On Fri Oct 9, 2026 at 7:44 PM CEST, Jim Cromie via B4 Relay wrote:
> From: Jim Cromie <[email protected]>
>
> Commit cb4c7603678c ("drm/nouveau/gsp/r570: Add support for
> INTERNAL_GCX_ENTRY_PREREQUISITE") hooked nvif_device_gcx_ready() into
> nouveau_pmops_runtime_suspend() to consult GSP before runtime suspend.
>
> On GPUs running without GSP-RM (e.g. Volta GV100, NvGspRm=0, or missing
> GSP firmware blobs), device->gsp is instantiated but gsp->rm remains
> NULL. nvkm_udevice_gcx_ready() checked "!gsp" instead of verifying
> nvkm_gsp_rm(gsp), allowing non-RM instances to invoke
> nvkm_gsp_gcx_ready(). nvkm_gsp_gcx_ready() then unconditionally
> dereferenced gsp->rm->api, causing a kernel oops:
>
> RIP: 0010:nvkm_gsp_gcx_ready+0x10/0x30 [nouveau]
> Code: ... 48 8b 87 70 0e 00 00 <48> 8b 40 18 ...
> RAX: 0000000000000000 RDI: ffff8c9555fca000
> Call Trace:
> nvkm_udevice_mthd+0xa5/0x150 [nouveau]
> nvkm_ioctl+0xcc/0x1d0 [nouveau]
> nvif_object_mthd+0x110/0x1e0 [nouveau]
> nvif_device_gcx_ready+0x36/0x60 [nouveau]
> nouveau_pmops_runtime_suspend+0x39/0x140 [nouveau]
>
> Instruction decode shows `mov rax, [rdi+0xe70]` loads gsp->rm (0x0),
> and `<48> 8b 40 18` (`mov rax, [rax+0x18]`) faults accessing rm->api.
>
> Fix by:
> 0. Checking !nvkm_gsp_rm(gsp) in nvkm_udevice_gcx_ready() so non-RM
> devices report GC6/GcOff ready and bypass the GSP call entirely.
> 1. Guarding gsp and gsp->rm in nvkm_gsp_gcx_ready() before inspecting
> gsp->rm->api->gsp->gcx_ready.
>
> Fixes: cb4c7603678c ("drm/nouveau/gsp/r570: Add support for
> INTERNAL_GCX_ENTRY_PREREQUISITE")
> Signed-off-by: Jim Cromie <[email protected]>
Isn't this fixed in commit 4d0e27041185 ("nouveau: check gsp->rm as well as gsp
pointer before gcx ready") already?
