arm_pmu_acpi.c 6.67 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13
/*
 * ACPI probing code for ARM performance counters.
 *
 * Copyright (C) 2017 ARM Ltd.
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License version 2 as
 * published by the Free Software Foundation.
 */

#include <linux/acpi.h>
#include <linux/cpumask.h>
#include <linux/init.h>
14 15
#include <linux/irq.h>
#include <linux/irqdesc.h>
16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
#include <linux/percpu.h>
#include <linux/perf/arm_pmu.h>

#include <asm/cputype.h>

static DEFINE_PER_CPU(struct arm_pmu *, probed_pmus);
static DEFINE_PER_CPU(int, pmu_irqs);

static int arm_pmu_acpi_register_irq(int cpu)
{
	struct acpi_madt_generic_interrupt *gicc;
	int gsi, trigger;

	gicc = acpi_cpu_get_madt_gicc(cpu);
	if (WARN_ON(!gicc))
		return -EINVAL;

	gsi = gicc->performance_interrupt;
34 35 36 37 38 39 40 41 42 43 44

	/*
	 * Per the ACPI spec, the MADT cannot describe a PMU that doesn't
	 * have an interrupt. QEMU advertises this by using a GSI of zero,
	 * which is not known to be valid on any hardware despite being
	 * valid per the spec. Take the pragmatic approach and reject a
	 * GSI of zero for now.
	 */
	if (!gsi)
		return 0;

45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91
	if (gicc->flags & ACPI_MADT_PERFORMANCE_IRQ_MODE)
		trigger = ACPI_EDGE_SENSITIVE;
	else
		trigger = ACPI_LEVEL_SENSITIVE;

	/*
	 * Helpfully, the MADT GICC doesn't have a polarity flag for the
	 * "performance interrupt". Luckily, on compliant GICs the polarity is
	 * a fixed value in HW (for both SPIs and PPIs) that we cannot change
	 * from SW.
	 *
	 * Here we pass in ACPI_ACTIVE_HIGH to keep the core code happy. This
	 * may not match the real polarity, but that should not matter.
	 *
	 * Other interrupt controllers are not supported with ACPI.
	 */
	return acpi_register_gsi(NULL, gsi, trigger, ACPI_ACTIVE_HIGH);
}

static void arm_pmu_acpi_unregister_irq(int cpu)
{
	struct acpi_madt_generic_interrupt *gicc;
	int gsi;

	gicc = acpi_cpu_get_madt_gicc(cpu);
	if (!gicc)
		return;

	gsi = gicc->performance_interrupt;
	acpi_unregister_gsi(gsi);
}

static int arm_pmu_acpi_parse_irqs(void)
{
	int irq, cpu, irq_cpu, err;

	for_each_possible_cpu(cpu) {
		irq = arm_pmu_acpi_register_irq(cpu);
		if (irq < 0) {
			err = irq;
			pr_warn("Unable to parse ACPI PMU IRQ for CPU%d: %d\n",
				cpu, err);
			goto out_err;
		} else if (irq == 0) {
			pr_warn("No ACPI PMU IRQ for CPU%d\n", cpu);
		}

92 93 94 95 96
		/*
		 * Log and request the IRQ so the core arm_pmu code can manage
		 * it. We'll have to sanity-check IRQs later when we associate
		 * them with their PMUs.
		 */
97
		per_cpu(pmu_irqs, cpu) = irq;
98
		armpmu_request_irq(irq, cpu);
99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137
	}

	return 0;

out_err:
	for_each_possible_cpu(cpu) {
		irq = per_cpu(pmu_irqs, cpu);
		if (!irq)
			continue;

		arm_pmu_acpi_unregister_irq(cpu);

		/*
		 * Blat all copies of the IRQ so that we only unregister the
		 * corresponding GSI once (e.g. when we have PPIs).
		 */
		for_each_possible_cpu(irq_cpu) {
			if (per_cpu(pmu_irqs, irq_cpu) == irq)
				per_cpu(pmu_irqs, irq_cpu) = 0;
		}
	}

	return err;
}

static struct arm_pmu *arm_pmu_acpi_find_alloc_pmu(void)
{
	unsigned long cpuid = read_cpuid_id();
	struct arm_pmu *pmu;
	int cpu;

	for_each_possible_cpu(cpu) {
		pmu = per_cpu(probed_pmus, cpu);
		if (!pmu || pmu->acpi_cpuid != cpuid)
			continue;

		return pmu;
	}

138
	pmu = armpmu_alloc_atomic();
139 140 141 142 143 144 145 146 147 148 149
	if (!pmu) {
		pr_warn("Unable to allocate PMU for CPU%d\n",
			smp_processor_id());
		return NULL;
	}

	pmu->acpi_cpuid = cpuid;

	return pmu;
}

150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178
/*
 * Check whether the new IRQ is compatible with those already associated with
 * the PMU (e.g. we don't have mismatched PPIs).
 */
static bool pmu_irq_matches(struct arm_pmu *pmu, int irq)
{
	struct pmu_hw_events __percpu *hw_events = pmu->hw_events;
	int cpu;

	if (!irq)
		return true;

	for_each_cpu(cpu, &pmu->supported_cpus) {
		int other_irq = per_cpu(hw_events->irq, cpu);
		if (!other_irq)
			continue;

		if (irq == other_irq)
			continue;
		if (!irq_is_percpu_devid(irq) && !irq_is_percpu_devid(other_irq))
			continue;

		pr_warn("mismatched PPIs detected\n");
		return false;
	}

	return true;
}

179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205
/*
 * This must run before the common arm_pmu hotplug logic, so that we can
 * associate a CPU and its interrupt before the common code tries to manage the
 * affinity and so on.
 *
 * Note that hotplug events are serialized, so we cannot race with another CPU
 * coming up. The perf core won't open events while a hotplug event is in
 * progress.
 */
static int arm_pmu_acpi_cpu_starting(unsigned int cpu)
{
	struct arm_pmu *pmu;
	struct pmu_hw_events __percpu *hw_events;
	int irq;

	/* If we've already probed this CPU, we have nothing to do */
	if (per_cpu(probed_pmus, cpu))
		return 0;

	irq = per_cpu(pmu_irqs, cpu);

	pmu = arm_pmu_acpi_find_alloc_pmu();
	if (!pmu)
		return -ENOMEM;

	per_cpu(probed_pmus, cpu) = pmu;

206 207 208 209 210 211 212
	if (pmu_irq_matches(pmu, irq)) {
		hw_events = pmu->hw_events;
		per_cpu(hw_events->irq, cpu) = irq;
	}

	cpumask_set_cpu(cpu, &pmu->supported_cpus);

213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266
	/*
	 * Ideally, we'd probe the PMU here when we find the first matching
	 * CPU. We can't do that for several reasons; see the comment in
	 * arm_pmu_acpi_init().
	 *
	 * So for the time being, we're done.
	 */
	return 0;
}

int arm_pmu_acpi_probe(armpmu_init_fn init_fn)
{
	int pmu_idx = 0;
	int cpu, ret;

	/*
	 * Initialise and register the set of PMUs which we know about right
	 * now. Ideally we'd do this in arm_pmu_acpi_cpu_starting() so that we
	 * could handle late hotplug, but this may lead to deadlock since we
	 * might try to register a hotplug notifier instance from within a
	 * hotplug notifier.
	 *
	 * There's also the problem of having access to the right init_fn,
	 * without tying this too deeply into the "real" PMU driver.
	 *
	 * For the moment, as with the platform/DT case, we need at least one
	 * of a PMU's CPUs to be online at probe time.
	 */
	for_each_possible_cpu(cpu) {
		struct arm_pmu *pmu = per_cpu(probed_pmus, cpu);
		char *base_name;

		if (!pmu || pmu->name)
			continue;

		ret = init_fn(pmu);
		if (ret == -ENODEV) {
			/* PMU not handled by this driver, or not present */
			continue;
		} else if (ret) {
			pr_warn("Unable to initialise PMU for CPU%d\n", cpu);
			return ret;
		}

		base_name = pmu->name;
		pmu->name = kasprintf(GFP_KERNEL, "%s_%d", base_name, pmu_idx++);
		if (!pmu->name) {
			pr_warn("Unable to allocate PMU name for CPU%d\n", cpu);
			return -ENOMEM;
		}

		ret = armpmu_register(pmu);
		if (ret) {
			pr_warn("Failed to register PMU for CPU%d\n", cpu);
267
			kfree(pmu->name);
268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292
			return ret;
		}
	}

	return 0;
}

static int arm_pmu_acpi_init(void)
{
	int ret;

	if (acpi_disabled)
		return 0;

	ret = arm_pmu_acpi_parse_irqs();
	if (ret)
		return ret;

	ret = cpuhp_setup_state(CPUHP_AP_PERF_ARM_ACPI_STARTING,
				"perf/arm/pmu_acpi:starting",
				arm_pmu_acpi_cpu_starting, NULL);

	return ret;
}
subsys_initcall(arm_pmu_acpi_init)