xref: /linux/drivers/perf/arm_pmu_acpi.c (revision 67f49869106f78882a8a09b736d4884be85aba18)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * ACPI probing code for ARM performance counters.
4  *
5  * Copyright (C) 2017 ARM Ltd.
6  */
7 
8 #include <linux/acpi.h>
9 #include <linux/cpumask.h>
10 #include <linux/init.h>
11 #include <linux/irq.h>
12 #include <linux/irqdesc.h>
13 #include <linux/percpu.h>
14 #include <linux/perf/arm_pmu.h>
15 
16 #include <asm/cpu.h>
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 
29 	gsi = gicc->performance_interrupt;
30 
31 	/*
32 	 * Per the ACPI spec, the MADT cannot describe a PMU that doesn't
33 	 * have an interrupt. QEMU advertises this by using a GSI of zero,
34 	 * which is not known to be valid on any hardware despite being
35 	 * valid per the spec. Take the pragmatic approach and reject a
36 	 * GSI of zero for now.
37 	 */
38 	if (!gsi)
39 		return 0;
40 
41 	if (gicc->flags & ACPI_MADT_PERFORMANCE_IRQ_MODE)
42 		trigger = ACPI_EDGE_SENSITIVE;
43 	else
44 		trigger = ACPI_LEVEL_SENSITIVE;
45 
46 	/*
47 	 * Helpfully, the MADT GICC doesn't have a polarity flag for the
48 	 * "performance interrupt". Luckily, on compliant GICs the polarity is
49 	 * a fixed value in HW (for both SPIs and PPIs) that we cannot change
50 	 * from SW.
51 	 *
52 	 * Here we pass in ACPI_ACTIVE_HIGH to keep the core code happy. This
53 	 * may not match the real polarity, but that should not matter.
54 	 *
55 	 * Other interrupt controllers are not supported with ACPI.
56 	 */
57 	return acpi_register_gsi(NULL, gsi, trigger, ACPI_ACTIVE_HIGH);
58 }
59 
60 static void arm_pmu_acpi_unregister_irq(int cpu)
61 {
62 	struct acpi_madt_generic_interrupt *gicc;
63 	int gsi;
64 
65 	gicc = acpi_cpu_get_madt_gicc(cpu);
66 
67 	gsi = gicc->performance_interrupt;
68 	if (gsi)
69 		acpi_unregister_gsi(gsi);
70 }
71 
72 #if IS_ENABLED(CONFIG_ARM_SPE_PMU)
73 static struct resource spe_resources[] = {
74 	{
75 		/* irq */
76 		.flags          = IORESOURCE_IRQ,
77 	}
78 };
79 
80 static struct platform_device spe_dev = {
81 	.name = ARMV8_SPE_PDEV_NAME,
82 	.id = -1,
83 	.resource = spe_resources,
84 	.num_resources = ARRAY_SIZE(spe_resources)
85 };
86 
87 /*
88  * For lack of a better place, hook the normal PMU MADT walk
89  * and create a SPE device if we detect a recent MADT with
90  * a homogeneous PPI mapping.
91  */
92 static void arm_spe_acpi_register_device(void)
93 {
94 	int cpu, hetid, irq, ret;
95 	bool first = true;
96 	u16 gsi = 0;
97 
98 	/*
99 	 * Sanity check all the GICC tables for the same interrupt number.
100 	 * For now, we only support homogeneous ACPI/SPE machines.
101 	 */
102 	for_each_possible_cpu(cpu) {
103 		struct acpi_madt_generic_interrupt *gicc;
104 
105 		gicc = acpi_cpu_get_madt_gicc(cpu);
106 		if (gicc->header.length < ACPI_MADT_GICC_SPE)
107 			return;
108 
109 		if (first) {
110 			gsi = gicc->spe_interrupt;
111 			if (!gsi)
112 				return;
113 			hetid = find_acpi_cpu_topology_hetero_id(cpu);
114 			first = false;
115 		} else if ((gsi != gicc->spe_interrupt) ||
116 			   (hetid != find_acpi_cpu_topology_hetero_id(cpu))) {
117 			pr_warn("ACPI: SPE must be homogeneous\n");
118 			return;
119 		}
120 	}
121 
122 	irq = acpi_register_gsi(NULL, gsi, ACPI_LEVEL_SENSITIVE,
123 				ACPI_ACTIVE_HIGH);
124 	if (irq < 0) {
125 		pr_warn("ACPI: SPE Unable to register interrupt: %d\n", gsi);
126 		return;
127 	}
128 
129 	spe_resources[0].start = irq;
130 	ret = platform_device_register(&spe_dev);
131 	if (ret < 0) {
132 		pr_warn("ACPI: SPE: Unable to register device\n");
133 		acpi_unregister_gsi(gsi);
134 	}
135 }
136 #else
137 static inline void arm_spe_acpi_register_device(void)
138 {
139 }
140 #endif /* CONFIG_ARM_SPE_PMU */
141 
142 static int arm_pmu_acpi_parse_irqs(void)
143 {
144 	int irq, cpu, irq_cpu, err;
145 
146 	for_each_possible_cpu(cpu) {
147 		irq = arm_pmu_acpi_register_irq(cpu);
148 		if (irq < 0) {
149 			err = irq;
150 			pr_warn("Unable to parse ACPI PMU IRQ for CPU%d: %d\n",
151 				cpu, err);
152 			goto out_err;
153 		} else if (irq == 0) {
154 			pr_warn("No ACPI PMU IRQ for CPU%d\n", cpu);
155 		}
156 
157 		/*
158 		 * Log and request the IRQ so the core arm_pmu code can manage
159 		 * it. We'll have to sanity-check IRQs later when we associate
160 		 * them with their PMUs.
161 		 */
162 		per_cpu(pmu_irqs, cpu) = irq;
163 		err = armpmu_request_irq(irq, cpu);
164 		if (err)
165 			goto out_err;
166 	}
167 
168 	return 0;
169 
170 out_err:
171 	for_each_possible_cpu(cpu) {
172 		irq = per_cpu(pmu_irqs, cpu);
173 		if (!irq)
174 			continue;
175 
176 		arm_pmu_acpi_unregister_irq(cpu);
177 
178 		/*
179 		 * Blat all copies of the IRQ so that we only unregister the
180 		 * corresponding GSI once (e.g. when we have PPIs).
181 		 */
182 		for_each_possible_cpu(irq_cpu) {
183 			if (per_cpu(pmu_irqs, irq_cpu) == irq)
184 				per_cpu(pmu_irqs, irq_cpu) = 0;
185 		}
186 	}
187 
188 	return err;
189 }
190 
191 static struct arm_pmu *arm_pmu_acpi_find_pmu(void)
192 {
193 	unsigned long cpuid = read_cpuid_id();
194 	struct arm_pmu *pmu;
195 	int cpu;
196 
197 	for_each_possible_cpu(cpu) {
198 		pmu = per_cpu(probed_pmus, cpu);
199 		if (!pmu || pmu->acpi_cpuid != cpuid)
200 			continue;
201 
202 		return pmu;
203 	}
204 
205 	return NULL;
206 }
207 
208 /*
209  * Check whether the new IRQ is compatible with those already associated with
210  * the PMU (e.g. we don't have mismatched PPIs).
211  */
212 static bool pmu_irq_matches(struct arm_pmu *pmu, int irq)
213 {
214 	struct pmu_hw_events __percpu *hw_events = pmu->hw_events;
215 	int cpu;
216 
217 	if (!irq)
218 		return true;
219 
220 	for_each_cpu(cpu, &pmu->supported_cpus) {
221 		int other_irq = per_cpu(hw_events->irq, cpu);
222 		if (!other_irq)
223 			continue;
224 
225 		if (irq == other_irq)
226 			continue;
227 		if (!irq_is_percpu_devid(irq) && !irq_is_percpu_devid(other_irq))
228 			continue;
229 
230 		pr_warn("mismatched PPIs detected\n");
231 		return false;
232 	}
233 
234 	return true;
235 }
236 
237 static void arm_pmu_acpi_associate_pmu_cpu(struct arm_pmu *pmu,
238 					   unsigned int cpu)
239 {
240 	int irq = per_cpu(pmu_irqs, cpu);
241 
242 	per_cpu(probed_pmus, cpu) = pmu;
243 
244 	if (pmu_irq_matches(pmu, irq)) {
245 		struct pmu_hw_events __percpu *hw_events;
246 		hw_events = pmu->hw_events;
247 		per_cpu(hw_events->irq, cpu) = irq;
248 	}
249 
250 	cpumask_set_cpu(cpu, &pmu->supported_cpus);
251 }
252 
253 /*
254  * This must run before the common arm_pmu hotplug logic, so that we can
255  * associate a CPU and its interrupt before the common code tries to manage the
256  * affinity and so on.
257  *
258  * Note that hotplug events are serialized, so we cannot race with another CPU
259  * coming up. The perf core won't open events while a hotplug event is in
260  * progress.
261  */
262 static int arm_pmu_acpi_cpu_starting(unsigned int cpu)
263 {
264 	struct arm_pmu *pmu;
265 
266 	/* If we've already probed this CPU, we have nothing to do */
267 	if (per_cpu(probed_pmus, cpu))
268 		return 0;
269 
270 	pmu = arm_pmu_acpi_find_pmu();
271 	if (!pmu) {
272 		pr_warn_ratelimited("Unable to associate CPU%d with a PMU\n",
273 				    cpu);
274 		return 0;
275 	}
276 
277 	arm_pmu_acpi_associate_pmu_cpu(pmu, cpu);
278 	return 0;
279 }
280 
281 static void arm_pmu_acpi_probe_matching_cpus(struct arm_pmu *pmu,
282 					     unsigned long cpuid)
283 {
284 	int cpu;
285 
286 	for_each_online_cpu(cpu) {
287 		unsigned long cpu_cpuid = per_cpu(cpu_data, cpu).reg_midr;
288 
289 		if (cpu_cpuid == cpuid)
290 			arm_pmu_acpi_associate_pmu_cpu(pmu, cpu);
291 	}
292 }
293 
294 int arm_pmu_acpi_probe(armpmu_init_fn init_fn)
295 {
296 	int pmu_idx = 0;
297 	unsigned int cpu;
298 	int ret;
299 
300 	ret = arm_pmu_acpi_parse_irqs();
301 	if (ret)
302 		return ret;
303 
304 	ret = cpuhp_setup_state_nocalls(CPUHP_AP_PERF_ARM_ACPI_STARTING,
305 					"perf/arm/pmu_acpi:starting",
306 					arm_pmu_acpi_cpu_starting, NULL);
307 	if (ret)
308 		return ret;
309 
310 	/*
311 	 * Initialise and register the set of PMUs which we know about right
312 	 * now. Ideally we'd do this in arm_pmu_acpi_cpu_starting() so that we
313 	 * could handle late hotplug, but this may lead to deadlock since we
314 	 * might try to register a hotplug notifier instance from within a
315 	 * hotplug notifier.
316 	 *
317 	 * There's also the problem of having access to the right init_fn,
318 	 * without tying this too deeply into the "real" PMU driver.
319 	 *
320 	 * For the moment, as with the platform/DT case, we need at least one
321 	 * of a PMU's CPUs to be online at probe time.
322 	 */
323 	for_each_online_cpu(cpu) {
324 		struct arm_pmu *pmu = per_cpu(probed_pmus, cpu);
325 		unsigned long cpuid;
326 		char *base_name;
327 
328 		/* If we've already probed this CPU, we have nothing to do */
329 		if (pmu)
330 			continue;
331 
332 		pmu = armpmu_alloc();
333 		if (!pmu) {
334 			pr_warn("Unable to allocate PMU for CPU%d\n",
335 				cpu);
336 			return -ENOMEM;
337 		}
338 
339 		cpuid = per_cpu(cpu_data, cpu).reg_midr;
340 		pmu->acpi_cpuid = cpuid;
341 
342 		arm_pmu_acpi_probe_matching_cpus(pmu, cpuid);
343 
344 		ret = init_fn(pmu);
345 		if (ret == -ENODEV) {
346 			/* PMU not handled by this driver, or not present */
347 			continue;
348 		} else if (ret) {
349 			pr_warn("Unable to initialise PMU for CPU%d\n", cpu);
350 			return ret;
351 		}
352 
353 		base_name = pmu->name;
354 		pmu->name = kasprintf(GFP_KERNEL, "%s_%d", base_name, pmu_idx++);
355 		if (!pmu->name) {
356 			pr_warn("Unable to allocate PMU name for CPU%d\n", cpu);
357 			return -ENOMEM;
358 		}
359 
360 		ret = armpmu_register(pmu);
361 		if (ret) {
362 			pr_warn("Failed to register PMU for CPU%d\n", cpu);
363 			kfree(pmu->name);
364 			return ret;
365 		}
366 	}
367 
368 	return ret;
369 }
370 
371 static int arm_pmu_acpi_init(void)
372 {
373 	if (acpi_disabled)
374 		return 0;
375 
376 	arm_spe_acpi_register_device();
377 
378 	return 0;
379 }
380 subsys_initcall(arm_pmu_acpi_init)
381