Re: use of malloc(9) for DMA buffers

[Michael Lorenz <macallan%netbsd.org@localhost>]

Hello,

On Aug 11, 2008, at 9:30 AM, Marco Trillo wrote:

> On Mon, Aug 11, 2008 at 2:40 PM, Michael Lorenz  
> <macallan%netbsd.org@localhost> wrote:
> > On Aug 11, 2008, at 8:17 AM, Marco Trillo wrote:
> > > I have a doubt regarding dbdma.c dbdma_alloc(). It uses malloc(9) to
> > > allocate the buffer for the DMA command descriptors:
> > >
> > >  122: dbdma_command_t        *
> > >  123: dbdma_alloc(size)
> > >  124:        int size;
> > >  125: {
> > >  126:        u_int buf;
> > >  127:
> > >  128:        buf = (u_int)malloc(size + 0x0f, M_DEVBUF, M_WAITOK);
> > >  129:        buf = (buf + 0x0f) & ~0x0f;
> > >  130:
> > >  131:        return (dbdma_command_t *)buf;
> > >  132: }
> > >
> > > Since these descriptors are directly accessed by hardware, wouldn't
> > > this cause problems if the buffer crosses a page boundary and so  
> > > it's
> > > not contiguous in physical memory? Or does this never happen?
> >
> > On macppc that's safe ( kernel virtual addresses are identical to  
> > physical
> > addresses ) but on most other platforms it would likely blow up in  
> > your
> > face.
>
> Hmm... I've done a test using buf = malloc(8192, M_DEVBUF) and  
> vtophys():
>
> buf = d0024000, vtophys(buf) = 2fb13000
> buf + NBPG = d0025000, vtophys(buf + NBPG) = 2fb12000

Err, forget what I said - that's only for stuff above 0x80000000. Not  
memory.

> Is the fact that malloc(9) returns page-aligned buffers even when
> allocating small quantities reliable for use in functions such as
> dbdma_alloc() ?

The malloc(9) man page says malloc 'returns data suitably aligned for  
storing any object'. I'd read that as page-aligned.

have fun
Michael