Merge tag 'sound-5.7-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/tiwai...
[linux-2.6-microblaze.git] / arch / x86 / events / amd / uncore.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2013 Advanced Micro Devices, Inc.
4  *
5  * Author: Jacob Shin <jacob.shin@amd.com>
6  */
7
8 #include <linux/perf_event.h>
9 #include <linux/percpu.h>
10 #include <linux/types.h>
11 #include <linux/slab.h>
12 #include <linux/init.h>
13 #include <linux/cpu.h>
14 #include <linux/cpumask.h>
15
16 #include <asm/cpufeature.h>
17 #include <asm/perf_event.h>
18 #include <asm/msr.h>
19 #include <asm/smp.h>
20
21 #define NUM_COUNTERS_NB         4
22 #define NUM_COUNTERS_L2         4
23 #define NUM_COUNTERS_L3         6
24 #define MAX_COUNTERS            6
25
26 #define RDPMC_BASE_NB           6
27 #define RDPMC_BASE_LLC          10
28
29 #define COUNTER_SHIFT           16
30
31 #undef pr_fmt
32 #define pr_fmt(fmt)     "amd_uncore: " fmt
33
34 static int num_counters_llc;
35 static int num_counters_nb;
36 static bool l3_mask;
37
38 static HLIST_HEAD(uncore_unused_list);
39
40 struct amd_uncore {
41         int id;
42         int refcnt;
43         int cpu;
44         int num_counters;
45         int rdpmc_base;
46         u32 msr_base;
47         cpumask_t *active_mask;
48         struct pmu *pmu;
49         struct perf_event *events[MAX_COUNTERS];
50         struct hlist_node node;
51 };
52
53 static struct amd_uncore * __percpu *amd_uncore_nb;
54 static struct amd_uncore * __percpu *amd_uncore_llc;
55
56 static struct pmu amd_nb_pmu;
57 static struct pmu amd_llc_pmu;
58
59 static cpumask_t amd_nb_active_mask;
60 static cpumask_t amd_llc_active_mask;
61
62 static bool is_nb_event(struct perf_event *event)
63 {
64         return event->pmu->type == amd_nb_pmu.type;
65 }
66
67 static bool is_llc_event(struct perf_event *event)
68 {
69         return event->pmu->type == amd_llc_pmu.type;
70 }
71
72 static struct amd_uncore *event_to_amd_uncore(struct perf_event *event)
73 {
74         if (is_nb_event(event) && amd_uncore_nb)
75                 return *per_cpu_ptr(amd_uncore_nb, event->cpu);
76         else if (is_llc_event(event) && amd_uncore_llc)
77                 return *per_cpu_ptr(amd_uncore_llc, event->cpu);
78
79         return NULL;
80 }
81
82 static void amd_uncore_read(struct perf_event *event)
83 {
84         struct hw_perf_event *hwc = &event->hw;
85         u64 prev, new;
86         s64 delta;
87
88         /*
89          * since we do not enable counter overflow interrupts,
90          * we do not have to worry about prev_count changing on us
91          */
92
93         prev = local64_read(&hwc->prev_count);
94         rdpmcl(hwc->event_base_rdpmc, new);
95         local64_set(&hwc->prev_count, new);
96         delta = (new << COUNTER_SHIFT) - (prev << COUNTER_SHIFT);
97         delta >>= COUNTER_SHIFT;
98         local64_add(delta, &event->count);
99 }
100
101 static void amd_uncore_start(struct perf_event *event, int flags)
102 {
103         struct hw_perf_event *hwc = &event->hw;
104
105         if (flags & PERF_EF_RELOAD)
106                 wrmsrl(hwc->event_base, (u64)local64_read(&hwc->prev_count));
107
108         hwc->state = 0;
109         wrmsrl(hwc->config_base, (hwc->config | ARCH_PERFMON_EVENTSEL_ENABLE));
110         perf_event_update_userpage(event);
111 }
112
113 static void amd_uncore_stop(struct perf_event *event, int flags)
114 {
115         struct hw_perf_event *hwc = &event->hw;
116
117         wrmsrl(hwc->config_base, hwc->config);
118         hwc->state |= PERF_HES_STOPPED;
119
120         if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
121                 amd_uncore_read(event);
122                 hwc->state |= PERF_HES_UPTODATE;
123         }
124 }
125
126 static int amd_uncore_add(struct perf_event *event, int flags)
127 {
128         int i;
129         struct amd_uncore *uncore = event_to_amd_uncore(event);
130         struct hw_perf_event *hwc = &event->hw;
131
132         /* are we already assigned? */
133         if (hwc->idx != -1 && uncore->events[hwc->idx] == event)
134                 goto out;
135
136         for (i = 0; i < uncore->num_counters; i++) {
137                 if (uncore->events[i] == event) {
138                         hwc->idx = i;
139                         goto out;
140                 }
141         }
142
143         /* if not, take the first available counter */
144         hwc->idx = -1;
145         for (i = 0; i < uncore->num_counters; i++) {
146                 if (cmpxchg(&uncore->events[i], NULL, event) == NULL) {
147                         hwc->idx = i;
148                         break;
149                 }
150         }
151
152 out:
153         if (hwc->idx == -1)
154                 return -EBUSY;
155
156         hwc->config_base = uncore->msr_base + (2 * hwc->idx);
157         hwc->event_base = uncore->msr_base + 1 + (2 * hwc->idx);
158         hwc->event_base_rdpmc = uncore->rdpmc_base + hwc->idx;
159         hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
160
161         if (flags & PERF_EF_START)
162                 amd_uncore_start(event, PERF_EF_RELOAD);
163
164         return 0;
165 }
166
167 static void amd_uncore_del(struct perf_event *event, int flags)
168 {
169         int i;
170         struct amd_uncore *uncore = event_to_amd_uncore(event);
171         struct hw_perf_event *hwc = &event->hw;
172
173         amd_uncore_stop(event, PERF_EF_UPDATE);
174
175         for (i = 0; i < uncore->num_counters; i++) {
176                 if (cmpxchg(&uncore->events[i], event, NULL) == event)
177                         break;
178         }
179
180         hwc->idx = -1;
181 }
182
183 /*
184  * Convert logical CPU number to L3 PMC Config ThreadMask format
185  */
186 static u64 l3_thread_slice_mask(int cpu)
187 {
188         u64 thread_mask, core = topology_core_id(cpu);
189         unsigned int shift, thread = 0;
190
191         if (topology_smt_supported() && !topology_is_primary_thread(cpu))
192                 thread = 1;
193
194         if (boot_cpu_data.x86 <= 0x18) {
195                 shift = AMD64_L3_THREAD_SHIFT + 2 * (core % 4) + thread;
196                 thread_mask = BIT_ULL(shift);
197
198                 return AMD64_L3_SLICE_MASK | thread_mask;
199         }
200
201         core = (core << AMD64_L3_COREID_SHIFT) & AMD64_L3_COREID_MASK;
202         shift = AMD64_L3_THREAD_SHIFT + thread;
203         thread_mask = BIT_ULL(shift);
204
205         return AMD64_L3_EN_ALL_SLICES | core | thread_mask;
206 }
207
208 static int amd_uncore_event_init(struct perf_event *event)
209 {
210         struct amd_uncore *uncore;
211         struct hw_perf_event *hwc = &event->hw;
212
213         if (event->attr.type != event->pmu->type)
214                 return -ENOENT;
215
216         /*
217          * NB and Last level cache counters (MSRs) are shared across all cores
218          * that share the same NB / Last level cache.  On family 16h and below,
219          * Interrupts can be directed to a single target core, however, event
220          * counts generated by processes running on other cores cannot be masked
221          * out. So we do not support sampling and per-thread events via
222          * CAP_NO_INTERRUPT, and we do not enable counter overflow interrupts:
223          */
224         hwc->config = event->attr.config & AMD64_RAW_EVENT_MASK_NB;
225         hwc->idx = -1;
226
227         if (event->cpu < 0)
228                 return -EINVAL;
229
230         /*
231          * SliceMask and ThreadMask need to be set for certain L3 events.
232          * For other events, the two fields do not affect the count.
233          */
234         if (l3_mask && is_llc_event(event))
235                 hwc->config |= l3_thread_slice_mask(event->cpu);
236
237         uncore = event_to_amd_uncore(event);
238         if (!uncore)
239                 return -ENODEV;
240
241         /*
242          * since request can come in to any of the shared cores, we will remap
243          * to a single common cpu.
244          */
245         event->cpu = uncore->cpu;
246
247         return 0;
248 }
249
250 static ssize_t amd_uncore_attr_show_cpumask(struct device *dev,
251                                             struct device_attribute *attr,
252                                             char *buf)
253 {
254         cpumask_t *active_mask;
255         struct pmu *pmu = dev_get_drvdata(dev);
256
257         if (pmu->type == amd_nb_pmu.type)
258                 active_mask = &amd_nb_active_mask;
259         else if (pmu->type == amd_llc_pmu.type)
260                 active_mask = &amd_llc_active_mask;
261         else
262                 return 0;
263
264         return cpumap_print_to_pagebuf(true, buf, active_mask);
265 }
266 static DEVICE_ATTR(cpumask, S_IRUGO, amd_uncore_attr_show_cpumask, NULL);
267
268 static struct attribute *amd_uncore_attrs[] = {
269         &dev_attr_cpumask.attr,
270         NULL,
271 };
272
273 static struct attribute_group amd_uncore_attr_group = {
274         .attrs = amd_uncore_attrs,
275 };
276
277 /*
278  * Similar to PMU_FORMAT_ATTR but allowing for format_attr to be assigned based
279  * on family
280  */
281 #define AMD_FORMAT_ATTR(_dev, _name, _format)                                \
282 static ssize_t                                                               \
283 _dev##_show##_name(struct device *dev,                                       \
284                 struct device_attribute *attr,                               \
285                 char *page)                                                  \
286 {                                                                            \
287         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                          \
288         return sprintf(page, _format "\n");                                  \
289 }                                                                            \
290 static struct device_attribute format_attr_##_dev##_name = __ATTR_RO(_dev);
291
292 /* Used for each uncore counter type */
293 #define AMD_ATTRIBUTE(_name)                                                 \
294 static struct attribute *amd_uncore_format_attr_##_name[] = {                \
295         &format_attr_event_##_name.attr,                                     \
296         &format_attr_umask.attr,                                             \
297         NULL,                                                                \
298 };                                                                           \
299 static struct attribute_group amd_uncore_format_group_##_name = {            \
300         .name = "format",                                                    \
301         .attrs = amd_uncore_format_attr_##_name,                             \
302 };                                                                           \
303 static const struct attribute_group *amd_uncore_attr_groups_##_name[] = {    \
304         &amd_uncore_attr_group,                                              \
305         &amd_uncore_format_group_##_name,                                    \
306         NULL,                                                                \
307 };
308
309 AMD_FORMAT_ATTR(event, , "config:0-7,32-35");
310 AMD_FORMAT_ATTR(umask, , "config:8-15");
311 AMD_FORMAT_ATTR(event, _df, "config:0-7,32-35,59-60");
312 AMD_FORMAT_ATTR(event, _l3, "config:0-7");
313 AMD_ATTRIBUTE(df);
314 AMD_ATTRIBUTE(l3);
315
316 static struct pmu amd_nb_pmu = {
317         .task_ctx_nr    = perf_invalid_context,
318         .event_init     = amd_uncore_event_init,
319         .add            = amd_uncore_add,
320         .del            = amd_uncore_del,
321         .start          = amd_uncore_start,
322         .stop           = amd_uncore_stop,
323         .read           = amd_uncore_read,
324         .capabilities   = PERF_PMU_CAP_NO_EXCLUDE | PERF_PMU_CAP_NO_INTERRUPT,
325 };
326
327 static struct pmu amd_llc_pmu = {
328         .task_ctx_nr    = perf_invalid_context,
329         .event_init     = amd_uncore_event_init,
330         .add            = amd_uncore_add,
331         .del            = amd_uncore_del,
332         .start          = amd_uncore_start,
333         .stop           = amd_uncore_stop,
334         .read           = amd_uncore_read,
335         .capabilities   = PERF_PMU_CAP_NO_EXCLUDE | PERF_PMU_CAP_NO_INTERRUPT,
336 };
337
338 static struct amd_uncore *amd_uncore_alloc(unsigned int cpu)
339 {
340         return kzalloc_node(sizeof(struct amd_uncore), GFP_KERNEL,
341                         cpu_to_node(cpu));
342 }
343
344 static int amd_uncore_cpu_up_prepare(unsigned int cpu)
345 {
346         struct amd_uncore *uncore_nb = NULL, *uncore_llc;
347
348         if (amd_uncore_nb) {
349                 uncore_nb = amd_uncore_alloc(cpu);
350                 if (!uncore_nb)
351                         goto fail;
352                 uncore_nb->cpu = cpu;
353                 uncore_nb->num_counters = num_counters_nb;
354                 uncore_nb->rdpmc_base = RDPMC_BASE_NB;
355                 uncore_nb->msr_base = MSR_F15H_NB_PERF_CTL;
356                 uncore_nb->active_mask = &amd_nb_active_mask;
357                 uncore_nb->pmu = &amd_nb_pmu;
358                 uncore_nb->id = -1;
359                 *per_cpu_ptr(amd_uncore_nb, cpu) = uncore_nb;
360         }
361
362         if (amd_uncore_llc) {
363                 uncore_llc = amd_uncore_alloc(cpu);
364                 if (!uncore_llc)
365                         goto fail;
366                 uncore_llc->cpu = cpu;
367                 uncore_llc->num_counters = num_counters_llc;
368                 uncore_llc->rdpmc_base = RDPMC_BASE_LLC;
369                 uncore_llc->msr_base = MSR_F16H_L2I_PERF_CTL;
370                 uncore_llc->active_mask = &amd_llc_active_mask;
371                 uncore_llc->pmu = &amd_llc_pmu;
372                 uncore_llc->id = -1;
373                 *per_cpu_ptr(amd_uncore_llc, cpu) = uncore_llc;
374         }
375
376         return 0;
377
378 fail:
379         if (amd_uncore_nb)
380                 *per_cpu_ptr(amd_uncore_nb, cpu) = NULL;
381         kfree(uncore_nb);
382         return -ENOMEM;
383 }
384
385 static struct amd_uncore *
386 amd_uncore_find_online_sibling(struct amd_uncore *this,
387                                struct amd_uncore * __percpu *uncores)
388 {
389         unsigned int cpu;
390         struct amd_uncore *that;
391
392         for_each_online_cpu(cpu) {
393                 that = *per_cpu_ptr(uncores, cpu);
394
395                 if (!that)
396                         continue;
397
398                 if (this == that)
399                         continue;
400
401                 if (this->id == that->id) {
402                         hlist_add_head(&this->node, &uncore_unused_list);
403                         this = that;
404                         break;
405                 }
406         }
407
408         this->refcnt++;
409         return this;
410 }
411
412 static int amd_uncore_cpu_starting(unsigned int cpu)
413 {
414         unsigned int eax, ebx, ecx, edx;
415         struct amd_uncore *uncore;
416
417         if (amd_uncore_nb) {
418                 uncore = *per_cpu_ptr(amd_uncore_nb, cpu);
419                 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
420                 uncore->id = ecx & 0xff;
421
422                 uncore = amd_uncore_find_online_sibling(uncore, amd_uncore_nb);
423                 *per_cpu_ptr(amd_uncore_nb, cpu) = uncore;
424         }
425
426         if (amd_uncore_llc) {
427                 uncore = *per_cpu_ptr(amd_uncore_llc, cpu);
428                 uncore->id = per_cpu(cpu_llc_id, cpu);
429
430                 uncore = amd_uncore_find_online_sibling(uncore, amd_uncore_llc);
431                 *per_cpu_ptr(amd_uncore_llc, cpu) = uncore;
432         }
433
434         return 0;
435 }
436
437 static void uncore_clean_online(void)
438 {
439         struct amd_uncore *uncore;
440         struct hlist_node *n;
441
442         hlist_for_each_entry_safe(uncore, n, &uncore_unused_list, node) {
443                 hlist_del(&uncore->node);
444                 kfree(uncore);
445         }
446 }
447
448 static void uncore_online(unsigned int cpu,
449                           struct amd_uncore * __percpu *uncores)
450 {
451         struct amd_uncore *uncore = *per_cpu_ptr(uncores, cpu);
452
453         uncore_clean_online();
454
455         if (cpu == uncore->cpu)
456                 cpumask_set_cpu(cpu, uncore->active_mask);
457 }
458
459 static int amd_uncore_cpu_online(unsigned int cpu)
460 {
461         if (amd_uncore_nb)
462                 uncore_online(cpu, amd_uncore_nb);
463
464         if (amd_uncore_llc)
465                 uncore_online(cpu, amd_uncore_llc);
466
467         return 0;
468 }
469
470 static void uncore_down_prepare(unsigned int cpu,
471                                 struct amd_uncore * __percpu *uncores)
472 {
473         unsigned int i;
474         struct amd_uncore *this = *per_cpu_ptr(uncores, cpu);
475
476         if (this->cpu != cpu)
477                 return;
478
479         /* this cpu is going down, migrate to a shared sibling if possible */
480         for_each_online_cpu(i) {
481                 struct amd_uncore *that = *per_cpu_ptr(uncores, i);
482
483                 if (cpu == i)
484                         continue;
485
486                 if (this == that) {
487                         perf_pmu_migrate_context(this->pmu, cpu, i);
488                         cpumask_clear_cpu(cpu, that->active_mask);
489                         cpumask_set_cpu(i, that->active_mask);
490                         that->cpu = i;
491                         break;
492                 }
493         }
494 }
495
496 static int amd_uncore_cpu_down_prepare(unsigned int cpu)
497 {
498         if (amd_uncore_nb)
499                 uncore_down_prepare(cpu, amd_uncore_nb);
500
501         if (amd_uncore_llc)
502                 uncore_down_prepare(cpu, amd_uncore_llc);
503
504         return 0;
505 }
506
507 static void uncore_dead(unsigned int cpu, struct amd_uncore * __percpu *uncores)
508 {
509         struct amd_uncore *uncore = *per_cpu_ptr(uncores, cpu);
510
511         if (cpu == uncore->cpu)
512                 cpumask_clear_cpu(cpu, uncore->active_mask);
513
514         if (!--uncore->refcnt)
515                 kfree(uncore);
516         *per_cpu_ptr(uncores, cpu) = NULL;
517 }
518
519 static int amd_uncore_cpu_dead(unsigned int cpu)
520 {
521         if (amd_uncore_nb)
522                 uncore_dead(cpu, amd_uncore_nb);
523
524         if (amd_uncore_llc)
525                 uncore_dead(cpu, amd_uncore_llc);
526
527         return 0;
528 }
529
530 static int __init amd_uncore_init(void)
531 {
532         int ret = -ENODEV;
533
534         if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD &&
535             boot_cpu_data.x86_vendor != X86_VENDOR_HYGON)
536                 return -ENODEV;
537
538         if (!boot_cpu_has(X86_FEATURE_TOPOEXT))
539                 return -ENODEV;
540
541         if (boot_cpu_data.x86 >= 0x17) {
542                 /*
543                  * For F17h and above, the Northbridge counters are
544                  * repurposed as Data Fabric counters. Also, L3
545                  * counters are supported too. The PMUs are exported
546                  * based on family as either L2 or L3 and NB or DF.
547                  */
548                 num_counters_nb           = NUM_COUNTERS_NB;
549                 num_counters_llc          = NUM_COUNTERS_L3;
550                 amd_nb_pmu.name           = "amd_df";
551                 amd_llc_pmu.name          = "amd_l3";
552                 format_attr_event_df.show = &event_show_df;
553                 format_attr_event_l3.show = &event_show_l3;
554                 l3_mask                   = true;
555         } else {
556                 num_counters_nb           = NUM_COUNTERS_NB;
557                 num_counters_llc          = NUM_COUNTERS_L2;
558                 amd_nb_pmu.name           = "amd_nb";
559                 amd_llc_pmu.name          = "amd_l2";
560                 format_attr_event_df      = format_attr_event;
561                 format_attr_event_l3      = format_attr_event;
562                 l3_mask                   = false;
563         }
564
565         amd_nb_pmu.attr_groups  = amd_uncore_attr_groups_df;
566         amd_llc_pmu.attr_groups = amd_uncore_attr_groups_l3;
567
568         if (boot_cpu_has(X86_FEATURE_PERFCTR_NB)) {
569                 amd_uncore_nb = alloc_percpu(struct amd_uncore *);
570                 if (!amd_uncore_nb) {
571                         ret = -ENOMEM;
572                         goto fail_nb;
573                 }
574                 ret = perf_pmu_register(&amd_nb_pmu, amd_nb_pmu.name, -1);
575                 if (ret)
576                         goto fail_nb;
577
578                 pr_info("%s NB counters detected\n",
579                         boot_cpu_data.x86_vendor == X86_VENDOR_HYGON ?
580                                 "HYGON" : "AMD");
581                 ret = 0;
582         }
583
584         if (boot_cpu_has(X86_FEATURE_PERFCTR_LLC)) {
585                 amd_uncore_llc = alloc_percpu(struct amd_uncore *);
586                 if (!amd_uncore_llc) {
587                         ret = -ENOMEM;
588                         goto fail_llc;
589                 }
590                 ret = perf_pmu_register(&amd_llc_pmu, amd_llc_pmu.name, -1);
591                 if (ret)
592                         goto fail_llc;
593
594                 pr_info("%s LLC counters detected\n",
595                         boot_cpu_data.x86_vendor == X86_VENDOR_HYGON ?
596                                 "HYGON" : "AMD");
597                 ret = 0;
598         }
599
600         /*
601          * Install callbacks. Core will call them for each online cpu.
602          */
603         if (cpuhp_setup_state(CPUHP_PERF_X86_AMD_UNCORE_PREP,
604                               "perf/x86/amd/uncore:prepare",
605                               amd_uncore_cpu_up_prepare, amd_uncore_cpu_dead))
606                 goto fail_llc;
607
608         if (cpuhp_setup_state(CPUHP_AP_PERF_X86_AMD_UNCORE_STARTING,
609                               "perf/x86/amd/uncore:starting",
610                               amd_uncore_cpu_starting, NULL))
611                 goto fail_prep;
612         if (cpuhp_setup_state(CPUHP_AP_PERF_X86_AMD_UNCORE_ONLINE,
613                               "perf/x86/amd/uncore:online",
614                               amd_uncore_cpu_online,
615                               amd_uncore_cpu_down_prepare))
616                 goto fail_start;
617         return 0;
618
619 fail_start:
620         cpuhp_remove_state(CPUHP_AP_PERF_X86_AMD_UNCORE_STARTING);
621 fail_prep:
622         cpuhp_remove_state(CPUHP_PERF_X86_AMD_UNCORE_PREP);
623 fail_llc:
624         if (boot_cpu_has(X86_FEATURE_PERFCTR_NB))
625                 perf_pmu_unregister(&amd_nb_pmu);
626         if (amd_uncore_llc)
627                 free_percpu(amd_uncore_llc);
628 fail_nb:
629         if (amd_uncore_nb)
630                 free_percpu(amd_uncore_nb);
631
632         return ret;
633 }
634 device_initcall(amd_uncore_init);