xref: /openbmc/linux/arch/x86/events/msr.c (revision cb1aaebe)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/perf_event.h>
3 #include <linux/nospec.h>
4 #include <asm/intel-family.h>
5 
6 enum perf_msr_id {
7 	PERF_MSR_TSC			= 0,
8 	PERF_MSR_APERF			= 1,
9 	PERF_MSR_MPERF			= 2,
10 	PERF_MSR_PPERF			= 3,
11 	PERF_MSR_SMI			= 4,
12 	PERF_MSR_PTSC			= 5,
13 	PERF_MSR_IRPERF			= 6,
14 	PERF_MSR_THERM			= 7,
15 	PERF_MSR_THERM_SNAP		= 8,
16 	PERF_MSR_THERM_UNIT		= 9,
17 	PERF_MSR_EVENT_MAX,
18 };
19 
20 static bool test_aperfmperf(int idx)
21 {
22 	return boot_cpu_has(X86_FEATURE_APERFMPERF);
23 }
24 
25 static bool test_ptsc(int idx)
26 {
27 	return boot_cpu_has(X86_FEATURE_PTSC);
28 }
29 
30 static bool test_irperf(int idx)
31 {
32 	return boot_cpu_has(X86_FEATURE_IRPERF);
33 }
34 
35 static bool test_therm_status(int idx)
36 {
37 	return boot_cpu_has(X86_FEATURE_DTHERM);
38 }
39 
40 static bool test_intel(int idx)
41 {
42 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL ||
43 	    boot_cpu_data.x86 != 6)
44 		return false;
45 
46 	switch (boot_cpu_data.x86_model) {
47 	case INTEL_FAM6_NEHALEM:
48 	case INTEL_FAM6_NEHALEM_G:
49 	case INTEL_FAM6_NEHALEM_EP:
50 	case INTEL_FAM6_NEHALEM_EX:
51 
52 	case INTEL_FAM6_WESTMERE:
53 	case INTEL_FAM6_WESTMERE_EP:
54 	case INTEL_FAM6_WESTMERE_EX:
55 
56 	case INTEL_FAM6_SANDYBRIDGE:
57 	case INTEL_FAM6_SANDYBRIDGE_X:
58 
59 	case INTEL_FAM6_IVYBRIDGE:
60 	case INTEL_FAM6_IVYBRIDGE_X:
61 
62 	case INTEL_FAM6_HASWELL_CORE:
63 	case INTEL_FAM6_HASWELL_X:
64 	case INTEL_FAM6_HASWELL_ULT:
65 	case INTEL_FAM6_HASWELL_GT3E:
66 
67 	case INTEL_FAM6_BROADWELL_CORE:
68 	case INTEL_FAM6_BROADWELL_XEON_D:
69 	case INTEL_FAM6_BROADWELL_GT3E:
70 	case INTEL_FAM6_BROADWELL_X:
71 
72 	case INTEL_FAM6_ATOM_SILVERMONT:
73 	case INTEL_FAM6_ATOM_SILVERMONT_X:
74 	case INTEL_FAM6_ATOM_AIRMONT:
75 
76 	case INTEL_FAM6_ATOM_GOLDMONT:
77 	case INTEL_FAM6_ATOM_GOLDMONT_X:
78 
79 	case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
80 
81 	case INTEL_FAM6_XEON_PHI_KNL:
82 	case INTEL_FAM6_XEON_PHI_KNM:
83 		if (idx == PERF_MSR_SMI)
84 			return true;
85 		break;
86 
87 	case INTEL_FAM6_SKYLAKE_MOBILE:
88 	case INTEL_FAM6_SKYLAKE_DESKTOP:
89 	case INTEL_FAM6_SKYLAKE_X:
90 	case INTEL_FAM6_KABYLAKE_MOBILE:
91 	case INTEL_FAM6_KABYLAKE_DESKTOP:
92 	case INTEL_FAM6_ICELAKE_MOBILE:
93 		if (idx == PERF_MSR_SMI || idx == PERF_MSR_PPERF)
94 			return true;
95 		break;
96 	}
97 
98 	return false;
99 }
100 
101 struct perf_msr {
102 	u64	msr;
103 	struct	perf_pmu_events_attr *attr;
104 	bool	(*test)(int idx);
105 };
106 
107 PMU_EVENT_ATTR_STRING(tsc,				evattr_tsc,		"event=0x00"	);
108 PMU_EVENT_ATTR_STRING(aperf,				evattr_aperf,		"event=0x01"	);
109 PMU_EVENT_ATTR_STRING(mperf,				evattr_mperf,		"event=0x02"	);
110 PMU_EVENT_ATTR_STRING(pperf,				evattr_pperf,		"event=0x03"	);
111 PMU_EVENT_ATTR_STRING(smi,				evattr_smi,		"event=0x04"	);
112 PMU_EVENT_ATTR_STRING(ptsc,				evattr_ptsc,		"event=0x05"	);
113 PMU_EVENT_ATTR_STRING(irperf,				evattr_irperf,		"event=0x06"	);
114 PMU_EVENT_ATTR_STRING(cpu_thermal_margin,		evattr_therm,		"event=0x07"	);
115 PMU_EVENT_ATTR_STRING(cpu_thermal_margin.snapshot,	evattr_therm_snap,	"1"		);
116 PMU_EVENT_ATTR_STRING(cpu_thermal_margin.unit,		evattr_therm_unit,	"C"		);
117 
118 static struct perf_msr msr[] = {
119 	[PERF_MSR_TSC]		= { 0,				&evattr_tsc,		NULL,			},
120 	[PERF_MSR_APERF]	= { MSR_IA32_APERF,		&evattr_aperf,		test_aperfmperf,	},
121 	[PERF_MSR_MPERF]	= { MSR_IA32_MPERF,		&evattr_mperf,		test_aperfmperf,	},
122 	[PERF_MSR_PPERF]	= { MSR_PPERF,			&evattr_pperf,		test_intel,		},
123 	[PERF_MSR_SMI]		= { MSR_SMI_COUNT,		&evattr_smi,		test_intel,		},
124 	[PERF_MSR_PTSC]		= { MSR_F15H_PTSC,		&evattr_ptsc,		test_ptsc,		},
125 	[PERF_MSR_IRPERF]	= { MSR_F17H_IRPERF,		&evattr_irperf,		test_irperf,		},
126 	[PERF_MSR_THERM]	= { MSR_IA32_THERM_STATUS,	&evattr_therm,		test_therm_status,	},
127 	[PERF_MSR_THERM_SNAP]	= { MSR_IA32_THERM_STATUS,	&evattr_therm_snap,	test_therm_status,	},
128 	[PERF_MSR_THERM_UNIT]	= { MSR_IA32_THERM_STATUS,	&evattr_therm_unit,	test_therm_status,	},
129 };
130 
131 static struct attribute *events_attrs[PERF_MSR_EVENT_MAX + 1] = {
132 	NULL,
133 };
134 
135 static struct attribute_group events_attr_group = {
136 	.name = "events",
137 	.attrs = events_attrs,
138 };
139 
140 PMU_FORMAT_ATTR(event, "config:0-63");
141 static struct attribute *format_attrs[] = {
142 	&format_attr_event.attr,
143 	NULL,
144 };
145 static struct attribute_group format_attr_group = {
146 	.name = "format",
147 	.attrs = format_attrs,
148 };
149 
150 static const struct attribute_group *attr_groups[] = {
151 	&events_attr_group,
152 	&format_attr_group,
153 	NULL,
154 };
155 
156 static int msr_event_init(struct perf_event *event)
157 {
158 	u64 cfg = event->attr.config;
159 
160 	if (event->attr.type != event->pmu->type)
161 		return -ENOENT;
162 
163 	/* unsupported modes and filters */
164 	if (event->attr.sample_period) /* no sampling */
165 		return -EINVAL;
166 
167 	if (cfg >= PERF_MSR_EVENT_MAX)
168 		return -EINVAL;
169 
170 	cfg = array_index_nospec((unsigned long)cfg, PERF_MSR_EVENT_MAX);
171 
172 	if (!msr[cfg].attr)
173 		return -EINVAL;
174 
175 	event->hw.idx		= -1;
176 	event->hw.event_base	= msr[cfg].msr;
177 	event->hw.config	= cfg;
178 
179 	return 0;
180 }
181 
182 static inline u64 msr_read_counter(struct perf_event *event)
183 {
184 	u64 now;
185 
186 	if (event->hw.event_base)
187 		rdmsrl(event->hw.event_base, now);
188 	else
189 		now = rdtsc_ordered();
190 
191 	return now;
192 }
193 
194 static void msr_event_update(struct perf_event *event)
195 {
196 	u64 prev, now;
197 	s64 delta;
198 
199 	/* Careful, an NMI might modify the previous event value: */
200 again:
201 	prev = local64_read(&event->hw.prev_count);
202 	now = msr_read_counter(event);
203 
204 	if (local64_cmpxchg(&event->hw.prev_count, prev, now) != prev)
205 		goto again;
206 
207 	delta = now - prev;
208 	if (unlikely(event->hw.event_base == MSR_SMI_COUNT)) {
209 		delta = sign_extend64(delta, 31);
210 		local64_add(delta, &event->count);
211 	} else if (unlikely(event->hw.event_base == MSR_IA32_THERM_STATUS)) {
212 		/* If valid, extract digital readout, otherwise set to -1: */
213 		now = now & (1ULL << 31) ? (now >> 16) & 0x3f :  -1;
214 		local64_set(&event->count, now);
215 	} else {
216 		local64_add(delta, &event->count);
217 	}
218 }
219 
220 static void msr_event_start(struct perf_event *event, int flags)
221 {
222 	u64 now = msr_read_counter(event);
223 
224 	local64_set(&event->hw.prev_count, now);
225 }
226 
227 static void msr_event_stop(struct perf_event *event, int flags)
228 {
229 	msr_event_update(event);
230 }
231 
232 static void msr_event_del(struct perf_event *event, int flags)
233 {
234 	msr_event_stop(event, PERF_EF_UPDATE);
235 }
236 
237 static int msr_event_add(struct perf_event *event, int flags)
238 {
239 	if (flags & PERF_EF_START)
240 		msr_event_start(event, flags);
241 
242 	return 0;
243 }
244 
245 static struct pmu pmu_msr = {
246 	.task_ctx_nr	= perf_sw_context,
247 	.attr_groups	= attr_groups,
248 	.event_init	= msr_event_init,
249 	.add		= msr_event_add,
250 	.del		= msr_event_del,
251 	.start		= msr_event_start,
252 	.stop		= msr_event_stop,
253 	.read		= msr_event_update,
254 	.capabilities	= PERF_PMU_CAP_NO_INTERRUPT | PERF_PMU_CAP_NO_EXCLUDE,
255 };
256 
257 static int __init msr_init(void)
258 {
259 	int i, j = 0;
260 
261 	if (!boot_cpu_has(X86_FEATURE_TSC)) {
262 		pr_cont("no MSR PMU driver.\n");
263 		return 0;
264 	}
265 
266 	/* Probe the MSRs. */
267 	for (i = PERF_MSR_TSC + 1; i < PERF_MSR_EVENT_MAX; i++) {
268 		u64 val;
269 
270 		/* Virt sucks; you cannot tell if a R/O MSR is present :/ */
271 		if (!msr[i].test(i) || rdmsrl_safe(msr[i].msr, &val))
272 			msr[i].attr = NULL;
273 	}
274 
275 	/* List remaining MSRs in the sysfs attrs. */
276 	for (i = 0; i < PERF_MSR_EVENT_MAX; i++) {
277 		if (msr[i].attr)
278 			events_attrs[j++] = &msr[i].attr->attr.attr;
279 	}
280 	events_attrs[j] = NULL;
281 
282 	perf_pmu_register(&pmu_msr, "msr", -1);
283 
284 	return 0;
285 }
286 device_initcall(msr_init);
287