| 1 | #define FIO_MAX_CPUS 512 /* From Hyper-V 2016's max logical processors */ |
| 2 | #define FIO_CPU_MASK_STRIDE 64 |
| 3 | #define FIO_CPU_MASK_ROWS (FIO_MAX_CPUS / FIO_CPU_MASK_STRIDE) |
| 4 | |
| 5 | typedef struct { |
| 6 | uint64_t row[FIO_CPU_MASK_ROWS]; |
| 7 | } os_cpu_mask_t; |
| 8 | |
| 9 | #define FIO_HAVE_CPU_ONLINE_SYSCONF |
| 10 | /* Return all processors regardless of processor group */ |
| 11 | static inline unsigned int cpus_online(void) |
| 12 | { |
| 13 | return GetActiveProcessorCount(ALL_PROCESSOR_GROUPS); |
| 14 | } |
| 15 | |
| 16 | static inline void print_mask(os_cpu_mask_t *cpumask) |
| 17 | { |
| 18 | for (int i = 0; i < FIO_CPU_MASK_ROWS; i++) |
| 19 | dprint(FD_PROCESS, "cpumask[%d]=%lu\n", i, cpumask->row[i]); |
| 20 | } |
| 21 | |
| 22 | /* Return the index of the least significant set CPU in cpumask or -1 if no |
| 23 | * CPUs are set */ |
| 24 | static inline int first_set_cpu(os_cpu_mask_t *cpumask) |
| 25 | { |
| 26 | int cpus_offset, mask_first_cpu, row; |
| 27 | |
| 28 | cpus_offset = 0; |
| 29 | row = 0; |
| 30 | mask_first_cpu = -1; |
| 31 | while (mask_first_cpu < 0 && row < FIO_CPU_MASK_ROWS) { |
| 32 | int row_first_cpu; |
| 33 | |
| 34 | row_first_cpu = __builtin_ffsll(cpumask->row[row]) - 1; |
| 35 | dprint(FD_PROCESS, "row_first_cpu=%d cpumask->row[%d]=%lu\n", |
| 36 | row_first_cpu, row, cpumask->row[row]); |
| 37 | if (row_first_cpu > -1) { |
| 38 | mask_first_cpu = cpus_offset + row_first_cpu; |
| 39 | dprint(FD_PROCESS, "first set cpu in mask is at index %d\n", |
| 40 | mask_first_cpu); |
| 41 | } else { |
| 42 | cpus_offset += FIO_CPU_MASK_STRIDE; |
| 43 | row++; |
| 44 | } |
| 45 | } |
| 46 | |
| 47 | return mask_first_cpu; |
| 48 | } |
| 49 | |
| 50 | /* Return the index of the most significant set CPU in cpumask or -1 if no |
| 51 | * CPUs are set */ |
| 52 | static inline int last_set_cpu(os_cpu_mask_t *cpumask) |
| 53 | { |
| 54 | int cpus_offset, mask_last_cpu, row; |
| 55 | |
| 56 | cpus_offset = (FIO_CPU_MASK_ROWS - 1) * FIO_CPU_MASK_STRIDE; |
| 57 | row = FIO_CPU_MASK_ROWS - 1; |
| 58 | mask_last_cpu = -1; |
| 59 | while (mask_last_cpu < 0 && row >= 0) { |
| 60 | int row_last_cpu; |
| 61 | |
| 62 | if (cpumask->row[row] == 0) |
| 63 | row_last_cpu = -1; |
| 64 | else { |
| 65 | uint64_t tmp = cpumask->row[row]; |
| 66 | |
| 67 | row_last_cpu = 0; |
| 68 | while (tmp >>= 1) |
| 69 | row_last_cpu++; |
| 70 | } |
| 71 | |
| 72 | dprint(FD_PROCESS, "row_last_cpu=%d cpumask->row[%d]=%lu\n", |
| 73 | row_last_cpu, row, cpumask->row[row]); |
| 74 | if (row_last_cpu > -1) { |
| 75 | mask_last_cpu = cpus_offset + row_last_cpu; |
| 76 | dprint(FD_PROCESS, "last set cpu in mask is at index %d\n", |
| 77 | mask_last_cpu); |
| 78 | } else { |
| 79 | cpus_offset -= FIO_CPU_MASK_STRIDE; |
| 80 | row--; |
| 81 | } |
| 82 | } |
| 83 | |
| 84 | return mask_last_cpu; |
| 85 | } |
| 86 | |
| 87 | static inline int mask_to_group_mask(os_cpu_mask_t *cpumask, int *processor_group, uint64_t *affinity_mask) |
| 88 | { |
| 89 | WORD online_groups, group, group_size; |
| 90 | bool found; |
| 91 | int cpus_offset, search_cpu, last_cpu, bit_offset, row, end; |
| 92 | uint64_t group_cpumask; |
| 93 | |
| 94 | search_cpu = first_set_cpu(cpumask); |
| 95 | if (search_cpu < 0) { |
| 96 | log_info("CPU mask doesn't set any CPUs\n"); |
| 97 | return 1; |
| 98 | } |
| 99 | |
| 100 | /* Find processor group first set CPU applies to */ |
| 101 | online_groups = GetActiveProcessorGroupCount(); |
| 102 | group = 0; |
| 103 | found = false; |
| 104 | cpus_offset = 0; |
| 105 | group_size = 0; |
| 106 | while (!found && group < online_groups) { |
| 107 | group_size = GetActiveProcessorCount(group); |
| 108 | dprint(FD_PROCESS, "group=%d group_start=%d group_size=%u search_cpu=%d\n", |
| 109 | group, cpus_offset, group_size, search_cpu); |
| 110 | if (cpus_offset + group_size > search_cpu) |
| 111 | found = true; |
| 112 | else { |
| 113 | cpus_offset += group_size; |
| 114 | group++; |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | if (!found) { |
| 119 | log_err("CPU mask contains processor beyond last active processor index (%d)\n", |
| 120 | cpus_offset - 1); |
| 121 | print_mask(cpumask); |
| 122 | return 1; |
| 123 | } |
| 124 | |
| 125 | /* Check all the CPUs in the mask apply to ONLY that processor group */ |
| 126 | last_cpu = last_set_cpu(cpumask); |
| 127 | if (last_cpu > (cpus_offset + group_size - 1)) { |
| 128 | log_info("CPU mask cannot bind CPUs (e.g. %d, %d) that are " |
| 129 | "in different processor groups\n", search_cpu, |
| 130 | last_cpu); |
| 131 | print_mask(cpumask); |
| 132 | return 1; |
| 133 | } |
| 134 | |
| 135 | /* Extract the current processor group mask from the cpumask */ |
| 136 | row = cpus_offset / FIO_CPU_MASK_STRIDE; |
| 137 | bit_offset = cpus_offset % FIO_CPU_MASK_STRIDE; |
| 138 | group_cpumask = cpumask->row[row] >> bit_offset; |
| 139 | end = bit_offset + group_size; |
| 140 | if (end > FIO_CPU_MASK_STRIDE && (row + 1 < FIO_CPU_MASK_ROWS)) { |
| 141 | /* Some of the next row needs to be part of the mask */ |
| 142 | int needed, needed_shift, needed_mask_shift; |
| 143 | uint64_t needed_mask; |
| 144 | |
| 145 | needed = end - FIO_CPU_MASK_STRIDE; |
| 146 | needed_shift = FIO_CPU_MASK_STRIDE - bit_offset; |
| 147 | needed_mask_shift = FIO_CPU_MASK_STRIDE - needed; |
| 148 | needed_mask = (uint64_t)-1 >> needed_mask_shift; |
| 149 | dprint(FD_PROCESS, "bit_offset=%d end=%d needed=%d needed_shift=%d needed_mask=%ld needed_mask_shift=%d\n", bit_offset, end, needed, needed_shift, needed_mask, needed_mask_shift); |
| 150 | group_cpumask |= (cpumask->row[row + 1] & needed_mask) << needed_shift; |
| 151 | } |
| 152 | group_cpumask &= (uint64_t)-1 >> (FIO_CPU_MASK_STRIDE - group_size); |
| 153 | |
| 154 | /* Return group and mask */ |
| 155 | dprint(FD_PROCESS, "Returning group=%d group_mask=%lu\n", group, group_cpumask); |
| 156 | *processor_group = group; |
| 157 | *affinity_mask = group_cpumask; |
| 158 | |
| 159 | return 0; |
| 160 | } |
| 161 | |
| 162 | static inline int fio_setaffinity(int pid, os_cpu_mask_t cpumask) |
| 163 | { |
| 164 | HANDLE handle = NULL; |
| 165 | int group, ret; |
| 166 | uint64_t group_mask = 0; |
| 167 | GROUP_AFFINITY new_group_affinity; |
| 168 | |
| 169 | ret = -1; |
| 170 | |
| 171 | if (mask_to_group_mask(&cpumask, &group, &group_mask) != 0) |
| 172 | goto err; |
| 173 | |
| 174 | handle = OpenThread(THREAD_QUERY_INFORMATION | THREAD_SET_INFORMATION, |
| 175 | TRUE, pid); |
| 176 | if (handle == NULL) { |
| 177 | log_err("fio_setaffinity: failed to get handle for pid %d\n", pid); |
| 178 | goto err; |
| 179 | } |
| 180 | |
| 181 | /* Set group and mask. |
| 182 | * Note: if the GROUP_AFFINITY struct's Reserved members are not |
| 183 | * initialised to 0 then SetThreadGroupAffinity will fail with |
| 184 | * GetLastError() set to ERROR_INVALID_PARAMETER */ |
| 185 | new_group_affinity.Mask = (KAFFINITY) group_mask; |
| 186 | new_group_affinity.Group = group; |
| 187 | new_group_affinity.Reserved[0] = 0; |
| 188 | new_group_affinity.Reserved[1] = 0; |
| 189 | new_group_affinity.Reserved[2] = 0; |
| 190 | if (SetThreadGroupAffinity(handle, &new_group_affinity, NULL) != 0) |
| 191 | ret = 0; |
| 192 | else { |
| 193 | log_err("fio_setaffinity: failed to set thread affinity " |
| 194 | "(pid %d, group %d, mask %" PRIx64 ", " |
| 195 | "GetLastError=%d)\n", pid, group, group_mask, |
| 196 | GetLastError()); |
| 197 | goto err; |
| 198 | } |
| 199 | |
| 200 | err: |
| 201 | if (handle) |
| 202 | CloseHandle(handle); |
| 203 | return ret; |
| 204 | } |
| 205 | |
| 206 | static inline void cpu_to_row_offset(int cpu, int *row, int *offset) |
| 207 | { |
| 208 | *row = cpu / FIO_CPU_MASK_STRIDE; |
| 209 | *offset = cpu << FIO_CPU_MASK_STRIDE * *row; |
| 210 | } |
| 211 | |
| 212 | static inline int fio_cpuset_init(os_cpu_mask_t *mask) |
| 213 | { |
| 214 | for (int i = 0; i < FIO_CPU_MASK_ROWS; i++) |
| 215 | mask->row[i] = 0; |
| 216 | return 0; |
| 217 | } |
| 218 | |
| 219 | /* |
| 220 | * fio_getaffinity() should not be called once a fio_setaffinity() call has |
| 221 | * been made because fio_setaffinity() may put the process into multiple |
| 222 | * processor groups |
| 223 | */ |
| 224 | static inline int fio_getaffinity(int pid, os_cpu_mask_t *mask) |
| 225 | { |
| 226 | int ret; |
| 227 | int row, offset, end, group, group_size, group_start_cpu; |
| 228 | DWORD_PTR process_mask, system_mask; |
| 229 | HANDLE handle; |
| 230 | PUSHORT current_groups; |
| 231 | USHORT group_count; |
| 232 | WORD online_groups; |
| 233 | |
| 234 | ret = -1; |
| 235 | current_groups = NULL; |
| 236 | handle = OpenProcess(PROCESS_QUERY_INFORMATION, TRUE, pid); |
| 237 | if (handle == NULL) { |
| 238 | log_err("fio_getaffinity: failed to get handle for pid %d\n", |
| 239 | pid); |
| 240 | goto err; |
| 241 | } |
| 242 | |
| 243 | group_count = 1; |
| 244 | /* |
| 245 | * GetProcessGroupAffinity() seems to expect more than the natural |
| 246 | * alignment for a USHORT from the area pointed to by current_groups so |
| 247 | * arrange for maximum alignment by allocating via malloc() |
| 248 | */ |
| 249 | current_groups = malloc(sizeof(USHORT)); |
| 250 | if (!current_groups) { |
| 251 | log_err("fio_getaffinity: malloc failed\n"); |
| 252 | goto err; |
| 253 | } |
| 254 | if (GetProcessGroupAffinity(handle, &group_count, current_groups) == 0) { |
| 255 | /* NB: we also fail here if we are a multi-group process */ |
| 256 | log_err("fio_getaffinity: failed to get single group affinity for pid %d\n", pid); |
| 257 | goto err; |
| 258 | } |
| 259 | GetProcessAffinityMask(handle, &process_mask, &system_mask); |
| 260 | |
| 261 | /* Convert group and group relative mask to full CPU mask */ |
| 262 | online_groups = GetActiveProcessorGroupCount(); |
| 263 | if (online_groups == 0) { |
| 264 | log_err("fio_getaffinity: error retrieving total processor groups\n"); |
| 265 | goto err; |
| 266 | } |
| 267 | |
| 268 | group = 0; |
| 269 | group_start_cpu = 0; |
| 270 | group_size = 0; |
| 271 | dprint(FD_PROCESS, "current_groups=%d group_count=%d\n", |
| 272 | current_groups[0], group_count); |
| 273 | while (true) { |
| 274 | group_size = GetActiveProcessorCount(group); |
| 275 | if (group_size == 0) { |
| 276 | log_err("fio_getaffinity: error retrieving size of " |
| 277 | "processor group %d\n", group); |
| 278 | goto err; |
| 279 | } else if (group >= current_groups[0] || group >= online_groups) |
| 280 | break; |
| 281 | else { |
| 282 | group_start_cpu += group_size; |
| 283 | group++; |
| 284 | } |
| 285 | } |
| 286 | |
| 287 | if (group != current_groups[0]) { |
| 288 | log_err("fio_getaffinity: could not find processor group %d\n", |
| 289 | current_groups[0]); |
| 290 | goto err; |
| 291 | } |
| 292 | |
| 293 | dprint(FD_PROCESS, "group_start_cpu=%d, group size=%u\n", |
| 294 | group_start_cpu, group_size); |
| 295 | if ((group_start_cpu + group_size) >= FIO_MAX_CPUS) { |
| 296 | log_err("fio_getaffinity failed: current CPU affinity (group " |
| 297 | "%d, group_start_cpu %d, group_size %d) extends " |
| 298 | "beyond mask's highest CPU (%d)\n", group, |
| 299 | group_start_cpu, group_size, FIO_MAX_CPUS); |
| 300 | goto err; |
| 301 | } |
| 302 | |
| 303 | fio_cpuset_init(mask); |
| 304 | cpu_to_row_offset(group_start_cpu, &row, &offset); |
| 305 | mask->row[row] = process_mask; |
| 306 | mask->row[row] <<= offset; |
| 307 | end = offset + group_size; |
| 308 | if (end > FIO_CPU_MASK_STRIDE) { |
| 309 | int needed; |
| 310 | uint64_t needed_mask; |
| 311 | |
| 312 | needed = FIO_CPU_MASK_STRIDE - end; |
| 313 | needed_mask = (uint64_t)-1 >> (FIO_CPU_MASK_STRIDE - needed); |
| 314 | row++; |
| 315 | mask->row[row] = process_mask; |
| 316 | mask->row[row] >>= needed; |
| 317 | mask->row[row] &= needed_mask; |
| 318 | } |
| 319 | ret = 0; |
| 320 | |
| 321 | err: |
| 322 | if (handle) |
| 323 | CloseHandle(handle); |
| 324 | if (current_groups) |
| 325 | free(current_groups); |
| 326 | |
| 327 | return ret; |
| 328 | } |
| 329 | |
| 330 | static inline void fio_cpu_clear(os_cpu_mask_t *mask, int cpu) |
| 331 | { |
| 332 | int row, offset; |
| 333 | cpu_to_row_offset(cpu, &row, &offset); |
| 334 | |
| 335 | mask->row[row] &= ~(1ULL << offset); |
| 336 | } |
| 337 | |
| 338 | static inline void fio_cpu_set(os_cpu_mask_t *mask, int cpu) |
| 339 | { |
| 340 | int row, offset; |
| 341 | cpu_to_row_offset(cpu, &row, &offset); |
| 342 | |
| 343 | mask->row[row] |= 1ULL << offset; |
| 344 | } |
| 345 | |
| 346 | static inline int fio_cpu_isset(os_cpu_mask_t *mask, int cpu) |
| 347 | { |
| 348 | int row, offset; |
| 349 | cpu_to_row_offset(cpu, &row, &offset); |
| 350 | |
| 351 | return (mask->row[row] & (1ULL << offset)) != 0; |
| 352 | } |
| 353 | |
| 354 | static inline int fio_cpu_count(os_cpu_mask_t *mask) |
| 355 | { |
| 356 | int count = 0; |
| 357 | |
| 358 | for (int i = 0; i < FIO_CPU_MASK_ROWS; i++) |
| 359 | count += hweight64(mask->row[i]); |
| 360 | |
| 361 | return count; |
| 362 | } |
| 363 | |
| 364 | static inline int fio_cpuset_exit(os_cpu_mask_t *mask) |
| 365 | { |
| 366 | return 0; |
| 367 | } |