Add Large pages support in GC

src/gc/CMakeLists.txt

@@ -84,7 +84,8 @@ if(WIN32)
   set (GC_LINK_LIBRARIES
     ${STATIC_MT_CRT_LIB}
     ${STATIC_MT_VCRT_LIB}
-    kernel32.lib)
+    kernel32.lib
+    advapi32.lib)
 else()
   set (GC_LINK_LIBRARIES)
 endif(WIN32)

src/gc/env/gcenv.os.h

@@ -275,6 +275,13 @@ class GCToOSInterface
     //  true if it has succeeded, false if it has failed
     static bool VirtualCommit(void *address, size_t size, uint16_t node = NUMA_NODE_UNDEFINED);
 
+    // Reserve and Commit virtual memory range for Large Pages
+    // Parameters:
+    //  size    - size of the virtual memory range
+    // Return:
+    //  Address of the allocated memory
+    static void* VirtualReserveAndCommitLargePages(size_t size);
+
     // Decomit virtual memory range.
     // Parameters:
     //  address - starting virtual address

src/gc/gc.cpp

@@ -2403,6 +2403,7 @@ void qsort1(uint8_t** low, uint8_t** high, unsigned int depth);
 #endif //USE_INTROSORT
 
 void* virtual_alloc (size_t size);
+void* virtual_alloc (size_t size, bool use_large_pages_p);
 void virtual_free (void* add, size_t size);
 
 /* per heap static initialization */
@@ -2818,6 +2819,7 @@ GCSpinLock gc_heap::gc_lock;
 
 size_t gc_heap::eph_gen_starts_size = 0;
 heap_segment* gc_heap::segment_standby_list;
+size_t        gc_heap::use_large_pages_p = 0;
 size_t        gc_heap::last_gc_index = 0;
 #ifdef SEG_MAPPING_TABLE
 size_t        gc_heap::min_segment_size = 0;
@@ -4263,7 +4265,7 @@ typedef struct
 
 initial_memory_details memory_details;
 
-BOOL reserve_initial_memory (size_t normal_size, size_t large_size, size_t num_heaps)
+BOOL reserve_initial_memory (size_t normal_size, size_t large_size, size_t num_heaps, bool use_large_pages_p)
 {
     BOOL reserve_success = FALSE;
 
@@ -4304,7 +4306,7 @@ BOOL reserve_initial_memory (size_t normal_size, size_t large_size, size_t num_h
 
     size_t requestedMemory = memory_details.block_count * (normal_size + large_size);
 
-    uint8_t* allatonce_block = (uint8_t*)virtual_alloc (requestedMemory);
+    uint8_t* allatonce_block = (uint8_t*)virtual_alloc (requestedMemory, use_large_pages_p);
     if (allatonce_block)
     {
         g_gc_lowest_address =  allatonce_block;
@@ -4324,10 +4326,10 @@ BOOL reserve_initial_memory (size_t normal_size, size_t large_size, size_t num_h
         // try to allocate 2 blocks
         uint8_t* b1 = 0;
         uint8_t* b2 = 0;
-        b1 = (uint8_t*)virtual_alloc (memory_details.block_count * normal_size);
+        b1 = (uint8_t*)virtual_alloc (memory_details.block_count * normal_size, use_large_pages_p);
         if (b1)
         {
-            b2 = (uint8_t*)virtual_alloc (memory_details.block_count * large_size);
+            b2 = (uint8_t*)virtual_alloc (memory_details.block_count * large_size, use_large_pages_p);
             if (b2)
             {
                 memory_details.allocation_pattern = initial_memory_details::TWO_STAGE;
@@ -4360,7 +4362,7 @@ BOOL reserve_initial_memory (size_t normal_size, size_t large_size, size_t num_h
                                      memory_details.block_size_normal :
                                      memory_details.block_size_large);
                 current_block->memory_base =
-                    (uint8_t*)virtual_alloc (block_size);
+                    (uint8_t*)virtual_alloc (block_size, use_large_pages_p);
                 if (current_block->memory_base == 0)
                 {
                     // Free the blocks that we've allocated so far
@@ -4468,6 +4470,11 @@ heap_segment* get_initial_segment (size_t size, int h_number)
 }
 
 void* virtual_alloc (size_t size)
+{
+    return virtual_alloc(size, false);
+}
+
+void* virtual_alloc (size_t size, bool use_large_pages_p)
 {
     size_t requested_size = size;
 
@@ -4488,7 +4495,8 @@ void* virtual_alloc (size_t size)
         flags = VirtualReserveFlags::WriteWatch;
     }
 #endif // !FEATURE_USE_SOFTWARE_WRITE_WATCH_FOR_GC_HEAP
-    void* prgmem = GCToOSInterface::VirtualReserve (requested_size, card_size * card_word_width, flags);
+
+    void* prgmem = use_large_pages_p ? GCToOSInterface::VirtualReserveAndCommitLargePages(requested_size) : GCToOSInterface::VirtualReserve(requested_size, card_size * card_word_width, flags);
     void *aligned_mem = prgmem;
 
     // We don't want (prgmem + size) to be right at the end of the address space 
@@ -5466,9 +5474,10 @@ bool gc_heap::virtual_commit (void* address, size_t size, int h_number, bool* ha
     }
 
     // If it's a valid heap number it means it's commiting for memory on the GC heap.
-    bool commit_succeeded_p = ((h_number >= 0) ? 
-        virtual_alloc_commit_for_heap (address, size, h_number) : 
-        GCToOSInterface::VirtualCommit(address, size));
+    // In addition if large pages is enabled, we set commit_succeeded_p to true because memory is already committed.
+    bool commit_succeeded_p = ((h_number >= 0) ? (use_large_pages_p ? true :
+                              virtual_alloc_commit_for_heap (address, size, h_number)) :
+                              GCToOSInterface::VirtualCommit(address, size));
 
     if (!commit_succeeded_p && heap_hard_limit)
     {
@@ -9219,7 +9228,7 @@ heap_segment* gc_heap::make_heap_segment (uint8_t* new_pages, size_t size, int h
     heap_segment_mem (new_segment) = start;
     heap_segment_used (new_segment) = start;
     heap_segment_reserved (new_segment) = new_pages + size;
-    heap_segment_committed (new_segment) = new_pages + initial_commit;
+    heap_segment_committed (new_segment) = (use_large_pages_p ? heap_segment_reserved(new_segment) : (new_pages + initial_commit));
     init_heap_segment (new_segment);
     dprintf (2, ("Creating heap segment %Ix", (size_t)new_segment));
     return new_segment;
@@ -9310,6 +9319,8 @@ void gc_heap::reset_heap_segment_pages (heap_segment* seg)
 void gc_heap::decommit_heap_segment_pages (heap_segment* seg,
                                            size_t extra_space)
 {
+    if (use_large_pages_p)
+        return;
     uint8_t*  page_start = align_on_page (heap_segment_allocated(seg));
     size_t size = heap_segment_committed (seg) - page_start;
     extra_space = align_on_page (extra_space);
@@ -10019,12 +10030,15 @@ HRESULT gc_heap::initialize_gc (size_t segment_size,
     block_count = 1;
 #endif //MULTIPLE_HEAPS
 
+    use_large_pages_p = false;
+
     if (heap_hard_limit)
     {
         check_commit_cs.Initialize();
+        use_large_pages_p = GCConfig::GetGCLargePages();
     }
 
-    if (!reserve_initial_memory(segment_size,heap_size,block_count))
+    if (!reserve_initial_memory(segment_size,heap_size,block_count,use_large_pages_p))
         return E_OUTOFMEMORY;
 
 #ifdef CARD_BUNDLE

src/gc/gcconfig.h

@@ -74,6 +74,7 @@ class GCConfigStringHolder
       "Specifies the name of the GC config log file")                                          \
   BOOL_CONFIG(GCNumaAware,   "GCNumaAware", true, "Enables numa allocations in the GC")        \
   BOOL_CONFIG(GCCpuGroup,    "GCCpuGroup", false, "Enables CPU groups in the GC")              \
+  BOOL_CONFIG(GCLargePages,  "GCLargePages", false, "Enables using Large Pages in the GC")     \
   INT_CONFIG(HeapVerifyLevel, "HeapVerify", HEAPVERIFY_NONE,                                   \
       "When set verifies the integrity of the managed heap on entry and exit of each GC")      \
   INT_CONFIG(LOHCompactionMode, "GCLOHCompact", 0, "Specifies the LOH compaction mode")        \

src/gc/gcpriv.h

@@ -3145,6 +3145,10 @@ class gc_heap
     PER_HEAP_ISOLATED
     size_t current_total_committed_gc_own;
 
+    // This is if large pages should be used.
+    PER_HEAP_ISOLATED
+    size_t use_large_pages_p;
+
     PER_HEAP_ISOLATED
     size_t last_gc_index;
 

src/gc/sample/CMakeLists.txt

@@ -24,6 +24,14 @@ set(SOURCES
     ../softwarewritewatch.cpp
 )
 
+if(WIN32)
+  set (GC_LINK_LIBRARIES
+    ${STATIC_MT_CRT_LIB}
+    ${STATIC_MT_VCRT_LIB}
+    kernel32.lib
+    advapi32.lib)
+endif(WIN32)
+
 if(WIN32)
     list(APPEND SOURCES
         ../windows/gcenv.windows.cpp)
@@ -36,3 +44,7 @@ endif()
 _add_executable(gcsample
     ${SOURCES}
 )
+
+if(WIN32)
+    target_link_libraries(gcsample ${GC_LINK_LIBRARIES})
+endif()

src/gc/unix/config.h.in

@@ -9,6 +9,7 @@
 #cmakedefine01 HAVE_SYS_MMAN_H
 #cmakedefine01 HAVE_PTHREAD_THREADID_NP
 #cmakedefine01 HAVE_PTHREAD_GETTHREADID_NP
+#cmakedefine01 HAVE_MAP_HUGETLB
 #cmakedefine01 HAVE_SCHED_GETCPU
 #cmakedefine01 HAVE_NUMA_H
 #cmakedefine01 HAVE_VM_ALLOCATE

src/gc/unix/configure.cmake

@@ -24,6 +24,15 @@ check_cxx_source_compiles("
     }
     " HAVE_PTHREAD_GETTHREADID_NP)
 
+check_cxx_source_compiles("
+    #include <sys/mman.h>
+
+    int main()
+    {
+        return MAP_HUGETLB;
+    }
+    " HAVE_MAP_HUGETLB)
+
 check_cxx_source_runs("
     #include <sched.h>
 

src/gc/unix/gcenv.unix.cpp

@@ -379,7 +379,7 @@ void GCToOSInterface::YieldThread(uint32_t switchCount)
 //  flags     - flags to control special settings like write watching
 // Return:
 //  Starting virtual address of the reserved range
-void* GCToOSInterface::VirtualReserve(size_t size, size_t alignment, uint32_t flags)
+static void* VirtualReserveInner(size_t size, size_t alignment, uint32_t flags, uint32_t hugePagesFlag = 0)
 {
     assert(!(flags & VirtualReserveFlags::WriteWatch) && "WriteWatch not supported on Unix");
     if (alignment == 0)
@@ -388,7 +388,7 @@ void* GCToOSInterface::VirtualReserve(size_t size, size_t alignment, uint32_t fl
     }
 
     size_t alignedSize = size + (alignment - OS_PAGE_SIZE);
-    void * pRetVal = mmap(nullptr, alignedSize, PROT_NONE, MAP_ANON | MAP_PRIVATE, -1, 0);
+    void * pRetVal = mmap(nullptr, alignedSize, PROT_NONE, MAP_ANON | MAP_PRIVATE | hugePagesFlag, -1, 0);
 
     if (pRetVal != NULL)
     {
@@ -413,6 +413,18 @@ void* GCToOSInterface::VirtualReserve(size_t size, size_t alignment, uint32_t fl
     return pRetVal;
 }
 
+// Reserve virtual memory range.
+// Parameters:
+//  size      - size of the virtual memory range
+//  alignment - requested memory alignment, 0 means no specific alignment requested
+//  flags     - flags to control special settings like write watching
+// Return:
+//  Starting virtual address of the reserved range
+void* GCToOSInterface::VirtualReserve(size_t size, size_t alignment, uint32_t flags)
+{
+    return VirtualReserveInner(size, alignment, flags);
+}
+
 // Release virtual memory range previously reserved using VirtualReserve
 // Parameters:
 //  address - starting virtual address
@@ -426,6 +438,28 @@ bool GCToOSInterface::VirtualRelease(void* address, size_t size)
     return (ret == 0);
 }
 
+// Commit virtual memory range.
+// Parameters:
+//  size      - size of the virtual memory range
+// Return:
+//  Starting virtual address of the committed range
+void* GCToOSInterface::VirtualReserveAndCommitLargePages(size_t size)
+{
+#if HAVE_MAP_HUGETLB
+    uint32_t largePagesFlag = MAP_HUGETLB;
+#else
+    uint32_t largePagesFlag = 0;
+#endif
+
+    void* pRetVal = VirtualReserveInner(size, OS_PAGE_SIZE, 0, largePagesFlag);
+    if (VirtualCommit(pRetVal, size, NUMA_NODE_UNDEFINED))
+    {
+        return pRetVal;
+    }
+
+    return nullptr;
+}
+
 // Commit virtual memory range. It must be part of a range reserved using VirtualReserve.
 // Parameters:
 //  address - starting virtual address

src/gc/windows/gcenv.windows.cpp

@@ -27,6 +27,8 @@ static size_t g_RestrictedPhysicalMemoryLimit = (size_t)UINTPTR_MAX;
 // memory on the machine/in the container, we need to restrict by the VM.
 static bool g_UseRestrictedVirtualMemory = false;
 
+static bool g_SeLockMemoryPrivilegeAcquired = false;
+
 static AffinitySet g_processAffinitySet;
 
 typedef BOOL (WINAPI *PIS_PROCESS_IN_JOB)(HANDLE processHandle, HANDLE jobHandle, BOOL* result);
@@ -114,6 +116,42 @@ DWORD LCM(DWORD u, DWORD v)
 }
 #endif
 
+bool InitLargePagesPrivilege()
+{
+    TOKEN_PRIVILEGES tp;
+    LUID luid;
+    if (!LookupPrivilegeValueW(nullptr, SE_LOCK_MEMORY_NAME, &luid))
+    {
+        return false;
+    }
+
+    tp.PrivilegeCount = 1;
+    tp.Privileges[0].Luid = luid;
+    tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
+
+    HANDLE token;
+    if (!OpenProcessToken(::GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES, &token))
+    {
+        return false;
+    }
+
+    BOOL retVal = AdjustTokenPrivileges(token, FALSE, &tp, 0, nullptr, 0);
+    DWORD gls = GetLastError();
+    CloseHandle(token);
+
+    if (!retVal)
+    {
+        return false;
+    }
+
+    if (gls != 0)
+    {
+        return false;
+    }
+
+    return true;
+}
+
 bool InitCPUGroupInfoArray()
 {
 #if (defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_))
@@ -699,6 +737,31 @@ bool GCToOSInterface::VirtualRelease(void* address, size_t size)
     return !!::VirtualFree(address, 0, MEM_RELEASE);
 }
 
+// Commit virtual memory range.
+// Parameters:
+//  size      - size of the virtual memory range
+// Return:
+//  Starting virtual address of the committed range
+void* GCToOSInterface::VirtualReserveAndCommitLargePages(size_t size)
+{
+    void* pRetVal = nullptr;
+
+    if (!g_SeLockMemoryPrivilegeAcquired)
+    {
+        if (!InitLargePagesPrivilege())
+        {
+            return nullptr;
+        }
+
+        g_SeLockMemoryPrivilegeAcquired = true;
+    }
+
+    SIZE_T largePageMinimum = GetLargePageMinimum();
+    size = (size + (largePageMinimum - 1)) & ~(largePageMinimum - 1);
+
+    return ::VirtualAlloc(nullptr, size, MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES, PAGE_READWRITE);
+}
+
 // Commit virtual memory range. It must be part of a range reserved using VirtualReserve.
 // Parameters:
 //  address - starting virtual address

src/inc/clrconfigvalues.h

@@ -330,6 +330,7 @@ RETAIL_CONFIG_STRING_INFO(EXTERNAL_GCName, W("GCName"), "")
 RETAIL_CONFIG_DWORD_INFO_DIRECT_ACCESS(EXTERNAL_GCHeapHardLimit, W("GCHeapHardLimit"), "Specifies the maximum commit size for the GC heap")
 RETAIL_CONFIG_DWORD_INFO_DIRECT_ACCESS(EXTERNAL_GCHeapHardLimitPercent, W("GCHeapHardLimitPercent"), "Specifies the GC heap usage as a percentage of the total memory")
 RETAIL_CONFIG_STRING_INFO(EXTERNAL_GCHeapAffinitizeRanges, W("GCHeapAffinitizeRanges"), "Specifies list of processors for Server GC threads. The format is a comma separated list of processor numbers or ranges of processor numbers. Example: 1,3,5,7-9,12")
+RETAIL_CONFIG_DWORD_INFO_DIRECT_ACCESS(EXTERNAL_GCLargePages, W("GCLargePages"), "Specifies whether large pages should be used when a heap hard limit is set")
 
 ///
 /// IBC

src/pal/inc/pal.h

@@ -2558,6 +2558,7 @@ SetErrorMode(
 #define MEM_MAPPED                      0x40000
 #define MEM_TOP_DOWN                    0x100000
 #define MEM_WRITE_WATCH                 0x200000
+#define MEM_LARGE_PAGES                 0x20000000
 #define MEM_RESERVE_EXECUTABLE          0x40000000 // reserve memory using executable memory allocator
 
 PALIMPORT