Skip to contentSkip to conversation

Userspace DMA

DISCUSSION Andrew Miklas 2 participants 2 messages Nov 16, 2003 – Nov 17, 2003

lore
Share & tools
lore
#0 lore raw

Hi,

Is there an accepted way of doing userspace DMA with Linux?

I've gone over a large number of posts, and all seem to recommend something different, and each method seems to involve certain drawbacks.

What I'd like to do is allow a process to be able to do operations like pci_alloc_consistent and pci_free_consistent. I'm trying to do this to allow a driver running under the Bochs emulator to interact with its (non-emulated) hardware using DMA.

I was originally going to do a pci_alloc_consistent and then mmap the space into the emulator's process. The process would then get the physical address of the memory through an ioctl, or I would leave the physical address at the start of the allocated buffer for the process to later retrieve.

However, it appears that remap_page_range won't work with addresses obtained from pci_alloc_consistent. I've also read that using the nopage callback can cause issues if the range I need to allocate is greater than one page.

Most of the approaches I've found seem to require that the hardware be capable of scatter-gather DMA (video4linux and bttv, for ex.). I've also seen one solution that had the driver do a pci_alloc_consistent on a certain ioctl, and then return the physical address of the buffer. The process would then mmap /dev/mem to access the buffer. However, someone warned that this method could have coherency issues.

Is there some other method that I should be looking at to accomplish this?

Thanks,

Andrew

#1 lore raw
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"Andrew" == Andrew Miklas <[email protected]> writes:

Andrew> Hi, Is there an accepted way of doing userspace DMA with Andrew> Linux?

I'm currently working on a framework for user-space device drivers, that allow suitably privileged user-space programs to set up and tear down dma regions -- essentially exposing pci_alloc_consistent and pci_{un}map_sg to userspace -- and also to handle interrupts in user space. The framework is implemented as a pair of filesystems, one for interrupts (actually, I hijack /proc/interrupts/*) and one for mappings --- mainly to allow easy cleanup when the userspace program dies.

It's not yet in a fit state to release, but contact me offline if you'd like a copy.

--
Dr Peter Chubb http://www.gelato.unsw.edu.au peterc AT gelato.unsw.edu.au The technical we do immediately, the political takes *forever*