Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 1 | /* |
| 2 | * ACPI probing code for ARM performance counters. |
| 3 | * |
| 4 | * Copyright (C) 2017 ARM Ltd. |
| 5 | * |
| 6 | * This program is free software; you can redistribute it and/or modify |
| 7 | * it under the terms of the GNU General Public License version 2 as |
| 8 | * published by the Free Software Foundation. |
| 9 | */ |
| 10 | |
| 11 | #include <linux/acpi.h> |
| 12 | #include <linux/cpumask.h> |
| 13 | #include <linux/init.h> |
| 14 | #include <linux/percpu.h> |
| 15 | #include <linux/perf/arm_pmu.h> |
| 16 | |
| 17 | #include <asm/cputype.h> |
| 18 | |
| 19 | static DEFINE_PER_CPU(struct arm_pmu *, probed_pmus); |
| 20 | static DEFINE_PER_CPU(int, pmu_irqs); |
| 21 | |
| 22 | static int arm_pmu_acpi_register_irq(int cpu) |
| 23 | { |
| 24 | struct acpi_madt_generic_interrupt *gicc; |
| 25 | int gsi, trigger; |
| 26 | |
| 27 | gicc = acpi_cpu_get_madt_gicc(cpu); |
| 28 | if (WARN_ON(!gicc)) |
| 29 | return -EINVAL; |
| 30 | |
| 31 | gsi = gicc->performance_interrupt; |
Wei Huang | 477c50e | 2017-05-30 11:56:22 +0100 | [diff] [blame] | 32 | |
| 33 | /* |
| 34 | * Per the ACPI spec, the MADT cannot describe a PMU that doesn't |
| 35 | * have an interrupt. QEMU advertises this by using a GSI of zero, |
| 36 | * which is not known to be valid on any hardware despite being |
| 37 | * valid per the spec. Take the pragmatic approach and reject a |
| 38 | * GSI of zero for now. |
| 39 | */ |
| 40 | if (!gsi) |
| 41 | return 0; |
| 42 | |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 43 | if (gicc->flags & ACPI_MADT_PERFORMANCE_IRQ_MODE) |
| 44 | trigger = ACPI_EDGE_SENSITIVE; |
| 45 | else |
| 46 | trigger = ACPI_LEVEL_SENSITIVE; |
| 47 | |
| 48 | /* |
| 49 | * Helpfully, the MADT GICC doesn't have a polarity flag for the |
| 50 | * "performance interrupt". Luckily, on compliant GICs the polarity is |
| 51 | * a fixed value in HW (for both SPIs and PPIs) that we cannot change |
| 52 | * from SW. |
| 53 | * |
| 54 | * Here we pass in ACPI_ACTIVE_HIGH to keep the core code happy. This |
| 55 | * may not match the real polarity, but that should not matter. |
| 56 | * |
| 57 | * Other interrupt controllers are not supported with ACPI. |
| 58 | */ |
| 59 | return acpi_register_gsi(NULL, gsi, trigger, ACPI_ACTIVE_HIGH); |
| 60 | } |
| 61 | |
| 62 | static void arm_pmu_acpi_unregister_irq(int cpu) |
| 63 | { |
| 64 | struct acpi_madt_generic_interrupt *gicc; |
| 65 | int gsi; |
| 66 | |
| 67 | gicc = acpi_cpu_get_madt_gicc(cpu); |
| 68 | if (!gicc) |
| 69 | return; |
| 70 | |
| 71 | gsi = gicc->performance_interrupt; |
| 72 | acpi_unregister_gsi(gsi); |
| 73 | } |
| 74 | |
| 75 | static int arm_pmu_acpi_parse_irqs(void) |
| 76 | { |
| 77 | int irq, cpu, irq_cpu, err; |
| 78 | |
| 79 | for_each_possible_cpu(cpu) { |
| 80 | irq = arm_pmu_acpi_register_irq(cpu); |
| 81 | if (irq < 0) { |
| 82 | err = irq; |
| 83 | pr_warn("Unable to parse ACPI PMU IRQ for CPU%d: %d\n", |
| 84 | cpu, err); |
| 85 | goto out_err; |
| 86 | } else if (irq == 0) { |
| 87 | pr_warn("No ACPI PMU IRQ for CPU%d\n", cpu); |
| 88 | } |
| 89 | |
| 90 | per_cpu(pmu_irqs, cpu) = irq; |
| 91 | } |
| 92 | |
| 93 | return 0; |
| 94 | |
| 95 | out_err: |
| 96 | for_each_possible_cpu(cpu) { |
| 97 | irq = per_cpu(pmu_irqs, cpu); |
| 98 | if (!irq) |
| 99 | continue; |
| 100 | |
| 101 | arm_pmu_acpi_unregister_irq(cpu); |
| 102 | |
| 103 | /* |
| 104 | * Blat all copies of the IRQ so that we only unregister the |
| 105 | * corresponding GSI once (e.g. when we have PPIs). |
| 106 | */ |
| 107 | for_each_possible_cpu(irq_cpu) { |
| 108 | if (per_cpu(pmu_irqs, irq_cpu) == irq) |
| 109 | per_cpu(pmu_irqs, irq_cpu) = 0; |
| 110 | } |
| 111 | } |
| 112 | |
| 113 | return err; |
| 114 | } |
| 115 | |
| 116 | static struct arm_pmu *arm_pmu_acpi_find_alloc_pmu(void) |
| 117 | { |
| 118 | unsigned long cpuid = read_cpuid_id(); |
| 119 | struct arm_pmu *pmu; |
| 120 | int cpu; |
| 121 | |
| 122 | for_each_possible_cpu(cpu) { |
| 123 | pmu = per_cpu(probed_pmus, cpu); |
| 124 | if (!pmu || pmu->acpi_cpuid != cpuid) |
| 125 | continue; |
| 126 | |
| 127 | return pmu; |
| 128 | } |
| 129 | |
| 130 | pmu = armpmu_alloc(); |
| 131 | if (!pmu) { |
| 132 | pr_warn("Unable to allocate PMU for CPU%d\n", |
| 133 | smp_processor_id()); |
| 134 | return NULL; |
| 135 | } |
| 136 | |
| 137 | pmu->acpi_cpuid = cpuid; |
| 138 | |
| 139 | return pmu; |
| 140 | } |
| 141 | |
| 142 | /* |
| 143 | * This must run before the common arm_pmu hotplug logic, so that we can |
| 144 | * associate a CPU and its interrupt before the common code tries to manage the |
| 145 | * affinity and so on. |
| 146 | * |
| 147 | * Note that hotplug events are serialized, so we cannot race with another CPU |
| 148 | * coming up. The perf core won't open events while a hotplug event is in |
| 149 | * progress. |
| 150 | */ |
| 151 | static int arm_pmu_acpi_cpu_starting(unsigned int cpu) |
| 152 | { |
| 153 | struct arm_pmu *pmu; |
| 154 | struct pmu_hw_events __percpu *hw_events; |
| 155 | int irq; |
| 156 | |
| 157 | /* If we've already probed this CPU, we have nothing to do */ |
| 158 | if (per_cpu(probed_pmus, cpu)) |
| 159 | return 0; |
| 160 | |
| 161 | irq = per_cpu(pmu_irqs, cpu); |
| 162 | |
| 163 | pmu = arm_pmu_acpi_find_alloc_pmu(); |
| 164 | if (!pmu) |
| 165 | return -ENOMEM; |
| 166 | |
| 167 | cpumask_set_cpu(cpu, &pmu->supported_cpus); |
| 168 | |
| 169 | per_cpu(probed_pmus, cpu) = pmu; |
| 170 | |
| 171 | /* |
| 172 | * Log and request the IRQ so the core arm_pmu code can manage it. In |
| 173 | * some situations (e.g. mismatched PPIs), we may fail to request the |
| 174 | * IRQ. However, it may be too late for us to do anything about it. |
| 175 | * The common ARM PMU code will log a warning in this case. |
| 176 | */ |
| 177 | hw_events = pmu->hw_events; |
| 178 | per_cpu(hw_events->irq, cpu) = irq; |
| 179 | armpmu_request_irq(pmu, cpu); |
| 180 | |
| 181 | /* |
| 182 | * Ideally, we'd probe the PMU here when we find the first matching |
| 183 | * CPU. We can't do that for several reasons; see the comment in |
| 184 | * arm_pmu_acpi_init(). |
| 185 | * |
| 186 | * So for the time being, we're done. |
| 187 | */ |
| 188 | return 0; |
| 189 | } |
| 190 | |
| 191 | int arm_pmu_acpi_probe(armpmu_init_fn init_fn) |
| 192 | { |
| 193 | int pmu_idx = 0; |
| 194 | int cpu, ret; |
| 195 | |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 196 | /* |
| 197 | * Initialise and register the set of PMUs which we know about right |
| 198 | * now. Ideally we'd do this in arm_pmu_acpi_cpu_starting() so that we |
| 199 | * could handle late hotplug, but this may lead to deadlock since we |
| 200 | * might try to register a hotplug notifier instance from within a |
| 201 | * hotplug notifier. |
| 202 | * |
| 203 | * There's also the problem of having access to the right init_fn, |
| 204 | * without tying this too deeply into the "real" PMU driver. |
| 205 | * |
| 206 | * For the moment, as with the platform/DT case, we need at least one |
| 207 | * of a PMU's CPUs to be online at probe time. |
| 208 | */ |
| 209 | for_each_possible_cpu(cpu) { |
| 210 | struct arm_pmu *pmu = per_cpu(probed_pmus, cpu); |
| 211 | char *base_name; |
| 212 | |
| 213 | if (!pmu || pmu->name) |
| 214 | continue; |
| 215 | |
| 216 | ret = init_fn(pmu); |
| 217 | if (ret == -ENODEV) { |
| 218 | /* PMU not handled by this driver, or not present */ |
| 219 | continue; |
| 220 | } else if (ret) { |
| 221 | pr_warn("Unable to initialise PMU for CPU%d\n", cpu); |
| 222 | return ret; |
| 223 | } |
| 224 | |
| 225 | base_name = pmu->name; |
| 226 | pmu->name = kasprintf(GFP_KERNEL, "%s_%d", base_name, pmu_idx++); |
| 227 | if (!pmu->name) { |
| 228 | pr_warn("Unable to allocate PMU name for CPU%d\n", cpu); |
| 229 | return -ENOMEM; |
| 230 | } |
| 231 | |
| 232 | ret = armpmu_register(pmu); |
| 233 | if (ret) { |
| 234 | pr_warn("Failed to register PMU for CPU%d\n", cpu); |
Arvind Yadav | a88dc7b | 2017-09-20 12:26:38 +0530 | [diff] [blame] | 235 | kfree(pmu->name); |
Mark Rutland | 45736a7 | 2017-04-11 09:39:55 +0100 | [diff] [blame] | 236 | return ret; |
| 237 | } |
| 238 | } |
| 239 | |
| 240 | return 0; |
| 241 | } |
| 242 | |
| 243 | static int arm_pmu_acpi_init(void) |
| 244 | { |
| 245 | int ret; |
| 246 | |
| 247 | if (acpi_disabled) |
| 248 | return 0; |
| 249 | |
| 250 | /* |
| 251 | * We can't request IRQs yet, since we don't know the cookie value |
| 252 | * until we know which CPUs share the same logical PMU. We'll handle |
| 253 | * that in arm_pmu_acpi_cpu_starting(). |
| 254 | */ |
| 255 | ret = arm_pmu_acpi_parse_irqs(); |
| 256 | if (ret) |
| 257 | return ret; |
| 258 | |
| 259 | ret = cpuhp_setup_state(CPUHP_AP_PERF_ARM_ACPI_STARTING, |
| 260 | "perf/arm/pmu_acpi:starting", |
| 261 | arm_pmu_acpi_cpu_starting, NULL); |
| 262 | |
| 263 | return ret; |
| 264 | } |
| 265 | subsys_initcall(arm_pmu_acpi_init) |