On 12/14/2013 02:02 PM, Rob Clark wrote:
On Sat, Dec 14, 2013 at 7:47 AM, Thomas Hellstrom thomas@shipmail.org wrote:
On 12/14/2013 01:37 PM, Thierry Reding wrote:
On Thu, Dec 12, 2013 at 11:30:23PM +0100, Daniel Vetter wrote:
On Thu, Dec 12, 2013 at 8:34 PM, Thomas Hellstrom thellstrom@vmware.com wrote:
On 12/12/2013 03:36 PM, Thierry Reding wrote:
This is a simple test module that can be used to allocate, export and delete DMA-BUF objects. It can be used to test DMA-BUF sharing in systems that lack a real second driver.
Looks nice. I wonder whether this could be extended to create a "streaming" dma-buf from a user space mapping. That could be used as a generic way to implement streaming (user) buffer objects, rather than to add explicit support for those in, for example, TTM.
Atm there's no way to get gpus to unbind their dma-buf mappings, so their essentially pinned forever from first use on.
Shouldn't this work by simply calling the GEM_CLOSE IOCTL on the handle returned by drmPrimeFDToHandle()? I mean that should drop the last reference on the GEM object and cause it to be cleaned up (which should include detaching the DMA-BUF).
Actually, while the GEM prime implementation appears to pin an exported dma-buf on first attach, from the dma-buf documentation it seems sufficient to pin it on map or cpu access.
But what I assume Daniel is referring to is that there is no way for exporters to tell importers to force unmap() the dma-buf, so that it can be unpinned?
yeah, or some way for importers to opportunistically keep around a mapping rather than map/unmap on each use..
maybe we need something shrinker-ish for dmabuf?
Yes, I think that's needed for both the memory-shortage case where we want to unpin, and I guess it would desirable for iommu space management as well.
/Thomas
BR, -R