freedreno/kgsl: opt in to linear dma-buf sharing

[lfdevs]

Also see: lfdevs/xwayland#1

Summary

This teaches the freedreno KGSL path to expose a narrow dma-buf import/export contract for surfaceless/XWayland consumers: shared linear buffers only.

The KGSL stack can allocate shareable buffers through dma-heap/ION and import them into KGSL, but native KGSL allocations are not directly exportable as dma-bufs. This change makes that distinction explicit, constrains exported/imported modifiers to LINEAR when the KGSL dma-buf path is enabled, and shadows non-shared resources into shared linear resources before export.

The KGSL dma-buf sharing behavior is intentionally opt-in, so the normal chroot DRM/GBM path keeps its previous behavior.

Changes

• Add an explicit FD_FEATURE_KGSL winsys feature and store it as screen->is_kgsl.
• Add a separate screen->kgsl_dmabuf flag for the opt-in dma-buf sharing path.
• Enable the KGSL dma-buf path with either:
  • FD_KGSL_ENABLE_DMABUF=1
  • XWAYLAND_FORCE_KGSL_SURFACELESS=1
• Advertise dma-buf import/export caps for KGSL only when screen->kgsl_dmabuf is enabled.
• Restrict KGSL dma-buf modifier queries to DRM_FORMAT_MOD_LINEAR.
• Force KGSL shared/scanout resources to use a linear layout when the opt-in path is enabled.
• Reject KGSL dma-buf imports with non-linear modifiers on the opt-in path.
• Make KGSL BO dma-buf export fail for native KGSL BOs instead of duplicating an invalid import fd.
• Shadow non-shared KGSL resources into shared linear resources before dma-buf export, avoiding the noisy failed native export attempt.

Why

XWayland/KWin surfaceless paths need a stable dma-buf contract for KGSL-backed rendering. The safe contract currently is:

• exported buffers are dma-heap/ION backed,
• modifiers are LINEAR,
• native KGSL-only BOs are not exported directly.

At the same time, enabling these rules unconditionally can break the chroot DRM/GBM path. In testing, the modified Mesa path made glmark2 render black under DRM/GBM while still producing a score, whereas forcing XWayland's KGSL surfaceless backend restored correct output.

For that reason, this keeps the surfaceless-oriented KGSL dma-buf behavior behind an explicit opt-in and leaves existing DRM/GBM behavior unchanged by default.

Testing

Tested with KGSL surfaceless/XWayland flow on Adreno 830 and 730:

• surfaceless EGL KGSL renderer initializes
• dma-buf export/import roundtrip succeeds
• XWayland glamor imports KGSL dma-bufs
• glmark2 runs through XWayland with hardware rendering

Also tested the chroot DRM/GBM path on Adreno 730:

• default DRM/GBM path remains separate from the opt-in KGSL dma-buf sharing path
• XWAYLAND_FORCE_KGSL_SURFACELESS=1 EGL_PLATFORM=surfaceless enables the surfaceless flow

The shadow-before-export change also removes the expected-but-noisy first export failure:

failed to get dmabuf fd: -1