[gdal-dev] [gdal 3.12beta] Normal chain of explicit destruction does not work with ComputedDataset

[Even Rouault]

Le 24/10/2025 à 13:33, Momtchil Momtchev a écrit :
>
> Is there any way to know whether this is a GDALComputedDataset if 
> everything I have is a GDALDataset pointer?
>
ds->GetRasterCount() > 0 && 
dynamic_cast<GDALComputedRasterBand*>(ds->GetRasterBand(1)) != nullptr 
maybe ?

but that assumes that the dataset is still alive If you do :

GDALDataset *dh = graylevel->GetDataset();
delete graylevel;

then dh is no longer valid after delete graylevel

There are other situations where the dataset returned by a band belongs 
to something else.

For example if you do 
ds->GetRasterBand(1)->GetOverview(0)->GetDataset(), this overview 
dataset is owned by ds and shouldn't be manually destroyed.

Only datasets returned by GDALDataset::Open() / GDALOpen() / 
GDALCreate() / GDALCreateCopy() should be manually destroyed.


> On 24/10/2025 12:55, Even Rouault wrote:
>>
>> Hi,
>>
>> you should not destroy the dataset of a GDALComputedRasterBand. The 
>> GDALComputedRasterBand destructor does that automatically.
>>
>> Even
>>
>> Le 24/10/2025 à 12:40, Momtchil Momtchev via gdal-dev a écrit :
>>>
>>> I have modified one of the algebra examples as follows:
>>>
>>> #include<gdal_priv.h>
>>> intmain() {
>>> GDALAllRegister();
>>> autopoDS=std::unique_ptr<GDALDataset>(GDALDataset::Open("test/data/a39se10.jpg"));
>>> auto&R=*(poDS->GetRasterBand(1));
>>> auto&G=*(poDS->GetRasterBand(2));
>>> auto&B=*(poDS->GetRasterBand(3));
>>> GDALRasterBand*graylevel=newGDALComputedRasterBand((0.299*R+0.587*G+0.114*B).AsType(GDT_Byte));
>>> autopoGTiffDrv=GetGDALDriverManager()->GetDriverByName("GTiff");
>>> std::unique_ptr<GDALDataset>(
>>> poGTiffDrv->CreateCopy("graylevel.tif", graylevel->GetDataset(), 
>>> false, nullptr, nullptr, nullptr))
>>> .reset();
>>> GDALDataset*dh=graylevel->GetDataset();
>>> deletegraylevel;
>>> dh->Close();
>>> return0;
>>> }
>>>
>>> I allocate graylevel on the heap, I obtain a pointer to the 
>>> GDALComputedDataset and I attempt to destroy first the 
>>> GDALComputedRasterBand then the GDALComputedDataset. This fails with 
>>> a crash in Close. Is this a bug or am I not supposed to do this? I 
>>> am doing it, because this is the usual chain of destruction events 
>>> for objects returned to JS. RasterBands hold a reference to their 
>>> Dataset in order to prevent the object from being GCed.
>>>
>>> -- 
>>> Momtchil Momtchev<momtchil at momtchev.com>
>>>
>>> _______________________________________________
>>> gdal-dev mailing list
>>> gdal-dev at lists.osgeo.org
>>> https://lists.osgeo.org/mailman/listinfo/gdal-dev
>> -- 
>> http://www.spatialys.com
>> My software is free, but my time generally not.
> -- 
> Momtchil Momtchev<momtchil at momtchev.com>

-- 
http://www.spatialys.com
My software is free, but my time generally not.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/gdal-dev/attachments/20251024/eb603738/attachment-0001.htm>