arch/arm/kernel/perf_event_cpu.c

   1 /*
   2  * This program is free software; you can redistribute it and/or modify
   3  * it under the terms of the GNU General Public License version 2 as
   4  * published by the Free Software Foundation.
   5  *
   6  * This program is distributed in the hope that it will be useful,
   7  * but WITHOUT ANY WARRANTY; without even the implied warranty of
   8  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
   9  * GNU General Public License for more details.
  10  *
  11  * You should have received a copy of the GNU General Public License
  12  * along with this program; if not, write to the Free Software
  13  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  14  *
  15  * Copyright (C) 2012 ARM Limited
  16  *
  17  * Author: Will Deacon <will.deacon@arm.com>
  18  */
  19 #define pr_fmt(fmt) "CPU PMU: " fmt
  20
  21 #include <linux/bitmap.h>
  22 #include <linux/export.h>
  23 #include <linux/kernel.h>
  24 #include <linux/of.h>
  25 #include <linux/platform_device.h>
  26 #include <linux/slab.h>
  27 #include <linux/spinlock.h>
  28 #include <linux/irq.h>
  29 #include <linux/irqdesc.h>
  30
  31 #include <asm/cputype.h>
  32 #include <asm/irq_regs.h>
  33 #include <asm/pmu.h>
  34
  35 /* Set at runtime when we know what CPU type we are. */
  36 static struct arm_pmu *__oprofile_cpu_pmu;
  37
  38 /*
  39  * Despite the names, these two functions are CPU-specific and are used
  40  * by the OProfile/perf code.
  41  */
  42 const char *perf_pmu_name(void)
  43 {
  44         if (!__oprofile_cpu_pmu)
  45                 return NULL;
  46
  47         return __oprofile_cpu_pmu->name;
  48 }
  49 EXPORT_SYMBOL_GPL(perf_pmu_name);
  50
  51 int perf_num_counters(void)
  52 {
  53         int max_events = 0;
  54
  55         if (__oprofile_cpu_pmu != NULL)
  56                 max_events = __oprofile_cpu_pmu->num_events;
  57
  58         return max_events;
  59 }
  60 EXPORT_SYMBOL_GPL(perf_num_counters);
  61
  62 /* Include the PMU-specific implementations. */
  63 #include "perf_event_v6.c"
  64 #include "perf_event_v7.c"
  65
  66 static void cpu_pmu_enable_percpu_irq(void *data)
  67 {
  68         int irq = *(int *)data;
  69
  70         enable_percpu_irq(irq, IRQ_TYPE_NONE);
  71 }
  72
  73 static void cpu_pmu_disable_percpu_irq(void *data)
  74 {
  75         int irq = *(int *)data;
  76
  77         disable_percpu_irq(irq);
  78 }
  79
  80 static void cpu_pmu_free_irq(struct arm_pmu *cpu_pmu)
  81 {
  82         int i, irq, irqs;
  83         struct platform_device *pmu_device = cpu_pmu->plat_device;
  84         struct pmu_hw_events __percpu *hw_events = cpu_pmu->hw_events;
  85
  86         irqs = min(pmu_device->num_resources, num_possible_cpus());
  87
  88         irq = platform_get_irq(pmu_device, 0);
  89         if (irq >= 0 && irq_is_percpu(irq)) {
  90                 on_each_cpu(cpu_pmu_disable_percpu_irq, &irq, 1);
  91                 free_percpu_irq(irq, &hw_events->percpu_pmu);
  92         } else {
  93                 for (i = 0; i < irqs; ++i) {
  94                         int cpu = i;
  95
  96                         if (cpu_pmu->irq_affinity)
  97                                 cpu = cpu_pmu->irq_affinity[i];
  98
  99                         if (!cpumask_test_and_clear_cpu(cpu, &cpu_pmu->active_irqs))
 100                                 continue;
 101                         irq = platform_get_irq(pmu_device, i);
 102                         if (irq >= 0)
 103                                 free_irq(irq, per_cpu_ptr(&hw_events->percpu_pmu, cpu));
 104                 }
 105         }
 106 }
 107
 108 static int cpu_pmu_request_irq(struct arm_pmu *cpu_pmu, irq_handler_t handler)
 109 {
 110         int i, err, irq, irqs;
 111         struct platform_device *pmu_device = cpu_pmu->plat_device;
 112         struct pmu_hw_events __percpu *hw_events = cpu_pmu->hw_events;
 113
 114         if (!pmu_device)
 115                 return -ENODEV;
 116
 117         irqs = min(pmu_device->num_resources, num_possible_cpus());
 118         if (irqs < 1) {
 119                 pr_warn_once("perf/ARM: No irqs for PMU defined, sampling events not supported\n");
 120                 return 0;
 121         }
 122
 123         irq = platform_get_irq(pmu_device, 0);
 124         if (irq >= 0 && irq_is_percpu(irq)) {
 125                 err = request_percpu_irq(irq, handler, "arm-pmu",
 126                                          &hw_events->percpu_pmu);
 127                 if (err) {
 128                         pr_err("unable to request IRQ%d for ARM PMU counters\n",
 129                                 irq);
 130                         return err;
 131                 }
 132                 on_each_cpu(cpu_pmu_enable_percpu_irq, &irq, 1);
 133         } else {
 134                 for (i = 0; i < irqs; ++i) {
 135                         int cpu = i;
 136
 137                         err = 0;
 138                         irq = platform_get_irq(pmu_device, i);
 139                         if (irq < 0)
 140                                 continue;
 141
 142                         if (cpu_pmu->irq_affinity)
 143                                 cpu = cpu_pmu->irq_affinity[i];
 144
 145                         /*
 146                          * If we have a single PMU interrupt that we can't shift,
 147                          * assume that we're running on a uniprocessor machine and
 148                          * continue. Otherwise, continue without this interrupt.
 149                          */
 150                         if (irq_set_affinity(irq, cpumask_of(cpu)) && irqs > 1) {
 151                                 pr_warn("unable to set irq affinity (irq=%d, cpu=%u)\n",
 152                                         irq, cpu);
 153                                 continue;
 154                         }
 155
 156                         err = request_irq(irq, handler,
 157                                           IRQF_NOBALANCING | IRQF_NO_THREAD, "arm-pmu",
 158                                           per_cpu_ptr(&hw_events->percpu_pmu, cpu));
 159                         if (err) {
 160                                 pr_err("unable to request IRQ%d for ARM PMU counters\n",
 161                                         irq);
 162                                 return err;
 163                         }
 164
 165                         cpumask_set_cpu(cpu, &cpu_pmu->active_irqs);
 166                 }
 167         }
 168
 169         return 0;
 170 }
 171
 172 /*
 173  * PMU hardware loses all context when a CPU goes offline.
 174  * When a CPU is hotplugged back in, since some hardware registers are
 175  * UNKNOWN at reset, the PMU must be explicitly reset to avoid reading
 176  * junk values out of them.
 177  */
 178 static int cpu_pmu_notify(struct notifier_block *b, unsigned long action,
 179                           void *hcpu)
 180 {
 181         int cpu = (unsigned long)hcpu;
 182         struct arm_pmu *pmu = container_of(b, struct arm_pmu, hotplug_nb);
 183
 184         if ((action & ~CPU_TASKS_FROZEN) != CPU_STARTING)
 185                 return NOTIFY_DONE;
 186
 187         if (!cpumask_test_cpu(cpu, &pmu->supported_cpus))
 188                 return NOTIFY_DONE;
 189
 190         if (pmu->reset)
 191                 pmu->reset(pmu);
 192         else
 193                 return NOTIFY_DONE;
 194
 195         return NOTIFY_OK;
 196 }
 197
 198 static int cpu_pmu_init(struct arm_pmu *cpu_pmu)
 199 {
 200         int err;
 201         int cpu;
 202         struct pmu_hw_events __percpu *cpu_hw_events;
 203
 204         cpu_hw_events = alloc_percpu(struct pmu_hw_events);
 205         if (!cpu_hw_events)
 206                 return -ENOMEM;
 207
 208         cpu_pmu->hotplug_nb.notifier_call = cpu_pmu_notify;
 209         err = register_cpu_notifier(&cpu_pmu->hotplug_nb);
 210         if (err)
 211                 goto out_hw_events;
 212
 213         for_each_possible_cpu(cpu) {
 214                 struct pmu_hw_events *events = per_cpu_ptr(cpu_hw_events, cpu);
 215                 raw_spin_lock_init(&events->pmu_lock);
 216                 events->percpu_pmu = cpu_pmu;
 217         }
 218
 219         cpu_pmu->hw_events      = cpu_hw_events;
 220         cpu_pmu->request_irq    = cpu_pmu_request_irq;
 221         cpu_pmu->free_irq       = cpu_pmu_free_irq;
 222
 223         /* Ensure the PMU has sane values out of reset. */
 224         if (cpu_pmu->reset)
 225                 on_each_cpu_mask(&cpu_pmu->supported_cpus, cpu_pmu->reset,
 226                          cpu_pmu, 1);
 227
 228         /* If no interrupts available, set the corresponding capability flag */
 229         if (!platform_get_irq(cpu_pmu->plat_device, 0))
 230                 cpu_pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
 231
 232         return 0;
 233
 234 out_hw_events:
 235         free_percpu(cpu_hw_events);
 236         return err;
 237 }
 238
 239 static void cpu_pmu_destroy(struct arm_pmu *cpu_pmu)
 240 {
 241         unregister_cpu_notifier(&cpu_pmu->hotplug_nb);
 242         free_percpu(cpu_pmu->hw_events);
 243 }
 244
 245 /*
 246  * PMU platform driver and devicetree bindings.
 247  */
 248 static const struct of_device_id cpu_pmu_of_device_ids[] = {
 249         {.compatible = "arm,cortex-a17-pmu",    .data = armv7_a17_pmu_init},
 250         {.compatible = "arm,cortex-a15-pmu",    .data = armv7_a15_pmu_init},
 251         {.compatible = "arm,cortex-a12-pmu",    .data = armv7_a12_pmu_init},
 252         {.compatible = "arm,cortex-a9-pmu",     .data = armv7_a9_pmu_init},
 253         {.compatible = "arm,cortex-a8-pmu",     .data = armv7_a8_pmu_init},
 254         {.compatible = "arm,cortex-a7-pmu",     .data = armv7_a7_pmu_init},
 255         {.compatible = "arm,cortex-a5-pmu",     .data = armv7_a5_pmu_init},
 256         {.compatible = "arm,arm11mpcore-pmu",   .data = armv6mpcore_pmu_init},
 257         {.compatible = "arm,arm1176-pmu",       .data = armv6_1176_pmu_init},
 258         {.compatible = "arm,arm1136-pmu",       .data = armv6_1136_pmu_init},
 259         {.compatible = "qcom,krait-pmu",        .data = krait_pmu_init},
 260         {.compatible = "qcom,scorpion-pmu",     .data = scorpion_pmu_init},
 261         {.compatible = "qcom,scorpion-mp-pmu",  .data = scorpion_mp_pmu_init},
 262         {},
 263 };
 264
 265 static struct platform_device_id cpu_pmu_plat_device_ids[] = {
 266         {.name = "arm-pmu"},
 267         {.name = "armv6-pmu"},
 268         {.name = "armv7-pmu"},
 269         {},
 270 };
 271
 272 static const struct pmu_probe_info pmu_probe_table[] = {
 273         ARM_PMU_PROBE(ARM_CPU_PART_ARM1136, armv6_1136_pmu_init),
 274         ARM_PMU_PROBE(ARM_CPU_PART_ARM1156, armv6_1156_pmu_init),
 275         ARM_PMU_PROBE(ARM_CPU_PART_ARM1176, armv6_1176_pmu_init),
 276         ARM_PMU_PROBE(ARM_CPU_PART_ARM11MPCORE, armv6mpcore_pmu_init),
 277         ARM_PMU_PROBE(ARM_CPU_PART_CORTEX_A8, armv7_a8_pmu_init),
 278         ARM_PMU_PROBE(ARM_CPU_PART_CORTEX_A9, armv7_a9_pmu_init),
 279         { /* sentinel value */ }
 280 };
 281
 282 /*
 283  * CPU PMU identification and probing.
 284  */
 285 static int probe_current_pmu(struct arm_pmu *pmu,
 286                              const struct pmu_probe_info *info)
 287 {
 288         int cpu = get_cpu();
 289         unsigned int cpuid = read_cpuid_id();
 290         int ret = -ENODEV;
 291
 292         pr_info("probing PMU on CPU %d\n", cpu);
 293
 294         for (; info->init != NULL; info++) {
 295                 if ((cpuid & info->mask) != info->cpuid)
 296                         continue;
 297                 ret = info->init(pmu);
 298                 break;
 299         }
 300
 301         put_cpu();
 302         return ret;
 303 }
 304
 305 static int of_pmu_irq_cfg(struct arm_pmu *pmu)
 306 {
 307         int i;
 308         struct platform_device *pdev = pmu->plat_device;
 309         int *irqs = kcalloc(pdev->num_resources, sizeof(*irqs), GFP_KERNEL);
 310
 311         if (!irqs)
 312                 return -ENOMEM;
 313
 314         for (i = 0; i < pdev->num_resources; ++i) {
 315                 struct device_node *dn;
 316                 int cpu;
 317
 318                 dn = of_parse_phandle(pdev->dev.of_node, "interrupt-affinity",
 319                                       i);
 320                 if (!dn) {
 321                         pr_warn("Failed to parse %s/interrupt-affinity[%d]\n",
 322                                 of_node_full_name(dn), i);
 323                         break;
 324                 }
 325
 326                 for_each_possible_cpu(cpu)
 327                         if (arch_find_n_match_cpu_physical_id(dn, cpu, NULL))
 328                                 break;
 329
 330                 of_node_put(dn);
 331                 if (cpu >= nr_cpu_ids) {
 332                         pr_warn("Failed to find logical CPU for %s\n",
 333                                 dn->name);
 334                         break;
 335                 }
 336
 337                 irqs[i] = cpu;
 338                 cpumask_set_cpu(cpu, &pmu->supported_cpus);
 339         }
 340
 341         if (i == pdev->num_resources) {
 342                 pmu->irq_affinity = irqs;
 343         } else {
 344                 kfree(irqs);
 345                 cpumask_setall(&pmu->supported_cpus);
 346         }
 347
 348         return 0;
 349 }
 350
 351 int arm_pmu_device_probe(struct platform_device *pdev,
 352                          const struct of_device_id *of_table,
 353                          const struct pmu_probe_info *probe_table)
 354 {
 355         const struct of_device_id *of_id;
 356         const int (*init_fn)(struct arm_pmu *);
 357         struct device_node *node = pdev->dev.of_node;
 358         struct arm_pmu *pmu;
 359         int ret = -ENODEV;
 360
 361         pmu = kzalloc(sizeof(struct arm_pmu), GFP_KERNEL);
 362         if (!pmu) {
 363                 pr_info("failed to allocate PMU device!\n");
 364                 return -ENOMEM;
 365         }
 366
 367         if (!__oprofile_cpu_pmu)
 368                 __oprofile_cpu_pmu = pmu;
 369
 370         pmu->plat_device = pdev;
 371
 372         if (node && (of_id = of_match_node(of_table, pdev->dev.of_node))) {
 373                 init_fn = of_id->data;
 374
 375                 ret = of_pmu_irq_cfg(pmu);
 376                 if (!ret)
 377                         ret = init_fn(pmu);
 378         } else {
 379                 ret = probe_current_pmu(pmu, probe_table);
 380                 cpumask_setall(&pmu->supported_cpus);
 381         }
 382
 383         if (ret) {
 384                 pr_info("failed to probe PMU!\n");
 385                 goto out_free;
 386         }
 387
 388         ret = cpu_pmu_init(pmu);
 389         if (ret)
 390                 goto out_free;
 391
 392         ret = armpmu_register(pmu, -1);
 393         if (ret)
 394                 goto out_destroy;
 395
 396         return 0;
 397
 398 out_destroy:
 399         cpu_pmu_destroy(pmu);
 400 out_free:
 401         pr_info("failed to register PMU devices!\n");
 402         kfree(pmu);
 403         return ret;
 404 }
 405
 406 static int cpu_pmu_device_probe(struct platform_device *pdev)
 407 {
 408         return arm_pmu_device_probe(pdev, cpu_pmu_of_device_ids,
 409                                     pmu_probe_table);
 410 }
 411
 412 static struct platform_driver cpu_pmu_driver = {
 413         .driver         = {
 414                 .name   = "arm-pmu",
 415                 .of_match_table = cpu_pmu_of_device_ids,
 416         },
 417         .probe          = cpu_pmu_device_probe,
 418         .id_table       = cpu_pmu_plat_device_ids,
 419 };
 420
 421 static int __init register_pmu_driver(void)
 422 {
 423         return platform_driver_register(&cpu_pmu_driver);
 424 }
 425 device_initcall(register_pmu_driver);