| 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 |
static int __maybe_unused |
| 73 |
arm_acpi_register_pmu_device(struct platform_device *pdev, u8 len, |
| 74 |
u16 (*parse_gsi)(struct acpi_madt_generic_interrupt *)) |
| 75 |
{ |
| 76 |
int cpu, this_hetid, hetid, irq, ret; |
| 77 |
u16 this_gsi = 0, gsi = 0; |
| 78 |
|
| 79 |
/* |
| 80 |
* Ensure that platform device must have IORESOURCE_IRQ |
| 81 |
* resource to hold gsi interrupt. |
| 82 |
*/ |
| 83 |
if (pdev->num_resources != 1) |
| 84 |
return -ENXIO; |
| 85 |
|
| 86 |
if (pdev->resource[0].flags != IORESOURCE_IRQ) |
| 87 |
return -ENXIO; |
| 88 |
|
| 89 |
/* |
| 90 |
* Sanity check all the GICC tables for the same interrupt |
| 91 |
* number. For now, only support homogeneous ACPI machines. |
| 92 |
*/ |
| 93 |
for_each_possible_cpu(cpu) { |
| 94 |
struct acpi_madt_generic_interrupt *gicc; |
| 95 |
|
| 96 |
gicc = acpi_cpu_get_madt_gicc(cpu); |
| 97 |
if (gicc->header.length < len) |
| 98 |
return gsi ? -ENXIO : 0; |
| 99 |
|
| 100 |
this_gsi = parse_gsi(gicc); |
| 101 |
this_hetid = find_acpi_cpu_topology_hetero_id(cpu); |
| 102 |
if (!gsi) { |
| 103 |
hetid = this_hetid; |
| 104 |
gsi = this_gsi; |
| 105 |
} else if (hetid != this_hetid || gsi != this_gsi) { |
| 106 |
pr_warn("ACPI: %s: must be homogeneous\n", pdev->name); |
| 107 |
return -ENXIO; |
| 108 |
} |
| 109 |
} |
| 110 |
|
| 111 |
if (!this_gsi) |
| 112 |
return 0; |
| 113 |
|
| 114 |
irq = acpi_register_gsi(NULL, gsi, ACPI_LEVEL_SENSITIVE, ACPI_ACTIVE_HIGH); |
| 115 |
if (irq < 0) { |
| 116 |
pr_warn("ACPI: %s Unable to register interrupt: %d\n", pdev->name, gsi); |
| 117 |
return -ENXIO; |
| 118 |
} |
| 119 |
|
| 120 |
pdev->resource[0].start = irq; |
| 121 |
ret = platform_device_register(pdev); |
| 122 |
if (ret) |
| 123 |
acpi_unregister_gsi(gsi); |
| 124 |
|
| 125 |
return ret; |
| 126 |
} |
| 127 |
|
| 128 |
#if IS_ENABLED(CONFIG_ARM_SPE_PMU) |
| 129 |
static struct resource spe_resources[] = { |
| 130 |
{ |
| 131 |
/* irq */ |
| 132 |
.flags = IORESOURCE_IRQ, |
| 133 |
} |
| 134 |
}; |
| 135 |
|
| 136 |
static struct platform_device spe_dev = { |
| 137 |
.name = ARMV8_SPE_PDEV_NAME, |
| 138 |
.id = -1, |
| 139 |
.resource = spe_resources, |
| 140 |
.num_resources = ARRAY_SIZE(spe_resources) |
| 141 |
}; |
| 142 |
|
| 143 |
static u16 arm_spe_parse_gsi(struct acpi_madt_generic_interrupt *gicc) |
| 144 |
{ |
| 145 |
return gicc->spe_interrupt; |
| 146 |
} |
| 147 |
|
| 148 |
/* |
| 149 |
* For lack of a better place, hook the normal PMU MADT walk |
| 150 |
* and create a SPE device if we detect a recent MADT with |
| 151 |
* a homogeneous PPI mapping. |
| 152 |
*/ |
| 153 |
static void arm_spe_acpi_register_device(void) |
| 154 |
{ |
| 155 |
int ret = arm_acpi_register_pmu_device(&spe_dev, ACPI_MADT_GICC_SPE, |
| 156 |
arm_spe_parse_gsi); |
| 157 |
if (ret) |
| 158 |
pr_warn("ACPI: SPE: Unable to register device\n"); |
| 159 |
} |
| 160 |
#else |
| 161 |
static inline void arm_spe_acpi_register_device(void) |
| 162 |
{ |
| 163 |
} |
| 164 |
#endif /* CONFIG_ARM_SPE_PMU */ |
| 165 |
|
| 166 |
#if IS_ENABLED(CONFIG_CORESIGHT_TRBE) |
| 167 |
static struct resource trbe_resources[] = { |
| 168 |
{ |
| 169 |
/* irq */ |
| 170 |
.flags = IORESOURCE_IRQ, |
| 171 |
} |
| 172 |
}; |
| 173 |
|
| 174 |
static struct platform_device trbe_dev = { |
| 175 |
.name = ARMV8_TRBE_PDEV_NAME, |
| 176 |
.id = -1, |
| 177 |
.resource = trbe_resources, |
| 178 |
.num_resources = ARRAY_SIZE(trbe_resources) |
| 179 |
}; |
| 180 |
|
| 181 |
static u16 arm_trbe_parse_gsi(struct acpi_madt_generic_interrupt *gicc) |
| 182 |
{ |
| 183 |
return gicc->trbe_interrupt; |
| 184 |
} |
| 185 |
|
| 186 |
static void arm_trbe_acpi_register_device(void) |
| 187 |
{ |
| 188 |
int ret = arm_acpi_register_pmu_device(&trbe_dev, ACPI_MADT_GICC_TRBE, |
| 189 |
arm_trbe_parse_gsi); |
| 190 |
if (ret) |
| 191 |
pr_warn("ACPI: TRBE: Unable to register device\n"); |
| 192 |
} |
| 193 |
#else |
| 194 |
static inline void arm_trbe_acpi_register_device(void) |
| 195 |
{ |
| 196 |
|
| 197 |
} |
| 198 |
#endif /* CONFIG_CORESIGHT_TRBE */ |
| 199 |
|
| 200 |
static int arm_pmu_acpi_parse_irqs(void) |
| 201 |
{ |
| 202 |
int irq, cpu, irq_cpu, err; |
| 203 |
|
| 204 |
for_each_possible_cpu(cpu) { |
| 205 |
irq = arm_pmu_acpi_register_irq(cpu); |
| 206 |
if (irq < 0) { |
| 207 |
err = irq; |
| 208 |
pr_warn("Unable to parse ACPI PMU IRQ for CPU%d: %d\n", |
| 209 |
cpu, err); |
| 210 |
goto out_err; |
| 211 |
} else if (irq == 0) { |
| 212 |
pr_warn("No ACPI PMU IRQ for CPU%d\n", cpu); |
| 213 |
} |
| 214 |
|
| 215 |
/* |
| 216 |
* Log and request the IRQ so the core arm_pmu code can manage |
| 217 |
* it. We'll have to sanity-check IRQs later when we associate |
| 218 |
* them with their PMUs. |
| 219 |
*/ |
| 220 |
per_cpu(pmu_irqs, cpu) = irq; |
| 221 |
err = armpmu_request_irq(&probed_pmus, irq, cpu); |
| 222 |
if (err) |
| 223 |
goto out_err; |
| 224 |
} |
| 225 |
|
| 226 |
return 0; |
| 227 |
|
| 228 |
out_err: |
| 229 |
for_each_possible_cpu(cpu) { |
| 230 |
irq = per_cpu(pmu_irqs, cpu); |
| 231 |
if (!irq) |
| 232 |
continue; |
| 233 |
|
| 234 |
arm_pmu_acpi_unregister_irq(cpu); |
| 235 |
|
| 236 |
/* |
| 237 |
* Blat all copies of the IRQ so that we only unregister the |
| 238 |
* corresponding GSI once (e.g. when we have PPIs). |
| 239 |
*/ |
| 240 |
for_each_possible_cpu(irq_cpu) { |
| 241 |
if (per_cpu(pmu_irqs, irq_cpu) == irq) |
| 242 |
per_cpu(pmu_irqs, irq_cpu) = 0; |
| 243 |
} |
| 244 |
} |
| 245 |
|
| 246 |
return err; |
| 247 |
} |
| 248 |
|
| 249 |
static struct arm_pmu *arm_pmu_acpi_find_pmu(void) |
| 250 |
{ |
| 251 |
unsigned long cpuid = read_cpuid_id(); |
| 252 |
struct arm_pmu *pmu; |
| 253 |
int cpu; |
| 254 |
|
| 255 |
for_each_possible_cpu(cpu) { |
| 256 |
pmu = per_cpu(probed_pmus, cpu); |
| 257 |
if (!pmu || pmu->acpi_cpuid != cpuid) |
| 258 |
continue; |
| 259 |
|
| 260 |
return pmu; |
| 261 |
} |
| 262 |
|
| 263 |
return NULL; |
| 264 |
} |
| 265 |
|
| 266 |
/* |
| 267 |
* Check whether the new IRQ is compatible with those already associated with |
| 268 |
* the PMU (e.g. we don't have mismatched PPIs). |
| 269 |
*/ |
| 270 |
static bool pmu_irq_matches(struct arm_pmu *pmu, int irq) |
| 271 |
{ |
| 272 |
struct pmu_hw_events __percpu *hw_events = pmu->hw_events; |
| 273 |
int cpu; |
| 274 |
|
| 275 |
if (!irq) |
| 276 |
return true; |
| 277 |
|
| 278 |
for_each_cpu(cpu, &pmu->supported_cpus) { |
| 279 |
int other_irq = per_cpu(hw_events->irq, cpu); |
| 280 |
if (!other_irq) |
| 281 |
continue; |
| 282 |
|
| 283 |
if (irq == other_irq) |
| 284 |
continue; |
| 285 |
if (!irq_is_percpu_devid(irq) && !irq_is_percpu_devid(other_irq)) |
| 286 |
continue; |
| 287 |
|
| 288 |
pr_warn("mismatched PPIs detected\n"); |
| 289 |
return false; |
| 290 |
} |
| 291 |
|
| 292 |
return true; |
| 293 |
} |
| 294 |
|
| 295 |
static void arm_pmu_acpi_associate_pmu_cpu(struct arm_pmu *pmu, |
| 296 |
unsigned int cpu) |
| 297 |
{ |
| 298 |
int irq = per_cpu(pmu_irqs, cpu); |
| 299 |
|
| 300 |
per_cpu(probed_pmus, cpu) = pmu; |
| 301 |
|
| 302 |
if (pmu_irq_matches(pmu, irq)) { |
| 303 |
struct pmu_hw_events __percpu *hw_events; |
| 304 |
hw_events = pmu->hw_events; |
| 305 |
per_cpu(hw_events->irq, cpu) = irq; |
| 306 |
} |
| 307 |
|
| 308 |
cpumask_set_cpu(cpu, &pmu->supported_cpus); |
| 309 |
} |
| 310 |
|
| 311 |
/* |
| 312 |
* This must run before the common arm_pmu hotplug logic, so that we can |
| 313 |
* associate a CPU and its interrupt before the common code tries to manage the |
| 314 |
* affinity and so on. |
| 315 |
* |
| 316 |
* Note that hotplug events are serialized, so we cannot race with another CPU |
| 317 |
* coming up. The perf core won't open events while a hotplug event is in |
| 318 |
* progress. |
| 319 |
*/ |
| 320 |
static int arm_pmu_acpi_cpu_starting(unsigned int cpu) |
| 321 |
{ |
| 322 |
struct arm_pmu *pmu; |
| 323 |
|
| 324 |
/* If we've already probed this CPU, we have nothing to do */ |
| 325 |
if (per_cpu(probed_pmus, cpu)) |
| 326 |
return 0; |
| 327 |
|
| 328 |
pmu = arm_pmu_acpi_find_pmu(); |
| 329 |
if (!pmu) { |
| 330 |
pr_warn_ratelimited("Unable to associate CPU%d with a PMU\n", |
| 331 |
cpu); |
| 332 |
return 0; |
| 333 |
} |
| 334 |
|
| 335 |
arm_pmu_acpi_associate_pmu_cpu(pmu, cpu); |
| 336 |
return 0; |
| 337 |
} |
| 338 |
|
| 339 |
static void arm_pmu_acpi_probe_matching_cpus(struct arm_pmu *pmu, |
| 340 |
unsigned long cpuid) |
| 341 |
{ |
| 342 |
int cpu; |
| 343 |
|
| 344 |
for_each_online_cpu(cpu) { |
| 345 |
unsigned long cpu_cpuid = per_cpu(cpu_data, cpu).reg_midr; |
| 346 |
|
| 347 |
if (cpu_cpuid == cpuid) |
| 348 |
arm_pmu_acpi_associate_pmu_cpu(pmu, cpu); |
| 349 |
} |
| 350 |
} |
| 351 |
|
| 352 |
int arm_pmu_acpi_probe(armpmu_init_fn init_fn) |
| 353 |
{ |
| 354 |
int pmu_idx = 0; |
| 355 |
unsigned int cpu; |
| 356 |
int ret; |
| 357 |
|
| 358 |
ret = arm_pmu_acpi_parse_irqs(); |
| 359 |
if (ret) |
| 360 |
return ret; |
| 361 |
|
| 362 |
ret = cpuhp_setup_state_nocalls(CPUHP_AP_PERF_ARM_ACPI_STARTING, |
| 363 |
"perf/arm/pmu_acpi:starting", |
| 364 |
arm_pmu_acpi_cpu_starting, NULL); |
| 365 |
if (ret) |
| 366 |
return ret; |
| 367 |
|
| 368 |
/* |
| 369 |
* Initialise and register the set of PMUs which we know about right |
| 370 |
* now. Ideally we'd do this in arm_pmu_acpi_cpu_starting() so that we |
| 371 |
* could handle late hotplug, but this may lead to deadlock since we |
| 372 |
* might try to register a hotplug notifier instance from within a |
| 373 |
* hotplug notifier. |
| 374 |
* |
| 375 |
* There's also the problem of having access to the right init_fn, |
| 376 |
* without tying this too deeply into the "real" PMU driver. |
| 377 |
* |
| 378 |
* For the moment, as with the platform/DT case, we need at least one |
| 379 |
* of a PMU's CPUs to be online at probe time. |
| 380 |
*/ |
| 381 |
for_each_online_cpu(cpu) { |
| 382 |
struct arm_pmu *pmu = per_cpu(probed_pmus, cpu); |
| 383 |
unsigned long cpuid; |
| 384 |
char *base_name; |
| 385 |
|
| 386 |
/* If we've already probed this CPU, we have nothing to do */ |
| 387 |
if (pmu) |
| 388 |
continue; |
| 389 |
|
| 390 |
pmu = armpmu_alloc(); |
| 391 |
if (!pmu) { |
| 392 |
pr_warn("Unable to allocate PMU for CPU%d\n", |
| 393 |
cpu); |
| 394 |
return -ENOMEM; |
| 395 |
} |
| 396 |
|
| 397 |
cpuid = per_cpu(cpu_data, cpu).reg_midr; |
| 398 |
pmu->acpi_cpuid = cpuid; |
| 399 |
|
| 400 |
arm_pmu_acpi_probe_matching_cpus(pmu, cpuid); |
| 401 |
|
| 402 |
ret = init_fn(pmu); |
| 403 |
if (ret == -ENODEV) { |
| 404 |
/* PMU not handled by this driver, or not present */ |
| 405 |
continue; |
| 406 |
} else if (ret) { |
| 407 |
pr_warn("Unable to initialise PMU for CPU%d\n", cpu); |
| 408 |
return ret; |
| 409 |
} |
| 410 |
|
| 411 |
base_name = pmu->name; |
| 412 |
pmu->name = kasprintf(GFP_KERNEL, "%s_%d", base_name, pmu_idx++); |
| 413 |
if (!pmu->name) { |
| 414 |
pr_warn("Unable to allocate PMU name for CPU%d\n", cpu); |
| 415 |
return -ENOMEM; |
| 416 |
} |
| 417 |
|
| 418 |
ret = armpmu_register(pmu); |
| 419 |
if (ret) { |
| 420 |
pr_warn("Failed to register PMU for CPU%d\n", cpu); |
| 421 |
kfree(pmu->name); |
| 422 |
return ret; |
| 423 |
} |
| 424 |
} |
| 425 |
|
| 426 |
return ret; |
| 427 |
} |
| 428 |
|
| 429 |
static int arm_pmu_acpi_init(void) |
| 430 |
{ |
| 431 |
if (acpi_disabled) |
| 432 |
return 0; |
| 433 |
|
| 434 |
arm_spe_acpi_register_device(); |
| 435 |
arm_trbe_acpi_register_device(); |
| 436 |
|
| 437 |
return 0; |
| 438 |
} |
| 439 |
subsys_initcall(arm_pmu_acpi_init) |