On 14.08.25 18:10, Andrew Davis wrote:
> Hello all,
>
> This series makes it so the udmabuf will sync the backing buffer
> with the set of attached devices as required for DMA-BUFs when
> doing {begin,end}_cpu_access.
Yeah the reason why we didn't do that is that this doesn't even work 100%
reliable in theory. So this patchset here might make your use case work but is
a bit questionable in general.
udmabuf is about turning a file descriptor created by memfd_create() into a
DMA-buf. Mapping that memory can happen through the memfd as well and so it is
perfectly valid to skip the DMA-buf begin_access and end_access callbacks.
Additional to that when CONFIG_DMABUF_DEBUG is enabled the DMA-buf code mangles
the page addresses in the sg table to prevent importers from abusing it. That
makes dma_sync_sgtable_for_cpu() and dma_sync_sgtable_for_device() on the
exporter side crash.
That's the reason why DMA-buf heaps uses a copy of the sg table for calling
dma_sync_sgtable_for_cpu()/dma_sync_sgtable_for_device().
It's basically a hack and should be removed, but for this we need to change all
clients which is tons of work.
Regards,
Christian.
>
> Thanks
> Andrew
>
> Changes for v2:
> - fix attachment table use-after-free
> - rebased on v6.17-rc1
>
> Andrew Davis (3):
> udmabuf: Keep track current device mappings
> udmabuf: Sync buffer mappings for attached devices
> udmabuf: Use module_misc_device() to register this device
>
> drivers/dma-buf/udmabuf.c | 133 +++++++++++++++++++-------------------
> 1 file changed, 67 insertions(+), 66 deletions(-)
>
