#include #include #include #include #include #include #include #include #include #ifdef __GNUC__ #define CPUINFO_ALLOCA __builtin_alloca #else #define CPUINFO_ALLOCA _alloca #endif static inline uint32_t bit_mask(uint32_t bits) { return (UINT32_C(1) << bits) - UINT32_C(1); } static inline uint32_t low_index_from_kaffinity(KAFFINITY kaffinity) { #if defined(_M_X64) || defined(_M_AMD64) unsigned long index; _BitScanForward64(&index, (unsigned __int64) kaffinity); return (uint32_t) index; #elif defined(_M_IX86) unsigned long index; _BitScanForward(&index, (unsigned long) kaffinity); return (uint32_t) index; #else #error Platform-specific implementation required #endif } static void cpuinfo_x86_count_caches( uint32_t processors_count, const struct cpuinfo_processor* processors, const struct cpuinfo_x86_processor* x86_processor, uint32_t* l1i_count_ptr, uint32_t* l1d_count_ptr, uint32_t* l2_count_ptr, uint32_t* l3_count_ptr, uint32_t* l4_count_ptr) { uint32_t l1i_count = 0, l1d_count = 0, l2_count = 0, l3_count = 0, l4_count = 0; uint32_t last_l1i_id = UINT32_MAX, last_l1d_id = UINT32_MAX; uint32_t last_l2_id = UINT32_MAX, last_l3_id = UINT32_MAX, last_l4_id = UINT32_MAX; for (uint32_t i = 0; i < processors_count; i++) { const uint32_t apic_id = processors[i].apic_id; cpuinfo_log_debug("APID ID %"PRIu32": logical processor %"PRIu32, apic_id, i); if (x86_processor->cache.l1i.size != 0) { const uint32_t l1i_id = apic_id & ~bit_mask(x86_processor->cache.l1i.apic_bits); if (l1i_id != last_l1i_id) { last_l1i_id = l1i_id; l1i_count++; } } if (x86_processor->cache.l1d.size != 0) { const uint32_t l1d_id = apic_id & ~bit_mask(x86_processor->cache.l1d.apic_bits); if (l1d_id != last_l1d_id) { last_l1d_id = l1d_id; l1d_count++; } } if (x86_processor->cache.l2.size != 0) { const uint32_t l2_id = apic_id & ~bit_mask(x86_processor->cache.l2.apic_bits); if (l2_id != last_l2_id) { last_l2_id = l2_id; l2_count++; } } if (x86_processor->cache.l3.size != 0) { const uint32_t l3_id = apic_id & ~bit_mask(x86_processor->cache.l3.apic_bits); if (l3_id != last_l3_id) { last_l3_id = l3_id; l3_count++; } } if (x86_processor->cache.l4.size != 0) { const uint32_t l4_id = apic_id & ~bit_mask(x86_processor->cache.l4.apic_bits); if (l4_id != last_l4_id) { last_l4_id = l4_id; l4_count++; } } } *l1i_count_ptr = l1i_count; *l1d_count_ptr = l1d_count; *l2_count_ptr = l2_count; *l3_count_ptr = l3_count; *l4_count_ptr = l4_count; } BOOL CALLBACK cpuinfo_x86_windows_init(PINIT_ONCE init_once, PVOID parameter, PVOID* context) { struct cpuinfo_processor* processors = NULL; struct cpuinfo_core* cores = NULL; struct cpuinfo_cluster* clusters = NULL; struct cpuinfo_package* packages = NULL; struct cpuinfo_cache* l1i = NULL; struct cpuinfo_cache* l1d = NULL; struct cpuinfo_cache* l2 = NULL; struct cpuinfo_cache* l3 = NULL; struct cpuinfo_cache* l4 = NULL; PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX processor_infos = NULL; HANDLE heap = GetProcessHeap(); struct cpuinfo_x86_processor x86_processor; ZeroMemory(&x86_processor, sizeof(x86_processor)); cpuinfo_x86_init_processor(&x86_processor); char brand_string[48]; cpuinfo_x86_normalize_brand_string(x86_processor.brand_string, brand_string); const uint32_t thread_bits_mask = bit_mask(x86_processor.topology.thread_bits_length); const uint32_t core_bits_mask = bit_mask(x86_processor.topology.core_bits_length); const uint32_t package_bits_offset = max( x86_processor.topology.thread_bits_offset + x86_processor.topology.thread_bits_length, x86_processor.topology.core_bits_offset + x86_processor.topology.core_bits_length); const uint32_t max_group_count = (uint32_t) GetMaximumProcessorGroupCount(); cpuinfo_log_debug("detected %"PRIu32" processor groups", max_group_count); uint32_t processors_count = 0; uint32_t* processors_per_group = (uint32_t*) CPUINFO_ALLOCA(max_group_count * sizeof(uint32_t)); for (uint32_t i = 0; i < max_group_count; i++) { processors_per_group[i] = GetMaximumProcessorCount((WORD) i); cpuinfo_log_debug("detected %"PRIu32" processors in group %"PRIu32, processors_per_group[i], i); processors_count += processors_per_group[i]; } uint32_t* processors_before_group = (uint32_t*) CPUINFO_ALLOCA(max_group_count * sizeof(uint32_t)); for (uint32_t i = 0, count = 0; i < max_group_count; i++) { processors_before_group[i] = count; cpuinfo_log_debug("detected %"PRIu32" processors before group %"PRIu32, processors_before_group[i], i); count += processors_per_group[i]; } processors = HeapAlloc(heap, HEAP_ZERO_MEMORY, processors_count * sizeof(struct cpuinfo_processor)); if (processors == NULL) { cpuinfo_log_error("failed to allocate %zu bytes for descriptions of %"PRIu32" logical processors", processors_count * sizeof(struct cpuinfo_processor), processors_count); goto cleanup; } DWORD cores_info_size = 0; if (GetLogicalProcessorInformationEx(RelationProcessorCore, NULL, &cores_info_size) == FALSE) { const DWORD last_error = GetLastError(); if (last_error != ERROR_INSUFFICIENT_BUFFER) { cpuinfo_log_error("failed to query size of processor cores information: error %"PRIu32, (uint32_t) last_error); goto cleanup; } } DWORD packages_info_size = 0; if (GetLogicalProcessorInformationEx(RelationProcessorPackage, NULL, &packages_info_size) == FALSE) { const DWORD last_error = GetLastError(); if (last_error != ERROR_INSUFFICIENT_BUFFER) { cpuinfo_log_error("failed to query size of processor packages information: error %"PRIu32, (uint32_t) last_error); goto cleanup; } } DWORD max_info_size = max(cores_info_size, packages_info_size); processor_infos = HeapAlloc(heap, 0, max_info_size); if (processor_infos == NULL) { cpuinfo_log_error("failed to allocate %"PRIu32" bytes for logical processor information", (uint32_t) max_info_size); goto cleanup; } if (GetLogicalProcessorInformationEx(RelationProcessorPackage, processor_infos, &max_info_size) == FALSE) { cpuinfo_log_error("failed to query processor packages information: error %"PRIu32, (uint32_t) GetLastError()); goto cleanup; } uint32_t packages_count = 0; PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX packages_info_end = (PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX) ((uintptr_t) processor_infos + packages_info_size); for (PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX package_info = processor_infos; package_info < packages_info_end; package_info = (PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX) ((uintptr_t) package_info + package_info->Size)) { if (package_info->Relationship != RelationProcessorPackage) { cpuinfo_log_warning("unexpected processor info type (%"PRIu32") for processor package information", (uint32_t) package_info->Relationship); continue; } /* We assume that packages are reported in APIC order */ const uint32_t package_id = packages_count++; /* Reconstruct package part of APIC ID */ const uint32_t package_apic_id = package_id << package_bits_offset; /* Iterate processor groups and set the package part of APIC ID */ for (uint32_t i = 0; i < package_info->Processor.GroupCount; i++) { const uint32_t group_id = package_info->Processor.GroupMask[i].Group; /* Global index of the first logical processor belonging to this group */ const uint32_t group_processors_start = processors_before_group[group_id]; /* Bitmask representing processors in this group belonging to this package */ KAFFINITY group_processors_mask = package_info->Processor.GroupMask[i].Mask; while (group_processors_mask != 0) { const uint32_t group_processor_id = low_index_from_kaffinity(group_processors_mask); const uint32_t processor_id = group_processors_start + group_processor_id; processors[processor_id].package = (const struct cpuinfo_package*) NULL + package_id; processors[processor_id].windows_group_id = (uint16_t) group_id; processors[processor_id].windows_processor_id = (uint16_t) group_processor_id; processors[processor_id].apic_id = package_apic_id; /* Reset the lowest bit in affinity mask */ group_processors_mask &= (group_processors_mask - 1); } } } max_info_size = max(cores_info_size, packages_info_size); if (GetLogicalProcessorInformationEx(RelationProcessorCore, processor_infos, &max_info_size) == FALSE) { cpuinfo_log_error("failed to query processor cores information: error %"PRIu32, (uint32_t) GetLastError()); goto cleanup; } uint32_t cores_count = 0; /* Index (among all cores) of the the first core on the current package */ uint32_t package_core_start = 0; uint32_t current_package_apic_id = 0; PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX cores_info_end = (PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX) ((uintptr_t) processor_infos + cores_info_size); for (PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX core_info = processor_infos; core_info < cores_info_end; core_info = (PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX) ((uintptr_t) core_info + core_info->Size)) { if (core_info->Relationship != RelationProcessorCore) { cpuinfo_log_warning("unexpected processor info type (%"PRIu32") for processor core information", (uint32_t) core_info->Relationship); continue; } /* We assume that cores and logical processors are reported in APIC order */ const uint32_t core_id = cores_count++; uint32_t smt_id = 0; /* Reconstruct core part of APIC ID */ const uint32_t core_apic_id = (core_id & core_bits_mask) << x86_processor.topology.core_bits_offset; /* Iterate processor groups and set the core & SMT parts of APIC ID */ for (uint32_t i = 0; i < core_info->Processor.GroupCount; i++) { const uint32_t group_id = core_info->Processor.GroupMask[i].Group; /* Global index of the first logical processor belonging to this group */ const uint32_t group_processors_start = processors_before_group[group_id]; /* Bitmask representing processors in this group belonging to this package */ KAFFINITY group_processors_mask = core_info->Processor.GroupMask[i].Mask; while (group_processors_mask != 0) { const uint32_t group_processor_id = low_index_from_kaffinity(group_processors_mask); const uint32_t processor_id = group_processors_start + group_processor_id; /* Check if this is the first core on a new package */ if (processors[processor_id].apic_id != current_package_apic_id) { package_core_start = core_id; current_package_apic_id = processors[processor_id].apic_id; } /* Core ID w.r.t package */ const uint32_t package_core_id = core_id - package_core_start; /* Update APIC ID with core and SMT parts */ processors[processor_id].apic_id |= ((smt_id & thread_bits_mask) << x86_processor.topology.thread_bits_offset) | ((package_core_id & core_bits_mask) << x86_processor.topology.core_bits_offset); cpuinfo_log_debug("reconstructed APIC ID 0x%08"PRIx32" for processor %"PRIu32" in group %"PRIu32, processors[processor_id].apic_id, group_processor_id, group_id); /* Set SMT ID (assume logical processors within the core are reported in APIC order) */ processors[processor_id].smt_id = smt_id++; processors[processor_id].core = (const struct cpuinfo_core*) NULL + core_id; /* Reset the lowest bit in affinity mask */ group_processors_mask &= (group_processors_mask - 1); } } } cores = HeapAlloc(heap, HEAP_ZERO_MEMORY, cores_count * sizeof(struct cpuinfo_core)); if (cores == NULL) { cpuinfo_log_error("failed to allocate %zu bytes for descriptions of %"PRIu32" cores", cores_count * sizeof(struct cpuinfo_core), cores_count); goto cleanup; } clusters = HeapAlloc(heap, HEAP_ZERO_MEMORY, packages_count * sizeof(struct cpuinfo_cluster)); if (clusters == NULL) { cpuinfo_log_error("failed to allocate %zu bytes for descriptions of %"PRIu32" core clusters", packages_count * sizeof(struct cpuinfo_cluster), packages_count); goto cleanup; } packages = HeapAlloc(heap, HEAP_ZERO_MEMORY, packages_count * sizeof(struct cpuinfo_package)); if (packages == NULL) { cpuinfo_log_error("failed to allocate %zu bytes for descriptions of %"PRIu32" physical packages", packages_count * sizeof(struct cpuinfo_package), packages_count); goto cleanup; } for (uint32_t i = processors_count; i != 0; i--) { const uint32_t processor_id = i - 1; struct cpuinfo_processor* processor = processors + processor_id; /* Adjust core and package pointers for all logical processors */ struct cpuinfo_core* core = (struct cpuinfo_core*) ((uintptr_t) cores + (uintptr_t) processor->core); processor->core = core; struct cpuinfo_cluster* cluster = (struct cpuinfo_cluster*) ((uintptr_t) clusters + (uintptr_t) processor->cluster); processor->cluster = cluster; struct cpuinfo_package* package = (struct cpuinfo_package*) ((uintptr_t) packages + (uintptr_t) processor->package); processor->package = package; /* This can be overwritten by lower-index processors on the same package */ package->processor_start = processor_id; package->processor_count += 1; /* This can be overwritten by lower-index processors on the same cluster */ cluster->processor_start = processor_id; cluster->processor_count += 1; /* This can be overwritten by lower-index processors on the same core*/ core->processor_start = processor_id; core->processor_count += 1; } /* Set vendor/uarch/CPUID information for cores */ for (uint32_t i = cores_count; i != 0; i--) { const uint32_t global_core_id = i - 1; struct cpuinfo_core* core = cores + global_core_id; const struct cpuinfo_processor* processor = processors + core->processor_start; struct cpuinfo_package* package = (struct cpuinfo_package*) processor->package; struct cpuinfo_cluster* cluster = (struct cpuinfo_cluster*) processor->cluster; core->cluster = cluster; core->package = package; core->core_id = core_bits_mask & (processor->apic_id >> x86_processor.topology.core_bits_offset); core->vendor = x86_processor.vendor; core->uarch = x86_processor.uarch; core->cpuid = x86_processor.cpuid; /* This can be overwritten by lower-index cores on the same cluster/package */ cluster->core_start = global_core_id; cluster->core_count += 1; package->core_start = global_core_id; package->core_count += 1; } for (uint32_t i = 0; i < packages_count; i++) { struct cpuinfo_package* package = packages + i; struct cpuinfo_cluster* cluster = clusters + i; cluster->package = package; cluster->vendor = cores[cluster->core_start].vendor; cluster->uarch = cores[cluster->core_start].uarch; cluster->cpuid = cores[cluster->core_start].cpuid; package->cluster_start = i; package->cluster_count = 1; cpuinfo_x86_format_package_name(x86_processor.vendor, brand_string, package->name); } /* Count caches */ uint32_t l1i_count, l1d_count, l2_count, l3_count, l4_count; cpuinfo_x86_count_caches(processors_count, processors, &x86_processor, &l1i_count, &l1d_count, &l2_count, &l3_count, &l4_count); /* Allocate cache descriptions */ if (l1i_count != 0) { l1i = HeapAlloc(heap, HEAP_ZERO_MEMORY, l1i_count * sizeof(struct cpuinfo_cache)); if (l1i == NULL) { cpuinfo_log_error("failed to allocate %zu bytes for descriptions of %"PRIu32" L1I caches", l1i_count * sizeof(struct cpuinfo_cache), l1i_count); goto cleanup; } } if (l1d_count != 0) { l1d = HeapAlloc(heap, HEAP_ZERO_MEMORY, l1d_count * sizeof(struct cpuinfo_cache)); if (l1d == NULL) { cpuinfo_log_error("failed to allocate %zu bytes for descriptions of %"PRIu32" L1D caches", l1d_count * sizeof(struct cpuinfo_cache), l1d_count); goto cleanup; } } if (l2_count != 0) { l2 = HeapAlloc(heap, HEAP_ZERO_MEMORY, l2_count * sizeof(struct cpuinfo_cache)); if (l2 == NULL) { cpuinfo_log_error("failed to allocate %zu bytes for descriptions of %"PRIu32" L2 caches", l2_count * sizeof(struct cpuinfo_cache), l2_count); goto cleanup; } } if (l3_count != 0) { l3 = HeapAlloc(heap, HEAP_ZERO_MEMORY, l3_count * sizeof(struct cpuinfo_cache)); if (l3 == NULL) { cpuinfo_log_error("failed to allocate %zu bytes for descriptions of %"PRIu32" L3 caches", l3_count * sizeof(struct cpuinfo_cache), l3_count); goto cleanup; } } if (l4_count != 0) { l4 = HeapAlloc(heap, HEAP_ZERO_MEMORY, l4_count * sizeof(struct cpuinfo_cache)); if (l4 == NULL) { cpuinfo_log_error("failed to allocate %zu bytes for descriptions of %"PRIu32" L4 caches", l4_count * sizeof(struct cpuinfo_cache), l4_count); goto cleanup; } } /* Set cache information */ uint32_t l1i_index = UINT32_MAX, l1d_index = UINT32_MAX, l2_index = UINT32_MAX, l3_index = UINT32_MAX, l4_index = UINT32_MAX; uint32_t last_l1i_id = UINT32_MAX, last_l1d_id = UINT32_MAX; uint32_t last_l2_id = UINT32_MAX, last_l3_id = UINT32_MAX, last_l4_id = UINT32_MAX; for (uint32_t i = 0; i < processors_count; i++) { const uint32_t apic_id = processors[i].apic_id; if (x86_processor.cache.l1i.size != 0) { const uint32_t l1i_id = apic_id & ~bit_mask(x86_processor.cache.l1i.apic_bits); processors[i].cache.l1i = &l1i[l1i_index]; if (l1i_id != last_l1i_id) { /* new cache */ last_l1i_id = l1i_id; l1i[++l1i_index] = (struct cpuinfo_cache) { .size = x86_processor.cache.l1i.size, .associativity = x86_processor.cache.l1i.associativity, .sets = x86_processor.cache.l1i.sets, .partitions = x86_processor.cache.l1i.partitions, .line_size = x86_processor.cache.l1i.line_size, .flags = x86_processor.cache.l1i.flags, .processor_start = i, .processor_count = 1, }; } else { /* another processor sharing the same cache */ l1i[l1i_index].processor_count += 1; } processors[i].cache.l1i = &l1i[l1i_index]; } else { /* reset cache id */ last_l1i_id = UINT32_MAX; } if (x86_processor.cache.l1d.size != 0) { const uint32_t l1d_id = apic_id & ~bit_mask(x86_processor.cache.l1d.apic_bits); processors[i].cache.l1d = &l1d[l1d_index]; if (l1d_id != last_l1d_id) { /* new cache */ last_l1d_id = l1d_id; l1d[++l1d_index] = (struct cpuinfo_cache) { .size = x86_processor.cache.l1d.size, .associativity = x86_processor.cache.l1d.associativity, .sets = x86_processor.cache.l1d.sets, .partitions = x86_processor.cache.l1d.partitions, .line_size = x86_processor.cache.l1d.line_size, .flags = x86_processor.cache.l1d.flags, .processor_start = i, .processor_count = 1, }; } else { /* another processor sharing the same cache */ l1d[l1d_index].processor_count += 1; } processors[i].cache.l1d = &l1d[l1d_index]; } else { /* reset cache id */ last_l1d_id = UINT32_MAX; } if (x86_processor.cache.l2.size != 0) { const uint32_t l2_id = apic_id & ~bit_mask(x86_processor.cache.l2.apic_bits); processors[i].cache.l2 = &l2[l2_index]; if (l2_id != last_l2_id) { /* new cache */ last_l2_id = l2_id; l2[++l2_index] = (struct cpuinfo_cache) { .size = x86_processor.cache.l2.size, .associativity = x86_processor.cache.l2.associativity, .sets = x86_processor.cache.l2.sets, .partitions = x86_processor.cache.l2.partitions, .line_size = x86_processor.cache.l2.line_size, .flags = x86_processor.cache.l2.flags, .processor_start = i, .processor_count = 1, }; } else { /* another processor sharing the same cache */ l2[l2_index].processor_count += 1; } processors[i].cache.l2 = &l2[l2_index]; } else { /* reset cache id */ last_l2_id = UINT32_MAX; } if (x86_processor.cache.l3.size != 0) { const uint32_t l3_id = apic_id & ~bit_mask(x86_processor.cache.l3.apic_bits); processors[i].cache.l3 = &l3[l3_index]; if (l3_id != last_l3_id) { /* new cache */ last_l3_id = l3_id; l3[++l3_index] = (struct cpuinfo_cache) { .size = x86_processor.cache.l3.size, .associativity = x86_processor.cache.l3.associativity, .sets = x86_processor.cache.l3.sets, .partitions = x86_processor.cache.l3.partitions, .line_size = x86_processor.cache.l3.line_size, .flags = x86_processor.cache.l3.flags, .processor_start = i, .processor_count = 1, }; } else { /* another processor sharing the same cache */ l3[l3_index].processor_count += 1; } processors[i].cache.l3 = &l3[l3_index]; } else { /* reset cache id */ last_l3_id = UINT32_MAX; } if (x86_processor.cache.l4.size != 0) { const uint32_t l4_id = apic_id & ~bit_mask(x86_processor.cache.l4.apic_bits); processors[i].cache.l4 = &l4[l4_index]; if (l4_id != last_l4_id) { /* new cache */ last_l4_id = l4_id; l4[++l4_index] = (struct cpuinfo_cache) { .size = x86_processor.cache.l4.size, .associativity = x86_processor.cache.l4.associativity, .sets = x86_processor.cache.l4.sets, .partitions = x86_processor.cache.l4.partitions, .line_size = x86_processor.cache.l4.line_size, .flags = x86_processor.cache.l4.flags, .processor_start = i, .processor_count = 1, }; } else { /* another processor sharing the same cache */ l4[l4_index].processor_count += 1; } processors[i].cache.l4 = &l4[l4_index]; } else { /* reset cache id */ last_l4_id = UINT32_MAX; } } /* Commit changes */ cpuinfo_processors = processors; cpuinfo_cores = cores; cpuinfo_clusters = clusters; cpuinfo_packages = packages; cpuinfo_cache[cpuinfo_cache_level_1i] = l1i; cpuinfo_cache[cpuinfo_cache_level_1d] = l1d; cpuinfo_cache[cpuinfo_cache_level_2] = l2; cpuinfo_cache[cpuinfo_cache_level_3] = l3; cpuinfo_cache[cpuinfo_cache_level_4] = l4; cpuinfo_processors_count = processors_count; cpuinfo_cores_count = cores_count; cpuinfo_clusters_count = packages_count; cpuinfo_packages_count = packages_count; cpuinfo_cache_count[cpuinfo_cache_level_1i] = l1i_count; cpuinfo_cache_count[cpuinfo_cache_level_1d] = l1d_count; cpuinfo_cache_count[cpuinfo_cache_level_2] = l2_count; cpuinfo_cache_count[cpuinfo_cache_level_3] = l3_count; cpuinfo_cache_count[cpuinfo_cache_level_4] = l4_count; cpuinfo_max_cache_size = cpuinfo_compute_max_cache_size(&processors[0]); cpuinfo_global_uarch = (struct cpuinfo_uarch_info) { .uarch = x86_processor.uarch, .cpuid = x86_processor.cpuid, .processor_count = processors_count, .core_count = cores_count, }; MemoryBarrier(); cpuinfo_is_initialized = true; processors = NULL; cores = NULL; clusters = NULL; packages = NULL; l1i = l1d = l2 = l3 = l4 = NULL; cleanup: if (processors != NULL) { HeapFree(heap, 0, processors); } if (cores != NULL) { HeapFree(heap, 0, cores); } if (clusters != NULL) { HeapFree(heap, 0, clusters); } if (packages != NULL) { HeapFree(heap, 0, packages); } if (l1i != NULL) { HeapFree(heap, 0, l1i); } if (l1d != NULL) { HeapFree(heap, 0, l1d); } if (l2 != NULL) { HeapFree(heap, 0, l2); } if (l3 != NULL) { HeapFree(heap, 0, l3); } if (l4 != NULL) { HeapFree(heap, 0, l4); } return TRUE; }