[gdal-dev] Under what (if any) conditions would GDALWriteBlock modify data buffer?

[Even Rouault]

Le 18/12/2023 à 21:15, Andrew C Aitchison a écrit :
> On Mon, 18 Dec 2023, Even Rouault via gdal-dev wrote:
>
>>
>> Le 18/12/2023 à 16:18, Andrew C Aitchison via gdal-dev a écrit :
>>> On Mon, 18 Dec 2023, Even Rouault via gdal-dev wrote:
>>>
>>>> Hi,
>>>>
>>>> interesting question. No easy answer as it is highly driver 
>>>> dependent. I believe that all drivers make sure that the content of 
>>>> the buffer before and after the call is the same, but some drivers 
>>>> might temporarily modify it, to do byte swapping. For example the 
>>>> HFA driver does that when run on big-endian hosts for non-Byte data 
>>>> type. I wouldn't exclude that for formats with MSB-byte ordering, a 
>>>> similar situation would happen for little endian hosts. So it is 
>>>> definitely not safe to use WriteBlock() with a buffer that would 
>>>> come from a read-only section of the calling program. Doc updated 
>>>> to reflect that in 
>>>> https://github.com/OSGeo/gdal/commit/ea321723dfc69ef3a422b1e3fe4dc9ee0832861d
>>>
>>> Did you mean to say
>>>   Note that even with eRWFlag==GF_Write, the content of the buffer
>>>   might be temporarily modified during the execution of this method
>>>   (and eventually restored back to its original content), so it is not
>>>   safe to use a buffer stored in a read-only section of the calling
>>>   program.
>>
>> Yes I meant that. I've less evidence in the RasterIO(GF_Write, ...) 
>> case than in the WriteBlock() case, but without checking all drivers, 
>> it is more prudent to assume that the buffer might be touched during 
>> RasterIO(GF_Write) by some drivers.
>
> In that case I don't understand
>    Note that even with eRWFlag==GF_Write...
> It suggests to me that the buffer may be modified for eRWFlag==GF_Write
> *as well as* for eRWFlag==GF_Read (which implies that it is *more* likely
> to happen for GF_Read) ?

Ah, perhaps non-idiomatic wording from mine. I struggle to make that 
clearer: if you've better wording to suggest, that's welcome

In the GF_Read case, modifying the buffer is obviously expected since it 
is then an output buffer, whereas in the GF_Write case it is a input 
buffer, and thus users could legitimately expect it to remain fully 
untouched, which isn't the case.


-- 
http://www.spatialys.com
My software is free, but my time generally not.