[j
].name
;
507 arc_pmu
->attr
[j
].attr
.attr
.mode
= VERIFY_OCTAL_PERMISSIONS(0444);
508 arc_pmu
->attr
[j
].attr
.show
= arc_pmu_events_sysfs_show
;
509 arc_pmu
->attr
[j
].id
= j
;
510 arc_pmu
->attrs
[j
] = &(arc_pmu
->attr
[j
].attr
.attr
);
513 static int arc_pmu_raw_alloc(struct device
*dev
)
515 arc_pmu
->attr
= devm_kmalloc_array(dev
, arc_pmu
->n_events
+ 1,
516 sizeof(*arc_pmu
->attr
), GFP_KERNEL
| __GFP_ZERO
);
520 arc_pmu
->attrs
= devm_kmalloc_array(dev
, arc_pmu
->n_events
+ 1,
521 sizeof(*arc_pmu
->attrs
), GFP_KERNEL
| __GFP_ZERO
);
525 arc_pmu
->raw_entry
= devm_kmalloc_array(dev
, arc_pmu
->n_events
,
526 sizeof(*arc_pmu
->raw_entry
), GFP_KERNEL
| __GFP_ZERO
);
527 if (!arc_pmu
->raw_entry
)
533 static inline bool event_in_hw_event_map(int i
, char *name
)
535 if (!arc_pmu_ev_hw_map
[i
])
538 if (!strlen(arc_pmu_ev_hw_map
[i
]))
541 if (strcmp(arc_pmu_ev_hw_map
[i
], name
))
547 static void arc_pmu_map_hw_event(int j
, char *str
)
551 /* See if HW condition has been mapped to a perf event_id */
552 for (i
= 0; i
< ARRAY_SIZE(arc_pmu_ev_hw_map
); i
++) {
553 if (event_in_hw_event_map(i
, str
)) {
554 pr_debug("mapping perf event %2d to h/w event \'%8s\' (idx %d)\n",
556 arc_pmu
->ev_hw_idx
[i
] = j
;
561 static int arc_pmu_device_probe(struct platform_device
*pdev
)
563 struct arc_reg_pct_build pct_bcr
;
564 struct arc_reg_cc_build cc_bcr
;
565 int i
, has_interrupts
;
566 int counter_size
; /* in bits */
573 char str
[ARCPMU_EVENT_NAME_LEN
];
577 READ_BCR(ARC_REG_PCT_BUILD
, pct_bcr
);
579 pr_err("This core does not have performance counters!\n");
582 BUILD_BUG_ON(ARC_PERF_MAX_COUNTERS
> 32);
583 if (WARN_ON(pct_bcr
.c
> ARC_PERF_MAX_COUNTERS
))
586 READ_BCR(ARC_REG_CC_BUILD
, cc_bcr
);
587 if (WARN(!cc_bcr
.v
, "Counters exist but No countable conditions?"))
590 arc_pmu
= devm_kzalloc(&pdev
->dev
, sizeof(struct arc_pmu
), GFP_KERNEL
);
594 arc_pmu
->n_events
= cc_bcr
.c
;
596 if (arc_pmu_raw_alloc(&pdev
->dev
))
599 has_interrupts
= is_isa_arcv2() ? pct_bcr
.i
: 0;
601 arc_pmu
->n_counters
= pct_bcr
.c
;
602 counter_size
= 32 + (pct_bcr
.s
<< 4);
604 arc_pmu
->max_period
= (1ULL << counter_size
) / 2 - 1ULL;
606 pr_info("ARC perf\t: %d counters (%d bits), %d conditions%s\n",
607 arc_pmu
->n_counters
, counter_size
, cc_bcr
.c
,
608 has_interrupts
? ", [overflow IRQ support]" : "");
610 cc_name
.str
[ARCPMU_EVENT_NAME_LEN
- 1] = 0;
611 for (i
= 0; i
< PERF_COUNT_ARC_HW_MAX
; i
++)
612 arc_pmu
->ev_hw_idx
[i
] = -1;
614 /* loop thru all available h/w condition indexes */
615 for (i
= 0; i
< cc_bcr
.c
; i
++) {
616 write_aux_reg(ARC_REG_CC_INDEX
, i
);
617 cc_name
.indiv
.word0
= le32_to_cpu(read_aux_reg(ARC_REG_CC_NAME0
));
618 cc_name
.indiv
.word1
= le32_to_cpu(read_aux_reg(ARC_REG_CC_NAME1
));
620 arc_pmu_map_hw_event(i
, cc_name
.str
);
621 arc_pmu_add_raw_event_attr(i
, cc_name
.str
);
624 arc_pmu_events_attr_gr
.attrs
= arc_pmu
->attrs
;
625 arc_pmu
->attr_groups
[ARCPMU_ATTR_GR_EVENTS
] = &arc_pmu_events_attr_gr
;
626 arc_pmu
->attr_groups
[ARCPMU_ATTR_GR_FORMATS
] = &arc_pmu_format_attr_gr
;
628 arc_pmu
->pmu
= (struct pmu
) {
629 .pmu_enable
= arc_pmu_enable
,
630 .pmu_disable
= arc_pmu_disable
,
631 .event_init
= arc_pmu_event_init
,
634 .start
= arc_pmu_start
,
635 .stop
= arc_pmu_stop
,
636 .read
= arc_pmu_read
,
637 .attr_groups
= arc_pmu
->attr_groups
,
640 if (has_interrupts
) {
641 int irq
= platform_get_irq(pdev
, 0);
644 pr_err("Cannot get IRQ number for the platform\n");
650 /* intc map function ensures irq_set_percpu_devid() called */
651 request_percpu_irq(irq
, arc_pmu_intr
, "ARC perf counters",
652 this_cpu_ptr(&arc_pmu_cpu
));
654 on_each_cpu(arc_cpu_pmu_irq_init
, &irq
, 1);
657 arc_pmu
->pmu
.capabilities
|= PERF_PMU_CAP_NO_INTERRUPT
;
660 * perf parser doesn't really like '-' symbol in events name, so let's
661 * use '_' in arc pct name as it goes to kernel PMU event prefix.
663 return perf_pmu_register(&arc_pmu
->pmu
, "arc_pct", PERF_TYPE_RAW
);
666 static const struct of_device_id arc_pmu_match
[] = {
667 { .compatible
= "snps,arc700-pct" },
668 { .compatible
= "snps,archs-pct" },
671 MODULE_DEVICE_TABLE(of
, arc_pmu_match
);
673 static struct platform_driver arc_pmu_driver
= {
676 .of_match_table
= of_match_ptr(arc_pmu_match
),
678 .probe
= arc_pmu_device_probe
,
681 module_platform_driver(arc_pmu_driver
);
683 MODULE_LICENSE("GPL");
684 MODULE_AUTHOR("Mischa Jonker <mjonker@synopsys.com>");
685 MODULE_DESCRIPTION("ARC PMU driver");