drm/nouveau/gr/tu102-: fix support for sw_bundle64_init

We weren't sending the high bits, though they're zero currently anyway.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf100.c

@@ -1389,9 +1389,15 @@ gf100_grctx_generate_main(struct gf100_gr_chan *chan)
 
 	if (gr->bundle_veid)
 		gf100_gr_icmd(gr, gr->bundle_veid);
-	if (grctx->sw_veid_bundle_init)
+	else
 		gf100_gr_icmd(gr, grctx->sw_veid_bundle_init);
 
+	if (gr->bundle64)
+		gf100_gr_icmd(gr, gr->bundle64);
+	else
+	if (grctx->sw_bundle64_init)
+		gf100_gr_icmd(gr, grctx->sw_bundle64_init);
+
 	if (grctx->r400088) grctx->r400088(gr, true);
 
 	nvkm_wr32(device, 0x404154, idle_timeout);

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf100.h

@@ -22,6 +22,7 @@ struct gf100_grctx_func {
 	const struct gf100_gr_pack *icmd;
 	const struct gf100_gr_pack *mthd;
 	const struct gf100_gr_pack *sw_veid_bundle_init;
+	const struct gf100_gr_pack *sw_bundle64_init;
 	/* bundle circular buffer */
 	void (*bundle)(struct gf100_gr_chan *, u64 addr, u32 size);
 	u32 bundle_size;

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxtu102.c

@@ -50,8 +50,8 @@ tu102_grctx_init_unknown_bundle_init_0[] = {
 };
 
 static const struct gf100_gr_pack
-tu102_grctx_pack_sw_veid_bundle_init[] = {
-	{ tu102_grctx_init_unknown_bundle_init_0 },
+tu102_grctx_pack_sw_bundle64_init[] = {
+	{ tu102_grctx_init_unknown_bundle_init_0, .type = 64 },
 	{}
 };
 
@@ -69,7 +69,7 @@ tu102_grctx = {
 	.unkn88c = gv100_grctx_unkn88c,
 	.main = gf100_grctx_generate_main,
 	.unkn = gv100_grctx_generate_unkn,
-	.sw_veid_bundle_init = tu102_grctx_pack_sw_veid_bundle_init,
+	.sw_bundle64_init = tu102_grctx_pack_sw_bundle64_init,
 	.bundle = gm107_grctx_generate_bundle,
 	.bundle_size = 0x3000,
 	.bundle_min_gpm_fifo_depth = 0x180,

drivers/gpu/drm/nouveau/nvkm/engine/gr/gf100.c

@@ -1097,7 +1097,7 @@ gf100_gr_icmd(struct gf100_gr *gr, const struct gf100_gr_pack *p)
 	struct nvkm_device *device = gr->base.engine.subdev.device;
 	const struct gf100_gr_pack *pack;
 	const struct gf100_gr_init *init;
-	u32 data = 0;
+	u64 data = 0;
 
 	nvkm_wr32(device, 0x400208, 0x80000000);
 
@@ -1107,6 +1107,8 @@ gf100_gr_icmd(struct gf100_gr *gr, const struct gf100_gr_pack *p)
 
 		if ((pack == p && init == p->init) || data != init->data) {
 			nvkm_wr32(device, 0x400204, init->data);
+			if (pack->type == 64)
+				nvkm_wr32(device, 0x40020c, upper_32_bits(init->data));
 			data = init->data;
 		}
 
@@ -2139,6 +2141,7 @@ gf100_gr_dtor(struct nvkm_gr *base)
 	nvkm_blob_dtor(&gr->gpccs.inst);
 	nvkm_blob_dtor(&gr->gpccs.data);
 
+	vfree(gr->bundle64);
 	vfree(gr->bundle_veid);
 	vfree(gr->bundle);
 	vfree(gr->method);

drivers/gpu/drm/nouveau/nvkm/engine/gr/gf100.h

@@ -91,6 +91,7 @@ struct gf100_gr {
 	struct gf100_gr_pack *sw_ctx;
 	struct gf100_gr_pack *bundle;
 	struct gf100_gr_pack *bundle_veid;
+	struct gf100_gr_pack *bundle64;
 	struct gf100_gr_pack *method;
 
 	struct gf100_gr_zbc_color zbc_color[NVKM_LTC_MAX_ZBC_COLOR_CNT];
@@ -286,7 +287,7 @@ struct gf100_gr_init {
 	u32 addr;
 	u8  count;
 	u32 pitch;
-	u32 data;
+	u64 data;
 };
 
 struct gf100_gr_pack {