Fix incorrect CPU topology initialization (#3304) (#3487)

The error handling in the global object that defines CPU topology in oneDAL/DAAL and in global daal::services::Environment object was improved.
Several related bugs were fixed.

- cpp/daal/src/services/service_topo.h and cpp/daal/src/services/service_topo.cpp files re-worked to make it less error prone.
- The logic in setChkProcessAffinityConsistency() was changed to allow affinity masks that contain zeros. Previously having the process affinity mask with zeros lead to undefined behavior because the global object that defines CPU topology ended up in non-initialized state in that case.
- Out-of-bound memory access issues in CPU topology initialization were fixed.
- The constructor of daal::services::Environment class was updated to always call getCpuId() method which initializes the Environment instance. Without this call an uninitialized version of Environment might be used.
- The behavior of daal::services::Environment::getInstance() function that returns the pointer to a global DAAL's Environment object was changed. Now the function returns nullptr in case the global Environment object is uninitialized. The respective null pointer checks were added into the calling code.
- The behavior of daal::services::Environment::initNumberOfThreads() was changed to handle the case of hyper-threading correctly, setting the maximal number of threads available to oneDAL equal to the number of physical CPU cores.

Author: Vika-F

cpp/daal/include/algorithms/algorithm_base_common.h

@@ -116,7 +116,9 @@ class AlgorithmIfaceImpl : public AlgorithmIface
 protected:
     services::Status getEnvironment()
     {
-        int cpuid = (int)daal::services::Environment::getInstance()->getCpuId();
+        daal::services::Environment * env = daal::services::Environment::getInstance();
+        if (!env) return services::Status(services::ErrorCpuNotSupported);
+        int cpuid = (int)env->getCpuId();
         if (cpuid < 0) return services::Status(services::ErrorCpuNotSupported);
         _env.cpuid           = cpuid;
         _env.cpuid_init_flag = true;

cpp/daal/include/services/daal_defines.h

@@ -41,6 +41,12 @@
     #define TARGET_RISCV64
 #endif
 
+#if !defined(TARGET_X86_64)
+    #ifndef DAAL_CPU_TOPO_DISABLED
+        #define DAAL_CPU_TOPO_DISABLED
+    #endif
+#endif
+
 #if (defined(__INTEL_COMPILER) || defined(__INTEL_LLVM_COMPILER)) && !defined(SYCL_LANGUAGE_VERSION)
     #define DAAL_INTEL_CPP_COMPILER
 #endif

cpp/daal/include/services/env_detect.h

@@ -65,6 +65,12 @@ class DAAL_EXPORT Environment : public Base
      */
     static Environment * getInstance();
 
+    /**
+     *  Returns the status of the environment instance initialization
+     *  \return Status of the environment instance initialization
+     */
+    static int getStatus();
+
     /**
      *  Decreases the instance counter
      *  \return The return code

cpp/daal/src/data_management/data_conversion.cpp

@@ -42,7 +42,9 @@ static bool tryToCopyFuncAVX512(const size_t nrows, const size_t ncols, void * d
 
     if (!ptr)
     {
-        int cpuid = (int)daal::services::Environment::getInstance()->getCpuId();
+        daal::services::Environment * env = daal::services::Environment::getInstance();
+        if (!env) return false; // Environment not initialized
+        int cpuid = (int)env->getCpuId();
 
         switch (cpuid)
         {

cpp/daal/src/externals/service_service_mkl.h

@@ -62,19 +62,40 @@ struct MklService
         // return memmove_s(dest, destSize, src, smax);
     }
 
-    static int serv_get_ht() { return (serv_get_ncorespercpu() > 1 ? 1 : 0); }
+    static int serv_get_ht()
+    {
+        const int ncorespercpu = serv_get_ncorespercpu();
+        if (ncorespercpu < 0)
+        {
+            // Failed to get the number of cores per CPU.
+            return -1;
+        }
+        return (ncorespercpu > 1 ? 1 : 0);
+    }
 
     static int serv_get_ncpus()
     {
-        unsigned int ncores = daal::services::internal::_internal_daal_GetProcessorCoreCount();
+        if (daal::services::internal::_internal_daal_GetStatus() != 0)
+        {
+            // CPU topology initialization failed;
+            return -1;
+        }
+        const unsigned int ncores = daal::services::internal::_internal_daal_GetProcessorCoreCount();
         return (ncores ? ncores : 1);
     }
 
     static int serv_get_ncorespercpu()
     {
-        unsigned int nlogicalcpu = daal::services::internal::_internal_daal_GetProcessorCoreCount();
-        unsigned int ncpus       = serv_get_ncpus();
-        return (ncpus > 0 && nlogicalcpu > 0 && nlogicalcpu > ncpus ? nlogicalcpu / ncpus : 1);
+        if (daal::services::internal::_internal_daal_GetStatus() != 0)
+        {
+            // CPU topology initialization failed;
+            return -1;
+        }
+        const unsigned int nlogicalcpu = daal::services::internal::_internal_daal_GetSysLogicalProcessorCount();
+        const unsigned int ncpus       = serv_get_ncpus();
+        // On hybrid systems, return 2 in case of hyper-threading enabled at least on some cores,
+        // 1 otherwise
+        return (ncpus > 0 && nlogicalcpu > 0 && nlogicalcpu > ncpus ? (nlogicalcpu - ncpus + 1) / ncpus : 1);
     }
 
     // TODO: The real call should be delegated to a backend library if the option is supported

cpp/daal/src/services/compiler/generic/env_detect_features.cpp

@@ -27,6 +27,11 @@
 
 #if defined(TARGET_X86_64)
     #include <immintrin.h>
+
+    #if defined(__GNUC__) || defined(__clang__)
+        #include <cpuid.h> // __cpuidex
+    #endif
+
 #elif defined(TARGET_ARM)
     #include <sys/auxv.h>
     #include <asm/hwcap.h>
@@ -54,10 +59,11 @@ void __daal_serv_CPUHasAVX512f_enable_it_mac();
 #if defined(TARGET_X86_64)
 void run_cpuid(uint32_t eax, uint32_t ecx, uint32_t * abcd)
 {
-    #if defined(_MSC_VER)
-    __cpuidex((int *)abcd, eax, ecx);
+    #if defined(_MSC_VER) || (defined(__clang__) && __clang_major__ > 18) || (defined(__GNUC__) && !defined(__clang__))
+    __cpuidex(reinterpret_cast<int *>(abcd), eax, ecx);
     #else
-    uint32_t ebx, edx;
+
+    uint32_t ebx = 0, edx = 0;
         #if defined(__i386__) && defined(__PIC__)
     /* in case of PIC under 32-bit EBX cannot be clobbered */
     __asm__("movl %%ebx, %%edi \n\t cpuid \n\t xchgl %%ebx, %%edi" : "=D"(ebx), "+a"(eax), "+c"(ecx), "=d"(edx));

cpp/daal/src/services/env_detect.cpp

@@ -54,7 +54,15 @@ void daal_free_buffers();
 DAAL_EXPORT daal::services::Environment * daal::services::Environment::getInstance()
 {
     static daal::services::Environment instance;
-    return &instance;
+    if (isInit)
+        return &instance;
+    else
+        return nullptr;
+}
+
+DAAL_EXPORT int daal::services::Environment::getStatus()
+{
+    return daal::services::internal::_internal_daal_GetStatus();
 }
 
 DAAL_EXPORT int daal::services::Environment::freeInstance()
@@ -129,6 +137,8 @@ DAAL_EXPORT daal::services::Environment::Environment() : _schedulerHandle {}, _g
 {
     _env.cpuid_init_flag = false;
     _env.cpuid           = -1;
+    // Initialize the Environment to prevent from using it uninitialized
+    getCpuId();
 }
 
 DAAL_EXPORT daal::services::Environment::Environment(const Environment & e) : daal::services::Environment::Environment() {}
@@ -145,15 +155,25 @@ DAAL_EXPORT void daal::services::Environment::initNumberOfThreads()
     daal::setSchedulerHandle(&_schedulerHandle);
 #endif
     /* if HT enabled - set _numThreads to physical cores num */
-    if (daal::internal::ServiceInst::serv_get_ht())
+    const int htStatus = daal::internal::ServiceInst::serv_get_ht();
+    if (htStatus < 0)
     {
-        /* Number of cores = number of cpu packages * number of cores per cpu package */
-        int ncores = daal::internal::ServiceInst::serv_get_ncpus() * daal::internal::ServiceInst::serv_get_ncorespercpu();
+        // Failed to get the number of CPUs. Environment cannot be initialized
+        return;
+    }
+    else if (htStatus > 0)
+    {
+        const int ncpus = daal::internal::ServiceInst::serv_get_ncpus();
+        if (ncpus < 0)
+        {
+            // Failed to get the number of CPUs. Environment cannot be initialized
+            return;
+        }
 
         /*  Re-set number of threads if ncores is valid and different to _numThreads */
-        if ((ncores > 0) && (ncores < _daal_threader_get_max_threads()))
+        if ((ncpus > 0) && (ncpus < _daal_threader_get_max_threads()))
         {
-            daal::services::Environment::setNumberOfThreads(ncores);
+            daal::services::Environment::setNumberOfThreads(ncpus);
         }
     }
     isInit = true;

cpp/daal/src/services/library_version_info.cpp

@@ -53,12 +53,25 @@ DAAL_EXPORT daal::services::LibraryVersionInfo::LibraryVersionInfo()
       productStatus(PRODUCTSTATUS),
       build(BUILD),
       build_rev(BUILD_REV),
-      name(PRODUCT_NAME_STR),
+      name(PRODUCT_NAME_STR)
+{
 #ifndef DAAL_REF
-      processor(cpu_long_names[daal::services::Environment::getInstance()->getCpuId()])
+    daal::services::Environment * env = daal::services::Environment::getInstance();
+    if (env)
+    {
+        processor = cpu_long_names[env->getCpuId()];
+    }
+    else
+    {
+    #if (!defined(DAAL_NOTHROW_EXCEPTIONS))
+        int error = daal::services::Environment::getStatus();
+        throw std::runtime_error("Environment not initialized, cannot get processor info, error code: " + std::to_string(error));
+    #else
+        processor = cpu_long_names[0];
+    #endif
+    }
 #else
-      processor(cpu_long_names[0])
+    processor = cpu_long_names[0];
 #endif
-{}
-
+}
 DAAL_EXPORT daal::services::LibraryVersionInfo::~LibraryVersionInfo() {}