Date: Thu, 15 Feb 96 16:21:02 GMT
From: Neil Jarvis <Neil DOT Jarvis AT proteon DOT com>
To: djgpp AT delorie DOT com
Subject: Re: how to allocate DMA pages ?
Reply-To: Neil DOT Jarvis AT proteon DOT com

>>>>> "Malcolm" == Malcolm Taylor <malcolm AT manawatu DOT gen DOT nz> writes:
>>>>> "Ulf" == Ulf Griesmann <ulfg AT groundhog DOT phy DOT nist DOT gov> writes:

    Ulf> Hello All,
    Ulf> The 8237 DMA controllers in a PC can only perform DMA
    Ulf> transfers to the lower 16MB of physical memory. This makes it
    Ulf> awkward to allocate buffers for DMA transfers on machines
    Ulf> with >16MB of memory because a simple 'malloc' is no longer
    Ulf> guaranteed to allocate DMAable memory.  Does anyone out there
    Ulf> know of a good strategy for allocating DMA buffers which
    Ulf> works on large memory machines ? (Will take source code...)

    Neil> It is actually worse than this - the 8237 DMA controller takes
    Neil> physical addresses, malloc returns virtual addresses. There
    Neil> is currently no mechanism to translate malloc'ed virtual
    Neil> addresses into physical addresses with which to program the
    Neil> DMA controller.

    Neil> Having said that, I came up with a solution that uses XMS to allocate
    Neil> and lock a region of memory above 1Mb - XMS will tell you
    Neil> the physical address of this region, allowing you to program
    Neil> DMA transfers. I then used DPMI to map a region of linear
    Neil> memory to this XMS region. The application can then access
    Neil> the data using the <sys/farptr.h> functions.

    Malcolm> There is actually an easier method than that (it has it's
    Malcolm> restrictions though). Under DPMI the first 1M of memory
    Malcolm> is always locked and has the linear addresses 0-1M. Just allocate
    Malcolm> some DOS memory from the DPMI provider and the get the
    Malcolm> linear address by segment*16+offset. This can be accessed
    Malcolm> with the far-ptr routines using the
    Malcolm> __go32_info_block.selector_for_linear_memory as the
    Malcolm> selector and the linear address as the offset. Usually
    Malcolm> you will want memory that starts on the DMA page
    Malcolm> boundaries, so just allocate the size+64K and figure from
    Malcolm> the address the start of the first page in the block.
    Malcolm> The only limitation of this that I know of is only a
    Malcolm> limited amount of memory can be allocated this way. This
    Malcolm> isn't normally a problem as the limit is high.

Agreed this is much simpler and I do use this when I can. My scheme
has a lot of management overhead.

Unfortunately my application requires the allocation of between 2 and
4Mb of DMA buffers. I just couldn't get them to fit into the first 1M
of memory :-)

-- Neil Jarvis, Proteon International, York R&D (Neil DOT Jarvis AT proteon DOT com)
--    Don't use commas, which aren't necessary.