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_xscale.c"
  64 #include "perf_event_v6.c"
  65 #include "perf_event_v7.c"
  66
  67 static void cpu_pmu_enable_percpu_irq(void *data)
  68 {
  69         int irq = *(int *)data;
  70
  71         enable_percpu_irq(irq, IRQ_TYPE_NONE);
  72 }
  73
  74 static void cpu_pmu_disable_percpu_irq(void *data)
  75 {
  76         int irq = *(int *)data;
  77
  78         disable_percpu_irq(irq);
  79 }
  80
  81 static void cpu_pmu_free_irq(struct arm_pmu *cpu_pmu)
  82 {
  83         int i, irq, irqs;
  84         struct platform_device *pmu_device = cpu_pmu->plat_device;
  85         struct pmu_hw_events __percpu *hw_events = cpu_pmu->hw_events;
  86
  87         irqs = min(pmu_device->num_resources, num_possible_cpus());
  88
  89         irq = platform_get_irq(pmu_device, 0);
  90         if (irq >= 0 && irq_is_percpu(irq)) {
  91                 on_each_cpu(cpu_pmu_disable_percpu_irq, &irq, 1);
  92                 free_percpu_irq(irq, &hw_events->percpu_pmu);
  93         } else {
  94                 for (i = 0; i < irqs; ++i) {
  95                         int cpu = i;
  96
  97                         if (cpu_pmu->irq_affinity)
  98                                 cpu = cpu_pmu->irq_affinity[i];
  99
 100                         if (!cpumask_test_and_clear_cpu(cpu, &cpu_pmu->active_irqs))
 101                                 continue;
 102                         irq = platform_get_irq(pmu_device, i);
 103                         if (irq >= 0)
 104                                 free_irq(irq, per_cpu_ptr(&hw_events->percpu_pmu, cpu));
 105                 }
 106         }
 107 }
 108
 109 static int cpu_pmu_request_irq(struct arm_pmu *cpu_pmu, irq_handler_t handler)
 110 {
 111         int i, err, irq, irqs;
 112         struct platform_device *pmu_device = cpu_pmu->plat_device;
 113         struct pmu_hw_events __percpu *hw_events = cpu_pmu->hw_events;
 114
 115         if (!pmu_device)
 116                 return -ENODEV;
 117
 118         irqs = min(pmu_device->num_resources, num_possible_cpus());
 119         if (irqs < 1) {
 120                 pr_warn_once("perf/ARM: No irqs for PMU defined, sampling events not supported\n");
 121                 return 0;
 122         }
 123
 124         irq = platform_get_irq(pmu_device, 0);
 125         if (irq >= 0 && irq_is_percpu(irq)) {
 126                 err = request_percpu_irq(irq, handler, "arm-pmu",
 127                                          &hw_events->percpu_pmu);
 128                 if (err) {
 129                         pr_err("unable to request IRQ%d for ARM PMU counters\n",
 130                                 irq);
 131                         return err;
 132                 }
 133                 on_each_cpu(cpu_pmu_enable_percpu_irq, &irq, 1);
 134         } else {
 135                 for (i = 0; i < irqs; ++i) {
 136                         int cpu = i;
 137
 138                         err = 0;
 139                         irq = platform_get_irq(pmu_device, i);
 140                         if (irq < 0)
 141                                 continue;
 142
 143                         if (cpu_pmu->irq_affinity)
 144                                 cpu = cpu_pmu->irq_affinity[i];
 145
 146                         /*
 147                          * If we have a single PMU interrupt that we can't shift,
 148                          * assume that we're running on a uniprocessor machine and
 149                          * continue. Otherwise, continue without this interrupt.
 150                          */
 151                         if (irq_set_affinity(irq, cpumask_of(cpu)) && irqs > 1) {
 152                                 pr_warn("unable to set irq affinity (irq=%d, cpu=%u)\n",
 153                                         irq, cpu);
 154                                 continue;
 155                         }
 156
 157                         err = request_irq(irq, handler,
 158                                           IRQF_NOBALANCING | IRQF_NO_THREAD, "arm-pmu",
 159                                           per_cpu_ptr(&hw_events->percpu_pmu, cpu));
 160                         if (err) {
 161                                 pr_err("unable to request IRQ%d for ARM PMU counters\n",
 162                                         irq);
 163                                 return err;
 164                         }
 165
 166                         cpumask_set_cpu(cpu, &cpu_pmu->active_irqs);
 167                 }
 168         }
 169
 170         return 0;
 171 }
 172
 173 /*
 174  * PMU hardware loses all context when a CPU goes offline.
 175  * When a CPU is hotplugged back in, since some hardware registers are
 176  * UNKNOWN at reset, the PMU must be explicitly reset to avoid reading
 177  * junk values out of them.
 178  */
 179 static int cpu_pmu_notify(struct notifier_block *b, unsigned long action,
 180                           void *hcpu)
 181 {
 182         int cpu = (unsigned long)hcpu;
 183         struct arm_pmu *pmu = container_of(b, struct arm_pmu, hotplug_nb);
 184
 185         if ((action & ~CPU_TASKS_FROZEN) != CPU_STARTING)
 186                 return NOTIFY_DONE;
 187
 188         if (!cpumask_test_cpu(cpu, &pmu->supported_cpus))
 189                 return NOTIFY_DONE;
 190
 191         if (pmu->reset)
 192                 pmu->reset(pmu);
 193         else
 194                 return NOTIFY_DONE;
 195
 196         return NOTIFY_OK;
 197 }
 198
 199 static int cpu_pmu_init(struct arm_pmu *cpu_pmu)
 200 {
 201         int err;
 202         int cpu;
 203         struct pmu_hw_events __percpu *cpu_hw_events;
 204
 205         cpu_hw_events = alloc_percpu(struct pmu_hw_events);
 206         if (!cpu_hw_events)
 207                 return -ENOMEM;
 208
 209         cpu_pmu->hotplug_nb.notifier_call = cpu_pmu_notify;
 210         err = register_cpu_notifier(&cpu_pmu->hotplug_nb);
 211         if (err)
 212                 goto out_hw_events;
 213
 214         for_each_possible_cpu(cpu) {
 215                 struct pmu_hw_events *events = per_cpu_ptr(cpu_hw_events, cpu);
 216                 raw_spin_lock_init(&events->pmu_lock);
 217                 events->percpu_pmu = cpu_pmu;
 218         }
 219
 220         cpu_pmu->hw_events      = cpu_hw_events;
 221         cpu_pmu->request_irq    = cpu_pmu_request_irq;
 222         cpu_pmu->free_irq       = cpu_pmu_free_irq;
 223
 224         /* Ensure the PMU has sane values out of reset. */
 225         if (cpu_pmu->reset)
 226                 on_each_cpu_mask(&cpu_pmu->supported_cpus, cpu_pmu->reset,
 227                          cpu_pmu, 1);
 228
 229         /* If no interrupts available, set the corresponding capability flag */
 230         if (!platform_get_irq(cpu_pmu->plat_device, 0))
 231                 cpu_pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
 232
 233         return 0;
 234
 235 out_hw_events:
 236         free_percpu(cpu_hw_events);
 237         return err;
 238 }
 239
 240 static void cpu_pmu_destroy(struct arm_pmu *cpu_pmu)
 241 {
 242         unregister_cpu_notifier(&cpu_pmu->hotplug_nb);
 243         free_percpu(cpu_pmu->hw_events);
 244 }
 245
 246 /*
 247  * PMU platform driver and devicetree bindings.
 248  */
 249 static const struct of_device_id cpu_pmu_of_device_ids[] = {
 250         {.compatible = "arm,cortex-a17-pmu",    .data = armv7_a17_pmu_init},
 251         {.compatible = "arm,cortex-a15-pmu",    .data = armv7_a15_pmu_init},
 252         {.compatible = "arm,cortex-a12-pmu",    .data = armv7_a12_pmu_init},
 253         {.compatible = "arm,cortex-a9-pmu",     .data = armv7_a9_pmu_init},
 254         {.compatible = "arm,cortex-a8-pmu",     .data = armv7_a8_pmu_init},
 255         {.compatible = "arm,cortex-a7-pmu",     .data = armv7_a7_pmu_init},
 256         {.compatible = "arm,cortex-a5-pmu",     .data = armv7_a5_pmu_init},
 257         {.compatible = "arm,arm11mpcore-pmu",   .data = armv6mpcore_pmu_init},
 258         {.compatible = "arm,arm1176-pmu",       .data = armv6_1176_pmu_init},
 259         {.compatible = "arm,arm1136-pmu",       .data = armv6_1136_pmu_init},
 260         {.compatible = "qcom,krait-pmu",        .data = krait_pmu_init},
 261         {.compatible = "qcom,scorpion-pmu",     .data = scorpion_pmu_init},
 262         {.compatible = "qcom,scorpion-mp-pmu",  .data = scorpion_mp_pmu_init},
 263         {},
 264 };
 265
 266 static struct platform_device_id cpu_pmu_plat_device_ids[] = {
 267         {.name = "arm-pmu"},
 268         {.name = "armv6-pmu"},
 269         {.name = "armv7-pmu"},
 270         {.name = "xscale-pmu"},
 271         {},
 272 };
 273
 274 static const struct pmu_probe_info pmu_probe_table[] = {
 275         ARM_PMU_PROBE(ARM_CPU_PART_ARM1136, armv6_1136_pmu_init),
 276         ARM_PMU_PROBE(ARM_CPU_PART_ARM1156, armv6_1156_pmu_init),
 277         ARM_PMU_PROBE(ARM_CPU_PART_ARM1176, armv6_1176_pmu_init),
 278         ARM_PMU_PROBE(ARM_CPU_PART_ARM11MPCORE, armv6mpcore_pmu_init),
 279         ARM_PMU_PROBE(ARM_CPU_PART_CORTEX_A8, armv7_a8_pmu_init),
 280         ARM_PMU_PROBE(ARM_CPU_PART_CORTEX_A9, armv7_a9_pmu_init),
 281         XSCALE_PMU_PROBE(ARM_CPU_XSCALE_ARCH_V1, xscale1pmu_init),
 282         XSCALE_PMU_PROBE(ARM_CPU_XSCALE_ARCH_V2, xscale2pmu_init),
 283         { /* sentinel value */ }
 284 };
 285
 286 /*
 287  * CPU PMU identification and probing.
 288  */
 289 static int probe_current_pmu(struct arm_pmu *pmu,
 290                              const struct pmu_probe_info *info)
 291 {
 292         int cpu = get_cpu();
 293         unsigned int cpuid = read_cpuid_id();
 294         int ret = -ENODEV;
 295
 296         pr_info("probing PMU on CPU %d\n", cpu);
 297
 298         for (; info->init != NULL; info++) {
 299                 if ((cpuid & info->mask) != info->cpuid)
 300                         continue;
 301                 ret = info->init(pmu);
 302                 break;
 303         }
 304
 305         put_cpu();
 306         return ret;
 307 }
 308
 309 static int of_pmu_irq_cfg(struct arm_pmu *pmu)
 310 {
 311         int i;
 312         struct platform_device *pdev = pmu->plat_device;
 313         int *irqs = kcalloc(pdev->num_resources, sizeof(*irqs), GFP_KERNEL);
 314
 315         if (!irqs)
 316                 return -ENOMEM;
 317
 318         for (i = 0; i < pdev->num_resources; ++i) {
 319                 struct device_node *dn;
 320                 int cpu;
 321
 322                 dn = of_parse_phandle(pdev->dev.of_node, "interrupt-affinity",
 323                                       i);
 324                 if (!dn) {
 325                         pr_warn("Failed to parse %s/interrupt-affinity[%d]\n",
 326                                 of_node_full_name(dn), i);
 327                         break;
 328                 }
 329
 330                 for_each_possible_cpu(cpu)
 331                         if (arch_find_n_match_cpu_physical_id(dn, cpu, NULL))
 332                                 break;
 333
 334                 of_node_put(dn);
 335                 if (cpu >= nr_cpu_ids) {
 336                         pr_warn("Failed to find logical CPU for %s\n",
 337                                 dn->name);
 338                         break;
 339                 }
 340
 341                 irqs[i] = cpu;
 342                 cpumask_set_cpu(cpu, &pmu->supported_cpus);
 343         }
 344
 345         if (i == pdev->num_resources) {
 346                 pmu->irq_affinity = irqs;
 347         } else {
 348                 kfree(irqs);
 349                 cpumask_setall(&pmu->supported_cpus);
 350         }
 351
 352         return 0;
 353 }
 354
 355 int arm_pmu_device_probe(struct platform_device *pdev,
 356                          const struct of_device_id *of_table,
 357                          const struct pmu_probe_info *probe_table)
 358 {
 359         const struct of_device_id *of_id;
 360         const int (*init_fn)(struct arm_pmu *);
 361         struct device_node *node = pdev->dev.of_node;
 362         struct arm_pmu *pmu;
 363         int ret = -ENODEV;
 364
 365         pmu = kzalloc(sizeof(struct arm_pmu), GFP_KERNEL);
 366         if (!pmu) {
 367                 pr_info("failed to allocate PMU device!\n");
 368                 return -ENOMEM;
 369         }
 370
 371         if (!__oprofile_cpu_pmu)
 372                 __oprofile_cpu_pmu = pmu;
 373
 374         pmu->plat_device = pdev;
 375
 376         if (node && (of_id = of_match_node(of_table, pdev->dev.of_node))) {
 377                 init_fn = of_id->data;
 378
 379                 ret = of_pmu_irq_cfg(pmu);
 380                 if (!ret)
 381                         ret = init_fn(pmu);
 382         } else {
 383                 ret = probe_current_pmu(pmu, probe_table);
 384                 cpumask_setall(&pmu->supported_cpus);
 385         }
 386
 387         if (ret) {
 388                 pr_info("failed to probe PMU!\n");
 389                 goto out_free;
 390         }
 391
 392         ret = cpu_pmu_init(pmu);
 393         if (ret)
 394                 goto out_free;
 395
 396         ret = armpmu_register(pmu, -1);
 397         if (ret)
 398                 goto out_destroy;
 399
 400         return 0;
 401
 402 out_destroy:
 403         cpu_pmu_destroy(pmu);
 404 out_free:
 405         pr_info("failed to register PMU devices!\n");
 406         kfree(pmu);
 407         return ret;
 408 }
 409
 410 static int cpu_pmu_device_probe(struct platform_device *pdev)
 411 {
 412         return arm_pmu_device_probe(pdev, cpu_pmu_of_device_ids,
 413                                     pmu_probe_table);
 414 }
 415
 416 static struct platform_driver cpu_pmu_driver = {
 417         .driver         = {
 418                 .name   = "arm-pmu",
 419                 .of_match_table = cpu_pmu_of_device_ids,
 420         },
 421         .probe          = cpu_pmu_device_probe,
 422         .id_table       = cpu_pmu_plat_device_ids,
 423 };
 424
 425 static int __init register_pmu_driver(void)
 426 {
 427         return platform_driver_register(&cpu_pmu_driver);
 428 }
 429 device_initcall(register_pmu_driver);