On Tue, 22 Nov 2022 at 20:34, Jason Gunthorpe jgg@ziepe.ca wrote:
On Tue, Nov 22, 2022 at 08:29:05PM +0100, Daniel Vetter wrote:
You nuke all the ptes. Drivers that move have slightly more than a bare struct file, they also have a struct address_space so that invalidate_mapping_range() works.
Okay, this is one of the ways that this can be made to work correctly, as long as you never allow GUP/GUP_fast to succeed on the PTEs. (this was the DAX mistake)
Hence this patch, to enforce that no dma-buf exporter gets this wrong. Which some did, and then blamed bug reporters for the resulting splats :-) One of the things we've reverted was the ttm huge pte support, since that doesn't have the pmd_special flag (yet) and so would let gup_fast through. -Daniel