Kyle Swenson | 8d8f654 | 2021-03-15 11:02:55 -0600 | [diff] [blame] | 1 | /* |
| 2 | * Copyright 2010 ARM Ltd. |
| 3 | * Copyright 2012 Advanced Micro Devices, Inc., Robert Richter |
| 4 | * |
| 5 | * Perf-events backend for OProfile. |
| 6 | */ |
| 7 | #include <linux/perf_event.h> |
| 8 | #include <linux/platform_device.h> |
| 9 | #include <linux/oprofile.h> |
| 10 | #include <linux/slab.h> |
| 11 | |
| 12 | /* |
| 13 | * Per performance monitor configuration as set via oprofilefs. |
| 14 | */ |
| 15 | struct op_counter_config { |
| 16 | unsigned long count; |
| 17 | unsigned long enabled; |
| 18 | unsigned long event; |
| 19 | unsigned long unit_mask; |
| 20 | unsigned long kernel; |
| 21 | unsigned long user; |
| 22 | struct perf_event_attr attr; |
| 23 | }; |
| 24 | |
| 25 | static int oprofile_perf_enabled; |
| 26 | static DEFINE_MUTEX(oprofile_perf_mutex); |
| 27 | |
| 28 | static struct op_counter_config *counter_config; |
| 29 | static DEFINE_PER_CPU(struct perf_event **, perf_events); |
| 30 | static int num_counters; |
| 31 | |
| 32 | /* |
| 33 | * Overflow callback for oprofile. |
| 34 | */ |
| 35 | static void op_overflow_handler(struct perf_event *event, |
| 36 | struct perf_sample_data *data, struct pt_regs *regs) |
| 37 | { |
| 38 | int id; |
| 39 | u32 cpu = smp_processor_id(); |
| 40 | |
| 41 | for (id = 0; id < num_counters; ++id) |
| 42 | if (per_cpu(perf_events, cpu)[id] == event) |
| 43 | break; |
| 44 | |
| 45 | if (id != num_counters) |
| 46 | oprofile_add_sample(regs, id); |
| 47 | else |
| 48 | pr_warning("oprofile: ignoring spurious overflow " |
| 49 | "on cpu %u\n", cpu); |
| 50 | } |
| 51 | |
| 52 | /* |
| 53 | * Called by oprofile_perf_setup to create perf attributes to mirror the oprofile |
| 54 | * settings in counter_config. Attributes are created as `pinned' events and |
| 55 | * so are permanently scheduled on the PMU. |
| 56 | */ |
| 57 | static void op_perf_setup(void) |
| 58 | { |
| 59 | int i; |
| 60 | u32 size = sizeof(struct perf_event_attr); |
| 61 | struct perf_event_attr *attr; |
| 62 | |
| 63 | for (i = 0; i < num_counters; ++i) { |
| 64 | attr = &counter_config[i].attr; |
| 65 | memset(attr, 0, size); |
| 66 | attr->type = PERF_TYPE_RAW; |
| 67 | attr->size = size; |
| 68 | attr->config = counter_config[i].event; |
| 69 | attr->sample_period = counter_config[i].count; |
| 70 | attr->pinned = 1; |
| 71 | } |
| 72 | } |
| 73 | |
| 74 | static int op_create_counter(int cpu, int event) |
| 75 | { |
| 76 | struct perf_event *pevent; |
| 77 | |
| 78 | if (!counter_config[event].enabled || per_cpu(perf_events, cpu)[event]) |
| 79 | return 0; |
| 80 | |
| 81 | pevent = perf_event_create_kernel_counter(&counter_config[event].attr, |
| 82 | cpu, NULL, |
| 83 | op_overflow_handler, NULL); |
| 84 | |
| 85 | if (IS_ERR(pevent)) |
| 86 | return PTR_ERR(pevent); |
| 87 | |
| 88 | if (pevent->state != PERF_EVENT_STATE_ACTIVE) { |
| 89 | perf_event_release_kernel(pevent); |
| 90 | pr_warning("oprofile: failed to enable event %d " |
| 91 | "on CPU %d\n", event, cpu); |
| 92 | return -EBUSY; |
| 93 | } |
| 94 | |
| 95 | per_cpu(perf_events, cpu)[event] = pevent; |
| 96 | |
| 97 | return 0; |
| 98 | } |
| 99 | |
| 100 | static void op_destroy_counter(int cpu, int event) |
| 101 | { |
| 102 | struct perf_event *pevent = per_cpu(perf_events, cpu)[event]; |
| 103 | |
| 104 | if (pevent) { |
| 105 | perf_event_release_kernel(pevent); |
| 106 | per_cpu(perf_events, cpu)[event] = NULL; |
| 107 | } |
| 108 | } |
| 109 | |
| 110 | /* |
| 111 | * Called by oprofile_perf_start to create active perf events based on the |
| 112 | * perviously configured attributes. |
| 113 | */ |
| 114 | static int op_perf_start(void) |
| 115 | { |
| 116 | int cpu, event, ret = 0; |
| 117 | |
| 118 | for_each_online_cpu(cpu) { |
| 119 | for (event = 0; event < num_counters; ++event) { |
| 120 | ret = op_create_counter(cpu, event); |
| 121 | if (ret) |
| 122 | return ret; |
| 123 | } |
| 124 | } |
| 125 | |
| 126 | return ret; |
| 127 | } |
| 128 | |
| 129 | /* |
| 130 | * Called by oprofile_perf_stop at the end of a profiling run. |
| 131 | */ |
| 132 | static void op_perf_stop(void) |
| 133 | { |
| 134 | int cpu, event; |
| 135 | |
| 136 | for_each_online_cpu(cpu) |
| 137 | for (event = 0; event < num_counters; ++event) |
| 138 | op_destroy_counter(cpu, event); |
| 139 | } |
| 140 | |
| 141 | static int oprofile_perf_create_files(struct dentry *root) |
| 142 | { |
| 143 | unsigned int i; |
| 144 | |
| 145 | for (i = 0; i < num_counters; i++) { |
| 146 | struct dentry *dir; |
| 147 | char buf[4]; |
| 148 | |
| 149 | snprintf(buf, sizeof buf, "%d", i); |
| 150 | dir = oprofilefs_mkdir(root, buf); |
| 151 | oprofilefs_create_ulong(dir, "enabled", &counter_config[i].enabled); |
| 152 | oprofilefs_create_ulong(dir, "event", &counter_config[i].event); |
| 153 | oprofilefs_create_ulong(dir, "count", &counter_config[i].count); |
| 154 | oprofilefs_create_ulong(dir, "unit_mask", &counter_config[i].unit_mask); |
| 155 | oprofilefs_create_ulong(dir, "kernel", &counter_config[i].kernel); |
| 156 | oprofilefs_create_ulong(dir, "user", &counter_config[i].user); |
| 157 | } |
| 158 | |
| 159 | return 0; |
| 160 | } |
| 161 | |
| 162 | static int oprofile_perf_setup(void) |
| 163 | { |
| 164 | raw_spin_lock(&oprofilefs_lock); |
| 165 | op_perf_setup(); |
| 166 | raw_spin_unlock(&oprofilefs_lock); |
| 167 | return 0; |
| 168 | } |
| 169 | |
| 170 | static int oprofile_perf_start(void) |
| 171 | { |
| 172 | int ret = -EBUSY; |
| 173 | |
| 174 | mutex_lock(&oprofile_perf_mutex); |
| 175 | if (!oprofile_perf_enabled) { |
| 176 | ret = 0; |
| 177 | op_perf_start(); |
| 178 | oprofile_perf_enabled = 1; |
| 179 | } |
| 180 | mutex_unlock(&oprofile_perf_mutex); |
| 181 | return ret; |
| 182 | } |
| 183 | |
| 184 | static void oprofile_perf_stop(void) |
| 185 | { |
| 186 | mutex_lock(&oprofile_perf_mutex); |
| 187 | if (oprofile_perf_enabled) |
| 188 | op_perf_stop(); |
| 189 | oprofile_perf_enabled = 0; |
| 190 | mutex_unlock(&oprofile_perf_mutex); |
| 191 | } |
| 192 | |
| 193 | #ifdef CONFIG_PM |
| 194 | |
| 195 | static int oprofile_perf_suspend(struct platform_device *dev, pm_message_t state) |
| 196 | { |
| 197 | mutex_lock(&oprofile_perf_mutex); |
| 198 | if (oprofile_perf_enabled) |
| 199 | op_perf_stop(); |
| 200 | mutex_unlock(&oprofile_perf_mutex); |
| 201 | return 0; |
| 202 | } |
| 203 | |
| 204 | static int oprofile_perf_resume(struct platform_device *dev) |
| 205 | { |
| 206 | mutex_lock(&oprofile_perf_mutex); |
| 207 | if (oprofile_perf_enabled && op_perf_start()) |
| 208 | oprofile_perf_enabled = 0; |
| 209 | mutex_unlock(&oprofile_perf_mutex); |
| 210 | return 0; |
| 211 | } |
| 212 | |
| 213 | static struct platform_driver oprofile_driver = { |
| 214 | .driver = { |
| 215 | .name = "oprofile-perf", |
| 216 | }, |
| 217 | .resume = oprofile_perf_resume, |
| 218 | .suspend = oprofile_perf_suspend, |
| 219 | }; |
| 220 | |
| 221 | static struct platform_device *oprofile_pdev; |
| 222 | |
| 223 | static int __init init_driverfs(void) |
| 224 | { |
| 225 | int ret; |
| 226 | |
| 227 | ret = platform_driver_register(&oprofile_driver); |
| 228 | if (ret) |
| 229 | return ret; |
| 230 | |
| 231 | oprofile_pdev = platform_device_register_simple( |
| 232 | oprofile_driver.driver.name, 0, NULL, 0); |
| 233 | if (IS_ERR(oprofile_pdev)) { |
| 234 | ret = PTR_ERR(oprofile_pdev); |
| 235 | platform_driver_unregister(&oprofile_driver); |
| 236 | } |
| 237 | |
| 238 | return ret; |
| 239 | } |
| 240 | |
| 241 | static void exit_driverfs(void) |
| 242 | { |
| 243 | platform_device_unregister(oprofile_pdev); |
| 244 | platform_driver_unregister(&oprofile_driver); |
| 245 | } |
| 246 | |
| 247 | #else |
| 248 | |
| 249 | static inline int init_driverfs(void) { return 0; } |
| 250 | static inline void exit_driverfs(void) { } |
| 251 | |
| 252 | #endif /* CONFIG_PM */ |
| 253 | |
| 254 | void oprofile_perf_exit(void) |
| 255 | { |
| 256 | int cpu, id; |
| 257 | struct perf_event *event; |
| 258 | |
| 259 | for_each_possible_cpu(cpu) { |
| 260 | for (id = 0; id < num_counters; ++id) { |
| 261 | event = per_cpu(perf_events, cpu)[id]; |
| 262 | if (event) |
| 263 | perf_event_release_kernel(event); |
| 264 | } |
| 265 | |
| 266 | kfree(per_cpu(perf_events, cpu)); |
| 267 | } |
| 268 | |
| 269 | kfree(counter_config); |
| 270 | exit_driverfs(); |
| 271 | } |
| 272 | |
| 273 | int __init oprofile_perf_init(struct oprofile_operations *ops) |
| 274 | { |
| 275 | int cpu, ret = 0; |
| 276 | |
| 277 | ret = init_driverfs(); |
| 278 | if (ret) |
| 279 | return ret; |
| 280 | |
| 281 | num_counters = perf_num_counters(); |
| 282 | if (num_counters <= 0) { |
| 283 | pr_info("oprofile: no performance counters\n"); |
| 284 | ret = -ENODEV; |
| 285 | goto out; |
| 286 | } |
| 287 | |
| 288 | counter_config = kcalloc(num_counters, |
| 289 | sizeof(struct op_counter_config), GFP_KERNEL); |
| 290 | |
| 291 | if (!counter_config) { |
| 292 | pr_info("oprofile: failed to allocate %d " |
| 293 | "counters\n", num_counters); |
| 294 | ret = -ENOMEM; |
| 295 | num_counters = 0; |
| 296 | goto out; |
| 297 | } |
| 298 | |
| 299 | for_each_possible_cpu(cpu) { |
| 300 | per_cpu(perf_events, cpu) = kcalloc(num_counters, |
| 301 | sizeof(struct perf_event *), GFP_KERNEL); |
| 302 | if (!per_cpu(perf_events, cpu)) { |
| 303 | pr_info("oprofile: failed to allocate %d perf events " |
| 304 | "for cpu %d\n", num_counters, cpu); |
| 305 | ret = -ENOMEM; |
| 306 | goto out; |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | ops->create_files = oprofile_perf_create_files; |
| 311 | ops->setup = oprofile_perf_setup; |
| 312 | ops->start = oprofile_perf_start; |
| 313 | ops->stop = oprofile_perf_stop; |
| 314 | ops->shutdown = oprofile_perf_stop; |
| 315 | ops->cpu_type = op_name_from_perf_id(); |
| 316 | |
| 317 | if (!ops->cpu_type) |
| 318 | ret = -ENODEV; |
| 319 | else |
| 320 | pr_info("oprofile: using %s\n", ops->cpu_type); |
| 321 | |
| 322 | out: |
| 323 | if (ret) |
| 324 | oprofile_perf_exit(); |
| 325 | |
| 326 | return ret; |
| 327 | } |