[gdal-dev] Multithread deadlock

[Even Rouault]

> I don't think this comes from GDALPam, but rather from the lower level
> machinery of GDALRasterBand & GDALRasterBlock. As I said before, multi-
> threaded writing has never properly worked and I'm not sure if the issue
> you see is just an implementation bug or a design bug.

OK, after investigation, it was just an implementation bug. A lack of mutex in 
GetLockedBlock() that caused IReadBlock() & IWriteBlock() to be called from 2 
different threads on the same dataset.

Fixed per https://trac.osgeo.org/gdal/ticket/6665

I think the issue was mostly due to using the low level  
GetLockedBlock()/MarkDirty()/DropLock() interfaces, and that people using 
RasterIO(GF_Read/GF_Write) should be immune to that bug due to locks being 
already taken at that stage.

Anyway fixed now (with a new test case similar to yours). Thanks for the 
reproducers.

Even

-- 
Spatialys - Geospatial professional services
http://www.spatialys.com