From: Jan Vorlicek Date: Wed, 19 Apr 2017 22:13:30 +0000 (-0700) Subject: Add NUMA and thread affinity support for Unix (dotnet/coreclr#10938) X-Git-Tag: submit/tizen/20210909.063632~11030^2~7191 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=d4ffe751d2c631dbdbcbacf9e8615974a207eb31;p=platform%2Fupstream%2Fdotnet%2Fruntime.git Add NUMA and thread affinity support for Unix (dotnet/coreclr#10938) * Add NUMA and thread affinity support for Unix This change adds new PAL functions for NUMA and thread affinity support for Unix and also enables related code in GC and VM for FEATURE_PAL. It doesn't reflect the limits imposed by CGROUPS on systems with CGROUPS enables yet. Commit migrated from https://github.com/dotnet/coreclr/commit/48ebab8270a96e505a2be128a4875f903ef1c4b2 --- diff --git a/src/coreclr/src/dlls/mscordac/mscordac_unixexports.src b/src/coreclr/src/dlls/mscordac/mscordac_unixexports.src index c2c96fa..b0c3b04 100644 --- a/src/coreclr/src/dlls/mscordac/mscordac_unixexports.src +++ b/src/coreclr/src/dlls/mscordac/mscordac_unixexports.src @@ -21,6 +21,7 @@ PAL_GetPALDirectoryW PAL_GetResourceString PAL_get_stdout PAL_get_stderr +PAL_GetCurrentThread PAL_GetSymbolModuleBase PAL_GetTransportPipeName PAL_InitializeDLL diff --git a/src/coreclr/src/gc/gc.cpp b/src/coreclr/src/gc/gc.cpp index 6d53c89..46820d1 100644 --- a/src/coreclr/src/gc/gc.cpp +++ b/src/coreclr/src/gc/gc.cpp @@ -5151,7 +5151,6 @@ void gc_heap::destroy_thread_support () } } -#if !defined(FEATURE_PAL) void set_thread_group_affinity_for_heap(int heap_number, GCThreadAffinity* affinity) { affinity->Group = GCThreadAffinity::None; @@ -5231,7 +5230,6 @@ void set_thread_affinity_mask_for_heap(int heap_number, GCThreadAffinity* affini } } } -#endif // !FEATURE_PAL bool gc_heap::create_gc_thread () { @@ -5241,7 +5239,6 @@ bool gc_heap::create_gc_thread () affinity.Group = GCThreadAffinity::None; affinity.Processor = GCThreadAffinity::None; -#if !defined(FEATURE_PAL) if (!gc_thread_no_affinitize_p) { // We are about to set affinity for GC threads. It is a good place to set up NUMA and @@ -5252,7 +5249,6 @@ bool gc_heap::create_gc_thread () else set_thread_affinity_mask_for_heap(heap_number, &affinity); } -#endif // !FEATURE_PAL return GCToOSInterface::CreateThread(gc_thread_stub, this, &affinity); } diff --git a/src/coreclr/src/inc/utilcode.h b/src/coreclr/src/inc/utilcode.h index 5394f1b..06133cf 100644 --- a/src/coreclr/src/inc/utilcode.h +++ b/src/coreclr/src/inc/utilcode.h @@ -1380,7 +1380,7 @@ public: static BOOL CanEnableGCNumaAware(); static void InitNumaNodeInfo(); -#if !defined(FEATURE_REDHAWK)&& !defined(FEATURE_PAL) +#if !defined(FEATURE_REDHAWK) private: // apis types //GetNumaHighestNodeNumber() @@ -1448,7 +1448,7 @@ public: static DWORD CalculateCurrentProcessorNumber(); //static void PopulateCPUUsageArray(void * infoBuffer, ULONG infoSize); -#if !defined(FEATURE_REDHAWK) && !defined(FEATURE_PAL) +#if !defined(FEATURE_REDHAWK) private: //GetLogicalProcessorInforomationEx() typedef BOOL diff --git a/src/coreclr/src/pal/inc/pal.h b/src/coreclr/src/pal/inc/pal.h index 5d2739e..9e97edc 100644 --- a/src/coreclr/src/pal/inc/pal.h +++ b/src/coreclr/src/pal/inc/pal.h @@ -4826,6 +4826,161 @@ RegisterEventSourceW ( #endif // !UNICODE // +// NUMA related APIs +// + +typedef enum _PROCESSOR_CACHE_TYPE { + CacheUnified, + CacheInstruction, + CacheData, + CacheTrace +} PROCESSOR_CACHE_TYPE; + +typedef struct _PROCESSOR_NUMBER { + WORD Group; + BYTE Number; + BYTE Reserved; +} PROCESSOR_NUMBER, *PPROCESSOR_NUMBER; + +typedef enum _LOGICAL_PROCESSOR_RELATIONSHIP { + RelationProcessorCore, + RelationNumaNode, + RelationCache, + RelationProcessorPackage, + RelationGroup, + RelationAll = 0xffff +} LOGICAL_PROCESSOR_RELATIONSHIP; + +typedef ULONG_PTR KAFFINITY; + +#define ANYSIZE_ARRAY 1 + +typedef struct _GROUP_AFFINITY { + KAFFINITY Mask; + WORD Group; + WORD Reserved[3]; +} GROUP_AFFINITY, *PGROUP_AFFINITY; + +typedef struct _PROCESSOR_GROUP_INFO { + BYTE MaximumProcessorCount; + BYTE ActiveProcessorCount; + BYTE Reserved[38]; + KAFFINITY ActiveProcessorMask; +} PROCESSOR_GROUP_INFO, *PPROCESSOR_GROUP_INFO; + +typedef struct _PROCESSOR_RELATIONSHIP { + BYTE Flags; + BYTE EfficiencyClass; + BYTE Reserved[21]; + WORD GroupCount; + GROUP_AFFINITY GroupMask[ANYSIZE_ARRAY]; +} PROCESSOR_RELATIONSHIP, *PPROCESSOR_RELATIONSHIP; + +typedef struct _GROUP_RELATIONSHIP { + WORD MaximumGroupCount; + WORD ActiveGroupCount; + BYTE Reserved[20]; + PROCESSOR_GROUP_INFO GroupInfo[ANYSIZE_ARRAY]; +} GROUP_RELATIONSHIP, *PGROUP_RELATIONSHIP; + +typedef struct _NUMA_NODE_RELATIONSHIP { + DWORD NodeNumber; + BYTE Reserved[20]; + GROUP_AFFINITY GroupMask; +} NUMA_NODE_RELATIONSHIP, *PNUMA_NODE_RELATIONSHIP; + +typedef struct _CACHE_RELATIONSHIP { + BYTE Level; + BYTE Associativity; + WORD LineSize; + DWORD CacheSize; + PROCESSOR_CACHE_TYPE Type; + BYTE Reserved[20]; + GROUP_AFFINITY GroupMask; +} CACHE_RELATIONSHIP, *PCACHE_RELATIONSHIP; + +typedef struct _SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX { + LOGICAL_PROCESSOR_RELATIONSHIP Relationship; + DWORD Size; + union { + PROCESSOR_RELATIONSHIP Processor; + NUMA_NODE_RELATIONSHIP NumaNode; + CACHE_RELATIONSHIP Cache; + GROUP_RELATIONSHIP Group; + }; +} SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX, *PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX; + + +PALIMPORT +BOOL +PALAPI +GetNumaHighestNodeNumber( + OUT PULONG HighestNodeNumber +); + +PALIMPORT +BOOL +PALAPI +GetNumaProcessorNodeEx( + IN PPROCESSOR_NUMBER Processor, + OUT PUSHORT NodeNumber +); + +PALIMPORT +LPVOID +PALAPI +VirtualAllocExNuma( + IN HANDLE hProcess, + IN OPTIONAL LPVOID lpAddress, + IN SIZE_T dwSize, + IN DWORD flAllocationType, + IN DWORD flProtect, + IN DWORD nndPreferred +); + +PALIMPORT +BOOL +PALAPI +GetLogicalProcessorInformationEx( + IN LOGICAL_PROCESSOR_RELATIONSHIP RelationshipType, + OUT OPTIONAL PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX Buffer, + IN OUT PDWORD ReturnedLength +); + +PALIMPORT +BOOL +PALAPI +SetThreadGroupAffinity( + IN HANDLE hThread, + IN const GROUP_AFFINITY *GroupAffinity, + OUT OPTIONAL PGROUP_AFFINITY PreviousGroupAffinity +); + +PALIMPORT +BOOL +PALAPI +GetThreadGroupAffinity( + IN HANDLE hThread, + OUT PGROUP_AFFINITY GroupAffinity +); + +PALIMPORT +VOID +PALAPI +GetCurrentProcessorNumberEx( + OUT PPROCESSOR_NUMBER ProcNumber +); + +PALIMPORT +BOOL +PALAPI +GetProcessAffinityMask( + IN HANDLE hProcess, + OUT PDWORD_PTR lpProcessAffinityMask, + OUT PDWORD_PTR lpSystemAffinityMask +); + +// // The types of events that can be logged. // #define EVENTLOG_SUCCESS 0x0000 diff --git a/src/coreclr/src/pal/src/CMakeLists.txt b/src/coreclr/src/pal/src/CMakeLists.txt index 5314cdf8..145c2c9 100644 --- a/src/coreclr/src/pal/src/CMakeLists.txt +++ b/src/coreclr/src/pal/src/CMakeLists.txt @@ -181,6 +181,7 @@ set(SOURCES misc/sysinfo.cpp misc/time.cpp misc/utils.cpp + numa/numa.cpp objmgr/palobjbase.cpp objmgr/shmobject.cpp objmgr/shmobjectmanager.cpp @@ -372,6 +373,12 @@ if(CMAKE_SYSTEM_NAME STREQUAL NetBSD) ) endif(CMAKE_SYSTEM_NAME STREQUAL NetBSD) +if(HAVE_NUMA_H) + target_link_libraries(coreclrpal + numa + ) +endif(HAVE_NUMA_H) + add_subdirectory(examples) if(FEATURE_EVENT_TRACE) diff --git a/src/coreclr/src/pal/src/config.h.in b/src/coreclr/src/pal/src/config.h.in index c2939f3..e3024ac 100644 --- a/src/coreclr/src/pal/src/config.h.in +++ b/src/coreclr/src/pal/src/config.h.in @@ -21,6 +21,8 @@ #cmakedefine01 HAVE_SYS_SYSCTL_H #cmakedefine01 HAVE_GNU_LIBNAMES_H #cmakedefine01 HAVE_PRCTL_H +#cmakedefine01 HAVE_NUMA_H +#cmakedefine01 HAVE_PTHREAD_NP_H #cmakedefine01 HAVE_KQUEUE #cmakedefine01 HAVE_GETPWUID_R @@ -33,6 +35,8 @@ #cmakedefine01 HAVE_PTHREAD_GETATTR_NP #cmakedefine01 HAVE_PTHREAD_GETCPUCLOCKID #cmakedefine01 HAVE_PTHREAD_SIGQUEUE +#cmakedefine01 HAVE_PTHREAD_GETAFFINITY_NP +#cmakedefine01 HAVE_CPUSET_T #cmakedefine01 HAVE_SIGRETURN #cmakedefine01 HAVE__THREAD_SYS_SIGRETURN #cmakedefine01 HAVE_COPYSIGN @@ -57,6 +61,7 @@ #cmakedefine01 HAS_SYSV_SEMAPHORES #cmakedefine01 HAS_PTHREAD_MUTEXES #cmakedefine01 HAVE_TTRACE +#cmakedefine01 HAVE_SCHED_GETAFFINITY #cmakedefine HAVE_UNW_GET_SAVE_LOC #cmakedefine HAVE_UNW_GET_ACCESSORS #cmakedefine01 HAVE_XSWDEV diff --git a/src/coreclr/src/pal/src/configure.cmake b/src/coreclr/src/pal/src/configure.cmake index 03c7343..f9a23e8 100644 --- a/src/coreclr/src/pal/src/configure.cmake +++ b/src/coreclr/src/pal/src/configure.cmake @@ -35,6 +35,8 @@ check_include_files(libunwind.h HAVE_LIBUNWIND_H) check_include_files(runetype.h HAVE_RUNETYPE_H) check_include_files(semaphore.h HAVE_SEMAPHORE_H) check_include_files(sys/prctl.h HAVE_PRCTL_H) +check_include_files(numa.h HAVE_NUMA_H) +check_include_files(pthread_np.h HAVE_PTHREAD_NP_H) if(NOT CMAKE_SYSTEM_NAME STREQUAL FreeBSD AND NOT CMAKE_SYSTEM_NAME STREQUAL NetBSD) set(CMAKE_REQUIRED_FLAGS "-ldl") @@ -69,6 +71,7 @@ check_library_exists(${PTHREAD_LIBRARY} pthread_attr_get_np "" HAVE_PTHREAD_ATTR check_library_exists(${PTHREAD_LIBRARY} pthread_getattr_np "" HAVE_PTHREAD_GETATTR_NP) check_library_exists(${PTHREAD_LIBRARY} pthread_getcpuclockid "" HAVE_PTHREAD_GETCPUCLOCKID) check_library_exists(${PTHREAD_LIBRARY} pthread_sigqueue "" HAVE_PTHREAD_SIGQUEUE) +check_library_exists(${PTHREAD_LIBRARY} pthread_getaffinity_np "" HAVE_PTHREAD_GETAFFINITY_NP) check_function_exists(sigreturn HAVE_SIGRETURN) check_function_exists(_thread_sys_sigreturn HAVE__THREAD_SYS_SIGRETURN) @@ -119,6 +122,14 @@ int main(int argc, char **argv) { }" HAVE_UNW_GET_ACCESSORS) set(CMAKE_REQUIRED_LIBRARIES) +check_cxx_source_compiles(" +#include +int main(int argc, char **argv) { + cpuset_t cpuSet; + + return 0; +}" HAVE_CPUSET_T) + check_struct_has_member ("struct stat" st_atimespec "sys/types.h;sys/stat.h" HAVE_STAT_TIMESPEC) check_struct_has_member ("struct stat" st_atimensec "sys/types.h;sys/stat.h" HAVE_STAT_NSEC) check_struct_has_member ("struct tm" tm_gmtoff time.h HAVE_TM_GMTOFF) diff --git a/src/coreclr/src/pal/src/include/pal/dbgmsg.h b/src/coreclr/src/pal/src/include/pal/dbgmsg.h index 7a49fc0..052c6fa 100644 --- a/src/coreclr/src/pal/src/include/pal/dbgmsg.h +++ b/src/coreclr/src/pal/src/include/pal/dbgmsg.h @@ -194,7 +194,7 @@ typedef enum #ifdef FEATURE_PAL_SXS DCI_SXS, #endif // FEATURE_PAL_SXS - + DCI_NUMA, DCI_LAST } DBG_CHANNEL_ID; diff --git a/src/coreclr/src/pal/src/include/pal/numa.h b/src/coreclr/src/pal/src/include/pal/numa.h new file mode 100644 index 0000000..4fb2308 --- /dev/null +++ b/src/coreclr/src/pal/src/include/pal/numa.h @@ -0,0 +1,39 @@ +// Licensed to the .NET Foundation under one or more agreements. +// The .NET Foundation licenses this file to you under the MIT license. +// See the LICENSE file in the project root for more information. + +/*++ + + + +Module Name: + + include/pal/numa.h + +Abstract: + + Header file for the NUMA functions. + + + +--*/ + +#ifndef _PAL_NUMA_H_ +#define _PAL_NUMA_H_ + +#ifdef __cplusplus +extern "C" +{ +#endif // __cplusplus + +BOOL +NUMASupportInitialize(); + +VOID +NUMASupportCleanup(); + +#ifdef __cplusplus +} +#endif // __cplusplus + +#endif /* _PAL_CRITSECT_H_ */ diff --git a/src/coreclr/src/pal/src/init/pal.cpp b/src/coreclr/src/pal/src/init/pal.cpp index 8b0e0f5..2fdafe4 100644 --- a/src/coreclr/src/pal/src/init/pal.cpp +++ b/src/coreclr/src/pal/src/init/pal.cpp @@ -42,6 +42,7 @@ SET_DEFAULT_DEBUG_CHANNEL(PAL); // some headers have code with asserts, so do th #include "pal/debug.h" #include "pal/locale.h" #include "pal/init.h" +#include "pal/numa.h" #include "pal/stackstring.hpp" #if HAVE_MACH_EXCEPTIONS @@ -523,6 +524,12 @@ Initialize( goto CLEANUP15; } + if (FALSE == NUMASupportInitialize()) + { + ERROR("Unable to initialize NUMA support\n"); + goto CLEANUP15; + } + TRACE("First-time PAL initialization complete.\n"); init_count++; @@ -548,6 +555,7 @@ Initialize( } goto done; + NUMASupportCleanup(); /* No cleanup required for CRTInitStdStreams */ CLEANUP15: FILECleanupStdHandles(); diff --git a/src/coreclr/src/pal/src/map/virtual.cpp b/src/coreclr/src/pal/src/map/virtual.cpp index d52ba1e..7e00843 100644 --- a/src/coreclr/src/pal/src/map/virtual.cpp +++ b/src/coreclr/src/pal/src/map/virtual.cpp @@ -1350,7 +1350,6 @@ done: return pRetVal; } - /*++ Function: VirtualFree diff --git a/src/coreclr/src/pal/src/misc/sysinfo.cpp b/src/coreclr/src/pal/src/misc/sysinfo.cpp index fff0518..70fe3e6 100644 --- a/src/coreclr/src/pal/src/misc/sysinfo.cpp +++ b/src/coreclr/src/pal/src/misc/sysinfo.cpp @@ -94,6 +94,39 @@ SET_DEFAULT_DEBUG_CHANNEL(MISC); #endif #endif // __APPLE__ +/*++ +Function: + GetNumberOfProcessors + +Return number of processors available for the current process +--*/ +int GetNumberOfProcessors() +{ + int nrcpus = 0; + +#if HAVE_SYSCONF + nrcpus = sysconf(_SC_NPROCESSORS_ONLN); + if (nrcpus < 1) + { + ASSERT("sysconf failed for _SC_NPROCESSORS_ONLN (%d)\n", errno); + } +#elif HAVE_SYSCTL + int rc; + size_t sz; + int mib[2]; + + sz = sizeof(nrcpus); + mib[0] = CTL_HW; + mib[1] = HW_NCPU; + rc = sysctl(mib, 2, &nrcpus, &sz, NULL, 0); + if (rc != 0) + { + ASSERT("sysctl failed for HW_NCPU (%d)\n", errno); + } +#endif // HAVE_SYSCONF + + return nrcpus; +} /*++ Function: @@ -137,27 +170,7 @@ GetSystemInfo( lpSystemInfo->dwPageSize = pagesize; lpSystemInfo->dwActiveProcessorMask_PAL_Undefined = 0; -#if HAVE_SYSCONF - nrcpus = sysconf(_SC_NPROCESSORS_ONLN); - if (nrcpus < 1) - { - ASSERT("sysconf failed for _SC_NPROCESSORS_ONLN (%d)\n", errno); - } -#elif HAVE_SYSCTL - int rc; - size_t sz; - int mib[2]; - - sz = sizeof(nrcpus); - mib[0] = CTL_HW; - mib[1] = HW_NCPU; - rc = sysctl(mib, 2, &nrcpus, &sz, NULL, 0); - if (rc != 0) - { - ASSERT("sysctl failed for HW_NCPU (%d)\n", errno); - } -#endif // HAVE_SYSCONF - + nrcpus = GetNumberOfProcessors(); TRACE("dwNumberOfProcessors=%d\n", nrcpus); lpSystemInfo->dwNumberOfProcessors = nrcpus; diff --git a/src/coreclr/src/pal/src/numa/numa.cpp b/src/coreclr/src/pal/src/numa/numa.cpp new file mode 100644 index 0000000..549c10a --- /dev/null +++ b/src/coreclr/src/pal/src/numa/numa.cpp @@ -0,0 +1,692 @@ +// Licensed to the .NET Foundation under one or more agreements. +// The .NET Foundation licenses this file to you under the MIT license. +// See the LICENSE file in the project root for more information. + +/*++ + + + +Module Name: + + numa.cpp + +Abstract: + + Implementation of NUMA related APIs + +--*/ + +#include "pal/dbgmsg.h" +SET_DEFAULT_DEBUG_CHANNEL(NUMA); + +#include "pal/palinternal.h" +#include "pal/dbgmsg.h" +#include "pal/numa.h" +#include "pal/corunix.hpp" +#include "pal/thread.hpp" + +#if HAVE_NUMA_H +#include +#include +#endif + +#if HAVE_PTHREAD_NP_H +#include +#endif + +#include + +using namespace CorUnix; + +#if HAVE_CPUSET_T +typedef cpuset_t cpu_set_t; +#endif + +int GetNumberOfProcessors(); + +// CPU affinity descriptor +struct CpuAffinity +{ + // NUMA node + BYTE Node; + // CPU number relative to the group the CPU is in + BYTE Number; + // CPU group + WORD Group; +}; + +// Array mapping global CPU index to its affinity +CpuAffinity *g_cpuToAffinity = NULL; + +// Array mapping CPU group and index in the group to the global CPU index +short *g_groupAndIndexToCpu = NULL; +// Array mapping CPU group to the corresponding affinity mask of the CPUs in the group +KAFFINITY *g_groupToCpuMask = NULL; +// Array mapping CPU group to the number of processors in the group +BYTE *g_groupToCpuCount = NULL; + +// Total number of processors in the system +int g_cpuCount = 0; +// Total number of CPU groups +int g_groupCount = 0; +// The highest NUMA node available +int g_highestNumaNode = 0; + +static const int MaxCpusPerGroup = 8 * sizeof(KAFFINITY); +static const WORD NO_GROUP = 0xffff; + +/*++ +Function: + AllocateLookupArrays + +Allocate CPU and group lookup arrays +--*/ +VOID +AllocateLookupArrays() +{ + g_groupAndIndexToCpu = (short*)malloc(g_groupCount * MaxCpusPerGroup * sizeof(short)); + g_cpuToAffinity = (CpuAffinity*)malloc(g_cpuCount * sizeof(CpuAffinity)); + g_groupToCpuMask = (KAFFINITY*)malloc(g_groupCount * sizeof(KAFFINITY)); + g_groupToCpuCount = (BYTE*)malloc(g_groupCount * sizeof(BYTE)); + + memset(g_groupAndIndexToCpu, 0xff, g_groupCount * MaxCpusPerGroup * sizeof(short)); + memset(g_cpuToAffinity, 0xff, g_cpuCount * sizeof(CpuAffinity)); + memset(g_groupToCpuMask, 0, g_groupCount * sizeof(KAFFINITY)); + memset(g_groupToCpuCount, 0, g_groupCount * sizeof(BYTE)); +} + +/*++ +Function: + FreeLookupArrays + +Free CPU and group lookup arrays +--*/ +VOID +FreeLookupArrays() +{ + free(g_groupAndIndexToCpu); + free(g_cpuToAffinity); + free(g_groupToCpuMask); + free(g_groupToCpuCount); + + g_groupAndIndexToCpu = NULL; + g_cpuToAffinity = NULL; + g_groupToCpuMask = NULL; + g_groupToCpuCount = NULL; +} + +/*++ +Function: + GetFullAffinityMask + +Get affinity mask for the specified number of processors with all +the processors enabled. +--*/ +KAFFINITY GetFullAffinityMask(int cpuCount) +{ + return ((KAFFINITY)1 << (cpuCount)) - 1; +} + +/*++ +Function: + NUMASupportInitialize + +Initialize data structures for getting and setting thread affinities to processors and +querying NUMA related processor information. +On systems with no NUMA support, it behaves as if there was a single NUMA node with +a single group of processors. +--*/ +BOOL +NUMASupportInitialize() +{ +#if HAVE_NUMA_H + if (numa_available() != -1) + { + struct bitmask *mask = numa_allocate_cpumask(); + int numaNodesCount = numa_max_node() + 1; + + g_cpuCount = numa_num_possible_cpus(); + g_groupCount = 0; + + for (int i = 0; i < numaNodesCount; i++) + { + int st = numa_node_to_cpus(i, mask); + // The only failure that can happen is that the mask is not large enough + // but that cannot happen since the mask was allocated by numa_allocate_cpumask + _ASSERTE(st == 0); + unsigned int nodeCpuCount = numa_bitmask_weight(mask); + unsigned int nodeGroupCount = (nodeCpuCount + MaxCpusPerGroup - 1) / MaxCpusPerGroup; + g_groupCount += nodeGroupCount; + } + + AllocateLookupArrays(); + + WORD currentGroup = 0; + int currentGroupCpus = 0; + + for (int i = 0; i < numaNodesCount; i++) + { + int st = numa_node_to_cpus(i, mask); + // The only failure that can happen is that the mask is not large enough + // but that cannot happen since the mask was allocated by numa_allocate_cpumask + _ASSERTE(st == 0); + unsigned int nodeCpuCount = numa_bitmask_weight(mask); + unsigned int nodeGroupCount = (nodeCpuCount + MaxCpusPerGroup - 1) / MaxCpusPerGroup; + for (int j = 0; j < g_cpuCount; j++) + { + if (numa_bitmask_isbitset(mask, j)) + { + if (currentGroupCpus == MaxCpusPerGroup) + { + g_groupToCpuCount[currentGroup] = MaxCpusPerGroup; + g_groupToCpuMask[currentGroup] = GetFullAffinityMask(MaxCpusPerGroup); + currentGroupCpus = 0; + currentGroup++; + } + g_cpuToAffinity[j].Node = i; + g_cpuToAffinity[j].Group = currentGroup; + g_cpuToAffinity[j].Number = currentGroupCpus; + g_groupAndIndexToCpu[currentGroup * MaxCpusPerGroup + currentGroupCpus] = j; + currentGroupCpus++; + } + } + + if (currentGroupCpus != 0) + { + g_groupToCpuCount[currentGroup] = currentGroupCpus; + g_groupToCpuMask[currentGroup] = GetFullAffinityMask(currentGroupCpus); + currentGroupCpus = 0; + currentGroup++; + } + } + + numa_free_cpumask(mask); + + g_highestNumaNode = numa_max_node(); + } + else +#endif // HAVE_NUMA_H + { + // No NUMA + g_cpuCount = GetNumberOfProcessors(); + g_groupCount = 1; + g_highestNumaNode = 0; + + AllocateLookupArrays(); + } + + return TRUE; +} + +/*++ +Function: + NUMASupportCleanup + +Cleanup of the NUMA support data structures +--*/ +VOID +NUMASupportCleanup() +{ + FreeLookupArrays(); +} + +/*++ +Function: + GetNumaHighestNodeNumber + +See MSDN doc. +--*/ +BOOL +PALAPI +GetNumaHighestNodeNumber( + OUT PULONG HighestNodeNumber +) +{ + PERF_ENTRY(GetNumaHighestNodeNumber); + ENTRY("GetNumaHighestNodeNumber(HighestNodeNumber=%p)\n", HighestNodeNumber); + *HighestNodeNumber = (ULONG)g_highestNumaNode; + + BOOL success = TRUE; + + LOGEXIT("GetNumaHighestNodeNumber returns BOOL %d\n", success); + PERF_EXIT(GetNumaHighestNodeNumber); + + return success; +} + +/*++ +Function: + GetNumaProcessorNodeEx + +See MSDN doc. +--*/ +BOOL +PALAPI +GetNumaProcessorNodeEx( + IN PPROCESSOR_NUMBER Processor, + OUT PUSHORT NodeNumber +) +{ + PERF_ENTRY(GetNumaProcessorNodeEx); + ENTRY("GetNumaProcessorNodeEx(Processor=%p, NodeNumber=%p)\n", Processor, NodeNumber); + + BOOL success = FALSE; + + if ((Processor->Group < g_groupCount) && + (Processor->Number < MaxCpusPerGroup) && + (Processor->Reserved == 0)) + { + short cpu = g_groupAndIndexToCpu[Processor->Group * MaxCpusPerGroup + Processor->Number]; + if (cpu != -1) + { + *NodeNumber = g_cpuToAffinity[cpu].Node; + success = TRUE; + } + } + + if (!success) + { + *NodeNumber = 0xffff; + SetLastError(ERROR_INVALID_PARAMETER); + } + + LOGEXIT("GetNumaProcessorNodeEx returns BOOL %d\n", success); + PERF_EXIT(GetNumaProcessorNodeEx); + + return success; +} + +/*++ +Function: + GetLogicalProcessorInformationEx + +See MSDN doc. +--*/ +BOOL +PALAPI +GetLogicalProcessorInformationEx( + IN LOGICAL_PROCESSOR_RELATIONSHIP RelationshipType, + OUT OPTIONAL PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX Buffer, + IN OUT PDWORD ReturnedLength +) +{ + PERF_ENTRY(GetLogicalProcessorInformationEx); + ENTRY("GetLogicalProcessorInformationEx(RelationshipType=%d, Buffer=%p, ReturnedLength=%p)\n", RelationshipType, Buffer, ReturnedLength); + + BOOL success = FALSE; + + if (RelationshipType == RelationGroup) + { + size_t requiredSize = __builtin_offsetof(SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX, Group); + requiredSize += __builtin_offsetof(GROUP_RELATIONSHIP, GroupInfo); + requiredSize += g_groupCount * sizeof(PROCESSOR_GROUP_INFO); + + if (*ReturnedLength >= requiredSize) + { + Buffer->Relationship = RelationGroup; + Buffer->Size = requiredSize; + Buffer->Group.MaximumGroupCount = g_groupCount; + Buffer->Group.ActiveGroupCount = g_groupCount; + for (int i = 0; i < g_groupCount; i++) + { + Buffer->Group.GroupInfo[i].MaximumProcessorCount = MaxCpusPerGroup; + Buffer->Group.GroupInfo[i].ActiveProcessorCount = g_groupToCpuCount[i]; + Buffer->Group.GroupInfo[i].ActiveProcessorMask = g_groupToCpuMask[i]; + } + + success = TRUE; + } + else + { + SetLastError(ERROR_INSUFFICIENT_BUFFER); + } + + *ReturnedLength = requiredSize; + } + else + { + // We only support the group relationship + SetLastError(ERROR_INVALID_PARAMETER); + } + + LOGEXIT("GetLogicalProcessorInformationEx returns BOOL %d\n", success); + PERF_EXIT(GetLogicalProcessorInformationEx); + + return success; +} + +/*++ +Function: + GetThreadGroupAffinityInternal + +Get the group affinity for the specified pthread +--*/ +BOOL +GetThreadGroupAffinityInternal( + IN pthread_t thread, + OUT PGROUP_AFFINITY GroupAffinity +) +{ + BOOL success = FALSE; + +#if HAVE_PTHREAD_GETAFFINITY_NP + cpu_set_t cpuSet; + + int st = pthread_getaffinity_np(thread, sizeof(cpu_set_t), &cpuSet); + + if (st == 0) + { + WORD group = NO_GROUP; + KAFFINITY mask = 0; + + for (int i = 0; i < g_cpuCount; i++) + { + if (CPU_ISSET(i, &cpuSet)) + { + WORD g = g_cpuToAffinity[i].Group; + // Unless the thread affinity was already set by SetThreadGroupAffinity, it is possible that + // the current thread has affinity with processors from multiple groups. So we report just the + // first group we find. + if (group == NO_GROUP || g == group) + { + group = g; + mask |= ((KAFFINITY)1) << g_cpuToAffinity[i].Number; + } + } + } + + GroupAffinity->Group = group; + GroupAffinity->Mask = mask; + success = TRUE; + } + else + { + SetLastError(ERROR_GEN_FAILURE); + } +#else // HAVE_PTHREAD_GETAFFINITY_NP + // There is no API to manage thread affinity, so let's return a group affinity + // with all the CPUs on the system. + GroupAffinity->Group = 0; + GroupAffinity->Mask = GetFullAffinityMask(g_cpuCount); + success = TRUE; +#endif // HAVE_PTHREAD_GETAFFINITY_NP + + return success; +} + +/*++ +Function: + GetThreadGroupAffinity + +See MSDN doc. +--*/ +BOOL +PALAPI +GetThreadGroupAffinity( + IN HANDLE hThread, + OUT PGROUP_AFFINITY GroupAffinity +) +{ + PERF_ENTRY(GetThreadGroupAffinity); + ENTRY("GetThreadGroupAffinity(hThread=%p, GroupAffinity=%p)\n", hThread, GroupAffinity); + + CPalThread *palThread = InternalGetCurrentThread(); + + BOOL success = GetThreadGroupAffinityInternal(palThread->GetPThreadSelf(), GroupAffinity); + + LOGEXIT("GetThreadGroupAffinity returns BOOL %d\n", success); + PERF_EXIT(GetThreadGroupAffinity); + + return success; +} + + +/*++ +Function: + SetThreadGroupAffinity + +See MSDN doc. +--*/ +BOOL +PALAPI +SetThreadGroupAffinity( + IN HANDLE hThread, + IN const GROUP_AFFINITY *GroupAffinity, + OUT OPTIONAL PGROUP_AFFINITY PreviousGroupAffinity +) +{ + PERF_ENTRY(SetThreadGroupAffinity); + ENTRY("SetThreadGroupAffinity(hThread=%p, GroupAffinity=%p, PreviousGroupAffinity=%p)\n", hThread, GroupAffinity, PreviousGroupAffinity); + + CPalThread *palThread = InternalGetCurrentThread(); + + pthread_t thread = palThread->GetPThreadSelf(); + + if (PreviousGroupAffinity != NULL) + { + GetThreadGroupAffinityInternal(thread, PreviousGroupAffinity); + } + +#if HAVE_PTHREAD_GETAFFINITY_NP + int groupStartIndex = GroupAffinity->Group * MaxCpusPerGroup; + KAFFINITY mask = 1; + cpu_set_t cpuSet; + CPU_ZERO(&cpuSet); + + for (int i = 0; i < MaxCpusPerGroup; i++, mask <<= 1) + { + if (GroupAffinity->Mask & mask) + { + int cpu = g_groupAndIndexToCpu[groupStartIndex + i]; + if (cpu != -1) + { + CPU_SET(cpu, &cpuSet); + } + } + } + + int st = pthread_setaffinity_np(thread, sizeof(cpu_set_t), &cpuSet); + + if (st == -1) + { + switch (errno) + { + case EINVAL: + // There is no processor in the mask that is allowed to execute the process + SetLastError(ERROR_INVALID_PARAMETER); + break; + case EPERM: + SetLastError(ERROR_ACCESS_DENIED); + break; + default: + SetLastError(ERROR_GEN_FAILURE); + break; + } + } + + BOOL success = (st == 0); +#else // HAVE_PTHREAD_GETAFFINITY_NP + // There is no API to manage thread affinity, so let's ignore the request + BOOL success = TRUE; +#endif // HAVE_PTHREAD_GETAFFINITY_NP + + LOGEXIT("SetThreadGroupAffinity returns BOOL %d\n", success); + PERF_EXIT(SetThreadGroupAffinity); + + return success; +} + +/*++ +Function: + GetCurrentProcessorNumberEx + +See MSDN doc. +--*/ +VOID +PALAPI +GetCurrentProcessorNumberEx( + OUT PPROCESSOR_NUMBER ProcNumber +) +{ + PERF_ENTRY(GetCurrentProcessorNumberEx); + ENTRY("GetCurrentProcessorNumberEx(ProcNumber=%p\n", ProcNumber); + + DWORD cpu = GetCurrentProcessorNumber(); + _ASSERTE(cpu < g_cpuCount); + ProcNumber->Group = g_cpuToAffinity[cpu].Group; + ProcNumber->Number = g_cpuToAffinity[cpu].Number; + + LOGEXIT("GetCurrentProcessorNumberEx\n"); + PERF_EXIT(GetCurrentProcessorNumberEx); +} + +/*++ +Function: + GetProcessAffinityMask + +See MSDN doc. +--*/ +BOOL +PALAPI +GetProcessAffinityMask( + IN HANDLE hProcess, + OUT PDWORD_PTR lpProcessAffinityMask, + OUT PDWORD_PTR lpSystemAffinityMask +) +{ + PERF_ENTRY(GetProcessAffinityMask); + ENTRY("GetProcessAffinityMask(hProcess=%p, lpProcessAffinityMask=%p, lpSystemAffinityMask=%p\n", hProcess, lpProcessAffinityMask, lpSystemAffinityMask); + + BOOL success = FALSE; + + if (hProcess == GetCurrentProcess()) + { + DWORD_PTR systemMask = GetFullAffinityMask(g_cpuCount); + +#if HAVE_SCHED_GETAFFINITY + int pid = getpid(); + cpu_set_t cpuSet; + int st = sched_getaffinity(pid, sizeof(cpu_set_t), &cpuSet); + if (st == 0) + { + WORD group = NO_GROUP; + DWORD_PTR processMask = 0; + + for (int i = 0; i < g_cpuCount; i++) + { + if (CPU_ISSET(i, &cpuSet)) + { + WORD g = g_cpuToAffinity[i].Group; + if (group == NO_GROUP || g == group) + { + group = g; + processMask |= ((DWORD_PTR)1) << g_cpuToAffinity[i].Number; + } + else + { + // The process has affinity in more than one group, in such case + // the function needs to return zero in both masks. + processMask = 0; + systemMask = 0; + group = NO_GROUP; + break; + } + } + } + + success = TRUE; + + *lpProcessAffinityMask = processMask; + *lpSystemAffinityMask = systemMask; + } + else + { + // We should not get any of the errors that the sched_getaffinity can return since none + // of them applies for the current thread, so this is an unexpected kind of failure. + SetLastError(ERROR_GEN_FAILURE); + } +#else // HAVE_SCHED_GETAFFINITY + // There is no API to manage thread affinity, so let's return both affinity masks + // with all the CPUs on the system set. + *lpSystemAffinityMask = systemMask; + *lpProcessAffinityMask = systemMask; + + success = TRUE; +#endif // HAVE_SCHED_GETAFFINITY + } + else + { + // PAL supports getting affinity mask for the current process only + SetLastError(ERROR_INVALID_PARAMETER); + } + + LOGEXIT("GetProcessAffinityMask returns BOOL %d\n", success); + PERF_EXIT(GetProcessAffinityMask); + + return success; +} + +/*++ +Function: + VirtualAllocExNuma + +See MSDN doc. +--*/ +LPVOID +PALAPI +VirtualAllocExNuma( + IN HANDLE hProcess, + IN OPTIONAL LPVOID lpAddress, + IN SIZE_T dwSize, + IN DWORD flAllocationType, + IN DWORD flProtect, + IN DWORD nndPreferred +) +{ + PERF_ENTRY(VirtualAllocExNuma); + ENTRY("VirtualAllocExNuma(hProcess=%p, lpAddress=%p, dwSize=%u, flAllocationType=%#x, flProtect=%#x, nndPreferred=%d\n", + hProcess, lpAddress, dwSize, flAllocationType, flProtect, nndPreferred); + + LPVOID result = NULL; + + if (hProcess == GetCurrentProcess()) + { + if (nndPreferred <= g_highestNumaNode) + { + result = VirtualAlloc(lpAddress, dwSize, flAllocationType, flProtect); +#if HAVE_NUMA_H + if (result != NULL) + { + int nodeMaskLength = (g_highestNumaNode + 1 + sizeof(unsigned long) - 1) / sizeof(unsigned long); + unsigned long *nodeMask = new unsigned long[nodeMaskLength]; + + memset(nodeMask, 0, nodeMaskLength); + + int index = nndPreferred / sizeof(unsigned long); + int mask = ((unsigned long)1) << (nndPreferred & (sizeof(unsigned long) - 1)); + nodeMask[index] = mask; + + int st = mbind(result, dwSize, MPOL_PREFERRED, nodeMask, g_highestNumaNode, 0); + free(nodeMask); + _ASSERTE(st == 0); + // If the mbind fails, we still return the allocated memory since the nndPreferred is just a hint + } +#endif // HAVE_NUMA_H + } + else + { + // The specified node number is larger than the maximum available one + SetLastError(ERROR_INVALID_PARAMETER); + } + } + else + { + // PAL supports allocating from the current process virtual space only + SetLastError(ERROR_INVALID_PARAMETER); + } + + LOGEXIT("VirtualAllocExNuma returns %p\n", result); + PERF_EXIT(VirtualAllocExNuma); + + return result; +} diff --git a/src/coreclr/src/utilcode/util.cpp b/src/coreclr/src/utilcode/util.cpp index c215a49..a8786de 100644 --- a/src/coreclr/src/utilcode/util.cpp +++ b/src/coreclr/src/utilcode/util.cpp @@ -728,7 +728,7 @@ BYTE * ClrVirtualAllocWithinRange(const BYTE *pMinAddr, //****************************************************************************** // NumaNodeInfo //****************************************************************************** -#if !defined(FEATURE_REDHAWK) && !defined(FEATURE_PAL) +#if !defined(FEATURE_REDHAWK) /*static*/ NumaNodeInfo::PGNHNN NumaNodeInfo::m_pGetNumaHighestNodeNumber = NULL; /*static*/ NumaNodeInfo::PVAExN NumaNodeInfo::m_pVirtualAllocExNuma = NULL; @@ -748,15 +748,19 @@ BYTE * ClrVirtualAllocWithinRange(const BYTE *pMinAddr, /*static*/ BOOL NumaNodeInfo::m_enableGCNumaAware = FALSE; /*static*/ BOOL NumaNodeInfo::InitNumaNodeInfoAPI() { -#if !defined(FEATURE_REDHAWK) && !defined(FEATURE_PAL) +#if !defined(FEATURE_REDHAWK) //check for numa support if multiple heaps are used ULONG highest = 0; if (CLRConfig::GetConfigValue(CLRConfig::UNSUPPORTED_GCNumaAware) == 0) return FALSE; +#ifndef FEATURE_PAL // check if required APIs are supported HMODULE hMod = GetModuleHandleW(WINDOWS_KERNEL32_DLLNAME_W); +#else + HMODULE hMod = GetCLRModule(); +#endif if (hMod == NULL) return FALSE; @@ -795,7 +799,7 @@ BYTE * ClrVirtualAllocWithinRange(const BYTE *pMinAddr, //****************************************************************************** // NumaNodeInfo //****************************************************************************** -#if !defined(FEATURE_REDHAWK) && !defined(FEATURE_PAL) +#if !defined(FEATURE_REDHAWK) /*static*/ CPUGroupInfo::PGLPIEx CPUGroupInfo::m_pGetLogicalProcessorInformationEx = NULL; /*static*/ CPUGroupInfo::PSTGA CPUGroupInfo::m_pSetThreadGroupAffinity = NULL; /*static*/ CPUGroupInfo::PGTGA CPUGroupInfo::m_pGetThreadGroupAffinity = NULL; @@ -848,8 +852,12 @@ BYTE * ClrVirtualAllocWithinRange(const BYTE *pMinAddr, } CONTRACTL_END; -#if !defined(FEATURE_REDHAWK) && defined(_TARGET_AMD64_) && !defined(FEATURE_PAL) +#if !defined(FEATURE_REDHAWK) && defined(_TARGET_AMD64_) +#ifndef FEATURE_PAL HMODULE hMod = GetModuleHandleW(WINDOWS_KERNEL32_DLLNAME_W); +#else + HMODULE hMod = GetCLRModule(); +#endif if (hMod == NULL) return FALSE; @@ -869,17 +877,19 @@ BYTE * ClrVirtualAllocWithinRange(const BYTE *pMinAddr, if (m_pGetCurrentProcessorNumberEx == NULL) return FALSE; +#ifndef FEATURE_PAL m_pGetSystemTimes = (PGST)GetProcAddress(hMod, "GetSystemTimes"); if (m_pGetSystemTimes == NULL) return FALSE; - +#endif + return TRUE; #else return FALSE; #endif } -#if !defined(FEATURE_REDHAWK) && defined(_TARGET_AMD64_) && !defined(FEATURE_PAL) +#if !defined(FEATURE_REDHAWK) && defined(_TARGET_AMD64_) // Calculate greatest common divisor DWORD GCD(DWORD u, DWORD v) { @@ -910,7 +920,7 @@ DWORD LCM(DWORD u, DWORD v) } CONTRACTL_END; -#if !defined(FEATURE_REDHAWK) && defined(_TARGET_AMD64_) && !defined(FEATURE_PAL) +#if !defined(FEATURE_REDHAWK) && defined(_TARGET_AMD64_) BYTE *bBuffer = NULL; SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *pSLPIEx = NULL; SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *pRecord = NULL; @@ -985,7 +995,7 @@ DWORD LCM(DWORD u, DWORD v) { LIMITED_METHOD_CONTRACT; -#if !defined(FEATURE_REDHAWK) && defined(_TARGET_AMD64_) && !defined(FEATURE_PAL) +#if !defined(FEATURE_REDHAWK) && defined(_TARGET_AMD64_) WORD begin = 0; WORD nr_proc = 0; @@ -1012,7 +1022,7 @@ DWORD LCM(DWORD u, DWORD v) } CONTRACTL_END; -#if !defined(FEATURE_REDHAWK) && defined(_TARGET_AMD64_) && !defined(FEATURE_PAL) +#if !defined(FEATURE_REDHAWK) && defined(_TARGET_AMD64_) BOOL enableGCCPUGroups = CLRConfig::GetConfigValue(CLRConfig::EXTERNAL_GCCpuGroup) != 0; BOOL threadUseAllCpuGroups = CLRConfig::GetConfigValue(CLRConfig::EXTERNAL_Thread_UseAllCpuGroups) != 0; @@ -1095,7 +1105,7 @@ retry: { LIMITED_METHOD_CONTRACT; -#if !defined(FEATURE_REDHAWK) && defined(_TARGET_AMD64_) && !defined(FEATURE_PAL) +#if !defined(FEATURE_REDHAWK) && defined(_TARGET_AMD64_) WORD bTemp = 0; WORD bDiff = processor_number - bTemp; @@ -1126,7 +1136,7 @@ retry: } CONTRACTL_END; -#if !defined(FEATURE_REDHAWK) && defined(_TARGET_AMD64_) && !defined(FEATURE_PAL) +#if !defined(FEATURE_REDHAWK) && defined(_TARGET_AMD64_) // m_enableGCCPUGroups and m_threadUseAllCpuGroups must be TRUE _ASSERTE(m_enableGCCPUGroups && m_threadUseAllCpuGroups); @@ -1147,7 +1157,7 @@ retry: #endif } -#if !defined(FEATURE_REDHAWK) && !defined(FEATURE_PAL) +#if !defined(FEATURE_REDHAWK) //Lock ThreadStore before calling this function, so that updates of weights/counts are consistent /*static*/ void CPUGroupInfo::ChooseCPUGroupAffinity(GROUP_AFFINITY *gf) { diff --git a/src/coreclr/src/vm/gcenv.os.cpp b/src/coreclr/src/vm/gcenv.os.cpp index d519c3c..5e820a6 100644 --- a/src/coreclr/src/vm/gcenv.os.cpp +++ b/src/coreclr/src/vm/gcenv.os.cpp @@ -329,11 +329,7 @@ bool GCToOSInterface::GetCurrentProcessAffinityMask(uintptr_t* processMask, uint { LIMITED_METHOD_CONTRACT; -#ifndef FEATURE_PAL return !!::GetProcessAffinityMask(GetCurrentProcess(), (PDWORD_PTR)processMask, (PDWORD_PTR)systemMask); -#else - return false; -#endif } // Get number of processors assigned to the current process