MS RFC 24: Mapscript memory management

Date:

2006/12/3

Author:

Umberto Nicoletti

Contact:

umberto.nicoletti@gmail.com

Last Edited:

2006/12/31

Status:

Done

Version:

MapServer 5.0

Tracker:

https://github.com/MapServer/MapServer/issues/2032

Tracker 3.2:

https://github.com/MapServer/MapServer/issues/2442

Id:

$Id$

Overview

Memory management in SWIG wrappers has a tradition of being difficult and error prone. The programmer of the wrapper has to deal with memory that can be allocated and then freed in two separate environments: the hosting language such as Java, C# or Perl and the wrapped native code.

Most modern languages implement garbage collection, so that the developer does not have to care about memory management. The programming language tracks memory (or objects, really) allocations and when an object goes out of scope (it is no more reachable from the running program) it marks it as eligible for garbage collection. A background process once in a while wakes up and frees the memory associated with marked objects. For the details on GC see this wikipedia entry: https://en.wikipedia.org/wiki/Garbage_collection_(computer_science)

What happens in most cases is that some memory is allocated in, say, Java, then another pointer is pointed to it by invoking some wrapped method. Eventually the GC runs and frees the memory. As soon as the other pointers are dereferenced the hosting language will crash because of a segmentation fault error (in Unix terms).

MapServer SWIG wrappers suffer from issues with garbage collections for example in dynamically adding layers to a map.

The purpose of this RFC is to address these issues and provide a solution that can be implemented in time for the release of MapServer 5.0.

This RFC does not address thread safety.

Problem description

This section gives an overview (along with examples) of errors in MapScript memory management. Most of the examples will be in Java, but they apply to all other MapScripts too. They can be reproduced against the latest CVS code of mapserver-HEAD as of 31st December 2006.

Object equality/identity

Consider the following Java MapScript code:

mapObj map=new mapObj("my.map");
layerObj layer=new layerObj(null);
// add layer to the map
int index=map.insertLayer(layer, -1);
// set its name
layer.setName("Change me");

// fetch from map
layerObj newLayer=map.getLayer(index);
// they should be the same...
System.out.println(newLayer.getName()+"=="+layer.getName());
// and this should print true (it is a reference comparison)
System.out.println(newLayer==layer);

when executed will produce the following output:

null==Change me
false

This happens because the current implementation strategy copies the layer when it is inserted into the map. The Java reference is not re-pointed to the new copy and is therefore ‚disconnected‘ from the actual memory area. This currently happens for the most used insert methods (i.e. insertClass).

Early garbage collection

Objects created through MapScript can be garbage-collected „early“, when there are live objects still referencing them. See this example in Java:

mapObj map=new mapObj("data/emptymap.map");

layerObj layer=new layerObj(map);
layer.setName("Layer 0");
classObj clazz=new classObj(null);
clazz.setName("Clazz 0 NULL");
int pos=layer.insertClass(clazz, -1);

map=null;
layer=null;

// force garbage collection
for(int i=0;i<100;i++)
        System.gc();

clazz.getLayer().getMap().draw();
// Java crashes because memory has been freed!

and its Perl equivalent:

use mapscript;

$map = new mapscript::mapObj("../../tests/test.map");
$layer = new mapscript::layerObj($map);

print "Before first draw for $layer->{map}\n";
$layer->{map}->draw();
print "Map drawn, now undef map\n";
$map = undef;
$map1=$layer->{map}->draw();
// perl interpreter segfault

Dynamically populated layers, classes, etc

See the following bug reports:

https://github.com/MapServer/MapServer/issues/1400

https://github.com/MapServer/MapServer/issues/1743

https://github.com/MapServer/MapServer/issues/1841

Please note that this issue can be difficult to reproduce, credits go to Tamas for pointing it out.

Proposed implementation

To solve the problems shown at items 2.1, 2.2 and 2.3 this RFC proposes that:

  1. a reference counter is added to each MapServer data structure that can be manipulated directly with MapScript AND inserted/added to another object (i.e. a layerObj)

  2. setters, getters and insert methods increase the counter whenever a reference to a MapServer data structure is created by MapScript

  3. all MapScript objects be always owned by SWIG or, more exactly, by the wrapper objects (swigCMemOwn is always true)

  4. the MapServer free* methods be modified so that the the underlying data structure is freed only when the counter is zero and decrease it otherwise

  5. wrapper objects be augmented to maintain a reference to their parents only and prevent early garbage collection

  6. all arrays of structures (map->layers and so on) be changed to arrays of pointers to eliminate the need for a copy operation on insert methods

By preliminary discussion it has been decided to drop the requirement to fully support object equality. As a result 2.1 will be implemented so that only the first comparison returns true.

This RFC should be applicable (with the necessary modifications) to all MapScript languages. Examples will be given for Perl or Java because of the author familiarity with these languages.

The items above are described in more detail in the following subsections. Subsections 3.4 and 3.5 offer an implementation example for the layerObj class. Please note that in the following we limit the scope of our analysis to the classes/layer relationship.

Implement a reference counter

The MapScript objects implementing this rfc will get a new int member called refcount.

Mapscript will keep read-only access to the reference counter which is useful for debugging

The reference counter increment and decrement will be implemented by the following macros:

#define MS_REF_INCR(obj) obj->refcount++
#define MS_REF_DECR(obj) (--(obj->refcount))

An alternative could be to keep the reference counting in a global hashmap, keyed by memory address. This will eliminate the need for a change to every object but might present a greater impact on performance. In particular the hash function must be chosen carefully.

http://en.wikipedia.org/wiki/Hash_table

The example implementation proposed at the tracker bug #2032 adopts the first strategy.

3.1.1 RefCounting in the CGI

The rfc should be modified to propose that the USE_MAPSCRIPT requirement is dropped on the following motivations:

  1. with the advent of use_mapscript the compilation process will be different if the user needs to build MapScript rather than the cgi. While this should not be a big deal for individuals it might place some extra burden on those maintaining binary distros like ms4w or fwtools

  2. having refcounting enabled does not harm the cgi which should only experience a tini tiny (if any at all) performance drop because of the extra if and ++

  3. we can always opt to introduce USE_MAPSCRIPT later

  4. it simplifies the build process maintaining exactly like it’s been until now

  5. it simplifies the developer’s life because of one less define to maintain

The following is the text (now obsolete) documenting USE_MAPSCRIPT.

Since this member will not be used by the CGI this RFC proposes that:

  1. it is wrapped by a new define USE_MAPSCRIPT

  2. a new configure option is added that is called –enable-mapscript

  3. starting from 5.0 on it will be required to specify –enable-mapscript to build any MapScript (with the exception of php, maybe)

Example:

/* CLASS OBJECT - basic symbolization and classification information */
typedef struct class_obj{
#ifndef USE_MAPSCRIPT
        #ifdef SWIG
        %immutable;
        #endif /* SWIG */
        int refcount;
        #ifdef SWIG
        %mutable;
        #endif /* SWIG */
#endif
        expressionObj expression; /* the expression to be matched */
#endif

Votes count:

+1: Umberto, Tamas and Daniel

Add references to the MapScript wrapper object

The MapScript objects will be modified so that they keep a reference to the other MapScript objects they are added to, like the C struct already does. This object can be hereafter referred to as the parent object.

In example, in the case of the layerObj the layerObj class will be extended to contain

  • a reference to the mapObj that contains the layer

The purpose of these changes is that the hosting language knows of the relationships between these objects and we solve the early garbage collection problem. This is also important to avoid unexpected crashes in the hosting language when the layer dereferences its parent object (grep ‚layer->map‘ *.c reports 105 usages).

As stated earlier, it has been decided to drop the requirement to fully support object equality/identity.

This item will be implemented in a second phase, after the basic refcounting is in place. Also the rfc proposes to implement a parent not-null check for the layer operations that use the parent map reference.

Change arrays of structures in arrays of pointers

This change will occur at the C-level and is quite an undertaking. Initially and for the purpose of this RFC the size of the arrays will still be fixed as it is now. The modification of the code will be made a way that future RFC addressing dynamically-sized arrays can build upon.

The strategy is as follows:

  1. all accesses to array elements could be wrapped by a convenient C macro. In this way implementation is abstracted away from the client code

  2. struct definitions and free/init methods will be modified to implement the new feature

  3. the macro will be modified to suit the implementation at previous item

To implement item 1 we will use a perl pie like the following:

perl -pi -e "s/([mM])ap->layers\[(.*?)\]\]\./GET_LAYER(\1ap, \2\])->/g" *.c
perl -pi -e "s/([mM])ap->layers\[(.*?)\]\./GET_LAYER(\1ap, \2)->/g" *.c
perl -pi -e "s/([mM])ap->layers\[(.*?)\]\]/GET_LAYER(\1ap, \2\])/g" *.c
perl -pi -e "s/([mM])ap->layers\[(.*?)\]/GET_LAYER(\1ap, \2)/g" *.c
perl -pi -e "s/dst->layers\[(.*?)\]\./GET_LAYER(dst, \1)->/g" *.c
perl -pi -e "s/src->layers\[(.*?)\]\./GET_LAYER(src, \1)->/g" *.c
perl -pi -e "s/dst->layers\[(.*?)\]/GET_LAYER(dst, \1)/g" *.c
perl -pi -e "s/src->layers\[(.*?)\]/GET_LAYER(src, \1)/g" *.c

and this is the macro that will be used:

#define GET_LAYER(map, pos) map->layers[pos]

This will leave only very few occurrences (about 4 or 5) out that must be edited by hand. The same approach will be used with other arrays of structures (classes and styles).

This item has been implemented for the classes without using the GET_CLASS macro.

Keep parent references and C internal structure in sync

For the mechanism described at the previous item to work some functions in the MapScript objects and in the native code must be modified so that the MapScript objects and the C data structures stay in sync.

Returning to the example of the layerObj the insertClass, getClass and removeClass methods will have to be modified to keep the parent reference in sync with the C data structures. The current methods will be modified by using specific typemaps. The constructor will also need to be modified to store the reference to the mapObj. Eventually also the native code actually performing the copy-and-insert operation must be modified to only perform the insert operation (layerobject.c, line 52, function msInsertClass).

The MapScript API will be backward compatible.

Destructors obey the reference counter

The various free* methods must check the counter before freeing memory. This will be implemented in the native code as in the following example:

void msFreeMap(mapObj *map) {
        if(!map) return;
        if ( MS_REF_DECR(obj) > 0) return;
        // go on destroying the object as usual
}

This will ensure that children will not be freed in case the parent is garbage collected before them. To avoid that the parent attempts to double free some of its children:

  1. if the parent is destroyed before its children the parent should NULLify its pointer in the children

  2. viceversa, when the children are GC’ed earlier than their parent, they must NULLify their pointers in the parent (we are evaluating whether this could ever happen once the parent reference is in place)

Always give object ownership to SWIG

For the reference count to work all object ownership must be given to SWIG. This is quite different from how it is today. The change however is straightforward because SWIG will acquire object ownership by default and is only a matter of removing all %newobject statements in the swig interface files.

At the moment there are 58 %newobject statements.

C# must also change the following three lines in csmodule.i:

csharp/csmodule.i:375:  if (map != null) this.swigCMemOwn = false;$excode
csharp/csmodule.i:379:  if (layer != null) this.swigCMemOwn = false;$excode
csharp/csmodule.i:383:  if (parent_class != null) this.swigCMemOwn = false;$excode

or drop the constructor customization altogether.

JAVA: MapScript code example for layerObj

Tamas has proposed a more object oriented approach to this problem, which can be adopted for those languages that support OOrientation.

Code example for layerObj (javamodule.i):

/*
   Modified constructor according to:
   - cache population and sync, item 3.2
*/
%typemap(javaconstruct) layerObj(mapObj map) %{ {
        this($imcall, true);
        if (map != null) {
                /* Store parent reference, item 3.2 */
                this.map=map;
        }
}
%}

%typemap(javaout) int insertClass {
        // call the C API, which needs to be modified
        // so that the classObj is not copied anymore
        int actualIndex=$jnicall;
        /* Store parent reference, item 3.2 */
        classobj.layer=this;
        return actualIndex;
}

%typemap(javacode) layerObj %{
        /* parent reference, item 3.2 */
        mapObj map=null;
%}

%typemap(javacode) classObj %{
        /* parent reference, item 3.2 */
        layerObj layer=null;
%}

PERL: MapScript code example for layerObj

Code example for layerObj (plmodule.i):

%feature("shadow") layerObj(mapObj *map)
%{
sub new {
        my $pkg = shift;
        my $self = mapscriptc::new_layerObj(@_);
        bless $self, $pkg if defined($self);
        if (defined($_[0])) {
                # parent reference
                mapscript::LAYER_ADD_MAP_REF($self, $_[0]);
        }
        return $self;
}
%}

%feature("shadow") ~layerObj()
%{
sub DESTROY {
        return unless $_[0]->isa('HASH');
        my $self = tied(%{$_[0]});
        return unless defined $self;
        delete $ITERATORS{$self};
        mapscriptc::delete_layerObj($self);
        # remove parent reference
        delete $mapscript::LAYERMAP{$self};
}
%}

%perlcode %{
        %LAYERMAP={};
        sub LAYER_ADD_MAP_REF {
                my ($layer, $map)=@_;
                #print "MAP key=" . tied(%$layer) . "\n";
                $LAYERMAP{ tied(%$layer) }=$map;
        }

        ##################
        # DEBUGGING ONLY #
        ##################
        sub getLayerFrom {
                my ($map, $idx)=@_;
                return $MAPLAYERS{tied(%$map)}->{$idx};
        }

        sub getMAPLAYERS {
                return \%MAPLAYERS;
        }
%}

Implementation plan

It seems that for most MapScripts (java, csharp, perl and python) there is enough functionality in SWIG to implement the features described in this RFC. For ruby we’ll probably have to go a different route and implement the %trackobjects feature to achieve 3.2. As of Tcl I currently don’t know if it’s possible.

The two following section describe in detail the required SWIG-MapScript features (injection of code and constructor customization). Each language gets then a specific section to deal with its own characteristics

Checking SWIG-Mapscript capabilities: %javacode

Swig provides the equivalent of %javacode for the following languages:

  1. perl through %perlcode

  2. python through %pythoncode

  3. csharp through %cscode

  4. ruby does not have any ruby code at all in its wrapper objects

  5. swig-Tcl doesn’t support a %tclcode construct

This swig construct will be used to inject in the wrapper the definition for the references described in 3.2 and the wrapper methods.

Checking SWIG-MapScript capabilities: constructor customization

The %csconstruct used to wrap and customize the costructor of MapScript objects (item 3.3) is only available in csharp and java. It should be possible to simulate its behavior with %pythonprepend or %pythonprepend in python and with %perlcode or %feature(„shadow“) in perl.

This swig construct will be used to populate parent backreferences.

Java and C-Sharp

SWIG-Java and SWIG-CSharp share a common ground and are therefore very similar. The names of SWIG-Java constructs can be roughly translated into their C-Sharp equivalents by changing the java prefix into cs (i.e. javacode in cscode, javaconstruct in csconstruct and javaout in csout).

The implementation should follow exactly this RCF or be based on the proposal made by Tamas in the last discussion thread.

Bemerkung

as of Jan 2008 Tamas decided to write his own implementation of MS RFC 24: Mapscript memory management for C#

Perl

As in the example above most of the perl customization can be done with the use of the shadow construct.

The implementation will follow this RFC exactly.

Python

Python enjoys first-grade support in SWIG so the RFC should be implemented exactly as described.

Ruby

Needs investigation, probably we’ll have to use rb_gc_* functions to mark objects and prevent their garbage collection or use %trackobjects. Ruby will not implement this RFC as of item 3.2.

Tcl

Needs investigation and a Tcl expert. At the time of this writing Tcl will probably not implement this RFC as of item 3.2.

PHP

PHP MapScript does not rely on SWIG, but since most of the code is native it should be possible to adopt this RFC.

Implementation checklist

The following table will be used to track the implementation status of this RFC. There is a table for each MapScript object and when a language has implemented this RFC for a given object the maintainer will populate the relative cell with one of the following marks:

  • < if the child->parent reference has been implemented

  • a plus sign (+) if object has had the reference counter increased (under the CNT=Counter column)

  • a minus sign (-) if object has had the reference counter decreased (under the CNT=Counter column)

  • =1 if the reference counter has been reset to 1 (under the CNT=Counter column)

mapObj

Method

CNT

java

C#

perl

python

tcl

ruby

mapObj (constructor)

+

getSymbolset

getFontset

getLabelcache

getExtent

setSaved_extent

getSaved_extent

getImagecolor

getOutputformat

getReference

getScalebar

getLegend

getQuerymap

getWeb

getConfigoptions

insertLayer

+

<

<

<

removeLayer

-

getLayer

+

<

<

<

getLayerByName

+

<

<

<

prepareImage

setOutputFormat

draw

drawQuery

drawLegend

drawScalebar

drawReferenceMap

getLabel

nextLabel

getOutputFormatByName

appendOutputFormat

removeOutputFormat

clone

=1

layerObj

Method

CNT

java

C#

perl

python

tcl

ruby

layerObj (constructor)

+

<

<

<

getMap

getOffsite

getMetadata

cloneLayer

=1

insertClass

+

<

<

<

removeClass

-

nextShape

getFeature

getShape

getResult

getClass

+

<

<

<

getResults

addFeature

getExtent

classObj

Method

CNT

java

C#

perl

python

tcl

ruby

classObj (constructor)

+

<

<

<

getLabel

getMetadata

getLayer

clone

=1

createLegendIcon

drawLegendIcon

getStyle

+

insertStyle

+

removeStyle

-

webObj

Method

java

C#

perl

python

tcl

ruby

webObj (constructor)

getMap

getExtent

setExtent

getMetadata

styleObj

For styleObjs it is enough to disown them when they are fetched from the container object. It is not necessary to add the reference pointing back to the container object.

Method

CNT

java

C#

perl

python

tcl

ruby

styleObj (constructor)

+

setColor

getColor

getBackgroundcolor

setBackgroundcolor

getOutlinecolor

setOutlinecolor

getMincolor

setMincolor

getMaxcolor

setMaxcolor

clone

=1

labelObj

For labelObjs it is enough to disown them when they are fetched from the container object. It is not necessary to add the reference pointing back to the container object.

Method

java

C#

perl

python

tcl

ruby

labelObj (constructor)

getColor

setColor

setOutlinecolor

getOutlinecolor

setShadowcolor

getShadowcolor

getBackgroundcolor

setBackgroundcolor

setBackgroundshadowcolor

getBackgroundshadowcolor

hashTableObj

For hashTableObjs it is enough to disown them when they are fetched from the container object (i.e. a layerObj). It is not necessary to add the reference pointing back to the container object.

Method

java

C#

perl

python

tcl

ruby

hashTableObj (constructor)

colorObj

For colorObjs it is enough to disown them when they are fetched from the container object. It is not necessary to add the reference pointing back to the container object.

Method

java

C#

perl

python

tcl

ruby

colorObj (constructor)

imageObj

For imageObjs it is enough to own them when they are fetched from the container object. It is not necessary to add the reference pointing back to the container object.

Method

java

C#

perl

python

tcl

ruby

imageObj (constructor)

shapeObj

For shapeObjs it is enough to set ownership properly when they are fetched from or added to the container object. It is not necessary to add the reference pointing back to the container object.

Method

java

C#

perl

python

tcl

ruby

shapeObj (constructor)

getLine

getBounds

get

add

clone

copy

buffer

convexHull

boundary

getCentroid

Union

intersection

difference

symDifference

lineObj

For lineObjs it is enough to set ownership properly when they are fetched from or added to the container object. It is not necessary to add the reference pointing back to the container object.

Method

java

C#

perl

python

tcl

ruby

lineObj (constructor)

getPoint

get

add

set

pointObj

For pointObjs it is enough to set ownership properly when they are fetched from or added to the container object. It is not necessary to add the reference pointing back to the container object.

Method

java

C#

perl

python

tcl

ruby

pointObj (constructors)

toShape

symbolsetObj

For symbolsetObjs it is not necessary to add the reference pointing back to the map.

Method

CNT

java

C#

perl

python

tcl

ruby

symbolsetObj (constructor)

setSymbol(symbolObj value)

getSymbol

+

getSymbol(int i)

+

getSymbolByName

+

index

appendSymbol

+

removeSymbol

-

symbolObj

For symbolObjs it is enough to set ownership properly when they are fetched from or added to the container object. It is not necessary to add the reference pointing back to the container object.

Method

CNT

java

C#

perl

python

tcl

ruby

symbolObj (constructor)

+

setPoints

getPoints

getImage

setImage

Open issues

The following issues should be discussed after this RFC has been adopted/implemented.

Multiple owners for an object

It is the case of layer that is added to more than one map. This should be prohibited because the layer has only one parent reference. On insertion the code should check whether the C parent reference is not null and in that case raise a errorObj which will be transformed by the hosting language in an exception.

Workaround: the user should instead clone the object and the add the clone to the second map.

API Compatibility

It is a top priority of this RFC to preserve the investment made by MapScript users by maintaining the API backwards compatible in both terms of method signatures and usage (i.e. order of invocation, types, return codes, etc).

If there will be any exception to this rule it will have to be justified and be described under this section.

Status

RFC opened for comments on Jan, the 10th 2007 with a post on mapserver-dev.

RFC undergoing revision after discussion on mapserver-dev.

New revision published on MapServer web.

RFC adopted with voting closed on April 4, 2007:

+1: Umberto, Pericles S. Nacionales, Howard Butler, Stephen Woodbridge

+0: Frank Warmerdam

Activity

The bug #2032 will be used to track activity related to this RFC.

2/20/2007: attached patch that converts map->layers in array of pointers with dynamic allocation (item 3.3)

2/24/2007: vote proposed on mapserver-dev

4/4/2007: RFC Adopted, undergoing implementation

1/10/2008: RFC implementation completed