2 * Copyright 2019 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
24 #include <linux/perf_event.h>
25 #include <linux/init.h>
27 #include "amdgpu_pmu.h"
29 #define PMU_NAME_SIZE 32
30 #define NUM_FORMATS_AMDGPU_PMU 4
31 #define NUM_FORMATS_DF_VEGA20 3
32 #define NUM_EVENTS_DF_VEGA20 8
33 #define NUM_EVENT_TYPES_VEGA20 1
34 #define NUM_EVENTS_VEGA20_XGMI 2
35 #define NUM_EVENTS_VEGA20_MAX NUM_EVENTS_VEGA20_XGMI
36 #define NUM_EVENT_TYPES_ARCTURUS 1
37 #define NUM_EVENTS_ARCTURUS_XGMI 6
38 #define NUM_EVENTS_ARCTURUS_MAX NUM_EVENTS_ARCTURUS_XGMI
40 struct amdgpu_pmu_event_attribute {
41 struct device_attribute attr;
42 const char *event_str;
46 /* record to keep track of pmu entry per pmu type per device */
47 struct amdgpu_pmu_entry {
48 struct list_head entry;
49 struct amdgpu_device *adev;
51 unsigned int pmu_perf_type;
53 char *pmu_file_prefix;
54 struct attribute_group fmt_attr_group;
55 struct amdgpu_pmu_event_attribute *fmt_attr;
56 struct attribute_group evt_attr_group;
57 struct amdgpu_pmu_event_attribute *evt_attr;
60 static ssize_t amdgpu_pmu_event_show(struct device *dev,
61 struct device_attribute *attr, char *buf)
63 struct amdgpu_pmu_event_attribute *amdgpu_pmu_attr;
65 amdgpu_pmu_attr = container_of(attr, struct amdgpu_pmu_event_attribute,
68 if (!amdgpu_pmu_attr->type)
69 return sprintf(buf, "%s\n", amdgpu_pmu_attr->event_str);
71 return sprintf(buf, "%s,type=0x%x\n",
72 amdgpu_pmu_attr->event_str, amdgpu_pmu_attr->type);
75 static LIST_HEAD(amdgpu_pmu_list);
78 struct amdgpu_pmu_attr {
83 struct amdgpu_pmu_type {
84 const unsigned int type;
85 const unsigned int num_of_type;
88 struct amdgpu_pmu_config {
89 struct amdgpu_pmu_attr *formats;
90 unsigned int num_formats;
91 struct amdgpu_pmu_attr *events;
92 unsigned int num_events;
93 struct amdgpu_pmu_type *types;
94 unsigned int num_types;
98 * Events fall under two categories:
100 * Events in /sys/bus/event_source/devices/amdgpu_<pmu_type>_<dev_num> have
101 * performance counter operations handled by one IP <pmu_type>. Formats and
102 * events should be defined by <pmu_type>_<asic_type>_formats and
103 * <pmu_type>_<asic_type>_events respectively.
105 * - Event config typed
106 * Events in /sys/bus/event_source/devices/amdgpu_<dev_num> have performance
107 * counter operations that can be handled by multiple IPs dictated by their
108 * "type" format field. Formats and events should be defined by
109 * amdgpu_pmu_formats and <asic_type>_events respectively. Format field
110 * "type" is generated in amdgpu_pmu_event_show and defined in
111 * <asic_type>_event_config_types.
114 static struct amdgpu_pmu_attr amdgpu_pmu_formats[NUM_FORMATS_AMDGPU_PMU] = {
115 { .name = "event", .config = "config:0-7" },
116 { .name = "instance", .config = "config:8-15" },
117 { .name = "umask", .config = "config:16-23"},
118 { .name = "type", .config = "config:56-63"}
122 static struct amdgpu_pmu_attr vega20_events[NUM_EVENTS_VEGA20_MAX] = {
123 { .name = "xgmi_link0_data_outbound",
124 .config = "event=0x7,instance=0x46,umask=0x2" },
125 { .name = "xgmi_link1_data_outbound",
126 .config = "event=0x7,instance=0x47,umask=0x2" }
129 static struct amdgpu_pmu_type vega20_types[NUM_EVENT_TYPES_VEGA20] = {
130 { .type = AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI,
131 .num_of_type = NUM_EVENTS_VEGA20_XGMI }
134 static struct amdgpu_pmu_config vega20_config = {
135 .formats = amdgpu_pmu_formats,
136 .num_formats = ARRAY_SIZE(amdgpu_pmu_formats),
137 .events = vega20_events,
138 .num_events = ARRAY_SIZE(vega20_events),
139 .types = vega20_types,
140 .num_types = ARRAY_SIZE(vega20_types)
143 /* Vega20 data fabric (DF) events */
144 static struct amdgpu_pmu_attr df_vega20_formats[NUM_FORMATS_DF_VEGA20] = {
145 { .name = "event", .config = "config:0-7" },
146 { .name = "instance", .config = "config:8-15" },
147 { .name = "umask", .config = "config:16-23"}
150 static struct amdgpu_pmu_attr df_vega20_events[NUM_EVENTS_DF_VEGA20] = {
151 { .name = "cake0_pcsout_txdata",
152 .config = "event=0x7,instance=0x46,umask=0x2" },
153 { .name = "cake1_pcsout_txdata",
154 .config = "event=0x7,instance=0x47,umask=0x2" },
155 { .name = "cake0_pcsout_txmeta",
156 .config = "event=0x7,instance=0x46,umask=0x4" },
157 { .name = "cake1_pcsout_txmeta",
158 .config = "event=0x7,instance=0x47,umask=0x4" },
159 { .name = "cake0_ftiinstat_reqalloc",
160 .config = "event=0xb,instance=0x46,umask=0x4" },
161 { .name = "cake1_ftiinstat_reqalloc",
162 .config = "event=0xb,instance=0x47,umask=0x4" },
163 { .name = "cake0_ftiinstat_rspalloc",
164 .config = "event=0xb,instance=0x46,umask=0x8" },
165 { .name = "cake1_ftiinstat_rspalloc",
166 .config = "event=0xb,instance=0x47,umask=0x8" }
169 static struct amdgpu_pmu_config df_vega20_config = {
170 .formats = df_vega20_formats,
171 .num_formats = ARRAY_SIZE(df_vega20_formats),
172 .events = df_vega20_events,
173 .num_events = ARRAY_SIZE(df_vega20_events),
178 /* Arcturus events */
179 static struct amdgpu_pmu_attr arcturus_events[NUM_EVENTS_ARCTURUS_MAX] = {
180 { .name = "xgmi_link0_data_outbound",
181 .config = "event=0x7,instance=0x4b,umask=0x2" },
182 { .name = "xgmi_link1_data_outbound",
183 .config = "event=0x7,instance=0x4c,umask=0x2" },
184 { .name = "xgmi_link2_data_outbound",
185 .config = "event=0x7,instance=0x4d,umask=0x2" },
186 { .name = "xgmi_link3_data_outbound",
187 .config = "event=0x7,instance=0x4e,umask=0x2" },
188 { .name = "xgmi_link4_data_outbound",
189 .config = "event=0x7,instance=0x4f,umask=0x2" },
190 { .name = "xgmi_link5_data_outbound",
191 .config = "event=0x7,instance=0x50,umask=0x2" }
194 static struct amdgpu_pmu_type arcturus_types[NUM_EVENT_TYPES_ARCTURUS] = {
195 { .type = AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI,
196 .num_of_type = NUM_EVENTS_ARCTURUS_XGMI }
199 static struct amdgpu_pmu_config arcturus_config = {
200 .formats = amdgpu_pmu_formats,
201 .num_formats = ARRAY_SIZE(amdgpu_pmu_formats),
202 .events = arcturus_events,
203 .num_events = ARRAY_SIZE(arcturus_events),
204 .types = arcturus_types,
205 .num_types = ARRAY_SIZE(arcturus_types)
208 /* initialize perf counter */
209 static int amdgpu_perf_event_init(struct perf_event *event)
211 struct hw_perf_event *hwc = &event->hw;
213 /* test the event attr type check for PMU enumeration */
214 if (event->attr.type != event->pmu->type)
217 /* update the hw_perf_event struct with config data */
218 hwc->config = event->attr.config;
219 hwc->config_base = AMDGPU_PMU_PERF_TYPE_NONE;
224 /* start perf counter */
225 static void amdgpu_perf_start(struct perf_event *event, int flags)
227 struct hw_perf_event *hwc = &event->hw;
228 struct amdgpu_pmu_entry *pe = container_of(event->pmu,
229 struct amdgpu_pmu_entry,
233 if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
236 if ((!pe->adev->df.funcs) ||
237 (!pe->adev->df.funcs->pmc_start))
240 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
243 switch (hwc->config_base) {
244 case AMDGPU_PMU_EVENT_CONFIG_TYPE_DF:
245 case AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI:
246 if (!(flags & PERF_EF_RELOAD)) {
247 target_cntr = pe->adev->df.funcs->pmc_start(pe->adev,
248 hwc->config, 0 /* unused */,
249 1 /* add counter */);
253 hwc->idx = target_cntr;
256 pe->adev->df.funcs->pmc_start(pe->adev, hwc->config,
263 perf_event_update_userpage(event);
266 /* read perf counter */
267 static void amdgpu_perf_read(struct perf_event *event)
269 struct hw_perf_event *hwc = &event->hw;
270 struct amdgpu_pmu_entry *pe = container_of(event->pmu,
271 struct amdgpu_pmu_entry,
275 if ((!pe->adev->df.funcs) ||
276 (!pe->adev->df.funcs->pmc_get_count))
280 prev = local64_read(&hwc->prev_count);
282 switch (hwc->config_base) {
283 case AMDGPU_PMU_EVENT_CONFIG_TYPE_DF:
284 case AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI:
285 pe->adev->df.funcs->pmc_get_count(pe->adev,
286 hwc->config, hwc->idx, &count);
292 } while (local64_cmpxchg(&hwc->prev_count, prev, count) != prev);
294 local64_add(count - prev, &event->count);
297 /* stop perf counter */
298 static void amdgpu_perf_stop(struct perf_event *event, int flags)
300 struct hw_perf_event *hwc = &event->hw;
301 struct amdgpu_pmu_entry *pe = container_of(event->pmu,
302 struct amdgpu_pmu_entry,
305 if (hwc->state & PERF_HES_UPTODATE)
308 if ((!pe->adev->df.funcs) ||
309 (!pe->adev->df.funcs->pmc_stop))
312 switch (hwc->config_base) {
313 case AMDGPU_PMU_EVENT_CONFIG_TYPE_DF:
314 case AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI:
315 pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, hwc->idx,
322 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
323 hwc->state |= PERF_HES_STOPPED;
325 if (hwc->state & PERF_HES_UPTODATE)
328 amdgpu_perf_read(event);
329 hwc->state |= PERF_HES_UPTODATE;
332 /* add perf counter */
333 static int amdgpu_perf_add(struct perf_event *event, int flags)
335 struct hw_perf_event *hwc = &event->hw;
336 int retval = 0, target_cntr;
337 struct amdgpu_pmu_entry *pe = container_of(event->pmu,
338 struct amdgpu_pmu_entry,
341 if ((!pe->adev->df.funcs) ||
342 (!pe->adev->df.funcs->pmc_start))
345 switch (pe->pmu_perf_type) {
346 case AMDGPU_PMU_PERF_TYPE_DF:
347 hwc->config_base = AMDGPU_PMU_EVENT_CONFIG_TYPE_DF;
349 case AMDGPU_PMU_PERF_TYPE_ALL:
350 hwc->config_base = (hwc->config >>
351 AMDGPU_PMU_EVENT_CONFIG_TYPE_SHIFT) &
352 AMDGPU_PMU_EVENT_CONFIG_TYPE_MASK;
356 event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
358 switch (hwc->config_base) {
359 case AMDGPU_PMU_EVENT_CONFIG_TYPE_DF:
360 case AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI:
361 target_cntr = pe->adev->df.funcs->pmc_start(pe->adev,
362 hwc->config, 0 /* unused */,
363 1 /* add counter */);
365 retval = target_cntr;
367 hwc->idx = target_cntr;
377 if (flags & PERF_EF_START)
378 amdgpu_perf_start(event, PERF_EF_RELOAD);
383 /* delete perf counter */
384 static void amdgpu_perf_del(struct perf_event *event, int flags)
386 struct hw_perf_event *hwc = &event->hw;
387 struct amdgpu_pmu_entry *pe = container_of(event->pmu,
388 struct amdgpu_pmu_entry,
390 if ((!pe->adev->df.funcs) ||
391 (!pe->adev->df.funcs->pmc_stop))
394 amdgpu_perf_stop(event, PERF_EF_UPDATE);
396 switch (hwc->config_base) {
397 case AMDGPU_PMU_EVENT_CONFIG_TYPE_DF:
398 case AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI:
399 pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, hwc->idx,
406 perf_event_update_userpage(event);
409 static void amdgpu_pmu_create_event_attrs_by_type(
410 struct attribute_group *attr_group,
411 struct amdgpu_pmu_event_attribute *pmu_attr,
412 struct amdgpu_pmu_attr events[],
419 pmu_attr += s_offset;
421 for (i = s_offset; i < e_offset; i++) {
422 attr_group->attrs[i] = &pmu_attr->attr.attr;
423 sysfs_attr_init(&pmu_attr->attr.attr);
424 pmu_attr->attr.attr.name = events[i].name;
425 pmu_attr->attr.attr.mode = 0444;
426 pmu_attr->attr.show = amdgpu_pmu_event_show;
427 pmu_attr->event_str = events[i].config;
428 pmu_attr->type = type;
433 static void amdgpu_pmu_create_attrs(struct attribute_group *attr_group,
434 struct amdgpu_pmu_event_attribute *pmu_attr,
435 struct amdgpu_pmu_attr events[],
438 amdgpu_pmu_create_event_attrs_by_type(attr_group, pmu_attr, events, 0,
439 num_events, AMDGPU_PMU_EVENT_CONFIG_TYPE_NONE);
443 static int amdgpu_pmu_alloc_pmu_attrs(
444 struct attribute_group *fmt_attr_group,
445 struct amdgpu_pmu_event_attribute **fmt_attr,
446 struct attribute_group *evt_attr_group,
447 struct amdgpu_pmu_event_attribute **evt_attr,
448 struct amdgpu_pmu_config *config)
450 *fmt_attr = kcalloc(config->num_formats, sizeof(**fmt_attr),
456 fmt_attr_group->attrs = kcalloc(config->num_formats + 1,
457 sizeof(*fmt_attr_group->attrs), GFP_KERNEL);
459 if (!fmt_attr_group->attrs)
460 goto err_fmt_attr_grp;
462 *evt_attr = kcalloc(config->num_events, sizeof(**evt_attr), GFP_KERNEL);
467 evt_attr_group->attrs = kcalloc(config->num_events + 1,
468 sizeof(*evt_attr_group->attrs), GFP_KERNEL);
470 if (!evt_attr_group->attrs)
471 goto err_evt_attr_grp;
477 kfree(fmt_attr_group->attrs);
483 /* init pmu tracking per pmu type */
484 static int init_pmu_entry_by_type_and_add(struct amdgpu_pmu_entry *pmu_entry,
485 struct amdgpu_pmu_config *config)
487 const struct attribute_group *attr_groups[] = {
488 &pmu_entry->fmt_attr_group,
489 &pmu_entry->evt_attr_group,
492 char pmu_name[PMU_NAME_SIZE];
493 int ret = 0, total_num_events = 0;
495 pmu_entry->pmu = (struct pmu){
496 .event_init = amdgpu_perf_event_init,
497 .add = amdgpu_perf_add,
498 .del = amdgpu_perf_del,
499 .start = amdgpu_perf_start,
500 .stop = amdgpu_perf_stop,
501 .read = amdgpu_perf_read,
502 .task_ctx_nr = perf_invalid_context,
505 ret = amdgpu_pmu_alloc_pmu_attrs(&pmu_entry->fmt_attr_group,
506 &pmu_entry->fmt_attr,
507 &pmu_entry->evt_attr_group,
508 &pmu_entry->evt_attr,
514 amdgpu_pmu_create_attrs(&pmu_entry->fmt_attr_group, pmu_entry->fmt_attr,
515 config->formats, config->num_formats);
517 if (pmu_entry->pmu_perf_type == AMDGPU_PMU_PERF_TYPE_ALL) {
520 for (i = 0; i < config->num_types; i++) {
521 amdgpu_pmu_create_event_attrs_by_type(
522 &pmu_entry->evt_attr_group,
527 config->types[i].num_of_type,
528 config->types[i].type);
529 total_num_events += config->types[i].num_of_type;
532 amdgpu_pmu_create_attrs(&pmu_entry->evt_attr_group,
534 config->events, config->num_events);
535 total_num_events = config->num_events;
538 pmu_entry->pmu.attr_groups = kmemdup(attr_groups, sizeof(attr_groups),
541 if (!pmu_entry->pmu.attr_groups) {
546 snprintf(pmu_name, PMU_NAME_SIZE, "%s_%d", pmu_entry->pmu_file_prefix,
547 adev_to_drm(pmu_entry->adev)->primary->index);
549 ret = perf_pmu_register(&pmu_entry->pmu, pmu_name, -1);
554 if (pmu_entry->pmu_perf_type != AMDGPU_PMU_PERF_TYPE_ALL)
555 pr_info("Detected AMDGPU %s Counters. # of Counters = %d.\n",
556 pmu_entry->pmu_type_name, total_num_events);
558 pr_info("Detected AMDGPU %d Perf Events.\n", total_num_events);
561 list_add_tail(&pmu_entry->entry, &amdgpu_pmu_list);
565 kfree(pmu_entry->pmu.attr_groups);
567 kfree(pmu_entry->fmt_attr_group.attrs);
568 kfree(pmu_entry->fmt_attr);
569 kfree(pmu_entry->evt_attr_group.attrs);
570 kfree(pmu_entry->evt_attr);
572 pr_warn("Error initializing AMDGPU %s PMUs.\n",
573 pmu_entry->pmu_type_name);
577 /* destroy all pmu data associated with target device */
578 void amdgpu_pmu_fini(struct amdgpu_device *adev)
580 struct amdgpu_pmu_entry *pe, *temp;
582 list_for_each_entry_safe(pe, temp, &amdgpu_pmu_list, entry) {
583 if (pe->adev != adev)
585 list_del(&pe->entry);
586 perf_pmu_unregister(&pe->pmu);
587 kfree(pe->pmu.attr_groups);
588 kfree(pe->fmt_attr_group.attrs);
590 kfree(pe->evt_attr_group.attrs);
596 static struct amdgpu_pmu_entry *create_pmu_entry(struct amdgpu_device *adev,
597 unsigned int pmu_type,
599 char *pmu_file_prefix)
601 struct amdgpu_pmu_entry *pmu_entry;
603 pmu_entry = kzalloc(sizeof(struct amdgpu_pmu_entry), GFP_KERNEL);
608 pmu_entry->adev = adev;
609 pmu_entry->fmt_attr_group.name = "format";
610 pmu_entry->fmt_attr_group.attrs = NULL;
611 pmu_entry->evt_attr_group.name = "events";
612 pmu_entry->evt_attr_group.attrs = NULL;
613 pmu_entry->pmu_perf_type = pmu_type;
614 pmu_entry->pmu_type_name = pmu_type_name;
615 pmu_entry->pmu_file_prefix = pmu_file_prefix;
620 /* init amdgpu_pmu */
621 int amdgpu_pmu_init(struct amdgpu_device *adev)
624 struct amdgpu_pmu_entry *pmu_entry, *pmu_entry_df;
626 switch (adev->asic_type) {
628 pmu_entry_df = create_pmu_entry(adev, AMDGPU_PMU_PERF_TYPE_DF,
634 ret = init_pmu_entry_by_type_and_add(pmu_entry_df,
642 pmu_entry = create_pmu_entry(adev, AMDGPU_PMU_PERF_TYPE_ALL,
646 amdgpu_pmu_fini(adev);
650 ret = init_pmu_entry_by_type_and_add(pmu_entry,
655 amdgpu_pmu_fini(adev);
661 pmu_entry = create_pmu_entry(adev, AMDGPU_PMU_PERF_TYPE_ALL,
666 ret = init_pmu_entry_by_type_and_add(pmu_entry,