xref: /openbmc/linux/drivers/perf/arm_pmu_acpi.c (revision be709d48)
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/irq.h>
15 #include <linux/irqdesc.h>
16 #include <linux/percpu.h>
17 #include <linux/perf/arm_pmu.h>
18 
19 #include <asm/cputype.h>
20 
21 static DEFINE_PER_CPU(struct arm_pmu *, probed_pmus);
22 static DEFINE_PER_CPU(int, pmu_irqs);
23 
24 static int arm_pmu_acpi_register_irq(int cpu)
25 {
26 	struct acpi_madt_generic_interrupt *gicc;
27 	int gsi, trigger;
28 
29 	gicc = acpi_cpu_get_madt_gicc(cpu);
30 	if (WARN_ON(!gicc))
31 		return -EINVAL;
32 
33 	gsi = gicc->performance_interrupt;
34 
35 	/*
36 	 * Per the ACPI spec, the MADT cannot describe a PMU that doesn't
37 	 * have an interrupt. QEMU advertises this by using a GSI of zero,
38 	 * which is not known to be valid on any hardware despite being
39 	 * valid per the spec. Take the pragmatic approach and reject a
40 	 * GSI of zero for now.
41 	 */
42 	if (!gsi)
43 		return 0;
44 
45 	if (gicc->flags & ACPI_MADT_PERFORMANCE_IRQ_MODE)
46 		trigger = ACPI_EDGE_SENSITIVE;
47 	else
48 		trigger = ACPI_LEVEL_SENSITIVE;
49 
50 	/*
51 	 * Helpfully, the MADT GICC doesn't have a polarity flag for the
52 	 * "performance interrupt". Luckily, on compliant GICs the polarity is
53 	 * a fixed value in HW (for both SPIs and PPIs) that we cannot change
54 	 * from SW.
55 	 *
56 	 * Here we pass in ACPI_ACTIVE_HIGH to keep the core code happy. This
57 	 * may not match the real polarity, but that should not matter.
58 	 *
59 	 * Other interrupt controllers are not supported with ACPI.
60 	 */
61 	return acpi_register_gsi(NULL, gsi, trigger, ACPI_ACTIVE_HIGH);
62 }
63 
64 static void arm_pmu_acpi_unregister_irq(int cpu)
65 {
66 	struct acpi_madt_generic_interrupt *gicc;
67 	int gsi;
68 
69 	gicc = acpi_cpu_get_madt_gicc(cpu);
70 	if (!gicc)
71 		return;
72 
73 	gsi = gicc->performance_interrupt;
74 	acpi_unregister_gsi(gsi);
75 }
76 
77 static int arm_pmu_acpi_parse_irqs(void)
78 {
79 	int irq, cpu, irq_cpu, err;
80 
81 	for_each_possible_cpu(cpu) {
82 		irq = arm_pmu_acpi_register_irq(cpu);
83 		if (irq < 0) {
84 			err = irq;
85 			pr_warn("Unable to parse ACPI PMU IRQ for CPU%d: %d\n",
86 				cpu, err);
87 			goto out_err;
88 		} else if (irq == 0) {
89 			pr_warn("No ACPI PMU IRQ for CPU%d\n", cpu);
90 		}
91 
92 		/*
93 		 * Log and request the IRQ so the core arm_pmu code can manage
94 		 * it. We'll have to sanity-check IRQs later when we associate
95 		 * them with their PMUs.
96 		 */
97 		per_cpu(pmu_irqs, cpu) = irq;
98 		armpmu_request_irq(irq, cpu);
99 	}
100 
101 	return 0;
102 
103 out_err:
104 	for_each_possible_cpu(cpu) {
105 		irq = per_cpu(pmu_irqs, cpu);
106 		if (!irq)
107 			continue;
108 
109 		arm_pmu_acpi_unregister_irq(cpu);
110 
111 		/*
112 		 * Blat all copies of the IRQ so that we only unregister the
113 		 * corresponding GSI once (e.g. when we have PPIs).
114 		 */
115 		for_each_possible_cpu(irq_cpu) {
116 			if (per_cpu(pmu_irqs, irq_cpu) == irq)
117 				per_cpu(pmu_irqs, irq_cpu) = 0;
118 		}
119 	}
120 
121 	return err;
122 }
123 
124 static struct arm_pmu *arm_pmu_acpi_find_alloc_pmu(void)
125 {
126 	unsigned long cpuid = read_cpuid_id();
127 	struct arm_pmu *pmu;
128 	int cpu;
129 
130 	for_each_possible_cpu(cpu) {
131 		pmu = per_cpu(probed_pmus, cpu);
132 		if (!pmu || pmu->acpi_cpuid != cpuid)
133 			continue;
134 
135 		return pmu;
136 	}
137 
138 	pmu = armpmu_alloc_atomic();
139 	if (!pmu) {
140 		pr_warn("Unable to allocate PMU for CPU%d\n",
141 			smp_processor_id());
142 		return NULL;
143 	}
144 
145 	pmu->acpi_cpuid = cpuid;
146 
147 	return pmu;
148 }
149 
150 /*
151  * Check whether the new IRQ is compatible with those already associated with
152  * the PMU (e.g. we don't have mismatched PPIs).
153  */
154 static bool pmu_irq_matches(struct arm_pmu *pmu, int irq)
155 {
156 	struct pmu_hw_events __percpu *hw_events = pmu->hw_events;
157 	int cpu;
158 
159 	if (!irq)
160 		return true;
161 
162 	for_each_cpu(cpu, &pmu->supported_cpus) {
163 		int other_irq = per_cpu(hw_events->irq, cpu);
164 		if (!other_irq)
165 			continue;
166 
167 		if (irq == other_irq)
168 			continue;
169 		if (!irq_is_percpu_devid(irq) && !irq_is_percpu_devid(other_irq))
170 			continue;
171 
172 		pr_warn("mismatched PPIs detected\n");
173 		return false;
174 	}
175 
176 	return true;
177 }
178 
179 /*
180  * This must run before the common arm_pmu hotplug logic, so that we can
181  * associate a CPU and its interrupt before the common code tries to manage the
182  * affinity and so on.
183  *
184  * Note that hotplug events are serialized, so we cannot race with another CPU
185  * coming up. The perf core won't open events while a hotplug event is in
186  * progress.
187  */
188 static int arm_pmu_acpi_cpu_starting(unsigned int cpu)
189 {
190 	struct arm_pmu *pmu;
191 	struct pmu_hw_events __percpu *hw_events;
192 	int irq;
193 
194 	/* If we've already probed this CPU, we have nothing to do */
195 	if (per_cpu(probed_pmus, cpu))
196 		return 0;
197 
198 	irq = per_cpu(pmu_irqs, cpu);
199 
200 	pmu = arm_pmu_acpi_find_alloc_pmu();
201 	if (!pmu)
202 		return -ENOMEM;
203 
204 	per_cpu(probed_pmus, cpu) = pmu;
205 
206 	if (pmu_irq_matches(pmu, irq)) {
207 		hw_events = pmu->hw_events;
208 		per_cpu(hw_events->irq, cpu) = irq;
209 	}
210 
211 	cpumask_set_cpu(cpu, &pmu->supported_cpus);
212 
213 	/*
214 	 * Ideally, we'd probe the PMU here when we find the first matching
215 	 * CPU. We can't do that for several reasons; see the comment in
216 	 * arm_pmu_acpi_init().
217 	 *
218 	 * So for the time being, we're done.
219 	 */
220 	return 0;
221 }
222 
223 int arm_pmu_acpi_probe(armpmu_init_fn init_fn)
224 {
225 	int pmu_idx = 0;
226 	int cpu, ret;
227 
228 	/*
229 	 * Initialise and register the set of PMUs which we know about right
230 	 * now. Ideally we'd do this in arm_pmu_acpi_cpu_starting() so that we
231 	 * could handle late hotplug, but this may lead to deadlock since we
232 	 * might try to register a hotplug notifier instance from within a
233 	 * hotplug notifier.
234 	 *
235 	 * There's also the problem of having access to the right init_fn,
236 	 * without tying this too deeply into the "real" PMU driver.
237 	 *
238 	 * For the moment, as with the platform/DT case, we need at least one
239 	 * of a PMU's CPUs to be online at probe time.
240 	 */
241 	for_each_possible_cpu(cpu) {
242 		struct arm_pmu *pmu = per_cpu(probed_pmus, cpu);
243 		char *base_name;
244 
245 		if (!pmu || pmu->name)
246 			continue;
247 
248 		ret = init_fn(pmu);
249 		if (ret == -ENODEV) {
250 			/* PMU not handled by this driver, or not present */
251 			continue;
252 		} else if (ret) {
253 			pr_warn("Unable to initialise PMU for CPU%d\n", cpu);
254 			return ret;
255 		}
256 
257 		base_name = pmu->name;
258 		pmu->name = kasprintf(GFP_KERNEL, "%s_%d", base_name, pmu_idx++);
259 		if (!pmu->name) {
260 			pr_warn("Unable to allocate PMU name for CPU%d\n", cpu);
261 			return -ENOMEM;
262 		}
263 
264 		ret = armpmu_register(pmu);
265 		if (ret) {
266 			pr_warn("Failed to register PMU for CPU%d\n", cpu);
267 			kfree(pmu->name);
268 			return ret;
269 		}
270 	}
271 
272 	return 0;
273 }
274 
275 static int arm_pmu_acpi_init(void)
276 {
277 	int ret;
278 
279 	if (acpi_disabled)
280 		return 0;
281 
282 	ret = arm_pmu_acpi_parse_irqs();
283 	if (ret)
284 		return ret;
285 
286 	ret = cpuhp_setup_state(CPUHP_AP_PERF_ARM_ACPI_STARTING,
287 				"perf/arm/pmu_acpi:starting",
288 				arm_pmu_acpi_cpu_starting, NULL);
289 
290 	return ret;
291 }
292 subsys_initcall(arm_pmu_acpi_init)
293