• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * @file op_model_ppro.h
3  * Family 6 perfmon and architectural perfmon MSR operations
4  *
5  * @remark Copyright 2002 OProfile authors
6  * @remark Copyright 2008 Intel Corporation
7  * @remark Read the file COPYING
8  *
9  * @author John Levon
10  * @author Philippe Elie
11  * @author Graydon Hoare
12  * @author Andi Kleen
13  */
14 
15 #include <linux/oprofile.h>
16 #include <linux/slab.h>
17 #include <asm/ptrace.h>
18 #include <asm/msr.h>
19 #include <asm/apic.h>
20 #include <asm/nmi.h>
21 #include <asm/intel_arch_perfmon.h>
22 
23 #include "op_x86_model.h"
24 #include "op_counter.h"
25 
26 static int num_counters = 2;
27 static int counter_width = 32;
28 
29 #define CTR_IS_RESERVED(msrs, c) (msrs->counters[(c)].addr ? 1 : 0)
30 #define CTR_OVERFLOWED(n) (!((n) & (1ULL<<(counter_width-1))))
31 
32 #define CTRL_IS_RESERVED(msrs, c) (msrs->controls[(c)].addr ? 1 : 0)
33 #define CTRL_READ(l, h, msrs, c) do {rdmsr((msrs->controls[(c)].addr), (l), (h)); } while (0)
34 #define CTRL_WRITE(l, h, msrs, c) do {wrmsr((msrs->controls[(c)].addr), (l), (h)); } while (0)
35 #define CTRL_SET_ACTIVE(n) (n |= (1<<22))
36 #define CTRL_SET_INACTIVE(n) (n &= ~(1<<22))
37 #define CTRL_CLEAR(x) (x &= (1<<21))
38 #define CTRL_SET_ENABLE(val) (val |= 1<<20)
39 #define CTRL_SET_USR(val, u) (val |= ((u & 1) << 16))
40 #define CTRL_SET_KERN(val, k) (val |= ((k & 1) << 17))
41 #define CTRL_SET_UM(val, m) (val |= (m << 8))
42 #define CTRL_SET_EVENT(val, e) (val |= e)
43 
44 static u64 *reset_value;
45 
ppro_fill_in_addresses(struct op_msrs * const msrs)46 static void ppro_fill_in_addresses(struct op_msrs * const msrs)
47 {
48 	int i;
49 
50 	for (i = 0; i < num_counters; i++) {
51 		if (reserve_perfctr_nmi(MSR_P6_PERFCTR0 + i))
52 			msrs->counters[i].addr = MSR_P6_PERFCTR0 + i;
53 		else
54 			msrs->counters[i].addr = 0;
55 	}
56 
57 	for (i = 0; i < num_counters; i++) {
58 		if (reserve_evntsel_nmi(MSR_P6_EVNTSEL0 + i))
59 			msrs->controls[i].addr = MSR_P6_EVNTSEL0 + i;
60 		else
61 			msrs->controls[i].addr = 0;
62 	}
63 }
64 
65 
ppro_setup_ctrs(struct op_msrs const * const msrs)66 static void ppro_setup_ctrs(struct op_msrs const * const msrs)
67 {
68 	unsigned int low, high;
69 	int i;
70 
71 	if (!reset_value) {
72 		reset_value = kmalloc(sizeof(reset_value[0]) * num_counters,
73 					GFP_ATOMIC);
74 		if (!reset_value)
75 			return;
76 	}
77 
78 	if (cpu_has_arch_perfmon) {
79 		union cpuid10_eax eax;
80 		eax.full = cpuid_eax(0xa);
81 
82 		/*
83 		 * For Core2 (family 6, model 15), don't reset the
84 		 * counter width:
85 		 */
86 		if (!(eax.split.version_id == 0 &&
87 			current_cpu_data.x86 == 6 &&
88 				current_cpu_data.x86_model == 15)) {
89 
90 			if (counter_width < eax.split.bit_width)
91 				counter_width = eax.split.bit_width;
92 		}
93 	}
94 
95 	/* clear all counters */
96 	for (i = 0 ; i < num_counters; ++i) {
97 		if (unlikely(!CTRL_IS_RESERVED(msrs, i)))
98 			continue;
99 		CTRL_READ(low, high, msrs, i);
100 		CTRL_CLEAR(low);
101 		CTRL_WRITE(low, high, msrs, i);
102 	}
103 
104 	/* avoid a false detection of ctr overflows in NMI handler */
105 	for (i = 0; i < num_counters; ++i) {
106 		if (unlikely(!CTR_IS_RESERVED(msrs, i)))
107 			continue;
108 		wrmsrl(msrs->counters[i].addr, -1LL);
109 	}
110 
111 	/* enable active counters */
112 	for (i = 0; i < num_counters; ++i) {
113 		if ((counter_config[i].enabled) && (CTR_IS_RESERVED(msrs, i))) {
114 			reset_value[i] = counter_config[i].count;
115 
116 			wrmsrl(msrs->counters[i].addr, -reset_value[i]);
117 
118 			CTRL_READ(low, high, msrs, i);
119 			CTRL_CLEAR(low);
120 			CTRL_SET_ENABLE(low);
121 			CTRL_SET_USR(low, counter_config[i].user);
122 			CTRL_SET_KERN(low, counter_config[i].kernel);
123 			CTRL_SET_UM(low, counter_config[i].unit_mask);
124 			CTRL_SET_EVENT(low, counter_config[i].event);
125 			CTRL_WRITE(low, high, msrs, i);
126 		} else {
127 			reset_value[i] = 0;
128 		}
129 	}
130 }
131 
132 
ppro_check_ctrs(struct pt_regs * const regs,struct op_msrs const * const msrs)133 static int ppro_check_ctrs(struct pt_regs * const regs,
134 			   struct op_msrs const * const msrs)
135 {
136 	u64 val;
137 	int i;
138 
139 	for (i = 0 ; i < num_counters; ++i) {
140 		if (!reset_value[i])
141 			continue;
142 		rdmsrl(msrs->counters[i].addr, val);
143 		if (CTR_OVERFLOWED(val)) {
144 			oprofile_add_sample(regs, i);
145 			wrmsrl(msrs->counters[i].addr, -reset_value[i]);
146 		}
147 	}
148 
149 	/* Only P6 based Pentium M need to re-unmask the apic vector but it
150 	 * doesn't hurt other P6 variant */
151 	apic_write(APIC_LVTPC, apic_read(APIC_LVTPC) & ~APIC_LVT_MASKED);
152 
153 	/* We can't work out if we really handled an interrupt. We
154 	 * might have caught a *second* counter just after overflowing
155 	 * the interrupt for this counter then arrives
156 	 * and we don't find a counter that's overflowed, so we
157 	 * would return 0 and get dazed + confused. Instead we always
158 	 * assume we found an overflow. This sucks.
159 	 */
160 	return 1;
161 }
162 
163 
ppro_start(struct op_msrs const * const msrs)164 static void ppro_start(struct op_msrs const * const msrs)
165 {
166 	unsigned int low, high;
167 	int i;
168 
169 	if (!reset_value)
170 		return;
171 	for (i = 0; i < num_counters; ++i) {
172 		if (reset_value[i]) {
173 			CTRL_READ(low, high, msrs, i);
174 			CTRL_SET_ACTIVE(low);
175 			CTRL_WRITE(low, high, msrs, i);
176 		}
177 	}
178 }
179 
180 
ppro_stop(struct op_msrs const * const msrs)181 static void ppro_stop(struct op_msrs const * const msrs)
182 {
183 	unsigned int low, high;
184 	int i;
185 
186 	if (!reset_value)
187 		return;
188 	for (i = 0; i < num_counters; ++i) {
189 		if (!reset_value[i])
190 			continue;
191 		CTRL_READ(low, high, msrs, i);
192 		CTRL_SET_INACTIVE(low);
193 		CTRL_WRITE(low, high, msrs, i);
194 	}
195 }
196 
ppro_shutdown(struct op_msrs const * const msrs)197 static void ppro_shutdown(struct op_msrs const * const msrs)
198 {
199 	int i;
200 
201 	for (i = 0 ; i < num_counters ; ++i) {
202 		if (CTR_IS_RESERVED(msrs, i))
203 			release_perfctr_nmi(MSR_P6_PERFCTR0 + i);
204 	}
205 	for (i = 0 ; i < num_counters ; ++i) {
206 		if (CTRL_IS_RESERVED(msrs, i))
207 			release_evntsel_nmi(MSR_P6_EVNTSEL0 + i);
208 	}
209 	if (reset_value) {
210 		kfree(reset_value);
211 		reset_value = NULL;
212 	}
213 }
214 
215 
216 struct op_x86_model_spec op_ppro_spec = {
217 	.num_counters		= 2,	/* can be overriden */
218 	.num_controls		= 2,	/* dito */
219 	.fill_in_addresses	= &ppro_fill_in_addresses,
220 	.setup_ctrs		= &ppro_setup_ctrs,
221 	.check_ctrs		= &ppro_check_ctrs,
222 	.start			= &ppro_start,
223 	.stop			= &ppro_stop,
224 	.shutdown		= &ppro_shutdown
225 };
226 
227 /*
228  * Architectural performance monitoring.
229  *
230  * Newer Intel CPUs (Core1+) have support for architectural
231  * events described in CPUID 0xA. See the IA32 SDM Vol3b.18 for details.
232  * The advantage of this is that it can be done without knowing about
233  * the specific CPU.
234  */
235 
arch_perfmon_setup_counters(void)236 void arch_perfmon_setup_counters(void)
237 {
238 	union cpuid10_eax eax;
239 
240 	eax.full = cpuid_eax(0xa);
241 
242 	/* Workaround for BIOS bugs in 6/15. Taken from perfmon2 */
243 	if (eax.split.version_id == 0 && current_cpu_data.x86 == 6 &&
244 		current_cpu_data.x86_model == 15) {
245 		eax.split.version_id = 2;
246 		eax.split.num_counters = 2;
247 		eax.split.bit_width = 40;
248 	}
249 
250 	num_counters = eax.split.num_counters;
251 
252 	op_arch_perfmon_spec.num_counters = num_counters;
253 	op_arch_perfmon_spec.num_controls = num_counters;
254 	op_ppro_spec.num_counters = num_counters;
255 	op_ppro_spec.num_controls = num_counters;
256 }
257 
258 struct op_x86_model_spec op_arch_perfmon_spec = {
259 	/* num_counters/num_controls filled in at runtime */
260 	.fill_in_addresses	= &ppro_fill_in_addresses,
261 	/* user space does the cpuid check for available events */
262 	.setup_ctrs		= &ppro_setup_ctrs,
263 	.check_ctrs		= &ppro_check_ctrs,
264 	.start			= &ppro_start,
265 	.stop			= &ppro_stop,
266 	.shutdown		= &ppro_shutdown
267 };
268