nouveau/tu102: flush all pdbs on vmm flush

[ Upstream commit cb9c919364653eeafb49e7ff5cd32f1ad64063ac ]

This is a hack around a bug exposed with the GSP code, I'm not sure
what is happening exactly, but it appears some of our flushes don't
result in proper tlb invalidation for out BAR2 and we get a BAR2
fault from GSP and it all dies.

Signed-off-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Danilo Krummrich <dakr@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231130010852.4034774-1-airlied@gmail.com
Signed-off-by: Sasha Levin <sashal@kernel.org>

diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmtu102.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmtu102.c
index 6cb5eefa45e9aa321d4c71ae73ba34d93457c951..5a08458fe1b7f543b5fbeb02b2f43505f3aa295d 100644 (file)
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmtu102.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmtu102.c
@@ -31,7 +31,7 @@ tu102_vmm_flush(struct nvkm_vmm *vmm, int depth)
 
        type |= 0x00000001; /* PAGE_ALL */
        if (atomic_read(&vmm->engref[NVKM_SUBDEV_BAR]))
-               type |= 0x00000004; /* HUB_ONLY */
+               type |= 0x00000006; /* HUB_ONLY | ALL PDB (hack) */
 
        mutex_lock(&vmm->mmu->mutex);