Re: rbuf starvation in the iwn driver

[Sverre Froyen <sverre%viewmark.com@localhost>]

On Wed April 7 2010 17:18:35 Thor Lancelot Simon wrote:
> On Wed, Apr 07, 2010 at 04:38:37PM -0600, Sverre Froyen wrote:
> > On Wed April 7 2010 15:54:49 Thor Lancelot Simon wrote:
> > > Doesn't this memory need to be dma-safe?
> >
> > It does. I assume it needs to be contiguous and properly aligned, which
> > malloc(9) seems to imply. Does it also need to be wired? I notice that
> > the malloc man page says it's deprecated and to use pool_cache(9) or
> > kmem(9) instead. kmem provides wired memory.
> 
> If you want memory you can DMA to/from, I think you need to use
> bus_dmamem_alloc().

Yes, reading the man page for bus_dmamap_load etc. I see that is required. Is 
this a requirement for NetBSD but not for OpenBSD? Look at this code fragment 
from the OpenBSD iwn driver:

                data->m = MCLGETI(NULL, M_DONTWAIT, NULL, IWN_RBUF_SIZE);
                if (data->m == NULL) {
                        printf("%s: could not allocate RX mbuf\n",
                            sc->sc_dev.dv_xname);
                        error = ENOBUFS;
                        goto fail;
                }

                error = bus_dmamap_load(sc->sc_dmat, data->map,
                    mtod(data->m, void *), IWN_RBUF_SIZE, NULL,
                    BUS_DMA_NOWAIT | BUS_DMA_READ);

MCLGETI returns an mbuf cluster where the mbuf data buffer is allocated like 
this (from uipc_mbuf.c):

                data->m->m_ext.ext_buf = pool_get(&mclpools[pi],
                    how == M_WAIT ? PR_WAITOK : 0);

This seems to imply that, at least for OpenBSD, pool_get returns an address 
that can be used by bus_dmamap_load etc.

I am probably missing something.

Thanks,
Sverre