On Fri, Dec 20, 2024 at 09:24:59AM -0500, Marek Olšák wrote: > > > > * Modifiers must uniquely encode buffer layout. In other words, a buffer > > must > > * match only a single modifier. > > > > That sentence is misleading and impossible to meet. Specifications are > sometimes changed to reflect the overwhelming reality. Multiple modifiers > can represent identical layouts - they already do between vendors, between > generations of the same vendor, and accidentally even between chips of the > same generation. Modifiers have already become 64-bit structures of > bitfields with currently 2^16 possible modifiers for some vendors, and > possibly exceeding 100k for all vendors combined. Encoding all linear > constraints into the 64 bits is one option. It needs more thought, but > encoding at least some constraints in the modifier is not totally off the > table. Yeah uniqueness is not required, we just try to avoid it since it causes some pain. Worst case all drivers that care about working sharing just need to make sure they advertise all overlapping variants they support. > The semi-functional LINEAR modifier needs to go. The idea of modifiers is > that nobody should have to expose one that is unsupported to keep things > working for a subset of apps. If the LINEAR modifier is a requirement > everywhere because of apps, and even drivers that can't support it must > expose it, that's a problem. It causes GBM/EGL to fail to import a DMABUF > for a random reason and it can't be prevented without, say, looking at PCI > IDs. If that happened for any other API, it would be considered unusable. > We can either fix it (by replacing/deprecating/removing LINEAR) or abandon > modifiers and replace them with something that works. I think Daniel's forward compatibility plan is more solid than trying to deprecate LINEAR itself, that seems like too much an uphill battle to me. -Sima -- Simona Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
