On 12/10/08 09:00, Dana H. Myers wrote:
> Garrett D'Amore wrote:
>> On 12/10/08 08:18, Andrew Gallatin wrote:
>>> Paul Durrant wrote:
>>>
>>>> Andrew Gallatin wrote:
>>>>
>>>>> The above suggests that all my allocation requests should fail,
>>>>> since all my allocation requests are for 1520 bytes, and that's
>>>>> not a multiple of 4095...
>>>>>
>>>>>
>>>> Drew,
>>>>
>>>> I didn't notice the 'multiple' rule when I started using it; to be
>>>> honest I just went off what I found in the source code ;-) (The
>>>> relevant bits of code are in the rootnex driver, if I remember
>>>> correctly).
>>>>
>>>
>>> Let's hope nobody "fixes" the code to match the man page :)
>>>
>>
>> Be very very careful here. The different platforms have different
>> implementations of the rootnex stuff, as I recently discovered on
>> 32-bit platforms. (Apparently I discovered a strange failure mode
>> when align == 2, but minxfer == 4. It worked on amd64, but caused an
>> allocation failure on 32-bit systems. Bumping align up to 4 fixed
>> the problem in my driver, although to be fair the problem was
>> *probably* a bug in the DDI DMA code for the i386.)
>>
>> I would avoid depending too much on specific allocation behaviors
>> like this, if I had a choice. Better, I think to just allocate one
>> or more pages and manage the the division yourself internally.
> ... and thus push complication into individual drivers? The desired
> functionality is
> reasonable, and the DMA API is demonstrably opaque here, since
> multiple people
> are looking at the man page, the code and empirically trying to get
> the API
> to do something it seems like it ought to be able to do easily.
The desired functionality is "efficiency of packing" of allocations.
There may be good reasons why a specific platform won't pack things in
the same way. The behavior of how DMA allocates objects is not
documented, other than they will conform to *at least* the constraints
specified by the dma_attr. If some specific relationship (such as
putting multiple objects in a physical page) is desired or required,
then I believe the better approach is for the driver to manage the
object allocations within a larger DMA object.
Relying on platform-specific behaviors just seems like a bad idea, to me.
-- Garrett
_______________________________________________
driver-discuss mailing list
driver-discuss@opensolaris.org
http://mail.opensolaris.org/mailman/listinfo/driver-discuss
