Thomas Gleixner | d2912cb | 2019-06-04 10:11:33 +0200 | [diff] [blame] | 1 | // SPDX-License-Identifier: GPL-2.0-only |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 2 | /* |
| 3 | * ACPI probing code for ARM performance counters. |
| 4 | * |
| 5 | * Copyright (C) 2017 ARM Ltd. |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 6 | */ |
| 7 | |
| 8 | #include <linux/acpi.h> |
| 9 | #include <linux/cpumask.h> |
| 10 | #include <linux/init.h> |
Mark Rutland | 43fc9a2 | 2018-02-05 16:41:59 +0000 | [diff] [blame] | 11 | #include <linux/irq.h> |
| 12 | #include <linux/irqdesc.h> |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 13 | #include <linux/percpu.h> |
| 14 | #include <linux/perf/arm_pmu.h> |
| 15 | |
| 16 | #include <asm/cputype.h> |
| 17 | |
| 18 | static DEFINE_PER_CPU(struct arm_pmu *, probed_pmus); |
| 19 | static DEFINE_PER_CPU(int, pmu_irqs); |
| 20 | |
| 21 | static int arm_pmu_acpi_register_irq(int cpu) |
| 22 | { |
| 23 | struct acpi_madt_generic_interrupt *gicc; |
| 24 | int gsi, trigger; |
| 25 | |
| 26 | gicc = acpi_cpu_get_madt_gicc(cpu); |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 27 | |
| 28 | gsi = gicc->performance_interrupt; |
Wei Huang | 477c50e | 2017-05-30 11:56:22 +0100 | [diff] [blame] | 29 | |
| 30 | /* |
| 31 | * Per the ACPI spec, the MADT cannot describe a PMU that doesn't |
| 32 | * have an interrupt. QEMU advertises this by using a GSI of zero, |
| 33 | * which is not known to be valid on any hardware despite being |
| 34 | * valid per the spec. Take the pragmatic approach and reject a |
| 35 | * GSI of zero for now. |
| 36 | */ |
| 37 | if (!gsi) |
| 38 | return 0; |
| 39 | |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 40 | if (gicc->flags & ACPI_MADT_PERFORMANCE_IRQ_MODE) |
| 41 | trigger = ACPI_EDGE_SENSITIVE; |
| 42 | else |
| 43 | trigger = ACPI_LEVEL_SENSITIVE; |
| 44 | |
| 45 | /* |
| 46 | * Helpfully, the MADT GICC doesn't have a polarity flag for the |
| 47 | * "performance interrupt". Luckily, on compliant GICs the polarity is |
| 48 | * a fixed value in HW (for both SPIs and PPIs) that we cannot change |
| 49 | * from SW. |
| 50 | * |
| 51 | * Here we pass in ACPI_ACTIVE_HIGH to keep the core code happy. This |
| 52 | * may not match the real polarity, but that should not matter. |
| 53 | * |
| 54 | * Other interrupt controllers are not supported with ACPI. |
| 55 | */ |
| 56 | return acpi_register_gsi(NULL, gsi, trigger, ACPI_ACTIVE_HIGH); |
| 57 | } |
| 58 | |
| 59 | static void arm_pmu_acpi_unregister_irq(int cpu) |
| 60 | { |
| 61 | struct acpi_madt_generic_interrupt *gicc; |
| 62 | int gsi; |
| 63 | |
| 64 | gicc = acpi_cpu_get_madt_gicc(cpu); |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 65 | |
| 66 | gsi = gicc->performance_interrupt; |
luanshi | 3ba52ad | 2020-02-26 13:45:10 +0800 | [diff] [blame] | 67 | if (gsi) |
| 68 | acpi_unregister_gsi(gsi); |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 69 | } |
| 70 | |
Jeremy Linton | d24a0c7 | 2019-06-26 16:37:17 -0500 | [diff] [blame] | 71 | #if IS_ENABLED(CONFIG_ARM_SPE_PMU) |
| 72 | static struct resource spe_resources[] = { |
| 73 | { |
| 74 | /* irq */ |
| 75 | .flags = IORESOURCE_IRQ, |
| 76 | } |
| 77 | }; |
| 78 | |
| 79 | static struct platform_device spe_dev = { |
| 80 | .name = ARMV8_SPE_PDEV_NAME, |
| 81 | .id = -1, |
| 82 | .resource = spe_resources, |
| 83 | .num_resources = ARRAY_SIZE(spe_resources) |
| 84 | }; |
| 85 | |
| 86 | /* |
| 87 | * For lack of a better place, hook the normal PMU MADT walk |
| 88 | * and create a SPE device if we detect a recent MADT with |
| 89 | * a homogeneous PPI mapping. |
| 90 | */ |
| 91 | static void arm_spe_acpi_register_device(void) |
| 92 | { |
| 93 | int cpu, hetid, irq, ret; |
| 94 | bool first = true; |
| 95 | u16 gsi = 0; |
| 96 | |
| 97 | /* |
| 98 | * Sanity check all the GICC tables for the same interrupt number. |
| 99 | * For now, we only support homogeneous ACPI/SPE machines. |
| 100 | */ |
| 101 | for_each_possible_cpu(cpu) { |
| 102 | struct acpi_madt_generic_interrupt *gicc; |
| 103 | |
| 104 | gicc = acpi_cpu_get_madt_gicc(cpu); |
| 105 | if (gicc->header.length < ACPI_MADT_GICC_SPE) |
| 106 | return; |
| 107 | |
| 108 | if (first) { |
| 109 | gsi = gicc->spe_interrupt; |
| 110 | if (!gsi) |
| 111 | return; |
| 112 | hetid = find_acpi_cpu_topology_hetero_id(cpu); |
| 113 | first = false; |
| 114 | } else if ((gsi != gicc->spe_interrupt) || |
| 115 | (hetid != find_acpi_cpu_topology_hetero_id(cpu))) { |
| 116 | pr_warn("ACPI: SPE must be homogeneous\n"); |
| 117 | return; |
| 118 | } |
| 119 | } |
| 120 | |
| 121 | irq = acpi_register_gsi(NULL, gsi, ACPI_LEVEL_SENSITIVE, |
| 122 | ACPI_ACTIVE_HIGH); |
| 123 | if (irq < 0) { |
| 124 | pr_warn("ACPI: SPE Unable to register interrupt: %d\n", gsi); |
| 125 | return; |
| 126 | } |
| 127 | |
| 128 | spe_resources[0].start = irq; |
| 129 | ret = platform_device_register(&spe_dev); |
| 130 | if (ret < 0) { |
| 131 | pr_warn("ACPI: SPE: Unable to register device\n"); |
| 132 | acpi_unregister_gsi(gsi); |
| 133 | } |
| 134 | } |
| 135 | #else |
| 136 | static inline void arm_spe_acpi_register_device(void) |
| 137 | { |
| 138 | } |
| 139 | #endif /* CONFIG_ARM_SPE_PMU */ |
| 140 | |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 141 | static int arm_pmu_acpi_parse_irqs(void) |
| 142 | { |
| 143 | int irq, cpu, irq_cpu, err; |
| 144 | |
| 145 | for_each_possible_cpu(cpu) { |
| 146 | irq = arm_pmu_acpi_register_irq(cpu); |
| 147 | if (irq < 0) { |
| 148 | err = irq; |
| 149 | pr_warn("Unable to parse ACPI PMU IRQ for CPU%d: %d\n", |
| 150 | cpu, err); |
| 151 | goto out_err; |
| 152 | } else if (irq == 0) { |
| 153 | pr_warn("No ACPI PMU IRQ for CPU%d\n", cpu); |
| 154 | } |
| 155 | |
Mark Rutland | 167e614 | 2017-10-09 17:09:05 +0100 | [diff] [blame] | 156 | /* |
| 157 | * Log and request the IRQ so the core arm_pmu code can manage |
| 158 | * it. We'll have to sanity-check IRQs later when we associate |
| 159 | * them with their PMUs. |
| 160 | */ |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 161 | per_cpu(pmu_irqs, cpu) = irq; |
Mark Rutland | 167e614 | 2017-10-09 17:09:05 +0100 | [diff] [blame] | 162 | armpmu_request_irq(irq, cpu); |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 163 | } |
| 164 | |
| 165 | return 0; |
| 166 | |
| 167 | out_err: |
| 168 | for_each_possible_cpu(cpu) { |
| 169 | irq = per_cpu(pmu_irqs, cpu); |
| 170 | if (!irq) |
| 171 | continue; |
| 172 | |
| 173 | arm_pmu_acpi_unregister_irq(cpu); |
| 174 | |
| 175 | /* |
| 176 | * Blat all copies of the IRQ so that we only unregister the |
| 177 | * corresponding GSI once (e.g. when we have PPIs). |
| 178 | */ |
| 179 | for_each_possible_cpu(irq_cpu) { |
| 180 | if (per_cpu(pmu_irqs, irq_cpu) == irq) |
| 181 | per_cpu(pmu_irqs, irq_cpu) = 0; |
| 182 | } |
| 183 | } |
| 184 | |
| 185 | return err; |
| 186 | } |
| 187 | |
| 188 | static struct arm_pmu *arm_pmu_acpi_find_alloc_pmu(void) |
| 189 | { |
| 190 | unsigned long cpuid = read_cpuid_id(); |
| 191 | struct arm_pmu *pmu; |
| 192 | int cpu; |
| 193 | |
| 194 | for_each_possible_cpu(cpu) { |
| 195 | pmu = per_cpu(probed_pmus, cpu); |
| 196 | if (!pmu || pmu->acpi_cpuid != cpuid) |
| 197 | continue; |
| 198 | |
| 199 | return pmu; |
| 200 | } |
| 201 | |
Mark Rutland | 0dc1a18 | 2018-02-05 16:41:58 +0000 | [diff] [blame] | 202 | pmu = armpmu_alloc_atomic(); |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 203 | if (!pmu) { |
| 204 | pr_warn("Unable to allocate PMU for CPU%d\n", |
| 205 | smp_processor_id()); |
| 206 | return NULL; |
| 207 | } |
| 208 | |
| 209 | pmu->acpi_cpuid = cpuid; |
| 210 | |
| 211 | return pmu; |
| 212 | } |
| 213 | |
| 214 | /* |
Mark Rutland | 43fc9a2 | 2018-02-05 16:41:59 +0000 | [diff] [blame] | 215 | * Check whether the new IRQ is compatible with those already associated with |
| 216 | * the PMU (e.g. we don't have mismatched PPIs). |
| 217 | */ |
| 218 | static bool pmu_irq_matches(struct arm_pmu *pmu, int irq) |
| 219 | { |
| 220 | struct pmu_hw_events __percpu *hw_events = pmu->hw_events; |
| 221 | int cpu; |
| 222 | |
| 223 | if (!irq) |
| 224 | return true; |
| 225 | |
| 226 | for_each_cpu(cpu, &pmu->supported_cpus) { |
| 227 | int other_irq = per_cpu(hw_events->irq, cpu); |
| 228 | if (!other_irq) |
| 229 | continue; |
| 230 | |
| 231 | if (irq == other_irq) |
| 232 | continue; |
| 233 | if (!irq_is_percpu_devid(irq) && !irq_is_percpu_devid(other_irq)) |
| 234 | continue; |
| 235 | |
| 236 | pr_warn("mismatched PPIs detected\n"); |
| 237 | return false; |
| 238 | } |
| 239 | |
| 240 | return true; |
| 241 | } |
| 242 | |
| 243 | /* |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 244 | * This must run before the common arm_pmu hotplug logic, so that we can |
| 245 | * associate a CPU and its interrupt before the common code tries to manage the |
| 246 | * affinity and so on. |
| 247 | * |
| 248 | * Note that hotplug events are serialized, so we cannot race with another CPU |
| 249 | * coming up. The perf core won't open events while a hotplug event is in |
| 250 | * progress. |
| 251 | */ |
| 252 | static int arm_pmu_acpi_cpu_starting(unsigned int cpu) |
| 253 | { |
| 254 | struct arm_pmu *pmu; |
| 255 | struct pmu_hw_events __percpu *hw_events; |
| 256 | int irq; |
| 257 | |
| 258 | /* If we've already probed this CPU, we have nothing to do */ |
| 259 | if (per_cpu(probed_pmus, cpu)) |
| 260 | return 0; |
| 261 | |
| 262 | irq = per_cpu(pmu_irqs, cpu); |
| 263 | |
| 264 | pmu = arm_pmu_acpi_find_alloc_pmu(); |
| 265 | if (!pmu) |
| 266 | return -ENOMEM; |
| 267 | |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 268 | per_cpu(probed_pmus, cpu) = pmu; |
| 269 | |
Mark Rutland | 43fc9a2 | 2018-02-05 16:41:59 +0000 | [diff] [blame] | 270 | if (pmu_irq_matches(pmu, irq)) { |
| 271 | hw_events = pmu->hw_events; |
| 272 | per_cpu(hw_events->irq, cpu) = irq; |
| 273 | } |
| 274 | |
| 275 | cpumask_set_cpu(cpu, &pmu->supported_cpus); |
| 276 | |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 277 | /* |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 278 | * Ideally, we'd probe the PMU here when we find the first matching |
| 279 | * CPU. We can't do that for several reasons; see the comment in |
| 280 | * arm_pmu_acpi_init(). |
| 281 | * |
| 282 | * So for the time being, we're done. |
| 283 | */ |
| 284 | return 0; |
| 285 | } |
| 286 | |
| 287 | int arm_pmu_acpi_probe(armpmu_init_fn init_fn) |
| 288 | { |
| 289 | int pmu_idx = 0; |
| 290 | int cpu, ret; |
| 291 | |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 292 | /* |
| 293 | * Initialise and register the set of PMUs which we know about right |
| 294 | * now. Ideally we'd do this in arm_pmu_acpi_cpu_starting() so that we |
| 295 | * could handle late hotplug, but this may lead to deadlock since we |
| 296 | * might try to register a hotplug notifier instance from within a |
| 297 | * hotplug notifier. |
| 298 | * |
| 299 | * There's also the problem of having access to the right init_fn, |
| 300 | * without tying this too deeply into the "real" PMU driver. |
| 301 | * |
| 302 | * For the moment, as with the platform/DT case, we need at least one |
| 303 | * of a PMU's CPUs to be online at probe time. |
| 304 | */ |
| 305 | for_each_possible_cpu(cpu) { |
| 306 | struct arm_pmu *pmu = per_cpu(probed_pmus, cpu); |
| 307 | char *base_name; |
| 308 | |
| 309 | if (!pmu || pmu->name) |
| 310 | continue; |
| 311 | |
| 312 | ret = init_fn(pmu); |
| 313 | if (ret == -ENODEV) { |
| 314 | /* PMU not handled by this driver, or not present */ |
| 315 | continue; |
| 316 | } else if (ret) { |
| 317 | pr_warn("Unable to initialise PMU for CPU%d\n", cpu); |
| 318 | return ret; |
| 319 | } |
| 320 | |
| 321 | base_name = pmu->name; |
| 322 | pmu->name = kasprintf(GFP_KERNEL, "%s_%d", base_name, pmu_idx++); |
| 323 | if (!pmu->name) { |
| 324 | pr_warn("Unable to allocate PMU name for CPU%d\n", cpu); |
| 325 | return -ENOMEM; |
| 326 | } |
| 327 | |
| 328 | ret = armpmu_register(pmu); |
| 329 | if (ret) { |
| 330 | pr_warn("Failed to register PMU for CPU%d\n", cpu); |
Arvind Yadav | a88dc7b | 2017-09-20 12:26:38 +0530 | [diff] [blame] | 331 | kfree(pmu->name); |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 332 | return ret; |
| 333 | } |
| 334 | } |
| 335 | |
| 336 | return 0; |
| 337 | } |
| 338 | |
| 339 | static int arm_pmu_acpi_init(void) |
| 340 | { |
| 341 | int ret; |
| 342 | |
| 343 | if (acpi_disabled) |
| 344 | return 0; |
| 345 | |
Jeremy Linton | d24a0c7 | 2019-06-26 16:37:17 -0500 | [diff] [blame] | 346 | arm_spe_acpi_register_device(); |
| 347 | |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 348 | ret = arm_pmu_acpi_parse_irqs(); |
| 349 | if (ret) |
| 350 | return ret; |
| 351 | |
| 352 | ret = cpuhp_setup_state(CPUHP_AP_PERF_ARM_ACPI_STARTING, |
| 353 | "perf/arm/pmu_acpi:starting", |
| 354 | arm_pmu_acpi_cpu_starting, NULL); |
| 355 | |
| 356 | return ret; |
| 357 | } |
| 358 | subsys_initcall(arm_pmu_acpi_init) |