Re: [PATCH] generic device DMA (dma_pool update)

[David Brownell <david-b@pacbell.net>]

Adam J. Richter wrote:

 > 	The pci_pool_alloc  in sa1111-buf.c is more interesting.
 > alloc_safe_buffer is used to implement an unusual version of
 > pci_map_single.  It is unclear to me whether this approach is optimal.
 > I'll look into this more.

Darn if I didn't mention SA1111 up front!  Here's where avoiding
mapping APIs is a win, in favor of allocating dma buffers that
will work from the get go.  (That is, dma_addr_t can usefully push
up the call stack, while gfp_flags pushes down.)

USB device drivers like "hid" and "usblp" do that already; later,
more drivers could convert.

Of course, the ohci-sa1111.c codepaths are also relevant here.
They use pci_pools for TD allocation.  The fact that they use
SLAB_ATOMIC (you missed one!), and once didn't, bothers me;
luckily that became an easy fix a few months back.

And the dma_pool_patch I posted would make all this code morph
to using the slab allocator underneath.  The slab code would be
smarter about managing it than pci_pool, so it'd be better at
sharing that 1 MByte of DMA memory with other devices.

- Dave