Am 24.01.19 um 10:13 schrieb Christoph Hellwig:
On Wed, Jan 23, 2019 at 05:52:50PM +0100, Ard Biesheuvel wrote:
But my concern is that it seems likely that non-cache coherent implementations are relying on this hack as well. There must be a reason that this hack is only disabled for PowerPC platforms if they are cache coherent, for instance, and I suspect that that reason is that the hack is the only thing ensuring that the CPU mapping attributes match the device ones used for these buffers (the vmap()ed ones), whereas the rings and other consistent data structures are using the DMA API as intended, and thus getting uncached attributes in the correct way.
Dave, who added that commit is on Cc together with just about everyone involved in the review chain. Based on the previous explanation that idea what we might want an uncached mapping for some non-coherent architectures for this to work at all makes sense, but then again the way to create those mappings is entirely architecture specific, and also need a cache flushing before creating the mapping to work properly. So my working theory is that this code never properly worked on architectures without DMA coherent for PCIe at all, but I'd love to be corrected by concrete examples including an explanation of how it actually ends up working.
Cache coherency is mandatory for modern GPU operation.
Otherwise you can't implement a bunch of the requirements of the userspace APIs.
In other words the applications doesn't inform the driver that the GPU or the CPU is accessing data, it just does it and assumes that it works.
Regards, Christian.