nvidia-drm: return -ENOSYS when PRIME sg_table is unavailable

[VaggelisGian]

Fixes the misleading error reported in #1302.

Problem

__nv_drm_gem_nvkms_memory_prime_get_sg_table() returns ERR_PTR(-ENOMEM) for NvKmsKapiMemory backed by video memory (pages_count == 0, no struct pages). The value propagates verbatim through dma_buf_map_attachment() to the importing driver: a cross-device PRIME import of such an object reports an out-of-memory condition that never happened. As reported in #1302, kwin_wayland treats this as real memory exhaustion and aborts instead of falling back or reporting an unsupported operation.

Change

One line: return ERR_PTR(-ENOSYS) instead of ERR_PTR(-ENOMEM) in that case.

-ENOSYS is what DRM core itself uses when a GEM object has no sg_table implementation:

• v5.11 commit d693def4fd1c ("drm: Remove obsolete GEM and PRIME callbacks from struct drm_driver"): WARN_ON(!obj->funcs->get_sg_table); return ERR_PTR(-ENOSYS);
• v6.4 commit 207395da5a97 ("drm/prime: reject DMA-BUF attach when get_sg_table is missing"): attach rejected with -ENOSYS.

The errno is the only change:

• No caller under kernel-open/nvidia-drm branches on the old value; DRM core forwards PTR_ERR unchanged.
• Real allocation failures on this path still return -ENOMEM from their own sites.
• pages_count == 0 occurs only for vidmem objects: sysmem objects whose page list cannot be built fail at object init (__nv_drm_nvkms_gem_obj_init).

Test Plan

Prepared on Windows without Linux kernel toolchain or NVIDIA GPU, so no build and no runtime test were possible; verification was static:

git diff --stat
 kernel-open/nvidia-drm/nvidia-drm-gem-nvkms-memory.c | 2 +-

git diff --check        # clean

Traced all four constructors of nvkms-memory GEM objects (dumb_create, import_nvkms_memory_ioctl, alloc_nvkms_memory_ioctl, prime_dup) to confirm pages_count == 0 implies vidmem; grepped all callers of nv_drm_gem_prime_get_sg_table for errno-specific branching (none). A hardware test on any multi-GPU system importing a vidmem-backed nvidia-drm buffer from another driver would exercise the changed path.

[CLAassistant]

All committers have signed the CLA.