[gdal-dev] Corrupted .shp error when reading shapefile from multiple threads

[Even Rouault]

Selon Martin Chapman <chapmanm at pixia.com>:

> Kedar,
>
> Think about it...how does GetNextFeature() know what feature is next?
> It's because somewhere inside the datasource handle there is probably a
> variable that is keeping track of what record is next.  If multiple
> threads are all using the same datasource handle then this variable is
> probably getting updated / read by different threads at the same time and
> then all hell will break loose.
>
> The answer to your problem is to open a separate datasource handle for
> every thread you have so there is no sharing across threads.   Right now
> your getZipCode() function is sharing a global datasource handle and the
> code is not synchronized.  If you synchronize this code then it will be no
> better than a single threaded application so if I was you I would either:
>
> 1. Open a datasource handle on each thread and pass it to the zipcode
> function from each thread ... or
>
> 2. Create a class that encapsulates the getZipCode() function and has a
> datasource member variable that opens the datasource on construction and
> then create a separate instance of this class for each thread on startup
> so each thread has its own copy of the class.   Or ....
>
> 3. See if there is a function on the handle that will get the next feature
> given an index (like GetFeature(int index)) that you pass into the
> function so there is no internal record keeping needed by the handle.
> This is a risky way to do it though because even though it will work there
> may be other global variables the handle is keeping so unless Frank can
> vouch for that method I would go with curtain number 2.

Martin, I concur with your points.

About 3), (using GetFeature(index)), this won't work either. Because
GetFeature() needs to seek in the .shp and .dbf files. So if 2 threads use
GetFeature() on the same layer object, which has a single file pointer, the
seeks may conflict with each other and you might end up reading the wrong
content. The only safe solution is to use a datasource object for each thread.