| 1 | /* SPDX-License-Identifier: (GPL-2.0-only OR LGPL-2.1-only) |
| 2 | * |
| 3 | * counter.c |
| 4 | * |
| 5 | * Copyright (C) 2020 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 6 | */ |
| 7 | |
| 8 | #include <linux/module.h> |
| 9 | #include <linux/slab.h> |
| 10 | #include <lttng/tracer.h> |
| 11 | #include <linux/cpumask.h> |
| 12 | #include <counter/counter.h> |
| 13 | #include <counter/counter-internal.h> |
| 14 | #include <wrapper/vmalloc.h> |
| 15 | #include <wrapper/limits.h> |
| 16 | |
| 17 | static size_t lttng_counter_get_dimension_nr_elements(struct lib_counter_dimension *dimension) |
| 18 | { |
| 19 | return dimension->max_nr_elem; |
| 20 | } |
| 21 | |
| 22 | static int lttng_counter_init_stride(const struct lib_counter_config *config, |
| 23 | struct lib_counter *counter) |
| 24 | { |
| 25 | size_t nr_dimensions = counter->nr_dimensions; |
| 26 | size_t stride = 1; |
| 27 | ssize_t i; |
| 28 | |
| 29 | for (i = nr_dimensions - 1; i >= 0; i--) { |
| 30 | struct lib_counter_dimension *dimension = &counter->dimensions[i]; |
| 31 | size_t nr_elem; |
| 32 | |
| 33 | nr_elem = lttng_counter_get_dimension_nr_elements(dimension); |
| 34 | dimension->stride = stride; |
| 35 | /* nr_elem should be minimum 1 for each dimension. */ |
| 36 | if (!nr_elem) |
| 37 | return -EINVAL; |
| 38 | stride *= nr_elem; |
| 39 | if (stride > SIZE_MAX / nr_elem) |
| 40 | return -EINVAL; |
| 41 | } |
| 42 | return 0; |
| 43 | } |
| 44 | |
| 45 | static int lttng_counter_layout_init(struct lib_counter *counter, int cpu) |
| 46 | { |
| 47 | struct lib_counter_layout *layout; |
| 48 | size_t counter_size; |
| 49 | size_t nr_elem = counter->allocated_elem; |
| 50 | |
| 51 | if (cpu == -1) |
| 52 | layout = &counter->global_counters; |
| 53 | else |
| 54 | layout = per_cpu_ptr(counter->percpu_counters, cpu); |
| 55 | switch (counter->config.counter_size) { |
| 56 | case COUNTER_SIZE_8_BIT: |
| 57 | case COUNTER_SIZE_16_BIT: |
| 58 | case COUNTER_SIZE_32_BIT: |
| 59 | case COUNTER_SIZE_64_BIT: |
| 60 | counter_size = (size_t) counter->config.counter_size; |
| 61 | break; |
| 62 | default: |
| 63 | return -EINVAL; |
| 64 | } |
| 65 | layout->counters = lttng_kvzalloc_node(ALIGN(counter_size * nr_elem, |
| 66 | 1 << INTERNODE_CACHE_SHIFT), |
| 67 | GFP_KERNEL | __GFP_NOWARN, |
| 68 | cpu_to_node(max(cpu, 0))); |
| 69 | if (!layout->counters) |
| 70 | return -ENOMEM; |
| 71 | layout->overflow_bitmap = lttng_kvzalloc_node(ALIGN(ALIGN(nr_elem, 8) / 8, |
| 72 | 1 << INTERNODE_CACHE_SHIFT), |
| 73 | GFP_KERNEL | __GFP_NOWARN, |
| 74 | cpu_to_node(max(cpu, 0))); |
| 75 | if (!layout->overflow_bitmap) |
| 76 | return -ENOMEM; |
| 77 | layout->underflow_bitmap = lttng_kvzalloc_node(ALIGN(ALIGN(nr_elem, 8) / 8, |
| 78 | 1 << INTERNODE_CACHE_SHIFT), |
| 79 | GFP_KERNEL | __GFP_NOWARN, |
| 80 | cpu_to_node(max(cpu, 0))); |
| 81 | if (!layout->underflow_bitmap) |
| 82 | return -ENOMEM; |
| 83 | return 0; |
| 84 | } |
| 85 | |
| 86 | static void lttng_counter_layout_fini(struct lib_counter *counter, int cpu) |
| 87 | { |
| 88 | struct lib_counter_layout *layout; |
| 89 | |
| 90 | if (cpu == -1) |
| 91 | layout = &counter->global_counters; |
| 92 | else |
| 93 | layout = per_cpu_ptr(counter->percpu_counters, cpu); |
| 94 | |
| 95 | lttng_kvfree(layout->counters); |
| 96 | lttng_kvfree(layout->overflow_bitmap); |
| 97 | lttng_kvfree(layout->underflow_bitmap); |
| 98 | } |
| 99 | |
| 100 | static |
| 101 | int lttng_counter_set_global_sum_step(struct lib_counter *counter, |
| 102 | int64_t global_sum_step) |
| 103 | { |
| 104 | if (global_sum_step < 0) |
| 105 | return -EINVAL; |
| 106 | |
| 107 | switch (counter->config.counter_size) { |
| 108 | case COUNTER_SIZE_8_BIT: |
| 109 | if (global_sum_step > S8_MAX) |
| 110 | return -EINVAL; |
| 111 | counter->global_sum_step.s8 = (int8_t) global_sum_step; |
| 112 | break; |
| 113 | case COUNTER_SIZE_16_BIT: |
| 114 | if (global_sum_step > S16_MAX) |
| 115 | return -EINVAL; |
| 116 | counter->global_sum_step.s16 = (int16_t) global_sum_step; |
| 117 | break; |
| 118 | case COUNTER_SIZE_32_BIT: |
| 119 | if (global_sum_step > S32_MAX) |
| 120 | return -EINVAL; |
| 121 | counter->global_sum_step.s32 = (int32_t) global_sum_step; |
| 122 | break; |
| 123 | case COUNTER_SIZE_64_BIT: |
| 124 | counter->global_sum_step.s64 = global_sum_step; |
| 125 | break; |
| 126 | default: |
| 127 | return -EINVAL; |
| 128 | } |
| 129 | |
| 130 | return 0; |
| 131 | } |
| 132 | |
| 133 | static |
| 134 | int validate_args(const struct lib_counter_config *config, |
| 135 | size_t nr_dimensions, |
| 136 | const size_t *max_nr_elem, |
| 137 | int64_t global_sum_step) |
| 138 | { |
| 139 | if (BITS_PER_LONG != 64 && config->counter_size == COUNTER_SIZE_64_BIT) { |
| 140 | WARN_ON_ONCE(1); |
| 141 | return -1; |
| 142 | } |
| 143 | if (!max_nr_elem) |
| 144 | return -1; |
| 145 | /* |
| 146 | * global sum step is only useful with allocating both per-cpu |
| 147 | * and global counters. |
| 148 | */ |
| 149 | if (global_sum_step && (!(config->alloc & COUNTER_ALLOC_GLOBAL) || |
| 150 | !(config->alloc & COUNTER_ALLOC_PER_CPU))) |
| 151 | return -1; |
| 152 | return 0; |
| 153 | } |
| 154 | |
| 155 | struct lib_counter *lttng_counter_create(const struct lib_counter_config *config, |
| 156 | size_t nr_dimensions, |
| 157 | const size_t *max_nr_elem, |
| 158 | int64_t global_sum_step) |
| 159 | { |
| 160 | struct lib_counter *counter; |
| 161 | size_t dimension, nr_elem = 1; |
| 162 | int cpu, ret; |
| 163 | |
| 164 | if (validate_args(config, nr_dimensions, max_nr_elem, global_sum_step)) |
| 165 | return NULL; |
| 166 | counter = kzalloc(sizeof(struct lib_counter), GFP_KERNEL); |
| 167 | if (!counter) |
| 168 | return NULL; |
| 169 | counter->config = *config; |
| 170 | if (lttng_counter_set_global_sum_step(counter, global_sum_step)) |
| 171 | goto error_sum_step; |
| 172 | counter->nr_dimensions = nr_dimensions; |
| 173 | counter->dimensions = kzalloc(nr_dimensions * sizeof(*counter->dimensions), GFP_KERNEL); |
| 174 | if (!counter->dimensions) |
| 175 | goto error_dimensions; |
| 176 | for (dimension = 0; dimension < nr_dimensions; dimension++) |
| 177 | counter->dimensions[dimension].max_nr_elem = max_nr_elem[dimension]; |
| 178 | if (config->alloc & COUNTER_ALLOC_PER_CPU) { |
| 179 | counter->percpu_counters = alloc_percpu(struct lib_counter_layout); |
| 180 | if (!counter->percpu_counters) |
| 181 | goto error_alloc_percpu; |
| 182 | } |
| 183 | |
| 184 | if (lttng_counter_init_stride(config, counter)) |
| 185 | goto error_init_stride; |
| 186 | //TODO saturation values. |
| 187 | for (dimension = 0; dimension < counter->nr_dimensions; dimension++) |
| 188 | nr_elem *= lttng_counter_get_dimension_nr_elements(&counter->dimensions[dimension]); |
| 189 | counter->allocated_elem = nr_elem; |
| 190 | if (config->alloc & COUNTER_ALLOC_GLOBAL) { |
| 191 | ret = lttng_counter_layout_init(counter, -1); /* global */ |
| 192 | if (ret) |
| 193 | goto layout_init_error; |
| 194 | } |
| 195 | if (config->alloc & COUNTER_ALLOC_PER_CPU) { |
| 196 | //TODO: integrate with CPU hotplug and online cpus |
| 197 | for (cpu = 0; cpu < num_possible_cpus(); cpu++) { |
| 198 | ret = lttng_counter_layout_init(counter, cpu); |
| 199 | if (ret) |
| 200 | goto layout_init_error; |
| 201 | } |
| 202 | } |
| 203 | return counter; |
| 204 | |
| 205 | layout_init_error: |
| 206 | if (config->alloc & COUNTER_ALLOC_PER_CPU) { |
| 207 | for (cpu = 0; cpu < num_possible_cpus(); cpu++) |
| 208 | lttng_counter_layout_fini(counter, cpu); |
| 209 | } |
| 210 | if (config->alloc & COUNTER_ALLOC_GLOBAL) |
| 211 | lttng_counter_layout_fini(counter, -1); |
| 212 | error_init_stride: |
| 213 | free_percpu(counter->percpu_counters); |
| 214 | error_alloc_percpu: |
| 215 | kfree(counter->dimensions); |
| 216 | error_dimensions: |
| 217 | error_sum_step: |
| 218 | kfree(counter); |
| 219 | return NULL; |
| 220 | } |
| 221 | EXPORT_SYMBOL_GPL(lttng_counter_create); |
| 222 | |
| 223 | void lttng_counter_destroy(struct lib_counter *counter) |
| 224 | { |
| 225 | struct lib_counter_config *config = &counter->config; |
| 226 | int cpu; |
| 227 | |
| 228 | if (config->alloc & COUNTER_ALLOC_PER_CPU) { |
| 229 | for (cpu = 0; cpu < num_possible_cpus(); cpu++) |
| 230 | lttng_counter_layout_fini(counter, cpu); |
| 231 | free_percpu(counter->percpu_counters); |
| 232 | } |
| 233 | if (config->alloc & COUNTER_ALLOC_GLOBAL) |
| 234 | lttng_counter_layout_fini(counter, -1); |
| 235 | kfree(counter->dimensions); |
| 236 | kfree(counter); |
| 237 | } |
| 238 | EXPORT_SYMBOL_GPL(lttng_counter_destroy); |
| 239 | |
| 240 | int lttng_counter_read(const struct lib_counter_config *config, |
| 241 | struct lib_counter *counter, |
| 242 | const size_t *dimension_indexes, |
| 243 | int cpu, int64_t *value, bool *overflow, |
| 244 | bool *underflow) |
| 245 | { |
| 246 | struct lib_counter_layout *layout; |
| 247 | size_t index; |
| 248 | |
| 249 | if (unlikely(lttng_counter_validate_indexes(config, counter, dimension_indexes))) |
| 250 | return -EOVERFLOW; |
| 251 | index = lttng_counter_get_index(config, counter, dimension_indexes); |
| 252 | |
| 253 | switch (config->alloc) { |
| 254 | case COUNTER_ALLOC_PER_CPU: |
| 255 | if (cpu < 0 || cpu >= num_possible_cpus()) |
| 256 | return -EINVAL; |
| 257 | layout = per_cpu_ptr(counter->percpu_counters, cpu); |
| 258 | break; |
| 259 | case COUNTER_ALLOC_PER_CPU | COUNTER_ALLOC_GLOBAL: |
| 260 | if (cpu >= 0) { |
| 261 | if (cpu >= num_possible_cpus()) |
| 262 | return -EINVAL; |
| 263 | layout = per_cpu_ptr(counter->percpu_counters, cpu); |
| 264 | } else { |
| 265 | layout = &counter->global_counters; |
| 266 | } |
| 267 | break; |
| 268 | case COUNTER_ALLOC_GLOBAL: |
| 269 | if (cpu >= 0) |
| 270 | return -EINVAL; |
| 271 | layout = &counter->global_counters; |
| 272 | break; |
| 273 | default: |
| 274 | return -EINVAL; |
| 275 | } |
| 276 | |
| 277 | switch (config->counter_size) { |
| 278 | case COUNTER_SIZE_8_BIT: |
| 279 | { |
| 280 | int8_t *int_p = (int8_t *) layout->counters + index; |
| 281 | *value = (int64_t) READ_ONCE(*int_p); |
| 282 | break; |
| 283 | } |
| 284 | case COUNTER_SIZE_16_BIT: |
| 285 | { |
| 286 | int16_t *int_p = (int16_t *) layout->counters + index; |
| 287 | *value = (int64_t) READ_ONCE(*int_p); |
| 288 | break; |
| 289 | } |
| 290 | case COUNTER_SIZE_32_BIT: |
| 291 | { |
| 292 | int32_t *int_p = (int32_t *) layout->counters + index; |
| 293 | *value = (int64_t) READ_ONCE(*int_p); |
| 294 | break; |
| 295 | } |
| 296 | #if BITS_PER_LONG == 64 |
| 297 | case COUNTER_SIZE_64_BIT: |
| 298 | { |
| 299 | int64_t *int_p = (int64_t *) layout->counters + index; |
| 300 | *value = READ_ONCE(*int_p); |
| 301 | break; |
| 302 | } |
| 303 | #endif |
| 304 | default: |
| 305 | WARN_ON_ONCE(1); |
| 306 | } |
| 307 | *overflow = test_bit(index, layout->overflow_bitmap); |
| 308 | *underflow = test_bit(index, layout->underflow_bitmap); |
| 309 | return 0; |
| 310 | } |
| 311 | EXPORT_SYMBOL_GPL(lttng_counter_read); |
| 312 | |
| 313 | int lttng_counter_aggregate(const struct lib_counter_config *config, |
| 314 | struct lib_counter *counter, |
| 315 | const size_t *dimension_indexes, |
| 316 | int64_t *value, bool *overflow, |
| 317 | bool *underflow) |
| 318 | { |
| 319 | int cpu, ret; |
| 320 | int64_t v, sum = 0; |
| 321 | bool of, uf; |
| 322 | |
| 323 | *overflow = false; |
| 324 | *underflow = false; |
| 325 | |
| 326 | switch (config->alloc) { |
| 327 | case COUNTER_ALLOC_GLOBAL: /* Fallthrough */ |
| 328 | case COUNTER_ALLOC_PER_CPU | COUNTER_ALLOC_GLOBAL: |
| 329 | /* Read global counter. */ |
| 330 | ret = lttng_counter_read(config, counter, dimension_indexes, |
| 331 | -1, &v, &of, &uf); |
| 332 | if (ret < 0) |
| 333 | return ret; |
| 334 | sum += v; |
| 335 | *overflow |= of; |
| 336 | *underflow |= uf; |
| 337 | break; |
| 338 | case COUNTER_ALLOC_PER_CPU: |
| 339 | break; |
| 340 | } |
| 341 | |
| 342 | switch (config->alloc) { |
| 343 | case COUNTER_ALLOC_GLOBAL: |
| 344 | break; |
| 345 | case COUNTER_ALLOC_PER_CPU | COUNTER_ALLOC_GLOBAL: /* Fallthrough */ |
| 346 | case COUNTER_ALLOC_PER_CPU: |
| 347 | //TODO: integrate with CPU hotplug and online cpus |
| 348 | for (cpu = 0; cpu < num_possible_cpus(); cpu++) { |
| 349 | int64_t old = sum; |
| 350 | |
| 351 | ret = lttng_counter_read(config, counter, dimension_indexes, |
| 352 | cpu, &v, &of, &uf); |
| 353 | if (ret < 0) |
| 354 | return ret; |
| 355 | *overflow |= of; |
| 356 | *underflow |= uf; |
| 357 | /* Overflow is defined on unsigned types. */ |
| 358 | sum = (int64_t) ((uint64_t) old + (uint64_t) v); |
| 359 | if (v > 0 && sum < old) |
| 360 | *overflow = true; |
| 361 | else if (v < 0 && sum > old) |
| 362 | *underflow = true; |
| 363 | } |
| 364 | break; |
| 365 | default: |
| 366 | return -EINVAL; |
| 367 | } |
| 368 | *value = sum; |
| 369 | return 0; |
| 370 | } |
| 371 | EXPORT_SYMBOL_GPL(lttng_counter_aggregate); |
| 372 | |
| 373 | static |
| 374 | int lttng_counter_clear_cpu(const struct lib_counter_config *config, |
| 375 | struct lib_counter *counter, |
| 376 | const size_t *dimension_indexes, |
| 377 | int cpu) |
| 378 | { |
| 379 | struct lib_counter_layout *layout; |
| 380 | size_t index; |
| 381 | |
| 382 | if (unlikely(lttng_counter_validate_indexes(config, counter, dimension_indexes))) |
| 383 | return -EOVERFLOW; |
| 384 | index = lttng_counter_get_index(config, counter, dimension_indexes); |
| 385 | |
| 386 | switch (config->alloc) { |
| 387 | case COUNTER_ALLOC_PER_CPU: |
| 388 | if (cpu < 0 || cpu >= num_possible_cpus()) |
| 389 | return -EINVAL; |
| 390 | layout = per_cpu_ptr(counter->percpu_counters, cpu); |
| 391 | break; |
| 392 | case COUNTER_ALLOC_PER_CPU | COUNTER_ALLOC_GLOBAL: |
| 393 | if (cpu >= 0) { |
| 394 | if (cpu >= num_possible_cpus()) |
| 395 | return -EINVAL; |
| 396 | layout = per_cpu_ptr(counter->percpu_counters, cpu); |
| 397 | } else { |
| 398 | layout = &counter->global_counters; |
| 399 | } |
| 400 | break; |
| 401 | case COUNTER_ALLOC_GLOBAL: |
| 402 | if (cpu >= 0) |
| 403 | return -EINVAL; |
| 404 | layout = &counter->global_counters; |
| 405 | break; |
| 406 | default: |
| 407 | return -EINVAL; |
| 408 | } |
| 409 | switch (config->counter_size) { |
| 410 | case COUNTER_SIZE_8_BIT: |
| 411 | { |
| 412 | int8_t *int_p = (int8_t *) layout->counters + index; |
| 413 | WRITE_ONCE(*int_p, 0); |
| 414 | break; |
| 415 | } |
| 416 | case COUNTER_SIZE_16_BIT: |
| 417 | { |
| 418 | int16_t *int_p = (int16_t *) layout->counters + index; |
| 419 | WRITE_ONCE(*int_p, 0); |
| 420 | break; |
| 421 | } |
| 422 | case COUNTER_SIZE_32_BIT: |
| 423 | { |
| 424 | int32_t *int_p = (int32_t *) layout->counters + index; |
| 425 | WRITE_ONCE(*int_p, 0); |
| 426 | break; |
| 427 | } |
| 428 | #if BITS_PER_LONG == 64 |
| 429 | case COUNTER_SIZE_64_BIT: |
| 430 | { |
| 431 | int64_t *int_p = (int64_t *) layout->counters + index; |
| 432 | WRITE_ONCE(*int_p, 0); |
| 433 | break; |
| 434 | } |
| 435 | #endif |
| 436 | default: |
| 437 | WARN_ON_ONCE(1); |
| 438 | } |
| 439 | clear_bit(index, layout->overflow_bitmap); |
| 440 | clear_bit(index, layout->underflow_bitmap); |
| 441 | return 0; |
| 442 | } |
| 443 | |
| 444 | int lttng_counter_clear(const struct lib_counter_config *config, |
| 445 | struct lib_counter *counter, |
| 446 | const size_t *dimension_indexes) |
| 447 | { |
| 448 | int cpu, ret; |
| 449 | |
| 450 | switch (config->alloc) { |
| 451 | case COUNTER_ALLOC_GLOBAL: /* Fallthrough */ |
| 452 | case COUNTER_ALLOC_PER_CPU | COUNTER_ALLOC_GLOBAL: |
| 453 | /* Clear global counter. */ |
| 454 | ret = lttng_counter_clear_cpu(config, counter, dimension_indexes, -1); |
| 455 | if (ret < 0) |
| 456 | return ret; |
| 457 | break; |
| 458 | case COUNTER_ALLOC_PER_CPU: |
| 459 | break; |
| 460 | } |
| 461 | |
| 462 | switch (config->alloc) { |
| 463 | case COUNTER_ALLOC_GLOBAL: |
| 464 | break; |
| 465 | case COUNTER_ALLOC_PER_CPU | COUNTER_ALLOC_GLOBAL: /* Fallthrough */ |
| 466 | case COUNTER_ALLOC_PER_CPU: |
| 467 | //TODO: integrate with CPU hotplug and online cpus |
| 468 | for (cpu = 0; cpu < num_possible_cpus(); cpu++) { |
| 469 | ret = lttng_counter_clear_cpu(config, counter, dimension_indexes, cpu); |
| 470 | if (ret < 0) |
| 471 | return ret; |
| 472 | } |
| 473 | break; |
| 474 | default: |
| 475 | return -EINVAL; |
| 476 | } |
| 477 | return 0; |
| 478 | } |
| 479 | EXPORT_SYMBOL_GPL(lttng_counter_clear); |
| 480 | |
| 481 | int lttng_counter_get_nr_dimensions(const struct lib_counter_config *config, |
| 482 | struct lib_counter *counter, |
| 483 | size_t *nr_dimensions) |
| 484 | { |
| 485 | *nr_dimensions = counter->nr_dimensions; |
| 486 | return 0; |
| 487 | } |
| 488 | EXPORT_SYMBOL_GPL(lttng_counter_get_nr_dimensions); |
| 489 | |
| 490 | int lttng_counter_get_max_nr_elem(const struct lib_counter_config *config, |
| 491 | struct lib_counter *counter, |
| 492 | size_t *max_nr_elem) /* array of size nr_dimensions */ |
| 493 | { |
| 494 | size_t dimension; |
| 495 | |
| 496 | for (dimension = 0; dimension < counter->nr_dimensions; dimension++) |
| 497 | max_nr_elem[dimension] = lttng_counter_get_dimension_nr_elements(&counter->dimensions[dimension]); |
| 498 | return 0; |
| 499 | } |
| 500 | EXPORT_SYMBOL_GPL(lttng_counter_get_max_nr_elem); |
| 501 | |
| 502 | MODULE_LICENSE("GPL and additional rights"); |
| 503 | MODULE_AUTHOR("Mathieu Desnoyers <mathieu.desnoyers@efficios.com>"); |
| 504 | MODULE_DESCRIPTION("LTTng counter library"); |
| 505 | MODULE_VERSION(__stringify(LTTNG_MODULES_MAJOR_VERSION) "." |
| 506 | __stringify(LTTNG_MODULES_MINOR_VERSION) "." |
| 507 | __stringify(LTTNG_MODULES_PATCHLEVEL_VERSION) |
| 508 | LTTNG_MODULES_EXTRAVERSION); |