| 1 | /* SPDX-License-Identifier: (GPL-2.0-only or LGPL-2.1-only) |
| 2 | * |
| 3 | * counter/counter-api.h |
| 4 | * |
| 5 | * LTTng Counters API, requiring counter/config.h |
| 6 | * |
| 7 | * Copyright (C) 2020 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 8 | */ |
| 9 | |
| 10 | #ifndef _LTTNG_COUNTER_API_H |
| 11 | #define _LTTNG_COUNTER_API_H |
| 12 | |
| 13 | #include <linux/types.h> |
| 14 | #include <linux/percpu.h> |
| 15 | #include <linux/bitops.h> |
| 16 | #include <counter/counter.h> |
| 17 | #include <counter/counter-internal.h> |
| 18 | #include <wrapper/limits.h> |
| 19 | |
| 20 | /* |
| 21 | * Using unsigned arithmetic because overflow is defined. |
| 22 | */ |
| 23 | static __always_inline int __lttng_counter_add(const struct lib_counter_config *config, |
| 24 | enum lib_counter_config_alloc alloc, |
| 25 | enum lib_counter_config_sync sync, |
| 26 | struct lib_counter *counter, |
| 27 | const size_t *dimension_indexes, int64_t v, |
| 28 | int64_t *remainder) |
| 29 | { |
| 30 | size_t index; |
| 31 | bool overflow = false, underflow = false; |
| 32 | struct lib_counter_layout *layout; |
| 33 | int64_t move_sum = 0; |
| 34 | |
| 35 | if (unlikely(lttng_counter_validate_indexes(config, counter, dimension_indexes))) |
| 36 | return -EOVERFLOW; |
| 37 | index = lttng_counter_get_index(config, counter, dimension_indexes); |
| 38 | |
| 39 | switch (alloc) { |
| 40 | case COUNTER_ALLOC_PER_CPU: |
| 41 | layout = per_cpu_ptr(counter->percpu_counters, smp_processor_id()); |
| 42 | break; |
| 43 | case COUNTER_ALLOC_GLOBAL: |
| 44 | layout = &counter->global_counters; |
| 45 | break; |
| 46 | default: |
| 47 | return -EINVAL; |
| 48 | } |
| 49 | |
| 50 | switch (config->counter_size) { |
| 51 | case COUNTER_SIZE_8_BIT: |
| 52 | { |
| 53 | int8_t *int_p = (int8_t *) layout->counters + index; |
| 54 | int8_t old, n, res; |
| 55 | int8_t global_sum_step = counter->global_sum_step.s8; |
| 56 | |
| 57 | res = *int_p; |
| 58 | switch (sync) { |
| 59 | case COUNTER_SYNC_PER_CPU: |
| 60 | { |
| 61 | do { |
| 62 | move_sum = 0; |
| 63 | old = res; |
| 64 | n = (int8_t) ((uint8_t) old + (uint8_t) v); |
| 65 | if (unlikely(n > (int8_t) global_sum_step)) |
| 66 | move_sum = (int8_t) global_sum_step / 2; |
| 67 | else if (unlikely(n < -(int8_t) global_sum_step)) |
| 68 | move_sum = -((int8_t) global_sum_step / 2); |
| 69 | n -= move_sum; |
| 70 | res = cmpxchg_local(int_p, old, n); |
| 71 | } while (old != res); |
| 72 | break; |
| 73 | } |
| 74 | case COUNTER_SYNC_GLOBAL: |
| 75 | { |
| 76 | do { |
| 77 | old = res; |
| 78 | n = (int8_t) ((uint8_t) old + (uint8_t) v); |
| 79 | res = cmpxchg(int_p, old, n); |
| 80 | } while (old != res); |
| 81 | break; |
| 82 | } |
| 83 | default: |
| 84 | return -EINVAL; |
| 85 | } |
| 86 | if (v > 0 && (v >= U8_MAX || n < old)) |
| 87 | overflow = true; |
| 88 | else if (v < 0 && (v <= -(s64) U8_MAX || n > old)) |
| 89 | underflow = true; |
| 90 | break; |
| 91 | } |
| 92 | case COUNTER_SIZE_16_BIT: |
| 93 | { |
| 94 | int16_t *int_p = (int16_t *) layout->counters + index; |
| 95 | int16_t old, n, res; |
| 96 | int16_t global_sum_step = counter->global_sum_step.s16; |
| 97 | |
| 98 | res = *int_p; |
| 99 | switch (sync) { |
| 100 | case COUNTER_SYNC_PER_CPU: |
| 101 | { |
| 102 | do { |
| 103 | move_sum = 0; |
| 104 | old = res; |
| 105 | n = (int16_t) ((uint16_t) old + (uint16_t) v); |
| 106 | if (unlikely(n > (int16_t) global_sum_step)) |
| 107 | move_sum = (int16_t) global_sum_step / 2; |
| 108 | else if (unlikely(n < -(int16_t) global_sum_step)) |
| 109 | move_sum = -((int16_t) global_sum_step / 2); |
| 110 | n -= move_sum; |
| 111 | res = cmpxchg_local(int_p, old, n); |
| 112 | } while (old != res); |
| 113 | break; |
| 114 | } |
| 115 | case COUNTER_SYNC_GLOBAL: |
| 116 | { |
| 117 | do { |
| 118 | old = res; |
| 119 | n = (int16_t) ((uint16_t) old + (uint16_t) v); |
| 120 | res = cmpxchg(int_p, old, n); |
| 121 | } while (old != res); |
| 122 | break; |
| 123 | } |
| 124 | default: |
| 125 | return -EINVAL; |
| 126 | } |
| 127 | if (v > 0 && (v >= U16_MAX || n < old)) |
| 128 | overflow = true; |
| 129 | else if (v < 0 && (v <= -(s64) U16_MAX || n > old)) |
| 130 | underflow = true; |
| 131 | break; |
| 132 | } |
| 133 | case COUNTER_SIZE_32_BIT: |
| 134 | { |
| 135 | int32_t *int_p = (int32_t *) layout->counters + index; |
| 136 | int32_t old, n, res; |
| 137 | int32_t global_sum_step = counter->global_sum_step.s32; |
| 138 | |
| 139 | res = *int_p; |
| 140 | switch (sync) { |
| 141 | case COUNTER_SYNC_PER_CPU: |
| 142 | { |
| 143 | do { |
| 144 | move_sum = 0; |
| 145 | old = res; |
| 146 | n = (int32_t) ((uint32_t) old + (uint32_t) v); |
| 147 | if (unlikely(n > (int32_t) global_sum_step)) |
| 148 | move_sum = (int32_t) global_sum_step / 2; |
| 149 | else if (unlikely(n < -(int32_t) global_sum_step)) |
| 150 | move_sum = -((int32_t) global_sum_step / 2); |
| 151 | n -= move_sum; |
| 152 | res = cmpxchg_local(int_p, old, n); |
| 153 | } while (old != res); |
| 154 | break; |
| 155 | } |
| 156 | case COUNTER_SYNC_GLOBAL: |
| 157 | { |
| 158 | do { |
| 159 | old = res; |
| 160 | n = (int32_t) ((uint32_t) old + (uint32_t) v); |
| 161 | res = cmpxchg(int_p, old, n); |
| 162 | } while (old != res); |
| 163 | break; |
| 164 | } |
| 165 | default: |
| 166 | return -EINVAL; |
| 167 | } |
| 168 | if (v > 0 && (v >= U32_MAX || n < old)) |
| 169 | overflow = true; |
| 170 | else if (v < 0 && (v <= -(s64) U32_MAX || n > old)) |
| 171 | underflow = true; |
| 172 | break; |
| 173 | } |
| 174 | #if BITS_PER_LONG == 64 |
| 175 | case COUNTER_SIZE_64_BIT: |
| 176 | { |
| 177 | int64_t *int_p = (int64_t *) layout->counters + index; |
| 178 | int64_t old, n, res; |
| 179 | int64_t global_sum_step = counter->global_sum_step.s64; |
| 180 | |
| 181 | res = *int_p; |
| 182 | switch (sync) { |
| 183 | case COUNTER_SYNC_PER_CPU: |
| 184 | { |
| 185 | do { |
| 186 | move_sum = 0; |
| 187 | old = res; |
| 188 | n = (int64_t) ((uint64_t) old + (uint64_t) v); |
| 189 | if (unlikely(n > (int64_t) global_sum_step)) |
| 190 | move_sum = (int64_t) global_sum_step / 2; |
| 191 | else if (unlikely(n < -(int64_t) global_sum_step)) |
| 192 | move_sum = -((int64_t) global_sum_step / 2); |
| 193 | n -= move_sum; |
| 194 | res = cmpxchg_local(int_p, old, n); |
| 195 | } while (old != res); |
| 196 | break; |
| 197 | } |
| 198 | case COUNTER_SYNC_GLOBAL: |
| 199 | { |
| 200 | do { |
| 201 | old = res; |
| 202 | n = (int64_t) ((uint64_t) old + (uint64_t) v); |
| 203 | res = cmpxchg(int_p, old, n); |
| 204 | } while (old != res); |
| 205 | break; |
| 206 | } |
| 207 | default: |
| 208 | return -EINVAL; |
| 209 | } |
| 210 | if (v > 0 && n < old) |
| 211 | overflow = true; |
| 212 | else if (v < 0 && n > old) |
| 213 | underflow = true; |
| 214 | break; |
| 215 | } |
| 216 | #endif |
| 217 | default: |
| 218 | return -EINVAL; |
| 219 | } |
| 220 | if (unlikely(overflow && !test_bit(index, layout->overflow_bitmap))) |
| 221 | set_bit(index, layout->overflow_bitmap); |
| 222 | else if (unlikely(underflow && !test_bit(index, layout->underflow_bitmap))) |
| 223 | set_bit(index, layout->underflow_bitmap); |
| 224 | if (remainder) |
| 225 | *remainder = move_sum; |
| 226 | return 0; |
| 227 | } |
| 228 | |
| 229 | static __always_inline int __lttng_counter_add_percpu(const struct lib_counter_config *config, |
| 230 | struct lib_counter *counter, |
| 231 | const size_t *dimension_indexes, int64_t v) |
| 232 | { |
| 233 | int64_t move_sum; |
| 234 | int ret; |
| 235 | |
| 236 | ret = __lttng_counter_add(config, COUNTER_ALLOC_PER_CPU, config->sync, |
| 237 | counter, dimension_indexes, v, &move_sum); |
| 238 | if (unlikely(ret)) |
| 239 | return ret; |
| 240 | if (unlikely(move_sum)) |
| 241 | return __lttng_counter_add(config, COUNTER_ALLOC_GLOBAL, COUNTER_SYNC_GLOBAL, |
| 242 | counter, dimension_indexes, move_sum, NULL); |
| 243 | return 0; |
| 244 | } |
| 245 | |
| 246 | static __always_inline int __lttng_counter_add_global(const struct lib_counter_config *config, |
| 247 | struct lib_counter *counter, |
| 248 | const size_t *dimension_indexes, int64_t v) |
| 249 | { |
| 250 | return __lttng_counter_add(config, COUNTER_ALLOC_GLOBAL, config->sync, counter, |
| 251 | dimension_indexes, v, NULL); |
| 252 | } |
| 253 | |
| 254 | static __always_inline int lttng_counter_add(const struct lib_counter_config *config, |
| 255 | struct lib_counter *counter, |
| 256 | const size_t *dimension_indexes, int64_t v) |
| 257 | { |
| 258 | switch (config->alloc) { |
| 259 | case COUNTER_ALLOC_PER_CPU: /* Fallthrough */ |
| 260 | case COUNTER_ALLOC_PER_CPU | COUNTER_ALLOC_GLOBAL: |
| 261 | return __lttng_counter_add_percpu(config, counter, dimension_indexes, v); |
| 262 | case COUNTER_ALLOC_GLOBAL: |
| 263 | return __lttng_counter_add_global(config, counter, dimension_indexes, v); |
| 264 | default: |
| 265 | return -EINVAL; |
| 266 | } |
| 267 | } |
| 268 | |
| 269 | static __always_inline int lttng_counter_inc(const struct lib_counter_config *config, |
| 270 | struct lib_counter *counter, |
| 271 | const size_t *dimension_indexes) |
| 272 | { |
| 273 | return lttng_counter_add(config, counter, dimension_indexes, 1); |
| 274 | } |
| 275 | |
| 276 | static __always_inline int lttng_counter_dec(const struct lib_counter_config *config, |
| 277 | struct lib_counter *counter, |
| 278 | const size_t *dimension_indexes) |
| 279 | { |
| 280 | return lttng_counter_add(config, counter, dimension_indexes, -1); |
| 281 | } |
| 282 | |
| 283 | #endif /* _LTTNG_COUNTER_API_H */ |