xref: /openbmc/linux/arch/powerpc/perf/8xx-pmu.c (revision 3dc4b6fb)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Performance event support - PPC 8xx
4  *
5  * Copyright 2016 Christophe Leroy, CS Systemes d'Information
6  */
7 
8 #include <linux/kernel.h>
9 #include <linux/sched.h>
10 #include <linux/perf_event.h>
11 #include <linux/percpu.h>
12 #include <linux/hardirq.h>
13 #include <asm/pmc.h>
14 #include <asm/machdep.h>
15 #include <asm/firmware.h>
16 #include <asm/ptrace.h>
17 #include <asm/code-patching.h>
18 
19 #define PERF_8xx_ID_CPU_CYCLES		1
20 #define PERF_8xx_ID_HW_INSTRUCTIONS	2
21 #define PERF_8xx_ID_ITLB_LOAD_MISS	3
22 #define PERF_8xx_ID_DTLB_LOAD_MISS	4
23 
24 #define C(x)	PERF_COUNT_HW_CACHE_##x
25 #define DTLB_LOAD_MISS	(C(DTLB) | (C(OP_READ) << 8) | (C(RESULT_MISS) << 16))
26 #define ITLB_LOAD_MISS	(C(ITLB) | (C(OP_READ) << 8) | (C(RESULT_MISS) << 16))
27 
28 extern unsigned long itlb_miss_counter, dtlb_miss_counter;
29 extern atomic_t instruction_counter;
30 
31 static atomic_t insn_ctr_ref;
32 static atomic_t itlb_miss_ref;
33 static atomic_t dtlb_miss_ref;
34 
35 static s64 get_insn_ctr(void)
36 {
37 	int ctr;
38 	unsigned long counta;
39 
40 	do {
41 		ctr = atomic_read(&instruction_counter);
42 		counta = mfspr(SPRN_COUNTA);
43 	} while (ctr != atomic_read(&instruction_counter));
44 
45 	return ((s64)ctr << 16) | (counta >> 16);
46 }
47 
48 static int event_type(struct perf_event *event)
49 {
50 	switch (event->attr.type) {
51 	case PERF_TYPE_HARDWARE:
52 		if (event->attr.config == PERF_COUNT_HW_CPU_CYCLES)
53 			return PERF_8xx_ID_CPU_CYCLES;
54 		if (event->attr.config == PERF_COUNT_HW_INSTRUCTIONS)
55 			return PERF_8xx_ID_HW_INSTRUCTIONS;
56 		break;
57 	case PERF_TYPE_HW_CACHE:
58 		if (event->attr.config == ITLB_LOAD_MISS)
59 			return PERF_8xx_ID_ITLB_LOAD_MISS;
60 		if (event->attr.config == DTLB_LOAD_MISS)
61 			return PERF_8xx_ID_DTLB_LOAD_MISS;
62 		break;
63 	case PERF_TYPE_RAW:
64 		break;
65 	default:
66 		return -ENOENT;
67 	}
68 	return -EOPNOTSUPP;
69 }
70 
71 static int mpc8xx_pmu_event_init(struct perf_event *event)
72 {
73 	int type = event_type(event);
74 
75 	if (type < 0)
76 		return type;
77 	return 0;
78 }
79 
80 static int mpc8xx_pmu_add(struct perf_event *event, int flags)
81 {
82 	int type = event_type(event);
83 	s64 val = 0;
84 
85 	if (type < 0)
86 		return type;
87 
88 	switch (type) {
89 	case PERF_8xx_ID_CPU_CYCLES:
90 		val = get_tb();
91 		break;
92 	case PERF_8xx_ID_HW_INSTRUCTIONS:
93 		if (atomic_inc_return(&insn_ctr_ref) == 1)
94 			mtspr(SPRN_ICTRL, 0xc0080007);
95 		val = get_insn_ctr();
96 		break;
97 	case PERF_8xx_ID_ITLB_LOAD_MISS:
98 		if (atomic_inc_return(&itlb_miss_ref) == 1) {
99 			unsigned long target = patch_site_addr(&patch__itlbmiss_perf);
100 
101 			patch_branch_site(&patch__itlbmiss_exit_1, target, 0);
102 #ifndef CONFIG_PIN_TLB_TEXT
103 			patch_branch_site(&patch__itlbmiss_exit_2, target, 0);
104 #endif
105 		}
106 		val = itlb_miss_counter;
107 		break;
108 	case PERF_8xx_ID_DTLB_LOAD_MISS:
109 		if (atomic_inc_return(&dtlb_miss_ref) == 1) {
110 			unsigned long target = patch_site_addr(&patch__dtlbmiss_perf);
111 
112 			patch_branch_site(&patch__dtlbmiss_exit_1, target, 0);
113 			patch_branch_site(&patch__dtlbmiss_exit_2, target, 0);
114 			patch_branch_site(&patch__dtlbmiss_exit_3, target, 0);
115 		}
116 		val = dtlb_miss_counter;
117 		break;
118 	}
119 	local64_set(&event->hw.prev_count, val);
120 	return 0;
121 }
122 
123 static void mpc8xx_pmu_read(struct perf_event *event)
124 {
125 	int type = event_type(event);
126 	s64 prev, val = 0, delta = 0;
127 
128 	if (type < 0)
129 		return;
130 
131 	do {
132 		prev = local64_read(&event->hw.prev_count);
133 		switch (type) {
134 		case PERF_8xx_ID_CPU_CYCLES:
135 			val = get_tb();
136 			delta = 16 * (val - prev);
137 			break;
138 		case PERF_8xx_ID_HW_INSTRUCTIONS:
139 			val = get_insn_ctr();
140 			delta = prev - val;
141 			if (delta < 0)
142 				delta += 0x1000000000000LL;
143 			break;
144 		case PERF_8xx_ID_ITLB_LOAD_MISS:
145 			val = itlb_miss_counter;
146 			delta = (s64)((s32)val - (s32)prev);
147 			break;
148 		case PERF_8xx_ID_DTLB_LOAD_MISS:
149 			val = dtlb_miss_counter;
150 			delta = (s64)((s32)val - (s32)prev);
151 			break;
152 		}
153 	} while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
154 
155 	local64_add(delta, &event->count);
156 }
157 
158 static void mpc8xx_pmu_del(struct perf_event *event, int flags)
159 {
160 	/* mfspr r10, SPRN_SPRG_SCRATCH0 */
161 	unsigned int insn = PPC_INST_MFSPR | __PPC_RS(R10) |
162 			    __PPC_SPR(SPRN_SPRG_SCRATCH0);
163 
164 	mpc8xx_pmu_read(event);
165 
166 	/* If it was the last user, stop counting to avoid useles overhead */
167 	switch (event_type(event)) {
168 	case PERF_8xx_ID_CPU_CYCLES:
169 		break;
170 	case PERF_8xx_ID_HW_INSTRUCTIONS:
171 		if (atomic_dec_return(&insn_ctr_ref) == 0)
172 			mtspr(SPRN_ICTRL, 7);
173 		break;
174 	case PERF_8xx_ID_ITLB_LOAD_MISS:
175 		if (atomic_dec_return(&itlb_miss_ref) == 0) {
176 			patch_instruction_site(&patch__itlbmiss_exit_1, insn);
177 #ifndef CONFIG_PIN_TLB_TEXT
178 			patch_instruction_site(&patch__itlbmiss_exit_2, insn);
179 #endif
180 		}
181 		break;
182 	case PERF_8xx_ID_DTLB_LOAD_MISS:
183 		if (atomic_dec_return(&dtlb_miss_ref) == 0) {
184 			patch_instruction_site(&patch__dtlbmiss_exit_1, insn);
185 			patch_instruction_site(&patch__dtlbmiss_exit_2, insn);
186 			patch_instruction_site(&patch__dtlbmiss_exit_3, insn);
187 		}
188 		break;
189 	}
190 }
191 
192 static struct pmu mpc8xx_pmu = {
193 	.event_init	= mpc8xx_pmu_event_init,
194 	.add		= mpc8xx_pmu_add,
195 	.del		= mpc8xx_pmu_del,
196 	.read		= mpc8xx_pmu_read,
197 	.capabilities	= PERF_PMU_CAP_NO_INTERRUPT |
198 			  PERF_PMU_CAP_NO_NMI,
199 };
200 
201 static int init_mpc8xx_pmu(void)
202 {
203 	mtspr(SPRN_ICTRL, 7);
204 	mtspr(SPRN_CMPA, 0);
205 	mtspr(SPRN_COUNTA, 0xffff);
206 
207 	return perf_pmu_register(&mpc8xx_pmu, "cpu", PERF_TYPE_RAW);
208 }
209 
210 early_initcall(init_mpc8xx_pmu);
211