2 * Copyright (C) 2013 Advanced Micro Devices, Inc.
4 * Author: Jacob Shin <jacob.shin@amd.com>
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 as
8 * published by the Free Software Foundation.
11 #include <linux/perf_event.h>
12 #include <linux/percpu.h>
13 #include <linux/types.h>
14 #include <linux/slab.h>
15 #include <linux/init.h>
16 #include <linux/cpu.h>
17 #include <linux/cpumask.h>
19 #include <asm/cpufeature.h>
20 #include <asm/perf_event.h>
23 #define NUM_COUNTERS_NB 4
24 #define NUM_COUNTERS_L2 4
25 #define MAX_COUNTERS NUM_COUNTERS_NB
27 #define RDPMC_BASE_NB 6
28 #define RDPMC_BASE_L2 10
30 #define COUNTER_SHIFT 16
39 cpumask_t *active_mask;
41 struct perf_event *events[MAX_COUNTERS];
42 struct amd_uncore *free_when_cpu_online;
45 static struct amd_uncore * __percpu *amd_uncore_nb;
46 static struct amd_uncore * __percpu *amd_uncore_l2;
48 static struct pmu amd_nb_pmu;
49 static struct pmu amd_l2_pmu;
51 static cpumask_t amd_nb_active_mask;
52 static cpumask_t amd_l2_active_mask;
54 static bool is_nb_event(struct perf_event *event)
56 return event->pmu->type == amd_nb_pmu.type;
59 static bool is_l2_event(struct perf_event *event)
61 return event->pmu->type == amd_l2_pmu.type;
64 static struct amd_uncore *event_to_amd_uncore(struct perf_event *event)
66 if (is_nb_event(event) && amd_uncore_nb)
67 return *per_cpu_ptr(amd_uncore_nb, event->cpu);
68 else if (is_l2_event(event) && amd_uncore_l2)
69 return *per_cpu_ptr(amd_uncore_l2, event->cpu);
74 static void amd_uncore_read(struct perf_event *event)
76 struct hw_perf_event *hwc = &event->hw;
81 * since we do not enable counter overflow interrupts,
82 * we do not have to worry about prev_count changing on us
85 prev = local64_read(&hwc->prev_count);
86 rdpmcl(hwc->event_base_rdpmc, new);
87 local64_set(&hwc->prev_count, new);
88 delta = (new << COUNTER_SHIFT) - (prev << COUNTER_SHIFT);
89 delta >>= COUNTER_SHIFT;
90 local64_add(delta, &event->count);
93 static void amd_uncore_start(struct perf_event *event, int flags)
95 struct hw_perf_event *hwc = &event->hw;
97 if (flags & PERF_EF_RELOAD)
98 wrmsrl(hwc->event_base, (u64)local64_read(&hwc->prev_count));
101 wrmsrl(hwc->config_base, (hwc->config | ARCH_PERFMON_EVENTSEL_ENABLE));
102 perf_event_update_userpage(event);
105 static void amd_uncore_stop(struct perf_event *event, int flags)
107 struct hw_perf_event *hwc = &event->hw;
109 wrmsrl(hwc->config_base, hwc->config);
110 hwc->state |= PERF_HES_STOPPED;
112 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
113 amd_uncore_read(event);
114 hwc->state |= PERF_HES_UPTODATE;
118 static int amd_uncore_add(struct perf_event *event, int flags)
121 struct amd_uncore *uncore = event_to_amd_uncore(event);
122 struct hw_perf_event *hwc = &event->hw;
124 /* are we already assigned? */
125 if (hwc->idx != -1 && uncore->events[hwc->idx] == event)
128 for (i = 0; i < uncore->num_counters; i++) {
129 if (uncore->events[i] == event) {
135 /* if not, take the first available counter */
137 for (i = 0; i < uncore->num_counters; i++) {
138 if (cmpxchg(&uncore->events[i], NULL, event) == NULL) {
148 hwc->config_base = uncore->msr_base + (2 * hwc->idx);
149 hwc->event_base = uncore->msr_base + 1 + (2 * hwc->idx);
150 hwc->event_base_rdpmc = uncore->rdpmc_base + hwc->idx;
151 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
153 if (flags & PERF_EF_START)
154 amd_uncore_start(event, PERF_EF_RELOAD);
159 static void amd_uncore_del(struct perf_event *event, int flags)
162 struct amd_uncore *uncore = event_to_amd_uncore(event);
163 struct hw_perf_event *hwc = &event->hw;
165 amd_uncore_stop(event, PERF_EF_UPDATE);
167 for (i = 0; i < uncore->num_counters; i++) {
168 if (cmpxchg(&uncore->events[i], event, NULL) == event)
175 static int amd_uncore_event_init(struct perf_event *event)
177 struct amd_uncore *uncore;
178 struct hw_perf_event *hwc = &event->hw;
180 if (event->attr.type != event->pmu->type)
184 * NB and L2 counters (MSRs) are shared across all cores that share the
185 * same NB / L2 cache. Interrupts can be directed to a single target
186 * core, however, event counts generated by processes running on other
187 * cores cannot be masked out. So we do not support sampling and
190 if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
193 /* NB and L2 counters do not have usr/os/guest/host bits */
194 if (event->attr.exclude_user || event->attr.exclude_kernel ||
195 event->attr.exclude_host || event->attr.exclude_guest)
198 /* and we do not enable counter overflow interrupts */
199 hwc->config = event->attr.config & AMD64_RAW_EVENT_MASK_NB;
205 uncore = event_to_amd_uncore(event);
210 * since request can come in to any of the shared cores, we will remap
211 * to a single common cpu.
213 event->cpu = uncore->cpu;
218 static ssize_t amd_uncore_attr_show_cpumask(struct device *dev,
219 struct device_attribute *attr,
223 cpumask_t *active_mask;
224 struct pmu *pmu = dev_get_drvdata(dev);
226 if (pmu->type == amd_nb_pmu.type)
227 active_mask = &amd_nb_active_mask;
228 else if (pmu->type == amd_l2_pmu.type)
229 active_mask = &amd_l2_active_mask;
233 n = cpulist_scnprintf(buf, PAGE_SIZE - 2, active_mask);
238 static DEVICE_ATTR(cpumask, S_IRUGO, amd_uncore_attr_show_cpumask, NULL);
240 static struct attribute *amd_uncore_attrs[] = {
241 &dev_attr_cpumask.attr,
245 static struct attribute_group amd_uncore_attr_group = {
246 .attrs = amd_uncore_attrs,
249 PMU_FORMAT_ATTR(event, "config:0-7,32-35");
250 PMU_FORMAT_ATTR(umask, "config:8-15");
252 static struct attribute *amd_uncore_format_attr[] = {
253 &format_attr_event.attr,
254 &format_attr_umask.attr,
258 static struct attribute_group amd_uncore_format_group = {
260 .attrs = amd_uncore_format_attr,
263 static const struct attribute_group *amd_uncore_attr_groups[] = {
264 &amd_uncore_attr_group,
265 &amd_uncore_format_group,
269 static struct pmu amd_nb_pmu = {
270 .attr_groups = amd_uncore_attr_groups,
272 .event_init = amd_uncore_event_init,
273 .add = amd_uncore_add,
274 .del = amd_uncore_del,
275 .start = amd_uncore_start,
276 .stop = amd_uncore_stop,
277 .read = amd_uncore_read,
280 static struct pmu amd_l2_pmu = {
281 .attr_groups = amd_uncore_attr_groups,
283 .event_init = amd_uncore_event_init,
284 .add = amd_uncore_add,
285 .del = amd_uncore_del,
286 .start = amd_uncore_start,
287 .stop = amd_uncore_stop,
288 .read = amd_uncore_read,
291 static struct amd_uncore *amd_uncore_alloc(unsigned int cpu)
293 return kzalloc_node(sizeof(struct amd_uncore), GFP_KERNEL,
297 static void amd_uncore_cpu_up_prepare(unsigned int cpu)
299 struct amd_uncore *uncore;
302 uncore = amd_uncore_alloc(cpu);
304 uncore->num_counters = NUM_COUNTERS_NB;
305 uncore->rdpmc_base = RDPMC_BASE_NB;
306 uncore->msr_base = MSR_F15H_NB_PERF_CTL;
307 uncore->active_mask = &amd_nb_active_mask;
308 uncore->pmu = &amd_nb_pmu;
309 *per_cpu_ptr(amd_uncore_nb, cpu) = uncore;
313 uncore = amd_uncore_alloc(cpu);
315 uncore->num_counters = NUM_COUNTERS_L2;
316 uncore->rdpmc_base = RDPMC_BASE_L2;
317 uncore->msr_base = MSR_F16H_L2I_PERF_CTL;
318 uncore->active_mask = &amd_l2_active_mask;
319 uncore->pmu = &amd_l2_pmu;
320 *per_cpu_ptr(amd_uncore_l2, cpu) = uncore;
324 static struct amd_uncore *
325 amd_uncore_find_online_sibling(struct amd_uncore *this,
326 struct amd_uncore * __percpu *uncores)
329 struct amd_uncore *that;
331 for_each_online_cpu(cpu) {
332 that = *per_cpu_ptr(uncores, cpu);
340 if (this->id == that->id) {
341 that->free_when_cpu_online = this;
351 static void amd_uncore_cpu_starting(unsigned int cpu)
353 unsigned int eax, ebx, ecx, edx;
354 struct amd_uncore *uncore;
357 uncore = *per_cpu_ptr(amd_uncore_nb, cpu);
358 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
359 uncore->id = ecx & 0xff;
361 uncore = amd_uncore_find_online_sibling(uncore, amd_uncore_nb);
362 *per_cpu_ptr(amd_uncore_nb, cpu) = uncore;
366 unsigned int apicid = cpu_data(cpu).apicid;
367 unsigned int nshared;
369 uncore = *per_cpu_ptr(amd_uncore_l2, cpu);
370 cpuid_count(0x8000001d, 2, &eax, &ebx, &ecx, &edx);
371 nshared = ((eax >> 14) & 0xfff) + 1;
372 uncore->id = apicid - (apicid % nshared);
374 uncore = amd_uncore_find_online_sibling(uncore, amd_uncore_l2);
375 *per_cpu_ptr(amd_uncore_l2, cpu) = uncore;
379 static void uncore_online(unsigned int cpu,
380 struct amd_uncore * __percpu *uncores)
382 struct amd_uncore *uncore = *per_cpu_ptr(uncores, cpu);
384 kfree(uncore->free_when_cpu_online);
385 uncore->free_when_cpu_online = NULL;
387 if (cpu == uncore->cpu)
388 cpumask_set_cpu(cpu, uncore->active_mask);
391 static void amd_uncore_cpu_online(unsigned int cpu)
394 uncore_online(cpu, amd_uncore_nb);
397 uncore_online(cpu, amd_uncore_l2);
400 static void uncore_down_prepare(unsigned int cpu,
401 struct amd_uncore * __percpu *uncores)
404 struct amd_uncore *this = *per_cpu_ptr(uncores, cpu);
406 if (this->cpu != cpu)
409 /* this cpu is going down, migrate to a shared sibling if possible */
410 for_each_online_cpu(i) {
411 struct amd_uncore *that = *per_cpu_ptr(uncores, i);
417 perf_pmu_migrate_context(this->pmu, cpu, i);
418 cpumask_clear_cpu(cpu, that->active_mask);
419 cpumask_set_cpu(i, that->active_mask);
426 static void amd_uncore_cpu_down_prepare(unsigned int cpu)
429 uncore_down_prepare(cpu, amd_uncore_nb);
432 uncore_down_prepare(cpu, amd_uncore_l2);
435 static void uncore_dead(unsigned int cpu, struct amd_uncore * __percpu *uncores)
437 struct amd_uncore *uncore = *per_cpu_ptr(uncores, cpu);
439 if (cpu == uncore->cpu)
440 cpumask_clear_cpu(cpu, uncore->active_mask);
442 if (!--uncore->refcnt)
444 *per_cpu_ptr(amd_uncore_nb, cpu) = NULL;
447 static void amd_uncore_cpu_dead(unsigned int cpu)
450 uncore_dead(cpu, amd_uncore_nb);
453 uncore_dead(cpu, amd_uncore_l2);
457 amd_uncore_cpu_notifier(struct notifier_block *self, unsigned long action,
460 unsigned int cpu = (long)hcpu;
462 switch (action & ~CPU_TASKS_FROZEN) {
464 amd_uncore_cpu_up_prepare(cpu);
468 amd_uncore_cpu_starting(cpu);
472 amd_uncore_cpu_online(cpu);
475 case CPU_DOWN_PREPARE:
476 amd_uncore_cpu_down_prepare(cpu);
479 case CPU_UP_CANCELED:
481 amd_uncore_cpu_dead(cpu);
491 static struct notifier_block amd_uncore_cpu_notifier_block = {
492 .notifier_call = amd_uncore_cpu_notifier,
493 .priority = CPU_PRI_PERF + 1,
496 static void __init init_cpu_already_online(void *dummy)
498 unsigned int cpu = smp_processor_id();
500 amd_uncore_cpu_starting(cpu);
501 amd_uncore_cpu_online(cpu);
504 static int __init amd_uncore_init(void)
509 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
512 if (!cpu_has_topoext)
515 if (cpu_has_perfctr_nb) {
516 amd_uncore_nb = alloc_percpu(struct amd_uncore *);
517 perf_pmu_register(&amd_nb_pmu, amd_nb_pmu.name, -1);
519 printk(KERN_INFO "perf: AMD NB counters detected\n");
523 if (cpu_has_perfctr_l2) {
524 amd_uncore_l2 = alloc_percpu(struct amd_uncore *);
525 perf_pmu_register(&amd_l2_pmu, amd_l2_pmu.name, -1);
527 printk(KERN_INFO "perf: AMD L2I counters detected\n");
535 /* init cpus already online before registering for hotplug notifier */
536 for_each_online_cpu(cpu) {
537 amd_uncore_cpu_up_prepare(cpu);
538 smp_call_function_single(cpu, init_cpu_already_online, NULL, 1);
541 register_cpu_notifier(&amd_uncore_cpu_notifier_block);
546 device_initcall(amd_uncore_init);