Call malloc_trim() regularly to improve deallocation behavior (#14707)

irr/include/IMeshBuffer.h

@@ -176,6 +176,32 @@ public:
 		}
 		return 0;
 	}
+
+	//! Calculate size of vertices and indices in memory
+	virtual size_t getSize() const
+	{
+		size_t ret = 0;
+		switch (getVertexType()) {
+			case video::EVT_STANDARD:
+				ret += sizeof(video::S3DVertex) * getVertexCount();
+				break;
+			case video::EVT_2TCOORDS:
+				ret += sizeof(video::S3DVertex2TCoords) * getVertexCount();
+				break;
+			case video::EVT_TANGENTS:
+				ret += sizeof(video::S3DVertexTangents) * getVertexCount();
+				break;
+		}
+		switch (getIndexType()) {
+			case video::EIT_16BIT:
+				ret += sizeof(u16) * getIndexCount();
+				break;
+			case video::EIT_32BIT:
+				ret += sizeof(u32) * getIndexCount();
+				break;
+		}
+		return ret;
+	}
 };
 
 } // end namespace scene

src/CMakeLists.txt

@@ -351,6 +351,12 @@ endif()
 include(CheckSymbolExists)
 check_symbol_exists(strlcpy "string.h" HAVE_STRLCPY)
 
+if(UNIX)
+	check_symbol_exists(malloc_trim "malloc.h" HAVE_MALLOC_TRIM)
+else()
+	set(HAVE_MALLOC_TRIM FALSE)
+endif()
+
 check_include_files(endian.h HAVE_ENDIAN_H)
 
 configure_file(

src/client/mapblock_mesh.cpp

@@ -800,11 +800,16 @@ MapBlockMesh::MapBlockMesh(Client *client, MeshMakeData *data, v3s16 camera_offs
 
 MapBlockMesh::~MapBlockMesh()
 {
+	size_t sz = 0;
 	for (scene::IMesh *m : m_mesh) {
+		for (u32 i = 0; i < m->getMeshBufferCount(); i++)
+			sz += m->getMeshBuffer(i)->getSize();
 		m->drop();
 	}
 	for (MinimapMapblock *block : m_minimap_mapblocks)
 		delete block;
+
+	porting::TrackFreedMemory(sz);
 }
 
 bool MapBlockMesh::animate(bool faraway, float time, int crack,

src/cmake_config.h.in

@@ -31,6 +31,7 @@
 #cmakedefine01 USE_REDIS
 #cmakedefine01 HAVE_ENDIAN_H
 #cmakedefine01 HAVE_STRLCPY
+#cmakedefine01 HAVE_MALLOC_TRIM
 #cmakedefine01 CURSES_HAVE_CURSES_H
 #cmakedefine01 CURSES_HAVE_NCURSES_H
 #cmakedefine01 CURSES_HAVE_NCURSES_NCURSES_H

src/mapblock.cpp

@@ -85,6 +85,7 @@ MapBlock::~MapBlock()
 #endif
 
 	delete[] data;
+	porting::TrackFreedMemory(sizeof(MapNode) * nodecount);
 }
 
 bool MapBlock::onObjectsActivation()

src/mapgen/mg_schematic.cpp

@@ -19,7 +19,6 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 */
 
 #include <fstream>
-#include <typeinfo>
 #include "mg_schematic.h"
 #include "server.h"
 #include "mapgen.h"
@@ -32,6 +31,7 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 #include "serialization.h"
 #include "filesys.h"
 #include "voxelalgorithms.h"
+#include "porting.h"
 
 ///////////////////////////////////////////////////////////////////////////////
 
@@ -80,6 +80,8 @@ Schematic::~Schematic()
 {
 	delete []schemdata;
 	delete []slice_probs;
+	u32 nodecount = size.X * size.Y * size.Z;
+	porting::TrackFreedMemory(nodecount * sizeof(MapNode));
 }
 
 ObjDef *Schematic::clone() const

src/porting.cpp

@@ -63,7 +63,11 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 	#include <FindDirectory.h>
 #endif
 
-#include "config.h"
+#if HAVE_MALLOC_TRIM
+	// glibc-only pretty much
+	#include <malloc.h>
+#endif
+
 #include "debug.h"
 #include "filesys.h"
 #include "log.h"
@@ -72,6 +76,7 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 #include <cstdarg>
 #include <cstdio>
 #include <signal.h>
+#include <atomic>
 
 #if !defined(SERVER) && defined(_WIN32)
 // On Windows export some driver-specific variables to encourage Minetest to be
@@ -903,4 +908,37 @@ double perf_freq = get_perf_freq();
 
 #endif
 
+#if HAVE_MALLOC_TRIM
+
+/*
+ * On Linux/glibc we found that after deallocating bigger chunks of data (esp. MapBlocks)
+ * the memory would not be given back to the OS and would stay at peak usage.
+ * This appears to be a combination of unfortunate allocation order/fragmentation
+ * and the fact that glibc does not call madvise(MADV_DONTNEED) on its own.
+ * Some other allocators were also affected, jemalloc and musl libc were not.
+ * read more: <https://forum.minetest.net/viewtopic.php?t=30509>
+ *
+ * As a workaround we track freed memory coarsely and call malloc_trim() once a
+ * certain amount is reached.
+ */
+
+static std::atomic<size_t> memory_freed;
+
+constexpr size_t MEMORY_TRIM_THRESHOLD = 128 * 1024 * 1024;
+
+void TrackFreedMemory(size_t amount)
+{
+	constexpr auto MO = std::memory_order_relaxed;
+	memory_freed.fetch_add(amount, MO);
+	if (memory_freed.load(MO) >= MEMORY_TRIM_THRESHOLD) {
+		// Synchronize call
+		if (memory_freed.exchange(0, MO) < MEMORY_TRIM_THRESHOLD)
+			return;
+		// Leave some headroom for future allocations
+		malloc_trim(1 * 1024 * 1024);
+	}
+}
+
+#endif
+
 } //namespace porting

src/porting.h

@@ -290,6 +290,17 @@ void osSpecificInit();
 // This attaches to the parents process console, or creates a new one if it doesnt exist.
 void attachOrCreateConsole();
 
+/**
+ * Call this after freeing bigger blocks of memory. Used on some platforms to
+ * properly give memory back to the OS.
+ * @param amount Number of bytes freed
+*/
+#if HAVE_MALLOC_TRIM
+void TrackFreedMemory(size_t amount);
+#else
+inline void TrackFreedMemory(size_t amount) { (void)amount; }
+#endif
+
 #ifdef _WIN32
 // Quotes an argument for use in a CreateProcess() commandline (not cmd.exe!!)
 std::string QuoteArgv(const std::string &arg);
@@ -298,6 +309,7 @@ std::string QuoteArgv(const std::string &arg);
 std::string ConvertError(DWORD error_code);
 #endif
 
+// snprintf wrapper
 int mt_snprintf(char *buf, const size_t buf_size, const char *fmt, ...);
 
 /**

src/voxel.cpp

@@ -23,6 +23,7 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 #include "nodedef.h"
 #include "util/directiontables.h"
 #include "util/timetaker.h"
+#include "porting.h"
 #include <cstring>  // memcpy, memset
 
 /*
@@ -40,12 +41,15 @@ VoxelManipulator::~VoxelManipulator()
 
 void VoxelManipulator::clear()
 {
-	// Reset area to volume=0
-	m_area = VoxelArea();
+	// Reset area to empty volume
+	VoxelArea old;
+	std::swap(m_area, old);
 	delete[] m_data;
 	m_data = nullptr;
 	delete[] m_flags;
 	m_flags = nullptr;
+
+	porting::TrackFreedMemory((sizeof(*m_data) + sizeof(*m_flags)) * old.getVolume());
 }
 
 void VoxelManipulator::print(std::ostream &o, const NodeDefManager *ndef,