optimized numa calls; better Linux support

CMakeLists.txt

@@ -1,6 +1,8 @@
 cmake_minimum_required(VERSION 3.0)
 project(libmimalloc C CXX)
 include("cmake/mimalloc-config-version.cmake")
+include("CheckIncludeFile")
+
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_CXX_STANDARD 17)
 
@@ -88,6 +90,16 @@ if(MI_USE_CXX MATCHES "ON")
   set_source_files_properties(src/static.c test/test-api.c PROPERTIES LANGUAGE CXX )
 endif()
 
+CHECK_INCLUDE_FILE("numaif.h" MI_HAVE_NUMA_H)
+if(MI_HAVE_NUMA_H)
+  list(APPEND mi_defines MI_HAS_NUMA)
+  list(APPEND mi_libraries numa)
+else()
+  if (NOT(WIN32))
+    message(WARNING "Compiling without using NUMA optimized allocation (on Linux, install libnuma-dev?)")
+  endif()
+endif()
+
 # Compiler flags
 if(CMAKE_C_COMPILER_ID MATCHES "AppleClang|Clang|GNU")
   list(APPEND mi_cflags -Wall -Wextra -Wno-unknown-pragmas)

include/mimalloc-internal.h

@@ -56,7 +56,7 @@ void       _mi_os_init(void);                                      // called fro
 void*      _mi_os_alloc(size_t size, mi_stats_t* stats);           // to allocate thread local data
 void       _mi_os_free(void* p, size_t size, mi_stats_t* stats);   // to free thread local data
 size_t     _mi_os_good_alloc_size(size_t size);
-int        _mi_os_numa_node(void);
+int        _mi_os_numa_node(mi_os_tld_t* tld);
 
 
 // memory.c

include/mimalloc-types.h

@@ -413,6 +413,7 @@ typedef struct mi_segments_tld_s {
 // OS thread local data
 typedef struct mi_os_tld_s {
   size_t              region_idx;   // start point for next allocation
+  int                 numa_node;    // numa node associated with this thread
   mi_stats_t*         stats;        // points to tld stats
 } mi_os_tld_t;
 

src/arena.c

@@ -267,7 +267,7 @@ void* _mi_arena_alloc_aligned(size_t size, size_t alignment,
   {
     size_t asize = _mi_align_up(size, MI_ARENA_BLOCK_SIZE);
     size_t bcount = asize / MI_ARENA_BLOCK_SIZE;
-    int numa_node = _mi_os_numa_node(); // current numa node
+    int numa_node = _mi_os_numa_node(tld); // current numa node
 
     mi_assert_internal(size <= bcount*MI_ARENA_BLOCK_SIZE);
     // try numa affine allocation

src/init.c

@@ -99,7 +99,7 @@ static mi_tld_t tld_main = {
   0, false,
   &_mi_heap_main,
   { { NULL, NULL }, {NULL ,NULL}, 0, 0, 0, 0, 0, 0, NULL, tld_main_stats }, // segments
-  { 0, tld_main_stats },       // os
+  { 0, -1, tld_main_stats },   // os
   { MI_STATS_NULL }            // stats
 };
 
@@ -218,6 +218,7 @@ static bool _mi_heap_init(void) {
     memset(tld, 0, sizeof(*tld));
     tld->heap_backing = heap;
     tld->segments.stats = &tld->stats;
+    tld->os.numa_node = -1;
     tld->os.stats = &tld->stats;
     _mi_heap_default = heap;
   }

src/memory.c

@@ -211,7 +211,7 @@ static bool mi_region_commit_blocks(mem_region_t* region, size_t idx, size_t bit
     if (mi_atomic_cas_strong(&region->info, info, 0)) {
       // update the region count
       region->arena_memid = arena_memid;
-      mi_atomic_write(&region->numa_node, _mi_os_numa_node() + 1);
+      mi_atomic_write(&region->numa_node, _mi_os_numa_node(tld) + 1);
       mi_atomic_increment(&regions_count);
     }
     else {
@@ -220,7 +220,7 @@ static bool mi_region_commit_blocks(mem_region_t* region, size_t idx, size_t bit
       for(size_t i = 1; i <= 4 && idx + i < MI_REGION_MAX; i++) {
         if (mi_atomic_cas_strong(&regions[idx+i].info, info, 0)) {
           regions[idx+i].arena_memid = arena_memid;
-          mi_atomic_write(&regions[idx+i].numa_node, _mi_os_numa_node() + 1);
+          mi_atomic_write(&regions[idx+i].numa_node, _mi_os_numa_node(tld) + 1);
           mi_atomic_increment(&regions_count);
           start = NULL;
           break;
@@ -430,7 +430,7 @@ void* _mi_mem_alloc_aligned(size_t size, size_t alignment, bool* commit, bool* l
   mi_assert_internal(blocks > 0 && blocks <= 8*MI_INTPTR_SIZE);
 
   // find a range of free blocks
-  int numa_node = _mi_os_numa_node();
+  int numa_node = _mi_os_numa_node(tld);
   void* p = NULL;
   size_t count = mi_atomic_read(&regions_count);
   size_t idx = tld->region_idx; // start at 0 to reuse low addresses? Or, use tld->region_idx to reduce contention?

src/os.c

@@ -884,7 +884,7 @@ void* _mi_os_alloc_huge_os_pages(size_t pages, int numa_node, size_t* psize) {
 }
 
 #ifdef WIN32
-static int mi_os_numa_nodex(void) {
+static int mi_os_numa_nodex() {
   PROCESSOR_NUMBER pnum;
   USHORT numa_node = 0;
   GetCurrentProcessorNumberEx(&pnum);
@@ -898,12 +898,42 @@ static int mi_os_numa_node_countx(void) {
   return (int)(numa_max + 1);
 }
 #elif MI_HAS_NUMA
-#include <numa.h>
+#include <dirent.h>
+#include <stdlib.h>
+#include <numaif.h>
 static int mi_os_numa_nodex(void) {
-  return numa_preferred();
+  #define MI_MAX_MASK (4)          // support at most 256 nodes
+  unsigned long mask[MI_MAX_MASK];
+  memset(mask,0,MI_MAX_MASK*sizeof(long));
+  int mode = 0;
+  long err = get_mempolicy(&mode, mask, MI_MAX_MASK*sizeof(long)*8, NULL, 0 /* thread policy */);
+  if (err != 0) return 0;
+  // find the lowest bit that is set
+  for(int i = 0; i < MI_MAX_MASK; i++) {
+    for(int j = 0; j < (int)(sizeof(long)*8); j++) {
+      if ((mask[i] & (1UL << j)) != 0) {
+        return (i*sizeof(long)*8 + j);
       }
+    }
+  }
+	return 0;
+}
+
 static int mi_os_numa_node_countx(void) {
-  return (numa_max_node() + 1);
+  DIR* d = opendir("/sys/devices/system/node");
+  if (d==NULL) return 1;
+  
+  struct dirent* de;
+  int max_node_num = 0;
+  while ((de = readdir(d)) != NULL) {
+  	int node_num;
+  	if (strncmp(de->d_name, "node", 4) == 0) {
+		  node_num = (int)strtol(de->d_name+4, NULL, 0);
+			if (max_node_num < node_num) max_node_num = node_num;
+    }
+  }
+  closedir(d);
+  return (max_node_num + 1);
 }
 #else
 static int mi_os_numa_nodex(void) {
@@ -915,18 +945,28 @@ static int mi_os_numa_node_countx(void) {
 #endif
 
 int _mi_os_numa_node_count(void) {
-  long ncount = mi_os_numa_node_countx();
+  static int numa_node_count = 0;
+  if (mi_unlikely(numa_node_count <= 0)) {
+    int ncount = mi_os_numa_node_countx();
     // never more than max numa node and at least 1
-  long nmax  = 1 + mi_option_get(mi_option_max_numa_node);
+    int nmax = 1 + (int)mi_option_get(mi_option_max_numa_node);
     if (ncount > nmax) ncount = nmax;
     if (ncount <= 0)   ncount = 1;
-  return ncount;
+    numa_node_count = ncount;
+  }
+  mi_assert_internal(numa_node_count >= 1);
+  return numa_node_count;
 }
 
-int _mi_os_numa_node(void) {
+int _mi_os_numa_node(mi_os_tld_t* tld) {
+  if (mi_unlikely(tld->numa_node < 0)) {
     int nnode = mi_os_numa_nodex();
     // never more than the node count
     int ncount = _mi_os_numa_node_count();
     if (nnode >= ncount) { nnode = nnode % ncount; }
-  return nnode;
+    if (nnode < 0) nnode = 0;
+    tld->numa_node = nnode;
+  }
+  mi_assert_internal(tld->numa_node >= 0 && tld->numa_node < _mi_os_numa_node_count());
+  return tld->numa_node;
 }