| 1 |
// SPDX-License-Identifier: GPL-2.0 |
| 2 |
/* |
| 3 |
* ARM CoreSight Architecture PMU driver. |
| 4 |
* |
| 5 |
* This driver adds support for uncore PMU based on ARM CoreSight Performance |
| 6 |
* Monitoring Unit Architecture. The PMU is accessible via MMIO registers and |
| 7 |
* like other uncore PMUs, it does not support process specific events and |
| 8 |
* cannot be used in sampling mode. |
| 9 |
* |
| 10 |
* This code is based on other uncore PMUs like ARM DSU PMU. It provides a |
| 11 |
* generic implementation to operate the PMU according to CoreSight PMU |
| 12 |
* architecture and ACPI ARM PMU table (APMT) documents below: |
| 13 |
* - ARM CoreSight PMU architecture document number: ARM IHI 0091 A.a-00bet0. |
| 14 |
* - APMT document number: ARM DEN0117. |
| 15 |
* |
| 16 |
* The user should refer to the vendor technical documentation to get details |
| 17 |
* about the supported events. |
| 18 |
* |
| 19 |
* Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved. |
| 20 |
* |
| 21 |
*/ |
| 22 |
|
| 23 |
#include <linux/acpi.h> |
| 24 |
#include <linux/cacheinfo.h> |
| 25 |
#include <linux/ctype.h> |
| 26 |
#include <linux/interrupt.h> |
| 27 |
#include <linux/io-64-nonatomic-lo-hi.h> |
| 28 |
#include <linux/module.h> |
| 29 |
#include <linux/mutex.h> |
| 30 |
#include <linux/of.h> |
| 31 |
#include <linux/perf_event.h> |
| 32 |
#include <linux/platform_device.h> |
| 33 |
|
| 34 |
#include "arm_cspmu.h" |
| 35 |
|
| 36 |
#define PMUNAME "arm_cspmu" |
| 37 |
#define DRVNAME "arm-cs-arch-pmu" |
| 38 |
|
| 39 |
#define ARM_CSPMU_CPUMASK_ATTR(_name, _config) \ |
| 40 |
ARM_CSPMU_EXT_ATTR(_name, arm_cspmu_cpumask_show, \ |
| 41 |
(unsigned long)_config) |
| 42 |
|
| 43 |
/* Each SET/CLR register supports up to 32 counters. */ |
| 44 |
#define ARM_CSPMU_SET_CLR_COUNTER_SHIFT 5 |
| 45 |
#define ARM_CSPMU_SET_CLR_COUNTER_NUM \ |
| 46 |
(1 << ARM_CSPMU_SET_CLR_COUNTER_SHIFT) |
| 47 |
|
| 48 |
/* Convert counter idx into SET/CLR register number. */ |
| 49 |
#define COUNTER_TO_SET_CLR_ID(idx) \ |
| 50 |
(idx >> ARM_CSPMU_SET_CLR_COUNTER_SHIFT) |
| 51 |
|
| 52 |
/* Convert counter idx into SET/CLR register bit. */ |
| 53 |
#define COUNTER_TO_SET_CLR_BIT(idx) \ |
| 54 |
(idx & (ARM_CSPMU_SET_CLR_COUNTER_NUM - 1)) |
| 55 |
|
| 56 |
#define ARM_CSPMU_ACTIVE_CPU_MASK 0x0 |
| 57 |
#define ARM_CSPMU_ASSOCIATED_CPU_MASK 0x1 |
| 58 |
|
| 59 |
/* |
| 60 |
* Maximum poll count for reading counter value using high-low-high sequence. |
| 61 |
*/ |
| 62 |
#define HILOHI_MAX_POLL 1000 |
| 63 |
|
| 64 |
static unsigned long arm_cspmu_cpuhp_state; |
| 65 |
|
| 66 |
static DEFINE_MUTEX(arm_cspmu_lock); |
| 67 |
|
| 68 |
static void arm_cspmu_set_ev_filter(struct arm_cspmu *cspmu, |
| 69 |
const struct perf_event *event); |
| 70 |
static void arm_cspmu_set_cc_filter(struct arm_cspmu *cspmu, |
| 71 |
const struct perf_event *event); |
| 72 |
|
| 73 |
static struct acpi_apmt_node *arm_cspmu_apmt_node(struct device *dev) |
| 74 |
{ |
| 75 |
struct acpi_apmt_node **ptr = dev_get_platdata(dev); |
| 76 |
|
| 77 |
return ptr ? *ptr : NULL; |
| 78 |
} |
| 79 |
|
| 80 |
/* |
| 81 |
* In CoreSight PMU architecture, all of the MMIO registers are 32-bit except |
| 82 |
* counter register. The counter register can be implemented as 32-bit or 64-bit |
| 83 |
* register depending on the value of PMCFGR.SIZE field. For 64-bit access, |
| 84 |
* single-copy 64-bit atomic support is implementation defined. APMT node flag |
| 85 |
* is used to identify if the PMU supports 64-bit single copy atomic. If 64-bit |
| 86 |
* single copy atomic is not supported, the driver treats the register as a pair |
| 87 |
* of 32-bit register. |
| 88 |
*/ |
| 89 |
|
| 90 |
/* |
| 91 |
* Read 64-bit register as a pair of 32-bit registers using hi-lo-hi sequence. |
| 92 |
*/ |
| 93 |
static u64 read_reg64_hilohi(const void __iomem *addr, u32 max_poll_count) |
| 94 |
{ |
| 95 |
u32 val_lo, val_hi; |
| 96 |
u64 val; |
| 97 |
|
| 98 |
/* Use high-low-high sequence to avoid tearing */ |
| 99 |
do { |
| 100 |
if (max_poll_count-- == 0) { |
| 101 |
pr_err("ARM CSPMU: timeout hi-low-high sequence\n"); |
| 102 |
return 0; |
| 103 |
} |
| 104 |
|
| 105 |
val_hi = readl(addr + 4); |
| 106 |
val_lo = readl(addr); |
| 107 |
} while (val_hi != readl(addr + 4)); |
| 108 |
|
| 109 |
val = (((u64)val_hi << 32) | val_lo); |
| 110 |
|
| 111 |
return val; |
| 112 |
} |
| 113 |
|
| 114 |
/* Check if cycle counter is supported. */ |
| 115 |
static inline bool supports_cycle_counter(const struct arm_cspmu *cspmu) |
| 116 |
{ |
| 117 |
return (cspmu->pmcfgr & PMCFGR_CC); |
| 118 |
} |
| 119 |
|
| 120 |
/* Get counter size, which is (PMCFGR_SIZE + 1). */ |
| 121 |
static inline u32 counter_size(const struct arm_cspmu *cspmu) |
| 122 |
{ |
| 123 |
return FIELD_GET(PMCFGR_SIZE, cspmu->pmcfgr) + 1; |
| 124 |
} |
| 125 |
|
| 126 |
/* Get counter mask. */ |
| 127 |
static inline u64 counter_mask(const struct arm_cspmu *cspmu) |
| 128 |
{ |
| 129 |
return GENMASK_ULL(counter_size(cspmu) - 1, 0); |
| 130 |
} |
| 131 |
|
| 132 |
/* Check if counter is implemented as 64-bit register. */ |
| 133 |
static inline bool use_64b_counter_reg(const struct arm_cspmu *cspmu) |
| 134 |
{ |
| 135 |
return (counter_size(cspmu) > 32); |
| 136 |
} |
| 137 |
|
| 138 |
ssize_t arm_cspmu_sysfs_event_show(struct device *dev, |
| 139 |
struct device_attribute *attr, char *buf) |
| 140 |
{ |
| 141 |
struct perf_pmu_events_attr *pmu_attr; |
| 142 |
|
| 143 |
pmu_attr = container_of(attr, typeof(*pmu_attr), attr); |
| 144 |
return sysfs_emit(buf, "event=0x%llx\n", pmu_attr->id); |
| 145 |
} |
| 146 |
EXPORT_SYMBOL_GPL(arm_cspmu_sysfs_event_show); |
| 147 |
|
| 148 |
/* Default event list. */ |
| 149 |
static struct attribute *arm_cspmu_event_attrs[] = { |
| 150 |
ARM_CSPMU_EVENT_ATTR(cycles, ARM_CSPMU_EVT_CYCLES_DEFAULT), |
| 151 |
NULL, |
| 152 |
}; |
| 153 |
|
| 154 |
static struct attribute ** |
| 155 |
arm_cspmu_get_event_attrs(const struct arm_cspmu *cspmu) |
| 156 |
{ |
| 157 |
struct attribute **attrs; |
| 158 |
|
| 159 |
attrs = devm_kmemdup(cspmu->dev, arm_cspmu_event_attrs, |
| 160 |
sizeof(arm_cspmu_event_attrs), GFP_KERNEL); |
| 161 |
|
| 162 |
return attrs; |
| 163 |
} |
| 164 |
|
| 165 |
static umode_t |
| 166 |
arm_cspmu_event_attr_is_visible(struct kobject *kobj, |
| 167 |
struct attribute *attr, int unused) |
| 168 |
{ |
| 169 |
struct device *dev = kobj_to_dev(kobj); |
| 170 |
struct arm_cspmu *cspmu = to_arm_cspmu(dev_get_drvdata(dev)); |
| 171 |
struct perf_pmu_events_attr *eattr; |
| 172 |
|
| 173 |
eattr = container_of(attr, typeof(*eattr), attr.attr); |
| 174 |
|
| 175 |
/* Hide cycle event if not supported */ |
| 176 |
if (!supports_cycle_counter(cspmu) && |
| 177 |
eattr->id == ARM_CSPMU_EVT_CYCLES_DEFAULT) |
| 178 |
return 0; |
| 179 |
|
| 180 |
return attr->mode; |
| 181 |
} |
| 182 |
|
| 183 |
static struct attribute *arm_cspmu_format_attrs[] = { |
| 184 |
ARM_CSPMU_FORMAT_EVENT_ATTR, |
| 185 |
ARM_CSPMU_FORMAT_FILTER_ATTR, |
| 186 |
ARM_CSPMU_FORMAT_FILTER2_ATTR, |
| 187 |
NULL, |
| 188 |
}; |
| 189 |
|
| 190 |
static struct attribute ** |
| 191 |
arm_cspmu_get_format_attrs(const struct arm_cspmu *cspmu) |
| 192 |
{ |
| 193 |
struct attribute **attrs; |
| 194 |
|
| 195 |
attrs = devm_kmemdup(cspmu->dev, arm_cspmu_format_attrs, |
| 196 |
sizeof(arm_cspmu_format_attrs), GFP_KERNEL); |
| 197 |
|
| 198 |
return attrs; |
| 199 |
} |
| 200 |
|
| 201 |
static u32 arm_cspmu_event_type(const struct perf_event *event) |
| 202 |
{ |
| 203 |
return event->attr.config & ARM_CSPMU_EVENT_MASK; |
| 204 |
} |
| 205 |
|
| 206 |
static bool arm_cspmu_is_cycle_counter_event(const struct perf_event *event) |
| 207 |
{ |
| 208 |
return (event->attr.config == ARM_CSPMU_EVT_CYCLES_DEFAULT); |
| 209 |
} |
| 210 |
|
| 211 |
static ssize_t arm_cspmu_identifier_show(struct device *dev, |
| 212 |
struct device_attribute *attr, |
| 213 |
char *page) |
| 214 |
{ |
| 215 |
struct arm_cspmu *cspmu = to_arm_cspmu(dev_get_drvdata(dev)); |
| 216 |
|
| 217 |
return sysfs_emit(page, "%s\n", cspmu->identifier); |
| 218 |
} |
| 219 |
|
| 220 |
static struct device_attribute arm_cspmu_identifier_attr = |
| 221 |
__ATTR(identifier, 0444, arm_cspmu_identifier_show, NULL); |
| 222 |
|
| 223 |
static struct attribute *arm_cspmu_identifier_attrs[] = { |
| 224 |
&arm_cspmu_identifier_attr.attr, |
| 225 |
NULL, |
| 226 |
}; |
| 227 |
|
| 228 |
static struct attribute_group arm_cspmu_identifier_attr_group = { |
| 229 |
.attrs = arm_cspmu_identifier_attrs, |
| 230 |
}; |
| 231 |
|
| 232 |
static const char *arm_cspmu_get_identifier(const struct arm_cspmu *cspmu) |
| 233 |
{ |
| 234 |
const char *identifier = |
| 235 |
devm_kasprintf(cspmu->dev, GFP_KERNEL, "%x", |
| 236 |
cspmu->impl.pmiidr); |
| 237 |
return identifier; |
| 238 |
} |
| 239 |
|
| 240 |
static const char *arm_cspmu_type_str[ACPI_APMT_NODE_TYPE_COUNT] = { |
| 241 |
"mc", |
| 242 |
"smmu", |
| 243 |
"pcie", |
| 244 |
"acpi", |
| 245 |
"cache", |
| 246 |
}; |
| 247 |
|
| 248 |
static const char *arm_cspmu_get_name(const struct arm_cspmu *cspmu) |
| 249 |
{ |
| 250 |
struct device *dev; |
| 251 |
struct acpi_apmt_node *apmt_node; |
| 252 |
u8 pmu_type; |
| 253 |
char *name; |
| 254 |
char acpi_hid_string[ACPI_ID_LEN] = { 0 }; |
| 255 |
static atomic_t pmu_idx[ACPI_APMT_NODE_TYPE_COUNT] = { 0 }; |
| 256 |
|
| 257 |
dev = cspmu->dev; |
| 258 |
apmt_node = arm_cspmu_apmt_node(dev); |
| 259 |
if (!apmt_node) |
| 260 |
return devm_kasprintf(dev, GFP_KERNEL, PMUNAME "_%u", |
| 261 |
atomic_fetch_inc(&pmu_idx[0])); |
| 262 |
|
| 263 |
pmu_type = apmt_node->type; |
| 264 |
|
| 265 |
if (pmu_type >= ACPI_APMT_NODE_TYPE_COUNT) { |
| 266 |
dev_err(dev, "unsupported PMU type-%u\n", pmu_type); |
| 267 |
return NULL; |
| 268 |
} |
| 269 |
|
| 270 |
if (pmu_type == ACPI_APMT_NODE_TYPE_ACPI) { |
| 271 |
memcpy(acpi_hid_string, |
| 272 |
&apmt_node->inst_primary, |
| 273 |
sizeof(apmt_node->inst_primary)); |
| 274 |
name = devm_kasprintf(dev, GFP_KERNEL, "%s_%s_%s_%u", PMUNAME, |
| 275 |
arm_cspmu_type_str[pmu_type], |
| 276 |
acpi_hid_string, |
| 277 |
apmt_node->inst_secondary); |
| 278 |
} else { |
| 279 |
name = devm_kasprintf(dev, GFP_KERNEL, "%s_%s_%d", PMUNAME, |
| 280 |
arm_cspmu_type_str[pmu_type], |
| 281 |
atomic_fetch_inc(&pmu_idx[pmu_type])); |
| 282 |
} |
| 283 |
|
| 284 |
return name; |
| 285 |
} |
| 286 |
|
| 287 |
static ssize_t arm_cspmu_cpumask_show(struct device *dev, |
| 288 |
struct device_attribute *attr, |
| 289 |
char *buf) |
| 290 |
{ |
| 291 |
struct pmu *pmu = dev_get_drvdata(dev); |
| 292 |
struct arm_cspmu *cspmu = to_arm_cspmu(pmu); |
| 293 |
struct dev_ext_attribute *eattr = |
| 294 |
container_of(attr, struct dev_ext_attribute, attr); |
| 295 |
unsigned long mask_id = (unsigned long)eattr->var; |
| 296 |
const cpumask_t *cpumask; |
| 297 |
|
| 298 |
switch (mask_id) { |
| 299 |
case ARM_CSPMU_ACTIVE_CPU_MASK: |
| 300 |
cpumask = &cspmu->active_cpu; |
| 301 |
break; |
| 302 |
case ARM_CSPMU_ASSOCIATED_CPU_MASK: |
| 303 |
cpumask = &cspmu->associated_cpus; |
| 304 |
break; |
| 305 |
default: |
| 306 |
return 0; |
| 307 |
} |
| 308 |
return cpumap_print_to_pagebuf(true, buf, cpumask); |
| 309 |
} |
| 310 |
|
| 311 |
static struct attribute *arm_cspmu_cpumask_attrs[] = { |
| 312 |
ARM_CSPMU_CPUMASK_ATTR(cpumask, ARM_CSPMU_ACTIVE_CPU_MASK), |
| 313 |
ARM_CSPMU_CPUMASK_ATTR(associated_cpus, ARM_CSPMU_ASSOCIATED_CPU_MASK), |
| 314 |
NULL, |
| 315 |
}; |
| 316 |
|
| 317 |
static struct attribute_group arm_cspmu_cpumask_attr_group = { |
| 318 |
.attrs = arm_cspmu_cpumask_attrs, |
| 319 |
}; |
| 320 |
|
| 321 |
static struct arm_cspmu_impl_match impl_match[] = { |
| 322 |
{ |
| 323 |
.module_name = "nvidia_cspmu", |
| 324 |
.pmiidr_val = ARM_CSPMU_IMPL_ID_NVIDIA, |
| 325 |
.pmiidr_mask = PMIIDR_IMPLEMENTER, |
| 326 |
.module = NULL, |
| 327 |
.impl_init_ops = NULL, |
| 328 |
}, |
| 329 |
{ |
| 330 |
.module_name = "ampere_cspmu", |
| 331 |
.pmiidr_val = ARM_CSPMU_IMPL_ID_AMPERE, |
| 332 |
.pmiidr_mask = PMIIDR_IMPLEMENTER, |
| 333 |
.module = NULL, |
| 334 |
.impl_init_ops = NULL, |
| 335 |
}, |
| 336 |
|
| 337 |
{0} |
| 338 |
}; |
| 339 |
|
| 340 |
static struct arm_cspmu_impl_match *arm_cspmu_impl_match_get(u32 pmiidr) |
| 341 |
{ |
| 342 |
struct arm_cspmu_impl_match *match = impl_match; |
| 343 |
|
| 344 |
for (; match->pmiidr_val; match++) { |
| 345 |
u32 mask = match->pmiidr_mask; |
| 346 |
|
| 347 |
if ((match->pmiidr_val & mask) == (pmiidr & mask)) |
| 348 |
return match; |
| 349 |
} |
| 350 |
|
| 351 |
return NULL; |
| 352 |
} |
| 353 |
|
| 354 |
static u32 arm_cspmu_get_pmiidr(struct arm_cspmu *cspmu) |
| 355 |
{ |
| 356 |
u32 pmiidr, pmpidr; |
| 357 |
|
| 358 |
pmiidr = readl(cspmu->base0 + PMIIDR); |
| 359 |
|
| 360 |
if (pmiidr != 0) |
| 361 |
return pmiidr; |
| 362 |
|
| 363 |
/* Construct PMIIDR value from PMPIDRs. */ |
| 364 |
|
| 365 |
pmpidr = readl(cspmu->base0 + PMPIDR0); |
| 366 |
pmiidr |= FIELD_PREP(PMIIDR_PRODUCTID_PART_0, |
| 367 |
FIELD_GET(PMPIDR0_PART_0, pmpidr)); |
| 368 |
|
| 369 |
pmpidr = readl(cspmu->base0 + PMPIDR1); |
| 370 |
pmiidr |= FIELD_PREP(PMIIDR_PRODUCTID_PART_1, |
| 371 |
FIELD_GET(PMPIDR1_PART_1, pmpidr)); |
| 372 |
pmiidr |= FIELD_PREP(PMIIDR_IMPLEMENTER_DES_0, |
| 373 |
FIELD_GET(PMPIDR1_DES_0, pmpidr)); |
| 374 |
|
| 375 |
pmpidr = readl(cspmu->base0 + PMPIDR2); |
| 376 |
pmiidr |= FIELD_PREP(PMIIDR_VARIANT, |
| 377 |
FIELD_GET(PMPIDR2_REVISION, pmpidr)); |
| 378 |
pmiidr |= FIELD_PREP(PMIIDR_IMPLEMENTER_DES_1, |
| 379 |
FIELD_GET(PMPIDR2_DES_1, pmpidr)); |
| 380 |
|
| 381 |
pmpidr = readl(cspmu->base0 + PMPIDR3); |
| 382 |
pmiidr |= FIELD_PREP(PMIIDR_REVISION, |
| 383 |
FIELD_GET(PMPIDR3_REVAND, pmpidr)); |
| 384 |
|
| 385 |
pmpidr = readl(cspmu->base0 + PMPIDR4); |
| 386 |
pmiidr |= FIELD_PREP(PMIIDR_IMPLEMENTER_DES_2, |
| 387 |
FIELD_GET(PMPIDR4_DES_2, pmpidr)); |
| 388 |
|
| 389 |
return pmiidr; |
| 390 |
} |
| 391 |
|
| 392 |
#define DEFAULT_IMPL_OP(name) .name = arm_cspmu_##name |
| 393 |
|
| 394 |
static int arm_cspmu_init_impl_ops(struct arm_cspmu *cspmu) |
| 395 |
{ |
| 396 |
int ret = 0; |
| 397 |
struct acpi_apmt_node *apmt_node = arm_cspmu_apmt_node(cspmu->dev); |
| 398 |
struct arm_cspmu_impl_match *match; |
| 399 |
|
| 400 |
/* Start with a default PMU implementation */ |
| 401 |
cspmu->impl.module = THIS_MODULE; |
| 402 |
cspmu->impl.pmiidr = arm_cspmu_get_pmiidr(cspmu); |
| 403 |
cspmu->impl.ops = (struct arm_cspmu_impl_ops) { |
| 404 |
DEFAULT_IMPL_OP(get_event_attrs), |
| 405 |
DEFAULT_IMPL_OP(get_format_attrs), |
| 406 |
DEFAULT_IMPL_OP(get_identifier), |
| 407 |
DEFAULT_IMPL_OP(get_name), |
| 408 |
DEFAULT_IMPL_OP(is_cycle_counter_event), |
| 409 |
DEFAULT_IMPL_OP(event_type), |
| 410 |
DEFAULT_IMPL_OP(set_cc_filter), |
| 411 |
DEFAULT_IMPL_OP(set_ev_filter), |
| 412 |
DEFAULT_IMPL_OP(event_attr_is_visible), |
| 413 |
}; |
| 414 |
|
| 415 |
/* Firmware may override implementer/product ID from PMIIDR */ |
| 416 |
if (apmt_node && apmt_node->impl_id) |
| 417 |
cspmu->impl.pmiidr = apmt_node->impl_id; |
| 418 |
|
| 419 |
/* Find implementer specific attribute ops. */ |
| 420 |
match = arm_cspmu_impl_match_get(cspmu->impl.pmiidr); |
| 421 |
|
| 422 |
/* Load implementer module and initialize the callbacks. */ |
| 423 |
if (match) { |
| 424 |
mutex_lock(&arm_cspmu_lock); |
| 425 |
|
| 426 |
if (match->impl_init_ops) { |
| 427 |
/* Prevent unload until PMU registration is done. */ |
| 428 |
if (try_module_get(match->module)) { |
| 429 |
cspmu->impl.module = match->module; |
| 430 |
cspmu->impl.match = match; |
| 431 |
ret = match->impl_init_ops(cspmu); |
| 432 |
if (ret) |
| 433 |
module_put(match->module); |
| 434 |
} else { |
| 435 |
WARN(1, "arm_cspmu failed to get module: %s\n", |
| 436 |
match->module_name); |
| 437 |
ret = -EINVAL; |
| 438 |
} |
| 439 |
} else { |
| 440 |
request_module_nowait(match->module_name); |
| 441 |
ret = -EPROBE_DEFER; |
| 442 |
} |
| 443 |
|
| 444 |
mutex_unlock(&arm_cspmu_lock); |
| 445 |
} |
| 446 |
|
| 447 |
return ret; |
| 448 |
} |
| 449 |
|
| 450 |
static struct attribute_group * |
| 451 |
arm_cspmu_alloc_event_attr_group(struct arm_cspmu *cspmu) |
| 452 |
{ |
| 453 |
struct attribute_group *event_group; |
| 454 |
struct device *dev = cspmu->dev; |
| 455 |
const struct arm_cspmu_impl_ops *impl_ops = &cspmu->impl.ops; |
| 456 |
|
| 457 |
event_group = |
| 458 |
devm_kzalloc(dev, sizeof(struct attribute_group), GFP_KERNEL); |
| 459 |
if (!event_group) |
| 460 |
return NULL; |
| 461 |
|
| 462 |
event_group->name = "events"; |
| 463 |
event_group->is_visible = impl_ops->event_attr_is_visible; |
| 464 |
event_group->attrs = impl_ops->get_event_attrs(cspmu); |
| 465 |
|
| 466 |
if (!event_group->attrs) |
| 467 |
return NULL; |
| 468 |
|
| 469 |
return event_group; |
| 470 |
} |
| 471 |
|
| 472 |
static struct attribute_group * |
| 473 |
arm_cspmu_alloc_format_attr_group(struct arm_cspmu *cspmu) |
| 474 |
{ |
| 475 |
struct attribute_group *format_group; |
| 476 |
struct device *dev = cspmu->dev; |
| 477 |
|
| 478 |
format_group = |
| 479 |
devm_kzalloc(dev, sizeof(struct attribute_group), GFP_KERNEL); |
| 480 |
if (!format_group) |
| 481 |
return NULL; |
| 482 |
|
| 483 |
format_group->name = "format"; |
| 484 |
format_group->attrs = cspmu->impl.ops.get_format_attrs(cspmu); |
| 485 |
|
| 486 |
if (!format_group->attrs) |
| 487 |
return NULL; |
| 488 |
|
| 489 |
return format_group; |
| 490 |
} |
| 491 |
|
| 492 |
static int arm_cspmu_alloc_attr_groups(struct arm_cspmu *cspmu) |
| 493 |
{ |
| 494 |
const struct attribute_group **attr_groups = cspmu->attr_groups; |
| 495 |
const struct arm_cspmu_impl_ops *impl_ops = &cspmu->impl.ops; |
| 496 |
|
| 497 |
cspmu->identifier = impl_ops->get_identifier(cspmu); |
| 498 |
cspmu->name = impl_ops->get_name(cspmu); |
| 499 |
|
| 500 |
if (!cspmu->identifier || !cspmu->name) |
| 501 |
return -ENOMEM; |
| 502 |
|
| 503 |
attr_groups[0] = arm_cspmu_alloc_event_attr_group(cspmu); |
| 504 |
attr_groups[1] = arm_cspmu_alloc_format_attr_group(cspmu); |
| 505 |
attr_groups[2] = &arm_cspmu_identifier_attr_group; |
| 506 |
attr_groups[3] = &arm_cspmu_cpumask_attr_group; |
| 507 |
|
| 508 |
if (!attr_groups[0] || !attr_groups[1]) |
| 509 |
return -ENOMEM; |
| 510 |
|
| 511 |
return 0; |
| 512 |
} |
| 513 |
|
| 514 |
static inline void arm_cspmu_reset_counters(struct arm_cspmu *cspmu) |
| 515 |
{ |
| 516 |
writel(PMCR_C | PMCR_P, cspmu->base0 + PMCR); |
| 517 |
} |
| 518 |
|
| 519 |
static inline void arm_cspmu_start_counters(struct arm_cspmu *cspmu) |
| 520 |
{ |
| 521 |
writel(PMCR_E, cspmu->base0 + PMCR); |
| 522 |
} |
| 523 |
|
| 524 |
static inline void arm_cspmu_stop_counters(struct arm_cspmu *cspmu) |
| 525 |
{ |
| 526 |
writel(0, cspmu->base0 + PMCR); |
| 527 |
} |
| 528 |
|
| 529 |
static void arm_cspmu_enable(struct pmu *pmu) |
| 530 |
{ |
| 531 |
bool disabled; |
| 532 |
struct arm_cspmu *cspmu = to_arm_cspmu(pmu); |
| 533 |
|
| 534 |
disabled = bitmap_empty(cspmu->hw_events.used_ctrs, |
| 535 |
cspmu->num_logical_ctrs); |
| 536 |
|
| 537 |
if (disabled) |
| 538 |
return; |
| 539 |
|
| 540 |
arm_cspmu_start_counters(cspmu); |
| 541 |
} |
| 542 |
|
| 543 |
static void arm_cspmu_disable(struct pmu *pmu) |
| 544 |
{ |
| 545 |
struct arm_cspmu *cspmu = to_arm_cspmu(pmu); |
| 546 |
|
| 547 |
arm_cspmu_stop_counters(cspmu); |
| 548 |
} |
| 549 |
|
| 550 |
static int arm_cspmu_get_event_idx(struct arm_cspmu_hw_events *hw_events, |
| 551 |
struct perf_event *event) |
| 552 |
{ |
| 553 |
int idx, ret; |
| 554 |
struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu); |
| 555 |
|
| 556 |
if (supports_cycle_counter(cspmu)) { |
| 557 |
if (cspmu->impl.ops.is_cycle_counter_event(event)) { |
| 558 |
/* Search for available cycle counter. */ |
| 559 |
if (test_and_set_bit(cspmu->cycle_counter_logical_idx, |
| 560 |
hw_events->used_ctrs)) |
| 561 |
return -EAGAIN; |
| 562 |
|
| 563 |
return cspmu->cycle_counter_logical_idx; |
| 564 |
} |
| 565 |
|
| 566 |
/* |
| 567 |
* Search a regular counter from the used counter bitmap. |
| 568 |
* The cycle counter divides the bitmap into two parts. Search |
| 569 |
* the first then second half to exclude the cycle counter bit. |
| 570 |
*/ |
| 571 |
idx = find_first_zero_bit(hw_events->used_ctrs, |
| 572 |
cspmu->cycle_counter_logical_idx); |
| 573 |
if (idx >= cspmu->cycle_counter_logical_idx) { |
| 574 |
idx = find_next_zero_bit( |
| 575 |
hw_events->used_ctrs, |
| 576 |
cspmu->num_logical_ctrs, |
| 577 |
cspmu->cycle_counter_logical_idx + 1); |
| 578 |
} |
| 579 |
} else { |
| 580 |
idx = find_first_zero_bit(hw_events->used_ctrs, |
| 581 |
cspmu->num_logical_ctrs); |
| 582 |
} |
| 583 |
|
| 584 |
if (idx >= cspmu->num_logical_ctrs) |
| 585 |
return -EAGAIN; |
| 586 |
|
| 587 |
if (cspmu->impl.ops.validate_event) { |
| 588 |
ret = cspmu->impl.ops.validate_event(cspmu, event); |
| 589 |
if (ret) |
| 590 |
return ret; |
| 591 |
} |
| 592 |
|
| 593 |
set_bit(idx, hw_events->used_ctrs); |
| 594 |
|
| 595 |
return idx; |
| 596 |
} |
| 597 |
|
| 598 |
static bool arm_cspmu_validate_event(struct pmu *pmu, |
| 599 |
struct arm_cspmu_hw_events *hw_events, |
| 600 |
struct perf_event *event) |
| 601 |
{ |
| 602 |
if (is_software_event(event)) |
| 603 |
return true; |
| 604 |
|
| 605 |
/* Reject groups spanning multiple HW PMUs. */ |
| 606 |
if (event->pmu != pmu) |
| 607 |
return false; |
| 608 |
|
| 609 |
return (arm_cspmu_get_event_idx(hw_events, event) >= 0); |
| 610 |
} |
| 611 |
|
| 612 |
/* |
| 613 |
* Make sure the group of events can be scheduled at once |
| 614 |
* on the PMU. |
| 615 |
*/ |
| 616 |
static bool arm_cspmu_validate_group(struct perf_event *event) |
| 617 |
{ |
| 618 |
struct perf_event *sibling, *leader = event->group_leader; |
| 619 |
struct arm_cspmu_hw_events fake_hw_events; |
| 620 |
|
| 621 |
if (event->group_leader == event) |
| 622 |
return true; |
| 623 |
|
| 624 |
memset(&fake_hw_events, 0, sizeof(fake_hw_events)); |
| 625 |
|
| 626 |
if (!arm_cspmu_validate_event(event->pmu, &fake_hw_events, leader)) |
| 627 |
return false; |
| 628 |
|
| 629 |
for_each_sibling_event(sibling, leader) { |
| 630 |
if (!arm_cspmu_validate_event(event->pmu, &fake_hw_events, |
| 631 |
sibling)) |
| 632 |
return false; |
| 633 |
} |
| 634 |
|
| 635 |
return arm_cspmu_validate_event(event->pmu, &fake_hw_events, event); |
| 636 |
} |
| 637 |
|
| 638 |
static int arm_cspmu_event_init(struct perf_event *event) |
| 639 |
{ |
| 640 |
struct arm_cspmu *cspmu; |
| 641 |
struct hw_perf_event *hwc = &event->hw; |
| 642 |
|
| 643 |
cspmu = to_arm_cspmu(event->pmu); |
| 644 |
|
| 645 |
if (event->attr.type != event->pmu->type) |
| 646 |
return -ENOENT; |
| 647 |
|
| 648 |
/* |
| 649 |
* Following other "uncore" PMUs, we do not support sampling mode or |
| 650 |
* attach to a task (per-process mode). |
| 651 |
*/ |
| 652 |
if (is_sampling_event(event)) { |
| 653 |
dev_dbg(cspmu->pmu.dev, |
| 654 |
"Can't support sampling events\n"); |
| 655 |
return -EOPNOTSUPP; |
| 656 |
} |
| 657 |
|
| 658 |
if (event->cpu < 0 || event->attach_state & PERF_ATTACH_TASK) { |
| 659 |
dev_dbg(cspmu->pmu.dev, |
| 660 |
"Can't support per-task counters\n"); |
| 661 |
return -EINVAL; |
| 662 |
} |
| 663 |
|
| 664 |
/* |
| 665 |
* Make sure the CPU assignment is on one of the CPUs associated with |
| 666 |
* this PMU. |
| 667 |
*/ |
| 668 |
if (!cpumask_test_cpu(event->cpu, &cspmu->associated_cpus)) { |
| 669 |
dev_dbg(cspmu->pmu.dev, |
| 670 |
"Requested cpu is not associated with the PMU\n"); |
| 671 |
return -EINVAL; |
| 672 |
} |
| 673 |
|
| 674 |
/* Enforce the current active CPU to handle the events in this PMU. */ |
| 675 |
event->cpu = cpumask_first(&cspmu->active_cpu); |
| 676 |
if (event->cpu >= nr_cpu_ids) |
| 677 |
return -EINVAL; |
| 678 |
|
| 679 |
if (!arm_cspmu_validate_group(event)) |
| 680 |
return -EINVAL; |
| 681 |
|
| 682 |
/* |
| 683 |
* The logical counter id is tracked with hw_perf_event.extra_reg.idx. |
| 684 |
* The physical counter id is tracked with hw_perf_event.idx. |
| 685 |
* We don't assign an index until we actually place the event onto |
| 686 |
* hardware. Use -1 to signify that we haven't decided where to put it |
| 687 |
* yet. |
| 688 |
*/ |
| 689 |
hwc->idx = -1; |
| 690 |
hwc->extra_reg.idx = -1; |
| 691 |
hwc->config = cspmu->impl.ops.event_type(event); |
| 692 |
|
| 693 |
return 0; |
| 694 |
} |
| 695 |
|
| 696 |
static inline u32 counter_offset(u32 reg_sz, u32 ctr_idx) |
| 697 |
{ |
| 698 |
return (PMEVCNTR_LO + (reg_sz * ctr_idx)); |
| 699 |
} |
| 700 |
|
| 701 |
static void arm_cspmu_write_counter(struct perf_event *event, u64 val) |
| 702 |
{ |
| 703 |
u32 offset; |
| 704 |
struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu); |
| 705 |
|
| 706 |
if (use_64b_counter_reg(cspmu)) { |
| 707 |
offset = counter_offset(sizeof(u64), event->hw.idx); |
| 708 |
|
| 709 |
if (cspmu->has_atomic_dword) |
| 710 |
writeq(val, cspmu->base1 + offset); |
| 711 |
else |
| 712 |
lo_hi_writeq(val, cspmu->base1 + offset); |
| 713 |
} else { |
| 714 |
offset = counter_offset(sizeof(u32), event->hw.idx); |
| 715 |
|
| 716 |
writel(lower_32_bits(val), cspmu->base1 + offset); |
| 717 |
} |
| 718 |
} |
| 719 |
|
| 720 |
static u64 arm_cspmu_read_counter(struct perf_event *event) |
| 721 |
{ |
| 722 |
u32 offset; |
| 723 |
const void __iomem *counter_addr; |
| 724 |
struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu); |
| 725 |
|
| 726 |
if (use_64b_counter_reg(cspmu)) { |
| 727 |
offset = counter_offset(sizeof(u64), event->hw.idx); |
| 728 |
counter_addr = cspmu->base1 + offset; |
| 729 |
|
| 730 |
return cspmu->has_atomic_dword ? |
| 731 |
readq(counter_addr) : |
| 732 |
read_reg64_hilohi(counter_addr, HILOHI_MAX_POLL); |
| 733 |
} |
| 734 |
|
| 735 |
offset = counter_offset(sizeof(u32), event->hw.idx); |
| 736 |
return readl(cspmu->base1 + offset); |
| 737 |
} |
| 738 |
|
| 739 |
/* |
| 740 |
* arm_cspmu_set_event_period: Set the period for the counter. |
| 741 |
* |
| 742 |
* To handle cases of extreme interrupt latency, we program |
| 743 |
* the counter with half of the max count for the counters. |
| 744 |
*/ |
| 745 |
static void arm_cspmu_set_event_period(struct perf_event *event) |
| 746 |
{ |
| 747 |
struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu); |
| 748 |
u64 val = counter_mask(cspmu) >> 1ULL; |
| 749 |
|
| 750 |
local64_set(&event->hw.prev_count, val); |
| 751 |
arm_cspmu_write_counter(event, val); |
| 752 |
} |
| 753 |
|
| 754 |
static void arm_cspmu_enable_counter(struct arm_cspmu *cspmu, int idx) |
| 755 |
{ |
| 756 |
u32 reg_id, reg_bit, inten_off, cnten_off; |
| 757 |
|
| 758 |
reg_id = COUNTER_TO_SET_CLR_ID(idx); |
| 759 |
reg_bit = COUNTER_TO_SET_CLR_BIT(idx); |
| 760 |
|
| 761 |
inten_off = PMINTENSET + (4 * reg_id); |
| 762 |
cnten_off = PMCNTENSET + (4 * reg_id); |
| 763 |
|
| 764 |
writel(BIT(reg_bit), cspmu->base0 + inten_off); |
| 765 |
writel(BIT(reg_bit), cspmu->base0 + cnten_off); |
| 766 |
} |
| 767 |
|
| 768 |
static void arm_cspmu_disable_counter(struct arm_cspmu *cspmu, int idx) |
| 769 |
{ |
| 770 |
u32 reg_id, reg_bit, inten_off, cnten_off; |
| 771 |
|
| 772 |
reg_id = COUNTER_TO_SET_CLR_ID(idx); |
| 773 |
reg_bit = COUNTER_TO_SET_CLR_BIT(idx); |
| 774 |
|
| 775 |
inten_off = PMINTENCLR + (4 * reg_id); |
| 776 |
cnten_off = PMCNTENCLR + (4 * reg_id); |
| 777 |
|
| 778 |
writel(BIT(reg_bit), cspmu->base0 + cnten_off); |
| 779 |
writel(BIT(reg_bit), cspmu->base0 + inten_off); |
| 780 |
} |
| 781 |
|
| 782 |
static void arm_cspmu_event_update(struct perf_event *event) |
| 783 |
{ |
| 784 |
struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu); |
| 785 |
struct hw_perf_event *hwc = &event->hw; |
| 786 |
u64 delta, prev, now; |
| 787 |
|
| 788 |
do { |
| 789 |
prev = local64_read(&hwc->prev_count); |
| 790 |
now = arm_cspmu_read_counter(event); |
| 791 |
} while (local64_cmpxchg(&hwc->prev_count, prev, now) != prev); |
| 792 |
|
| 793 |
delta = (now - prev) & counter_mask(cspmu); |
| 794 |
local64_add(delta, &event->count); |
| 795 |
} |
| 796 |
|
| 797 |
static inline void arm_cspmu_set_event(struct arm_cspmu *cspmu, |
| 798 |
struct hw_perf_event *hwc) |
| 799 |
{ |
| 800 |
u32 offset = PMEVTYPER + (4 * hwc->idx); |
| 801 |
|
| 802 |
writel(hwc->config, cspmu->base0 + offset); |
| 803 |
} |
| 804 |
|
| 805 |
static void arm_cspmu_set_ev_filter(struct arm_cspmu *cspmu, |
| 806 |
const struct perf_event *event) |
| 807 |
{ |
| 808 |
u32 filter = event->attr.config1 & ARM_CSPMU_FILTER_MASK; |
| 809 |
u32 filter2 = event->attr.config2 & ARM_CSPMU_FILTER_MASK; |
| 810 |
u32 offset = 4 * event->hw.idx; |
| 811 |
|
| 812 |
writel(filter, cspmu->base0 + PMEVFILTR + offset); |
| 813 |
writel(filter2, cspmu->base0 + PMEVFILT2R + offset); |
| 814 |
} |
| 815 |
|
| 816 |
static void arm_cspmu_set_cc_filter(struct arm_cspmu *cspmu, |
| 817 |
const struct perf_event *event) |
| 818 |
{ |
| 819 |
u32 filter = event->attr.config1 & ARM_CSPMU_FILTER_MASK; |
| 820 |
|
| 821 |
writel(filter, cspmu->base0 + PMCCFILTR); |
| 822 |
} |
| 823 |
|
| 824 |
static void arm_cspmu_start(struct perf_event *event, int pmu_flags) |
| 825 |
{ |
| 826 |
struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu); |
| 827 |
struct hw_perf_event *hwc = &event->hw; |
| 828 |
|
| 829 |
/* We always reprogram the counter */ |
| 830 |
if (pmu_flags & PERF_EF_RELOAD) |
| 831 |
WARN_ON(!(hwc->state & PERF_HES_UPTODATE)); |
| 832 |
|
| 833 |
arm_cspmu_set_event_period(event); |
| 834 |
|
| 835 |
if (event->hw.extra_reg.idx == cspmu->cycle_counter_logical_idx) { |
| 836 |
cspmu->impl.ops.set_cc_filter(cspmu, event); |
| 837 |
} else { |
| 838 |
arm_cspmu_set_event(cspmu, hwc); |
| 839 |
cspmu->impl.ops.set_ev_filter(cspmu, event); |
| 840 |
} |
| 841 |
|
| 842 |
hwc->state = 0; |
| 843 |
|
| 844 |
arm_cspmu_enable_counter(cspmu, hwc->idx); |
| 845 |
} |
| 846 |
|
| 847 |
static void arm_cspmu_stop(struct perf_event *event, int pmu_flags) |
| 848 |
{ |
| 849 |
struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu); |
| 850 |
struct hw_perf_event *hwc = &event->hw; |
| 851 |
|
| 852 |
if (hwc->state & PERF_HES_STOPPED) |
| 853 |
return; |
| 854 |
|
| 855 |
arm_cspmu_disable_counter(cspmu, hwc->idx); |
| 856 |
|
| 857 |
if (cspmu->impl.ops.reset_ev_filter) |
| 858 |
cspmu->impl.ops.reset_ev_filter(cspmu, event); |
| 859 |
|
| 860 |
arm_cspmu_event_update(event); |
| 861 |
|
| 862 |
hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE; |
| 863 |
} |
| 864 |
|
| 865 |
static inline u32 to_phys_idx(struct arm_cspmu *cspmu, u32 idx) |
| 866 |
{ |
| 867 |
return (idx == cspmu->cycle_counter_logical_idx) ? |
| 868 |
ARM_CSPMU_CYCLE_CNTR_IDX : idx; |
| 869 |
} |
| 870 |
|
| 871 |
static int arm_cspmu_add(struct perf_event *event, int flags) |
| 872 |
{ |
| 873 |
struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu); |
| 874 |
struct arm_cspmu_hw_events *hw_events = &cspmu->hw_events; |
| 875 |
struct hw_perf_event *hwc = &event->hw; |
| 876 |
int idx; |
| 877 |
|
| 878 |
if (WARN_ON_ONCE(!cpumask_test_cpu(smp_processor_id(), |
| 879 |
&cspmu->associated_cpus))) |
| 880 |
return -ENOENT; |
| 881 |
|
| 882 |
idx = arm_cspmu_get_event_idx(hw_events, event); |
| 883 |
if (idx < 0) |
| 884 |
return idx; |
| 885 |
|
| 886 |
hw_events->events[idx] = event; |
| 887 |
hwc->idx = to_phys_idx(cspmu, idx); |
| 888 |
hwc->extra_reg.idx = idx; |
| 889 |
hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE; |
| 890 |
|
| 891 |
if (flags & PERF_EF_START) |
| 892 |
arm_cspmu_start(event, PERF_EF_RELOAD); |
| 893 |
|
| 894 |
/* Propagate changes to the userspace mapping. */ |
| 895 |
perf_event_update_userpage(event); |
| 896 |
|
| 897 |
return 0; |
| 898 |
} |
| 899 |
|
| 900 |
static void arm_cspmu_del(struct perf_event *event, int flags) |
| 901 |
{ |
| 902 |
struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu); |
| 903 |
struct arm_cspmu_hw_events *hw_events = &cspmu->hw_events; |
| 904 |
struct hw_perf_event *hwc = &event->hw; |
| 905 |
int idx = hwc->extra_reg.idx; |
| 906 |
|
| 907 |
arm_cspmu_stop(event, PERF_EF_UPDATE); |
| 908 |
|
| 909 |
hw_events->events[idx] = NULL; |
| 910 |
|
| 911 |
clear_bit(idx, hw_events->used_ctrs); |
| 912 |
|
| 913 |
perf_event_update_userpage(event); |
| 914 |
} |
| 915 |
|
| 916 |
static void arm_cspmu_read(struct perf_event *event) |
| 917 |
{ |
| 918 |
arm_cspmu_event_update(event); |
| 919 |
} |
| 920 |
|
| 921 |
static struct arm_cspmu *arm_cspmu_alloc(struct platform_device *pdev) |
| 922 |
{ |
| 923 |
struct acpi_apmt_node *apmt_node; |
| 924 |
struct arm_cspmu *cspmu; |
| 925 |
struct device *dev = &pdev->dev; |
| 926 |
|
| 927 |
cspmu = devm_kzalloc(dev, sizeof(*cspmu), GFP_KERNEL); |
| 928 |
if (!cspmu) |
| 929 |
return NULL; |
| 930 |
|
| 931 |
cspmu->dev = dev; |
| 932 |
platform_set_drvdata(pdev, cspmu); |
| 933 |
|
| 934 |
apmt_node = arm_cspmu_apmt_node(dev); |
| 935 |
if (apmt_node) { |
| 936 |
cspmu->has_atomic_dword = apmt_node->flags & ACPI_APMT_FLAGS_ATOMIC; |
| 937 |
} else { |
| 938 |
u32 width = 0; |
| 939 |
|
| 940 |
device_property_read_u32(dev, "reg-io-width", &width); |
| 941 |
cspmu->has_atomic_dword = (width == 8); |
| 942 |
} |
| 943 |
|
| 944 |
return cspmu; |
| 945 |
} |
| 946 |
|
| 947 |
static int arm_cspmu_init_mmio(struct arm_cspmu *cspmu) |
| 948 |
{ |
| 949 |
struct device *dev; |
| 950 |
struct platform_device *pdev; |
| 951 |
|
| 952 |
dev = cspmu->dev; |
| 953 |
pdev = to_platform_device(dev); |
| 954 |
|
| 955 |
/* Base address for page 0. */ |
| 956 |
cspmu->base0 = devm_platform_ioremap_resource(pdev, 0); |
| 957 |
if (IS_ERR(cspmu->base0)) { |
| 958 |
dev_err(dev, "ioremap failed for page-0 resource\n"); |
| 959 |
return PTR_ERR(cspmu->base0); |
| 960 |
} |
| 961 |
|
| 962 |
/* Base address for page 1 if supported. Otherwise point to page 0. */ |
| 963 |
cspmu->base1 = cspmu->base0; |
| 964 |
if (platform_get_resource(pdev, IORESOURCE_MEM, 1)) { |
| 965 |
cspmu->base1 = devm_platform_ioremap_resource(pdev, 1); |
| 966 |
if (IS_ERR(cspmu->base1)) { |
| 967 |
dev_err(dev, "ioremap failed for page-1 resource\n"); |
| 968 |
return PTR_ERR(cspmu->base1); |
| 969 |
} |
| 970 |
} |
| 971 |
|
| 972 |
cspmu->pmcfgr = readl(cspmu->base0 + PMCFGR); |
| 973 |
|
| 974 |
cspmu->num_logical_ctrs = FIELD_GET(PMCFGR_N, cspmu->pmcfgr) + 1; |
| 975 |
|
| 976 |
cspmu->cycle_counter_logical_idx = ARM_CSPMU_MAX_HW_CNTRS; |
| 977 |
|
| 978 |
if (supports_cycle_counter(cspmu)) { |
| 979 |
/* |
| 980 |
* The last logical counter is mapped to cycle counter if |
| 981 |
* there is a gap between regular and cycle counter. Otherwise, |
| 982 |
* logical and physical have 1-to-1 mapping. |
| 983 |
*/ |
| 984 |
cspmu->cycle_counter_logical_idx = |
| 985 |
(cspmu->num_logical_ctrs <= ARM_CSPMU_CYCLE_CNTR_IDX) ? |
| 986 |
cspmu->num_logical_ctrs - 1 : |
| 987 |
ARM_CSPMU_CYCLE_CNTR_IDX; |
| 988 |
} |
| 989 |
|
| 990 |
cspmu->num_set_clr_reg = |
| 991 |
DIV_ROUND_UP(cspmu->num_logical_ctrs, |
| 992 |
ARM_CSPMU_SET_CLR_COUNTER_NUM); |
| 993 |
|
| 994 |
cspmu->hw_events.events = |
| 995 |
devm_kcalloc(dev, cspmu->num_logical_ctrs, |
| 996 |
sizeof(*cspmu->hw_events.events), GFP_KERNEL); |
| 997 |
|
| 998 |
if (!cspmu->hw_events.events) |
| 999 |
return -ENOMEM; |
| 1000 |
|
| 1001 |
return 0; |
| 1002 |
} |
| 1003 |
|
| 1004 |
static inline int arm_cspmu_get_reset_overflow(struct arm_cspmu *cspmu, |
| 1005 |
u32 *pmovs) |
| 1006 |
{ |
| 1007 |
int i; |
| 1008 |
u32 pmovclr_offset = PMOVSCLR; |
| 1009 |
u32 has_overflowed = 0; |
| 1010 |
|
| 1011 |
for (i = 0; i < cspmu->num_set_clr_reg; ++i) { |
| 1012 |
pmovs[i] = readl(cspmu->base1 + pmovclr_offset); |
| 1013 |
has_overflowed |= pmovs[i]; |
| 1014 |
writel(pmovs[i], cspmu->base1 + pmovclr_offset); |
| 1015 |
pmovclr_offset += sizeof(u32); |
| 1016 |
} |
| 1017 |
|
| 1018 |
return has_overflowed != 0; |
| 1019 |
} |
| 1020 |
|
| 1021 |
static irqreturn_t arm_cspmu_handle_irq(int irq_num, void *dev) |
| 1022 |
{ |
| 1023 |
int idx, has_overflowed; |
| 1024 |
struct perf_event *event; |
| 1025 |
struct arm_cspmu *cspmu = dev; |
| 1026 |
DECLARE_BITMAP(pmovs, ARM_CSPMU_MAX_HW_CNTRS); |
| 1027 |
bool handled = false; |
| 1028 |
|
| 1029 |
arm_cspmu_stop_counters(cspmu); |
| 1030 |
|
| 1031 |
has_overflowed = arm_cspmu_get_reset_overflow(cspmu, (u32 *)pmovs); |
| 1032 |
if (!has_overflowed) |
| 1033 |
goto done; |
| 1034 |
|
| 1035 |
for_each_set_bit(idx, cspmu->hw_events.used_ctrs, |
| 1036 |
cspmu->num_logical_ctrs) { |
| 1037 |
event = cspmu->hw_events.events[idx]; |
| 1038 |
|
| 1039 |
if (!event) |
| 1040 |
continue; |
| 1041 |
|
| 1042 |
if (!test_bit(event->hw.idx, pmovs)) |
| 1043 |
continue; |
| 1044 |
|
| 1045 |
arm_cspmu_event_update(event); |
| 1046 |
arm_cspmu_set_event_period(event); |
| 1047 |
|
| 1048 |
handled = true; |
| 1049 |
} |
| 1050 |
|
| 1051 |
done: |
| 1052 |
arm_cspmu_start_counters(cspmu); |
| 1053 |
return IRQ_RETVAL(handled); |
| 1054 |
} |
| 1055 |
|
| 1056 |
static int arm_cspmu_request_irq(struct arm_cspmu *cspmu) |
| 1057 |
{ |
| 1058 |
int irq, ret; |
| 1059 |
struct device *dev; |
| 1060 |
struct platform_device *pdev; |
| 1061 |
|
| 1062 |
dev = cspmu->dev; |
| 1063 |
pdev = to_platform_device(dev); |
| 1064 |
|
| 1065 |
/* Skip IRQ request if the PMU does not support overflow interrupt. */ |
| 1066 |
irq = platform_get_irq_optional(pdev, 0); |
| 1067 |
if (irq < 0) |
| 1068 |
return irq == -ENXIO ? 0 : irq; |
| 1069 |
|
| 1070 |
ret = devm_request_irq(dev, irq, arm_cspmu_handle_irq, |
| 1071 |
IRQF_NOBALANCING | IRQF_NO_THREAD, dev_name(dev), |
| 1072 |
cspmu); |
| 1073 |
if (ret) { |
| 1074 |
dev_err(dev, "Could not request IRQ %d\n", irq); |
| 1075 |
return ret; |
| 1076 |
} |
| 1077 |
|
| 1078 |
cspmu->irq = irq; |
| 1079 |
|
| 1080 |
return 0; |
| 1081 |
} |
| 1082 |
|
| 1083 |
#if defined(CONFIG_ACPI) && defined(CONFIG_ARM64) |
| 1084 |
#include <acpi/processor.h> |
| 1085 |
|
| 1086 |
static inline int arm_cspmu_find_cpu_container(int cpu, u32 container_uid) |
| 1087 |
{ |
| 1088 |
struct device *cpu_dev; |
| 1089 |
struct acpi_device *acpi_dev; |
| 1090 |
|
| 1091 |
cpu_dev = get_cpu_device(cpu); |
| 1092 |
if (!cpu_dev) |
| 1093 |
return -ENODEV; |
| 1094 |
|
| 1095 |
acpi_dev = ACPI_COMPANION(cpu_dev); |
| 1096 |
while (acpi_dev) { |
| 1097 |
if (acpi_dev_hid_uid_match(acpi_dev, ACPI_PROCESSOR_CONTAINER_HID, container_uid)) |
| 1098 |
return 0; |
| 1099 |
|
| 1100 |
acpi_dev = acpi_dev_parent(acpi_dev); |
| 1101 |
} |
| 1102 |
|
| 1103 |
return -ENODEV; |
| 1104 |
} |
| 1105 |
|
| 1106 |
static int arm_cspmu_acpi_get_cpus(struct arm_cspmu *cspmu) |
| 1107 |
{ |
| 1108 |
struct acpi_apmt_node *apmt_node; |
| 1109 |
int affinity_flag; |
| 1110 |
int cpu; |
| 1111 |
|
| 1112 |
apmt_node = arm_cspmu_apmt_node(cspmu->dev); |
| 1113 |
affinity_flag = apmt_node->flags & ACPI_APMT_FLAGS_AFFINITY; |
| 1114 |
|
| 1115 |
if (affinity_flag == ACPI_APMT_FLAGS_AFFINITY_PROC) { |
| 1116 |
for_each_possible_cpu(cpu) { |
| 1117 |
if (apmt_node->proc_affinity == |
| 1118 |
get_acpi_id_for_cpu(cpu)) { |
| 1119 |
cpumask_set_cpu(cpu, &cspmu->associated_cpus); |
| 1120 |
break; |
| 1121 |
} |
| 1122 |
} |
| 1123 |
} else { |
| 1124 |
for_each_possible_cpu(cpu) { |
| 1125 |
if (arm_cspmu_find_cpu_container( |
| 1126 |
cpu, apmt_node->proc_affinity)) |
| 1127 |
continue; |
| 1128 |
|
| 1129 |
cpumask_set_cpu(cpu, &cspmu->associated_cpus); |
| 1130 |
} |
| 1131 |
} |
| 1132 |
|
| 1133 |
return 0; |
| 1134 |
} |
| 1135 |
#else |
| 1136 |
static int arm_cspmu_acpi_get_cpus(struct arm_cspmu *cspmu) |
| 1137 |
{ |
| 1138 |
return -ENODEV; |
| 1139 |
} |
| 1140 |
#endif |
| 1141 |
|
| 1142 |
static int arm_cspmu_of_get_cpus(struct arm_cspmu *cspmu) |
| 1143 |
{ |
| 1144 |
struct of_phandle_iterator it; |
| 1145 |
int ret, cpu; |
| 1146 |
|
| 1147 |
of_for_each_phandle(&it, ret, dev_of_node(cspmu->dev), "cpus", NULL, 0) { |
| 1148 |
cpu = of_cpu_node_to_id(it.node); |
| 1149 |
if (cpu < 0) |
| 1150 |
continue; |
| 1151 |
cpumask_set_cpu(cpu, &cspmu->associated_cpus); |
| 1152 |
} |
| 1153 |
return ret == -ENOENT ? 0 : ret; |
| 1154 |
} |
| 1155 |
|
| 1156 |
static int arm_cspmu_get_cpus(struct arm_cspmu *cspmu) |
| 1157 |
{ |
| 1158 |
int ret = 0; |
| 1159 |
|
| 1160 |
if (arm_cspmu_apmt_node(cspmu->dev)) |
| 1161 |
ret = arm_cspmu_acpi_get_cpus(cspmu); |
| 1162 |
else if (device_property_present(cspmu->dev, "cpus")) |
| 1163 |
ret = arm_cspmu_of_get_cpus(cspmu); |
| 1164 |
else |
| 1165 |
cpumask_copy(&cspmu->associated_cpus, cpu_possible_mask); |
| 1166 |
|
| 1167 |
if (!ret && cpumask_empty(&cspmu->associated_cpus)) { |
| 1168 |
dev_dbg(cspmu->dev, "No cpu associated with the PMU\n"); |
| 1169 |
ret = -ENODEV; |
| 1170 |
} |
| 1171 |
return ret; |
| 1172 |
} |
| 1173 |
|
| 1174 |
static int arm_cspmu_register_pmu(struct arm_cspmu *cspmu) |
| 1175 |
{ |
| 1176 |
int ret, capabilities; |
| 1177 |
|
| 1178 |
ret = arm_cspmu_alloc_attr_groups(cspmu); |
| 1179 |
if (ret) |
| 1180 |
return ret; |
| 1181 |
|
| 1182 |
ret = cpuhp_state_add_instance(arm_cspmu_cpuhp_state, |
| 1183 |
&cspmu->cpuhp_node); |
| 1184 |
if (ret) |
| 1185 |
return ret; |
| 1186 |
|
| 1187 |
capabilities = PERF_PMU_CAP_NO_EXCLUDE; |
| 1188 |
if (cspmu->irq == 0) |
| 1189 |
capabilities |= PERF_PMU_CAP_NO_INTERRUPT; |
| 1190 |
|
| 1191 |
cspmu->pmu = (struct pmu){ |
| 1192 |
.task_ctx_nr = perf_invalid_context, |
| 1193 |
.module = cspmu->impl.module, |
| 1194 |
.parent = cspmu->dev, |
| 1195 |
.pmu_enable = arm_cspmu_enable, |
| 1196 |
.pmu_disable = arm_cspmu_disable, |
| 1197 |
.event_init = arm_cspmu_event_init, |
| 1198 |
.add = arm_cspmu_add, |
| 1199 |
.del = arm_cspmu_del, |
| 1200 |
.start = arm_cspmu_start, |
| 1201 |
.stop = arm_cspmu_stop, |
| 1202 |
.read = arm_cspmu_read, |
| 1203 |
.attr_groups = cspmu->attr_groups, |
| 1204 |
.capabilities = capabilities, |
| 1205 |
}; |
| 1206 |
|
| 1207 |
/* Hardware counter init */ |
| 1208 |
arm_cspmu_reset_counters(cspmu); |
| 1209 |
|
| 1210 |
ret = perf_pmu_register(&cspmu->pmu, cspmu->name, -1); |
| 1211 |
if (ret) { |
| 1212 |
cpuhp_state_remove_instance(arm_cspmu_cpuhp_state, |
| 1213 |
&cspmu->cpuhp_node); |
| 1214 |
} |
| 1215 |
|
| 1216 |
return ret; |
| 1217 |
} |
| 1218 |
|
| 1219 |
static int arm_cspmu_device_probe(struct platform_device *pdev) |
| 1220 |
{ |
| 1221 |
int ret; |
| 1222 |
struct arm_cspmu *cspmu; |
| 1223 |
|
| 1224 |
cspmu = arm_cspmu_alloc(pdev); |
| 1225 |
if (!cspmu) |
| 1226 |
return -ENOMEM; |
| 1227 |
|
| 1228 |
ret = arm_cspmu_init_mmio(cspmu); |
| 1229 |
if (ret) |
| 1230 |
return ret; |
| 1231 |
|
| 1232 |
ret = arm_cspmu_request_irq(cspmu); |
| 1233 |
if (ret) |
| 1234 |
return ret; |
| 1235 |
|
| 1236 |
ret = arm_cspmu_get_cpus(cspmu); |
| 1237 |
if (ret) |
| 1238 |
return ret; |
| 1239 |
|
| 1240 |
ret = arm_cspmu_init_impl_ops(cspmu); |
| 1241 |
if (ret) |
| 1242 |
return ret; |
| 1243 |
|
| 1244 |
ret = arm_cspmu_register_pmu(cspmu); |
| 1245 |
|
| 1246 |
/* Matches arm_cspmu_init_impl_ops() above. */ |
| 1247 |
if (cspmu->impl.module != THIS_MODULE) |
| 1248 |
module_put(cspmu->impl.module); |
| 1249 |
|
| 1250 |
return ret; |
| 1251 |
} |
| 1252 |
|
| 1253 |
static void arm_cspmu_device_remove(struct platform_device *pdev) |
| 1254 |
{ |
| 1255 |
struct arm_cspmu *cspmu = platform_get_drvdata(pdev); |
| 1256 |
|
| 1257 |
perf_pmu_unregister(&cspmu->pmu); |
| 1258 |
cpuhp_state_remove_instance(arm_cspmu_cpuhp_state, &cspmu->cpuhp_node); |
| 1259 |
} |
| 1260 |
|
| 1261 |
static const struct platform_device_id arm_cspmu_id[] = { |
| 1262 |
{DRVNAME, 0}, |
| 1263 |
{ }, |
| 1264 |
}; |
| 1265 |
MODULE_DEVICE_TABLE(platform, arm_cspmu_id); |
| 1266 |
|
| 1267 |
static const struct of_device_id arm_cspmu_of_match[] = { |
| 1268 |
{ .compatible = "arm,coresight-pmu" }, |
| 1269 |
{} |
| 1270 |
}; |
| 1271 |
MODULE_DEVICE_TABLE(of, arm_cspmu_of_match); |
| 1272 |
|
| 1273 |
static struct platform_driver arm_cspmu_driver = { |
| 1274 |
.driver = { |
| 1275 |
.name = DRVNAME, |
| 1276 |
.of_match_table = arm_cspmu_of_match, |
| 1277 |
.suppress_bind_attrs = true, |
| 1278 |
}, |
| 1279 |
.probe = arm_cspmu_device_probe, |
| 1280 |
.remove = arm_cspmu_device_remove, |
| 1281 |
.id_table = arm_cspmu_id, |
| 1282 |
}; |
| 1283 |
|
| 1284 |
static void arm_cspmu_set_active_cpu(int cpu, struct arm_cspmu *cspmu) |
| 1285 |
{ |
| 1286 |
cpumask_set_cpu(cpu, &cspmu->active_cpu); |
| 1287 |
if (cspmu->irq) |
| 1288 |
WARN_ON(irq_set_affinity(cspmu->irq, &cspmu->active_cpu)); |
| 1289 |
} |
| 1290 |
|
| 1291 |
static int arm_cspmu_cpu_online(unsigned int cpu, struct hlist_node *node) |
| 1292 |
{ |
| 1293 |
struct arm_cspmu *cspmu = |
| 1294 |
hlist_entry_safe(node, struct arm_cspmu, cpuhp_node); |
| 1295 |
|
| 1296 |
if (!cpumask_test_cpu(cpu, &cspmu->associated_cpus)) |
| 1297 |
return 0; |
| 1298 |
|
| 1299 |
/* If the PMU is already managed, there is nothing to do */ |
| 1300 |
if (!cpumask_empty(&cspmu->active_cpu)) |
| 1301 |
return 0; |
| 1302 |
|
| 1303 |
/* Use this CPU for event counting */ |
| 1304 |
arm_cspmu_set_active_cpu(cpu, cspmu); |
| 1305 |
|
| 1306 |
return 0; |
| 1307 |
} |
| 1308 |
|
| 1309 |
static int arm_cspmu_cpu_teardown(unsigned int cpu, struct hlist_node *node) |
| 1310 |
{ |
| 1311 |
unsigned int dst; |
| 1312 |
|
| 1313 |
struct arm_cspmu *cspmu = |
| 1314 |
hlist_entry_safe(node, struct arm_cspmu, cpuhp_node); |
| 1315 |
|
| 1316 |
/* Nothing to do if this CPU doesn't own the PMU */ |
| 1317 |
if (!cpumask_test_and_clear_cpu(cpu, &cspmu->active_cpu)) |
| 1318 |
return 0; |
| 1319 |
|
| 1320 |
/* Choose a new CPU to migrate ownership of the PMU to */ |
| 1321 |
dst = cpumask_any_and_but(&cspmu->associated_cpus, |
| 1322 |
cpu_online_mask, cpu); |
| 1323 |
if (dst >= nr_cpu_ids) |
| 1324 |
return 0; |
| 1325 |
|
| 1326 |
/* Use this CPU for event counting */ |
| 1327 |
perf_pmu_migrate_context(&cspmu->pmu, cpu, dst); |
| 1328 |
arm_cspmu_set_active_cpu(dst, cspmu); |
| 1329 |
|
| 1330 |
return 0; |
| 1331 |
} |
| 1332 |
|
| 1333 |
static int __init arm_cspmu_init(void) |
| 1334 |
{ |
| 1335 |
int ret; |
| 1336 |
|
| 1337 |
ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, |
| 1338 |
"perf/arm/cspmu:online", |
| 1339 |
arm_cspmu_cpu_online, |
| 1340 |
arm_cspmu_cpu_teardown); |
| 1341 |
if (ret < 0) |
| 1342 |
return ret; |
| 1343 |
arm_cspmu_cpuhp_state = ret; |
| 1344 |
return platform_driver_register(&arm_cspmu_driver); |
| 1345 |
} |
| 1346 |
|
| 1347 |
static void __exit arm_cspmu_exit(void) |
| 1348 |
{ |
| 1349 |
platform_driver_unregister(&arm_cspmu_driver); |
| 1350 |
cpuhp_remove_multi_state(arm_cspmu_cpuhp_state); |
| 1351 |
} |
| 1352 |
|
| 1353 |
int arm_cspmu_impl_register(const struct arm_cspmu_impl_match *impl_match) |
| 1354 |
{ |
| 1355 |
struct arm_cspmu_impl_match *match; |
| 1356 |
int ret = 0; |
| 1357 |
|
| 1358 |
match = arm_cspmu_impl_match_get(impl_match->pmiidr_val); |
| 1359 |
|
| 1360 |
if (match) { |
| 1361 |
mutex_lock(&arm_cspmu_lock); |
| 1362 |
|
| 1363 |
if (!match->impl_init_ops) { |
| 1364 |
match->module = impl_match->module; |
| 1365 |
match->impl_init_ops = impl_match->impl_init_ops; |
| 1366 |
} else { |
| 1367 |
/* Broken match table may contain non-unique entries */ |
| 1368 |
WARN(1, "arm_cspmu backend already registered for module: %s, pmiidr: 0x%x, mask: 0x%x\n", |
| 1369 |
match->module_name, |
| 1370 |
match->pmiidr_val, |
| 1371 |
match->pmiidr_mask); |
| 1372 |
|
| 1373 |
ret = -EINVAL; |
| 1374 |
} |
| 1375 |
|
| 1376 |
mutex_unlock(&arm_cspmu_lock); |
| 1377 |
|
| 1378 |
if (!ret) |
| 1379 |
ret = driver_attach(&arm_cspmu_driver.driver); |
| 1380 |
} else { |
| 1381 |
pr_err("arm_cspmu reg failed, unable to find a match for pmiidr: 0x%x\n", |
| 1382 |
impl_match->pmiidr_val); |
| 1383 |
|
| 1384 |
ret = -EINVAL; |
| 1385 |
} |
| 1386 |
|
| 1387 |
return ret; |
| 1388 |
} |
| 1389 |
EXPORT_SYMBOL_GPL(arm_cspmu_impl_register); |
| 1390 |
|
| 1391 |
static int arm_cspmu_match_device(struct device *dev, const void *match) |
| 1392 |
{ |
| 1393 |
struct arm_cspmu *cspmu = platform_get_drvdata(to_platform_device(dev)); |
| 1394 |
|
| 1395 |
return (cspmu && cspmu->impl.match == match) ? 1 : 0; |
| 1396 |
} |
| 1397 |
|
| 1398 |
void arm_cspmu_impl_unregister(const struct arm_cspmu_impl_match *impl_match) |
| 1399 |
{ |
| 1400 |
struct device *dev; |
| 1401 |
struct arm_cspmu_impl_match *match; |
| 1402 |
|
| 1403 |
match = arm_cspmu_impl_match_get(impl_match->pmiidr_val); |
| 1404 |
|
| 1405 |
if (WARN_ON(!match)) |
| 1406 |
return; |
| 1407 |
|
| 1408 |
/* Unbind the driver from all matching backend devices. */ |
| 1409 |
while ((dev = driver_find_device(&arm_cspmu_driver.driver, NULL, |
| 1410 |
match, arm_cspmu_match_device))) { |
| 1411 |
device_release_driver(dev); |
| 1412 |
put_device(dev); |
| 1413 |
} |
| 1414 |
|
| 1415 |
mutex_lock(&arm_cspmu_lock); |
| 1416 |
|
| 1417 |
match->module = NULL; |
| 1418 |
match->impl_init_ops = NULL; |
| 1419 |
|
| 1420 |
mutex_unlock(&arm_cspmu_lock); |
| 1421 |
} |
| 1422 |
EXPORT_SYMBOL_GPL(arm_cspmu_impl_unregister); |
| 1423 |
|
| 1424 |
module_init(arm_cspmu_init); |
| 1425 |
module_exit(arm_cspmu_exit); |
| 1426 |
|
| 1427 |
MODULE_DESCRIPTION("ARM CoreSight Architecture Performance Monitor Driver"); |
| 1428 |
MODULE_LICENSE("GPL v2"); |