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 *cpu_pmu;
  37
  38 static DEFINE_PER_CPU(struct arm_pmu *, percpu_pmu);
  39 static DEFINE_PER_CPU(struct perf_event * [ARMPMU_MAX_HWEVENTS], hw_events);
  40 static DEFINE_PER_CPU(unsigned long [BITS_TO_LONGS(ARMPMU_MAX_HWEVENTS)], used_mask);
  41 static DEFINE_PER_CPU(struct pmu_hw_events, cpu_hw_events);
  42
  43 /*
  44  * Despite the names, these two functions are CPU-specific and are used
  45  * by the OProfile/perf code.
  46  */
  47 const char *perf_pmu_name(void)
  48 {
  49         if (!cpu_pmu)
  50                 return NULL;
  51
  52         return cpu_pmu->name;
  53 }
  54 EXPORT_SYMBOL_GPL(perf_pmu_name);
  55
  56 int perf_num_counters(void)
  57 {
  58         int max_events = 0;
  59
  60         if (cpu_pmu != NULL)
  61                 max_events = cpu_pmu->num_events;
  62
  63         return max_events;
  64 }
  65 EXPORT_SYMBOL_GPL(perf_num_counters);
  66
  67 /* Include the PMU-specific implementations. */
  68 #include "perf_event_xscale.c"
  69 #include "perf_event_v6.c"
  70 #include "perf_event_v7.c"
  71
  72 static struct pmu_hw_events *cpu_pmu_get_cpu_events(void)
  73 {
  74         return this_cpu_ptr(&cpu_hw_events);
  75 }
  76
  77 static void cpu_pmu_enable_percpu_irq(void *data)
  78 {
  79         struct arm_pmu *cpu_pmu = data;
  80         struct platform_device *pmu_device = cpu_pmu->plat_device;
  81         int irq = platform_get_irq(pmu_device, 0);
  82
  83         enable_percpu_irq(irq, IRQ_TYPE_NONE);
  84         cpumask_set_cpu(smp_processor_id(), &cpu_pmu->active_irqs);
  85 }
  86
  87 static void cpu_pmu_disable_percpu_irq(void *data)
  88 {
  89         struct arm_pmu *cpu_pmu = data;
  90         struct platform_device *pmu_device = cpu_pmu->plat_device;
  91         int irq = platform_get_irq(pmu_device, 0);
  92
  93         cpumask_clear_cpu(smp_processor_id(), &cpu_pmu->active_irqs);
  94         disable_percpu_irq(irq);
  95 }
  96
  97 static void cpu_pmu_free_irq(struct arm_pmu *cpu_pmu)
  98 {
  99         int i, irq, irqs;
 100         struct platform_device *pmu_device = cpu_pmu->plat_device;
 101
 102         irqs = min(pmu_device->num_resources, num_possible_cpus());
 103
 104         irq = platform_get_irq(pmu_device, 0);
 105         if (irq >= 0 && irq_is_percpu(irq)) {
 106                 on_each_cpu(cpu_pmu_disable_percpu_irq, cpu_pmu, 1);
 107                 free_percpu_irq(irq, &percpu_pmu);
 108         } else {
 109                 for (i = 0; i < irqs; ++i) {
 110                         if (!cpumask_test_and_clear_cpu(i, &cpu_pmu->active_irqs))
 111                                 continue;
 112                         irq = platform_get_irq(pmu_device, i);
 113                         if (irq >= 0)
 114                                 free_irq(irq, cpu_pmu);
 115                 }
 116         }
 117 }
 118
 119 static int cpu_pmu_request_irq(struct arm_pmu *cpu_pmu, irq_handler_t handler)
 120 {
 121         int i, err, irq, irqs;
 122         struct platform_device *pmu_device = cpu_pmu->plat_device;
 123
 124         if (!pmu_device)
 125                 return -ENODEV;
 126
 127         irqs = min(pmu_device->num_resources, num_possible_cpus());
 128         if (irqs < 1) {
 129                 printk_once("perf/ARM: No irqs for PMU defined, sampling events not supported\n");
 130                 return 0;
 131         }
 132
 133         irq = platform_get_irq(pmu_device, 0);
 134         if (irq >= 0 && irq_is_percpu(irq)) {
 135                 err = request_percpu_irq(irq, handler, "arm-pmu", &percpu_pmu);
 136                 if (err) {
 137                         pr_err("unable to request IRQ%d for ARM PMU counters\n",
 138                                 irq);
 139                         return err;
 140                 }
 141                 on_each_cpu(cpu_pmu_enable_percpu_irq, cpu_pmu, 1);
 142         } else {
 143                 for (i = 0; i < irqs; ++i) {
 144                         err = 0;
 145                         irq = platform_get_irq(pmu_device, i);
 146                         if (irq < 0)
 147                                 continue;
 148
 149                         /*
 150                          * If we have a single PMU interrupt that we can't shift,
 151                          * assume that we're running on a uniprocessor machine and
 152                          * continue. Otherwise, continue without this interrupt.
 153                          */
 154                         if (irq_set_affinity(irq, cpumask_of(i)) && irqs > 1) {
 155                                 pr_warning("unable to set irq affinity (irq=%d, cpu=%u)\n",
 156                                             irq, i);
 157                                 continue;
 158                         }
 159
 160                         err = request_irq(irq, handler,
 161                                           IRQF_NOBALANCING | IRQF_NO_THREAD, "arm-pmu",
 162                                           cpu_pmu);
 163                         if (err) {
 164                                 pr_err("unable to request IRQ%d for ARM PMU counters\n",
 165                                         irq);
 166                                 return err;
 167                         }
 168
 169                         cpumask_set_cpu(i, &cpu_pmu->active_irqs);
 170                 }
 171         }
 172
 173         return 0;
 174 }
 175
 176 static void cpu_pmu_init(struct arm_pmu *cpu_pmu)
 177 {
 178         int cpu;
 179         for_each_possible_cpu(cpu) {
 180                 struct pmu_hw_events *events = &per_cpu(cpu_hw_events, cpu);
 181                 events->events = per_cpu(hw_events, cpu);
 182                 events->used_mask = per_cpu(used_mask, cpu);
 183                 raw_spin_lock_init(&events->pmu_lock);
 184                 per_cpu(percpu_pmu, cpu) = cpu_pmu;
 185         }
 186
 187         cpu_pmu->get_hw_events  = cpu_pmu_get_cpu_events;
 188         cpu_pmu->request_irq    = cpu_pmu_request_irq;
 189         cpu_pmu->free_irq       = cpu_pmu_free_irq;
 190
 191         /* Ensure the PMU has sane values out of reset. */
 192         if (cpu_pmu->reset)
 193                 on_each_cpu(cpu_pmu->reset, cpu_pmu, 1);
 194
 195         /* If no interrupts available, set the corresponding capability flag */
 196         if (!platform_get_irq(cpu_pmu->plat_device, 0))
 197                 cpu_pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
 198 }
 199
 200 /*
 201  * PMU hardware loses all context when a CPU goes offline.
 202  * When a CPU is hotplugged back in, since some hardware registers are
 203  * UNKNOWN at reset, the PMU must be explicitly reset to avoid reading
 204  * junk values out of them.
 205  */
 206 static int cpu_pmu_notify(struct notifier_block *b, unsigned long action,
 207                           void *hcpu)
 208 {
 209         if ((action & ~CPU_TASKS_FROZEN) != CPU_STARTING)
 210                 return NOTIFY_DONE;
 211
 212         if (cpu_pmu && cpu_pmu->reset)
 213                 cpu_pmu->reset(cpu_pmu);
 214         else
 215                 return NOTIFY_DONE;
 216
 217         return NOTIFY_OK;
 218 }
 219
 220 static struct notifier_block cpu_pmu_hotplug_notifier = {
 221         .notifier_call = cpu_pmu_notify,
 222 };
 223
 224 /*
 225  * PMU platform driver and devicetree bindings.
 226  */
 227 static struct of_device_id cpu_pmu_of_device_ids[] = {
 228         {.compatible = "arm,cortex-a17-pmu",    .data = armv7_a17_pmu_init},
 229         {.compatible = "arm,cortex-a15-pmu",    .data = armv7_a15_pmu_init},
 230         {.compatible = "arm,cortex-a12-pmu",    .data = armv7_a12_pmu_init},
 231         {.compatible = "arm,cortex-a9-pmu",     .data = armv7_a9_pmu_init},
 232         {.compatible = "arm,cortex-a8-pmu",     .data = armv7_a8_pmu_init},
 233         {.compatible = "arm,cortex-a7-pmu",     .data = armv7_a7_pmu_init},
 234         {.compatible = "arm,cortex-a5-pmu",     .data = armv7_a5_pmu_init},
 235         {.compatible = "arm,arm11mpcore-pmu",   .data = armv6mpcore_pmu_init},
 236         {.compatible = "arm,arm1176-pmu",       .data = armv6_1176_pmu_init},
 237         {.compatible = "arm,arm1136-pmu",       .data = armv6_1136_pmu_init},
 238         {.compatible = "qcom,krait-pmu",        .data = krait_pmu_init},
 239         {},
 240 };
 241
 242 static struct platform_device_id cpu_pmu_plat_device_ids[] = {
 243         {.name = "arm-pmu"},
 244         {},
 245 };
 246
 247 /*
 248  * CPU PMU identification and probing.
 249  */
 250 static int probe_current_pmu(struct arm_pmu *pmu)
 251 {
 252         int cpu = get_cpu();
 253         unsigned long implementor = read_cpuid_implementor();
 254         unsigned long part_number = read_cpuid_part_number();
 255         int ret = -ENODEV;
 256
 257         pr_info("probing PMU on CPU %d\n", cpu);
 258
 259         /* ARM Ltd CPUs. */
 260         if (implementor == ARM_CPU_IMP_ARM) {
 261                 switch (part_number) {
 262                 case ARM_CPU_PART_ARM1136:
 263                         ret = armv6_1136_pmu_init(pmu);
 264                         break;
 265                 case ARM_CPU_PART_ARM1156:
 266                         ret = armv6_1156_pmu_init(pmu);
 267                         break;
 268                 case ARM_CPU_PART_ARM1176:
 269                         ret = armv6_1176_pmu_init(pmu);
 270                         break;
 271                 case ARM_CPU_PART_ARM11MPCORE:
 272                         ret = armv6mpcore_pmu_init(pmu);
 273                         break;
 274                 case ARM_CPU_PART_CORTEX_A8:
 275                         ret = armv7_a8_pmu_init(pmu);
 276                         break;
 277                 case ARM_CPU_PART_CORTEX_A9:
 278                         ret = armv7_a9_pmu_init(pmu);
 279                         break;
 280                 }
 281         /* Intel CPUs [xscale]. */
 282         } else if (implementor == ARM_CPU_IMP_INTEL) {
 283                 switch (xscale_cpu_arch_version()) {
 284                 case ARM_CPU_XSCALE_ARCH_V1:
 285                         ret = xscale1pmu_init(pmu);
 286                         break;
 287                 case ARM_CPU_XSCALE_ARCH_V2:
 288                         ret = xscale2pmu_init(pmu);
 289                         break;
 290                 }
 291         }
 292
 293         put_cpu();
 294         return ret;
 295 }
 296
 297 static int cpu_pmu_device_probe(struct platform_device *pdev)
 298 {
 299         const struct of_device_id *of_id;
 300         const int (*init_fn)(struct arm_pmu *);
 301         struct device_node *node = pdev->dev.of_node;
 302         struct arm_pmu *pmu;
 303         int ret = -ENODEV;
 304
 305         if (cpu_pmu) {
 306                 pr_info("attempt to register multiple PMU devices!");
 307                 return -ENOSPC;
 308         }
 309
 310         pmu = kzalloc(sizeof(struct arm_pmu), GFP_KERNEL);
 311         if (!pmu) {
 312                 pr_info("failed to allocate PMU device!");
 313                 return -ENOMEM;
 314         }
 315
 316         cpu_pmu = pmu;
 317         cpu_pmu->plat_device = pdev;
 318
 319         if (node && (of_id = of_match_node(cpu_pmu_of_device_ids, pdev->dev.of_node))) {
 320                 init_fn = of_id->data;
 321                 ret = init_fn(pmu);
 322         } else {
 323                 ret = probe_current_pmu(pmu);
 324         }
 325
 326         if (ret) {
 327                 pr_info("failed to probe PMU!");
 328                 goto out_free;
 329         }
 330
 331         cpu_pmu_init(cpu_pmu);
 332         ret = armpmu_register(cpu_pmu, PERF_TYPE_RAW);
 333
 334         if (!ret)
 335                 return 0;
 336
 337 out_free:
 338         pr_info("failed to register PMU devices!");
 339         kfree(pmu);
 340         return ret;
 341 }
 342
 343 static struct platform_driver cpu_pmu_driver = {
 344         .driver         = {
 345                 .name   = "arm-pmu",
 346                 .pm     = &armpmu_dev_pm_ops,
 347                 .of_match_table = cpu_pmu_of_device_ids,
 348         },
 349         .probe          = cpu_pmu_device_probe,
 350         .id_table       = cpu_pmu_plat_device_ids,
 351 };
 352
 353 static int __init register_pmu_driver(void)
 354 {
 355         int err;
 356
 357         err = register_cpu_notifier(&cpu_pmu_hotplug_notifier);
 358         if (err)
 359                 return err;
 360
 361         err = platform_driver_register(&cpu_pmu_driver);
 362         if (err)
 363                 unregister_cpu_notifier(&cpu_pmu_hotplug_notifier);
 364
 365         return err;
 366 }
 367 device_initcall(register_pmu_driver);